content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{sec:intro}
Spatial filtering techniques are a widely used means to spatially focus on a target source by exploiting spatial information of a wave field which is sampled by several sensors at different positions in space.
When spatial filtering techniques are applied to a robot audition scenario, i.e., when the microphones are mounted on a humanoid robot's head, the influence of the head on the sound field has to be taken into account by the beamformer design in order to obtain a satisfying spatial filtering performance. To this end, \acp{HRTF}\footnote{In the context of this work, \acp{HRTF} only model the direct propagation path between a source and a microphone mounted on a humanoid robot's head, but no reverberation components.} can be incorporated into the beamformer design as steering vectors, see, e.g., \cite{Pedersen_ICASSP_2004,Maazaoui_EURASIP_2012,Maazaoui_IWSSIP_2012}.
In \cite{lnt2009-22}, Mabande et al.~proposed a \ac{RLSFI} beamformer design which allows the user to directly control the tradeoff between the beamformer's spatial selectivity and its robustness. Recently, we extended this design to an \ac{HRTF}-based \ac{RLSFI} beamformer design by following the approach described above \cite{lnt2015-31}. Despite all advantages of the beamformer designs in \cite{lnt2009-22, lnt2015-31}, a clear disadvantage is that whenever the beamformer is steered to another direction, a new optimization problem has to be solved which makes it unattractive for real-time processing. To overcome this limitation, Mabande et al.~proposed a \ac{RLSFIP} beamformer design \cite{lnt2010-46} as extension of \cite{lnt2009-22}, which allows for a flexible steering of the beamformer.
In this work, we extend the \ac{HRTF}-based \ac{RLSFI} beamformer design \cite{lnt2015-31} to the concept of polynomial beamforming in order to allow for a flexible steering of the \ac{HRTF}-based beamformer in a robot audition scenario.
The remainder of this article is structured as follows: In Section~\ref{sec:HRTFbasedRobustPolynomialbeamforming}, the \ac{HRTF}-based \ac{RLSFIP} beamformer design is introduced. Then, an evaluation of the new \ac{HRTF}-based polynomial beamformer design is presented in Section~\ref{sec:Evaluation}. Finally, conclusions and an outlook to future work are given in Section~\ref{sec:Conclusion}.
\section{HRTF-based robust polynomial beamforming}
\label{sec:HRTFbasedRobustPolynomialbeamforming}
\subsection{Concept of polynomial beamforming}
\label{subsec:conceptPolynomialBeamforming}
In Fig.~\ref{fig:polynomialFSB}, the block diagram of a \ac{PFSB}, as presented in \cite{lnt2010-46, Kajala_ICASSP_2001, lnt2014-74}, is illustrated. It consists of a beamforming stage containing $P+1$ \acp{FSU}, followed by a \ac{PPF}. The output signal $y_{p}[k]$ of the $p$-th \ac{FSU} at time instant $k$ is obtained by convolving the microphone signals $x_n[k], \, n \in \{0, \ldots, N-1\}$ with the \ac{FSU}'s \ac{FIR} filters $\bb{w}_{n,p} = [w_{np,0}, \ldots, w_{np,L-1}]^\text{T}$ of length $L$, followed by a summation over all $N$ channels. Operator $(\cdot)^\text{T}$ represents the transpose of vectors or matrices, respectively.
In the \ac{PPF}, the output $y_{D}[k]$ of the \ac{PFSB} is obtained by weighting the output of each \ac{FSU} by a factor $D^{p}$ and summing them up:
\begin{equation}
y_{D}[k] = y_{0}[k] + Dy_{1}[k] + D^{2}y_{2}[k] + \ldots + D^{P}y_{P}[k].
\label{eq:y_PFSB}
\end{equation}
Hence, the output signal of each \ac{FSU} can be seen as one coefficient of a polynomial of order $P$ with variable $D$.
The advantage of a \ac{PFSB} is that the steering of the main beam is accomplished by simply changing the scalar value $D$, whereas the \ac{FIR} filters of the \acp{FSU} can be designed beforehand and remain fixed during runtime. A more detailed explanation of how the steering is controlled by $D$ is given in Section~\ref{subsec:HRTFbasedRLSFIPbeamforming}.
\begin{figure}[t]
\centering
\scriptsize
\psfrag{x0}[cl][cl]{$x_{0}[k]$}
\psfrag{x1}[cl][cl]{$x_{1}[k]$}
\psfrag{xN}[cl][cl]{$x_{N-1}[k]$}
\psfrag{y0}[cl][cl]{$y_{0}[k]$}
\psfrag{y1}[cl][cl]{$y_{1}[k]$}
\psfrag{yP}[cl][cl]{$y_{P}[k]$}
\psfrag{yD}[cl][cl]{$y_{D}[k]$}
\psfrag{D}[cl][cl]{$D$}
\psfrag{w00}[c][c]{$\bb{w}_{0,0}$}
\psfrag{w10}[c][c]{$\bb{w}_{1,0}$}
\psfrag{wN0}[c][c]{$\bb{w}_{N-1,0}$}
\psfrag{w01}[c][c]{$\bb{w}_{0,1}$}
\psfrag{w11}[c][c]{$\bb{w}_{1,1}$}
\psfrag{wN1}[c][c]{$\bb{w}_{N-1,1}$}
\psfrag{w0P}[c][c]{$\bb{w}_{0,P}$}
\psfrag{w1P}[c][c]{$\bb{w}_{1,P}$}
\psfrag{wNP}[c][c]{$\bb{w}_{N-1,P}$}
\psfrag{p1}[cl][cl]{
\psfrag{p2}[cl][cl]{
\psfrag{pP}[cl][cl]{
\psfrag{+}[c][c]{$+$}
\psfrag{F}[c][c]{FSUs}
\psfrag{PPF}[c][c]{PPF}
\includegraphics[scale = .95]{Illustration_polynomialFSB.eps}
\caption{Illustration of a polynomial filter-and-sum beamformer after \cite{lnt2010-46}.}
\label{fig:polynomialFSB}
\vspace{-6mm}
\end{figure}
The beamformer response of the \ac{PFSB} is given as \cite{lnt2010-46}:
\begin{equation}
B_{D}(\omega, \phi, \theta) = \sum\limits_{p=0}^{P} D^{p} \sum\limits_{n=0}^{N-1} W_{n,p}(\omega) g_{n}(\omega, \phi, \theta),
\label{eq:BFResponse_PBF}
\end{equation}
where $W_{n,p}(\omega) = \sum_{l=0}^{L-1} w_{np,l} e^{-j \omega l}$ is the \ac{DTFT} representation of $\mathbf{w}_{n,p}$, and $g_{n}(\omega, \phi, \theta)$ is the sensor response of the $n$-th microphone to a plane wave with frequency $\omega$ traveling in the direction $(\phi, \theta)$. Variables $\phi$ and $\theta$ denote azimuth and elevation angle, and are measured with respect to the positive x-axis and the positive z-axis, respectively, as in \cite{VanTrees:2004}.
\subsection{HRTF-based robust least-squares frequency-invariant polynomial beamforming}
\label{subsec:HRTFbasedRLSFIPbeamforming}
The main goal of the proposed \ac{HRTF}-based \ac{RLSFIP} beamformer design is to jointly approximate $I$ desired beamformer responses $\hat{B}_{D_{i}}(\omega, \phi, \theta)$, each with a different \ac{PLD} $(\phi_{i},\theta_{i}), \, i = 0, \ldots, I-1$, by the actual beamformer response $B_{D_{i}}(\omega, \phi, \theta)$, where $D_{i} = (\phi_{i}-90)/90$, in the \ac{LS} sense. Hence, $D_{i}$ lies in the interval $-1 \leq D_{i} \leq 1$, where, for example, $D=0$ and $D=-1$ steer the main beam towards $\phi=90^\circ$ and $\phi=0^\circ$, respectively. For values of $D$ which do not correspond to one of the \acp{PLD}, the \ac{PPF} will interpolate between them, as expressed in (\ref{eq:y_PFSB}). In this work, we apply polynomial beamforming only in the horizontal dimension. Thus, $D_{i}$ only depends on the azimuth angle, whereas $\theta_{i}$ is constant for all \acp{PLD}. The extension to two-dimensional beam steering is an aspect of future work.
In addition to the \ac{LS} approximation, a distortionless response constraint and a constraint on the \ac{WNG} is imposed on each of the $I$ \acp{PLD}. The approximation is carried out for a discrete set of $Q$ frequencies $\omega_{q}, \, q \in \{0, \ldots, Q-1\}$ and $M$ look directions $(\phi_{m},\theta_{m}), \, m \in \{0, \ldots, M-1\}$ (where, in this work, $\theta_{m}$ remains fixed) in order to obtain a numerical solution. Hence, the optimization problem of the \ac{HRTF}-based \ac{RLSFIP} beamformer design can be expressed as:
\begin{equation}
\argmin\limits_{\bb{w}_\mrm{f}(\omega_{q})} \sum\limits_{i=0}^{I-1} \Vert \bb{G}(\omega_{q}) \bb{D}_{i} \bb{w}_\mrm{f}(\omega_{q}) - \bb{\hat{b}}_{i} \Vert_{2}^{2},
\label{eq:OP_LSApproximation}
\end{equation}
subject to $I$ constraints on the corresponding WNG and response in the desired look direction, respectively:
\begin{align}
\frac{ |\bb{a}^\text{T}_{i}(\omega_{q}) \bb{D}_{i} \bb{w}_\mrm{f}(\omega_{q})|^{2}}{\Vert \bb{D}_{i}\bb{w}_\mrm{f}(\omega_{q}) \Vert_{2}^{2}} & \ge \gamma > 0, \quad \bb{a}^\text{T}_{i}(\omega_{q}) \bb{D}_{i} \bb{w}_\mrm{f}(\omega_{q}) = 1,\nonumber \\
& \forall i = 0, \ldots, I-1.
\label{eq:OP_Constraints}
\end{align}
where $\displaystyle \bb{\hat{b}}_{i} = [\hat{B}_{D_i}(\phi_{0},\theta_{0}), \ldots, \hat{B}_{D_i}(\phi_{M-1},\theta_{M-1})]^\text{T}$ is a vector of dimension $M \times 1$ containing the $i$-th desired response for all $M$ angles, matrix $\displaystyle [\bb{G}(\omega_{q})]_{mn} = g_{n}(\omega_{q}, \phi_{m}, \theta_{m})$, vector $\bb{a}_{i}(\omega_{q})=[g_{0}(\omega_{q}, \phi_{i}, \theta_{i}), \ldots, g_{N-1}(\omega_{q}, \phi_{i}, \theta_{i})]^\text{T}$ is the steering vector which contains the sensor responses for the $i$-th \ac{PLD} $(\phi_{i},\theta_{i})$, and vector $\displaystyle \bb{w}_\mathrm{f}(\omega_{q}) = [W_{0,0}(\omega_{q}), \ldots,$ $W_{N-1,P}(\omega_{q})]^\text{T}$ of dimension $N(P+1) \times 1$ contains all filter coefficients. Furthermore, $\bb{D}_{i}=\bb{I}_{N} \otimes [D^{0}_{i}, \ldots, D^{P}_{i}]$ is an $N \times N(P+1)$ matrix, where $\bb{I}_{N}$ is an $N \times N$ identity matrix and $\otimes$ denotes the Kronecker product. Operator $\Vert \cdot \Vert_{2}$ denotes the Euclidean norm of a vector.
The optimization problem in (\ref{eq:OP_LSApproximation}), (\ref{eq:OP_Constraints}) can be interpreted as follows: Equation (\ref{eq:OP_LSApproximation}) describes the \ac{LS} approximation of the $I$ desired responses $\hat{B}_{D_{i}}(\omega_{q}, \phi_{m}, \theta_{m})$ by the actual beamformer response.
The first part of (\ref{eq:OP_Constraints}) represents the \ac{WNG} constraint which is imposed on each of the $I$ \acp{PLD}. $\gamma$ is the lower bound on the \ac{WNG} and has to be defined by the user. Hence, the user has the possibility to directly control the beamformer's robustness against small random errors like sensor mismatch or position errors of the microphones. The second part of (\ref{eq:OP_Constraints}) ensures a distortionless beamformer response for each of the $I$ \acp{PLD}.
As in \cite{lnt2015-31}, we include measured \acp{HRTF} in (\ref{eq:OP_LSApproximation}) and (\ref{eq:OP_Constraints}) instead of the free-field-based steering vectors (which are only based on the microphone positions and the look directions). By doing this, the beamformer design can account for the influence of the humanoid robot's head on the sound field which would not be the case if we used free-field-based steering vectors as in \cite{lnt2010-46}. The sensor responses are given as $g_{n}(\omega_{q}, \phi_{m}, \theta_{m}) = h_{mn}(\omega_{q})$,
where $h_{mn}(\omega_{q})$ is the \ac{HRTF} modeling the propagation between the $m$-th source position and the $n$-th microphone, mounted at the humanoid robot's head, at frequency $\omega_{q}$. Matrix $\bb{G}(\omega_{q})$ consists of all \acp{HRTF} between the $M$ look directions and the $N$ microphones, and $\bb{a}_{i}(\omega_{q})$ contains the \acp{HRTF} corresponding to the $i$-th \ac{PLD}.
The optimization problem has to be solved for each frequency $\omega_{q}$ separately. We use the same desired response for all frequencies for the design of the polynomial beamformer, which is indicated by the frequency-independent entries of $\hat{\bb{b}}_{i}$ \cite{lnt2009-22, lnt2015-31,lnt2010-46}. The optimization problem in (\ref{eq:OP_LSApproximation}), (\ref{eq:OP_Constraints}) is formulated as a convex optimization problem \cite{lnt2010-46} and we use \textrm{CVX}, a package for specifying and solving convex programs in \textrm{Matlab} \cite{cvx}, to solve it. After the optimum filter weights at each frequency $\omega_{q}$ have been found, \ac{FIR} filters of length $L$ are obtained by \ac{FIR} approximation, see, e.g., \cite{Oppenheim:1999:DSP}, of the optimum filter weights using the \textrm{fir2} method provided by \textrm{Matlab} \cite{fir2}.
\section{Evaluation}
\label{sec:Evaluation}
In the following, we evaluate the proposed \ac{HRTF}-based \ac{RLSFIP} beamformer design and compare it to the \ac{HRTF}-based \ac{RLSFI} beamformer design proposed in \cite{lnt2015-31}. At first, the experimental setup is introduced. Then, the two beamformer designs are compared with respect to their approximation errors of the desired beamformer response. Eventually, the signal enhancement performance is evaluated in terms of \acp{WER} of an \ac{ASR} system.
\subsection{Setup and parameters}
\label{subsec:SetupAndParameters}
The evaluated beamformers were designed for the five-microphone robot head array in Fig.~\ref{fig:setup_headArray}, using a \ac{WNG} constraint of $\gamma_\text{dB}=-20\text{dB}$ and a filter length of $L=1024$.
For the design of the polynomial beamformer, we used $I=5$ \acp{PLD} $\phi_{i} \in \{0^\circ, 45^\circ, 90^\circ, 135^\circ,$ $180^\circ\}$ and a \ac{PPF} of order $P=4$.
The set of \acp{HRTF} which is required for the \ac{HRTF}-based beamformer design was measured in a low-reverberation chamber ($T_{60} \approx 50$ms) using maximum-length sequences, see, e.g., \cite{Schroeder_JASA_1979,Holters_DAF_2009}.
The \acp{HRTF} were measured for the same five-microphone array shown in Fig.~\ref{fig:setup_headArray} for a robot-loudspeaker distance of $1.1$m. The loudspeaker was at an elevation angle of $\theta = 56.4^\circ$ with respect to the robot. We chose this setup to simulate a taller human interacting with the NAO robot which is of height $0.57 \, \text{m}$. The measurements were carried out for the robot looking towards broadside $(\phi,\theta)=(90^\circ,90^\circ)$.
\begin{figure}
\subfigure[Microphone positions.]{
\includegraphics[width = 4cm]{microphones_annotated_5mics_close_greencircles_new.eps}
\label{fig:setup_headArray}
}
\hfill
\subfigure[Source positions.]{
\psfrag{inta}[cl][cl]{\parbox[t]{2cm}{\scriptsize \color{red} 1) $\phi_\mrm{int}=70^\circ$}}
\psfrag{intb}[cl][cl]{\parbox[t]{2cm}{\scriptsize \color{red} 2) $\phi_\mrm{int}=170^\circ$}}
\psfrag{t}[c][c]{\scriptsize \color{green} target}
\psfrag{30}[c][c]{\scriptsize $30^\circ$}
\psfrag{d}[cr][cr]{\scriptsize $1.1\, \text{m}$}
\psfrag{x1}[cl][cl]{\scriptsize $x_{N-1}$}
\psfrag{x0}[cl][cl]{\scriptsize $x_0$}
\includegraphics[width = 3.5cm]{Illustration_scenario.eps}
\label{fig:evaluation_scenarios}
}
\vspace{-2mm}
\caption{Illustration of the employed microphone positions (green circles) at the humanoid robot's head and the source positions of the two-speaker scenario.}
\vspace{-4mm}
\end{figure}
\subsection{Evaluation of HRTF-based polynomial beamformer design}
\label{subsec:EsperimentalResults}
In this section, we investigate how well the desired beamformer response $\hat{B}_{D_{i}}(\phi, \theta)$ is approximated by the beamformer response of either the \ac{HRTF}-based \ac{RLSFI} or the \ac{HRTF}-based \ac{RLSFIP} beamformer.
Ideally, the polynomial beamformer should be as good as the \ac{RLSFI} beamformer in the best case, because it approximates the latter, i.e., the performance of both beamformers should be similar when steered towards one of the $I$ \acp{PLD}.
Fig.~\ref{fig:designexample_prototpyelookDirection} shows the beampatterns of the \ac{HRTF}-based \ac{RLSFI} beamformer and of the \ac{HRTF}-based \ac{RLSFIP} beamformer in Figs~\ref{fig:designexample_prototpyelookDirection}(a) and \ref{fig:designexample_prototpyelookDirection}(b), respectively, steered towards $(\phi_\text{ld}, \theta_\text{ld})=(135^\circ, 56.4^\circ)$. The resulting \ac{WNG} of both beamformer designs is shown in Fig.~\ref{fig:designexample_prototpyelookDirection}(c). Please note that the beampatterns were computed with \acp{HRTF} modeling the acoustic system. Thus, they effectively show the transfer function between source position and beamformer output. A comparison of the beampatterns of the \ac{HRTF}- and free-field-based \ac{RLSFI} beamformer can be found in \cite{lnt2015-31}, illustrating the effect of the humanoid robot's head as scatterer on the sound field.
From Fig.~\ref{fig:designexample_prototpyelookDirection} it can be seen that the beampatterns of both beamformers look almost identical. This is because the actual look direction of the beamformers is equal to one of the five \acp{PLD} of the polynomial beamformer design.
One can also see that the \ac{WNG} is successfully constrained for both beamformer designs across the entire frequency range (with some slight deviations due to the \ac{FIR} approximation with finite filter length). Comparison of Figs~\ref{fig:designexample_prototpyelookDirection}(a) and \ref{fig:designexample_prototpyelookDirection}(b) with Fig~\ref{fig:designexample_prototpyelookDirection}(c) reveals that the beamformer's main beam broadens when the \ac{WNG} reaches its lower bound. This points to the tradeoff between robustness and spatial selectivity which the user can control via $\gamma$ in (\ref{eq:OP_Constraints}).
\begin{figure}[t]
\subfigure{
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\node at (-0.975,1.575) {\scriptsize (a)};
\node at (6.725,1.875) {\scriptsize $\text{BP}_\text{dB}/\text{dB}$};
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-1mm},
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
axis on top,
enlargelimits=false,
xmin=300, xmax=5000, ymin=0, ymax=180,
xtick={300,1000,2000,3000,4000,5000},
xticklabels={\empty},
ytick={0,45,90,135,180},
ylabel={$\phi/\circ\,\rightarrow$},
colorbar horizontal, colormap/jet,
colorbar style={
at={(0,1.15)}, anchor=north west, font=\tiny, width=6cm, height=0.15cm, xticklabel pos=upper
},
point meta min=-40, point meta max=0]
\addplot graphics [xmin=270, xmax=5035, ymin=0, ymax=185] {BP_hrtf_RLSFI_nmic_5_az_135_WNGlim_-20.eps};
\end{axis}
\end{tikzpicture}
}
\\[-4mm]
\subfigure{
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\node at (-0.975,1.575) {\scriptsize (b)};
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-1mm},
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
axis on top,
enlargelimits=false,
xmin=300, xmax=5000, ymin=0, ymax=180,
xtick={300,1000,2000,3000,4000,5000},
xticklabels={\empty},
ytick={0,45,90,135,180},
ylabel={$\phi/\circ\,\rightarrow$}]
\addplot graphics [xmin=270, xmax=5035, ymin=0, ymax=185] {BP_hrtf_RLSFIP_nmic_5_az_135_WNGlim_-20.eps};
\end{axis}
\end{tikzpicture}
}\\[-3.5mm]
\subfigure{
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\node at (-0.975,1.575) {\scriptsize (c)};
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-2mm},
legend style={font=\scriptsize, yshift=0.25mm, at={(.515,0.97)}},
legend columns = -1,
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
xtick={300,1000,2000,3000,4000,5000},
xticklabels={$300$,$1000$,$2000$,$3000$,$4000$,$5000$},
xlabel={$f/\text{Hz} \, \rightarrow$},
ytick={-20, -15, -10, -5, 0},
ylabel={$\text{WNG}/\text{dB}\,\rightarrow$},
ymin=-22.5, ymax=2.5, xmin=300,xmax=5000]
\addplot[thick,blue,solid] table [x index=0, y index=1]{WNG_RLSFI_hrtf_nmic_5_az_135_WNGlim_-20.dat}; \addlegendentry{RLSFI$\,\,\,$};
\addplot[thick,red,dashed] table [x index=0, y index=1]{WNG_RLSFIP_hrtf_nmic_5_az_135_WNGlim_-20.dat}; \addlegendentry{RLSFIP};
\end{axis}
\end{tikzpicture}
}
\vspace{-8mm}
\caption{Illustration of beampatterns of (a) the \ac{HRTF}-based \ac{RLSFI} beamformer and (b) the \ac{HRTF}-based \ac{RLSFIP} beamformer when the polynomial beamformer's look direction coincides with a \ac{PLD}. The beamformers were designed for the five-microphone robot head array in Fig.~\ref{fig:setup_headArray} with look direction $(\phi_\text{ld}, \theta_\text{ld})=(135^\circ, 56.4^\circ)$ and \ac{WNG} constraint $\gamma_\text{dB}=-20\, \text{dB}$. The resulting \ac{WNG} is illustrated in Subfigure (c).}
\label{fig:designexample_prototpyelookDirection}
\vspace{-6mm}
\end{figure}
In Fig.~\ref{fig:designexample_offPrototpyelookDirection} the beampatterns of the \ac{HRTF}-based \ac{RLSFI} and \ac{RLSFIP} beamformers are shown for the look direction $(\phi_\text{ld}, \theta_\text{ld})=(110^\circ, 56.4^\circ)$, which lies roughly half-way between two \acp{PLD} and can be expected to exhibit a large deviation from the desired response. The two beampatterns now look different, which is due to the interpolation between the \acp{PLD} by the polynomial beamformer. While for the lower frequencies the two main beams still look similar, the main beam of the polynomial beamformer is degraded for higher frequencies.
Moreover, it can be observed that the polynomial beamformer cannot maintain a distortionless response in the desired look direction across the entire frequency range.
The mismatch between \ac{RLSFI} and \ac{RLSFIP} beamformer also becomes obvious when looking at the \ac{WNG} in Fig.~\ref{fig:designexample_offPrototpyelookDirection}(c). The \ac{WNG} of the \ac{RLSFIP} beamformer is generally lower than that of the \ac{RLSFI} beamformer. In addition, the polynomial beamformer also exhibits a stronger violation of the \ac{WNG} constraint than the \ac{RLSFI} beamformer for $f < 500$Hz.
\begin{figure}[t]
\subfigure{
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\node at (-0.975,1.575) {\scriptsize (a)};
\node at (6.725,1.875) {\scriptsize $\text{BP}_\text{dB}/\text{dB}$};
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-1mm},
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
axis on top,
enlargelimits=false,
xmin=300, xmax=5000, ymin=0, ymax=180,
xtick={300,1000,2000,3000,4000,5000},
xticklabels={\empty},
ytick={0,45,90,135,180},
ylabel={$\phi/\circ\,\rightarrow$},
colorbar horizontal, colormap/jet,
colorbar style={
at={(0,1.15)}, anchor=north west, font=\tiny, width=6cm, height=0.15cm, xticklabel pos=upper
},
point meta min=-40, point meta max=0]
\addplot graphics [xmin=270, xmax=5035, ymin=0, ymax=185] {BP_hrtf_RLSFI_nmic_5_az_110_WNGlim_-20.eps};
\end{axis}
\end{tikzpicture}
}
\\[-4mm]
\subfigure{
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\node at (-0.975,1.575) {\scriptsize (b)};
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-1mm},
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
axis on top,
enlargelimits=false,
xmin=300, xmax=5000, ymin=0, ymax=180,
xtick={300,1000,2000,3000,4000,5000},
xticklabels={\empty},
ytick={0,45,90,135,180},
ylabel={$\phi/\circ\,\rightarrow$}]
\addplot graphics [xmin=270, xmax=5035, ymin=0, ymax=185] {BP_hrtf_RLSFIP_nmic_5_az_110_WNGlim_-20.eps};
\end{axis}
\end{tikzpicture}
}\\[-3.5mm]
\subfigure{
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\node at (-0.975,1.575) {\scriptsize (c)};
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-2mm},
legend style={font=\scriptsize, yshift=0.25mm, at={(.515,0.97)}},
legend columns = -1,
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
xtick={300,1000,2000,3000,4000,5000},
xticklabels={$300$,$1000$,$2000$,$3000$,$4000$,$5000$},
xlabel={$f/\text{Hz} \, \rightarrow$},
ytick={-20, -15, -10, -5, 0},
ylabel={$\text{WNG}/\text{dB} \,\rightarrow$},
ymin=-22.5, ymax=2.5, xmin=300,xmax=5000]
\addplot[thick,blue,solid] table [x index=0, y index=1]{WNG_RLSFI_hrtf_nmic_5_az_110_WNGlim_-20.dat}; \addlegendentry{RLSFI$\,\,\,\,$};
\addplot[thick,red,dashed] table [x index=0, y index=1]{WNG_RLSFIP_hrtf_nmic_5_az_110_WNGlim_-20.dat}; \addlegendentry{RLSFIP};
\end{axis}
\end{tikzpicture}
}
\vspace{-8.5mm}
\caption{Illustration of beampatterns of (a) the \ac{HRTF}-based \ac{RLSFI} beamformer and (b) the \ac{HRTF}-based \ac{RLSFIP} beamformer when the polynomial beamformer's look direction does not coincide with one of the \acp{PLD}. The beamformers were designed for the five-microphone robot head array in Fig.~\ref{fig:setup_headArray} with look direction $(\phi_\text{ld}, \theta_\text{ld})=(110^\circ, 56.4^\circ)$ and \ac{WNG} constraint $\gamma_\text{dB}=-20\, \text{dB}$. Subfigure (c) shows the resulting \ac{WNG}.}
\label{fig:designexample_offPrototpyelookDirection}
\end{figure}
In the following, we measure the approximation error of the desired response $\hat{B}_{D_\text{ld}}(\phi, \theta)$ for a certain look direction $\phi_\text{ld}$ by the actual beamformer response $B_{D_\text{ld}}(\omega, \phi, \theta)$, where $D_\text{ld} = (\phi_\text{ld}-90)/90$, of either the \ac{RLSFI} or \ac{RLSFIP} beamformer by calculating the \ac{MSE} which is defined as \cite{lnt2014-74}:
\begin{equation}
\text{MSE}(\phi_\text{ld}) = \sum\limits_{q=0}^{Q-1}\sum\limits_{m=0}^{M-1} \frac{\left( |B_{D_\text{ld}}(\omega_{q}, \phi_{m}, \theta_{m})|-|\hat{B}_{D_\text{ld}}(\phi_{m}, \theta_{m})| \right)^{2}}{Q \cdot M}.
\label{eq:MSE_BFResponse}
\end{equation}
Fig.~\ref{fig:MSE_I_5_P_4} depicts the \ac{MSE} of the \ac{HRTF}-based \ac{RLSFI} and \ac{RLSFIP} beamformer designs, calculated in steps of five degrees over the entire steering range $0^\circ \leq \phi_\text{ld} \leq 180^\circ$.
When steered to one of the five \acp{PLD}, i.e., when $\phi_\text{ld}=\phi_{i}$, the \ac{RLSFIP} beamformer design yields a similar \ac{MSE} as the \ac{RLSFI} beamformer design. In between those \acp{PLD}, the \ac{MSE} of the polynomial beamformer design is usually larger than that of the \ac{RLSFI} beamformer design. The lower \ac{MSE} of the polynomial beamformer for $\phi_\text{ld} \in \{5^\circ, 175^\circ\}$ may be explained by side lobes of the polynomial beamformer which are less pronounced at higher frequencies than those of the \ac{RLSFI} beamformer for these two particular look directions.
\begin{figure}
\hspace{8mm}
\begin{tikzpicture}[scale=1,trim axis left]
\begin{axis}[
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-1mm},
legend style={font=\scriptsize, yshift=0.25mm, at={(.515,0.97)}},
legend columns = -1,
width=8.91cm,height=3.25cm,grid=major,grid style = {dotted,black},
x dir=reverse,
xtick={0, 45, 90, 135, 180},
xlabel={$\phi_\text{ld}/\circ$},
ytick={0, 0.1, 0.2, 0.3},
ylabel={$\text{MSE} \,\rightarrow$},
ymin=0, ymax=0.35, xmin=0,xmax=180]
\addplot[thick,blue,solid] table [x index=0, y index=1]{MSE_RLSFI_hrtf_nmic_5_WNGlim_-20.dat}; \addlegendentry{RLSFI$\,\,\,\,$};
\addplot[thick,red,dashed] table [x index=0, y index=1]{MSE_RLSFIP_hrtf_nmic_5_WNGlim_-20.dat}; \addlegendentry{RLSFIP};
\end{axis}
\end{tikzpicture}
\vspace{-4.5mm}
\caption{Illustration of the \ac{MSE} (\ref{eq:MSE_BFResponse}) of the \ac{HRTF}-based \ac{RLSFI} (blue curve) and \ac{HRTF}-based \ac{RLSFIP} (red curve) beamformer designs, calculated in steps of five degrees over the entire steering range.}
\label{fig:MSE_I_5_P_4}
\vspace{-3mm}
\end{figure}
\subsection{Evaluation of signal enhancement performance}
\label{subsec:EsperimentalResults}
In this section, we evaluate the overall quality of the enhanced signals at the outputs of the \ac{HRTF}-based \ac{RLSFI} and \ac{RLSFIP} beamformers. In addition, we also evaluate the original free-field-based \ac{RLSFIP} beamformer \cite{lnt2010-46} which assumes free-field propagation of sound waves and, therefore, cannot account for the influence of robot's head on the sound field. To this end, we use \acp{WER} of an automatic speech recognizer to evaluate the overall quality of the enhanced signals at the beamformer outputs, since a high speech recognition accuracy is the main goal in robot audition. As \ac{ASR} engine, we employed PocketSphinx \cite{Huggins:2006} with a \ac{HMM}-\ac{GMM}-based acoustic model which was trained on clean speech from the GRID corpus \cite{Cooke:2006}, using MFCC+$\Delta$+$\Delta \Delta$ features and cepstral mean normalization. For the computation of the \ac{WER} scores, only the letter and the number in the utterance were evaluated, as in the CHiME challenge \cite{ChristensenBMG10}. Our test signal contained $200$ utterances. Note that since the \ac{ASR} system was trained on clean speech, we implicitly measure the amount of target signal distortion and interferer suppression.
We evaluated the signal enhancement in a two-speaker scenario, where the target signal was located at positions between $\phi_\text{ld}=0^\circ$ and $\phi_\text{ld}=180^\circ$ in steps of $30^\circ$. The \ac{DOA} of the target signal was assumed to be known for the experiments, i.e., no localization algorithm was applied. An investigation of the \ac{HRTF}-based beamformer's robustness against \ac{DOA} estimation errors can be found in \cite{lnt2016-16}. For each target position, seven interfering speaker positions between $\phi_\text{int}=15^\circ$ and $\phi_\text{int}=165^\circ$ in steps of $30^\circ$ were evaluated. An overview over all source positions is given in Fig.~\ref{fig:evaluation_scenarios}, where target and interfering sources are represented by black circles and red crosses, respectively.
We created the microphone signals by convolving clean speech signals with \acp{RIR} which we measured in a lab room with a reverberation time of $T_{60} \approx 190$ms and a critical distance \cite{kuttruff2000room} of approximately $1.2$m. The \acp{RIR} were measured with the same configuration as was used for the \ac{HRTF} measurements described above.
The \acp{WER} were calculated for each combination of target and interfering source position and averaged over the \acp{WER} obtained for the different interferer positions. The resulting average target source position-specific \acp{WER} are depicted in Fig.~\ref{fig:results_WER}.
The obtained \acp{WER} show that both \ac{HRTF}-based beamformers significantly improve the speech recognition accuracy of the input signal. Moreover, they also outperform the free-field-based \ac{RLSFIP} beamformer significantly, which emphasizes the necessity to incorporate the effect of the robot's head on the sound field into the beamformer design. It is interesting to see that the \ac{HRTF}-based \ac{RLSFIP} beamformer performs as well as the \ac{HRTF}-based \ac{RLSFI} beamformer whenever the target source is located in one of the \acp{PLD} which were used for designing the polynomial beamformer. When this is not the case, only a slightly higher average \ac{WER} is obtained. This confirms that the polynomial interpolation of the \ac{HRTF}-based \ac{RLSFI} beamformer design works reasonably well such that it can be used in a robot audition scenario.
\begin{figure}
\begin{tikzpicture}
\begin{axis}[
width=9.25cm,height=4cm,grid=major,grid style = {dotted,black},
label style = {font=\scriptsize},
tick label style = {font=\tiny},
ylabel style={yshift=-2mm},
ybar=0.5pt,
bar width=6pt,
enlargelimits=0.075,
ymin=0, ymax = 65,
ytick={0,20,40,60},
ylabel={$\text{WER}/\%\, \rightarrow$},
xtick=data,
xmin=0, xmax = 180,
xlabel={$\phi_\text{ld}/\circ$},
xlabel shift={-1mm},
x dir=reverse,
every node near coord/.append style={font=\tiny,
rotate=90, anchor=west,
/tikz/.cd},
nodes near coords, nodes near coords align={vertical},
legend style={at={(0.5,1.225)},anchor=north,legend columns=-1,font=\scriptsize},
legend entries={$\text{Input}\quad$, $\text{RLSFI}_\text{HRTF}\quad$, $\text{RLSFIP}_\text{HRTF}\quad$, $\text{RLSFIP}_\text{Free-field}$},
]
\addplot[black,fill=red,postaction={pattern=north east lines}] coordinates {(0,50.5) (30,48.9) (60,48.0) (90,47.1) (120,48.0) (150,48.6) (180,50.7)};
\addplot[black,fill=green] coordinates {(0,33.7) (30,33.8) (60,32.1) (90,32.4) (120,34.5) (150,34.7) (180,35.7)};
\addplot[black,fill=yellow, postaction={pattern=crosshatch dots}] coordinates {(0,33.8) (30,36.3) (60,33.3) (90,32.5) (120,36.1) (150,37) (180,35.6)};
\addplot[black,fill=cyan, postaction={pattern=north west lines}] coordinates {(0,46.9) (30,39.6) (60,38.2) (90,40.0) (120,39.7) (150,43.5) (180,44.1)};
\end{axis}
\end{tikzpicture}
\vspace{-4.5mm}
\caption{Illustration of average target source position-specific \acp{WER} in $\%$, obtained at the input (red bars) and at the output of the \ac{HRTF}-based \ac{RLSFI} (green bars), \ac{HRTF}-based \ac{RLSFIP} (yellow bars), and free-field-based \ac{RLSFIP} (cyan bars) beamformers.}
\label{fig:results_WER}
\vspace{-5mm}
\end{figure}
\vspace{-3mm}
\section{Conclusion}
\label{sec:Conclusion}
In this work, we proposed an \ac{HRTF}-based \ac{RLSFIP} beamformer design which allows for a flexible steering of a previously proposed robust \ac{HRTF}-based \ac{RLSFI} beamformer.
We evaluated both beamformer designs with respect to their corresponding approximation error of the desired beamformer response and with respect to their signal enhancement performance which was evaluated by means of \acp{WER} of an \ac{ASR} system. The results showed that the polynomial beamformer design provides a good approximation of the \ac{RLSFI} beamformer design and, therefore, can be used successfully in a robot audition scenario instead of the computationally much more complex \ac{RLSFI} beamformer design.
Future work includes an investigation of the proposed \ac{HRTF}-based polynomial beamformer design for more irregular sensor arrangements
as well as an evaluation with a state-of-the-art \ac{DNN}-based \ac{ASR} system.
An extension of the \ac{RLSFIP} beamformer design to allow for a flexible steering of the main beam in two dimensions is another aspect of future work.
\bibliographystyle{IEEEbib}
|
\section{\label{Sec: Introduction}Introduction}
Laser scattering involving stimulated Brillouin scattering (SBS), i.e., the coupling of a large amplitude light wave into a scattered light wave plus an ion acoustic wave (IAW), \cite{Kruer} play an important role in the successful ignition goal of inertial confinement fusion (ICF) \cite{(1), (9)_37}.
Many methods have been used to reduce the SBS \cite{(9)_49, (9)_46} scattering level, such as increasing the Landau damping of ion-acoustic waves\cite{Liu_8} and the saturation of ion-acoustic wave \cite{(9)_6,(9)_7}.
Various mechanisms have been put forward to explain ion-acoustic wave saturation, including frequency detuning induced by particle-trapping,\cite{(9)_6_2,(9)_6_4,(9)_6_5} increased linear Landau damping due to kinetic ion heating,\cite{(9)_6_9,(9)_6_10} nonlinear damping induced by wave-breaking and trapping\cite{(9)_6_11,(9)_6_12} or coupling with higher harmonics\cite{(9)_6_13,(9)_6_14}.
Multiple ion species are contained in the laser fusion program. The presence of multiple ion species can add additional branches to the IAW dispersion relation and change the total Landau damping significantly, which may provide the possibility of controlling SBS by increasing the linear damping of the IAW.\cite{Liu_8} When the external driving electric field is on, the trapping of particles will reduce Landau damping, if the time of the driver is long enough, the Landau damping will be nearly zero.\cite{(3)} At the same time, the dispersion of the small amplitude nonlinear IAW will be altered. The linear kinetic theory of ion acoustic waves in two ion species plasmas was researched by E. A. Williams et al. in detail.\cite{(7)} There, $k\lambda_{De}=0.1$ was taken, the linear frequency of the IAW was thought to be the resonance frequency of the IAW. In this paper, we can see when $k\lambda_{De}=0.1$, the linear frequency of the IAW was close to the frequency of the small amplitude nonlinear IAW (discussed later). In 2013, T. Chapman et al. researched the kinetic nonlinear frequency shift (KNFS) of the nonlinear ion acoustic wave in CH plasmas.\cite{T. Chapman_PRL} The fundamental frequency $\omega_0$ for Vlasov results was given by simulation. They took the frequency of the IAW of zero amplitude by extrapolating from the lowest measured amplitude cases, i. e., $\omega_{0}=\omega(\phi\rightarrow0)$. In 2001, Rose and Russell\cite{Rose_2001POP} derived the expressions for the nonlinear dielectric function of Langmuir waves, thus the nonlinear dispersion relation. Then on this basis, in 2003, Rose presented the estimates to the spatial gain rate coefficients of the backward stimulated Raman scatter (BSRS) and backward stimulated Brillouin scatter (BSBS) due to trapping effects.\cite{Rose_2003POP} And also, Strozzi et al. \cite{Strozzi_2007POP} paid attention to the linear modes with distributions modified by electron trapping and researched the kinetic effects such as electron trapping in stimulated Raman backscatter by one-dimensional Eulerian Vlasov-Maxwell simulations. Then, they presented a framework for estimating when electron trapping nonlinearity was expected to be important in Langmuir waves.\cite{Strozzi_2012POP} In the work of Rose and Russell,\cite{Rose_2001POP} the finite wave amplitude $\phi$ was considered and the real part of the nonlinear dielectric function determined the nonlinear resonance condition. When the wave amplitude was small enough, the solutions in the limit $\phi\rightarrow0$ corresponded to the infinitesimal amplitude Bernstein-Greene-Kruskal (BGK) modes\cite{BGK}, which was consistent to that found by Holloway and Dorning.\cite{Holloway_1991} Their results were for the electron plasma waves while in this paper the resonance condition is applied to IAWs in multi species plasmas especially for the infinitesimal amplitude BGK modes. In this paper, the dispersion relation with no damping, corresponding to the real part of the linear dielectric function with no damping, is given to calculate the fundamental frequency $\omega_0$ of the small-amplitude nonlinear IAW which is undamped plasma wave\cite{Holloway_1991}. The electrons and ions are all kinetic, but the Landau damping of the nonlinear IAW can be ignored as a result of particle trapping. We can see when $k\lambda_{De}$ is not large, the deviation between the frequency calculated by the linear dispersion relation (defined as the linear frequency) and that calculated by the dispersion relation with no damping of the infinitesimal amplitude nonlinear IAW is close to each other, but with $k\lambda_{De}$ increasing, the deviation becomes larger. By Vlasov simulation, the frequency calculated by the dispersion relation with no damping is verified much closer to the resonance frequency of the small-amplitude nonlinear IAW.
This paper discusses the change of the linear Landau damping rate of the fast IAW mode (the phase velocity of which is faster than any ion's thermal velocity) and the slow IAW mode (the phase velocity of which is close to the thermal velocity of one of the plasma species) in CH plasmas when $ k\lambda_{De} $ varies. When $k\lambda_{De}$ increases, the linear Landau damping of the fast mode will increase obviously in the region of $T_i/T_e\lesssim0.2$. However, this result is for the linear situation. When the driving electric field is on, the distribution of the particles will keep flat at the phase velocity of the IAW, thus turning off the linear Landau damping. As a result, the deviation between the frequency calculated by the linear dispersion relation and that by the dispersion relation with no damping becomes larger with $k\lambda_{De}$ increasing.
When $k\lambda_{De}$ is not large, the nonlinear IAW can be excited by the driver with the linear frequency of the modes, which indicates that the frequency of the nonlinear IAW can be approximated by the linear dispersion relation such as $k\lambda_{De}=0.1, 0.3, 0.5$. However, when $k\lambda_{De}=0.7$, the linear dispersion relation can not be applied to exciting the nonlinear IAW, while the dispersion relation with no damping can be applied to exciting the nonlinear IAW.
This paper is organized as follows:
In Sec. \ref{Sec:Theory analysis}, the linear dispersion relation and the dispersion relation with no damping are presented. The change of the phase velocity and Landau damping of the IAW modes with $k\lambda_{De}$ is discussed. And the dispersion relation with no damping is provided to calculate the frequency of the small-amplitude nonlinear IAW in CH plasmas.
In Sec. \ref{Sec:Vlasov simulation}, the excitation of the nonlinear ion-acoustic wave by an external electric field with the linear frequency and the frequency calculated by the dispersion relation with no damping is demonstrated by one-dimensional (1D) Vlasov simulation. And the Vlasov simulation results of the excitation of the nonlinear IAW both for the fast mode and the slow mode are shown when $k\lambda_{De}$ varies.
In Sec. \ref{Sec:Discussion}, the discussion of the effect of the nonlinear frequency shift is given.
In Sec. \ref{Sec:Summary}, our main results are summarized.
\section{\label{Sec:Theory analysis}Theoretical analysis}
The linear kinetic theory of ion acoustic wave in a non-magnetized, homogeneous plasma consisting of multi-species ions is considered. Considering a neutral, fully ionized, unmagnetized CH plasmas (1:1 mixed) with the same temperature of all ion species ($T_H=T_C=T_i$), the linear frequencies of the plasmas are then given by the zeros of the plasma dielectric function, i.e., $\epsilon(k,\omega)=0$, which gives the linear dispersion relation of the ion acoustic wave in multi-ion species plasmas
\begin{equation}
\label{Eq:Dispersion}
\epsilon_L(\omega,k)=1+\chi_e+\sum_\beta \chi_{i\beta}=0,
\end{equation}
where $\chi_e$ is the susceptibility of electron and $\chi_{i\beta}$ is the susceptibility of species $\beta$ (H ion or C ion). $\chi_j$ (j present for electron, H ion or C ion) can be expressed by $Z$ function
\begin{equation}
\label{Eq:Chi}
\chi_j=\frac{1}{(k\lambda_{Dj})^2}(1+\xi_jZ(\xi_j)),
\end{equation}
where a Maxwellian velocity distribution for all species is assumed. $\xi_j=\omega/(\sqrt{2}kv_{tj})$ is generally complex, $\omega=\omega_r+i\gamma$, $k$ is the frequency and the wave number of the given mode (such as the fast or slow IAW mode). $v_{tj}=\sqrt{T_j/m_j}$ ($T_j$, $m_j$ are the temperature and the mass of particle $j$) is the thermal velocity of particle $j$. $\lambda_{Dj}$ is the Debye length $\lambda_{Dj}=\sqrt{T_j/4\pi n_jZ_j^2e^2}$, i.e., $\lambda_{Dj}=v_{tj}/\omega_{pj}$ ($\omega_{pj}=\sqrt{4\pi n_jZ_j^2e^2/m_j}$ is the plasma frequency for specie j). $Z$ function is the dispersion function
\begin{equation}
\label{Eq:Zeta}
\begin{aligned}
Z(\xi_j) & =\frac{1}{\sqrt{\pi}}\int_{-\infty}^{+\infty}\frac{e^{-v^2}}{v-\xi_j}dv\\
& =i\sqrt{\pi}e^{-\xi_j^2}(1+erf(i\xi_j)),
\end{aligned}
\end{equation}
in which $erf(i\xi_j)=2/\sqrt{\pi}\int_0^{i\xi_j} e^{-\eta^2}d\eta$ is the error function, $\xi_j$ is a complex variable. The direct numerical solution to the Eq. (\ref{Eq:Dispersion}) can be solved by Newton-Raphson iterative method. The newly developed accurate algorithm of the Faddeyeva (plasma dispersion) function $\omega(z)=e^{-z^2}(1+erf(iz))$ (where $z=x+iy$ is a complex variable) is showed by M. R. Zaghloul\cite{ACM}.
\begin{figure}[!tp]
\includegraphics[width=1.0\columnwidth]{Fig1.eps}
\caption{\label{Fig:LandauDamping}(Color online) (a) The phase velocity and (b) the linear Landau damping of the fast mode and the slow mode as a function of $T_i/T_e$ when $k\lambda_{De}=0.1, 0.3, 0.5, 0.7$. }
\end{figure}
\begin{table*}
\caption{\label{table1} The main results of the linear dispersion relation and dispersion relation with no damping}
\begin{ruledtabular}
\begin{tabular}{|c|cccc|cccc|}
\hline
& \multicolumn{4}{c|}{\bf{fast mode}}&\multicolumn{4}{c|}{\bf{slow mode}}\\
\hline
$\bf{T_i/T_e}$& \multicolumn{4}{c|} {{\bf0.1}} & \multicolumn{4}{c|}{{\bf0.5}} \\
\hline
$\bf{k\lambda_{De}}$& {\bf0.1} & {\bf0.3} &{\bf 0.5} & {\bf0.7}
& {\bf0.1} & {\bf0.3} & {\bf0.5} & {\bf0.7} \\
\hline
$\bf{Re(\omega_L)/10^{-3}\omega_{pe}}$& ${\bf 1.965}$&${\bf5.700}$& ${\bf8.930}$ &${\bf 11.47}$& $\bf1.746$& $\bf5.121$
&$\bf8.250$ &$\bf10.91$ \\
$\bf{\omega_N/10^{-3}\omega_{pe}}$& ${\bf 1.912}$& ${\bf5.505}$& ${\bf 8.478}$ &${\bf 10.60}$&$\bf1.746$ &$\bf5.094$
& $\bf8.079$ &$\bf10.60$ \\
$\bf{|Im(\omega_L)|/10^{-3}\omega_{pe}}$& {0.1561} & {0.5150}& {1.018}& {1.767}& 0.1707 & 0.4857
& 0.7682& 1.055 \\
$\bf{|Im(\omega_L)/Re(\omega_L)|}$& {0.07946} & {0.09035}& {0.114}& {0.154}& 0.09775 & 0.09485
& 0.09312& 0.09667 \\
\hline
\end{tabular}
\end{ruledtabular}
\end{table*}
Figs. \ref{Fig:LandauDamping}(a) and \ref{Fig:LandauDamping}(b) show the phase velocity and the linear Landau damping rate of the fast mode and the slow mode in the condition of $k\lambda_{De}=0.1, 0.3, 0.5, 0.7$. With $k\lambda_{De}$ increasing, the phase velocity of the fast mode will decrease and be closer to the thermal velocity of H ions especially in the region of $T_i/T_e\lesssim0.2$ in which the Landau damping rate of the fast mode is weak. Thus, the Landau damping rate of the fast mode increases obviously with $k\lambda_{De}$ increasing especially in the region of $T_i/T_e\lesssim0.2$.\\
When the external driving electric field is turned on, the total electric potential of the system including the electric field of the driver and the electric field of IAW will trap particles, thus making the distribution flat at the wave phase velocity, which will reduce Landau damping.\cite{(3)} If the duration time of the driver is long enough, the Landau damping will be turned off, i.e., $Im(\omega)\approx0$. By retaining only the real part of Eq. (1), the dispersion relation with no damping of the infinitesimal amplitude nonlinear modes (described briefly below as \textquotedblleft the dispersion relation with no damping\textquotedblright ) is given by
\begin{equation}
\label{Eq:Nonlinear}
Re(\epsilon_L(Re(\omega), k))=0,
\end{equation}
in which $Re(\omega)$ is the frequency of the infinitesimal amplitude nonlinear mode by taking $Im(\omega)=0$. $\epsilon_L$ is defined in Eq. (1). Where Eq. (\ref{Eq:Nonlinear}) describes only small-amplitude nonlinear IAW. A Maxwellian distribution for all species is used and the width of the plateau, where $\partial f_0/\partial v=0$, is infinitesimal.
\begin{figure}[!tp]
\includegraphics[width=\columnwidth]{Fig2.eps}
\caption{\label{Fig:LinearNonlinear}(Color online) The dispersion relations of the fast mode and the slow mode calculated by linear dispersion relation and dispersion relation with no damping in the condition of (a) $T_i/T_e=0.1$, (b) $T_i/T_e=0.5$. Note that \textquotedblleft Nonlinear mode" presents for the \textquotedblleft infinitesimal amplitude nonlinear mode" (red lines in (a) and (b)). The linear Landau damping of (c) the fast mode in the condition of $T_i/T_e=0.1$ and (d) the slow mode in the condition of $T_i/T_e=0.5$.}
\end{figure}
As $k\lambda_{De}$ increases, the linear Landau damping of the fast mode increases obviously as shown in Fig. \ref{Fig:LinearNonlinear}(c), therefore, the frequency of the linear fast mode deviates from that of the nonlinear IAW mode (Fig. \ref{Fig:LinearNonlinear}(a), the upper branch of the red line). When $k\lambda_{De}=0.7$, the deviation between the frequency of the linear fast mode and the nonlinear mode, which is calculated by the dispersion relation with no damping, is nearly as large as $7\%$. The same analysis of the slow mode in the condition of $T_i/T_e=0.5$ is given in Figs. \ref{Fig:LinearNonlinear}(b) and \ref{Fig:LinearNonlinear}(d). As shown in Fig. \ref{Fig:LinearNonlinear}(d), the linear Landau damping rate $|Im(\omega)/Re(\omega)|$ of the slow mode decreases first and then increases as $k\lambda_{De}$ increases. The deviation between the frequency of the linear slow mode and the nonlinear IAW mode (Fig. \ref{Fig:LinearNonlinear}(b), the upper branch of the red line) becomes larger as $k\lambda_{De}$ increases as shown in Fig. \ref{Fig:LinearNonlinear}(b). When $k\lambda_{De}=0.7$, the deviation is nearly $3\%$. For $k\lambda_{De}=0.1, 0.3, 0.5$, the frequency of the linear fast mode in the condition of $T_i/T_e=0.1$ and the linear slow mode in the condition of $T_i/T_e=0.5$ are near the frequency of the nonlinear IAW mode as shown in Figs. \ref{Fig:LinearNonlinear}(a) and \ref{Fig:LinearNonlinear}(b), so the nonlinear IAW can be excited by the driver with the linear frequency of the IAW.
These cases are calculated by linear dispersion relation and dispersion relation with no damping to get the frequency of the IAW ${Re(\omega_L)}$, $\omega_N$, the linear Landau damping $|Im(\omega_L)|$, the relative linear Landau damping rate $|Im(\omega_L)/Re(\omega_L)|$ of the fast mode in the condition of $T_i/T_e=0.1$ and the slow mode in the condition of $T_i/T_e=0.5$. The results of the linear dispersion relation and the dispersion relation with no damping of different $k\lambda_{De}$ are shown in Table I. As shown in Fig. \ref{Fig:LandauDamping}(b), when $T_i/T_e=0.1$, the less damped mode is the fast mode, when $T_i/T_e=0.5$, the less damped mode is the slow mode. In this paper, only the weakly damped modes are considered, i.e., a fast mode in the condition of $T_i/T_e=0.1$ and a slow mode in the condition of $T_i/T_e=0.5$ are taken as representations.
\section{\label{Sec:Vlasov simulation}Vlasov simulation}
In the one-dimensional (1D) Vlasov code,\cite{Liu-2009POP,Liu-2009POP_1} assuming that the external driving electric field $E_d$ is along the x direction. For particle specie s (s presents for electrons, H ions or C ions in this paper), then the motion of specie s can be described by Vlasov-Poisson equations:
\begin{equation}
\label{Eq:Vlasov}
\frac{\partial f_s}{\partial t}+v_{xs}\frac{\partial f_s}{\partial x}+\frac{q_s}{m_s}(E_x+E_d)\frac{\partial f_s}{\partial v_x}=0,
\end{equation}
\begin{equation}
\label{Eq:Poisson}
\frac{\partial E_x}{\partial x}=4\pi \sum\limits_{s}n_{s0}q_s\int\limits_{-\infty}^{+\infty}f_sdv,
\end{equation}
where all of the particles including electrons and ions are taken as kinetic particles and $f_s$ is the normalized distribution function for specie s. The initial equilibrium distribution function for each specie particles $f_{s0}$ is normalized Maxwellian distribution satisfying $\int_{-\infty}^{+\infty}f_{s0}dv=1$. $E_x$ is the self-consistent electrostatic field in plasmas and is calculated from Poisson's equation. And $q_s$, $m_s$ and $n_{s0}$ present for the charge, mass and the background equilibrium density of particle s.
To solve Vlasov equation (\ref{Eq:Vlasov}), we split the time-stepping operator into free-streaming in x and motion in $v_x$,\cite{7_5-1976JCP,7-2004CPC} then we can get the advection equations:
\begin{equation}
\begin{aligned}
\label{Eq:split1}
&\frac{\partial f_s}{\partial t}+v_{xs}\frac{\partial f_s}{\partial x}=0,\\
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\label{Eq:split2}
&\frac{\partial f_s}{\partial t}+\frac{q_s}{m_s}(E_x+E_d)\frac{\partial f_s}{\partial v_x}=0,
\end{aligned}
\end{equation}
a third order Van Leer scheme (VL3)\cite{VL3, 10-2006POP} is taken to solve the advection equations (\ref{Eq:split1}) and (\ref{Eq:split2}). To solve the Poisson equation (\ref{Eq:Poisson}), fast Fourier transform (FFT) method\cite{1965FFT, FFTsolvePoisson} is taken to calculate the electrostatic field.
The excitation of the nonlinear IAW is simulated by 1D1V Vlasov code. In this code, the electrons and ions are all kinetic. To solve the electrons and ions kinetic behavior, the phase space domain is $[0, L_x]\times[-v_{max}, v_{max}]$, where $L_x=2\pi/k$ is the longest wavelength that fits into the simulation box and $v_{max}=8v_{tj}$, i.e., for kinetic electrons $v_{max}=8v_{te}$, for kinetic ions $v_{max}=8v_{ti}$ ($i$ presents for H ions or C ions in this paper). The phase space is discretized with $N_x=128$ grid points in the spatial domain and $N_v=256$ in the velocity domain. The periodic boundary condition is taken in the spatial domain. The time step is $dt=0.1\omega_{pe}^{-1}$. The envelope of the external driver is given by
\begin{equation}
\tilde{E}_d(t)=\tilde{E}_d\frac{1}{1+(\frac{t-t_0}{\frac{1}{2}t_0})^{10}},
\end{equation}
where $\tilde{E}_d=eE_d\lambda_{De}/T_e$ is the maximum amplitude of the external driver, tilde presents for normalization. $t_0$ is the duration time of the peak electric field.
To generate a driven IAW, one thus considers
\begin{equation}
\tilde{E}_d(x,t)=\tilde{E}_d(t)sin(kx-\omega_d t),
\end{equation}
where $\omega_d, k$ are the frequency and the wave number of the external driver. In this paper, to excite the nonlinear IAW, $\omega_d$ is chosen as the frequency of the modes calculated by linear dispersion relation or dispersion relation with no damping.
\subsection{\label{Subsec:A. Response of IAW}Response of IAW to the driver}
\begin{figure}[htd]
\includegraphics[width=\columnwidth]{Fig3.eps}
\caption{\label{Fig:Resonance}(Color online) The resonance curve of the fast mode ($T_i/T_e=0.1$) in the condition of (a) $k\lambda_{De}=0.1$, (b) $k\lambda_{De}=0.3$. The amplitude of the driver $eE_d\lambda_{De}/T_e=1\times10^{-3}$.}
\end{figure}
\begin{figure}[htd]
\includegraphics[width=\columnwidth]{Fig4.eps}
\caption{\label{Fig:Resonance2}(Color online) The resonance curve for (a) the fast mode, $T_i/T_e=0.1$, (b) the slow mode, $T_i/T_e=0.5$ in the condition of $k\lambda_{De}=0.5$, the black dashed line presents for the amplitude of the driver $eE_d\lambda_{De}/T_e=7\times10^{-3}$.}
\end{figure}
\begin{figure}[htd]
\includegraphics[width=\columnwidth]{Fig5.eps}
\caption{\label{Fig:Resonance3}(Color online) The resonance curve for (a) the fast mode, $T_i/T_e=0.1$, (b) the slow mode, $T_i/T_e=0.5$ in the condition of $k\lambda_{De}=0.7$, the black dashed line presents for the amplitude of the driver $eE_d\lambda_{De}/T_e=7\times10^{-3}$.}
\end{figure}
\begin{figure*}[!tp]
\includegraphics[width=2\columnwidth]{Fig6.eps}
\caption{\label{Fig:SlowMode}(Color online) The total electric field and the envelope curve of the external driving electric field (driver) for the slow mode, $T_i/T_e=0.5$, in the condition of (a) $k\lambda_{De}=0.5$, (b) $k\lambda_{De}=0.7$. The phase picture of (c)-(e) electrons, (f)-(h) H ions, (i)-(k) C ions at the time of $T_1$ ($\sim4\times10^5\omega_{pe}^{-1}, k\lambda_{De}=0.5$), $T_2$ ($\sim1\times10^5\omega_{pe}^{-1}, k\lambda_{De}=0.7$), $T_3$ ($\sim4\times10^5\omega_{pe}^{-1}, k\lambda_{De}=0.7$). The driver is with the linear frequency of the slow mode calculated by the linear dispersion relation.}
\end{figure*}
To get the resonance curve, $k\lambda_{De}$ and other conditions are fixed, the frequency of the driver varies. The resonance curves of the fast mode ($T_i/T_e=0.1$) in the condition of $k\lambda_{De}=0.1, 0.3$ are shown in Fig. \ref{Fig:Resonance}. The external driving electric field is turned on from the time of 0 to $2\times10^5\omega_{pe}^{-1}$ with the duration time $t_0=1\times10^5\omega_{pe}^{-1}$ of the maximum amplitude $eE_d\lambda_{De}/T_e=1.0\times10^{-3}$. When $k\lambda_{De}=0.1$, for the harmonics have much larger growth rate than the fundamental mode, the harmonics will carry a large part of energy of the IAW and prevent the fundamental mode from growing up. Then all modes will start to saturate and the IAW amplitude will saturate at a smaller level than the driver amplitude, which is similar to the results of the electron acoustic wave (EAW) researched by Xiao et al.\cite{Xiao_2014POP}. After $t=2\times10^5\omega_{pe}^{-1}$, the driver is turned off, the amplitude of the electric field of the driving IAW, $eE\lambda_{De}/T_e$ (or $eE_{IAW}\lambda_{De}/T_e$), is obtained at the time of $t\sim4\times10^{5}\omega_{pe}^{-1}$ (long after the driver has been turned off). As shown in Fig. \ref{Fig:Resonance}(a), in the condition of $k\lambda_{De}=0.1$ for the fast mode, $T_i/T_e=0.1$, the resonance frequency is near $\omega_N=1.912\times10^{-3}\omega_{pe}$ which is the frequency calculated by the dispersion relation with no damping rather than the frequency calculated by the linear dispersion relation $\omega_L=1.965\times10^{-3}\omega_{pe}$. The same results are obtained in the condition of $k\lambda_{De}=0.3$ for the fast mode ($T_i/T_e=0.1$) as shown in Fig. \ref{Fig:Resonance}(b). The resonance frequency is very close to the frequency calculated by the dispersion relation with no damping $\omega_N=5.505\times10^{-3}\omega_{pe}$ and the resonance peak is much larger than the amplitude of the driving IAW by the driver with the frequency calculated by the linear dispersion relation $\omega_L=5.7\times10^{-3}\omega_{pe}$.
When the nonlinear IAW excited by the driver is weak, the nonlinear frequency shift (NFS)\cite{(9)} of the IAW is small, thus the resonance peak is very close to the frequency calculated by the dispersion relation with no damping.
In the same way, when $k\lambda_{De}=0.5$, the resonance curves of the fast mode and the slow mode are given as shown in Fig. \ref{Fig:Resonance2}. It's also found that the resonance peak is close to the frequency calculated by the dispersion relation with no damping $\omega_N$. The linear frequency is higher but not much too higher than the resonance frequency, thus the IAW can be excited by the driver with the linear frequency especially when the driver amplitude is large.
When $k\lambda_{De}=0.7$, as shown in Fig. \ref{Fig:Resonance3}, the resonance frequency is lower than the frequency calculated by the dispersion relation with no damping $\omega_N$ and also much lower than the linear frequency $\omega_L$. When $k\lambda_{De}$ is as large as 0.7, the harmonics are very weak and the fluid NFS from harmonic generation is negligible. It's caused by the negative nonlinear frequency shift due to the particles trapping when the IAW amplitude is large.\cite{T. Chapman_PRL, (9)} This phenomenon is common in the nonlinear IAW in the stimulated Brillouin scattering (SBS)\cite{Liu-2011POP} and other BGK modes such as nonlinear electron acoustic waves (EAWs)\cite{Valentini-2006POP}, Langmuir waves\cite{Rose_2001POP} and kinetic electrostatic electron nonlinear (KEEN) waves\cite{Johnston-2009POP}. For the linear frequency $\omega_L$ is far away from the resonance frequency, no matter how large of the driver amplitude, the IAW can nearly not be excited by the driver with the linear frequency.
\subsection{\label{Subsec:B. Large amplitude driver}Large amplitude driver with the linear frequency of the IAW}
\begin{figure*}[!tp]
\includegraphics[width=2\columnwidth]{Fig7.eps}
\caption{\label{Fig:FastMode}(Color online) The phase picture of (a)-(d) electrons, (e)-(h) H ions, (i)-(l) C ions for the fast mode ($T_i/T_e=0.1$) in the condition of $k\lambda_{De}=0.1, 0.3, 0.5, 0.7$ when the external driving electric field is off. The driver is with the frequency of the fast mode calculated by the linear dispersion relation.}
\end{figure*}
To show more clearly the difference of the excitation of the IAW and the phase pictures of particles in different $k\lambda_{De}$, the driver amplitude takes a large value $eE_d\lambda_{De}/T_e=5.33\times10^{-2}$.
As shown in Figs. \ref{Fig:SlowMode}(a) and \ref{Fig:SlowMode}(b), the external driving electric field (driver) is turned on from the time of 0 to $2\times10^5\omega_{pe}^{-1}$ with the duration time $t_0=1\times10^5\omega_{pe}^{-1}$ as the same. The frequency of the driver takes the linear frequency of the IAW as shown in Table I. For example, in Fig. \ref{Fig:SlowMode}(a), $k\lambda_{De}=0.5$, $T_i/T_e=0.5$, so the linear frequency of the slow mode is $Re(\omega_s)=8.250\times10^{-3}\omega_{pe}$, in the same way, in Fig. \ref{Fig:SlowMode}(b), $k\lambda_{De}=0.7$, $T_i/T_e=0.5$, so the frequency of the driver is $1.091\times10^{-2}\omega_{pe}$. After the driver is turned off, the amplitude of the electric field in the system in the condition of $k\lambda_{De}=0.5$ is obviously larger than that in the condition of $k\lambda_{De}=0.7$.
As shown in Fig. \ref{Fig:SlowMode}(a), in the condition of $k\lambda_{De}=0.5$, when the driver is turned on, the electric field amplitude may exhibit slow oscillations on the ion bounce time scale ($\tau_{bi}=2\pi/\sqrt{kq_iE/m_i}$) as the wave and resonant particles exchange energy. This process allows phase locking of the mode to the driver, which will make the distribution flat at the phase velocity thereby reducing Landau damping.\cite{(3)} After the driver is turned off, for the frequency of the driver is near the resonance frequency of the nonlinear IAW as shown in Fig. \ref{Fig:Resonance2}(b), the driver with the linear frequency will couple significant energy into a BGK mode especially when the driver amplitude is large, thus the undriven BGK-like mode with relatively large amplitude will be established, known as auto-resonance which doesn't require feedback to maintain resonance. Thus, the trapping of electrons, H ions and C ions is very obvious as shown in Figs. \ref{Fig:SlowMode}(c)-\ref{Fig:SlowMode}(e) after the driver is turned off.
However, the driver is added into the CH plasmas systems in Fig. \ref{Fig:SlowMode}(b) as same as in Fig. \ref{Fig:SlowMode}(a). For Fig. \ref{Fig:SlowMode}(b), $k\lambda_{De}=0.7$, for the linear frequency $\omega_L$ is far away from the resonance frequency of the slow mode as shown in Fig. \ref{Fig:Resonance3}(b), the driver with the linear frequency $\omega_L$ can nearly not excite a large amplitude IAW by the way of resonance excitation no matter how large of the driver amplitude. As a result, after the driver is off, the IAW amplitude will fall down to a very small level and stabilize at a small but finite amplitude. In the same way, the small amplitude IAW will trap particles although the trapping width is very small, thus making the distribution of particles flat at the phase velocity thereby reducing Landau damping to nearly zero after several bounce periods. Therefore, the IAW stabilize at a very small but nonzero amplitude. This process allow the driver with the linear frequency $\omega_L$ to couple small but nonzero energy into a BGK mode. At the time of $T_3$ (as shown in Fig. \ref{Fig:SlowMode}(b), $T_3$ is the time after the diver is off in the case of $k\lambda_{De}=0.7$, $T_3\sim4\times10^{5}\omega_{pe}^{-1}$), the ion bounce periodic time $\tau_{bi}$ shows several times of the IAW period $\tau_{pi}=2\pi/\omega_{pi}$ when the internal IAW amplitude $eE^{int}\lambda_{De}/T_e$ is as low as $2\times10^{-3}$. Figs. \ref{Fig:SlowMode}(f)-\ref{Fig:SlowMode}(h) show the phase picture of electrons, H ions and C ions at the time of $T_2$ when the driver is on. The trapping of the particles is obvious due to the interaction of the external driving electric field and particles when the driver is on. This process makes the distribution of particles flat at the phase velocity of IAW, $v_\phi$, thereby reducing the Landau damping. When the driver is off, the distribution keeps flat at $v_\phi$, but the trapping width of the particles at the time of $T_3$ (see Figs. \ref{Fig:SlowMode}(i)-\ref{Fig:SlowMode}(k)) is narrower than that at the time of $T_1$ (see Figs. \ref{Fig:SlowMode}(c)-\ref{Fig:SlowMode}(e), where $T_1$ is the time after the driver is off in the case of $k\lambda_{De}=0.5$, $T_1\sim4\times10^{5}\omega_{pe}^{-1}$) as a result of the lower amplitude of IAW electric field when the driver is off.
Fig. \ref{Fig:FastMode} shows the trapping of the particles including electrons, H ions and C ions in the cases of $k\lambda_{De}=0.1$, 0.3, 0.5, 0.7 for the fast mode when the driver is off. The electric fields of the system are shown in Fig. \ref{Fig:Electric_FastMode} (discussed later), where the maximum amplitude of the driver $eE_d\lambda_{De}/T_e=5.33\times10^{-2}$ is taken as an example and the driver frequency takes the linear frequency of the IAW. Here, $v_\phi$, the phase velocity of the modes, are marked in the phase pictures as shown in Fig. \ref{Fig:FastMode}. For the phase velocity of the fast mode, $v_{\phi f}$, is much larger than the thermal velocity of C ions, $v_{tC}$, C ions can not be trapped in all cases for the fast mode (see Figs. \ref{Fig:FastMode}(i)-\ref{Fig:FastMode}(l)), which is different from the trapping of C ions for slow mode (see Figs. \ref{Fig:SlowMode}(e), \ref{Fig:SlowMode}(k)). For the slow mode, the phase velocity, $v_{\phi s}$, is nearly 3-4 times of the thermal velocity of C ions, $v_{tC}$, so C ions can be trapped by the slow mode.
We can find when $k\lambda_{De}=0.1, 0.3, 0.5$, the particles including electrons and H ions show a large trapping width when the driver with the linear frequency of the fast mode is off. In these cases, the linear Landau damping is relatively weak and the linear frequency is not much larger than the resonance frequency as shown in Figs. \ref{Fig:Resonance}(a), \ref{Fig:Resonance}(b) and \ref{Fig:Resonance2}(a), thus the IAW can be excited to a large amplitude especially when the driver amplitude is large. When $k\lambda_{De}=0.7$, particles including electrons, H ions and C ions can nearly not be trapped (see Figs. \ref{Fig:FastMode}(d), \ref{Fig:FastMode}(h) and \ref{Fig:FastMode}(l)). As the relative (or absolute) Landau damping rate of the fast mode, $|Im(\omega_f)/Re(\omega_f)|$ (or $|Im(\omega_f)|$), increases obviously with $k\lambda_{De}$ increasing (see Fig. \ref{Fig:LinearNonlinear}(c)) which will lead to a larger deviation between the linear frequency and the frequency calculated by the dispersion relation with no damping (shown in Fig. \ref{Fig:LinearNonlinear}(a)). In this case of $k\lambda_{De}=0.7$, the deviation between the linear frequency and the frequency calculated by the dispersion relation with no damping is nearly $7\%$ as shown in Fig. \ref{Fig:LinearNonlinear}(a) and the linear frequency is far away from the resonance frequency as shown in Fig. \ref{Fig:Resonance3}(a), thus no matter how large of the driver amplitude, the driver with the linear frequency couples very small energy into the fast mode and the IAW can nearly not be excited to a large amplitude.
\begin{figure}[!tp]
\includegraphics[width=\columnwidth]{Fig8.eps}
\caption{\label{Fig:Electric_FastMode}(Color online) The total electric field of the system (blue line) and the envelope of the external driving electric field (driver, red line) for the fast mode ($T_i/T_e=0.1$) in the condition of (a) $k\lambda_{De}=0.1$, (b) $k\lambda_{De}=0.3$, (c) $k\lambda_{De}=0.5$, (d) $k\lambda_{De}=0.7$, which is corresponding to Fig. \ref{Fig:FastMode}. The driver is with the linear frequency of the fast mode calculated by the linear dispersion relation. }
\end{figure}
To clarify the excitation of the nonlinear ion acoustic wave (IAW) by the driver with the linear frequency of the fast mode when $k\lambda_{De}$ changes, the electric field of the system for the fast mode in some cases of $k\lambda_{De}$ is given in Fig. \ref{Fig:Electric_FastMode}. In the condition of $k\lambda_{De}=0.7$ for the fast mode (shown in Fig. \ref{Fig:Electric_FastMode}(d)), when the driver is turned on, the electric field of the system will show a large response to the driver because of the internal static electric field generated by the interaction of the driving electric field and the particles, which is the same as the process in the condition of $k\lambda_{De}=0.1, 0.3, 0.5$. However, in the condition of $k\lambda_{De}=0.7$, when the driver is off, the driver with the linear frequency appears to couple small but nonzero energy into the fast mode and the amplitude of the internal static electric field will fall down to a very low level (see Fig. \ref{Fig:Electric_FastMode}(d)). For the linear Landau damping of the fast mode increases with $k\lambda_{De}$ as shown in Fig. \ref{Fig:LinearNonlinear}(c), the linear Landau damping of the fast mode in the case of $k\lambda_{De}=0.7$ is relatively stronger than that in the case of $k\lambda_{De}=0.5, 0.3$ or 0.1, which will lead to a larger deviation between the linear frequency and the frequency calculated by the dispersion relation with no damping (as shown in Fig. \ref{Fig:LinearNonlinear}(a)), as a result, the linear frequency of the fast mode is far away from the resonance frequency as shown in Fig. \ref{Fig:Resonance3}(a), which is different from that in the condition of $k\lambda_{De}=0.1, 0.3, 0.5$. As shown in Figs. \ref{Fig:Electric_FastMode}(a)- \ref{Fig:Electric_FastMode}(c), when the driver is off, the amplitude of the internal electric field in CH plasmas keeps on a relatively larger level which leads to the trapping of particles with a large trapping width. This indicates that when $k\lambda_{De}$ is not large, the nonlinear IAW can be excited by the driver with the linear frequency of the modes. However, in the condition of large $k\lambda_{De}$, especially when $k\lambda_{De}$ is as large as 0.7, the nonlinear ion acoustic wave can nearly not be excited by the driver with linear frequency of the modes.
\subsection{\label{Subsec:C}Driver with the linear frequency and the frequency calculated by the dispersion with no damping of the IAW}
\begin{figure}[!tp]
\includegraphics[width=\columnwidth]{Fig9.eps}
\caption{\label{Fig:k=0.7}(Color online) The excitation of the nonlinear ion acoustic wave by the driver with (a), (b) the linear frequency and (c), (d) the frequency calculated by the dispersion with no damping. (a), (c) are the excitation of the fast mode, $T_i/T_e=0.1$, (b), (d) are the excitation of the slow mode, $T_i/T_e=0.5$. All the cases are in the condition of $k\lambda_{De}=0.7$ and with the same driver with the maximum amplitude $eE_d\lambda_{De}/T_e=7\times10^{-3}$. Where the vertical axis $E_x=eE\lambda_{De}/T_e$.}
\end{figure}
To verify whether the driver with the frequency calculated by the dispersion relation with no damping can excite a large amplitude nonlinear IAW especially when $k\lambda_{De}=0.7$, the cases of the driver with the frequency calculated by the dispersion with no damping and the linear frequency of the IAW are taken to excite the IAW. As shown in Fig. \ref{Fig:k=0.7}, in the condition of $k\lambda_{De}=0.7$, the driver with the linear frequency of the IAW ($\omega_d=\omega_L=0.01147\omega_{pe}$ for the fast mode, $\omega_d=\omega_L=0.01091\omega_{pe}$ for the slow mode) can nearly not excite the IAW as shown in Figs. \ref{Fig:k=0.7}(a) and \ref{Fig:k=0.7}(b). However, the driver with the frequency calculated by the dispersion with no damping of the IAW ($\omega_d=\omega_N=0.01060\omega_{pe}$ for both the fast mode and the slow mode, the frequencies of the fast mode and the slow mode in this condition are the same by coincidence, see Table \ref{table1}) can excite a large amplitude nonlinear IAW after several trapping periods as shown in Figs. \ref{Fig:k=0.7}(c) and \ref{Fig:k=0.7}(d). Although the frequency has a negative shift due to the particles trapping as shown in Fig. \ref{Fig:Resonance3}, the frequency calculated by the dispersion relation with no damping $\omega_N$ is also closer to the resonance frequency than the linear frequency $\omega_L$, thus the driver with $\omega_N$ can couple significant energy into the BGK modes as shown in Figs. \ref{Fig:k=0.7}(c) and \ref{Fig:k=0.7}(d). This illustrates that the linear frequency of the modes is far away from the resonance frequency and the frequency calculated by the dispersion relation with no damping is closer to the resonance frequency of the nonlinear IAW than that calculated by the linear dispersion relation. If the IAW amplitude is small enough, the NFS will be very small and the resonance frequency will be very close to the frequency calculated by the dispersion relation with no damping, as shown in Fig. \ref{Fig:Resonance}.
\section{\label{Sec:Discussion}Discussion}
When the nonlinear IAW amplitude excited by the driver is weak, the nonlinear frequency shift (NFS) of the nonlinear IAW is small. The nonlinear frequency shift from particles trapping is proportional to the square root of the potential amplitude of the nonlinear IAW, $\delta \omega_N^{kin} \propto \sqrt{e\phi/T_e}$, where $\phi=E/k$, and the NFS from harmonic generation is proportional to the square of the potential amplitude of the nonlinear IAW, $\delta \omega_N^{flu} \propto (e\phi/T_e)^2$.\cite{(9)} The maximum amplitude of the driver $eE_d^{max}\lambda_{De}/T_e=1\times10^{-3}$ is taken in the cases of the resonance curve of the fast mode in the condition of $k\lambda_{De}=0.1, 0.3$ as shown in Fig. \ref{Fig:Resonance}. The corresponding peak amplitude of nonlinear IAW is very small, $eE\lambda_{De}/T_e\sim3\times10^{-4}$ in the condition of $k\lambda_{De}=0.1$ and $eE\lambda_{De}/T_e\sim6\times10^{-3}$ in the condition of $k\lambda_{De}=0.3$. In these cases, the nonlinear frequency shift is very small, which ensures that the NFS have a small effect on the resonance frequency of the nonlinear IAWs. These cases as shown in Fig. \ref{Fig:Resonance} can illustrate that the frequency calculated by the dispersion relation with no damping is much closer to the resonance frequency than the linear frequency when the NFS can be ignored.
However, when the large amplitude nonlinear IAW is excited by the large amplitude external driving electric field, the NFS may need to be considered. Figs. \ref{Fig:SlowMode}-\ref{Fig:Electric_FastMode} show the nonlinear IAW with large amplitude $eE\lambda_{De}/T_e\sim3\times10^{-2}$. In the condition of $k\lambda_{De}=0.1, 0.3, 0.5$, for the linear frequency $Re(\omega_L)$ is near the resonance frequency $\omega_N+\delta \omega$, the nonlinear IAW can be excited to a large amplitude. In this paper, we think the frequency calculated by the dispersion relation with no damping $\omega_N$ is the fundamental frequency, $\delta\omega$ is the quantity of the nonlinear frequency shift relative to the fundamental frequency. While in the condition of $k\lambda_{De}=0.7$, the harmonic effect is weak and the positive fluid NFS from harmonic generation is ignored. The NFS of the nonlinear IAW mainly comes from the kinetic NFS due to particles trapping, and in this condition, the kinetic NFS is negative. As shown in Fig. \ref{Fig:Resonance3}, the frequency calculated by the dispersion relation with no damping $\omega_N$ is much closer to the resonance frequency than the linear frequency $\omega_L$.
Therefore, the driver with $\omega_N$ can excite a large amplitude IAW as shown in Figs. \ref{Fig:k=0.7}(c) and \ref{Fig:k=0.7}(d), while the driver with the linear frequecy $\omega_L$ couples very small energy into a BGK mode and can nearly not excite a large amplitude IAW no matter how large of the driver amplitude (the fast mode as shown in Figs. \ref{Fig:Electric_FastMode} (d), \ref{Fig:k=0.7} (a), the slow mode as shown in Figs. \ref{Fig:SlowMode} (b), \ref{Fig:k=0.7} (b)).
\section{\label{Sec:Summary}Summary}
The excitation of the nonlinear ion acoustic modes including the fast mode and the slow mode by the driving electric field with the linear frequency or the frequency calculated by the dispersion with no damping of the IAW when $k\lambda_{De}$ varies has been shown. For the IAW with low Landau damping can be excited preferentially, only the IAWs in the low Landau damping region are considered, such as the fast mode in the condition of $T_i/T_e=0.1$, the slow mode in the condition of $T_i/T_e=0.5$.
When $k\lambda_{De}$ increases, the linear Landau damping, $|Im(\omega)/Re(\omega)|$, of the fast mode increases obviously in the region of low Landau damping ($T_i/T_e\lesssim0.2$). This provides the possibility of the suppression of SBS in the region of large wave number , i.e., the region of high temperature and low density of the electrons. However, this is the linear result. Considering the particle trapping, the distribution of particles will keep flat at the phase velocity, thus turning off Landau damping. As a result, the dispersion relation with no damping is provided to calculate the frequency of the nonlinear IAWs. When $k\lambda_{De}$ increases, the frequency calculated by the linear dispersion relation and that calculated by the dispersion relation with no damping deviate each other larger and larger. Especially, in the condition of $k\lambda_{De}=0.7$ for the fast mode, the deviation is as large as nearly $7\%$.
By Vlasov simulation, the resonance curves show that the resonance frequency is much closer to the frequency calculated by the dispersion relation with no damping than that calculated by the linear dispersion relation. When $k\lambda_{De}$ is not large, such as $k\lambda_{De}=0.1, 0.3, 0.5$, the frequency of the linear fast mode in the condition of $T_i/T_e=0.1$ and the linear slow mode in the condition of $T_i/T_e=0.5$ are near the resonance frequency of the nonlinear IAW modes, thus, the nonlinear IAW can be excited by the driver with the linear frequency of the modes. However, in the condition of $k\lambda_{De}=0.7$, for the linear frequency is far away from the resonance frequency of the nonlinear IAW, the nonlinear IAW can nearly not be excited by the external driver with the linear frequency of the IAW (no matter how large of the amplitude of the driver). While the driver with the frequency calculated by the dispersion relation with no damping can excite the large amplitude nonlinear IAW.
Our findings indicate: The frequency calculated by the dispersion relation with no damping is much closer to the resonance frequency of the small amplitude nonlinear IAW than the linear frequency. When $k\lambda_{De}$ is large, such as $k\lambda_{De}=0.7$, the linear frequency can not be applied to exciting the nonlinear IAW, while the frequency calculated by the dispersion relation with no damping can be applied to exciting the nonlinear IAW.
\begin{acknowledgments}
We are pleased to acknowledge useful discussions with L. H. Cao. This research was supported by the National Natural Science Foundation of China (Grant Nos. 11575035, 11475030 and 11435011) and National Basic Research
Program of China (Grant No. 2013CB834101).
\end{acknowledgments}
|
\section{Introduction}
Recent years have seen tremendous progress on the understanding of
planar models in statistical physics, particularly the (classical)
Ising model at criticality. A major breakthrough in this area was
the definition, and proof of convergence to conformally covariant
scaling limits,
of fermionic observables in the critical Ising
model, first on the square lattice by Smirnov~\cite{smi-towards,smi-fk},
and later on all isoradial graphs
by Chelkak and Smirnov~\cite{ch-sm}.
The fermionic observables enjoy a crucial property
called s-holo{-}morphicity, a strong form of discrete analyticity.
Besides satisfying a discrete version of the Cauchy--Riemann
relations, if a function $F_\d$
is s-holomorphic then
one may define a discrete primitive
$H_\d=\Im\big(\int^\d F_\d^2\big)$ of its
\emph{square}.
Moreover this function $H_\d$ is very close to being
(discrete) harmonic.
When combined with control of the behaviour of $H_\d$ at the boundary
of the domain, this allows to deduce convergence of the
fermionic observables from convergence of solutions
to discrete boundary-value problems.
The identification of these and related observables and their scaling limits has
subsequently led to some outstanding results on the
critical planar Ising model, settling several predictions from
conformal field theory. This includes convergence of
the energy-density~\cite{en-dens},
correlation functions~\cite{correlations}, as well as
interfaces to SLE-curves~\cite{interfaces} and
loops to CLE-processes~\cite{ben-hong,ke-sm}, to mention but a few.
There has also been work on extending the definition of
s-holomorphicity to general graphs~\cite{cimasoni}.
In this note we start to consider similar questions in the context of
one-dimensional quantum spin-systems, specifically the
transverse-field (quantum) Ising model, hereafter abbreviated {\sc tfim}.
This model has Hamiltonian given by
\begin{equation}\label{ham_eq}
-\mathcal{H}_N=J\sum_{x=1}^{N-1} \sigma_x^{(3)}\sigma_{x+1}^{(3)}
+h\sum_{x=1}^N \sigma_x^{(1)},
\mbox{ acting on } \otimes_{x=1}^N \mathbb{C}^2,
\end{equation}
where
$\sigma^{(3)}=\big(\begin{smallmatrix}
1 & 0 \\
0 & -1
\end{smallmatrix}\big)$ and
$\sigma^{(1)}=\big(\begin{smallmatrix}
0 & 1 \\
1 & 0
\end{smallmatrix}\big)$
are the spin-$\tfrac12$ Pauli matrices,
and $J,h>0$ give the coupling- and transverse-field-strengths,
respectively. (For $h=0$ this is just the classical Ising model.)
We will be working with the ground-state (zero
temperature), where the model is known to undergo a phase-transition
as the ratio $h/J$ is varied, at the critical point $h/J=1$~\cite{pfeuty}.
The phase-transition is continuous~\cite{bjogr}.
It is well-known that the {{\sc tfim}}
in $d$ dimensions possesses a graphical, probabilistic
representation in $\mathbb{Z}^d\times\RR$, and it behaves in many ways like a
classical Ising model in $d+1$ dimensions, see e.g.\ the results
in~\cite{B-irb,B-van}. One may thus ask if
the results mentioned above, on conformal invariance in the
two-dimensional classical Ising model at criticality,
have analogs in the one-dimensional
quantum model?
This note is a first step in this direction. We introduce a notion of
s-holomorphicity for functions on
$\mathbb{Z}+i\RR\subseteq\mathbb{C}$; we
show that functions that satisfy this enjoy (analogs of) the key
properties that hold in the classical case;
and we define two observables in the critical
{{\sc tfim}} which we show to be
s-holomorphic.
The graphical representations
that we consider may be obtained as
limits of classical counterparts on $\mathbb{Z}+i(\varepsilon\mathbb{Z})$ as
$\varepsilon\to0$. The latter graphs are all isoradial, and some of the key
quantities we work with can be interpreted as limits of the
corresponding quantities for isoradial graphs~\cite{ch-sm}.
We give examples of this
in Section~\ref{isorad-sec}.
However, for all our definitions and results we work
directly in the `continuous' setting $\mathbb{Z}+i\RR$ and the
rescaled version $\d\mathbb{Z}+i\RR$.
We do not go into the details for scaling limits (as $\d\to0$)
of our observables here, but we expect this
to be very similar to the classical case. As we discuss in
Section~\ref{disc-sec}, we expect
analogous reasoning and estimates to show
that our
observables converge to the \emph{same} scaling-limits as their
classical counterparts.
\subsection*{Outline and main contributions}
After reviewing the graphical
representations of the {{\sc tfim}} in Section~\ref{graph-sec}, we
give our definition of s-holomorphicity in Section~\ref{s-hol-sec},
and prove some key properties of s-holomorphic functions in
Proposition~\ref{H-prop}.
We introduce and study our two fermionic observables in
Sections~\ref{fk-obs-sec} and~\ref{spin-obs-sec}, respectively. The
main results
are that these observables satisfy our definition of s-holomorphicity,
stated precisely in Theorems~\ref{s-hol-thm-fk}
and~\ref{s-hol-thm-spin}.
\subsection*{Bibliographical remark}
Shortly after this paper was made public, Li~\cite{li}
announced a complete proof of convergence of
the {{\sc fk}}-observable considered here, as well as the
{{\sc fk}}-interface to SLE$_{16/3}$, in the scaling limit.
Li independently arrived at equivalent definitions of the
{{\sc fk}}-observable and s-holomorphicity as presented here,
and supplied the details necessary to prove convergence.
He does not consider the spin-observable. Most likely
his results are useful for proving convergence of that
observable as well.
\section{Graphical representations of the TFIM}
\label{graph-sec}
We briefly review three graphical representations of the {{\sc tfim}}.
They may be obtained using a Lie--Trotter expansion, see
e.g.~\cite{akn,bjo_phd,bjogr,ioffe_geom} for details.
We also present a version of the Kramers--Wannier duality; as for the
classical case, this allows us to easily identify the critical parameters
of the model (but for rigorous proofs
see~\cite{pfeuty,bjogr}).
We write the partition function
$\mathcal{Z}_{N,\beta}=\mathcal{Z}_{N,\beta}(h,J)=\mathrm{tr}(e^{-\beta \mathcal{H}_N})$
where $\mathcal{H}_N$ is the Hamiltonian~\eqref{ham_eq}, and
$\beta>0$ is the inverse-temperature. For illustration we will also
consider the two-point correlation
\[
\el\sigma^{(3)}_x\sigma^{(3)}_y\rangle_{N,\beta}=
\mathrm{tr}(\sigma^{(3)}_x\sigma^{(3)}_ye^{-\beta \mathcal{H}_N})/\mathcal{Z}_{N,\beta}.
\]
Thermodynamic limits are obtained for $N\to\infty$, and
the ground-state is obtained by also letting
$\beta\to\infty$.
The {{\sc tfim}} on $\{1,\dotsc,N\}$
maps onto stochastic models in the rectangular domain
$\Om=[1,N]+i[0,\beta]\subseteq\mathbb{C}$. We write
\begin{equation}
\Om^\b=\{1,\dotsc,N\}+i[0,\beta],\quad
\Om^\circ=(\sfrac12+\{1,\dotsc,N-1\})+i[0,\beta].
\end{equation}
We will let $\xi^\b$ and $\xi^\circ$ denote independent Poisson processes
on $\Om^\b$ and $\Om^\circ$, respectively. Their respective rates
will be denoted $r^\b$ and $r^\circ$ and will be functions of $h$ and
$J$. We
write $\EE_{r^\b,r^\circ}[\cdot]$ for the law (expectation operator)
governing them, and $\xi=\xi^\b\cup\xi^\circ$. Elements of $\xi^\b$ will
be represented graphically by $\times$ and called `cuts'; an element
$(x+\sfrac12)+it$ of $\xi^\circ$ will be represented as a horizontal
line-segment between $x+it$ and $(x+1)+it$ and called a `bridge'.
The interpretation of these objects will
differ slightly for the three different representations, as we now
describe. See
Figures~\ref{fk-rep-fig} and~\ref{kw_fig} for examples.
\subsection{FK-representation}
\label{fk-sec}
For this representation we set $r^\b=h$ and $r^\circ=2J$. We interpret
the cuts $x+it\in\xi^\b$ as \emph{severing} a line-segment
$x+i[0,\beta]$, and the
bridges $\xi^\circ$ as \emph{connecting} neighbouring line segments. Thus the
configuration $\xi$ is a partly continuous
percolation-configuration. The
maximal connected subsets of $\Om^\b$ are called \emph{components},
and their number is denoted $k^\b(\xi)$. The components may be defined
with respect to various different boundary conditions,
but for now we only consider the
`vertically periodic' boundary condition, meaning that the points at
the top and bottom of $\Om^\b$ are identified (i.e.\ we treat
$[0,\beta]$ as a circle).
See Figure~\ref{fk-rep-fig}.
The {\sc fk}-representation expresses
\begin{equation}\label{fk}
\mathcal{Z}_{N,\beta}=e^{\beta J(N-1)} \EE_{h,2J}[2^{k^\b(\xi)}],
\quad
\el\sigma^{(3)}_x\sigma^{(3)}_y\rangle_{N,\beta}=
\frac{\EE_{h,2J}[\hbox{\rm 1\kern-.27em I}\{x\leftrightarrow y\}2^{k^\b(\xi)}]}
{\EE_{h,2J}[2^{k^\b(\xi)}]},
\end{equation}
where $\{x\leftrightarrow y\}$ denotes the event that $x,y\in\{1,\dotsc,N\}$
belong to the same connected component.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.44}\hspace{2cm}
\includegraphics{fermionic.45}
\caption{
\emph{Left:} Illustration of the {\sc fk}-representation. Cuts ($\times$)
disconnect, bridges (horizontal line segments) connect, and top and
bottom of the intervals are identified. The number $k(\xi)$ of
components is 5.
\emph{Right:} The same {\sc fk}-sample $\xi$ (solid) with its dual $\xi'$
(dashed).
}
\label{fk-rep-fig}
\end{figure}
With an {\sc fk}-configuration $\xi$ we can associate a \emph{dual}
configuration $\xi'$, whose connected components are subsets of
$\Om^\circ$ rather than $\Om^\b$. For simplicity we describe this in
the case when $\xi^\b$ has no cuts on the left- or rightmost intervals
$1+i[0,\beta]$ and $N+i[0,\beta]$. We obtain $\xi'$ by drawing a
bridge from $(x-\sfrac12)+it$ to $(x+\sfrac12)+it$ for each cut
$x+it\in\xi^\b$, and placing a cut $\times$ at $(x+\sfrac12)+it$
whenever $\xi^\circ$ has a bridge there. See Figure~\ref{fk-rep-fig}.
Objects, such as cuts, bridges and components, pertaining to $\xi'$
will be referred to as \emph{dual} and those of $\xi$ as \emph{primal}
when a distinction needs to be made. The number of dual components
will be denoted $k^\circ(\xi)$.
It turns out that $\xi'$ also has the law of a {\sc fk}-configuration, with
adjusted parameters.
We will return to this construction when we define the {\sc fk}-observable
in Section~\ref{fk-obs-sec}.
\subsection{Random-parity representation}
For this representation we set $r^\b=0$ and $r^\circ=J$, thus there are
only bridges. We use auxiliary configurations $\psi\in\{0,1\}^N$
together with a fixed, finite subset $A\subseteq\Om^\b$ of
\emph{sources}. The
configuration $\psi$ is extended to a function
$\psi_A:\Om^\b\to\{0,1\}$,
in a way which depends on $\xi^\circ$ and $A$, using the following
rules.
The function $\psi_A(x+it)$ is equal to $\psi(x)$ for $t$ from 0 to the
first time of either a bridge $(x\pm\sfrac12)+it\in\xi^\circ$,
\emph{or} a source $x+it\in A$. At such a point it switches
to $1-\psi(x)$. Then it stays at that value until it encounters
another bridge-endpoint or source, where it switches back to
$\psi(x)$; and so
on. See Figure~\ref{kw_fig} for an example.
The subset of $\Om^\b$ where $\psi_A$ takes value 1 is denoted
$I(\psi_A)=\psi_A^{-1}(1)$, and will for definiteness be taken to be
closed. We denote its total length $|I(\psi_A)|$. We will only be
considering the cases when either $A=\varnothing$ or $A$ consists of two
points; in the former case $I(\psi_A)$ consists of a collection of
loops, in the latter case loops plus a unique path connecting the two
points of $A$.
We impose the periodicity constraint that
$\psi(x+i\beta)=\psi(x)$ for all $x\in\{1,\dotsc,N\}$;
if $x\in A$ then the correct
interpretation is $\psi(x+i\beta)=1-\psi(x)$ due to the
switching-rule.
Hence we discount some
configurations $\xi$, specifically those where some line
$x+i[0,\beta]$ meets an odd number of switching-points.
As we will see presently, this discounting can be done
formally by redefining
$|I(\psi_A)|=\infty$ when the constraint is violated.
The random-parity representation expresses
\begin{equation}\begin{split}\label{rpr}
&\mathcal{Z}_{N,\beta}=e^{\beta h N+\beta J(N-1)}
\EE_{0,J}\Big[\sum_{\psi\in\{0,1\}^N}\exp(-2h|I(\psi_\varnothing)|)\Big],\\
&\el\sigma^{(3)}_x\sigma^{(3)}_y\rangle_{N,\beta}=\frac{
\EE_{0,J}\Big[\sum_{\psi\in\{0,1\}^N}\exp(-2h|I(\psi_{\{x,y\}})|)\Big]
}
{\EE_{0,J}\Big[\sum_{\psi\in\{0,1\}^N}\exp(-2h|I(\psi_\varnothing)|)\Big]}.
\end{split}\end{equation}
This representation is a quantum version of Aizenman's
random-current representation~\cite{aiz82}.
There is a notion of planar duality also for this representation,
mapping onto the space--time spin representation, which we describe
now.
\subsection{Space--time spin representation}
This representation plays a less prominent role in this note,
and is mainly interesting since it is dual to the random-parity
representation.
We now set $r^\b=h$ and $r^\circ=0$, thus there
are only cuts. We let $\S(\xi)$ denote the set of functions
$\sigma:\Om^\b\to\{-1,+1\}$ which are constant between points of $\xi^\b$,
change value at the points of $\xi^\b$, and satisfy the periodicity
constraint $\sigma(x)=\sigma(x+i\beta)$ for all $x\in\{1,\dotsc,N\}$.
See Figure~\ref{kw_fig}.
(For definiteness we may take $\sigma^{-1}(+1)$ to be closed; also note
that for some $\xi$ we have $\S(\xi)=\varnothing$.)
For readability we also write $\sigma_x(t)$ for $\sigma(x+it)$.
The space--time spin representation expresses
\begin{equation}\label{stim}
\begin{split}
&\mathcal{Z}_{N,\beta}=e^{\beta h N}
\EE_{h,0}\Big[
\sum_{\sigma\in\S(\xi)}\exp\Big( J\sum_{z=1}^{N-1}\int_0^\beta
\sigma_z(t)\sigma_{z+1}(t)\, dt\Big)\Big],\\
&\el\sigma^{(3)}_x\sigma^{(3)}_y\rangle_{N,\beta}=\frac{
\EE_{h,0}\Big[\sum_{\sigma\in\S(\xi)}
\sigma(x)\sigma(y)\exp\Big( J\sum_{z=1}^{N-1}\int_0^\beta
\sigma_z(t)\sigma_{z+1}(t)\, dt\Big)\Big]
}{
\EE_{h,0}\Big[
\sum_{\sigma\in\S(\xi)}\exp\Big( J\sum_{z=1}^{N-1}\int_0^\beta
\sigma_z(t)\sigma_{z+1}(t)\, dt\Big)\Big]}.
\end{split}
\end{equation}
\subsection{Kramers--Wannier duality}
We now describe a duality between the random-parity and
spin-representations. We will associate (in a reversible way) to a
spin-configuration $\sigma:\Om^\b\to\{-1,+1\}$ a `dual'
random-parity-configuration
$\psi=\psi_\varnothing:\Om^\circ\to\{0,1\}$. Note that the domain of $\psi$ is
$\Om^\circ$ rather than $\Om^\b$.
We impose the `wired' boundary condition
\begin{equation}\begin{split}
\sigma(1+it)=\sigma(N+it)=\sigma(x)&=\sigma(x+i\beta)=+1,\\
&\mbox{for all } t\in[0,\beta], x\in\{1,\dotsc,N\}.
\end{split}\end{equation}
As we will see, this will automatically lead to the boundary condition
\begin{equation}
\psi(x+\sfrac12)=\psi((x+\sfrac12)+i\beta)=0,
\mbox{ for all } x\in\{1,\dotsc,N\}.
\end{equation}
Subject to the boundary conditions,
the sums over $\sigma$ in~\eqref{stim}
and $\psi$ in~\eqref{rpr}
contribute with at most one nonzero term each, hence they will not be
written out.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.46}\hspace{2cm}
\includegraphics{fermionic.4}
\caption{
\emph{Left:} Sample of the random-parity representation with source
set $A=\{a,b\}$.
Intervals where $\psi=1$ are drawn bold, with red
for the unique path between $a$ and $b$ and blue for the loops.
\emph{Right:} Duality between the space--time spin and random-parity
representations. Values $+$ and $-$ indicate the value of $\sigma(z)$
on the corresponding interval in $\Om^\b$,
and these values flip at cuts $\times$.
Blue vertical intervals mark where $\psi(z)=1$.}
\label{kw_fig}
\end{figure}
We construct $\psi$ from $\sigma$ as follows, see Figure~\ref{kw_fig}.
If two neighbouring points $x+it$ and
$(x+1)+it$ have the same spin-value,
$\sigma(x+it)=\sigma((x+1)+it)$, then we set
$\psi((x+\sfrac12)+it)=0$;
otherwise if $\sigma(x+it)\neq\sigma((x+1)+it)$,
then we set $\psi((x+\sfrac12)+it)=1$.
If $x+it\in \xi^\b$ is a
point of spin-flip for $\sigma$, we draw a bridge between
$(x-\sfrac12)+it$
and $(x+\sfrac12)+it$. Thus the bridges form a Poisson
process of rate $h$.
Writing $\mathcal{Z}^+_{N,\beta}(h,J)$ for the partition function~\eqref{stim}
associated with the spin-configurations,
we have that
\begin{equation}\begin{split}
\mathcal{Z}^+_{N,\beta}(h,J)
&=e^{\beta h N}\EE_{h,0}\Big[
\exp\Big(J\sum_{x=1}^{N-1}\int_0^\beta
\sigma_x(t)\sigma_{x+1}(t)\, dt\Big)\Big]\\
&=e^{\beta h N}\EE_{h,0}\Big[
\exp\Big(J\sum_{x=1}^{N-1}\int_0^\beta
[1-2\psi((x+\sfrac12)+it)]\, dt\Big)\Big]\\
&= e^{\beta h N+\beta J (N-1)}
\EE_{h,0}\big[\exp(-2J |I(\psi)|)\big].
\end{split}\end{equation}
Comparing with~\eqref{rpr}, we see that the last factor
\begin{equation}
\EE_{h,0}\big[\exp(-2J |I(\psi)|)\big]=
e^{-\beta J(N-1)-\beta h (N-2)}
\mathcal{Z}^{0}_{N-1,\beta}(J,h),
\end{equation}
where $\mathcal{Z}^{0}_{N-1,\beta}(J,h)$ is the partition function associated with
the $\psi$:s with the prescribed boundary condition. Note that the
order of the parameters $h,J$ is swapped.
We conclude that
\begin{equation}
\mathcal{Z}^+_{N,\beta}(h,J)=e^{2\beta h} \mathcal{Z}^0_{N-1,\beta}(J,h).
\end{equation}
Assuming (as can be justified) the existence of the limit as well as its
independence of the boundary condition, we deduce that the free energy
$f(h,J)=\lim_{N,\beta\to\infty} \tfrac{1}{\beta N} \log \mathcal{Z}_{N,\beta}(h,J)$
satisfies $f(h,J)=f(J,h)$.
This symmetry is consistent with a phase-transition at
$h=J$.
In the rest of this note we consider only the critical case,
$h=J$.
\section{S-holomorphic functions}
\label{s-hol-sec}
\subsection{Discrete domains}
\label{dom-sec}
As indicated above, we will be considering functions on
(bounded subsets of) $\d\mathbb{Z}+i\RR\subseteq\mathbb{C}$.
We use the notation
\[
\mathbb{C}_\d^\b=\d \mathbb{Z}+ i\RR,\quad
\mathbb{C}_\d^\circ=\mathbb{C}_\d^\b+\sfrac\d2, \quad \mbox{and}\quad
\mathbb{C}^\diamondsuit_\d=(\mathbb{C}^\b_\d\cup\mathbb{C}^\circ_\d)+\sfrac\d4.
\]
We will sometimes refer to points of $\mathbb{C}^\b_\d$ as
\emph{primal} or \emph{black}, points of $\mathbb{C}^\circ_\d$ as
\emph{dual} or \emph{white}, and points of
$\mathbb{C}^\diamondsuit_\d$ as \emph{medial}.
See Figure~\ref{dom-fig}
for illustrations of the definitions that follow.
Let $\partial_\d:[0,1]\to\mathbb{C}$ be
a simple closed rectangular path, consisting of
vertical and horizontal line segments, whose
vertical segments are restricted to $\mathbb{C}_\d^\b$.
Let $\Om_\d$ denote the bounded component of
$\mathbb{C}\setminus\partial_\d[0,1]$.
Such a domain $\Om_\d$ will be referred to as
a \emph{primal (discrete) domain}. We also write,
for $\ast\in\{\b,\circ\}$,
\begin{equation}\begin{split}\label{doms-not}
\Om_\d^\ast=\overline{\Om_\d}\cap\mathbb{C}^\ast_\d,\quad
\partial\Om_\d^\ast=\Om_\d^\ast\cap\partial\Om_\d,\quad
\Om_\d^{\ast,\mathrm{int}}=\Om_\d^\ast\setminus\partial\Om_\d^\ast.
\end{split}\end{equation}
Note that $\Om_\d^\ast$ consists of a collection of
vertical line segments, and $\partial\Om_\d^\ast$ of
vertical line segments together with a finite number
of points (forming the hortizontal part of the boundary).
We similarly define a \emph{dual (discrete) domain} $\Om_\d$
by shifting the above definition by
$\sfrac\d2$ (thus swapping $\mathbb{C}^\b_\d$ and $\mathbb{C}^\circ_\d$).
We will also consider Dobrushin domains. For this we let
$a_\d,b_\d\in\mathbb{C}_\d^\diamondsuit$ be two distinct medial points, and
let $\partial_\d:[0,1]\to\mathbb{C}$ be a simple closed
\emph{positively oriented} rectangular path, satisfying
\[
\partial_\d(0)=\partial_\d(1)=a_\d,\quad
\partial_\d(\sfrac12)=b_\d.
\]
We define $\partial_\d^\b,\partial_\d^\circ:[0,1]\to\mathbb{C}$ by
\[
\partial_\d^\circ(t)=\partial_\d(t/2),\quad
\partial_\d^\b(t)=\partial_\d(1-t/2),\quad t\in[0,1].
\]
Thus $\partial_\d^\circ$ goes from $a_\d$ to $b_\d$ in the
counter-clockwise direction, and
$\partial_\d^\b$ goes from $a_\d$ to $b_\d$ in the
clockwise direction.
Finally we assume that the vertical segments of
$\partial_\d^\b$ and $\partial_\d^\circ$ belong to
$\mathbb{C}^\b_\d$ and $\mathbb{C}^\circ_\d$, respectively.
Again we write $\Om_\d$ for the bounded component of
$\mathbb{C}\setminus\partial_\d[0,1]$, and we refer to the triple
$(\Om_\d,a_\d,b_\d)$ as a \emph{discrete Dobrushin domain}.
We define $\Om_\d^\b$, $\partial\Om_\d^\b$,
$\Om_\d^{\b,\mathrm{int}}$, as well as
$\Om_\d^\circ$, $\partial\Om_\d^\circ$,
$\Om_\d^{\circ,\mathrm{int}}$, as in~\eqref{doms-not}.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.1}\hspace{1cm}
\includegraphics{fermionic.2}
\caption{
\emph{Left:}
A primal domain $\Om_\d$. The boundary is drawn with solid
black lines, while $\Om_\d^\b$ consists of the solid black and gray
vertical lines and $\Om_\d^\circ$ of the dashed gray vertical lines.
\emph{Right:} A Dobrushin domain $\Om_\d$ with $\partial^\b_\d$ drawn
solid and $\partial^\circ_\d$ dashed.
}
\label{dom-fig}
\end{figure}
For a primal, dual or Dobrushin domain
$\Om_\d$, and $\ast\in\{\b,\circ\}$,
we define the \emph{vertical}
and \emph{horizontal} parts of the boundary $\partial\Om_\d^\ast$ by
\begin{equation}\begin{split}
\partial^\mathrm{v}\Om^\ast_\d&=\{z\in\partial\Om^\ast_\d:
z+\varepsilon\not\in\Om_\d\mbox{ or } z-\varepsilon\not\in\Om_\d
\mbox{ for small enough }\varepsilon>0\},\\
\partial^\mathrm{h}\Om^\ast_\d&=\{z\in\partial\Om^\ast_\d:
z+ i\varepsilon\not\in\Om_\d\mbox{ or }
z- i\varepsilon\not\in\Om_\d\mbox{ for small enough }\varepsilon>0\}.
\end{split}\end{equation}
We also let
$\partial^\mathrm{v}\Om_\d=\partial^\mathrm{v}\Om^\b_\d\cup \partial^\mathrm{v}\Om^\circ_\d$
and
$\partial^\mathrm{h}\Om_\d=\partial^\mathrm{h}\Om^\b_\d\cup \partial^\mathrm{h}\Om^\circ_\d$.
In words, $\partial^\mathrm{v}\Om_\d$ consists of the vertical
segments of $\partial\Om_\d$, and $\partial^\mathrm{h}\Om_\d$ of the
endpoints of segments in $\overline{\Om_\d}$.
We finally make the assumption on $\Om_\d$ that if
$z\in\partial^\mathrm{v}\Om_\d$ then at least one of $z\pm\sfrac\d2$ belongs
to the interior $\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}}$.
In what follows we will consider triples $(\Om_\d,a_\d,b_\d)$
which are either discrete Dobrushin domains,
alternatively discrete
primal or dual domains with two marked points
$a_\d,b_\d\in\partial\Om_\d$.
One may think of these as approximating
a simply connected domain $\Om\subseteq\mathbb{C}$ with
two marked points $a,b$ on its boundary.
\subsection{S-holomorphic functions}
Let $\Om_\d$ be a discrete domain, as above, and
$F:\Om_\d\to\mathbb{C}$ a function. We will be using the notation
\begin{equation}
\dot F(z):=\lim_{\varepsilon\to 0}\frac{F(z+i\varepsilon)-F(z)}{\varepsilon},
\quad \mbox{with }\varepsilon\in\RR,
\end{equation}
for the derivative of $F$ in the `vertical' direction, when it exists.
We similarly write $\ddot F(z)$ for the second derivative.
For a complex number $\zeta$, with $|\zeta|=1$,
and $z\in\mathbb{C}$, we write
\begin{equation}\label{proj-def}
\mathrm{Proj}[z;\zeta]=
\mathrm{Proj}[z;\zeta\RR]=\tfrac12(z+\overline z \zeta^2)
\end{equation}
for the projection of $z$ onto (the straight line through 0 and)
$\zeta$. The cases when $\zeta=e^{\pm i\pi/4}$ will be particularly
important in what follows, and
we will write $\ell(\uparrow)=e^{-i\pi/4}\RR$ and
$\ell(\downarrow)=e^{i\pi/4}\RR$. (This choice of notation will be motivated
below, in the context of the {\sc fk}-observable).
We define
\begin{equation}\label{up-dn-notation}
F^\uparrow(z)=\mathrm{Proj}[F(z);\ell(\uparrow)],\quad
F^\downarrow(z)=\mathrm{Proj}[F(z);\ell(\downarrow)].
\end{equation}
Note that $F(z)=F^\uparrow(z)+F^\downarrow(z)$ since
$\ell(\uparrow)\perp\ell(\downarrow)$.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.42}\hspace{2cm}
\includegraphics{fermionic.43}
\caption{
\emph{Left:} The lines $\ell(\uparrow)$ and $\ell(\downarrow)$.
\emph{Right:} Illustration of conditions~\eqref{s-hol-def-w-1}
and~\eqref{s-hol-def-u-1} in Definition~\ref{s-hol-def}. For a pair
of adjacent black and white points,
separated by an arrow in direction $\a\in\{\uparrow,\downarrow\}$,
the projections of $F$ onto $\ell(\a)$ are the same.
}
\label{s-hol-fig}
\end{figure}
\begin{definition}[s-holomorphic]\label{s-hol-def}
A function $F:\Om^\b_\d\cup\Om^\circ_\d\to\mathbb{C}$ is s-holomorphic at a
point $w\in\Om_\d^{\circ,\mathrm{int}}$ if the following hold:
\begin{equation}\label{s-hol-def-w-1}
F^\uparrow(w)=F^\uparrow(w-\sfrac\d2),\quad
F^\downarrow(w)=F^\downarrow(w+\sfrac\d2),\quad\mbox{and}
\end{equation}
\begin{equation}\label{s-hol-def-w-2}
\begin{split}
\dot F^\uparrow(w)&=\tfrac i\d
\big(F^\downarrow(w+\sfrac\d2)-F^\downarrow(w-\sfrac\d2)\big),\\
\dot F^\downarrow(w)&=\tfrac i\d
\big(F^\uparrow(w+\sfrac\d2)-F^\uparrow(w-\sfrac\d2)\big).
\end{split}
\end{equation}
It is s-holomorphic at a point
$u\in\Om_\d^{\b,\mathrm{int}}$ if the following hold:
\begin{equation}\label{s-hol-def-u-1}
F^\uparrow(u)=F^\uparrow(u+\sfrac\d2),\quad
F^\downarrow(u)=F^\downarrow(u-\sfrac\d2),\quad\mbox{and}
\end{equation}
\begin{equation}\label{s-hol-def-u-2}
\begin{split}
\dot F^\uparrow(u)&=\tfrac i\d
\big(F^\downarrow(u+\sfrac\d2)-F^\downarrow(u-\sfrac\d2)\big),\\
\dot F^\downarrow(u)&=\tfrac i\d
\big(F^\uparrow(u+\sfrac\d2)-F^\uparrow(u-\sfrac\d2)\big).
\end{split}
\end{equation}
If $F$ is s-holomorphic at every point
$z\in\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}}$ then we simply say that
$F$ is s-holomorphic in $\Om_\d$.
\end{definition}
The choice of the term s-holomorphic is mainly motivated by
Proposition~\ref{H-prop} below, which is completely analogous to the
classical case (e.g.\ Proposition~3.6 of~\cite{ch-sm}).
It is easy to see that a function $F$ which is s-holomorphic at a point
$z\in\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}}$
satisfies the following natural preholomorphicity condition:
\begin{equation}\label{prehol}
\tfrac1\d\big(F(z+\sfrac\d2)-F(z-\sfrac\d2)\big)
+i\dot F(z)=0.
\end{equation}
However, as for the classical case, the main benefit of s-holomorphic
functions $F$ is that they have well-behaved discrete analogs of
$\Im\big(\int F^2\big)$.
In the next result we write $\D_\d$ for the appropriate Laplacian
operator given by
\begin{equation}\label{lap-def}
[\D_\d f](z)=\ddot f(z)
+\tfrac1{\d^2}\big(f(z+\d)+f(z-\d)-2f(z)\big).
\end{equation}
We say that a function $h$ is $\D_\d$-harmonic (respectively,
$\D_\d$-sub- or $\D_\d$-super-harmonic) at a point
$z\in\mathbb{C}_\d^\b\cup\mathbb{C}_\d^\circ$ if
$[\D_\d h](z)=0$ (respectively, $[\D_\d h](z)\geq0$ or $[\D_\d h](z)\leq0$).
\begin{proposition}\label{H-prop}
Let $F$ be s-holomorphic in $\Om_\d$. Then there
is a function
$H:\Om_\d^\b\cup\Om_\d^\circ\to\RR$, unique
up to an additive constant, satisfying the following.
Firstly, for $z$ s.t.\ $[z,z+\sfrac\d2]\subseteq\Om_\d$,
\begin{equation}\label{discr-diff-eq}
H(z+\sfrac\d2)-H(z)=
\left\{
\begin{array}{ll}
+|F^\downarrow(z)|^2, & \mbox{if } z\in\Om_\d^\circ,\\
-|F^\uparrow(z)|^2, & \mbox{if } z\in\Om_\d^\b,
\end{array}
\right.
\end{equation}
and secondly, for any $z\in\Om_\d^\b\cup\Om_\d^\circ$,
\begin{equation}\label{der-eq}
\dot H(z)=\tfrac2\d F^\uparrow(z) F^\downarrow(z).
\end{equation}
Moreover, we have for all $u\in\Om_\d^{\b,\mathrm{int}}$ and
$w\in\Om_\d^{\circ,\mathrm{int}}$ that
\begin{equation}\label{H-sub-sup}
[\D_\d H](u)=|\dot F(u)|^2
\quad\mbox{and}\quad
[\D_\d H](w)=-|\dot F(w)|^2.
\end{equation}
Hence
$H$ is $\D_\d$-sub-harmonic in $\Om_\d^{\b,\mathrm{int}}$ and
$\D_\d$-super-harmonic in $\Om_\d^{\circ,\mathrm{int}}$.
\end{proposition}
\begin{remark}
The function $\d H(z)$ is a discrete analog
of $\Im\big(\int^z F^2\big)$.
Indeed, since
$2F^\uparrow(z)F^\downarrow(z)=\Re[F(z)^2],$ we see that
if $u,u'\in\Om_\d^\b$ with
$u=x+iy$ and $u'=x+iy'$ for some $x\in\d\mathbb{Z}$ and $y,y'\in\RR$ such
that $[u,u']\subseteq\Om_\d^\b$, then
\begin{equation}
\begin{split}
H(u')-H(u)&=\int_{y}^{y'} \dot H(x+it) \,dt
=\frac1\d\int_y^{y'} \Re[F(x+it)^2] \,dt\\
&=\frac1\d\int_y^{y'} \Im[iF(x+it)^2] \,dt
=\tfrac1\d\Im\Big[\int_u^{u'} F(z)^2 \,dz\Big].
\end{split}
\end{equation}
Similarly, if $v=u+\d\in\Om_\d^\b$
and $w=u+\sfrac\d2=v-\sfrac\d2$ is midway between $u$ and $v$
then, using
$F^\uparrow(z)^2+F^\downarrow(z)^2=i\,\Im[F(z)^2]$, we have
\begin{equation}
\begin{split}
H(v)-H(u)&=H(w+\sfrac\d2)-H(w)+H(w)-H(w-\sfrac\d2)\\
&=|F^\downarrow(w)|^2-|F^\uparrow(w)|^2
=\tfrac1i(F^\uparrow(w)^2+F^\downarrow(w)^2)\\
&=\Im[F(w)^2].
\end{split}
\end{equation}
\end{remark}
\begin{proof}[Proof of Proposition~\ref{H-prop}]
Uniqueness up to an additive constant follows since if
we fix $H(u)$ for some point $u$, then for $v\neq u$
we may obtain the value $H(v)$ by integrating
using~\eqref{discr-diff-eq} and~\eqref{der-eq}.
To see that $H$ is well-defined,
consider a situation such as in Figure~\ref{H-def-fig}.
It suffices to show that the total increment of
$H$ around the blue (left) contour
and around the green (right) contour are both equal to 0.
We prove this for the green (right) contour, the other one being
similar.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.3}
\caption{Contours in the proof of Proposition~\ref{H-prop}.}
\label{H-def-fig}
\end{figure}
Let us write, for $a,b,c,t_1,t_2\in\RR$ and $j=1,2$,
$u_j=a+i t_j$, $w_j=b+i t_j$ and $v_j=c+i t_j$.
We have
\begin{equation}\begin{split}
[H(v_2)-&H(v_1)]+[H(w_1)-H(w_2)]=
\int_{t_1}^{t_2}\big(\dot H(c+it)-\dot H(b+it)\big)dt\\
&=\tfrac2\d\int_{t_1}^{t_2}F^\downarrow(b+it)\big(
F^\uparrow(c+it)-F^\uparrow(b+it)\big)dt\\
&=\tfrac1i\int_{t_1}^{t_2} 2F^\downarrow(b+it)\dot F^\downarrow(b+it) dt\\
&=\tfrac1i\big(F^\downarrow(w_2)^2-F^\downarrow(w_1)^2\big)
=|F^\downarrow(w_2)|^2-|F^\downarrow(w_1)|^2\\
&=[H(v_2)-H(w_2)]-[H(v_1)-H(w_1)].
\end{split}\end{equation}
That is, the increments around the green contour satisfy
\[
[H(v_2)-H(v_1)]+[H(w_2)-H(v_2)]+
[H(w_1)-H(w_2)]+[H(v_1)-H(w_1)]=0,
\]
as required.
We turn now to the statement~\eqref{H-sub-sup}.
We give the details for $u\in\Om_\d^\b$, the case $w\in\Om_\d^\circ$
being similar.
Since $\dot H(u)=\tfrac2\d F^\uparrow(u)F^\downarrow(u)$
we have
\begin{equation}
\ddot H(u)=\tfrac2\d\big(
\dot F^\uparrow(u)F^\downarrow(u)+F^\uparrow(u)\dot F^\downarrow(u)
\big).
\end{equation}
Using s-holomorphicity we deduce that
\begin{equation}
\begin{split}
\ddot H(u)&=2\tfrac{i}{\d^2}\big(
[F^\downarrow(u+\sfrac\d2)-F^\downarrow(u-\sfrac\d2)]F^\downarrow(u)\\
&\qquad\qquad
+[F^\uparrow(u+\sfrac\d2)-F^\uparrow(u-\sfrac\d2)]F^\uparrow(u)
\big)\\
&=\tfrac{i}{\d^2} \big(
2F^\downarrow(u-\sfrac\d2)F^\downarrow(u+\sfrac\d2)-2F^\downarrow(u-\sfrac\d2)^2\\
&\qquad\qquad+2F^\uparrow(u+\sfrac\d2)^2-
2F^\uparrow(u-\sfrac\d2)F^\uparrow(u+\sfrac\d2)
\big).
\end{split}
\end{equation}
Next,
\begin{equation}
\begin{split}
H(u-\d)-H(u)&=
|F^\uparrow(u-\sfrac\d2)|^2-|F^\downarrow(u-\sfrac\d2)|^2\\
&=i(F^\uparrow(u-\sfrac\d2)^2+F^\downarrow(u-\sfrac\d2)^2),\mbox{ and }\\
H(u+\d)-H(u)&=
|F^\downarrow(u+\sfrac\d2)|^2-|F^\uparrow(u+\sfrac\d2)|^2\\
&=-i(F^\downarrow(u+\sfrac\d2)^2+F^\uparrow(u+\sfrac\d2)^2).
\end{split}
\end{equation}
It follows that
\begin{equation}
\begin{split}
\d^2&\ddot H(u)+
[H(u-\d)-H(u)]+[H(u+\d)-H(u)]\\
&=i\big(
[F^\uparrow(u-\sfrac\d2)-F^\uparrow(u+\sfrac\d2)]^2-
[F^\downarrow(u-\sfrac\d2)-F^\downarrow(u+\sfrac\d2)]^2
\big).
\end{split}
\end{equation}
Writing $F^\uparrow(u-\sfrac\d2)=ae^{-i\pi/4}$,
$F^\uparrow(u+\sfrac\d2)=be^{-i\pi/4}$,
$F^\downarrow(u-\sfrac\d2)=ce^{i\pi/4}$
and $F^\downarrow(u+\sfrac\d2)=de^{i\pi/4}$,
for $a,b,c,d\in\RR$, the right-hand-side
equals
\begin{equation}
i\big(\tfrac1i[a-b]^2-i[c-d]^2
\big)=(a-b)^2+(c-d)^2.
\end{equation}
But we also have that
\begin{equation}\begin{split}
&|F(u-\sfrac\d2)-F(u+\sfrac\d2)|^2\\
&\quad=
|(F^\uparrow(u-\sfrac\d2)-F^\uparrow(u+\sfrac\d2))+
(F^\uparrow(u-\sfrac\d2)-F^\uparrow(u+\sfrac\d2))|^2\\
&\quad=(a-b)^2+(c-d)^2, \mbox{ since } \ell(\uparrow)\perp\ell(\downarrow).
\end{split}\end{equation}
Thus, using also~\eqref{prehol},
\begin{equation}
\d^2 [\D_\d H](u)=
|F(u-\sfrac\d2)-F(u+\sfrac\d2)|^2=\d^2|\dot F(u)|^2,
\end{equation}
as claimed.
\end{proof}
\section{The FK-observable}\label{fk-obs-sec}
\subsection{Definition}
Let $(\Om_\d,a_\d,b_\d)$ be a Dobrushin domain
(see Section~\ref{dom-sec} for notation).
We will consider {\sc fk}-configurations $\xi$ in $\Om_\d$ and
their duals $\xi'$. These are defined as in
Section~\ref{fk-sec}
with some adaptations of the boundary condition.
We take $\xi=\xi^\b\cup\xi^\circ$ with
$\xi^\b\subseteq\Om^{\b,\mathrm{int}}_\d$ and $\xi^\circ\subseteq\Om^{\circ,\mathrm{int}}_\d$ finite
subsets. Note that we do not allow $\xi^\b$ to have any points on
the black part $\partial_\d^\b$ of the boundary, nor do we allow
$\xi^\circ$ to have any points on the white part $\partial^\circ_\d$.
Instead of applying periodic boundary conditions, we let
horizontal
segments in $\partial^\b_\d$ and $\partial^\circ_\d$ count as primal and
dual bridges, respectively. Thus, in essence, we have separately
wired together the black and white parts $\partial^\b_\d$ and
$\partial^\circ_\d$ of the boundary.
See Figure~\ref{domain-interface-fig}.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.5}\hspace{1cm}
\includegraphics{fermionic.6}
\caption{Dobrushin domain $(\Om_\d,a_\d,b_\d)$
with an \textsc{fk} configuration $\xi$ and its dual $\xi'$,
as well as the interface $\g$
(left) and the $L(\xi)=5$ loops (right).
We have omitted the $\times$-marks for
cuts.}
\label{domain-interface-fig}
\end{figure}
We adjust the
locations of the points $a_\d$ and $b_\d$ slightly compared to
Section~\ref{dom-sec}, as follows.
Firstly, we assume that $a_\d$ is placed
so that the first point of $\mathbb{C}_\d^\b\cup\mathbb{C}_\d^\circ$ visited by
$\partial_\d^\b$ (as it travels clockwise from $a_\d$ to $b_\d$)
belongs to $\mathbb{C}_\d^\b$. Thus $a_\d$ is of the form $u+\sfrac\d4$
for some $u\in\Om_\d^\b$ if $\Om_\d$ is `above' $a_\d$, or of the
form $u-\sfrac\d4$ if $\Om_\d$ is `below' $a_\d$.
With this assumption, an {\sc fk}-configuration $\xi$ together
with its dual $\xi'$ define an interface $\g$ from $a_\d$ to $b_\d$,
separating the (primal) component of $\partial_\d^\b$ from the (dual)
component of $\partial_\d^\circ$, and $\g$ always has black on the left and
white on the right as it travels from $a_\d$ to $b_\d$.
We take $\g$ to travel in the directions $\uparrow,\downarrow$
on the medial lattice $\mathbb{C}_\d^\diamondsuit$ between bridges, and
in the directions $\leftarrow,\rightarrow$ at bridges (if $\g$ passes the same bridge
twice we slightly separate the points where it passes).
We also shift $b_\d$ left or right by $\sfrac\d4$ so
that the interface $\g$ ends pointing in the direction $\rightarrow$ into
$b_\d$. See Figure~\ref{domain-interface-fig} again.
Apart from the interface $\g$, we also draw a loop around each
(primal and dual) component which is disjoint from the boundary.
We let $L(\xi)$ denote the number of such loops.
Let $\EE_\d(\cdot)$ denote the probability measure under which
$\xi^\b$ and $\xi^\circ$ are independent Poisson processes on
$\Om_\d^{\b,\mathrm{int}}$ and $\Om_\d^{\circ,\mathrm{int}}$, respectively,
\emph{both with the same rate} $\tfrac1{\d\sqrt2}$. By~\eqref{fk}, the
appropriate density of a random {\sc fk}-configuration $\xi$ with respect
to $\EE_\d(\cdot)$ is proportional to
\begin{equation}\label{fk-dens-1}
2^{k^\b(\xi)} h^{|\xi^\b|} (2J)^{|\xi^\circ|} (\d\sqrt2)^{|\xi^\b|+|\xi^\circ|}.
\end{equation}
Using the Euler-relation one may see that
$k^\b(\xi)-|\xi^\b|=k^\circ(\xi)-|\xi^\circ|+\mathrm{cst}$ for some constant
not depending on $\xi$. Also, $L(\xi)=k^\b(\xi)+k^\circ(\xi)-2$.
We choose the parameters
\begin{equation}
h=J=\tfrac1{2\d}.
\end{equation}
It then follows that the density~\eqref{fk-dens-1} is
proportional to simply $(\sqrt2)^{L(\xi)}$.
We write $\hat\EE_\d=\hat\EE_{(\Om_\d,a_\d,b_\d)}$
for the critical {\sc fk}-law in $(\Om_\d,a_\d,b_\d)$ given by
\begin{equation}\label{fk-dens}
\frac{d\hat\EE_\d}{d\EE_\d}(\xi)=
\frac{(\sqrt{2})^{L(\xi)}}{Z_\d},
\mbox{ where } Z_\d=\EE_\d[(\sqrt2)^{L(\xi)}].
\end{equation}
Now let $z\in\Om_\d^\b\cup \Om_\d^\circ$ be arbitrary.
For $\a\in\{\uparrow,\downarrow,\leftarrow,\rightarrow\}$, define the event
\[
\Gamma^\a_z=\{\xi: \g(\xi)\mbox{ passes by $z$ in direction }\a\}.
\]
For $\a\in\{\uparrow,\downarrow\}$ we count both the case when $\g$ passes
on the left side of $z$ (i.e.\ goes through $z-\sfrac\d4$) and
when it passes on the right side (i.e.\ goes through $z+\sfrac\d4$).
Similarly, for $\a\in\{\leftarrow,\rightarrow\}$ we count both the cases when
$\g$ passes `just below' $z$ and `just above' $z$.
Assuming that $\Gamma^\a_z$ happens, let
$W^\a_\g(z)$ denote the winding-angle (in radians) of $\g$
from $z$ to the exit $b_\d$;
if $\g$ passes $z$ twice, in opposite directions, we count here the
winding angle from when it passes in direction $\a$.
Note that $W^\a_\g(z)$ is deterministic up to a multiple
of $2\pi$.
We define the four
(random) functions $\varphi^\uparrow(\xi; z)$,
$\varphi^\downarrow(\xi; z)$,
$\varphi^\leftarrow(\xi; z)$ and
$\varphi^\rightarrow(\xi; z)$ by
\begin{equation}\label{phi-om-def}
\varphi^\a(\xi; z)=\hbox{\rm 1\kern-.27em I}_{\Gamma^\a_z}(\xi)
\exp(\tfrac{i}{2}W^\a_{\g(\xi)}(z)).
\end{equation}
Note that the supports of $\varphi^\leftarrow(\xi; z)$ and of
$\varphi^\rightarrow(\xi; z)$ are discrete sets contained in
$\xi\cup\partial^\mathrm{h}\Om_\d$,
whereas the supports of $\varphi^\uparrow(\xi; z)$ and
$\varphi^\downarrow(\xi; z)$ are disjoint from $\xi$.
Also note
that if $u\in\Om_\d^\b$ is black
and $w=u+\sfrac\d2$ is the white neighbour
of $u$ on the right, then
$\varphi^\uparrow(\xi;u)=\varphi^\uparrow(\xi;w)$,
whereas if $w'=u-\sfrac\d2$ is the white neighbour of $u$
on the left then
$\varphi^\downarrow(\xi;u)=\varphi^\downarrow(\xi;w')$.
(Here we assume that $u\pm\sfrac\d2\in\Om_\d^\circ$ in the appropriate
cases.)
\begin{definition}\label{fk-obs-def}
Write
\begin{equation}\label{phi-updn}
\Phi_\d^\uparrow(z)=\hat\EE_\d[\varphi^\uparrow(\xi;z)],\quad
\Phi_\d^\downarrow(z)=\hat\EE_\d[\varphi^\downarrow(\xi;z)].
\end{equation}
We define the \textsc{fk}--Ising observable
$F_\d(z)=F^{\mathrm{FK}}_\d(z)$ by
\begin{equation}\label{fk-obs-eq}
F_\d(z)=\Phi_\d^\uparrow(z)+\Phi_\d^\downarrow(z),\quad
z\in\Om_\d^\b\cup\Om_\d^\circ.
\end{equation}
\end{definition}
We remark that the notation used here is consistent with our previous
notation~\eqref{up-dn-notation} for the projections
$F^\uparrow$, $F^\downarrow$ of a function
$F$ onto $\ell(\uparrow)=e^{-i\pi/4}\RR$ and
$\ell(\downarrow)=e^{i\pi/4}\RR$, in the sense that
\[
F_\d^\uparrow(z)=\Phi_\d^\uparrow(z)\mbox{ and } F_\d^\downarrow(z)=\Phi_\d^\downarrow(z).
\]
Indeed, if we identify arrows
$\a\in\{\uparrow,\downarrow,\leftarrow,\rightarrow\}$ with complex numbers by
the rules
\begin{equation}
\rightarrow\,=1=i^0,\quad
\uparrow\,=i=i^1,\quad
\leftarrow\,=-1=i^2,\quad
\downarrow\,=-i=i^3,
\end{equation}
then we have that
\begin{equation}
W_{\g(\xi)}^\a(z)=-\arg(\a)+2\pi n(\xi)
\end{equation}
for some random $n(\xi)\in\mathbb{Z}$. Thus
$\varphi^\a(\xi;z)$ is a real multiple of $\sqrt{\overline{\a}}$
i.e.\ belongs to $\ell(\a)$.
Note that the line $\sqrt{\overline{\a}}\RR$ does not
depend on the choice of square-root.
\subsection{Comparison with isoradial graphs}
\label{isorad-sec}
For readers familiar with the work of Chelkak and
Smirnov~\cite{ch-sm} on the
classical Ising model on isoradial graphs, the following brief
discussion may be useful. Let $0<\varepsilon\ll\d$ and consider a rhombic
tiling of $\mathbb{C}$ where all the rhombi have two vertices in each of
$\mathbb{C}^\b_\d$ and $\mathbb{C}^\circ_\d$, and acute angle $2\varepsilon$,
as in Figure~\ref{squeeze-2_fig}.
This corresponds to an isoradial embedding of $\mathbb{Z}^2$ with common
radius $\tfrac\d{2\cos(\varepsilon)}$ and vertices restricted to
$\mathbb{C}^\b_\d$.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.7}
\caption{Isoradial approximation of the {\sc fk}-representation of the
{\sc tfim}.}
\label{squeeze-2_fig}
\end{figure}
Let $\hat\EE_{\d,\varepsilon}(\cdot)$ denote the law of the critical
(classical) {\sc fk}--Ising model in some Dobrushin-domain in this graph,
as given in eq.~(2.1)
of~\cite{ch-sm}, and let $\g_\varepsilon$ denote the interface.
It is well-known that the laws $\hat\EE_{\d,\varepsilon}$
converge weakly to $\hat\EE_\d$ as $\varepsilon\to0$.
In this setting, the interface $\g_\varepsilon$ is taken
to cross the rhombus-sides perpendicularly,
i.e.\ roughly speaking in the directions
$\nearrow$, $\searrow$, $\nwarrow$
and $\swarrow$. If we specify a rhombus as well as one of these four
directions of travel, this corresponds to a unique edge of the
rhombus, hence
the edge-observables~\cite[eq.~(2.2)]{ch-sm}
of Chelkak and Smirnov can be indexed
as $F_{\d,\varepsilon}^\nearrow(z), F_{\d,\varepsilon}^\searrow(z), \dotsc$
for rhombus centres $z$. Using notation similar
to~\eqref{phi-om-def}, we have (up to a real factor)
\begin{equation}
F_{\d,\varepsilon}^\a(z)=\hat\EE_{\d,\varepsilon}[\hbox{\rm 1\kern-.27em I}_{\Gamma_z^\a}
\exp(\tfrac{i}{2}W^\a_{\g_\varepsilon}(z))],\quad
\a\in\{\nearrow,\searrow,\nwarrow,\swarrow\}.
\end{equation}
We may further take $\g_\varepsilon$
to pass `closest' to rhombus centres $z$
in the directions $\uparrow$, $\downarrow$, $\leftarrow$ or $\rightarrow$.
This allows us to define more observables:
\begin{equation}
\Phi_{\d,\varepsilon}^\a(z)=\hat\EE_{\d,\varepsilon}[\hbox{\rm 1\kern-.27em I}_{\Gamma_z^\a}
\exp(\tfrac{i}{2}W^\a_{\g_\varepsilon}(z))],\quad
\a\in\{\uparrow,\downarrow,\leftarrow,\rightarrow\}.
\end{equation}
Clearly each $\Phi_{\d,\varepsilon}^\a(z)\in\ell(\a)$
as is the case for the $\Phi_\d^\a(z)$ (provided we assume that
$\g_\varepsilon$ exits the domain in the direction $\rightarrow$).
Referring to Figure~\ref{squeeze-2_fig}, we see for example that
if $\g_\varepsilon$ enters the
rhombus of $z$ in direction $\nwarrow$
(edge on the lower right of $z$), then
it passes closest to $z$ in \emph{either}
direction $\uparrow$ as depicted, or directon $\leftarrow$,
\emph{but not both} ($\uparrow$ if there is a black vertical
edge at $z$, and $\leftarrow$ if there is a white horizontal
edge).
Similar considerations apply at all rhombus centres, and this allows
us to derive linear relations for the
$F_{\d,\varepsilon}^\a(z)$ in terms of the $\Phi_{\d,\varepsilon}^\a(z)$.
Writing $\varepsilon^\star=\tfrac\pi2-\varepsilon$ we have:
\begin{equation}\label{mtrx}
\begin{pmatrix}
F_{\d,\varepsilon}^\nwarrow(z) \\
F_{\d,\varepsilon}^\nearrow(z) \\
F_{\d,\varepsilon}^\searrow(z) \\
F_{\d,\varepsilon}^\swarrow(z)
\end{pmatrix}=
\begin{pmatrix}
e^{-\tfrac i2 \varepsilon} & 0 & 0 & e^{\tfrac i2 \varepsilon^\star} \\
e^{\tfrac i2 \varepsilon} & e^{-\tfrac i2 \varepsilon^\star} & 0 & 0 \\
0 & e^{\tfrac i2 \varepsilon^\star} & e^{-\tfrac i2 \varepsilon} & 0 \\
0 & 0 & e^{\tfrac i2 \varepsilon} & e^{-\tfrac i2 \varepsilon^\star}
\end{pmatrix}
\begin{pmatrix}
\Phi_{\d,\varepsilon}^\uparrow(z) \\
\Phi_{\d,\varepsilon}^\rightarrow(z) \\
\Phi_{\d,\varepsilon}^\downarrow(z) \\
\Phi_{\d,\varepsilon}^\leftarrow(z)
\end{pmatrix}.
\end{equation}
The {\sc fk}--Ising observable~\cite[eq.~(2.4)]{ch-sm} of Chelkak and
Smirnov is given by
\begin{equation}
\begin{split}
F_{\d,\varepsilon}(z)&=\tfrac12\textstyle{\sum}_\a F_{\d,\varepsilon}^\a(z)\\
&=\cos(\varepsilon/2)[\Phi_\varepsilon^\uparrow(z)+ \Phi_\varepsilon^\downarrow(z)]
+\cos(\varepsilon^\star/2)[\Phi_\varepsilon^\leftarrow(z)+ \Phi_\varepsilon^\rightarrow(z)],
\end{split}
\end{equation}
where the second line uses~\eqref{mtrx}.
Assuming the limits
\begin{equation}
\Phi_\d^\uparrow(z)=\lim_{\varepsilon\to 0} \Phi_{\d,\varepsilon}^\uparrow(z),\quad
\Phi_\d^\downarrow(z)=\lim_{\varepsilon\to 0} \Phi_{\d,\varepsilon}^\downarrow(z),
\end{equation}
as well as $\Phi_{\d,\varepsilon}^\a(z)=O(\varepsilon)$ for
$\a\in\{\leftarrow,\rightarrow\}$, we get
\begin{equation}
\lim_{\varepsilon\to0} F_{\d,\varepsilon}(z)=\Phi_\d^\uparrow(z)+ \Phi_\d^\downarrow(z),
\end{equation}
which is how we defined our observable $F_\d(z)$.
\subsection{S-holomorphicity}
In this section we show the following result:
\begin{theorem}\label{s-hol-thm-fk}
Let $F_\d=F^{\mathrm{FK}}_\d$ be the {\sc fk}-observable in a Dobrushin
domain $(\Om_\d,a_\d,b_\d)$. Then $F_\d$ is s-holomorphic in
$\Om_\d$.
\end{theorem}
It is immediate that $F_\d$ satisfies the
conditions~\eqref{s-hol-def-w-1} and~\eqref{s-hol-def-u-1}
in the definition of s-holomorphicity, see
the discussion just above Definition~\ref{fk-obs-def}.
We thus need to show that also~\eqref{s-hol-def-w-2}
and~\eqref{s-hol-def-u-2} are satisfied.
In the proof we drop the subscript $\d$ from $\EE$ and $\hat\EE$.
For $z\in\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}}$ we
let $\xi_z=\xi\triangle\{z\}$ and
we define the auxiliary observables
\begin{equation}\label{phi-ltrt}
\begin{split}
\Phi_\d^\leftarrow(z)&=\hat\EE[(\sqrt{2})^{L(\xi_z)-L(\xi)}\varphi^\leftarrow(\xi_z;z)],\\
\Phi_\d^\rightarrow(z)&=\hat\EE[(\sqrt{2})^{L(\xi_z)-L(\xi)}\varphi^\rightarrow(\xi_z;z)].
\end{split}
\end{equation}
If $z\in\partial^\mathrm{v}\Om_\d$ is in the vertical part of the boundary
then we set $\Phi_\d^\leftarrow(z)=\Phi_\d^\rightarrow(z)=0$, whereas if
$z\in\partial^\mathrm{h}\Om_\d$ is in the horizontal part we define them as
in~\eqref{phi-ltrt} but with $\xi_z$ replaced by $\xi$.
As we remarked above we have that $\Phi_\d^\rightarrow(z)\in\ell(\rightarrow)=\RR$ and
$\Phi_\d^\leftarrow(z)\in\ell(\leftarrow)=i\RR$.
We now claim the following:
\begin{lemma}\label{turn-lem}
For all $z\in\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}}$ we have that
\begin{equation}\label{turn-eq}
\begin{split}
\Phi_\d^\uparrow(z)&=\tfrac{1}{\sqrt2}\big(
e^{i\pi/4}\Phi_\d^\leftarrow(z)+e^{-i\pi/4}\Phi_\d^\rightarrow(z)\big),\\
\Phi_\d^\downarrow(z)&=\tfrac{1}{\sqrt2}\big(
e^{i\pi/4}\Phi_\d^\rightarrow(z)+e^{-i\pi/4}\Phi_\d^\leftarrow(z)\big).
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
We prove the statement for $\Phi_\d^\uparrow(z)$ in the case when
$z\in\Om_\d^{\circ,\mathrm{int}}$ is white, the other cases are similar.
We refer to Figures~\ref{prehol_fig-1},
\ref{prehol_fig-2} and \ref{prehol_fig-3}.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.8}\hspace{1cm}
\includegraphics{fermionic.9}
\caption{In $\xi$ the interface $\g$ passes $z$ in direction $\uparrow$ only,
in $\xi_z$ it passes in directions $\rightarrow$ and $\leftarrow$.}
\label{prehol_fig-1}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.10}\hspace{1cm}
\includegraphics{fermionic.11}
\caption{In $\xi$ the interface $\g$ passes $z$ in directions $\uparrow$
and $\downarrow$, in $\xi_z$ it passes in direction $\leftarrow$.}
\label{prehol_fig-2}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.12}\hspace{1cm}
\includegraphics{fermionic.13}
\caption{In $\xi$ the interface $\g$ passes $z$ in directions $\uparrow$
and $\downarrow$, in $\xi_z$ it passes in direction $\rightarrow$.}
\label{prehol_fig-3}
\end{figure}
Let $A$ denote the event
that $\g$ passes $z$ only once, in the direction $\uparrow$, as depicted on
the left in Figure~\ref{prehol_fig-1}. Let $A'$ denote the event that
$\g$ passes $z$ once in the direction $\rightarrow$ and once in the direction
$\leftarrow$, with $\rightarrow$ coming first,
as depicted on the right in Figure~\ref{prehol_fig-1}.
Similarly, let $B$ and $B'$ denote the events depicted in
Figure~\ref{prehol_fig-2}. Explicitly, $B$ is the event that $\g$
passes $z$ both going $\uparrow$ and $\downarrow$, with $\downarrow$ coming first, and
$B'$ is the event that $\g$ passes $z$ in direction $\leftarrow$ only.
Finally, let $C$ and $C'$ be as in Figure~\ref{prehol_fig-3}:
$C$ is the event that $\g$ passes $z$ in direction $\uparrow$ and later in
direction $\downarrow$, and $C'$ is the event that it passes in direction
$\rightarrow$ only.
We note the following facts. Firstly,
\begin{equation}
\begin{split}
&\xi\in A\Leftrightarrow \xi_z\in A',\mbox{ and then }
L(\xi)=L(\xi_z)+1,\\
&\xi\in B\Leftrightarrow \xi_z\in B',\mbox{ and then }
L(\xi)=L(\xi_z)-1,\\
&\xi\in C\Leftrightarrow \xi_z\in C', \mbox{ and then }
L(\xi)=L(\xi_z)-1.
\end{split}
\end{equation}
Secondly, the event $\Gamma^\uparrow_z=\{\g\mbox{ passes $z$ going }\uparrow\}$
satisfies
\begin{equation}
\hbox{\rm 1\kern-.27em I}_{\Gamma^\uparrow_z}(\xi)=\hbox{\rm 1\kern-.27em I}_A(\xi)+\hbox{\rm 1\kern-.27em I}_B(\xi)+\hbox{\rm 1\kern-.27em I}_C(\xi)
\end{equation}
and the events $\Gamma^\leftarrow_z$ and $\Gamma^\rightarrow_z$ satisfy
\begin{equation}
\begin{split}
\hbox{\rm 1\kern-.27em I}_{\Gamma^\leftarrow_z}(\xi_z)&=\hbox{\rm 1\kern-.27em I}_{A'}(\xi_z)+\hbox{\rm 1\kern-.27em I}_{B'}(\xi_z),
\mbox{ and }\\
\hbox{\rm 1\kern-.27em I}_{\Gamma^\rightarrow_z}(\xi_z)&=\hbox{\rm 1\kern-.27em I}_{A'}(\xi_z)+\hbox{\rm 1\kern-.27em I}_{C'}(\xi_z).
\end{split}
\end{equation}
Thirdly, the winding angles are related by
\begin{equation}\label{angles-1}
\begin{split}
W^\uparrow_{\g(\xi)}(z)&=W^\leftarrow_{\g(\xi_z)}(z)+\pi/2,
\mbox{ for } \xi\in A\cup B;\\
W^\uparrow_{\g(\xi)}(z)&=W^\rightarrow_{\g(\xi_z)}(z)-\pi/2,
\mbox{ for } \xi\in A\cup C.
\end{split}
\end{equation}
Using these facts, we obtain:
\begin{equation}
\begin{split}
\varphi^\uparrow(\xi;z)&=(\hbox{\rm 1\kern-.27em I}_A(\xi)+\hbox{\rm 1\kern-.27em I}_B(\xi)+\hbox{\rm 1\kern-.27em I}_C(\xi))
\exp\big(\tfrac i2 W^\uparrow_{\g(\xi)}(z)\big)\\
&=\tfrac12\hbox{\rm 1\kern-.27em I}_{A'}(\xi_z)
\exp\big(\tfrac i2 W^\leftarrow_{\g(\xi_z)}(z)\big) e^{i\pi/4}\\
&\quad+\tfrac12\hbox{\rm 1\kern-.27em I}_{A'}(\xi_z)
\exp\big(\tfrac i2 W^\rightarrow_{\g(\xi_z)}(z)\big) e^{-i\pi/4}\\
&\quad+\hbox{\rm 1\kern-.27em I}_{B'}(\xi_z)
\exp\big(\tfrac i2 W^\leftarrow_{\g(\xi_z)}(z)\big) e^{i\pi/4}\\
&\quad+\hbox{\rm 1\kern-.27em I}_{C'}(\xi_z)
\exp\big(\tfrac i2 W^\rightarrow_{\g(\xi_z)}(z)\big) e^{-i\pi/4}.
\end{split}
\end{equation}
Thus
\begin{equation}
\begin{split}
&(\sqrt2)^{L(\xi)}\varphi^\uparrow(\xi;z)\\
&\quad=(\sqrt2)^{L(\xi)-2}\hbox{\rm 1\kern-.27em I}_{A'}(\xi_z)
\big\{\exp\big(\tfrac i2 W^\leftarrow_{\g(\xi_z)}(z)\big) e^{i\pi/4}+
\exp\big(\tfrac i2 W^\rightarrow_{\g(\xi_z)}(z)\big) e^{-i\pi/4}\big\}\\
&\quad\quad + (\sqrt2)^{L(\xi)}\hbox{\rm 1\kern-.27em I}_{B'}(\xi_z)
\exp\big(\tfrac i2 W^\leftarrow_{\g(\xi_z)}(z)\big) e^{i\pi/4}\\
&\quad\quad +
(\sqrt2)^{L(\xi)}\hbox{\rm 1\kern-.27em I}_{C'}(\xi_z)
\exp\big(\tfrac i2 W^\rightarrow_{\g(\xi_z)}(z)\big) e^{-i\pi/4}\\
&\quad=(\sqrt2)^{L(\xi_z)-1}
\big[\varphi^\leftarrow(\xi_z;z)e^{i\pi/4}+
\varphi^\rightarrow(\xi_z;z)e^{-i\pi/4}\big].
\end{split}
\end{equation}
Taking the $\EE$-expectation,
\begin{equation}\begin{split}
&\EE[(\sqrt2)^{L(\xi)}\varphi^\uparrow(\xi;z)]\\
&\quad=\tfrac{1}{\sqrt2}\big(
\EE[(\sqrt2)^{L(\xi_z)}\varphi^\leftarrow(\xi_z;z)]e^{i\pi/4}+
\EE[(\sqrt2)^{L(\xi_z)}\varphi^\rightarrow(\xi_z;z)]e^{-i\pi/4}
\big).
\end{split}\end{equation}
This readily gives the claim~\eqref{turn-eq}
for $\Phi^\uparrow(z)$.
\end{proof}
We now calculate $\dot\Phi_\d^\uparrow$ and $\dot\Phi_\d^\downarrow$.
We will use
the notation $\xi(z,z+i\varepsilon)$ for the number of elements of
$\xi$ in the interval $(z,z+i\varepsilon)$.
For a function $f(\xi,z)$ we write
$f(\xi,t\pm)=\lim_{\varepsilon\downarrow 0}f(\xi,t\pm i\varepsilon)$.
Recall that $\xi_z=\xi\triangle\{z\}$.
\begin{lemma}\label{phi-dot-lem}
Let $w\in\Om_\d^{\circ,\mathrm{int}}$ and write $u=w-\sfrac\d2$
and $v=w+\sfrac\d2$. Then
\begin{equation}\label{phi-up-dot}
\begin{split}
\dot\Phi_\d^\uparrow(w)=\dot\Phi_\d^\uparrow(u)
&=\tfrac1{\d\sqrt2}\big(
e^{i\pi/4}\Phi_\d^\leftarrow(w)-e^{-i\pi/4}\Phi_\d^\rightarrow(w)\big)\\
&\quad +
\tfrac1{\d\sqrt2}\big(
e^{-i\pi/4}\Phi_\d^\rightarrow(u)-e^{i\pi/4}\Phi_\d^\leftarrow(u)\big).
\end{split}
\end{equation}
and
\begin{equation}\label{phi-dn-dot}
\begin{split}
\dot\Phi_\d^\downarrow(w)=\dot\Phi_\d^\downarrow(v)
&=\tfrac1{\d\sqrt2}\big(
e^{-i\pi/4}\Phi_\d^\leftarrow(w)-e^{i\pi/4}\Phi_\d^\rightarrow(w)\big)\\
&\quad +
\tfrac1{\d\sqrt2}\big(
e^{i\pi/4}\Phi_\d^\rightarrow(v)-e^{-i\pi/4}\Phi_\d^\leftarrow(v)\big).
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
The first equalities in~\eqref{phi-up-dot} and~\eqref{phi-dn-dot}
hold since $\Phi_\d^\uparrow(w)=\Phi_\d^\uparrow(u)$ and
$\Phi_\d^\downarrow(w)=\Phi_\d^\downarrow(v)$. We prove~\eqref{phi-up-dot}, the
other claim~\eqref{phi-dn-dot} is similar.
We have that
\begin{equation}\label{der-1}
Z_\d\frac{\Phi_\d^\uparrow(w+i\varepsilon)-\Phi_\d^\uparrow(w)}{\varepsilon}
=\tfrac1\varepsilon\EE[(\sqrt2)^{L(\xi)}
(\varphi^\uparrow(\xi;w+i\varepsilon)-\varphi^\uparrow(\xi;w))].
\end{equation}
Note that $\varphi^\uparrow(\xi;w+i\varepsilon)-\varphi^\uparrow(\xi;w)=0$
unless either $\xi(w,w+i\varepsilon)>0$ or $\xi(u,u+i\varepsilon)>0$.
The probability that both these happen is $O(\varepsilon^2)$ and may therefore
be ignored. Also recall that $\varphi^\uparrow(\xi;w)=\varphi^\uparrow(\xi;u)$
for $w$ and $u$ as specified. Thus the right-hand-side
of~\eqref{der-1} equals
\begin{equation}
\begin{split}
&\tfrac1\varepsilon\EE[(\sqrt2)^{L(\xi)}
(\varphi^\uparrow(\xi;w+i\varepsilon)-\varphi^\uparrow(\xi;w))
\hbox{\rm 1\kern-.27em I}\{\xi(w,w+i\varepsilon)>0\}]\\
&\quad+
\tfrac1\varepsilon\EE[(\sqrt2)^{L(\xi)}
(\varphi^\uparrow(\xi;u+i\varepsilon)-\varphi^\uparrow(\xi;u))
\hbox{\rm 1\kern-.27em I}\{\xi(u,u+i\varepsilon)>0\}]+o(1).
\end{split}
\end{equation}
This converges to
\begin{equation}
\begin{split}
&\tfrac1{\d\sqrt2}\EE[(\sqrt2)^{L(\xi_w)}
(\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-))]\\
&\quad+
\tfrac1{\d\sqrt2}\EE[(\sqrt2)^{L(\xi_u)}
(\varphi^\uparrow(\xi_u;u+)-\varphi^\uparrow(\xi_u;u-))].
\end{split}
\end{equation}
Consider $\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-)$.
We refer again to Figures~\ref{prehol_fig-1}, \ref{prehol_fig-2} and
\ref{prehol_fig-3} and the events $A,B,C,A',B',C'$ depicted there, as
well as the relation~\eqref{angles-1} between winding angles.
We have that
\begin{equation}
\begin{split}
\mbox{for }\xi\in A,\quad
&\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-)=0\\
&\quad=\varphi^\leftarrow(\xi_w;w)e^{i\pi/4}-\varphi^\rightarrow(\xi_w;w)e^{-i\pi/4},
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\mbox{for }\xi\in B,\quad
&\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-)=
\varphi^\uparrow(\xi_w;w+)\\
&\quad=\varphi^\leftarrow(\xi_w;w)e^{i\pi/4}\\
&\quad=\varphi^\leftarrow(\xi_w;w)e^{i\pi/4}-\varphi^\rightarrow(\xi_w;w)e^{-i\pi/4},
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\mbox{for }\xi\in C,\quad
&\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-)=
-\varphi^\uparrow(\xi_w;w-)\\
&\quad=-\varphi^\rightarrow(\xi_w;w)e^{-i\pi/4}\\
&\quad=\varphi^\leftarrow(\xi_w;w)e^{i\pi/4}-\varphi^\rightarrow(\xi_w;w)e^{-i\pi/4}.
\end{split}
\end{equation}
That is to say, we have the identity
\begin{equation}
\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-)=
\varphi^\leftarrow(\xi_w;w)e^{i\pi/4}-\varphi^\rightarrow(\xi_w;w)e^{-i\pi/4}.
\end{equation}
This gives
\begin{equation}
\EE\big[(\sqrt2)^{L(\xi_w)}
\big(\varphi^\uparrow(\xi_w;w+)-\varphi^\uparrow(\xi_w;w-)\big)\big]=
Z_\d(\Phi_\d^\leftarrow(w)e^{i\pi/4}-\Phi_\d^\rightarrow(w)e^{-i\pi/4}).
\end{equation}
Similar considerations give
\begin{equation}
\EE\big[(\sqrt2)^{L(\xi_u)}
\big(\varphi^\uparrow(\xi_u;u+)-\varphi^\uparrow(\xi_u;u-)\big)\big]=
Z_\d(\Phi_\d^\rightarrow(w)e^{-i\pi/4}-\Phi_\d^\leftarrow(w)e^{i\pi/4}).
\end{equation}
Combining these and dividing by $Z_\d$ gives the
claim~\eqref{phi-up-dot}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{s-hol-thm-fk}]
As already noted,
properties~\eqref{s-hol-def-w-1} and~\eqref{s-hol-def-u-1}
are immediate, so we need to establish~\eqref{s-hol-def-w-2}
and~\eqref{s-hol-def-u-2}.
We check the case $z=w\in\Om_\d^{\circ,\mathrm{int}}$, the case
$z\in\Om_\d^{\b,\mathrm{int}}$ being similar.
Writing $u=w-\sfrac\d2$
and $v=w+\sfrac\d2$, we need to show that
\begin{equation}\label{phi-dots}
\begin{split}
\dot\Phi_\d^\uparrow(w)&=\dot\Phi_\d^\uparrow(u)=
\tfrac{i}{\d}\big(\Phi_\d^\downarrow(w)-\Phi_\d^\downarrow(u)\big),\\
\dot\Phi_\d^\downarrow(w)&=\dot\Phi_\d^\downarrow(v)=
\tfrac{i}{\d}\big(\Phi_\d^\uparrow(v)-\Phi_\d^\uparrow(w)\big).
\end{split}
\end{equation}
But for any
$z\in\Om_\d^{\circ,\mathrm{int}}\cup\Om_\d^{\b,\mathrm{int}}$
we have, by Lemma~\ref{turn-lem}, firstly
\begin{equation}\begin{split}
e^{i\pi/4}\Phi_\d^\leftarrow(z)-e^{-i\pi/4}\Phi_\d^\rightarrow(z)
&=e^{i\pi/2} e^{-i\pi/4}\Phi_\d^\leftarrow(z)-e^{-i\pi/2}e^{i\pi/4}\Phi_\d^\rightarrow(z)\\
&=i\cdot (e^{-i\pi/4}\Phi_\d^\leftarrow(z)+e^{i\pi/4}\Phi_\d^\rightarrow(z))\\
&=i\sqrt2\cdot \Phi_\d^\downarrow(z),
\end{split}\end{equation}
and secondly
\begin{equation}\begin{split}
e^{i\pi/4}\Phi_\d^\rightarrow(z)-e^{-i\pi/4}\Phi_\d^\leftarrow(z)
&=e^{i\pi/2} e^{-i\pi/4}\Phi_\d^\rightarrow(z)-e^{-i\pi/2}e^{i\pi/4}\Phi_\d^\leftarrow(z)\\
&=i\cdot (e^{-i\pi/4}\Phi_\d^\rightarrow(z)+e^{i\pi/4}\Phi_\d^\leftarrow(z))\\
&=i\sqrt2\cdot \Phi_\d^\uparrow(z).
\end{split}\end{equation}
Putting these into Lemma~\ref{phi-dot-lem} gives the
result.
\end{proof}
\section{The spin-observable}
\label{spin-obs-sec}
\subsection{Definition}
Let $\Om_\d$ be a discrete \emph{dual} domain (see
Section~\ref{dom-sec}).
We work with the random-parity representation~\eqref{rpr}
in $\Om_\d^\b$, and as before we set $h=J=\tfrac{1}{2\d}$.
Recall that the set $\xi=\xi^\circ\subseteq\Om_\d^\circ$ of bridges
is a Poisson process with rate $J$.
Define the `lower boundary' of $\Om_\d^\b$ as
\[
\partial^-\Om_\d^\b=\{z\in\partial\Om_\d^\b:
z-i\varepsilon\not\in\Om_\d^\b
\mbox{ for all $\varepsilon>0$ small enough}\},
\]
and similarly the `upper boundary' of $\Om_\d^\b$ as
\[
\partial^+\Om_\d^\b=\{z\in\partial\Om_\d^\b:
z+i\varepsilon\not\in\Om_\d^\b
\mbox{ for all $\varepsilon>0$ small enough}\},
\]
Thus
$\partial^-\Om_\d^\b\cup \partial^+\Om_\d^\b=\partial^\mathrm{h}\Om_\d^\b$.
We take two distinct points $a_\d,b_\d$
on the boundary $\partial\Om_\d$ with
$a_\d\in\partial^\mathrm{v}\Om_\d^\circ\cup\partial^\mathrm{h}\Om_\d^\b$ either a
white point on the `sides' or a black point on the `top or bottom',
and $b_\d\in\partial^-\Om_\d^\b$ on the lower boundary.
In the case when $a_\d\in\partial^\mathrm{v}\Om_\d^\circ$ we let
$a_\d^\mathrm{int}=a_\d\pm\sfrac\d2\in\Om_\d^\b$
be the black point in $\Om_\d$ next to
$a_\d$, so that $(a_\d,a_\d^\mathrm{int})$ is a directed half-edge
pointing horizontally into $\Om_\d$,
as in Figure~\ref{spin-dom-fig-1}.
If $a_\d\in\partial^\mathrm{h}\Om_\d^\b$ we
let $a_\d^\mathrm{int}=a_\d$ but sometimes interpret
$a_\d^\mathrm{int}=a_\d\pm0i$ as a point `just inside' $\Om_\d^{\b,\mathrm{int}}$.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.47}
\hspace{1cm}
\includegraphics{fermionic.48}
\caption{
Dual domain $\Om_\d$ with $\partial\Om_\d$ drawn dashed and
$\Om_\d^\b$ drawn solid.
\emph{Left:} A labelling $\psi_{a_\d,b_\d}^\xi$, with
points $u$ satisfying $\psi(u)=1$ marked fat, with blue colour
for loops and red for the path $\g$. In this case
$a_\d\in\partial^\mathrm{v}\Om_\d^\circ$.
\emph{Right:} Same domain with a labelling $\psi_{a_\d,z}^\xi$ for
$z\in\Om_\d^{\circ,\mathrm{int}}$. In this case $a_\d\in\partial^+\Om_\d^\b$.}
\label{spin-dom-fig-1}
\end{figure}
For $a_\d$ as above and for a fixed
$z\in\Om_\d^\b$, possibly $z=b_\d$, we let
$\psi=\psi_{a_\d,z}^\xi:\Om_\d^\b\to\{0,1\}$ be
a function satisfying the following:
\begin{enumerate}
\item $\psi(a_\d^\mathrm{int})=\psi(z)=1$ if $z\neq a_\d^\mathrm{int}$,
respectively $=0$ if $z=a_\d^\mathrm{int}$,
\item $\psi(u)=0$ for all $u\in\partial^\mathrm{h}\Om^\b_\d\setminus\{a_\d,z\}$,
\item for $u\in\Om_\d^{\b,\mathrm{int}}$ we have that
$\psi(u+\varepsilon i)=1-\psi(u-\varepsilon i)$ for all small enough $\varepsilon>0$
if \emph{either} $u\pm\sfrac\d2\in\xi$
(that is, $u$ is an endpoint of a bridge) \emph{or}
$u\in(\{a_\d^\mathrm{int}\}\triangle\{z\})$; and
\item the set $I(\psi)=\{u\in\Om_\d^\b:\psi(u)=1\}$ is closed.
\end{enumerate}
Thus $\psi$ is a random-parity configuration with sources
$A=\{a_\d^\mathrm{int},z\}$ and boundary condition 0 on
$\partial^\mathrm{h}\Om_\d^\b$.
It is easy to see that there is at most one function $\psi_{a_\d,z}^\xi$
satisfying the above constraints, for each given $\xi$ (and $a_\d,z$).
We let $\mathcal{A}(a_\d,z)$
be the event (set of $\xi$:s) such that there
exists such a $\psi$. We also extend the definition of $\mathcal{A}(a_\d,z)$ to
allow $z\in\Om_\d^{\circ}$ by letting
\begin{equation}
\mathcal{A}(a_\d,z)=\mathcal{A}(a_\d,z-\sfrac\d2)\cup\mathcal{A}(a_\d,z+\sfrac\d2)
\mbox{ if } z\in\Om_\d^{\circ}.
\end{equation}
Note that this union is disjoint.
It is worth stating precisely a (necessary and sufficient) condition for
$\xi$ to belong to $\mathcal{A}({a_\d,z})$ when $z\in\Om_\d^\b$.
To state the condition, let
\[
V(u)=\{u'\in\Om_\d^\b:[u,u']\subseteq\Om_\d^\b\},
\mbox{ for } u\in\Om_\d^\b,
\]
be the maximal vertical line contained in $\Om_\d^\b$ and containing
$u$. Let
\begin{equation}
S_{a_\d,z}^\xi(u)=\{v\in V(u): v\pm\sfrac\d2\in\xi\}
\cup (\{a_\d^\mathrm{int}\}\triangle\{z\})
\end{equation}
be the set of points in $V(u)$ where $\psi$ is required to change value.
Then, for $z\in\Om_\d^\b$,
\begin{equation}
\xi\in\mathcal{A}({a_\d,z})\Leftrightarrow
|S_{a_\d,z}^\xi(u)|\mbox{ is even for all } u\in\Om_\d^\b.
\end{equation}
In words,
$\psi$ must switch (from 0 to 1 or from 1 to 0)
an even number of times on each line $V(u)$.
If $u\in\Om_\d^\b\setminus\{a_\d^\mathrm{int}\}$
and $\xi\in\mathcal{A}({a_\d,u})$ then $\psi=\psi^\xi_{a_\d,u}$
contains a unique path $\g(\xi)$ from $a_\d$ to $u$ which traverses
the half-edge $(a_\d,a_\d^\mathrm{int})$ if $a_\d\in\partial^\mathrm{v}\Om_\d^\circ$,
intervals along which $\psi=1$, as well as bridges of $\xi$.
For $w\in\Om_\d^{\circ}$ and $\xi\in\mathcal{A}({a_\d,w})$
we complete $\g$ to form a path to $w$ by including the half-edge
$(w-\sfrac\d2)\to w$ (if $\xi\in\mathcal{A}(a_\d,w-\sfrac\d2)$)
respectively
$w\leftarrow(w+\sfrac\d2)$ (if $\xi\in\mathcal{A}(a_\d,w+\sfrac\d2)$).
See Figure~\ref{spin-dom-fig-1}.
In the cases when $z\in\{a_\d,a_\d^\mathrm{int}\}$ the path $\g$ is
degenerate, and we interpret it as a small arrow (or half-edge)
pointing from $a_\d$ to $a_\d^\mathrm{int}$ if $z=a_\d^\mathrm{int}$, alternatively as
a small path making an angle $\pi$ turn if $z=a_\d$.
We define $W^{a_\d,z}_{\g(\xi)}$ to be the
winding-angle of $\g(\xi)$ from $a_\d$ to $z$
(with $W^{a_\d,a_\d^\mathrm{int}}=0$ and $W^{a_\d,a_\d}=\pi$). It is important to note
that, in the case when $z=b_\d$ is on the boundary, then
$W^{a_\d,b_\d}_{\g(\xi)}$ does not depend on $\xi$ (one cannot wind around the
boundary, and $a_\d,b_\d$ have fixed orientations), i.e.\ it takes a fixed
value which we denote $W^{a_\d,b_\d}$.
Write $1^\circ(z)$ for the indicator that $\g$ ends with a half-edge
(i.e.\ either $z\in\Om_\d^\circ$ or $a_\d\in\partial^\mathrm{v}\Om_\d^\circ$
and $z=a_\d^\mathrm{int}$).
Define the random variable
\begin{equation}
X^{a_\d,z}(\xi)=\hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,z)}(\xi)
\exp(-2h|I(\psi_{a_\d,z}^\xi)|)
(\tfrac1{\sqrt2})^{1^\circ(z)}.
\end{equation}
\begin{definition}\label{spin-obs-def}
Write $\EE=\EE_{0,1/2\d}$ for the law of $\xi=\xi^\circ$ and let
$(\Om_\d,a_\d,b_\d)$ be as above.
Define the spin-observable
\begin{equation}
F^{\mathrm{sp}}_\d(z)=
\frac{\EE[\exp(-\tfrac{i}2W^{a_\d,z}_{\g(\xi)}) X_{\g(\xi)}^{a_\d,z}]}
{\EE[\exp(-\tfrac{i}2W^{a_\d,b_\d}_{\g(\xi)}) X_{\g(\xi)}^{a_\d,b_\d}]},
\quad z\in\Om_\d^\b\cup\Om_\d^{\circ}.
\end{equation}
\end{definition}
Note that we have defined this observable using the random-parity
representation, whose classical analogue is the random-current
representation of~\cite{aiz82} rather than the high-temperature expansion
used by Chelkak and Smirnov~\cite{ch-sm}.
The high-temperature expansion is essentially the random-current
representation `modulo two'.
\subsection{S-holomorphicity}
In this section we show the following result.
\begin{theorem}\label{s-hol-thm-spin}
Let $F^{\mathrm{sp}}_\d$ be the spin-observable in a primal domain
$(\Om_\d,a_\d,b_\d)$ with two marked points on the
boundary, as above. Then $F^{\mathrm{sp}}_\d$ is s-holomorphic at
all $z\in(\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}})\setminus\{a_\d^\mathrm{int},a_\d^\mathrm{int}\pm\sfrac\d2\}$.
\end{theorem}
Regarding the behaviour near $a_\d^\mathrm{int}$, we note that half of
condition~\eqref{s-hol-def-u-1} in Definition~\ref{s-hol-def}
holds at $a_\d^\mathrm{int}$, but condition~\eqref{s-hol-def-u-2} fails.
Since $a_\d$ and $b_\d$ are fixed
we will use the shorthands
\begin{equation}
\cW^z(\xi)=W^{a_\d,b_\d}-W^{a_\d,z}_{\g(\xi)},
\qquad
X^z(\xi)=X^{a_\d,z}(\xi).
\end{equation}
Note that $F^{\mathrm{sp}}_\d(z)$ is a real multiple of
\begin{equation}\label{sp-obs-pf-eq}
F_\d(z)=\EE\big[\exp(\tfrac{i}2 \cW^z(\xi))
X^{z}(\xi)\big],
\end{equation}
so it suffices to show s-holomorphicity of this $F_\d(z)$.
It will be useful to note the
following interpretation of the quantity $\cW^{z}(\xi)$.
Imagine that we augment $\g$ with a curve $\hat\g$ in $\Om_\d$
which starts at $z$
in the same direction that $\g$ ends, and which finishes at $b_\d$
(pointing down). Let $W^{z,b_\d}_{\hat\g}$ denote its winding angle.
Then $\Gamma=\g\cup\hat\g$ is a curve in
$\Om_\d^\b$ from $a_\d$ to $b_\d$,
thus $\Gamma$ has winding angle $W^{a_\d,b_\d}$,
meaning that $\cW^{z}(\xi)$ is the winding-angle from $z$ to $b_\d$.
We now turn to the proof of Theorem~\ref{s-hol-thm-spin}.
Recall that we define $F_\d^\a(z)$ by
\[
F_\d^\a(z)=\mathrm{Proj}[ F_\d(z); \ell(\a)],\quad
\a\in\{\uparrow,\downarrow,\leftarrow,\rightarrow\}.
\]
This means that the $F_\d^\a$
automatically satisfy the relations of the $\Phi_\d^\a$
in Lemma~\ref{turn-lem}, that is:
\begin{equation}\label{turn-eq-2}
\begin{split}
F_\d^\uparrow(z)&=\tfrac{1}{\sqrt2}\big(
e^{i\pi/4}F_\d^\leftarrow(z)+e^{-i\pi/4}F_\d^\rightarrow(z)\big),\\
F_\d^\downarrow(z)&=\tfrac{1}{\sqrt2}\big(
e^{i\pi/4}F_\d^\rightarrow(z)+e^{-i\pi/4}F_\d^\leftarrow(z)\big).
\end{split}
\end{equation}
If $z=u\in\Om_\d^\b\setminus\{a_\d^\mathrm{int}\}$
then $\g$ reaches $u$ either
from below or from above; we write these events pictorially as
\[
\Big\{\lower 2mm \hbox{\includegraphics{fermionic.32}}\Big\}
\qquad\mbox{and}\qquad
\Big\{\lower 2mm \hbox{\includegraphics{fermionic.33}}\Big\}.
\]
Similarly, if $z=w\in\Om_\d^\circ$ then $\g$ reaches $w$ either from the
left or the right, pictorially represented as
\[
\big\{\hbox{\includegraphics{fermionic.35}}\big\} \qquad\mbox{and}\qquad
\big\{\hbox{\includegraphics{fermionic.34}}\big\}.
\]
\begin{lemma}
If $u\in\Om_\d^\b\setminus\{a_\d^\mathrm{int}\}$ then
\begin{equation}\label{Phi-u-proj}
\begin{split}
F_\d^\leftarrow(u)&=\EE\Big[
\exp\big(\tfrac{i}2\cW^{u}\big) X^{u}
\hbox{\rm 1\kern-.27em I}\Big\{\lower 2mm \hbox{\includegraphics{fermionic.32}}\Big\}
\Big],\mbox{ and}\\
F_\d^\rightarrow(u)&=\EE\Big[
\exp\big(\tfrac{i}2\cW^{u}\big) X^{u}
\hbox{\rm 1\kern-.27em I} \Big\{\lower 2mm \hbox{\includegraphics{fermionic.33}}\Big\}
\Big],
\end{split}
\end{equation}
and if $w\in\Om_\d^\circ$,
\begin{equation}\label{Phi-w-proj}
\begin{split}
F_\d^\uparrow(w)&=\EE\Big[
\exp\big(\tfrac{i}2\cW^{w}\big) X^{w}
\hbox{\rm 1\kern-.27em I}\big\{\hbox{\includegraphics{fermionic.34}}\big\}
\Big],\mbox{ and}\\
F_\d^\downarrow(w)&=\EE\Big[
\exp\big(\tfrac{i}2\cW^{w}\big) X^{w}
\hbox{\rm 1\kern-.27em I}\big\{\hbox{\includegraphics{fermionic.35}}\big\}
\Big].
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
We show~\eqref{Phi-u-proj}, the argument for~\eqref{Phi-w-proj} is
similar.
Certainly the two terms on the right-hand-sides of~\eqref{Phi-u-proj}
sum to $F_\d(u)$. Moreover, if $\g$ reaches $u$ from below
then $\cW^{u}=\pi+2\pi n$ for some $n=n(\xi)\in\mathbb{Z}$, and if
$\g$ reaches $u$ from above
then $W^{u}=0+2\pi n$ for some $n=n(\xi)\in\mathbb{Z}$.
Thus the two terms belong to
$\ell(\leftarrow)$ and $\ell(\rightarrow)$ respectively, and these two lines being
perpendicular, the claim~\eqref{Phi-u-proj} follows.
\end{proof}
\begin{proposition}
Conditions~\eqref{s-hol-def-w-1} and~\eqref{s-hol-def-u-1} in
Definition~\ref{s-hol-def} hold at all
$z\in(\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}})\setminus\{a_\d^\mathrm{int},a_\d^\mathrm{int}\pm\sfrac\d2\}$.
\end{proposition}
\begin{proof}
We give details for the case $z=w\in\Om_\d^{\circ,\mathrm{int}}$, the case
$z\in\Om_\d^{\b,\mathrm{int}}$ being similar.
Writing $u=w-\sfrac\d2$, $v=w+\sfrac\d2$, we need to show that
(when neither $u$ nor $v$ equals $a_\d^\mathrm{int}$)
\[
F_\d^\uparrow(w)=F_\d^\uparrow(u) \mbox{ and }
F_\d^\downarrow(w) =F_\d^\downarrow(v).
\]
We give details for the case $\uparrow$ only, the claim for $\downarrow$ again
being similar.
Consider the terms in~\eqref{Phi-u-proj}.
Inside the expectations we have
\begin{equation} \label{Phi-u-wind-1}
\mbox{in }F_\d^\leftarrow(u),\quad
X^{u}(\xi)={\sqrt2} X^{w}(\xi)
\mbox{ and }
W^{a_\d,u}_{\g(\xi)}=W^{a_\d,w}_{\g(\xi)}+\tfrac\pi2,
\end{equation}
since if we add the half-edge from $u$ to $w$,
this puts an additional factor
$\sfrac1{\sqrt2}$ into $X$, and $\g$ does an additional
$-\sfrac\pi2$ turn. Similarly,
\begin{equation} \label{Phi-u-wind-2}
\mbox{in }F_\d^\rightarrow(u),\quad
X^{u}(\xi)={\sqrt2} X^{w}(\xi)
\mbox{ and }
W^{a_\d,u}_{\g(\xi)}=W^{a_\d,w}_{\g(\xi)}-\tfrac\pi2.
\end{equation}
We use the symbolic notation
\[
\Big\{\lower 2mm \hbox{\includegraphics{fermionic.36}}\Big\} \qquad\mbox{and}\qquad
\Big\{\lower 2mm \hbox{\includegraphics{fermionic.37}}\Big\}
\]
for the events that $\g$ ends with a right- or left-turn at $u$ into
$w$, respectively. Using~\eqref{turn-eq-2}, \eqref{Phi-u-proj},
\eqref{Phi-u-wind-1} and \eqref{Phi-u-wind-2}, we have for the case
when neither $u$ nor $v$ equals $a_\d^\mathrm{int}$:
\begin{equation}
\begin{split}
F_\d^\uparrow(u)&=\tfrac{1}{\sqrt2}\big(
e^{i\pi/4}F_\d^\leftarrow(u)+e^{-i\pi/4}F_\d^\rightarrow(u)\big)\\
&=\tfrac1{\sqrt2}\EE\Big[
\exp\big(\tfrac{i}2(\cW^{u}+\tfrac\pi2)\big)
X^{u} \hbox{\rm 1\kern-.27em I}\Big\{\lower 2mm \hbox{\includegraphics{fermionic.32}}\Big\}
\Big]\\
&\qquad+\tfrac1{\sqrt2}\EE\Big[
\exp\big(\tfrac{i}2(\cW^{u}-\tfrac\pi2)\big)
X^{u} \hbox{\rm 1\kern-.27em I}\Big\{\lower 2mm \hbox{\includegraphics{fermionic.33}}\Big\}
\Big]\\
&=\EE\Big[
\exp\big(\tfrac{i}2\cW^{w}\big) X^{w}
\hbox{\rm 1\kern-.27em I}\Big\{\lower 2mm \hbox{\includegraphics{fermionic.36}}\Big\}
\Big] \\ &\qquad
+\EE\Big[
\exp\big(\tfrac{i}2\cW^{w}\big) X^{w}
\hbox{\rm 1\kern-.27em I}\Big\{\lower 2mm \hbox{\includegraphics{fermionic.37}}\Big\}
\Big]\\
&=\EE\Big[
\exp\big(\tfrac{i}2\cW^{w}\big) X^{w}
\hbox{\rm 1\kern-.27em I}\big\{\hbox{\includegraphics{fermionic.34}}\big\}
\Big]\\
&=F_\d^\uparrow(w),\mbox{ as required.}\qedhere
\end{split}
\end{equation}
\end{proof}
The remaining conditions for s-holomorphicity take more work to
verify.
Theorem~\ref{s-hol-thm-spin} follows once we establish the following:
\begin{proposition}
Conditions~\eqref{s-hol-def-w-2} and~\eqref{s-hol-def-u-2} in
Definition~\ref{s-hol-def} hold at all
$z\in(\Om_\d^{\b,\mathrm{int}}\cup\Om_\d^{\circ,\mathrm{int}})\setminus\{a_\d^\mathrm{int},a_\d^\mathrm{int}\pm\sfrac\d2\}$.
\end{proposition}
\begin{proof}
Again we give details only for $z=w\in\Om_\d^{\circ,\mathrm{int}}$. Writing
$u=w-\sfrac\d2$, $v=w+\sfrac\d2$, we need to show
(as long as neither $u$ nor $v$ equals $a_\d^\mathrm{int}$) that
\[
\begin{split}
\dot F_\d^\uparrow(w)=\tfrac i\d
\big(F_\d^\downarrow(v)-F_\d^\downarrow(u)\big)\mbox{ and }
\dot F_\d^\downarrow(w)=\tfrac i\d
\big(F_\d^\uparrow(v)-F_\d^\uparrow(u)\big).
\end{split}
\]
We give details only for the case of $\dot F_\d^\uparrow(w)$.
Take $\varepsilon>0$ small, and consider
$F_\d^\uparrow(w+i\varepsilon)-F_\d^\uparrow(w)$. Note from~\eqref{Phi-w-proj} that
\begin{equation}
F_\d^\uparrow(w)=\tfrac1{\sqrt2}\EE\Big[
\exp\big(\tfrac{i}2\cW^w\big)
\exp\big(-2h|I(\psi_{a_\d,u})|\big)\hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,u)}
\Big].
\end{equation}
Also note that $\mathcal{A}(a_\d,u)=\mathcal{A}(a_\d,u+i\varepsilon)$ for $\varepsilon>0$ small.
We may thus write
\begin{equation}\label{phi-dot-exp-1}
\begin{split}
&F_\d^\uparrow(w+i\varepsilon)-F_\d^\uparrow(w)\\
&=\tfrac1{\sqrt2} \EE\Big[
\Big(e^{\sfrac{i}2\cW^{w+i\varepsilon}}
e^{-2h|I(\psi_{a_\d,u+i\varepsilon})|} - e^{\sfrac{i}2\cW^{w}}
e^{-2h|I(\psi_{a_\d,u})|}\Big) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,u)}
\Big].
\end{split}
\end{equation}
We will split the expectation into the two cases:
(i) $\xi(w,w+i\varepsilon)=0$, and
(ii) $\xi(w,w+i\varepsilon)>0$, i.e.\ according to whether there is
a bridge in the interval $(w,w+i\varepsilon)$ or not.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.16}\hspace{1cm}
\includegraphics{fermionic.17}
\caption{The curve $\g$ finishes with a right turn at $u$ (in
$\psi^\xi_{a,w}$, displayed left)
respectively $u+i\varepsilon$ (in $\psi^\xi_{a_\d,w+i\varepsilon}$, displayed right).
The winding angle is the same in both cases.}
\label{spin-der-fig-1}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.18}\hspace{1cm}
\includegraphics{fermionic.19}
\caption{Here $u+i\varepsilon$ is contained in a loop (in
$\psi^\xi_{a_\d,w}$, displayed left), which becomes part of
$\g$ (in $\psi^\xi_{a_\d,w+i\varepsilon}$, displayed right).
The winding angle is still the same in
both cases.}
\label{spin-der-fig-2}
\end{figure}
The first case, when there is no bridge,
is illustrated in Figures~\ref{spin-der-fig-1}
and~\ref{spin-der-fig-2}.
In this case we have that
$W^{a_\d,w+i\varepsilon}_\g=W^{a_\d,w}_\g$
and hence $\cW^w=\cW^{w+i\varepsilon}$.
Let
\begin{equation}
\hat\varepsilon=|I(\psi_{a_\d,u})|-|I(\psi_{a_\d,u+i\varepsilon})|,
\end{equation}
and note that $-\varepsilon\leq\hat\varepsilon\leq\varepsilon$.
We may thus write the contribution from case (i) to the
expectation in~\eqref{phi-dot-exp-1} as
\begin{equation}\label{phi-dot-exp-2}
\tfrac1{\sqrt2}\EE\Big[
e^{\sfrac{i}2\cW^w}
e^{-2h|I(\psi_{a_\d,u})|}
\big(e^{2h\hat\varepsilon}-1\big) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,u)}
\hbox{\rm 1\kern-.27em I}\{\xi(w,w+i\varepsilon)=0\}
\Big].
\end{equation}
Since the factor $e^{2h\hat\varepsilon}-1$ is of order $O(\varepsilon)$ we can (up
to an error of order $O(\varepsilon^2)$) ignore events of probability
$O(\varepsilon)$. Thus we may assume that there is no bridge
in $(w-\d,w-\d+i\varepsilon)$ (i.e.\ we have a situation as in
Figure~\ref{spin-der-fig-1}, not as in
Figure~\ref{spin-der-fig-2}). Under the latter assumption we have
that
\begin{equation}
\hat\varepsilon=\left\{\begin{array}{ll}
+\varepsilon, & \mbox{if $\g$ comes from above},\\
-\varepsilon, & \mbox{if $\g$ comes from below}.
\end{array}\right.
\end{equation}
Thus, up to an error of order $O(\varepsilon^2)$, the integrand
in~\eqref{phi-dot-exp-2} equals
\begin{equation}\begin{split}
2h\varepsilon\Big( e^{\sfrac{i}2\cW^w}
e^{-2h|I(\psi_{a_\d,u})|}\hbox{\rm 1\kern-.27em I}\{\lower 2mm \hbox{\includegraphics{fermionic.37}}\}
- e^{\sfrac{i}2\cW^w}
e^{-2h|I(\psi_{a_\d,u})|}\hbox{\rm 1\kern-.27em I}\{\lower 2mm \hbox{\includegraphics{fermionic.36}}\}\Big).
\end{split}\end{equation}
In the first term we have $W_\g^{a_\d,w}=W_\g^{a_\d,u}+\pi/2$
and in the second term we have $W_\g^{a_\d,w}=W_\g^{a_\d,u}-\pi/2$.
Dividing by $\varepsilon$ and letting $\varepsilon\downarrow0$,
it follows that the contribution to $\dot F_\d^\uparrow(w)$ from case
(i) is
\begin{equation}
\begin{split}
& h\sqrt2 (e^{-i\pi/4}F_\d^\rightarrow(u) -e^{i\pi/4}F_\d^\leftarrow(u) )\\
&\quad=-ih\sqrt2 (e^{i\pi/4}F_\d^\rightarrow(u) +e^{-i\pi/4}F_\d^\leftarrow(u) )\\
&\quad=-2ih F_\d^\downarrow(u)= -\tfrac{i}{\d}F_\d^\downarrow(u).
\end{split}
\end{equation}
We now turn to case (ii), when there is a bridge in $(w,w+i\varepsilon)$.
We need to show that the contribution
from this case is $\tfrac i\d F_\d^\downarrow(w)=\tfrac i\d F_\d^\downarrow(v)$.
We start by noting that, up to an error of order $O(\varepsilon^2)$, we may
in fact assume that $\xi$ belongs to the event
\begin{equation}\label{B}
B=\left\{\begin{array}{l}
\xi(w,w+i\varepsilon)=1,\\
\xi(w-\d,w-\d+i\varepsilon)=0,\mbox{ and}\\
\xi(w+\d,w+\d+i\varepsilon)=0.
\end{array}\right\}
\end{equation}
The possible scenarios are illustrated in Figures~\ref{case-iia-fig},
\ref{case-iib-fig}, \ref{case-iic-fig} and \ref{case-iid-fig}.
We write $\hat w$ for the location of the unique bridge in
$(w,w+i\varepsilon)$. Recall the notation
$\xi_{\hat w}=\xi\triangle\{\hat w\}$ for the
configuration obtained by removing
the bridge at $\hat w$ from $\xi$.
We have that
\begin{equation}
\xi\in\mathcal{A}(a_\d,u) \Leftrightarrow \xi\in\mathcal{A}(a_\d,u+i\varepsilon)
\Leftrightarrow \xi_{\hat w}\in\mathcal{A}(a_\d,v).
\end{equation}
Moreover, we have that the quantities
\begin{equation}
\begin{split}
\hat\varepsilon_1&=|I(\psi_{a_\d,v}^{\xi_{\hat w}})|
-|I(\psi_{a_\d,u+i\varepsilon}^{\xi})|\\
\hat\varepsilon_2&=|I(\psi_{a_\d,v}^{\xi_{\hat w}})|
-|I(\psi_{a_\d,u}^{\xi})|
\end{split}
\end{equation}
satisfy $-2\varepsilon\leq \hat\varepsilon_1,\hat\varepsilon_2\leq 2\varepsilon$.
The contribution from case (ii) to the expectation
in~\eqref{phi-dot-exp-1} may thus, up to an error of order
$O(\varepsilon^2)$, be written as
\begin{equation}\label{phi-dot-exp-3}
\begin{split}
&\EE\Big[\hbox{\rm 1\kern-.27em I}_B(\xi) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,v)}(\xi_{\hat w})
X^{w}(\xi_{\hat w})
\Big(e^{\sfrac{i}2\cW^{w+i\varepsilon}(\xi)}
e^{2h\hat \varepsilon_1} -
e^{\sfrac{i}2\cW^{w}(\xi)}
e^{2h\hat \varepsilon_2} \Big) \Big]\\
&=\EE\Big[\hbox{\rm 1\kern-.27em I}_B(\xi) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,v)}(\xi_{\hat w})
X^{w}(\xi_{\hat w})
\Big(e^{\sfrac{i}2\cW^{w+i\varepsilon}(\xi)}
- e^{\sfrac{i}2\cW^{w}(\xi)}
\Big) \Big]+O(\varepsilon^2).
\end{split}
\end{equation}
We used that the event $B$ has probability $O(\varepsilon)$ and that both
$e^{2h\hat\varepsilon_1}=1+O(\varepsilon)$ and $e^{2h\hat\varepsilon_2}=1+O(\varepsilon)$.
It remains to understand the factor
\[
\hbox{\rm 1\kern-.27em I}_B(\xi) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,v)}(\xi_{\hat w})
\big(
e^{\sfrac{i}2\cW^{w+i\varepsilon}(\xi)}
- e^{\sfrac{i}2\cW^{w}(\xi)}
\big)
\]
We claim that, for $\xi\in B$ and $\xi_{\hat w}\in\mathcal{A}(a_\d,v)$,
\begin{equation}\label{phi-dot-exp-4}
e^{\sfrac{i}2\cW^{w+i\varepsilon}(\xi)}
- e^{\sfrac{i}2\cW^{w}(\xi)}
=2i\cdot
e^{\sfrac{i}2\cW^{w}(\xi_{\hat w})}.
\end{equation}
Before showing this, we explain how to finish the proof.
From~\eqref{phi-dot-exp-3}, and assuming~\eqref{phi-dot-exp-4}, the
contribution to $\dot F_\d^\uparrow(w)$ from case (ii) is
\begin{equation}\label{phi-dot-exp-5}
\begin{split}
&2i\cdot \lim_{\varepsilon\dn0} \tfrac1\varepsilon
\EE\Big[\hbox{\rm 1\kern-.27em I}_B(\xi) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,v)}(\xi_{\hat w})
X^{w}(\xi_{\hat w})
e^{\sfrac{i}2\cW^{w}(\xi_{\hat w})}
\Big]\\
&=
2i J \EE\Big[
e^{\sfrac{i}2\cW^{w}(\xi)}
X^{w}(\xi) \hbox{\rm 1\kern-.27em I}_{\mathcal{A}(a_\d,v)}(\xi)
\Big]\\
&=\tfrac i\d F_\d^\downarrow(w),
\end{split}
\end{equation}
as required (we used~\eqref{Phi-w-proj}).
It remains to show~\eqref{phi-dot-exp-4}.
There are 4 sub-cases to consider, depending on whether $\g$ traverses
$\hat w$ (in $\psi_{a_\d,u}^\xi$), in which direction, et.c.
The first case, which we call case (a), is defined by the condition
$\psi_{a_\d,u}^\xi(u+0i)=1$ and is depicted in Figure~\ref{case-iia-fig}.
In this case $\g(\xi)$ necessarily traverses $\hat w$ from right to
left.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.20}\hspace{1cm}
\includegraphics{fermionic.21}\hspace{1cm}
\includegraphics{fermionic.22}
\caption{Case (ii)(a), with $\psi^\xi_{a_\d,w}$ to the left,
$\psi^\xi_{a_\d,w+i\varepsilon}$ in the middle, and
$\psi^{\xi_{\hat w}}_{a_\d,w}$ to the right.}
\label{case-iia-fig}
\end{figure}
It is not hard to see that we get
\begin{equation}
\mbox{case (a):}\quad
W^{a_\d,w}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}+\pi,\quad
W^{a_\d,w+i\varepsilon}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}-\pi.
\end{equation}
This establishes~\eqref{phi-dot-exp-4} for case (a).
In the remaining 3 cases we will have $\psi^\xi_{a_\d,u}(u+0i)=0$,
meaning that (in $\psi^\xi_{a_\d,u}$) $\g$ can traverse $\hat w$
from right to left (case (b)), from left to right (case (c)), or not
at all (case (d)). The cases are depicted in
Figures~\ref{case-iib-fig},
\ref{case-iic-fig} and \ref{case-iid-fig},
respectively. We get the following:
\begin{equation}
\mbox{case (b):}\quad
W^{a_\d,w}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}+\pi,\quad
W^{a_\d,w+i\varepsilon}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}-\pi.
\end{equation}
\begin{equation}
\mbox{case (c):}\quad
W^{a_\d,w}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}-3\pi,\quad
W^{a_\d,w+i\varepsilon}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}-\pi.
\end{equation}
\begin{equation}
\mbox{case (d):}\quad
W^{a_\d,w}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}+\pi,\quad
W^{a_\d,w+i\varepsilon}_{\g(\xi)}= W^{a_\d,w}_{\g(\xi_{\hat w})}+3\pi.
\end{equation}
In all cases we see that~\eqref{phi-dot-exp-4} holds, as claimed.
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.23}\hspace{1cm}
\includegraphics{fermionic.24}\hspace{1cm}
\includegraphics{fermionic.25}
\caption{Case (ii)(b), with $\psi^\xi_{a_\d,w}$ to the left,
$\psi^\xi_{a_\d,w+i\varepsilon}$ in the middle, and
$\psi^{\xi_{\hat w}}_{a_\d,w}$ to the right.}
\label{case-iib-fig}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.26}\hspace{1cm}
\includegraphics{fermionic.27}\hspace{1cm}
\includegraphics{fermionic.28}
\caption{Case (ii)(c), with $\psi^\xi_{a_\d,w}$ to the left,
$\psi^\xi_{a_\d,w+i\varepsilon}$ in the middle, and
$\psi^{\xi_{\hat w}}_{a_\d,w}$ to the right.}
\label{case-iic-fig}
\end{figure}
\begin{figure}[hbt]
\centering
\includegraphics{fermionic.29}\hspace{1cm}
\includegraphics{fermionic.30}\hspace{1cm}
\includegraphics{fermionic.31}
\caption{Case (ii)(d), with $\psi^\xi_{a_\d,w}$ to the left,
$\psi^\xi_{a_\d,w+i\varepsilon}$ in the middle, and
$\psi^{\xi_{\hat w}}_{a_\d,w}$ to the right.}
\label{case-iid-fig}
\end{figure}
\end{proof}
\newpage
\section{Discussion}\label{disc-sec}
\subsection{Convergence of the observables}
As mentioned in the Introduction we expect that
both the {\sc fk}- and
spin-observables, suitably rescaled, converge as $\d\to0$.
We sketch an outline of a possible argument,
following the arguments for the classical case
(see~\cite{ch-sm,dum-cop,smi-fk}).
As also mentioned, the details in the case of the
{\sc fk}-observable were supplied by Li~\cite{li}
shortly after this paper was finished.
We take the discrete domains $(\Om_\d,a_\d,b_\d)$ to approximate a
continuous
domain $(\Om,a,b)$ (e.g.\ in the Carath\'eodory sense, i.e.\ convergence on
compact subsets of suitably normalized conformal maps from the upper
half-plane into the domains, see~\cite[Definition~3.10]{dum-cop}).
The two main steps are to show (i) precompactness of
sequences of s-holomorphic
functions $(F_\d)_{\d>0}$, and (ii) convergence of the auxiliary
functions $(H_\d)_{\d>0}$ given in Proposition~\ref{H-prop}.
For (i), note first that
preholomorphic functions, and hence in particular
s-holomorphic functions, are $\D_\d$-harmonic. Indeed,
if $F_\d$ satisfies~\eqref{prehol} at $z$ and $z\pm\sfrac\d2$, then
differentiating twice using~\eqref{prehol} gives
\begin{equation}\begin{split}
\ddot F_\d(z)&=\tfrac1{i\d}\big(
\dot F_\d(z-\sfrac\d2)-\dot F_\d(z+\sfrac\d2) \big)\\
&=-\tfrac1{\d^2}
\big(F_\d(z-\d)+F_\d(z+\d)-2F_\d(z)\big).
\end{split}\end{equation}
Thus precompactness of s-holomorphic functions would follow from
Lipschitzness of $\D_\d$-harmonic functions combined with a suitable
boundedness condition, using the Arzela--Ascoli theorem as
in~\cite[Proposition~8.7]{dum-cop}.
Completing this argument would require estimates
for the Green's function $G_\d(\cdot)$ in $\mathbb{C}_\d^\b$,
in particular a suitable form of the asymptotics of
$G_\d(z)$ as $|z|\to\infty$ as in~\cite{kenyon} and~\cite{bucking,ch-sm-cplx}.
See Section~3.4 of Li's paper~\cite{li} for details in the
present context.
For (ii), consider the sub- and superharmonic functions
$H^\b_\d=H_\d\!\mid_{\Om_\d^\b}$ and
$H^\circ_\d=H_\d\!\mid_{\Om_\d^\circ}$ (see Proposition~\ref{H-prop}). It is
not hard to partly determine the behaviour of these functions on the
boundary. In the case of the {\sc fk}-observable we can choose the
additive constant so that $H^\b_\d=1$
on the black part $\partial^\b_\d$ and $H_\d^\circ=0$ on the white part
$\partial^\circ_\d$.
In the case of the spin-observable
the constant can be chosen so
that $H_\d^\circ(w)=0$ for all $w\in\partial\Om_\d^\circ\setminus\{a_\d\}$
(note also that $\nu(z)^{1/2} F_\d^\mathrm{sp}(z)\in\RR$ for all
$z\in\partial^\mathrm{v}\Om_\d^\circ\cup\partial^\mathrm{h}\Om_\d^\b$
where $\nu(z)$ is the counter-clockwise
oriented unit tangent).
To fully determine the boundary-behaviour one could try to use a
variant of the `boundary modification trick' of~\cite{ch-sm}
(this is the approach taken by Li~\cite{li}).
In the case of
the {\sc fk}-observable one could alternatively note that the difference of
$H_\d$ on the boundary
and `just inside' the boundary is proportional to a
percolation-probability which converges to zero away from $a_\d,b_\d$,
like in the original argument for the square-lattice case~\cite{smi-fk}
(this uses that the phase-transition is continuous~\cite{bjogr}).
Having determined the boundary-values of $H^\b_\d$ and $H^\circ_\d$ one
would show that these functions are close to the harmonic function $h$
in $\Om$
with the corresponding boundary-values.
In the case of the {\sc fk}-observable we have $h=1$ on the clockwise arc
from $a$ to $b$ and $h=0$ on the counter-clockwise arc, whereas for
the spin-observable we have $h=0$ on $\partial\Om\setminus\{a\}$.
Since these are also the boundary-conditions for the
classical case~\cite{ch-sm,smi-fk}, we expect the
observables to converge to the same limits under the same
rescaling, namely
\begin{equation}
\tfrac1{\sqrt{\d}} F_\d^\mathrm{FK}(\cdot)\to \sqrt{\phi'(\cdot)},
\quad
F_\d^\mathrm{sp}(\cdot)\to\sqrt{\tfrac{\psi'(\cdot)}{\psi'(b)}},
\end{equation}
where $\phi$ is a conformal map from $\Om$ to
$\RR+i(0,1)$ mapping $a$ to $-\infty$ and $b$ to $+\infty$,
and $\psi$ is a conformal map from $\Om$ to the upper half-plane
mapping $a$ to $\infty$ and $b$ to 0.
As mentioned, the first of these limits has now
been established by Li~\cite{li}.
\subsection{Parafermionic observables}
Recall from~\eqref{fk-dens} that the {\sc fk}--Ising model at the critical
parameters $h=J=1/2\d$ has density proportional to $(\sqrt2)^{L(\xi)}$
with respect to a Poisson law, where $L(\xi)$ is
the number of loops. It is natural to ask also about measures with
density $(\sqrt q)^{L(\xi)}$ for other $q>0$.
Such measures arise in
the Aizenman--Nachtergaele representation~\cite{aizenman_nacht}
of a class of quantum spin
systems which includes the (spin-$\tfrac12$) Heisenberg
antiferromagnet as the case $q=4$. One may define an analog of the
{\sc fk}--Ising observable~\eqref{fk-obs-eq}
which is also a direct analog of Smirnov's parafermionic
observable for critical random-cluster models~\cite{smi-fk}. We
briefly describe this now.
Let $(\Om_\d,a_\d,b_\d)$ be a Dobrushin-domain as in
Section~\ref{fk-obs-sec} and let $\sigma$ satisfy
$\sin(\sigma\tfrac\pi2)=\tfrac12\sqrt q$.
Thus $\sigma=\tfrac12$ for $q=2$ ({\sc tfim}) and $\sigma=1$ for $q=4$ (Heisenberg
model).
Recall the events
$\Gamma^\a_z=\{\g\mbox{ passes $z$ in direction }\a\}$
and the winding-angle $W^\a_\g(z)$ of the interface
to the exit. We now define
\b
\varphi^\a(\xi; z)=\hbox{\rm 1\kern-.27em I}_{\Gamma^\a_z}(\xi)\exp(i\sigma W^\a_{\g(\xi)}(z)).
\end{equation}
Let $\hat\EE_\d$ denote the measure with density proportional to
$(\sqrt q)^{L(\xi)}$ with respect to the Poisson law with rate
$\tfrac1{\d\sqrt q}$.
Similarly to before we define observables
\
\Phi_\d^\uparrow(z)=\hat\EE_\d[\varphi^\uparrow(\xi;z)],\quad
\Phi_\d^\downarrow(z)=\hat\EE_\d[\varphi^\downarrow(\xi;z)],
\]
as well as
$F_\d(z)=\Phi_\d^\uparrow(z)+\Phi_\d^\downarrow(z)$.
Some properties of these quantities are immediate, e.g.\ for
$w\in\Om_\d^{\circ,\mathrm{int}}$ we still have
$\Phi^\uparrow_\d(w)=\Phi_\d^\uparrow(w-\sfrac\d2)$ and
$\Phi^\downarrow_\d(w)=\Phi_\d^\downarrow(w+\sfrac\d2)$, and also a version of
Lemma~\ref{phi-dot-lem} holds. It might be interesting to investigate
these observables further, especially due to the connection with the
Heisenberg antiferromagnet.
\subsection*{Acknowledgement}
This work was mainly carried out while the author
was at the University of Copenhagen in Denmark.
The author is now supported by
Vetenskapsr{\aa}det grant 2015-05195.
|
\section{Introduction}
For a natural number $m$, the $x$-th generalized $m$-gonal number is given by $p_m(x)=\frac{(m-2)x^2-(m-4)x}{2}$ where $x\in \mathbb Z$. In 1638, Fermat claimed that every natural number may be written as the sum of at most $3$ triangular numbers, $4$ squares, $5$ pentagonal numbers, and in general $m$ $m$-gonal numbers. Lagrange proved the four squares theorem (the $m=4$ case) in 1770, Gauss proved the the triangular number theorem (the $m=3$ case) in 1796, and Cauchy proved the full claim in 1813 \cite{Cauchy}. Guy \cite{Guy} later investigated the minimal number $r_m\in\mathbb{N}$ chosen such that every natural number may be written as the sum of $r_m$ generalized $m$-gonal numbers. For $m\geq 8$, Guy noted that an elementary argument shows that one needs $m-4$ generalized $m$-gonal numbers to represent $m-4$, so $m-4\leq r_m\leq m$. However, he pointed out that for large enough $n\in\mathbb{N}$, one could likely represent $n$ with significantly fewer generalized $m$-gonal numbers. In this paper, we investigate for which $m$ every sufficiently large $n\in\mathbb{N}$ is the sum of three $m$-gonal numbers. That is to say, we study representations of natural numbers by the ternary sum
\[
P_m(x,y,z):=p_m(x)+p_m(y)+p_m(z),
\]
and we ask for which $m$ the form $P_m$ is almost universal; a form is called \begin{it}almost universal\end{it} if it represents all but finitely many natural numbers. In other words, we would like to determine the set of $m$ for which the set
\[
\mathcal{S}_m:=\!\left\{n\in\mathbb{N}: \not\exists (x,y,z)\in \mathbb{Z}^3\text{ with } P_m(x,y,z)=n\right\}
\]
is finite. The set $\mathcal{S}_m$ is those positive integers which are not represented by $P_m$, and we call $P_m$ \begin{it}almost universal\end{it} if $\mathcal{S}_m$ is finite.
\begin{theorem}\label{thm:main}
If $m\not\equiv 2\pmod{3}$ and $4\nmid m$, then $P_m$ is almost universal.
\end{theorem}
\begin{remarks}
\noindent
\noindent
\begin{enumerate}[leftmargin=*,label={\rm(\arabic*)}]
\item
Theorem \ref{thm:main} states that for $m\not\equiv 2\pmod{3}$ and $4\nmid m$, every sufficiently large natural number may be written as the sum of at most three generalized $m$-gonal numbers. However, its proof relies on Siegel's ineffective bound \cite{Siegel} for the class numbers of imaginary quadratic orders, so the result does not give an explicit bound $n_m$ such that every $n>n_m$ may be written as the sum of three generalized $m$-gonal numbers.
\item
Questions of almost universality have recently been studied by a number of authors, but in a slightly different way. In most cases, $(m_1,m_2,m_3)\in\mathbb{N}_{\geq 3}^3$ has been fixed and authors investigated representations by weighted sums of the type
\[
ap_{m_1}(x)+bp_{m_2}(y)+cp_{m_3}(z).
\]
In particular, authors worked on the classification of $a,b,c$ for which the above form is almost universal; on the contrary, in Theorem \ref{thm:main} we fix $a$, $b$, and $c$ and vary $m$. In \cite{KS08}, for example, a classification of such $a,b,c$ was given for $m_1=m_2=4$ and $m_3=3$. In the case of a weighted sum of triangular numbers, a partial answer was given in \cite{KS08}, and the characterization of such almost universal sums was completed by Chan and Oh in \cite{CO09}. The most general result to date appears in \cite{H14}, where a characterization of almost universal weighted sums of $m$-gonal numbers is given for $m-2=2p$ with $p$ an odd prime. In this last case, the results in \cite{H14} imply that $P_m$ is not almost universal (note that $m\equiv 0\pmod{4}$, so this is partially complementary to the result in Theorem \ref{thm:main}).
\end{enumerate}
\end{remarks}
It turns out that the restrictions $m\not\equiv 2\pmod{3}$ and $4\nmid m$ are both necessary in Theorem \ref{thm:main}, but are of a very different nature. If $4\mid m$, then there is a \begin{it}local obstruction\end{it} to $P_m$ being almost universal, i.e., there is an entire congruence class $A\mathbb{N}_0+B\subseteq \mathcal{S}_m$ because it is not even represented modulo $A$. Details of these local obstructions may be found in Lemma \ref{lem:local}.
The restriction $m\not\equiv 2\pmod{3}$ (with $4\nmid m$) is much more delicate, and a seemingly deep connection between the analytic and algebraic theory lies beneath this case. In this case, there are no local obstructions, but $\mathcal{S}_m$ is not necessarily finite. To get a better understanding of the set $\mathcal{S}_m$, for $m$ even we define
\[
\mathcal{S}_{m,3}^{\operatorname{e}}:=\!\left\{n\in \mathcal{S}_m: \exists r\in\mathbb{Z}\text{ with }2(m-2)n+3\left(\frac{m-4}{2}\right)^2 = 3r^2\right\}
\]
and for $m$ odd we define
\[
\mathcal{S}_{m,3}^{\operatorname{o}}:=\!\left\{n\in \mathcal{S}_m: \exists r\in\mathbb{Z}\text{ with }8(m-2)n+3(m-4)^2 = 3r^2\right\}.
\]
We next see that if $m\equiv 2\pmod{3}$, then most of the exceptional set $\mathcal{S}_m$ is contained in $\mathcal{S}_{m,3}^{\operatorname{e}}$ if $m\equiv 2\pmod{4}$ and contained in $\mathcal{S}_{m,3}^{\operatorname{o}}$ if $m$ is odd.
\begin{theorem}\label{thm:2mod3}
\noindent
\noindent
\begin{enumerate}[leftmargin=*,label={\rm(\arabic*)}]
\item
If $m\equiv 2\pmod{12}$, then $\mathcal{S}_m\setminus\mathcal{S}_{m,3}^{\operatorname{e}}$ is finite.
\item
If $m\equiv 2\pmod{3}$ and $m$ is odd, then $\mathcal{S}_m\setminus\mathcal{S}_{m,3}^{\operatorname{o}}$ is finite.
\end{enumerate}
\end{theorem}
The proof of Theorem \ref{thm:main} mainly uses the analytic approach and relies on ineffective bounds of class numbers. However, this approach fails to gain any control in determining the sets $\mathcal{S}_{m,3}^{\operatorname{e}}$ and $\mathcal{S}_{m,3}^{\operatorname{o}}$. One is hence motivated to blend the two approaches together in order to investigate these sets. From the algebraic point of view, representations of an integer $n$ by $P_m$ is equivalent to representations of a related integer by a lattice coset $L+\nu$, where $L=L_{(m)}$ and $\nu=\nu_{(m)}$ are completely determined by $m$ (for the precise formulation of $L$ and $\nu$ see Section \ref{sec:prelim}). To explain why combining the algebraic and analytic theories may be beneficial, we recall an important interplay between the analytic and algebraic theories which occurs when $P_m$ is replaced with the quadratic form $Q$ on the positive-definite ternary lattice $L$, called the \begin{it}norm\end{it} on $L$, which we later emulate. To understand the link, for such a lattice $L$, let $\mathcal{L}$ denote the primitive elements of $L$ (those which are not non-zero integral multiples of other elements of $L$) and set
\[
\mathcal{S}_{L}:=\{n\in\mathbb{N} : \not\exists \alpha\in \mathcal{L}\text{ with }Q(\alpha)=n\}.
\]
Since $L$ is a positive-definite lattice, there will always be local obstructions at an odd number of finite primes, but our main consideration is those $n\in\mathcal{S}_{L}$ which are locally represented, which we refer to as \begin{it}locally admissible\end{it}. Moreover, there are finitely many primes $p$ (known as anisotropic primes) for which every $\nu\in L$ with $Q(\nu)$ highly divisible by $p$ is necessarily imprimitive (i.e., $\nu=p\nu'$ for some $\nu'\in L$). Therefore, if $\text{ord}_p(n)$ is large, we immediately conclude that $n\in \mathcal{S}_L$, so we restrict $\text{ord}_p(n)$. Using the analytic theory, one can show that for the special case where $L$ is a lattice, the subset of $n\in \mathcal{S}_L$ which are both locally admissible and have bounded $p$-adic order (by a specific constant depending on $p$) at all anisotropic primes $p$ is finite outside of finitely many square classes $t_1\mathbb{Z}^2,\dots,t_{\ell}\mathbb{Z}^2$. This follows from a result of Duke and Schulze-Pillot in \cite{DS-P90}. The behavior inside these square classes is explained via the spinor norm map in the algebraic theory; this occurs by realizing $t_j\mathbb{Z}^2$ as a spinor exceptional square class; the {\em primitive spinor exceptions} for the genus of $L$ are those integers which are primitively represented by some but not all of the spinor genera in the genus of $L$. These primitive spinor exceptions are determined by Earnest, Hsia, and Hung in \cite{EH94}. With additional investigation one can use these results to determine the existence of infinite subsets of admissible elements of $\mathcal{S}_{L}$; from this one can determine that if the subset of admissible elements of $\mathcal{S}_L$ with bounded divisibility at the anisotropic primes is infinite, then there is at least one spinor exceptional square class.
Returning to our case $P_m$, one would expect a similar theory of spinor exceptional square classes to emerge if one could link the algebraic and analytic approaches. It is revealing that for $4\nmid m$ the only possibly infinite part of $n\in \mathcal{S}_m$ occurs when $An+B$ is within the square class $3\mathbb{Z}^2$, hinting at a synthesis between the approaches yet to be investigated. In order to state a conjectural link in our case, we next recall the link between the two approaches in a little more detail.
The main synthesis between the analytic and algebraic theories goes through the Siegel--Weil (mass) formula. For a lattice $L_0$ in a positive-definite space, let $\mathbb{G}}%\operatorname{gen}(L_0)$ be a set of representatives of the classes in the genus $\operatorname{gen}(L_0)$ of $L_0$. One version of the Siegel--Weil formula states that
\[
\mathcal{E}_{\operatorname{gen}\!\left(L_0\right)}:=\frac{1}{\sum_{L\in\mathbb{G}}%\operatorname{gen}\!\left(L_0\right)} \omega_L^{-1}}\sum_{L\in \mathbb{G}}%\operatorname{gen}\!\left(L_0\right)} \frac{\Theta_L}{\omega_L}
\]
is a certain Eisenstein series. Here $\Theta_L$ is the theta function associated to $L$ (i.e., the generating function for the elements of $L$ of a given norm; see \eqref{eqn:thetadef}) and $\omega_L$ is the number of \begin{it}automorphs\end{it} of $L$ (i.e., the number of linear \begin{it}isometries\end{it} from $L$ to itself; these are invertible linear maps on the vector space $\mathbb{Q} L$ which fix $L$ and preserve the associated quadratic form $Q$). If $L_0$ has rank $3$, then if one instead takes the associated sum over a set $\mathbb{S}} %\operatorname{spn}(L_0)$ of representatives of the classes of lattices
in the spinor genus $\operatorname{spn}(L_0)$, one obtains
\begin{equation}\label{eqn:spnquadratic}
\frac{1}{\sum_{L\in\mathbb{S}} %\operatorname{spn}(L_0)} \omega_L^{-1}}\sum_{L\in \mathbb{S}} %\operatorname{spn}(L_0)} \frac{\Theta_L}{\omega_L}= \mathcal{E}_{\operatorname{gen}(L_0)} + \mathcal{U}_{\operatorname{spn}(L_0)},
\end{equation}
where $\mathcal{U}_{\operatorname{spn}(L_0)}$ is a linear combination of unary theta functions \cite{SP,SP2}. The Fourier coefficients of $\mathcal{U}_{\operatorname{spn}(L_0)}$ count the excess or deficiency of the weighted average of the number of representations by the spinor genus of $L_0$ when compared with the weighted average of the number of representations by the genus, giving a direct connection back to the algebraic theory, the spinor norm map, and spinor exceptions. The key observation which makes \eqref{eqn:spnquadratic} useful is that the left-hand side is a weighted average of modular forms all of whose coefficients are non-negative. Hence if the $n$-th coefficient of this sum is zero, then the $n$-th coefficient of each summand must also be zero, and these coefficients count the number of representations of $n$. On the other hand, the functions appearing on the right-hand side of \eqref{eqn:spnquadratic} are special types of modular forms whose Fourier coefficients may be explicitly computed.
After rewriting the question about representations by $P_m$ as a question about representations by a particular lattice coset $L_{(m)}+\nu_{(m)}$ (defined in \eqref{eqn:Ldef} and \eqref{eqn:nudef}), one would expect such a theory to hold in our case as well. Indeed, the Siegel--Weil formula for the genus of every lattice coset $L+\nu$ was proven by van der Blij \cite{vanderBlij} and then later independently by Shimura \cite{Shimura2004}, who showed that
\[
\Theta_{\operatorname{gen}(L+\nu)}=\mathcal{E}_{\operatorname{gen}(L+\nu)}:=\frac{1}{\sum_{M+\nu'\in \mathbb{G}}%\operatorname{gen}(L+\nu)} \omega_{M+\nu'}^{-1}}\sum_{M+\nu'\in \mathbb{G}}%\operatorname{gen}(L+\nu)} \frac{\Theta_{M+\nu'}}{\omega_{M+\nu'}}
\]
is an Eisenstein series, where $\omega_{M+\nu'}$ is the number of automorphs of the lattice coset and $\mathbb{G}}%\operatorname{gen}(L+\nu)$ denotes a complete set of representatives of the classes in the genus of $L+\nu$. Kneser further showed in \cite{Kneser2} how this formula for the genus of lattice cosets follows by investigating the Haar measure on the orthogonal group, but we do not take that perspective in this paper. We conjecture that the expected link holds in the same way for spinor genera of lattice cosets.
\begin{conjecture}\label{conj:SiegelWeil}
We have
\[
\Theta_{\operatorname{spn}(L+\nu)}:=\frac{1}{\sum_{M+\nu'\in \mathbb{S}} %\operatorname{spn}(L+\nu)} \omega_{M+\nu'}^{-1}}\sum_{M+\nu'\in \mathbb{S}} %\operatorname{spn}(L+\nu)} \frac{\Theta_{M+\nu'}}{\omega_{M+\nu'}}=\mathcal{E}_{\operatorname{gen}(L+\nu)} + \mathcal{U}_{\operatorname{spn}(L+\nu)},
\]
where $\mathcal{U}_{\operatorname{spn}(L+\nu)}$ is a linear combination of unary theta functions and $\mathbb{S}} %\operatorname{spn}(L+\nu)$ denotes a set of representatives of the classes in the spinor genus of $L+\nu$.
\end{conjecture}
Conjecture \ref{conj:SiegelWeil} is useful in two different ways. Firstly, it shows that the number of representations by the spinor genus is usually the same as the number of representations by the genus, and secondly it is useful for showing that certain integers in the support of the unary theta functions are \underline{not} represented by a given lattice coset. To better understand the utility of Conjecture \ref{conj:SiegelWeil} and to motivate why we believe it to be true, we return to $P_m$. In particular, for $m=14$, a finite calculation yields the following.
\begin{proposition}\label{prop:SiegelWeil}
The theta function $\Theta_{\operatorname{spn}(L_{(14)}+\nu_{(14)})}$ satisfies Conjecture \ref{conj:SiegelWeil}.
\end{proposition}
As stated above, one of the main advantages of Proposition \ref{prop:SiegelWeil} is that one can use it to show that the Fourier coefficients of $\Theta_{\operatorname{spn}(L_{14}+\nu_{14})}$ usually agree with those of
\[
\mathcal{E}_m:=\mathcal{E}_{\operatorname{gen}(L_{(m)}+\nu_{(m)})}.
\]
However, we specifically use Proposition \ref{prop:SiegelWeil} to investigate the coefficients supported by the unary theta functions to prove that infinitely many coefficients of $\Theta_{14}$ in these square classes vanish, where
\[
\Theta_m:=\Theta_{L_{(m)}+\nu_{(m)}}.
\]
We then build off of this to use Proposition \ref{prop:SiegelWeil} to prove that $P_{m}$ is not almost universal for every $m\equiv 2\pmod{12}$.
\begin{theorem}\label{thm:P14}
For every $m\equiv 2\pmod{12}$, the form $P_{m}$ is not almost universal.
\end{theorem}
\begin{remarks}
\noindent
\noindent
\begin{enumerate}[leftmargin=*,label={\rm(\arabic*)}]
\item
For any given lattice coset $L+\nu$, one can check Conjecture \ref{conj:SiegelWeil} with a (possibly long) finite calculation. To show that Conjecture \ref{conj:SiegelWeil} is true for all lattice cosets, one would need to develop the algebraic theory further to determine spinor exceptions (resp. primitive spinor exceptions) for lattice cosets, proving a theorem analogous to Schulze-Pillot's results in \cite{SP80} (resp. Earnest, Hsia, and Hung's results in \cite{EH94}).
\item
It is natural to ask whether one expects the forms $P_m$ to be almost universal in the case that $m\equiv 2\pmod{3}$ is odd. Guy showed in \cite{Guy} that $P_5$ is not only almost universal, but indeed universal. Computer calculations indicate that $P_{11}$ is also almost universal. In order to prove that any given $P_m$ in this family is almost universal, it suffices to decompose the associated theta function into an Eisenstein series, a linear combination of unary theta functions, and a cusp form which is orthogonal to unary theta functions. If the contribution from unary theta functions is trivial, then form will be almost universal. Following Conjecture \ref{conj:SiegelWeil}, one expects the unary theta function contribution to directly appear from the theta function associated to the spinor genus.
\end{enumerate}
\end{remarks}
The paper is organized as follows. We first give some preliminary definitions and known results in Section \ref{sec:prelim}. In Section \ref{sec:algebraic}, we use algebraic methods to establish the local behavior of $P_m$. In Section \ref{sec:analytic}, we give a proof Theorem \ref{thm:main} using analytic methods. We then finally blend the two approaches together in Section \ref{sec:link} in order to prove Proposition \ref{prop:SiegelWeil} and Theorem \ref{thm:P14}.
\section*{Acknowledgements}
The authors thank Wai Kiu Chan for helpful discussion and Rainer Schulze-Pillot for pointing out the work of van der Blij \cite{vanderBlij} and Kneser \cite{Kneser2} related to the Siegel--Weil formula for shifted lattices after seeing a preliminary version of this paper. The authors also thank the anonymous referee for many corrections and also suggestions that improved the exposition of the paper.
\section{Preliminaries}\label{sec:prelim}
In this section, we introduce the necessary objects used in the algebraic proofs.
\subsection{Setup for the algebraic approach: Lattice theory}\label{sec:prelimlattice}
For the algebraic approach, we adopt the language of quadratic spaces and lattices as set forth in \cite{OM}. If $L$ is a lattice and $A$ is the Gram matrix for $L$ with respect to some basis, we write $L\cong A$. When $A$ is a diagonal matrix with entries $a_1,...,a_n$ on the diagonal, then $A$ is written as $\langle a_1,...,a_n\rangle$. For a lattice $L$, we let $V$ denote the underlying quadratic space; that is, $V=\mathbb{Q} L$. In this case, we say that $L$ is a lattice on the quadratic space $V$. For a lattice $L$ we define the localization of $L$ by $L_p=L\otimes_\mathbb{Z}\Z_p$, where now $L_p$ is a $\mathbb{Z}_p$-lattice on $V_p:=V\otimes_{\mathbb{Q}} \mathbb{Q}_p$.
Given a lattice $L$ and a vector $\nu\in V$, we have the lattice coset $L+\nu$. If we define a lattice $M=L+\mathbb{Z}\nu$, then $L+\nu$ can be regarded as coset inside the lattice quotient $M/L$. Elements in $L+\nu$ are simply vectors of the form $\nu+x$, where $x\in L$.
We are considering representations of an integer $n$ by the sum $P_m$, which upon completing the square, is seen to be equivalent to the condition that
\[
\ell_n:=\begin{cases}
3\left(\frac{(m-4)}{2}\right)^2+2(m-2)n & \text{ for $m$ even}\\
3\left(m-4\right)^2+8(m-2)n & \text{ for $m$ odd}
\end{cases}
\]
is represented by the lattice coset $L+\nu$, where $L=L_{(m)}$ is defined as the $\mathbb{Z}$-lattice
\begin{equation}\label{eqn:Ldef}
\begin{cases}
\langle(m-2)^2,(m-2)^2,(m-2)^2\rangle & \text{ for $m$ even}\\
\lan4(m-2)^2,4(m-2)^2,4(m-2)^2\rangle & \text{ for $m$ odd},\\
\end{cases}
\end{equation}
in the orthogonal basis $\{e_1,e_2,e_3\}$, and
\begin{equation}\label{eqn:nudef}
\nu=\nu_{(m)}:=
\frac{(m-4)}{2(m-2)}\!\left(e_1+e_2+e_3\right).
\end{equation}
To prove Theorem \ref{thm:main}, we need to show that all but finitely many $\ell_n$ are represented by the lattice coset $L+\nu$.
In order to approach this problem from the algebraic side, we need to develop some algebraic notion of the class, spinor genus, and genus of a lattice coset. Following the definitions that originally appear in \cite{CO13} the {\em class} of $L+\nu$ is defined as
\begin{equation}\label{eqn:clscosetdef}
\operatorname{cls}(L+\nu):=\text{the orbit of $L+\nu$ under the action of $SO(V)$,}
\end{equation}
the {\em spinor genus} of $L+\nu$ as
\begin{equation}\label{eqn:spncosetdef}
\operatorname{spn}(L+\nu):=\text{the orbit of $L+\nu$ under the action of $SO(V)O'_\mathbb{A}(V)$,}
\end{equation}
and the {\em genus} of the lattice coset $L+\nu$ by
\begin{equation}\label{eqn:gencosetdef}
\operatorname{gen}(L+\nu):=\text{ the orbit of $L+\nu$ under the action of $SO_\mathbb{A}(V)$,}
\end{equation}
and where $O'_\mathbb{A}(V)$ denotes the adeles of the kernel of the spinor norm map, $\theta:SO(V)\rightarrow \mathbb Q^\times/{\mathbb Q^\times}^2$ as defined in \cite[\S 55]{OM}. Note that what we refer to as the genus (resp. spinor genus or class) above is often called the \begin{it}proper genus\end{it} (resp. proper spinor genus or proper class), and is commonly denoted with a superscript $+$; e.g., the proper genus is written $\operatorname{gen}^+(L+\nu)$, while the (non-proper) genus (resp. spinor genus and class) are usually defined with the corresponding special orthogonal groups (e.g., $SO_{\mathbb{A}}(V)$) replaced by the orthogonal groups (e.g. $O_{\mathbb{A}}(V)$). Although the genus and proper genus are always equal in the cases of lattices (see \cite[\S102 A]{OM}) this is not always true for lattice cosets. In particular, if $O(L_p+\nu)$ does not contain an improper isometry (an element of the orthogonal group with determinant $-1$) at some finite prime $p$, then $\text{gen}^+(L+\nu)\subsetneq\text{gen}(L+\nu)$. For an example of this phenomenon, we direct the reader to \cite[Example 4.5]{CO13}. In our case we are guaranteed that $O(L_p+\nu)$ contains a symmetry at every prime $p$. This is obvious at primes $p$ not dividing $2(m-2)$ since in this case $L_p+\nu=L_p$ is just a diagonal lattice. At other primes we can take the symmetry $\tau_{e_1-e_2}$, which switches the basis elements $e_1$ and $e_2$ while fixing $e_3$. Therefore the $\text{gen}(L+\nu)=\text{gen}^+(L+\nu)$. Since it will not make a difference in this setting, we choose to define the class, genus, and spinor genus above in terms of the special orthogonal group so that our notation matches that given in \cite{Xu}, which will be helpful to us in what follows. We let $\mathbb{G}}%\operatorname{gen}(L+\nu)$ (resp. $\mathbb{S}} %\operatorname{spn}(L+\nu)$) denote a set of representatives of the classes in $\operatorname{gen}(L+\nu)$ (resp. $\operatorname{spn}(L+\nu)$). For any further unexplained notation, the reader is directed to \cite{OM}.
The general strategy will be to show first that there are no local obstructions, i.e. that $Q(\nu)+2(m-2)n$ is represented by the $\operatorname{gen}(L+\nu)$. Next we will determine conditions under which the spinor genus and genus coincide. An essential ingredient here will be to count the number of spinor genera in the genus of a lattice coset. For this we turn to a formula given by Xu in \cite{Xu}, counting the number of spinor genera in $\operatorname{gen}(L+\nu)$,
\begin{eqnarray}\label{spincount}
[J_\mathbb{Q}:\mathbb{Q}^\times\prod_{p\in \Omega}\theta(SO(L_p+\nu))]
\end{eqnarray}
where $J_\mathbb{Q}$ is the set of ideles of $\mathbb{Q}$ and $\Omega$ is the set of primes in $\mathbb{Q}$ and $SO(L_p+\nu)$ is the stabilizer of $L_p+\nu$ in $SO(V_p)$. One easily checks that (cf. \cite{CO13})
\begin{equation}\label{orthog}
SO(L_p+\nu)=\{\sigma\in SO(V_p):\sigma(L_p)=L_p \text{ and }\sigma(\nu)\equiv \nu\mod L_p\}.
\end{equation}
In Theorem \ref{prop:spngen} we will explicitly compute the image of $SO(L_p+\nu)$ under the spinor norm map and count the number of spinor genera.
\subsection{Setup for the analytic approach: Modular forms theory}
We require some results about (classical holomorphic) modular forms.
\subsubsection{Basic definitions}
Let $\H$ denote the \begin{it}upper half-plane\end{it}, i.e., those $\tau=u+iv\in \mathbb{C}$ with $u\in\mathbb{R}$ and $v>0$. The matrices $\gamma=\left(\begin{smallmatrix} a&b\\ c&d\end{smallmatrix}\right)\in\operatorname{SL}_2(\mathbb{Z})$ (the space of two-by-two integral matrices with determinant $1$) act on $\H$ via \begin{it}fractional linear transformations\end{it} $\gamma \tau:=\frac{a\tau+b}{c\tau+d}$. For
\[
j(\gamma,\tau):=c\tau+d,
\]
a \begin{it}multiplier system\end{it} for a subgroup $\Gamma\subseteq \operatorname{SL}_2(\mathbb{Z})$ and \begin{it}weight\end{it} $r\in \mathbb{R}$ is a function $\nu:\Gamma\mapsto \mathbb{C}$ such that for all $\gamma,M\in\Gamma$ (cf. \cite[(2a.4)]{Pe1})
\[
\nu(M \gamma) j(M\gamma,\tau)^r = \nu(M)j(M,\gamma \tau)^r \nu(\gamma)j(\gamma,\tau)^r.
\]
The \begin{it}slash operator\end{it} $|_{r,\nu}$ of weight $r$ and multiplier system $\nu$ is then
\[
f|_{r,\nu}\gamma (\tau):=\nu(\gamma)^{-1} j(\gamma,\tau)^{-r} f(\gamma \tau).
\]
A \begin{it}(holomorphic) modular form\end{it} of weight $r\in\mathbb{R}$ and multiplier system $\nu$ for $\Gamma$ is a function $f:\H\to\mathbb{C}$ satisfying the following criteria:
\noindent
\noindent
\begin{enumerate}[leftmargin=*]
\item
The function $f$ is holomorphic on $\H$.
\item
For every $\gamma\in\Gamma$, we have
\begin{equation}\label{eqn:modularity}
f|_{r,\nu}\gamma= f.
\end{equation}
\item
The function $f$ is bounded towards every \begin{it}cusp\end{it} (i.e., those elements of $\Gamma\backslash(\mathbb{Q}\cup\{i\infty\})$). This means that at each cusp $\varrho$ of $\Gamma\backslash \H$, the function $f_{\varrho}(\tau):=f|_{r,\nu}\gamma_{\varrho}(\tau)$ is bounded as $v\to \infty$, where $\gamma_{\varrho}\in \operatorname{SL}_2(\mathbb{Z})$ sends $i\infty$ to $\varrho$.
\end{enumerate}
Furthermore, if $f$ vanishes at every cusp (i.e., $\lim_{\tau\to i\infty} f_{\varrho}(\tau)=0$), then we call $f$ a \begin{it}cusp form\end{it}.
\subsubsection{Half-integral weight forms}
We are particularly interested in the case where $r=k+1/2$ with $k\in\mathbb{N}_0$ and
\[
\Gamma=\Gamma_1(M):=\left\{ \left(\begin{matrix}a&b\\ c&d\end{matrix}\right)\in\operatorname{SL}_2(\mathbb{Z}): M\mid c, a\equiv d\equiv 1\pmod{M}\right\}
\]
for some $M\in\mathbb{N}$ divisible by $4$. The multiplier system we are particularly interested in is given in \cite[Proposition 2.1]{Shimura}, although we do not need the explicit form of the multiplier for this paper.
If $T^N\in \Gamma$ with $T:=\left(\begin{smallmatrix} 1&1\\ 0 &1\end{smallmatrix}\right)$, then by \eqref{eqn:modularity} we have $f(\tau+N)=f(\tau)$, and hence $f$ has a Fourier expansion ($a_{f}(n)\in\mathbb{C}$)
\begin{equation}\label{eqn:fexp}
f(\tau)=\sum_{n\geq 0} a_{f}(n) e^{\frac{2\pi i n \tau}{N}}.
\end{equation}
The restriction $n\geq 0$ follows from the fact that $f$ is bounded as $\tau\to i\infty$. One commonly sets $q:=e^{2\pi i \tau}$ and associates the above expansion with the corresponding formal power series, using them interchangeably unless explicit analytic properties of the function $f$ are required.
\subsubsection{Theta functions for quadratic polynomials}
In \cite[(2.0)]{Shimura}, Shimura defined theta functions associated to lattice cosets $L+\nu$ (for a lattice $L$ of rank $n$) and polynomials $P$ on lattice points. Namely, he defined
\[
\Theta_{L+\nu,P}(\tau):=\sum_{\boldsymbol{x}\in L+\nu} P(\boldsymbol{x}) q^{Q(\boldsymbol{x})},
\]
where $Q$ is the quadratic map on the associated quadratic space. We omit $P$ when it is trivial. In this case, we may write $r_{L+\nu}(\ell)$ for the number of elements in $L+\nu$ of norm $\ell$ and we get
\begin{equation}\label{eqn:thetadef}
\Theta_{L+\nu}(\tau)=\sum_{\ell\geq 0} r_{L+\nu}(\ell) q^{\ell}.
\end{equation}
Shimura then showed (see \cite[Proposition 2.1]{Shimura}) that $\Theta_{L+\nu}$ is a modular form of weight $n/2$ for $\Gamma_1(4N^2)$ (for some $N$ which depends on $L$ and $\nu$) and a particular multiplier. Note that we have taken $\tau\mapsto 2N\tau$ in Shimura's definition. To show the modularity properties, for $\gamma=\left(\begin{smallmatrix}a&b\\ c&d\end{smallmatrix}\right)\in \Gamma_1(4N^2)$, we compute
\[
2N\gamma(\tau)=2N\frac{a\tau+b}{c\tau+d} = \frac{a(2N\tau) + 2Nb}{\frac{c}{2N} (2N\tau)+d} = \left(\begin{matrix} a & 2Nb\\ \frac{c}{2N} & d\end{matrix}\right) (2N\tau).
\]
Since $\gamma\in \Gamma_{1}(4N^2)$, we have
\[
\left(\begin{matrix} a & 2Nb\\ \frac{c}{2N} & d\end{matrix}\right)\in \Gamma(2N):=\left\{ \gamma= \left(\begin{matrix}a&b\\ c&d\end{matrix}\right)\in\operatorname{SL}_2(\mathbb{Z}): \gamma\equiv I_{2}\pmod {N}\right\}\subset \Gamma_{1}(2N),
\]
so we may then use \cite[Proposition 2.1]{Shimura}. Specifically, the multiplier is the same multiplier as $\Theta^3$, where $\Theta(\tau):=\sum_{n\in\mathbb{Z}} q^{n^2}$ is the classical Jacobi theta function.
We only require the associated polynomial in one case. Namely, for $n=1$ and $P(x)=x$, we require the \begin{it}unary theta functions\end{it} (see \cite[(2.0)]{Shimura} with $N\mapsto 2N^2/t$, $P(m)=m$, $A=(2N^2/t)$, and $h\mapsto 2Nh$, multiplied by $(2N)^{-1}$)
\begin{equation}\label{eqn:unarydef}
\vartheta_{h,t}(\tau)=\vartheta_{h,t,N}(\tau):=\sum_{\substack{r\in\mathbb{Z}\\ r\equiv h\pmod{\frac{N}{t}}}} r q^{t r^2},
\end{equation}
where $h$ may be chosen modulo $N/t$ and $t$ is a squarefree divisor of $N$. These are weight $3/2$ modular forms on $\Gamma_1(4N^2)$ the same multliplier system as $\Theta_{L+\nu}$.
\section{Algebraic Approach}\label{sec:algebraic}
As seen in Section \ref{sec:prelimlattice}, a natural number $n$ is represented by $P_m(x,y,z)$ if and only if $\ell_n$ is represented by the lattice coset $L+\nu$. In this section, we check for local obstructions; i.e., we check whether $\ell_n$ may be represented by $L_p+\nu$ for every prime $p$.
\begin{lemma}\label{lem:local}
If $m\equiv 0\pmod 4$ then $P_m$ is not almost universal.
\end{lemma}
\begin{proof}
When $m=2p+2$ for an odd prime $p$, then the claim follows immediately from \cite[Theorem 7]{H14}. Otherwise it can be easily verified that if $m\equiv 0\pmod 4$ then $P_m(x,y,z)$ always fails to represent an entire square class modulo 8, and is therefore not almost universal. Specifically, if $m\equiv 4\pmod{8}$, then $P_m(x,y,z)$ does not represent any integer congruent to $-1$ modulo $8$, while if $m\equiv 0\pmod{8}$, then $P_m(x,y,z)$ does not represent any integer congruent to $4$ modulo $8$.
\end{proof}
In order for $P_m$ to be almost universal, a necessary condition is that every integer $\ell_n$ is represented by $\operatorname{gen}(L+\nu)$. Since it will be helpful in much of what follows, we define the ternary lattice $M:=L+\mathbb{Z}\nu$ and note that this lattice has a basis $\{\nu,e_1,e_2\}$. We will also define $T:=\{p\text{ prime }:p\mid (m-2)\}$.
\begin{lemma}\label{lem:Lp}
For any odd prime $p\notin T$, we have $M_p=L_p=L_p+\nu$.
\end{lemma}
\begin{proof}
This follows immediately from the fact that $\nu\in L_p$.
\end{proof}
\begin{lemma}\label{genrep1}
If $m\not\equiv 0\pmod{4}$, then $\ell_n$ is represented by $L_p+\nu$ for every prime $p$.
\end{lemma}
\begin{proof}
For odd $p\notin T$, Lemma \ref{lem:Lp} implies that $L_p+\nu=L_p$ and since $L_p$ is unimodular, it represents every integer in $\mathbb{Z}_p$ (cf. \cite[92:1b]{OM}).
For odd $p\in T$ and $p\neq 3$, $\ell_n$ is a unit in $\mathbb{Z}_p$, since
\[
Q(\nu)=\begin{cases}
3\left(\frac{m-4}{2}\right)^2 & \text{ when $m-2$ is even}\\
3(m-4)^2 & \text{ when $m-2$ is odd}
\end{cases}
\]
is never divisible by $p$. Therefore, since $Q(\nu)$ is represented by $M_p$, it follows from the local square theorem that $\ell_n$ is represented by $M_p$ for every choice of $n$. Suppose that $\ell_n$ is represented by an arbitrary coset $L_p+t\nu$ of $L_p$ in $M_p$, where $t\in \{0,..,p^k-1\}$. Then
\[
Q(\nu)\equiv Q(\omega+t\nu)\equiv t^2Q(\nu)\pmod {p^k}
\]
for $\omega\in L_p$. Consequently, $t=\pm 1$, since the multiplicative group $(\mathbb{Z}/p^k\mathbb{Z})^\times$ contains at most one subgroup of order 2. Therefore $\ell_n$ is represented by the coset $L_p+\nu$.
Finally, when $p=2$, we will proceed by showing that in fact every integer in $\mathbb{Z}_2$ can be written as an $m$-gonal number when $m-2\equiv 0\mod 4$. We may suppose that $\text{ord}_2(m-2)=k+1$ where $k>0$. Therefore, $(m-2)=2^{k+1}\epsilon$ and $(m-4)=2\gamma$ where $\epsilon,\gamma\in \mathbb{Z}_2^\times$. Then an integer $n$ can be written as an $m$-gonal number precisely when there exists $x\in \mathbb{Z}_2$ such that
\begin{equation}\label{T2univ}
n=\frac{(m-2)x^2-(m-4)x}{2}=2^{k}\epsilon x^2-\gamma x.
\end{equation}
The $x$ in \eqref{T2univ} (in the algebraic closure of $\mathbb{Z}_2$) is given by
\begin{equation}\label{solveforx}
x=\frac{\gamma\pm\sqrt{\gamma^2-4(2^k\epsilon)(-n)}}{2^{k+1}\epsilon}
=\frac{1\pm\sqrt{1+2^{k+2}\alpha n}}{2^{k+1}\beta}
\end{equation}
where $\alpha=\epsilon/\gamma^2$ and $\beta=\epsilon/\gamma$. By the local square theorem, we know that $1+2^{k+2}\alpha n$ is the square of a unit in $\mathbb{Z}_2$. Therefore,
\[
1+2^{k+2}\alpha n=(1+2^s \delta)^2=1+2^{s+1}\delta+2^{2s}\delta^2=1+2^{s+1}(\delta+2^{s-1}\delta),
\]
where $s>0$ and $\delta\in \mathbb{Z}_2^\times$, and
\[
x=\frac{1\pm \sqrt{1+2^{k+2}\alpha n}}{2^{k+1}\beta}=\frac{1\pm\sqrt{(1+2^s \delta)^2}}{2^{k+1}\beta}=\frac{1\pm(1+2^{s}\delta)}{2^{k+1}\beta}.
\]
When $s>1$, since $\left| 2^{k+2}\alpha n\right|_2=\left| 2^{s+1}(\delta+2^{s-1}\delta)\right|_2$, it follows that $k+2+r=s+1$ where $r=\text{ord}_2(n)$. Therefore,
\[
x=\frac{1-(1+2^s\delta)}{2^{k+1}\beta}=\frac{2^s\delta}{2^{k+1}\beta}=\frac{2^r\delta}{\beta}\in \mathbb{Z}_2,
\]
since $s=k+1+r$. On the other hand, when $s=1$, then $k+2+r=2+\text{ord}_2(1+\delta)$, and therefore,
\[
x=\frac{1+\sqrt{(1+2 \delta)^2}}{2^{k+1}\beta}=\frac{1+(1+2\delta)}{2^{k+1}\beta}=\frac{2+2\delta}{2^{k+1}\beta}=\frac{1+\delta}{2^{k}\beta}=\frac{2^r\delta}{\beta}\in \mathbb{Z}_2,
\]
since $k+r=\text{ord}_2(1+\delta)$. Therefore, since every $2$-adic integer can be expressed as an $m$-gonal number, it follows that every $\ell_n$ is represented by the coset $L_2+\nu$.
When $m-2$ is odd a similar argument follows, by letting $(m-2)=\epsilon$ and $(m-4)=\gamma$ where $\epsilon,\gamma\in \mathbb{Z}_2^\times$ and then simply replacing equation (\ref{T2univ}) with
\[
2n=(m-2)x^2-(m-4)x= \epsilon x^2-\gamma x.
\]
Hence equation (\ref{solveforx}) becomes
\[
x=\frac{\gamma\pm \sqrt{\gamma^2-4(\epsilon)(-2n)}}{2\epsilon}=\frac{\gamma\pm \sqrt{\gamma^2+8\alpha n}}{2\beta}
\]
where$\alpha =\epsilon/\gamma^2$ and $\beta=\epsilon/\gamma$, and the result follows as above.
\end{proof}
Having established the local conditions, we next calculate the number of spinor genera for $L+\nu$. Recall from (\ref{spincount}), the number of spinor genera in the genus of the coset is given by
\begin{eqnarray*}
[J_\mathbb{Q}:\mathbb{Q}^\times\prod_{p\in \Omega}\theta(SO(L_p+\nu))]
\end{eqnarray*}
where $J_\mathbb{Q}$ is the set of ideles of $\mathbb{Q}$ and $\Omega$ is the set of primes in $\mathbb{Q}$. From this formula, we see that much like in the case of lattices, $\mathbb{Z}_p^\times\subseteq \theta(SO(L_p+\nu))$ for every prime $p$ is sufficient, though certainly not necessary to guarantee that $\operatorname{gen}(L+\nu)$ and $\operatorname{spn}(L+\nu)$ coincide.
\begin{proposition}\label{prop:spngen}
\noindent
\begin{enumerate}
\item If $m\equiv 2\pmod 4$ and $m\not\equiv 2\pmod {12}$, then $\operatorname{spn}(L+\nu)=\operatorname{gen}(L+\nu)$.
\item For $m\equiv 2\mod 12$, there are two spinor genera in the genus of $L+\nu$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) For primes $p\not\in T$, it is immediate that $\mathbb{Z}_p^\times\subseteq \theta(SO(L_p+\nu))$ since $L_p+\nu=L_p\cong\langle 1,1,1\rangle$. For primes $p\in T$, we have $(m-2)=p^k\epsilon$ and $\frac{m-4}{2}=\gamma$ where $k\geq 1$ and $\epsilon,\gamma\in \mathbb{Z}_p^\times$. Then, in the basis $\{\nu,e_1,e_2\}$ we have
\[
M_p\cong\begin{bmatrix}
3\gamma^2 & p^k\epsilon\gamma & p^k\epsilon\gamma \\
p^k\epsilon\gamma & p^{2k}\epsilon^2 & 0\\
p^k\epsilon\gamma & 0 & p^{2k}\epsilon^2,
\end{bmatrix}
\]
and by a change of basis to $\{\nu,p^k\epsilon\nu-3\gamma e_1,p^k\epsilon\nu-3\gamma e_2\}$ we obtain
\[
M_p\cong \lan3\gamma^2\rangle\perp 3p^{2k}\epsilon^2\gamma^2\begin{bmatrix}
6 & -3\\-3 & 6
\end{bmatrix}.
\]
From this we clearly see that $\mathbb{Z}_p[\nu]$ splits $M_p$ as an orthogonal summand; in other words, $M_p$ is the orthogonal sum $M_p =\mathbb{Z}_p[\nu]\perp K_p$, where
\[
K_p\cong 9p^{2k}\epsilon^2\gamma^2\begin{bmatrix}
2 & -1\\-1 & 2
\end{bmatrix}
\]
is a binary modular lattice. From here it follows immediately from \cite[Satz 3]{K56} that $\mathbb{Z}_p^\times\subseteq \theta(SO(K_p))$ for odd prime $p$. When $p=2$ the result follows from \cite[Lemma 1]{H75}. On the other hand, setting
\[
\omega:=\frac{p^k\epsilon}{\gamma}\nu=e_1+e_2+e_3
\]
the set of vectors $\{\omega,\omega-3e_1,\omega-3e_2\}$ form a basis for $L_p$, and in this basis we obtain
\[
L_p= \mathbb{Z}_p[\omega]\perp K_p.
\]
Any isometry $SO(K_p)$ can be extended to an isometry $\sigma\in SO(M_p)$ which simultaneously satisfies $\sigma(\nu)=\nu$ and $\sigma(L_p)=L_p$, and therefore $\sigma(L_p+\nu)=L_p+\nu$. Hence $\sigma\in SO(L_p+\nu)$, from which we may conclude that $\theta(SO(K_p))\subseteq \theta(SO(L_p+\nu))$ and hence $\mathbb{Z}_p^\times\subseteq \theta(SO(L_p+\nu))$.
Now for any $\vec{x}=(x_p)\in J_\mathbb{Q}$ we know that $x_p$ is a unit at almost every prime. Therefore, multiplying by a suitable element of $a\in \mathbb{Q}^\times$ we can assume $a\vec{x}=(ax_p)$ is a unit at every prime. Moreover, since for the infinite prime $\theta(L_{\infty}+\nu)=\theta(SO(V_{\infty}))=\mathbb{R}^{\times^2}$, we only need to chose $a$ to have the same sign as $x_\infty$. Chosen in this way, $a\vec{x}$ is an element in the restricted product.
(2) When $m\equiv 2\mod 12$, then $L\cong \langle (m-2)^2,(m-2)^2,(m-2)^2\rangle$ in the basis $\{e_1,e_2,e_3\}$ and $\nu=\frac{m-4}{2(m-2)}[e_1+e_2+e_3]$. For primes away from $T$, we once again know that $L_p+\nu=L_p\cong\langle 1,1,1\rangle$, and hence $\mathbb{Z}_p^\times\subseteq \theta(SO(L_p+\nu))$. Moreover, for primes $p\neq 3$ in $T$, the argument from above is still sufficient to show that $\mathbb{Z}_p^\times\subseteq \theta(SO(L_p+\nu))$. When $p=2$ we make one further observation, namely that in this case $\mathbb{Z}_2^\times= \theta(SO(L_2+\nu))$. If $\sigma\in SO(L_2+\nu)$ then we know $\sigma(\nu)=\nu+x$ for $x\in L_2$, hence
\[
Q(\nu)=Q(\sigma(\nu))=Q(\nu+x)=Q(\nu)+Q(x)+2B(\nu,x),
\]
by a simple congruence argument we see that no nontrivial $x$ can satisfy this equality. Therefore the only isometries of $L_2+\nu$ are those fixing $\nu$, and hence are precisely the isometries of $K_2$ described above. In particular, it follows from \cite[Lemma 1]{H75} that $\mathbb{Z}_2^\times{\mathbb{Q}_2^\times}^2=\theta(SO(K_2))=\theta(SO(L_2+\nu))$.
When $p=3$, then we consider the generalized lattice $M/L$, as defined in \cite{T97}, which has the orthogonal group
\[
O(M_3/L_3)=\{\sigma\in O(V_3):\sigma(x)\in x+L_3 \text{ for all }x\in M_3\},
\]
also defined in \cite{T97}. An isometry $\sigma$ is in $O(M_3/L_3)$ precisely when $\sigma(L_3)=L_3$ and $\sigma(\nu)\equiv \nu\mod L_3$. Therefore, from (\ref{orthog}), we see that $O(M_3/L_3)=O(L_3+\nu)$ and hence $SO(M_3/L_3)=SO(L_3+\nu)$. However, from \cite[Theorem 2]{T97} we know that $\theta(SO(M_3/L_3))$, and hence $\theta(SO(L_3+\nu))$ is generated by pairs of symmetries coming from $O(M_3/L_3)$. If $\tau$ is a symmetry in $O(M_3/L_3)$, then there is some $\omega=e_1x_1+e_2x_2+e_3x_3\in L_3$ such that
\[
\tau(y)=\tau_\omega(y)=y-\frac{2B(\omega,y)}{Q(\omega)}\omega
\]
for every $y\in L_3$. We may assume that $x_1,x_2,x_3\in \mathbb{Z}_3$, and without loss of generality, that $x_1\in \mathbb{Z}_3^\times$. But now
\[
\tau_\omega(e_1)=e_1-\frac{2B(\omega,e_1)}{Q(\omega)}\omega=e_1-\frac{2\cdot (m-2)^2}{(m-2)^2(x_1^2+x_2^2+x_3^2)}\omega=e_1-\frac{2}{(x_1^2+x_2^2+x_3^2)}\omega\in L_3
\]
and hence $x^2+y^2+z^2\not \equiv 0\mod 3$. This means that at least one of $x_2$ and $x_3$ is not a unit, without loss of generality, say $x_3\not\in \mathbb{Z}_3^\times$. On the other hand,
\[
\tau_\omega(\nu)=\nu-\frac{2B(\omega,\nu)}{Q(\omega)}\omega=\nu-\frac{(m-4)(m-2) (x_1+x_2+x_3)}{(m-2)^2(x_1^2+x_2^2+x_3^2)}\omega=\nu-\frac{(m-4)(x_1+x_2+x_3)}{(m-2)(x_1^2+x_2^2+x_3^2)}\omega
\]
and since $\tau_\omega(\nu)\equiv\nu\mod L_3$ it must follow that $x_1+x_2+x_3\equiv 0\mod 3$. Therefore the only possibility is that $x_2\in \mathbb{Z}_3^\times$ and $x_1\not\equiv x_2\mod 3$. Therefore, $(m-2)^{-2}Q(\omega)\equiv 2\mod 3$, and consequently $\theta(SO(M_3/L_3))$, and hence $\theta(SO(L_3+\nu))$, contains no nontrivial elements. That is, $2\not \in \theta(SO(L_3+\nu))$.
Finally, we will show that the number of spinor genera in the genus of $L+\nu$, in this case, is equal to 2. In order to show that
\[
\left[J_\mathbb{Q}:\mathbb{Q}^\times\prod_{p\in \Omega}\theta(SO(L_p+\nu))\right]=2,
\]
we prove that the principal idele $1$ and the idele $\iota$, given by
\[
\iota_p:=\begin{cases} 1 & \text{if } p\neq 3,\\ 2 &\text{if }p=3,\end{cases}
\]
are inequivalent and the cosets $[1]$ and $[\iota]$ are a full set of representatives of the quotient space.
For any $\vec{x}=(x_p)\in J_\mathbb{Q}$, we know that $x_p$ is a unit for almost every $p$. Multiplying by a suitable element $a$ in $\mathbb{Q}^\times$ (where $a$ has the same sign as $x_\infty$) if necessary, we may assume that $ax_p$ is a unit at every prime $p$ (including the infinite prime). Since $\mathbb{Z}_p^{\times}\subseteq \theta(SO(L_p+\nu))$ for $p\neq 3$, the coset of $\vec{x}$ is completely determined by the congruence class of $ax_3$. If $ax_3\equiv 1\mod 3$ then $\vec{x}\in [1]$ and if $ax_3\equiv 2\mod 3$ then $\vec{x}\in [\iota]$.
\end{proof}
Although Proposition \ref{prop:spngen} (1) doesn't directly lead to a proof of Theorem \ref{thm:main}, it gives a strong expectation for the results given in Theorem \ref{thm:main}. Namely, there is a result of Duke and Schulze-Pillot \cite{DS-P90} which used the analytic theory to obtain the conclusion in the case of lattices that every sufficiently large integer primitively represented by the spinor genus is also represented by the lattice. Since the lattice and genus coincide by Proposition \ref{prop:spngen} (1), one may expect a result similar to Duke and Schulze-Pillot's to imply Theorem \ref{thm:main}. Since no analogous theorem has yet been developed, we turn to a trick in the analytic theory to prove Theorem \ref{thm:main}.
\section{Analytic approach}\label{sec:analytic}
In this section, we use the analytic proof to show Theorem \ref{thm:main}, Theorem \ref{thm:2mod3} (2), and the first statement of Theorem \ref{thm:2mod3} (1). These are rewritten in the following theorem.
\begin{theorem}\label{thm:m2mod4}
Suppose that $m\not\equiv 0\pmod{4}$. Then we have the following.
\noindent
\begin{enumerate}[leftmargin=*,label={\rm(\arabic*)}]
\item
If $m\not\equiv 2\pmod{3}$, then every sufficiently large $n$ may be represented in the form
$$
n=p_m(x)+p_m(y)+p_m(z)
$$
for some $x,y,z\in\mathbb{Z}$. That is to say, $P_m$ is almost universal.
\item
If $m\equiv 2\pmod{12}$, then every sufficiently large $n\notin \mathcal{S}_{m,3}^{\operatorname{e}}$ may be represented in the form
$$
n=p_m(x)+p_m(y)+p_m(z)
$$
for some $x,y,z\in\mathbb{Z}$.
\item
If $m\equiv 2\pmod{3}$ is odd, then every sufficiently large $n\notin \mathcal{S}_{m,3}^{\operatorname{o}}$ may be represented in the form
$$
n=p_m(x)+p_m(y)+p_m(z)
$$
for some $x,y,z\in\mathbb{Z}$.
\end{enumerate}
\end{theorem}
\begin{proof}
We split the proof into four pieces. First we assume that $m\equiv 2\pmod{4}$ and then split depending on the congruence class of $m$ modulo $3$, and we will later assume that $m$ is odd.
By completing the square, a solution to the given representation is equivalent to a solution to
$$
2(m-2)n+3\!\left(\frac{m-4}{2}\right)^2= \!\left((m-2)x-\frac{m-4}{2}\right)^2+\!\left((m-2)y-\frac{m-4}{2}\right)^2+\!\left((m-2)z-\frac{m-4}{2}\right)^2.
$$
We set $N:=(m-2)$ and $\ell=\ell_n:=2(m-2)n+3\left(\frac{m-4}{2}\right)^2$. Denoting by $r_{m}(\ell)$ the number of such solutions (with $r_m(\ell):=0$ if $\ell$ is not in the correct congruence class), we hence consider the generating function
\begin{equation}
\Theta_m(\tau):=\sum_{n\geq 0} r_{m}\!\left(\ell_n\right) q^{\ell_n},
\end{equation}
where $q:=e^{2\pi i \tau}$. The function $\Theta_m$ is a theta series for a lattice coset. Since the gram matrix $A=N^2\cdot I_3$ associated to the lattice is diagonal with even entries, \cite[Proposition 2.1]{Shimura} (with $P(m)=1$, $\tau\mapsto 2N\tau$, and $h=((m-4)/2,(m-4)/2,(m-4)/2)^T$) implies that $\Theta_m$ is a weight $3/2$ modular form on $\Gamma_1(4N^2)$ with some multiplier. As usual, we decompose
\begin{equation}\label{eqn:decompose}
\Theta_m(\tau)=\mathcal{E}_m(\tau)+\mathcal{U}_m(\tau)+f_m(\tau),
\end{equation}
where $\mathcal{E}_m$ is in the space spanned by Eisenstein series, $\mathcal{U}_m$ is in the space spanned by unary theta functions, and $f_m$ is a cusp form which is orthogonal to unary theta functions. Of course, each term in the decomposition is modular of weight $3/2$ on $\Gamma_1(4N^2)$ with the same multiplier.
We now follow an argument of Duke and Schulze-Pillot \cite{DS-P90}, who proved that sufficiently large integers are primitively represented by quadratic forms if and only if they are primitively represented by the spinor genus of the quadratic form (i.e., they investigated the coefficients of theta series with $\Theta_m$ replaced with the theta series for a lattice). By work of Duke \cite{Duke}, the Fourier coefficients of $f_m$ grow at most like $\ell^{3/7+\varepsilon}$, while the coefficients of $\mathcal{E}_m$ are certain class numbers and by Siegel's (ineffective) bound \cite{Siegel} for class numbers, they grow like $\gg \ell^{1/2-\varepsilon}$ for any $\ell$ supported on the coefficients of $\mathcal{E}_m$ and which are primitively represented by the genus. The requirement that the representations are primitive comes from the fact that there are certain primes $p$ for which the $p^r\ell$-th coefficients of Eisenstein series do not grow as a function of $r$; this phenomenon is explained on the algebraic side in the case of lattices by realizing $p$ as an anisotropic prime. When the power of such primes is bounded, the coefficients of $\mathcal{E}_{m}$ grow faster than the coefficients of $f_m$. We may hence disregard $f_m$ completely whenever the $\ell$-th coefficient is not supported in $\mathcal{U}_m$ and the power of bad primes $p$ dividing $\ell$ is bounded. However, Shimura \cite{Shimura2004} used the Siegel--Weil formula for inhomogeneous quadratic forms (i.e., quadratic polynomials) to show that $\mathcal{E}_m$ is the weighted average of representations by members of the genus of the lattice coset and simultaneously the product of the local densities. For $p\mid m-2$ with $p\neq 3$, the power of $p$ dividing $\ell$ for $p\mid m-2$ is bounded from above by the congruence conditions. Similarly, if $\text{ord}_3(m-2)> 1$, then $\text{ord}_3(\ell)$ is bounded. In the special case that $\text{ord}_3(m-2)=1$, we note that $X^2+Y^2+Z^2\equiv 0\pmod{3}$ implies that either $3$ divides each of $X$, $Y$, and $Z$ or none of them. Hence, in this case, the local density at $3$ for representations of $\ell$ by the lattice coset equals the local density at $3$ corresponding to the number of primitive representations by the lattice. Since $3$ is not an anisotropic prime for the lattice $\langle 1,1,1\rangle$, we conclude that the local densities grow as expected. Since $L_p+\nu=L_p$ for $p \nmid m-2$ by Lemma \ref{lem:Lp}, the coefficients of $\mathcal{E}_m$ grow $\gg \ell^{\frac{1}{2}-\varepsilon}$ whenever they are represented (due to the fact that $\ell$ is primitively represented by the lattice). Hence the congruence conditions for $\mathcal{E}_m$ are equivalent to checking that the integer is represented locally, or in other words that the genus represents the given integer, which follows immediately from Lemma \ref{genrep1}.
We claim furthermore that $\mathcal{U}_m$ is identically zero, from which the claim will follow. We may decompose $\mathcal{U}_m(\tau)$ into a linear combination of finitely many unary theta functions (defined in \eqref{eqn:unarydef}). The goal now is to determine the possible $\vartheta_{h,t}$ in the decomposition with non-zero coefficient. We do so by restricting the possible choices of $t$ with congruence conditions on $tr^2$ implied by the definition of $\Theta_m$. The $\ell$ upon which the coefficients of $q^{\ell}$ in $\vartheta_{h,t}$ are supported must satisfy
$$
\ell =tr^2 \equiv 0\pmod{t}.
$$
However, the coefficients of the unary theta function are supported on the same integers as the original theta series $\Theta_m$, since the Eisenstein series are also supported on these coefficients and there would hence otherwise be integers $\ell$ upon which the coefficient of $\Theta_m$ is negative, contradicting the fact that it is a generating function for the non-negative integers $r_{m}(\ell)$. Therefore, we conclude that
$$
2(m-2)n+3\left(\frac{m-4}{2}\right)^2=\ell\equiv 0\pmod{t}.
$$
Since $2(m-2)$ is even and $3\left((m-4)/2\right)^2$ is odd, we conclude that $t$ must be odd. The congruence then becomes
$$
3\left(\frac{m-4}{2}\right)^2\equiv 0\pmod{t},
$$
where $t$ is some divisor of (the odd part of) $m-2$. Rewritten, this implies that
$$
t\; \Big| \left(m-2,3\left(\frac{m-4}{2}\right)^2\right).
$$
Now note that if $p\mid m-2$ and $p\mid m-4$, then
$$
p\mid m-2-(m-4)=2\implies p=2.
$$
Thus, since $(m-4)/2$ is odd,
$$
\left(m-2,3\left(\frac{m-4}{2}\right)^2\right)=(m-2,3).
$$
We now split the proof into two cases to prove (1) and (2).
To prove (1) for $m$ even, we assume that $m\not\equiv 2\pmod{3}$, so
$$
t\mid (m-2,3)=1.
$$
We conclude that $t=1$. However, since $m-2$ is even and $(m-4)/2$ is odd, we also have
$$
\ell =\left((m-2)x+\frac{m-4}{2}\right)^2+\left((m-2)y+\frac{m-4}{2}\right)^2+\left((m-2)z+\frac{m-4}{2}\right)^2\equiv 3\pmod{8}.
$$
In particular, $\ell=tr^2\equiv 3\pmod{8}$ implies that $t=1$ is impossible. Since there are no possible choices of $t$, we conclude that $\mathcal{U}_m=0$. This gives the first claim in the case $m\equiv 2\pmod{4}$.
To prove (2), we assume that $m\equiv 2\pmod{3}$ (i.e., $m\equiv 2\pmod{12}$), so that
$$
t\mid (m-2,3)=3
$$
implies that $t=1$ or $t=3$. The case $t=1$ is again impossible by the congruence condition modulo $8$ considered in part (1). It follows that $\mathcal{U}_m$ is a linear combination of forms all of which have $t=3$. Hence every $n$ suffiently large for which the corresponding $\ell$ is not of the form $3r^2$ must be represented as the sum of three $m$-gonal numbers.
We now consider the case $m$ odd. In this case,
\[
n\!=\!p_m(x)+p_m(y)+p_m(z)
\]
is equivalent to
\begin{multline}\label{eqn:modd}
8n(m-2)+3(m-4)^2 = \left(2(m-2)x+(m-4)\right)^2 + \left(2(m-2)y+(m-4)\right)^2 +\\
+\left(2(m-2)z+(m-4)\right)^2.
\end{multline}
Letting $R_{m}(\ell)$ be the number of solutions to \eqref{eqn:modd} with $\ell=\ell_n:=8n(m-2)+3(m-4)^2$, by \cite[Proposition 2.1]{Shimura} we see that
\[
\Theta_m'(\tau):=\sum_{n\geq 0} R_{m}\!\left(\ell_n\right) q^{\frac{\ell_n}{4N}}
\]
is a weight $3/2$ modular form on $\Gamma_1(4N^2)$ with some multiplier. Here $N=(m-2)$, as in the $m\equiv 2\pmod{4}$ case above. Firstly, Lemma \ref{genrep1} implies that $\ell_n$ is represented locally, or equivalently, by the genus of the lattice coset. We conclude that the relevant coefficients of the Eisenstein series $\mathcal{E}_m'$ in the decomposition \eqref{eqn:decompose} are positive and it remains to again determine the unary theta functions $\vartheta_{h,t,2N}$ which may occur in the decomposition \eqref{eqn:decompose}. Arguing as before, we have $\ell \equiv 0\pmod{t}$ and $t\mid 2N$. However, since $m$ is odd, $m-4$ is also odd, so $\ell\equiv 3\pmod{8}$. It follows that $t$ is odd, and hence $t\mid N$. We conclude that
\[
t\mid \left(m-2, 3(m-4)^2\right).
\]
Since $(m-2,m-4)=1$, we conclude that $t\mid (m-2,3)$.
Now we complete the proof of (1) when $m$ is odd. If $m\not\equiv 2\pmod{3}$, then necessarily $t=1$. However, $\ell\equiv 3\pmod{8}$ implies that $\ell$ is not a square, and hence $t=1$ is impossible.
To prove (3), we assume $m$ is odd and $m\equiv 2\mod 3$. Then, since $t$ is a divisor of $1$ or $3$, we conclude that $t=1$ or $t=3$. However, $t=1$ is again impossible because $\ell_n\equiv 3\pmod{8}$ (as defined before the definitions of $\Theta_m$ and $\Theta_m'$ above) for every $n\in\mathbb{N}_0$. Therefore we have $t=3$ and the only possible exceptions are in the square class $3\mathbb{Z}^2$.
\end{proof}
\section{Linking the analytic and algebraic theories and forms which are not almost universal}\label{sec:link}
In this section, we prove Proposition \ref{prop:SiegelWeil} and Theorem \ref{thm:P14}, establishing that $P_{m}$ is not almost universal for all $m\equiv 2\pmod{12}$. We draw on intuition from Proposition \ref{prop:spngen} to both motivate the proof of Theorem \ref{thm:P14} and explain the statement. For $m\equiv 2\pmod{4}$ but $m\not\equiv 2\pmod{12}$, Proposition \ref{prop:spngen} (1) implies that there is only one spinor genus in the genus, and we proved in Theorem \ref{thm:main} that $P_m$ is indeed almost universal in this case. On the other hand, for $m\equiv 2\pmod{12}$, Proposition \ref{prop:spngen} (2) implies that there are two spinor genera. It is hence natural to search for ``primitive spinor exceptions'' for the lattice coset by studying whether there are families of exceptions in certain square classes; from the point of view of modular forms, we are searching for the component of the cuspidal part coming from unary theta functions, and Proposition \ref{prop:SiegelWeil} gives us a way to discover a unary theta function.
In order to prove Proposition \ref{prop:SiegelWeil}, we explicitly determine the genus and spinor genus of a lattice coset. Let $L+\nu$ be the lattice coset associated to $P_{14}$ as in Section \ref{sec:algebraic}. That is to say, $L:=\langle 12^2,12^2,12^2\rangle$ and $\nu:=
\frac{5}{12}(e_1+e_2+e_3)$.
\begin{lemma}\label{lem:genspnm=14}
\noindent
\noindent
\begin{enumerate}[leftmargin=*,label={\rm(\arabic*)}]
\item
Defining
\[
\mu:=\frac{1}{12}\!\left(5e_1+e_2+e_3\right),
\]
the classes in the genus of $L+\nu$ are then represented by $L+\nu$, $L+5\nu$, $L+\mu$, and $L+5\mu$.
\item
The cosets $L+\nu$ and $L+\mu$ form one spinor genus and the cosets $L+5\nu$ and $L+5\mu$ form the other spinor genus.
\end{enumerate}
\end{lemma}
\begin{proof}
\noindent
\noindent
(1) Suppose that $M+\mu'\in \mathbb{G}}%\operatorname{gen}(L+\nu)$. The conductor, $c$, as defined in \cite{H14}, is the smallest positive integer for which $c\nu\in L$, or equivalently,
\[
c=\prod_p[L_p+\mathbb{Z}_p[\nu]:L_p]=12.
\]
Since any local isometry from $L_p+\nu$ to $M_p+\mu'$ must send $M_p$ to $L_p$, the conductor is an invariant of the genus and hence $c$ is also the minimal positive integer for which $c\mu'\in M$. Moreover, since $M_p\cong L_p$ at every prime $p$ it must follow that $M\cong L$ since $L$ has class number 1. Therefore each class in the genus of $L+\nu$ contains a coset of the form $L+\mu'$, and it just remains to determine the possible values of $\mu'$. From here, given that the conductor of the genus is $c=12$,
enumerating the possibilities yields a finite set of possibilities for the class representatives in the genus. Many of these classes are immediately seen not to be in the same genus as $L+\nu$ simply by comparing the numbers locally represented by these classes. We further restrict the set by explicitly finding elements of $SO(V)$ between different cosets. From this, one concludes that the representatives for the classes are a subset of the four claimed lattice cosets. Furthermore, one easily checks that the theta functions of the four cosets are different (i.e., they each represent integers a different number of times), from which one concludes that they cannot be equivalent under the action of $SO(V)$. Finally, we construct $\sigma\in SO_{\mathbb{A}}(V)$ mapping each of the cosets to each other.
For $p\neq 2,3$, we have $\nu,\mu\in L_p$ and hence
\begin{align*}
L_p+5\nu&=L_p=L_p+\nu,\\
L_p+5\mu&=L_p=L_p+\mu,
\end{align*}
so the identity map suffices in this case. When $p=2$, we observe that $4\nu=\frac{5}{3}(e_1+e_2+e_3)\in L_2$ and $4\mu=\frac{1}{3}(5e_1+e_2+e_3)\in L_2$, so that
\begin{align*}
\nu&=-4\nu+5\nu\in L_2+5\nu,\\
\mu&= -4\mu+ 5\mu\in L_2+5\mu,
\end{align*}
and $\frac{5}{3}e_1\in L_2$ implies that
\[
\nu=\frac{5}{12}(e_1+e_2+e_3)=5\mu-\frac{5}{3}(e_1)\in L_2+5\mu,
\]
implying $L_2+5\mu = L_2+\mu = L_2+\nu=L_2+5\nu$. When $p=3$, then we consider the symmetries $\tau_{e_i}$ of $L_3$ which negate the vector $e_i$, and the symmetry $\tau_{e_2-e_3}$ which switch $e_2$ and $e_3$ and fix everything else. Then
\[
\tau_{e_2}\circ \tau_{e_3}(\nu)=\frac{5}{12}(e_1-e_2-e_3)=\mu-\frac{1}{2}(e_2+e_3)
\]
so $L_3+\nu\cong L_3+\mu$ and the same isometry can be used to show that $L_3+5\nu\cong L_3+5\mu$.
Similarly,
\[
\tau_{e_1}\circ \tau_{e_2-e_3}(\mu)=\tau_{e_1}(\mu)=\frac{1}{12}(-5e_1+e_2+e_3)=5\nu-\frac{5}{2}e_1-2e_2-2e_3,
\]
and hence $L_3+\mu$ is isometric to $L_3+5\nu$. From here we conclude that $L+\nu$, $L+5\nu$, $L+\mu$ and $L+5\mu$ are in the same genus.
\noindent
\vspace{0.05in}
\noindent
(2) By Proposition \ref{prop:spngen} (2), there are precisely 2 spinor genera in the genus of $L+\nu$. Now it only remains to find representatives for the classes in the two spinor genera. To do this, we need only find a map $\sigma=(\sigma_2,\sigma_3,...,\sigma_p,...)\in O'_\mathbb{A}(V)$ for which $\sigma_p(L_p+\nu)=L_p+\mu$ at every prime $p$. For primes away from $2$ and $3$ we have
\[
L_p+\mu=L_p=L_p+\nu,
\]
and so we can let $\sigma_p$ be the identity map for $p\neq 2,3$. Moreover, when $p=2$, then
\[
\nu=\mu+\frac{1}{3}e_2+\frac{1}{3}e_3
\]
so in fact $L_2+\nu=L_2+\mu$, and hence $\sigma_2$ can also be taken to be the identity map. When $p=3$ we consider the symmetries $\tau_{e_2}$ and $\tau_{e_3}$, then
\[
\tau_{e_2}\circ \tau_{e_3}(\nu)=\frac{5}{12}(e_1-e_2-e_3)=\mu-\frac{1}{2}e_2-\frac{1}{2}e_3,
\]
and therefore we let $\sigma_3=\tau_{e_2}\circ\tau_{e_3}$. Then clearly $\sigma$ is in the kernel of the adelic spinor norm map, since $Q(e_2)=Q(e_3)$, and this map sends $L+\nu$ to $L+\mu$. A similar argument can be used to show that $L+5\nu$ and $L+5\mu$ are representatives for the two classes in the spinor genus of $L+5\nu$.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:SiegelWeil}]
The number of automorphs of either $L+\nu$ or $L+5\nu$ is $6$, while the number of automorphs of either $L+\mu$ or $L+5\mu$ is $2$. Thus we conclude by Lemma \ref{lem:genspnm=14} that
\[
\Theta_{\operatorname{spn}(L+\nu)}= \frac{3}{2}\left(\frac{\Theta_{L+\nu}}{6} + \frac{\Theta_{L+\mu}}{2}\right)
\]
and
\begin{equation}\label{eqn:Thetagenm=14}
\mathcal{E}_{(L+\nu)}=\Theta_{\operatorname{gen}(L+\nu)} = \frac{3}{4}\left( \frac{\Theta_{L+5\nu}}{6} + \frac{\Theta_{L+5\mu}}{2}+\frac{\Theta_{L+\nu}}{6} + \frac{\Theta_{L+\mu}}{2}\right).
\end{equation}
We claim that
\begin{equation}\label{eqn:Thetaspnm=14}
\Theta_{\operatorname{spn}(L+\nu)}(\tau) = \Theta_{\operatorname{gen}(L+\nu)}(\tau) -\frac{1}{8}\vartheta_{1,3,12}(\tau),
\end{equation}
which would imply the claim. Both sides are modular forms of weight $3/2$ on $\Gamma_1(4\cdot 12^2)$ with the usual $\Theta^3$-multiplier.
Recall now that by the valence formula, a modular form of weight $k$ for $\Gamma\subseteq\operatorname{SL}_2(\mathbb{Z})$ with some multiplier is uniquely determined by the first
\[
\frac{k}{12}[\operatorname{SL}_2(\mathbb{Z}):\Gamma]
\]
Fourier coefficients, where $[\operatorname{SL}_2(\mathbb{Z}):\Gamma]$ is the index of $\Gamma$ in $\operatorname{SL}_2(\mathbb{Z})$. Since (cf. \cite[p. 2]{OnoBook})
\[
\left[\operatorname{SL}_2(\mathbb{Z}):\Gamma_1(N)\right]=N^2\prod_{p\mid N}\left(1-\frac{1}{p^2}\right),
\]
we have
\[
\left[\operatorname{SL}_2(\mathbb{Z}):\Gamma_1(24^2)\right]=24^4\left(1-\frac{1}{4}\right)\left(1-\frac{1}{9}\right)= 221184.
\]
Hence we only need to check $\frac{3}{24}\cdot 221184 =27648$ Fourier coefficients to verify \eqref{eqn:Thetaspnm=14}. This is easily verified with a computer by computing the relevant theta series.
\end{proof}
In order to prove Theorem \ref{thm:P14}, we use the $m=14$ case as a springboard from which the other cases follow. In particular, we show the following theorem.
\begin{theorem}\label{thm:XYZmod12}
\noindent
\noindent
\begin{enumerate}[leftmargin=*,label={\rm(\arabic*)}]
\item If $\ell\equiv 1\pmod{12}$ is an odd prime, then
\[
X^2+Y^2+Z^2=3\ell^2
\]
has no solutions in $X,Y,Z\in\mathbb{Z}$ with $X\equiv Y\equiv Z\equiv 5\pmod{12}$.
\item If $\ell\equiv 7\pmod{12}$ is an odd prime, then
\[
X^2+Y^2+Z^2=3\ell^2
\]
has no solutions in $X,Y,Z\in\mathbb{Z}$ with $X\equiv Y\equiv Z\equiv 1\pmod{12}$.
\end{enumerate}
\end{theorem}
\begin{proof}
\noindent
\noindent
(1) As in the proof of Proposition \ref{prop:SiegelWeil} (and as in Section \ref{sec:algebraic}), we let $L+\nu$ be the lattice coset associated to $P_{14}$. The claim is equivalent to the statement that $L+\nu$ does not represent $3\ell^2$ for all $\ell\equiv 1\pmod{12}$. Since Conjecture \ref{conj:SiegelWeil} is true for the spinor genus of $L+\nu$ by Proposition \ref{prop:SiegelWeil} and the Fourier coefficients of each $\Theta_{M+\nu'}$ are non-negative, the $3\ell^2$-th coefficient of the theta function $\Theta_{\operatorname{spn}(L+\nu)}$ is zero if and only if the $3\ell^2$-th coefficient of $\Theta_{M+\nu'}$ is zero for all $M+\nu'\in \mathbb{S}} %\operatorname{spn}(L+\nu)$. In particular, this implies that if the $3\ell^2$-th coefficient of $\Theta_{\operatorname{spn}(L+\nu)}$ always vanishes, then the claim is true.
We next show that these coefficients of $\Theta_{\operatorname{spn}(L+\nu)}$ do indeed vanish. In order to show this, we explicitly compute the Eisenstein series $\Theta_{\operatorname{gen}(L+\nu)}=\mathcal{E}_{\operatorname{gen}(L+\nu)}$ and the linear combination of unary theta functions $\mathcal{U}_{\operatorname{spn}(L+\nu)}$. By \eqref{eqn:Thetaspnm=14}, we have
\[
\mathcal{U}_{\operatorname{spn}(L+\nu)}=-\frac{1}{8}\vartheta_{1,3,12}.
\]
We next use \eqref{eqn:Thetagenm=14} to compute the Eisenstein series component $\mathcal{E}_{\operatorname{gen}(L+\nu)}$. To ease notation, we define the \begin{it}sieve operator\end{it} $S_{N,\ell}$, acting on Fourier expansions $f(\tau)=\sum_{n\geq 0 }a_f(n) q^n$ by
\[
f|S_{N,\ell}(\tau):=\sum_{\substack{n\geq 0\\ n\equiv \ell\pmod{N}}} a_f(n)q^n.
\]
Then a straightforward elementary calculation (by splitting the representations $x^2+y^2+z^2=24n+3$ via the congruence classes of $x$, $y$, and $z$) yields
\[
\Theta^3\big|S_{24,3}(\tau)= 48\left( \frac{\Theta_{L+5\nu}(\tau)}{6} + \frac{\Theta_{L+5\mu}(\tau)}{2}+\frac{\Theta_{L+\nu}(\tau)}{6} + \frac{\Theta_{L+\mu}(\tau)}{2}\right) + 8\Theta^3(3\tau)\big|S_{24,3}.
\]
Hence by \eqref{eqn:Thetagenm=14}, we have
\[
\mathcal{E}_{\operatorname{gen}(L+\nu)}=\frac{1}{64}\left(\Theta^3\big|S_{24,3}(\tau)-\frac{1}{8}\Theta^3(3\tau)\big|S_{24,3}\right)
\]
In particular, the $3\ell^2$-th coefficient of $\mathcal{E}_{\operatorname{gen}(L+\nu)}$ is exactly the number of ways to write $3\ell^2$ as the sum of $3$ squares. By \cite[Theorem 86]{Jones}, since the quadratic form $Q(x,y,z)=x^2+y^2+z^2$ has class number $1$, this coefficient is given by
\[
24H\!\left(3\ell^2\right),
\]
where
\[
H(d):=\sum_{\substack{f\in\mathbb{N}\\ -\frac{d}{f^2}\equiv 0,1\pmod{4}}}\frac{h\!\left(-\frac{d}{f^2}\right)}{u\!\left(-\frac{d}{f^2}\right)}
\]
denotes the Hurwitz class number, with $h(D)$ denoting the usual class number and $u(D)$ being half the size of the automorphism group of the order of discriminant $D$ in $\mathbb{Q}(\sqrt{D})$.
However, for $d=3\ell^2$, the class number formula \cite[Corollary 7.28, page 148]{Cox} and $h(-3)=1$ (as well as the fact that $u(-3)=3$ and $u(-3r^2)=1$ for $r>1$) imply that (for $\ell$ prime)
\begin{equation}\label{eqn:Heval}
H\!\left(3\ell^2\right) = \frac{h(-3)}{3} + h\!\left(-3\ell^2\right) = \frac{1}{3}+ \frac{1}{3}\left(\ell-\!\left(\frac{-3}{\ell}\right)\right)= \frac{1}{3}\!\left(\ell +1-\!\left(\frac{-3}{\ell}\right)\right).
\end{equation}
Here $(-3/\ell)$ is the Kronecker--Jacobi--Legendre symbol, which for $\ell\equiv 1\pmod{3}$ is $1$ in particular. Thus for prime $\ell\equiv 1\pmod{3}$, the coefficient of $3\ell^2$ in $\mathcal{E}_{\operatorname{gen}(L+\nu)}$ is
\begin{equation}\label{eqn:Eiscoeffm=14}
\frac{1}{64}\cdot 24 \cdot \frac{\ell}{3}= \frac{\ell}{8}.
\end{equation}
At the same time,
\begin{equation}\label{eqn:hrewrite}
\vartheta_{1,3,12}(\tau)= \sum_{\substack{n\in\mathbb{Z}\\ n\equiv 1\pmod{4}}} n q^{3n^2} = \sum_{n\geq 0} \left(\frac{-4}{n}\right) n q^{3n^2}.
\end{equation}
Thus the $3n^2$-th coefficient of $\vartheta_{1,3,12}(\tau)/8$ is $\left(\frac{-4}{n}\right) n/8$. For $n=\ell\equiv 1\pmod{4}$, we specifically have $\ell/8$, which cancels with the coefficient from $\mathcal{E}_{\operatorname{gen}(L+\nu)}$. Hence by Proposition \ref{prop:SiegelWeil} (in particular, see \eqref{eqn:Thetaspnm=14}), the $3\ell^2$-th coefficient of $\Theta_{\operatorname{spn}(L+\nu)}$ is zero. This yields the claim.
\noindent
\vspace{0.05in}
\noindent
(2) We argue similarly to part (1), except this time we use the lattice coset $L+5\nu$ instead of $L+\nu$ (the statement is a rewording of the claim that $L+5\nu$ does not represent any integer of the form $3\ell^2$ with $\ell\equiv 7\pmod{12}$ prime). The classes in the spinor genus of $L+5\nu$ are given by $L+5\nu$ and $L+5\mu$. Moreover, by \eqref{eqn:Thetagenm=14} we have
\[
\Theta_{\operatorname{gen}(L+5\nu)}=\Theta_{\operatorname{gen}(L+\nu)}= \frac{1}{2}\left(\Theta_{\operatorname{spn}(L+\nu)}+\Theta_{\operatorname{spn}(L+5\nu)}\right).
\]
Rearranging and plugging in \eqref{eqn:Thetaspnm=14}, we have
\[
\Theta_{\operatorname{spn}(L+5\nu)} = 2\Theta_{\operatorname{gen}(L+\nu)} -\Theta_{\operatorname{spn}(L+\nu)} \overset{\eqref{eqn:Thetaspnm=14}}{=} \Theta_{\operatorname{gen}(L+\nu)}+\frac{1}{8}\vartheta_{1,3,12}(\tau).
\]
We then use \eqref{eqn:Heval} and \eqref{eqn:hrewrite} to compute the $3\ell^2$-th coefficient of each side. For $\ell\equiv 7\pmod{12}$ prime, we have $(-3/\ell)=1$ so that \eqref{eqn:Eiscoeffm=14} yields that the $3\ell^2$-th coefficient of the Eisenstein series $\Theta_{\operatorname{gen}(L+\nu)}$ is precisely $\ell/8$. Since $(-4/\ell)=-1$ for $\ell\equiv 7\pmod{12}$, the coefficient of $\vartheta_{1,3,12}(\tau)/8$ is $-\ell/8$, giving cancellation. We conclude that the spinor genus of $L+5\nu$ does not represent $3\ell^2$ by Proposition \ref{prop:SiegelWeil}.
\end{proof}
We are now ready to prove Theorem \ref{thm:P14}.
\begin{proof}[Proof of Theorem \ref{thm:P14}]
For $m\equiv 2\pmod{12}$, we write $m=12r+2$. We claim that, in particular, $P_m$ does not represent $n$ whenever
\[
2(m-2)n + 3\left(\frac{m-4}{2}\right)^2 = 3\ell^2,
\]
where $\ell$ is any prime satisfying
\begin{equation}\label{eqn:ellmod12}
\begin{cases} \ell\equiv 1\pmod{12} & \text{if $r$ is odd,}\\
\ell\equiv 7\pmod{12} & \text{if $r$ is even.}
\end{cases}
\end{equation}
Note first that $P_m$ represents $n$ if and only if there exist $x,y,z\in\mathbb{Z}$ such that
\begin{align}
\nonumber 24rn+3(6r-1)^2 &= 2(m-2)n+3\!\left(\frac{m-4}{2}\right)^2\\
\nonumber&=\!\left((m-2)x+\frac{m-4}{2}\right)^2\!\!+\!\left((m-2)y+\frac{m-4}{2}\right)^2\!\!+\!\left((m-2)z+\frac{m-4}{2}\right)^2\\
\label{eqn:badident}&= \left(12rx+6r-1\right)^2+\left(12ry+6r-1\right)^2+\left(12rz+6r-1\right)^2.
\end{align}
Notice that since $(6r-1)^2\equiv 1\pmod{24}$, the left hand side of \eqref{eqn:badident} is congruent to $3$ modulo $24$, so we may write it in the shape $24n'+3$ for some $n'$. Writing $X:=12rx+6r-1$, $Y:=12ry+6r-1$, and $Z:=12rz+6r-1$, if \eqref{eqn:badident} holds, then there hence exist $X,Y,Z\in\mathbb{Z}$ with $X\equiv Y\equiv Z\equiv 6r-1 \pmod{12}$ for which $X^2+Y^2+Z^2=24n'+3$. In particular, if $24n'+3=3\ell^2$ with $\ell$ a prime satisfying \eqref{eqn:ellmod12}, then Theorem \ref{thm:XYZmod12} implies that \eqref{eqn:badident} is not solvable.
Therefore $P_m$ is not almost universal if there are infinitely many primes $\ell$ satisfying \eqref{eqn:ellmod12} for which $3\ell^2$ is in the set
\[
\mathcal{S}:=\left\{ 2(m-2)n+3\!\left(\frac{m-4}{2}\right)^2: n\in\mathbb{N}_0\right\}=\left\{ 24rn+3(6r-1)^2: n\in\mathbb{N}_0\right\}.
\]
Hence, we need to find infinitely many $\ell$ satisfying \eqref{eqn:ellmod12} and $3\ell^2\equiv 3\pmod{24r}$, or in other words, we want $\ell^2\equiv 1\pmod{8r}$ and $\ell\equiv 1\pmod{12}$ if $r$ is odd and $\ell\equiv 7\pmod{12}$ if $r$ is even. For $r$ odd, we take $\ell\equiv 1\pmod{12r}$ sufficiently large. For $r=2^a3^br'$ with $a>0$, we require $\ell\equiv 7\pmod{12}$ and $\ell^2\equiv 1\pmod{8r}$. By the Chinese Remainder Theorem and Hensel's Lemma, there are infinitely many $\ell\equiv 1\pmod{r'}$, $\ell\equiv 1\pmod{3^{b+1}}$, and $\ell\equiv -1\pmod{4}$ such that $\ell^2\equiv 1\pmod{2^{3+a}}$, and these $\ell$ satisfy the desired congruences. Therefore, there are infinitely many $\ell$ satisfying \eqref{eqn:ellmod12} for which $3\ell^2\in\mathcal{S}$ by the existence of infinitely many primes in arithmetic progressions. Each such $\ell$ corresponds to some $n$ which is not represented by $P_m$, yielding the claim.
\end{proof}
|
\section{Application Example: Robust Quadratic Programming}\label{sec:Application}
Our framework is general and can be applied to many robust convex optimization problems. In this section we walk through the setup and resulting convergence rates of our framework for a robust feasibility problem of a quadratically constrained quadratic program (QP) with ellipsoidal uncertainty. To be precise, our deterministic feasibility problem is
\begin{align*}
\text{find}~~ x\in X~~\text{s.t.}~~ \|A_i x\|_2^2 \leq b_i^\top x +c_i,\quad \forall i\in[m],
\end{align*}
where $X \subseteq {\mathbb{R}}^n$ is the unit Euclidean ball, $A_i\in{\mathbb{R}}^{n\times n}$, $b_i\in{\mathbb{R}}^n$, and $c_i\in{\mathbb{R}}$ for all $i\in[m]$.
We consider the robust quadratic feasibility problem given by
\begin{equation}\label{eqn:robustQP-feas}
\text{find}~~ x\in X~~\text{s.t.}~~\sup_{u \in \widehat{U}} \left\| \left( A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)} \right) x \right\|_2^2 - b_i^\top x -c_i \leq 0,\quad \forall i\in[m],
\end{equation}
where $P_1^i,\ldots,P_K^i$ are uncertainty matrices for each constraint $i\in[m]$, for simplicity we assume uncertainty sets $U^i = \widehat{U}= \{u \in {\mathbb{R}}^{K} : \|u\|_2 \leq 1\}$ for all $i\in[m]$, and $u^{(k)}$ denotes the $k$-th entry of $u$.
It is well known that the robust counterpart of this feasibility problem is a semidefinite program \cite{BenTalelGhaouiNemirovski2009,BertsimasBrownCaramanis2011}. Because current state-of-the-art QP solvers can handle two to three orders
of magnitude larger QPs than semidefinite programs (SDPs), Ben-Tal et al.\@ \cite[Section 4.2]{BenTalHazan2015} suggest an approach that avoids solving SDPs associated with robust QPs. Their approach relies on running a probabilistic OCO algorithm in which a trust region subproblem (TRS)---a class of well-studied nonconvex QPs---is solved in each iteration.
Our results here further enhance this approach. In particular, we show that we can achieve the same rate of convergence in our framework while working with a deterministic OCO algorithm and only carrying out first-order updates in each iteration. In fact, the most expensive operation involved with each iteration of our approach is a maximum eigenvalue computation. Because maximum eigenvalue computation is much cheaper than solving a TRS, we not only present a deterministic approach but also strikingly reduce the cost of each iteration.
To simplify our exposition, let us introduce some notation. For each $i\in[m]$, we define the matrix ${\cal P}_x^i \in {\mathbb{R}}^{n \times K}$ whose columns are given by the vectors $P_k^i \, x$ for $k \in [K]$ together with
\begin{align*}
Q_x^i := ({\cal P}_x^i)^\top {\cal P}_x^i \in {\mathbb{S}}^K_+,\quad
r_x^i := ({\cal P}_x^i)^\top A_i x \in {\mathbb{R}}^K, \quad \text{and}\quad
s_x^i := \|A_i x\|_2^2 - b_i^\top x -c_i \in {\mathbb{R}};
\end{align*}
then it is easy to check that for all $i\in[m]$ and $u\in{\mathbb{R}}^K$ we have
\[ \left\| \left( A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)} \right) x \right\|_2^2 - b_i^\top x -c_i = u^\top Q_x^i u + 2(r_x^i)^\top u + s_x^i. \]
For each $i\in[m]$, we define $f^i: X \times \widehat{U} \to {\mathbb{R}}$ as
\begin{align}
f^i(x,u) &:= \left\| \left( A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)} \right) x \right\|_2^2 - b_i^\top x -c_i + \lambda_{\max}(Q_x^i)\left( 1 - \|u\|_2^2 \right) \nonumber\\
&= u^\top Q_x^i u + 2(r_x^i)^\top u + s_x^i + \lambda_{\max}(Q_x^i)\left( 1 - \|u\|_2^2 \right). \label{eqn:f-forQP}
\end{align}
\begin{lemma}\label{lem:QPreformulation}
For each $i\in[m]$, the function $f^i(x,u)$ defined in \eqref{eqn:f-forQP} is convex in $x$ for any fixed $u\in \widehat{U}$ and concave in $u$ for any given $x$. Moreover, for all $i\in[m]$ and for any $x\in X$,
\[ \sup_{u \in \widehat{U}} \left\| \left( A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)} \right) x \right\|_2^2 - b_i^\top x -c_i =\sup_{u \in \widehat{U}} f^i(x,u) . \]
\end{lemma}
{\emph{#1}}{Proof.}
Fix $i\in[m]$.
By rearranging terms in \eqref{eqn:f-forQP}, we obtain $f^i(x,u)=u^\top (Q_x^i-\lambda_{\max}(Q_x^i) I_K) u + 2(r_x^i)^\top u +s_x^i$. Since $Q_x^i-\lambda_{\max}(Q_x^i) I_K\in{\mathbb{S}}^K_+$ for any given $x$, $f^i(x,u)$ is concave in $u$ for any given $x$.
Now consider a fixed $u\in \widehat{U}$. Note that
\[\lambda_{\max}(Q_x^i) = \max_{\|v\|_2 \leq 1} v^\top (Q_x^i) v = \max_{\|v\|_2 \leq 1} \sum_{1 \leq j,k \leq K} v^{(j)} v^{(k)} x^\top \!(P_j^i)^\top P_k^i\, x = \max_{\|v\|_2 \leq 1} x^\top\! \left(\! \sum_{k=1}^{K} P_k^i\, v^{(k)} \!\!\right)^\top\!\!\! \left(\! \sum_{k=1}^{K} P_k^i\, v^{(k)} \!\!\right) x.\]
Because $ \left( \sum_{k=1}^{K} P_k^i\, v^{(k)} \right)^\top \! \left( \sum_{k=1}^{K} P_k^i\, v^{(k)} \right)\in{\mathbb{S}}^n_+$, then $\lambda_{\max}(Q_x^i)$ is a maximum of convex quadratic functions of $x$ and hence is convex in $x$. Thus, for fixed $u \in \widehat{U}$, $f^i(x,u)$ is convex in $x$.
Reformulation of the nonconvex QP over an ellipsoid into a convex QP over the ellipsoid
via the relation between $u^\top\! Q_x^i u + 2(r_x^i)^\top\! u + s_x^i$ and $f^i(x,u)$ in \eqref{eqn:f-forQP} follows from \cite[Theorem 2.1]{JeyakumarLi2013}.
Lemma~\ref{lem:QPreformulation} implies that $\sup_{u \in \widehat{U}} f^i(x,u) \leq 0$ is an alternate representation of our robust quadratic constraint. We next state the convergence rate in our framework for the associated feasibility problem. For this, we define the quantities
\begin{align}\label{eqn:QPquantities}
&\sigma^2 := \max_{i\in[m]} \sum_{k=1}^{K} \|P_k^i\|_{\mathop{\hbox{\rm\scriptsize Fro}}}^2, \quad
\chi := \max_{i\in[m]} \max_{k \in [K]} \|P_k^i\|_{\mathop{\hbox{\rm\scriptsize Spec}}}, \quad \text{and}\quad \notag\\
&
\rho := \max_{i\in[m]}\|A_i\|_{\mathop{\hbox{\rm\scriptsize Spec}}}, \qquad~~
\beta := \max_{i\in[m]}\|b_i\|_2.
\end{align}
Note that $\chi\leq \sigma$. Furthermore, \cite[Lemma 7]{BenTalHazan2015} proves that $\|Q_x^i\|_{\mathop{\hbox{\rm\scriptsize Fro}}}\leq \sigma^2$ and $\|r_x^i\|_2\leq \sigma\rho$ holds for all $x$ such that $\|x\|_2\leq 1$.
\begin{corollary}\label{cor:robustQP}
Let our domain be given by $X = \{x\in{\mathbb{R}}^n:\; \|x\|_2 \leq 1\}$.
The customization of our OFO-based approach to the problem~\eqref{eqn:robustQP-feas} ensures that within $O\left( ((\rho + \sqrt{K} \sigma)^2 + \beta)^2 \right) \epsilon^{-2}$ iterations, we obtain a robust feasibility/infeasibility certificate.
Moreover, each iteration in our framework relies on a first-order update where the most expensive operation in the case of \eqref{eqn:robustQP-feas} is computing $\lambda_{\max}(Q_x^i)$, which can be done efficiently.
\end{corollary}
{\emph{#1}}{Proof.}
In order to apply OFO-based approach, we need to customize our proximal setup. Given that the sets $X$ and $ \widehat{U}$ are Euclidean balls, we set the proximal setup for generating the iterates $\{x_t,u^i_t\}_{t=1}^T$ to be the standard Euclidean d.g.f.\ with $\|\cdot\|_2$-norm, and thus $\Omega_X=\Omega_{\widehat{U}}={1\over 2}$. We must bound the magnitude of the gradients measured by the $\|\cdot\|_2$-norm.
Note that for any $i\in[m]$, the gradients of $f^i$ are given by
\begin{align*}
\ensuremath{\nabla}_{u} f^i(x,u) &= 2 \left( Q_x^i - \lambda_{\max}(Q_x^i) I_K \right) u + 2 r_x^i\\
\ensuremath{\nabla}_x f^i(x,u) &= 2 \left(\!\! A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)} \!\!\right)^\top \!\!\left(\!\! A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)} \!\!\right) x + 2 \left(1 - \|u\|_2^2\right)\! \left(\! \sum_{k=1}^{K} P_k^i\, v^{(k)} \!\right)^\top\!\! \left(\! \sum_{k=1}^{K} P_k^i\, v^{(k)} \!\right) x -b_i,
\end{align*}
where $v \in \widehat{U}$ is an eigenvector of $Q_x^i$ corresponding to $\lambda_{\max}(Q_x^i)$.
Let us fix an $i\in[m]$. We first bound $\|\ensuremath{\nabla}_u f^i(x,u)\|_2$ for any $u\in\widehat{U}$ as follows:
\begin{align*}
\|\ensuremath{\nabla}_u f^i(x,u)\|_2 &= 2 \left\| \left( Q_x^i - \lambda_{\max}(Q_x^i) I_K \right) u + r_x^i \right\|_2\\
&\leq 2\left(\left\| \left( Q_x^i - \lambda_{\max}(Q_x^i) I_K \right) u\right\|_2 + \left\| r_x^i \right\|_2\right)
\leq 2 \lambda_{\max}(Q_x^i) \|u\|_2 + 2 \sigma \rho
\leq 2 (\sigma^2 + \sigma \rho),
\end{align*}
where the second inequality follows from $\left\| Q_x^i - \lambda_{\max}(Q_x^i) I_K \right\|_{\mathop{\hbox{\rm\scriptsize Spec}}} \leq \lambda_{\max}(Q_x^i)$ and $\| r_x^i\|_2\leq \sigma \rho$ which is implied by \cite[Lemma 7]{BenTalHazan2015}, and the last inequality follows from the facts that $u\in\widehat{U}$, the definitions given in \eqref{eqn:QPquantities}, and $\lambda_{\max}(Q_x^i) = \|{\cal P}_x^i\|_{\mathop{\hbox{\rm\scriptsize Spec}}}^2 \leq \|{\cal P}_x^i\|_{\mathop{\hbox{\rm\scriptsize Fro}}}^2 \leq \sum_{k=1}^{K} \|P_k^i\|_{\mathop{\hbox{\rm\scriptsize Fro}}}^2 \leq \sigma^2$ for any $x\in X$.
Therefore, we deduce from Theorem~\ref{thm:OCO-non-smooth} with uniform weights $\theta_t= 1/T$ that the rate of convergence for bounding the weighted regret associated with constraint $i\in[m]$ using the online mirror descent algorithm is
\[ \sup_{u \in U} \frac{1}{T} \sum_{t=1}^{T} f^i(x_t,u) - \frac{1}{T} \sum_{t=1}^{T} f^i(x_t,u_t) \leq \frac{2(\sigma^2 + \sigma \rho)}{\sqrt{T}}.\]
This implies that $r_u(\epsilon) = O((\sigma^2 + \sigma \rho)^2 \epsilon^{-2})$.
We next bound the weighted regret of the functions $\varphi_t(x) = \max_{i \in [m]} f^i(x,u_t^i)$, i.e., the term $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ by bounding the $\|\cdot\|_2$-norm of $\ensuremath{\nabla}_x \varphi_t(x)$. Notice that
\begin{align*}
\left\| \ensuremath{\nabla}_x \varphi_t(x) \right\|_2 &\leq \max_{i \in [m]} \|\ensuremath{\nabla}_x f^i(x,u_t^i)\|_2.
\end{align*}
Thus, we must bound $\|\ensuremath{\nabla}_x f^i(x,u)\|_2$ for all $x \in X$, $u \in \widehat{U}$. To this end, note that for any $u\in\widehat{U}$
\[ \left\| \sum_{k=1}^{K} P_k^i\, u^{(k)} \right\|_{\mathop{\hbox{\rm\scriptsize Spec}}} \leq \sum_{k=1}^{K} \|P_k^i\|_{\mathop{\hbox{\rm\scriptsize Spec}}}\, |u^{(k)}| \leq \sqrt{K} \max_{k \in [K]} \|P_k^i\|_{\mathop{\hbox{\rm\scriptsize Spec}}} \leq \sqrt{K} \chi, \]
where the second inequality holds because $\|u\|_1\leq\sqrt{K}\|u\|_2\leq\sqrt{K}$ holds for all $u\in\widehat{U}$.
Then for any $x\in X$, $u\in\widehat{U}$, and eigenvector $v\in\widehat{U}$, we have
\begin{align*}
\|\ensuremath{\nabla}_x f^i(x,u)\|_2
&\leq 2\left\| A_i + \sum_{k=1}^{K} P_k^i\, u^{(k)}\right\|_{\mathop{\hbox{\rm\scriptsize Spec}}}^2 \|x\|_2 + 2\left(1 - \|u\|_2^2\right)\left\| \sum_{k=1}^{K} P_k^i\, v^{(k)} \right\|_{\mathop{\hbox{\rm\scriptsize Spec}}}^2 \|x\|_2 +\|b_i\|_2\\
&\leq 2 (\rho + \sqrt{K} \chi )^2 + 2 K \chi^2 +\beta\\
&\leq 4 (\rho + \sqrt{K} \sigma )^2 +\beta.
\end{align*}
Hence, $\left\| \ensuremath{\nabla}_x \varphi_t(x) \right\|_2 \leq 4 (\rho + \sqrt{K} \sigma )^2 +\beta$. Then Theorem~\ref{thm:OCO-non-smooth} with weights $\theta_t=1/T$ implies
\[ \sum_{t=1}^T \theta_t \max_{i \in [m]} f^i(x_t,u_t^i) - \inf_{x \in X} \sum_{t=1}^T \max_{i \in [m]} f^i(x,u_t^i) \leq \frac{\left( 4 (\rho + \sqrt{K} \sigma )^2 +\beta \right)}{\sqrt{T}}. \]
Thus, $r_x(\epsilon) = O\left( ((\rho + \sqrt{K} \sigma)^2 + \beta)^2 \right) \epsilon^{-2}$. Therefore, the number of iterations required for our OFO-based approach to obtain a robust feasibility/infeasibility certificate is $T = \max\{r_x(\epsilon),r_u(\epsilon)\}$.
Note that each iteration of our approach requires a first-order update that is composed of computing the gradients $\ensuremath{\nabla}_x f^i(x,u)$ and $\ensuremath{\nabla}_u f^i(x,u)$ and prox computations. Because our domains involve only direct products of Euclidean balls and simplices, they admit efficient prox computations which take $O(Km+mn)$ time.
In order to evaluate the gradients $\ensuremath{\nabla}_x f^i(x,u)$ and $\ensuremath{\nabla}_u f^i(x,u)$, in addition to the elementary matrix vector operations, we need to compute $\lambda_{\max}(Q_x^i)$ which is the most expensive operation in our first-order update. Fortunately, computing the maximum eigenvalue of a matrix is a well-studied problem and can be computed very efficiently.
In the case of robust QP feasibility problem~\eqref{eqn:robustQP-feas}, \cite[Corollary 3]{BenTalHazan2015} states that with probability $1-\delta$, their framework returns robust feasibility/infeasibility certificates in at most \sloppy $O\left( K^2 \sigma^2(\rho^2 + \sigma^2) \mathop{{\rm log}}({m/\delta})\epsilon^{-2} \right)$ calls (iterations) to their oracle. In each call to their oracle, a nominal feasibility problem is solved to the accuracy ${\epsilon/2}$. In comparison we deduce from Corollary~\ref{cor:robustQP} that our framework requires comparable number of iterations as the approach of Ben-Tal et al.\@ \cite{BenTalHazan2015}. Even so, there are a number of reasons that considerably favor our approach. First, our approach is deterministic as opposed to the high $1-\delta$ probability guarantee of \cite{BenTalHazan2015} which requires using an adaptation of the follow-the-perturbed-leader type OCO. Second, each iteration of their approach requires solving a nominal feasibility problem for solution oracle as well as solving TRSs for the computation of noises $u_t$. In contrast to this, in each iteration we carry out mainly elementary operations such as matrix vector multiplications and our most computationally expensive operation is the maximum eigenvalue computations $\lambda_{\max}(Q_x^i)$. While there are established algorithms to solve the TRS, it is inherently more complicated than finding the maximum eigenvalue of a positive semidefinite matrix. Moreover, \cite{BenTalHazan2015} suffers from the additional computational cost of their solution oracle which solves the nominal feasibility problem. Hence, our approach, while requiring a comparable number of iterations, reduces the cost per iteration remarkably.
\section{Conclusion}\label{sec:Conclusions}
In this paper, we advance the line of research in \cite{BertsimasSim2003,MutapcicBoyd2009,BenTalHazan2015} that aims to solve robust optimization problems via iterative techniques, i.e., without transforming them into their equivalent robust counterparts. Thus far, the literature on iterative methods for RO has relied on more expensive nominal feasibility or pessimization oracles. However, in many applications of robust convex optimization, the original deterministic problem comes equipped with first-order oracles that provide gradient/subgradient information on the constraint functions $f^i$. In this paper, we present an efficient framework that can both work with cheap online first-order oracles and also capture the prior oracle-based approaches of \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015}.
We further show that working with these OFO oracles essentially does not increase the worst case theoretical bound on number of overall oracle calls, i.e., the worst case bound on number of main iterations of our approach is better than or comparable to the prior approaches. Moreover, when OFO oracles are utilized in our framework, the resulting overall arithmetic complexity including all of the basic operations in each iteration is remarkably cheaper than the prior approaches. The resulting framework is simple, easy-to-implement, flexible, and it can easily be customized to many applications. We demonstrate our framework via an illustrative robust QP example, where the most expensive operation in each iteration of our framework is a maximum eigenvalue computation. We further illustrate this with a preliminary numerical study on robust portfolio optimization problem.
Our framework is amenable to exploiting favorable structural properties of the functions $f^i$ such as strong concavity, smoothness, etc., through which better convergence rates can be achieved. For example, when $f^i$ are strongly concave in $u^i$, by exploiting this structural information and using a customization of the weighted regret online mirror descent for strongly convex functions, it is possible to achieve a better convergence rate of $O(1/\epsilon)$ in both our online first-order oracle setup and the nominal feasibility oracle framework of \cite{BenTalHazan2015}. This then partially resolves/refines an open question stated in \cite{BenTalHazan2015} for the lower bound on the number of iterations/calls needed in their nominal feasibility oracle based framework. However, it remains open whether $O(1/\epsilon^2)$ bound is tight when no further favorable structure is present in $f^i$ or the tightness of $O(1/\epsilon)$ in the favorable case.
There are several other compelling avenues for future research. From a practical perspective, it is well-known, and also confirmed by our preliminary proof-of-concept computational experiments, that the computation of gradients/subgradients constitute a major bottleneck in the practical performance of FOMs. Thus, as a step to reduce the efforts involved in such computations, possible incorporation of stochastic \cite{RobbinsMonro1951,NemJudLanShapiro2009} and/or randomized FOMs \cite{JKKN13,BenTalNemirovski2015} working with stochastic subgradients into our framework is of great practical and theoretical interest. A critical assumption in our approach as well as others, e.g., see \cite{BenTalHazan2015} and references therein, is that the domain $X$ is convex. Removing the convexity requirement on the domain $X$ will be an important theoretical development on its own. Besides, this will open up possibilities for more principled approaches to solving robust combinatorial optimization problems (see \cite{BertsimasBrownCaramanis2011,BertsimasSim2003}) where such a convexity assumption on $X$ is not satisfied. Finally, another attractive research direction is develop analogous frameworks for multi-stage RO problems such as robust Markov decision processes (see \cite{NilimElGhaoui2005,Iyengar2005}).
\section{Convex-Concave Saddle Point Reformulation}\label{sec:concaveSP}
The SP problem~\eqref{eqn:convex-nonconcaveSP} based on the function $\Phi(x,u)$ which is not necessarily concave in $u$ admits a convex-concave SP representation in a lifted space via perspective transformations. To present this reformulation, we start by defining the following sets with additional variables $y\in{\mathbb{R}}^m_+$ and new variables $v^i$ for $i\in[m]$:
\begin{align*}
V^i &= \left\{ [v^i;y^{(i)}] :~ 0 < y^{(i)} \leq 1,~ \frac{v^i}{y^{(i)}} \in U^i \right\}\quad\forall i\in[m], \\
W &= \left\{ w = [v^1;\ldots;v^m;y] :~ [v^i; y^{(i)}] \in \cl(V^i),\ i \in[m],~ \sum_{i=1}^{m} y^{(i)} = 1 \right\}.
\end{align*}
Note that for all $i\in[m]$, $\cl(V^i) = V^i \cup \{[0;0]\}$ because we assumed $U^i$ to be closed sets.
For the point $[v^i;y^{(i)}] = [0;0]$, we set $y^{(i)} f^i\left( x, \frac{v^i}{y^{(i)}} \right) = 0$ for any $x\in X$. Note that setting $y^{(i)} f^i\left( x, \frac{v^i}{y^{(i)}} \right) = 0$ for $[v^i;y^{(i)}] = [0;0]$ is well-defined as the continuation since from Assumption~\ref{ass:f-concave-u}, $f^i(x,u^i)$ is continuous and finite-valued on $U^i$, and $U^i$ is compact, so we deduce that $f^i(x,u^i)$ will be bounded on $U^i$. We also define the function $\psi:X \times W \to {\mathbb{R}}$ as
\[
\psi(x,w) = \psi(x,v,y) := \sum_{i=1}^{m} y^{(i)} f^i\left( x, \frac{v^i}{y^{(i)}} \right).
\]
\begin{lemma}\label{lem:convexSPdef}
For fixed $w\in W$, the function $\psi(x,w)$ is convex in $x$ over $X$, and $\psi(x,w)$ is a concave function of $w$ over $W$ for any fixed $x$. Moreover, $W$ is closed, and when $U^i$ for $i\in[m]$ are convex, the sets $V^i$ for $i\in[m]$ and $W$ are all convex.
\end{lemma}
{\emph{#1}}{Proof.}
For any $w = [v^1;\ldots;v^m;y]$, the function $\psi$ is convex in $x$ since in all of the nonzero terms in the summation over all $i\in[m]$ defining $\psi$, we have $y^{(i)}>0$ and in each such nonzero term each function $f^i\left( x, \frac{v^i}{y^{(i)}} \right)$ is convex in $x$ for the given $\frac{v^i}{y^{(i)}}\in U^i$ (see Assumption~\ref{ass:f-concave-u}). In addition, for any given $x\in X$, the function $\psi$ is jointly concave in $v$ and $y$ because it is written as a sum of the perspective functions of functions $f^i$ which are concave in $u^i$ (see Assumption~\ref{ass:f-concave-u}).
The closedness of $W$ is immediate, and the convexity of the sets $V^i$ and $W$ follows immediately from their definition and the convexity assumption on $U^i$.
With these definitions and Lemma~\ref{lem:convexSPdef}, we observe that \eqref{eqn:convex-nonconcaveSP} is equivalent to evaluating the convex-concave SP problem defined by the function $\psi$ over the convex domains $X$ and $W$:
\begin{equation}\label{eqn:concaveSP}
\inf_{x \in X} \sup_{w \in W} \psi(x,w) \leq \epsilon \quad \text{or} \quad \inf_{x \in X} \sup_{w \in W} \psi(x,w) > 0.
\end{equation}
We state this formally in the following lemma.
\begin{lemma}\label{lem:concaveSPReformulation}
For any $\epsilon > 0$ and $\bar{x} \in X$,
\[
\max_{i\in[m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) \leq \epsilon \quad\text{if and only if}\quad \sup_{w \in W} \psi(\bar{x},w) \leq \epsilon.
\]
As a result,
\[
\inf_{x \in X} \max_{i\in[m]} \sup_{u^i \in U^i} f^i(x,u^i) \leq \epsilon \quad\text{if and only if}\quad \inf_{x \in X} \sup_{w \in W} \psi(x,w) \leq \epsilon.
\]
\end{lemma}
{\emph{#1}}{Proof.}
Fix $\bar{x} \in X$ and $\epsilon > 0$. Suppose $\max_{i\in[m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) \leq \epsilon$; then for all $u^i \in U^i$, $i \in [m]$, we have $f^i(\bar{x},u^i) \leq \epsilon$. Now consider any $w = [v^1;\ldots;v^m;y] \in W$. Then $0 \leq y^{(i)} \leq 1$ for all $i \in [m]$ and $\sum_{i=1}^m y^{(i)}=1$. For all $i \in [m]$, define $u^i = {v^i\over y^{(i)}} \in U^i$ whenever $y^{(i)}>0$.
Then $y^{(i)} f^i(\bar{x},{v^i\over y^{(i)}}) = y^{(i)} f^i(\bar{x},u^i) \leq y^{(i)} \epsilon$ for $0 <y^{(i)} \leq 1$. In addition, when $y^{(i)}=0$, because $w\in W$ we must have $v^i=0$ and then by definition we have $y^{(i)} f^i(\bar{x},{v^i\over y^{(i)}})=0$. Therefore, from $\sum_{i=1}^my^{(i)}=1$, we deduce $\psi(\bar{x},w) = \sum_{i=1}^{m} y^{(i)} f^i\left(\bar{x}, \frac{v^i}{y^{(i)}} \right)\leq \epsilon$ holds for any $w\in W$.
Now suppose that $\sup_{w \in W} \psi(\bar{x},w) \leq \epsilon$ holds. Given $i \in [m]$ and $u^i \in U^i$, set $w$ to have components $y^{(i)} = 1$, $v^i = u^i$, and $[v^j;y^{(j)}] = [0;0]$ for $j \neq i$. Then $f^i(\bar{x},u^i) = \psi(\bar{x},w) \leq \epsilon$. Hence, $\max_{i\in[m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) \leq \epsilon$ follows.
\begin{remark}\label{rem:m=1pers}
When $m=1$, i.e., we have only one function $f^1(x,u^1)$ and only one uncertainty set $U^1$, hence $W = U^1$ and $\inf_{x\in X} \sup_{v \in W} \psi(x,w) = \inf_{x\in X} \sup_{u^1 \in U^1} f^1(x,u^1)$. Also, under Assumption~\ref{ass:f-concave-u}, $\psi(x,w)$ is convex in $x$ and concave in $u^1$. Thus, the preceding perspective transformation resulting in \eqref{eqn:concaveSP} directly generalizes this case of a convex-concave SP formulation for $m=1$ discussed in Remark~\ref{rem:SPm=1}.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
As a result, Lemma~\ref{lem:concaveSPReformulation} and Theorem~\ref{thm:saddlept-certificate} combined with any FOM that provides bounds on the saddle point gap $\epsilon_{\sad}^\psi(\bar{x},\bar{w})$ lead to an efficient way of verifying robust feasibility of \eqref{eqn:convex-nonconcaveSP} as follows:
\begin{theorem}\label{thm:single-robust-constraint-feas-check}
Suppose $\bar{x}\in X$, $\bar{w}\in W$, and $\tau\in(0,1)$ are such that $\epsilon_{\sad}^\psi(\bar{x},\bar{w}) \leq \tau\epsilon$. If $\psi(\bar{x},\bar{w}) \leq (1-\tau)\epsilon$, then \sloppy $\max_{i \in [m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) \leq \epsilon$. If $\psi(\bar{x},\bar{w}) > (1-\tau)\epsilon$ and $\tau\leq{1\over 2}$, then $\inf_{x \in X} \max_{i \in [m]} \sup_{u^i \in U^i} f^i(x,u^i) > 0$.
\end{theorem}
{\emph{#1}}{Proof.}
Suppose $\psi(\bar{x},\bar{w}) \leq \tau\epsilon$. By Theorem~\ref{thm:saddlept-certificate}, we have $\inf_{x \in X} \sup_{w \in W} \psi(x,w) \leq \sup_{w\in W} \psi(\bar{x},w) \leq \epsilon$. By Lemma~\ref{lem:concaveSPReformulation}, $\max_{i \in [m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) \leq \epsilon$ as well.
On the other hand, when $\psi(\bar{x},\bar{w}) > (1-\tau)\epsilon$ and $\tau\leq{1\over 2}$, Theorem~\ref{thm:saddlept-certificate} implies \sloppy $\inf_{x\in X} \sup_{w\in W} \psi(x,w) \geq \inf_{x\in X} \psi(x,\bar{w}) > 0$. Then by Lemma~\ref{lem:concaveSPReformulation}, $\max_{i \in [m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) > 0$ follows.
Because of the existence of efficient FOMs to solve convex-concave SP problems, Theorem~\ref{thm:single-robust-constraint-feas-check} suggests a possible advantage of using the convex-concave SP problem given in \eqref{eqn:concaveSP}.
Nevertheless, working with the SP reformulation given by \eqref{eqn:concaveSP} in the extended space $X\times W$ presents a number of critical challenges.
First, efficient FOMs associated with convex-concave SP problems often require computing prox operations or projections onto the domains $X$ and $W$. Unfortunately, even if projection (or prox-mappings) onto $U^i$ admits a closed form solution or an efficient procedure, it is unclear how to extend such projections onto $W$.
Furthermore, while the perspective transformations involved in constructing the function $\psi$ preserves certain desirable properties of the functions $f^i$, such as Lipschitz continuity and smoothness, the parameters associated with $\psi$ are in general larger than those associated with the original functions $f^i$. Such parameters are critical for FOM convergence rates, and thus the FOMs when applied to solve \eqref{eqn:concaveSP} will have slower convergence rates.
To address the issues outlined above, in the main paper we discuss how to obtain robust feasibility/infeasibility certificates for the convex-nonconcave SP problem \eqref{eqn:convex-nonconcaveSP} directly, i.e., we work with the functions $f^i$ and the sets $U^i$ directly. This direct approach in particular allows us to take greater advantage of the structure of the original formulation such as the availability of efficient projection (prox) computations over domains, and/or better parameters for smoothness, Lipschitz continuity, etc., of the functions.
\section{Introduction}\label{sec:intro}
Robust optimization (RO) is one of the leading modeling paradigms for optimization problems under uncertainty. As opposed to the other approaches, RO seeks a solution that is immunized against \emph{all} possible realizations of uncertain model parameters (noises) from a given uncertainty set. It is widely adopted in practice mainly because of its computational tractability. We refer the reader to the paper by Ben-Tal and Nemirovski \cite{BenTalNem1998}, the book by Ben-Tal et al.\@ \cite{BenTalelGhaouiNemirovski2009} and surveys \cite{BenTalNemirovski2002,BenTalNemirovski2008,BertsimasBrownCaramanis2011,CaramanisMannorXu2011} for a detailed account of RO theory and numerous applications.
Recently, fascinating connections have been established between problems from the statistics and machine learning domains and robust optimization. More precisely, it is demonstrated that RO can be used to achieve desirable statistical properties such as stability, sparsity, and consistency.
For example, for linear regression problems, El Ghaoui and Lebret \cite{ElGhaouiLebret1997} and Xu et al.\@ \cite{XuCaramanisMannor2010} respectively establish the equivalence of the ridge regression and Lasso to specific RO formulations of unregularized regression problems. Moreover, Xu et al.\@ \cite{XuCaramanisMannor2009} exhibit similar results in the context of regularizing support vector machines (SVMs), and \cite{XuCaramanisMannor2009,XuCaramanisMannor2010} validate the statistical consistency of methods such as SVM and Lasso via RO methodology.
In addition to these RO interpretations of regularization techniques used in statistics and machine learning, robust versions of many problems from these domains are gaining traction. For example, \cite{Shivaswamy2006} examines robust variants of SVMs and other classification problems, and \cite{Ben-TalRobustSVM2012} explores a robust formulation for kernel classification problems. We refer the reader to \cite{CaramanisMannorXu2011,BenTalHazan2015} and references therein for further examples and details on connections between robust optimization and statistics and machine learning.
These recent connections not only highlight the importance of RO methodology but also present algorithmic challenges where the scalability of RO algorithms with problem dimension becomes crucial. The primary method for solving a robust convex optimization problem is to transform it into an equivalent deterministic problem called the \emph{robust counterpart}. Under mild assumptions, this yields a convex and tractable robust counterpart problem (see \cite{BenTalelGhaouiNemirovski2009,BertsimasBrownCaramanis2011,BenTalDenHertog2015}), which can then be solved using existing convex optimization software and tools.
This traditional approach has seen much success in decision making domain, nevertheless it has a major drawback that the reformulated robust counterpart is often not as scalable as the deterministic nominal program. In particular, the robust counterpart can easily belong to a different class of optimization problems as opposed to the underlying original deterministic problem. For example, a linear program (LP) with ellipsoidal uncertainty is equivalent to a convex quadratic program (QP), and similarly, a conic-quadratic program with ellipsoidal uncertainty is equivalent to a semidefinite program (SDP) (see e.g., \cite{BenTalelGhaouiNemirovski2009, BertsimasBrownCaramanis2011}). It is well-known that convex QPs as opposed to LPs, and SDPs as opposed to convex QPs are much less scalable in practice. This then presents a critical challenge in applying RO methodology in big data applications frequently encountered in machine learning and statistics, where even solving the original deterministic nominal problem to high accuracy is prohibitively time-consuming.
The iterative schemes that alternate between the generation/update of candidate solutions and the realizations of noises offer a convenient remedy to the scalability issues associated with the robust counterpart approach. Thus far, such approaches \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015} have relied on two oracles: $(i)$ \emph{solution oracles} to solve instances of extended (or nominal) problems with constraint structures similar to (or the same as) the deterministic problem, and $(ii)$ \emph{noise oracles} to generate/update particular realizations of the uncertain parameters. At each iteration of these schemes, both solution and noise oracles are called, and their outputs are used to update the inputs of each other oracle in the next iteration. Because solution oracles rely on a solver of the same class capable of solving the deterministic problem, these iterative approaches circumvent the issue of the robust counterpart approach potentially relying on a different solver. Nevertheless, these iterative approaches still suffer from a serious drawback: the solution oracles in \cite{MutapcicBoyd2009,BenTalHazan2015} themselves can be expensive as they require solving extended or nominal optimization problems completely. While solving the nominal problem is not as computationally demanding as solving the robust counterpart, the overall procedure depending on repeated calls to such oracles can be prohibitive. In fact, each such call to a solution oracle may endure a significant computational cost, which is at least as much as the computational cost of solving an instance of the deterministic nominal problem. Note that, to ensure scalability, most applications in machine learning and statistics already need to rely on cheap first-order methods for solving deterministic nominal problems.
In this paper, we propose an efficient iterative framework for solving robust convex optimization problems which can rely on, in an \emph{online} fashion, much cheaper \emph{first-order oracles} in place of full solution and noise oracles. In particular, in each iteration, instead of solving a complete optimization problem within the solution and/or noise oracles, we show that simple simultaneous updates on the solution and noise in an online fashion using only first-order information from the deterministic constraint structure is sufficient to solve robust convex optimization problems. Moreover, we show that the number of calls to such online first-order (OFO) oracles is not only at most that of the state-of-the-art iterative approaches utilizing full optimization based oracles for solution and/or noise, but also almost independent of the dimension of the problem. Therefore, this makes our approach especially attractive for applications in statistics and machine learning domains where it is critical to maintain that the overall approach has both gracious dependence on the dimension of the problem and cheap iterations. We outline our contribution more concretely after discussing the most relevant literature.
\subsubsection*{Related Work}
Thus far, the iterative approaches, which bypass the restrictions of the robust counterparts, work with \emph{extended} nominal problems that belong to the same class as the deterministic nominal one by carefully controlling the constraints included in the formulation corresponding to noise realizations.
For robust binary linear optimization problems with only objective function uncertainty and a polyhedral uncertainty set, Bertsimas and Sim \cite{BertsimasSim2003} suggest an approach which relies on solving $n+1$ number of instances of the nominal problem, where $n$ is the dimension of the problem.
For robust convex optimization problems, Calafiore and Campi \cite{CalafioreCampi2004} study a `constraint sampling' approach based on forming a single extended nominal problem of the same class as the deterministic one via i.i.d.\@ sampling of noise realizations. They show that the optimal solution to this extended nominal problem is robust feasible with high probability where the probability depends on the sampling procedure, the number of samples drawn, and the dimension.
Mutapcic and Boyd \cite{MutapcicBoyd2009} follow a `cutting-plane' type approach where in each iteration, a solution oracle is called to solve an extended nominal problem of the same class as the deterministic problem and a noise oracle, referred to as \emph{pessimization oracle}, is invoked to iteratively expand and refine the extended nominal problem. Given a candidate solution, a pessimization oracle either certifies its feasibility with respect to the robust constraints or returns a new noise realization from the uncertainty set for which the solution is infeasible; then the nominal constraint associated with that particular noise realization is included in the extended problem. This process is repeated until a robust feasible solution is found or the last extended problem is found to be infeasible. In the overall procedure, the number of iterations (or calls to the pessimization oracle) can be exponential in the dimension. Despite this, \cite{MutapcicBoyd2009} reports impressive computational results. The cutting-plane approach is also further tested on mixed integer linear problems in \cite{Bertsimas2016} and it is demonstrated that the same computational phenomenon holds.
Both of the approaches from \cite{CalafioreCampi2004} and \cite{MutapcicBoyd2009} pose issues for high-dimensional problems.
In \cite{CalafioreCampi2004}, as the dimension grows, an extended problem with linearly more nominal constraints is required to ensure the high probability guarantee on finding a good quality solution. In \cite{MutapcicBoyd2009} at each iteration, a nominal constraint is added to the extended nominal problem. The theoretical bound on the number of constraints that need to be added is exponential, so the extended problem in \cite{MutapcicBoyd2009} can grow to be exponentially large. Moreover, in both cases the extended nominal problem may no longer have certain favorable problem structure of the deterministic nominal problem, such as a network flow structure.
To address these issues, in particular, the issue of solving extended nominal problems that are not only larger-in-size than the deterministic problem but also may lack certain favorable problem structure of the deterministic problem, Ben-Tal et al.\@ \cite{BenTalHazan2015} introduce a new iterative approach to approximately solve robust feasibility problems via a \emph{nominal feasibility oracle} and running an online learning algorithm to choose noise realizations. Given a particular noise realization, the nominal feasibility oracle solves an instance of the deterministic nominal feasibility problem obtained by simply fixing the noise to the given value. Hence, the problem solved by this oracle has the \emph{same} number of constraints and the \emph{same structure} as the original nominal problem; in particular its size does not grow in each iteration. This is an important distinguishing feature of this approach. The other distinguishing feature is that Ben-Tal et al.\@ \cite{BenTalHazan2015} replace the pessimization oracle of \cite{MutapcicBoyd2009} by employing an online learning algorithm, which simply requires first-order information of the noise from the constraint functions. Moreover, \cite{BenTalHazan2015} provides a dimension independent bound on the number of iterations (nominal feasibility oracle calls).
Because the approaches of both \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015} are closely related to our work, we give a detailed summary of these in Sections~\ref{sec:pessimization-oracle}~and~\ref{sec:oracle-based-approach} respectively and highlight their connections to our work. In fact, we show that they both can be seen as special cases of our framework.
We close with a brief summary of the assumptions on the computational requirements of these methods. The constraint sampling approach of \cite{CalafioreCampi2004} requires access to a sampling procedure on the uncertainty sets as well as an oracle capable of solving the extended nominal problem. The cutting plane approach of \cite{MutapcicBoyd2009} replaces the sampling procedure of \cite{CalafioreCampi2004} with a noise oracle, namely the pessimization oracle that works with the uncertainty sets but still requires the same type of optimization oracle as a solution oracle to solve the extended problems. Ben-Tal et al.\@ \cite{BenTalHazan2015} substitute the pessimization oracle with an online learning-based procedure, which requires merely first-order information from the constraint functions and simple projection type operations on the associated uncertainty sets, but it still relies on a solution oracle capable of solving the original nominal problem, which is essentially the same (up to log factors) as the optimization oracles in \cite{CalafioreCampi2004} and \cite{MutapcicBoyd2009}. If the deterministic problem admits special structure such as network flows etc., a specific solver can be used in the framework of \cite{BenTalHazan2015}, but this is not possible for \cite{CalafioreCampi2004} and \cite{MutapcicBoyd2009}.
\subsubsection*{Summary of Our Contributions}
It is possible to view all of these iterative approaches as two iterative processes that run simultaneously and in conjunction with each other to generate/update solutions and noise realizations. This naturally leads to a dynamic game environment where in each round Player 1 chooses a solution and Player 2 chooses a realization of uncertain parameters. In this framework, the policies employed by these players in their decision making determine the nature of the final approach. In the case of \cite{MutapcicBoyd2009}, Player 1 considers all of the previous noise realizations when making his decision, whereas Player 2 simply reacts to the current solution when choosing the noise. In \cite{BenTalHazan2015}, Player 1 reacts to only the current noise in generating/updating the solution while Player 2 minimizes the regret associated with past solutions in choosing noise.
In this paper, we further analyze this interaction between Player 1 and Player 2, with the aim of deriving a simpler and computationally much less demanding iterative approach to solving RO problems. Our contributions can be summarized as follows.
\begin{enumerate}
\item We build a \emph{general and flexible framework} for iteratively solving robust feasibility problems, and demonstrate its flexibility by describing it as a meta-template. By customizing our framework appropriately, we modify the pessimization oracle-based approach of \cite{MutapcicBoyd2009} by replacing the extended nominal solver used in \cite{MutapcicBoyd2009} with efficient first-order updates. We call this the \emph{FO-based pessimization approach}, and demonstrate that as opposed to \cite{MutapcicBoyd2009} it has both a much better bound on the number of oracle calls and far superior practical performance. We also provide a new interpretation of the nominal feasibility oracle-based approach of \cite{BenTalHazan2015} as a special case within our framework. Furthermore, we extend the analysis of the approach of \cite{BenTalHazan2015} under the same assumptions, e.g., access to a nominal optimization oracle, and show that it can solve the robust optimization problem directly without relying on a binary search (see Remark \ref{rem:oracle-based-optimal-sol}).
\item When the original deterministic problem admits first-order oracles capable of providing gradient/subgradient information on each constraint function, we demonstrate that \emph{online first-order} (OFO) algorithms can be used to iteratively generate/update solutions and noise realizations simultaneously in an online manner leading to robust feasibility/infeasibility certificates within our framework. In contrast to the approaches of \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015}, which rely on full nominal feasibility oracles to generate points, our OFO-based approach only requires simple update rules in each iteration and thus has much lower per-iteration cost. Besides, our noise oracle generates a realization of the noise in an online learning fashion as was done in \cite{BenTalHazan2015}, and hence it is less expensive than the pessimization oracle of \cite{MutapcicBoyd2009}.
\item In our framework, the number of iterations (or oracle calls) needed to obtain approximate robust solution or a robust infeasibility certificate is a function of the approximation guarantee $\epsilon$ and the complexities of the domains for the solution and the uncertainty set; in particular, our convergence rate is (almost) independent of both the number of robust constraints and the dimension of the deterministic problem. We also demonstrate that the iteration complexity of our OFO-based approach is at least as good as that of the efficient approach of \cite{BenTalHazan2015}, and better than the exponential complexity of \cite{MutapcicBoyd2009}. Overall, our OFO-based approach leads to computational savings over the approach of \cite{BenTalHazan2015} by a factor as large as $O(1/(\epsilon^2 \mathop{{\rm log}}(1/\epsilon)))$ arithmetic operations when the number updates of the solution is smaller than or equal to the number of updates of the noise realization, which is the case in many applications. For further comparisons and discussion, see Section~\ref{sec:RateDiscussion}. In addition, our framework is amenable to exploiting favorable structural properties of the constraint functions such as strong concavity, smoothness, etc., through which better convergence rates can be achieved.
\item Our framework is based on formulating the robust feasibility problem as a convex-nonconcave saddle point (SP) problem, and explicitly analyzing its structure. While convex-concave SP problems are well-studied in the literature, and many efficient first-order algorithms exist for these (see for example \cite{Nesterov2005,JuditNem2012Pt1,JuditNem2012Pt2}), the convex-\emph{nonconcave} SP problem is not as well-studied. To our knowledge, an explicit study of convex-nonconcave SP problems and their relation to RO has not been conducted previously; in this respect, the most closely related work \cite{BenTalHazan2015} neither provides an explicit connection between robust feasibility and SP problems, nor analyzes their structure explicitly.
\end{enumerate}
To demonstrate the application and effectiveness of our proposed framework, we walk through a detailed example on robust QPs. In particular, for robust QPs, we are able to leverage a recent convex QP-based reformulation of the classical trust region subproblem \cite{JeyakumarLi2013,Ho-NguyenKK2017TRS}
in order to avoid working with a nonconvex reformulation in a lifted space as in \cite[Section 4.2]{BenTalHazan2015} and relying on a probabilistic follow-the-perturbed-leader type algorithm \cite[Section 3.2]{BenTalHazan2015}. While using such nonconvex techniques will work within our framework, our convex reformulation allows us to work directly in the original space of the variables with a deterministic subgradient-based algorithm while still achieving asymptotically similar iteration complexity guarantees as \cite{BenTalHazan2015}. Moreover, each iteration of our approach requires only first-order updates where the most expensive operation is the computation of a maximum eigenvector; thus our per-iteration cost is significantly less.
We also conduct a preliminary numerical study on the comparison of our approach with other iterative approaches \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015} on robust QPs arising in portfolio optimization. Our results show that when the problem size is small, the nominal solver approaches of \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015} are more efficient. However, when problem size increases, replacing the nominal solvers with first-order updates using our framework allows us to achieve faster solution times. This highlights the benefits and potential of investigating first-order based approaches such as ours in iterative RO methods.
\subsubsection*{Outline}
The rest of the paper is organized as follows. We begin with some notation and preliminaries in Section~\ref{sec:prelim}. We introduce our robust feasibility problem and robust feasibility/infeasibility certificates in Section~\ref{sec:R-feas-prelim}, convex-concave SP problems in Section~\ref{sec:SPprelim}, and briefly summarize important online convex optimization (OCO) tools as well as a useful OFO algorithm in Section~\ref{sec:OCO}. We formulate the robust feasibility problem as a convex-nonconcave SP problem in Section~\ref{sec:SPFormulation}; this formulation and certain bounds associated with its SP gap function form the basis of our general framework for solving robust feasibility problems. In Section~\ref{sec:OCOforRO} we specify an assortment of approaches obtained in our general framework by using different oracles. We examine our OFO-based approach in Section~\ref{sec:RegretforRO} by interpreting various terms in our framework in the context of OCO. In Section \ref{sec:pessimization-oracle}, we modify the pessimization oracle-based approach of \cite{MutapcicBoyd2009} to obtain an efficient bound on the number of iterations required. In Section \ref{sec:oracle-based-approach} we show how the nominal feasibility oracle-based approach of \cite{BenTalHazan2015} fits within our framework.
Finally, we discuss the convergence rates and accelerations attainable in our framework and compare our work with the existing approaches in Section~\ref{sec:RateDiscussion}. In Section~\ref{sec:Application} we illustrate our OFO-based approach through an example application on robust QPs.
We provide in Section~\ref{sec:numerical} a preliminary numerical study comparing our framework with other iterative approaches \cite{MutapcicBoyd2009} and \cite{BenTalHazan2015}. We close with a summary of our results and a few compelling further research directions in Section~\ref{sec:Conclusions}.
In Appendix~\ref{sec:concaveSP} we give an alternative formulation of the robust feasibility problem as a convex-concave SP problem in an extended space, and discuss its advantages and disadvantages over the convex-nonconcave SP formulation.
\section{Notation and Preliminaries}\label{sec:prelim}
Given $a\in{\mathbb{R}}$, $\sign(a)$ denotes the sign of the number $a$. For a positive integer $n\in{\mathbb{N}}$, we let $[n]=\{1,\ldots,n\}$ and define $\Delta_n:=\{x\in{\mathbb{R}}^n_+:~\sum_{i\in[n]} x_i=1\}$ to be the standard simplex. Throughout the paper, the superscript, e.g., $f^i, u^i, U^i$, is used to attribute items to the $i$-th constraint, whereas the subscript, e.g., $x_t,f_t,\phi_t$, is used to attribute items to the $t$-th iteration. Therefore, we sometimes use $u^i$, $x_t$, as well as $u^i_t$ to denote vectors in ${\mathbb{R}}^n$. We use the notation $\{x_t\}_{t=1}^T$ to denote the collection of items $\{x_1,\ldots,x_T\}$. Given a vector $x\in{\mathbb{R}}^n$, we let $x^{(k)}$ denote its $k$-th coordinate for $k\in[n]$. One exception we make to this notation is that we always denote the convex combination weights $\theta\in\Delta_T$ with $\theta_t$. For $x\in{\mathbb{R}}^n$ and $p\in[1,\infty]$, we use $\|x\|_p$ to denote the $\ell_p$-norm of $x$ defined as
\[
\|x\|_p=\begin{cases} \left(\sum_{i\in[n]} |x^{(i)}|^p \right)^{1/p} &\text{if } p\in[1,\infty) \\ \max_{i\in[n]}|x^{(i)}| &\text{if } p=\infty\end{cases} .
\]
Throughout this paper, we use Matlab notation to denote vectors and matrices, i.e., $[x;y]$ denotes the concatenation of two column vectors $x$, $y$.
${\mathbb{S}}^n$ denotes the space of $n\times n$ symmetric matrices, and we let ${\mathbb{S}}^n_+$ be the positive semidefinite cone in ${\mathbb{S}}^n$. We let $I_n$ denote the identity matrix in ${\mathbb{S}}^n$. For a matrix $A\in{\mathbb{S}}^n$, $\lambda_{\max}(A)$, $\|A\|_{\mathop{\hbox{\rm\scriptsize Fro}}}$, and $\|A\|_{\mathop{\hbox{\rm\scriptsize Spec}}}$ correspond to its maximal eigenvalue, Frobenius norm, and spectral norm, respectively.
Given a set $V$, we denote its closure by $\cl(V)$. We abuse notation slightly by denoting $\ensuremath{\nabla} f(x)$ for both the gradient of function $f$ at $x$ if $f$ is differentiable and a subgradient of $f$ at $x$, even if $f$ is not differentiable. If $f$ is of the form $f(x,u)$, then $\ensuremath{\nabla}_x f(x,u)$ denotes the subgradient of $f$ at $x$ while keeping the other variables fixed at $u$.
\subsection{Robust Feasibility Problem}\label{sec:R-feas-prelim}
Consider a convex \emph{deterministic} or \emph{nominal} mathematical program
\begin{equation}\label{eqn:deterministic-opt}
\min_x \left\{ f^0(x):~ x \in X, \;~f^i(x,u^i) \leq 0,\ \forall i\in[m] \right\},
\end{equation}
where the domain $X\subset{\mathbb{R}}^n$ is closed and convex, the functions $f^0(x)$ and $f^i(x,u^i)$ for $i\in[m]$ are convex functions of $x$, and $u = (u^{1},\ldots,u^{m})$ is a fixed parameter vector. Without loss of generality we assume the objective function $f^0(x)$ does not have uncertainty. The \emph{robust convex optimization problem} associated with \eqref{eqn:deterministic-opt} is
\begin{equation}\label{eqn:robust-opt-intro}
\Opt:=\min_x \left\{ f^0(x):~ x \in X, \;~\sup_{u^i \in U^i} f^i(x,u^i) \leq 0,\ \forall i\in[m] \right\},
\end{equation}
where $U^1,\ldots,U^m$ are the \emph{uncertainty sets} given for the parameter $u^i$ of constraint $i\in[m]$. Because we assume formulation~\eqref{eqn:deterministic-opt} is convex, the overall optimization problem in \eqref{eqn:robust-opt-intro} is convex.
In this paper, we work under the following mild regularity assumption:
\begin{assumption}\label{ass:f-concave-u}
The constraint functions $f^i(x,u^i)$ for all $i\in[m]$ are finite-valued on the domain $X\times U^i$, convex in $x$ and concave in $u^i$. $X$, the domain for $x$, is closed and convex, and $U^i$, the domains for $u^i$, are closed and bounded.
\end{assumption}
We take Assumption~\ref{ass:f-concave-u} as given for all our results and proofs. Without loss of generality, we assume that the uncertainty set has a Cartesian product form $U^1 \times \ldots \times U^m$, see e.g., \cite{BenTalNemirovski2002}; we let $U = U^1 \times \ldots \times U^m$ and write $u = [u^1;\ldots;u^m] \in U$. We do not further assume that the sets $U^i$ are convex. However, for some algorithms we consider, convexity of $U^i$ for $i\in[m]$ will be required.
A convex optimization problem can be solved by solving a polynomial number of associated feasibility problems in a standard way, via a binary search over its optimal value. In particular, let $[\underline{\upsilon}_0,\overline{\upsilon}_0]$ be an initial interval containing the optimal value of \eqref{eqn:robust-opt-intro}. At each iteration $k$ of the binary search, we update the domain $X_k:=X\cap\{x:\;f^0(x)\leq \upsilon_k$\} for some $\upsilon_k\in[\underline{\upsilon}_k,\overline{\upsilon}_k]$ and arrive at the following robust feasibility problem:
\begin{equation}\label{eqn:robust-feas}
\text{find}\ \ x \in X_k \quad\text{s.t.}\quad \sup_{u^i\in U^i} f^i(x,u^i) \leq 0 \quad \forall i\in[m].
\end{equation}
Then based on the feasibility/infeasibility status of \eqref{eqn:robust-feas}, we update our range $[\underline{\upsilon}_{k+1},\overline{\upsilon}_{k+1}]$ and go to iteration $k+1$. In this scheme, we are guaranteed to find a solution $x^* \in X$ whose objective value is within $\delta>0$ of the optimum value of \eqref{eqn:robust-opt-intro} in at most $\left\lfloor \mathop{{\rm log}}_2\left(\frac{\overline{\upsilon}_0-\underline{\upsilon}_0}{\delta}\right) \right\rfloor$ iterations. Therefore, one can equivalently study the complexity of solving robust feasibility problem~\eqref{eqn:robust-feas} as opposed to \eqref{eqn:robust-opt-intro}.
From now on, we focus on solving robust feasibility problem and assume that the constraint on the objective function $f^0(x)$ is already included in the domain $X$ for simplicity in our notation.
Given functional constraints $f^i(x) \leq 0$, $i \in [m]$, most convex optimization methods will declare infeasibility or return an approximate solution $x \in X$ such that $f^i(x) \leq \epsilon$ for $i \in [m]$ for some tolerance level $\epsilon > 0$. Therefore, we consider the following \emph{robust approximate feasibility problem}:
\begin{equation}\label{eqn:approx-robust-feas}
\begin{cases}
\text{\emph{Either}: find}\ \ x \in X \quad\text{s.t.}\quad \sup_{u^i\in U^i} f^i(x,u^i) \leq \epsilon \quad \forall i\in[m];\\
\text{\emph{or}: declare infeasibility, } \forall x \in X,\ \exists i \in [m] \quad \text{s.t.} \quad \sup_{u^i\in U^i} f^i(x,u^i) > 0.
\end{cases}
\end{equation}
We refer to any feasible solution $x$ to \eqref{eqn:approx-robust-feas}, i.e., $x \in X$ such that $\sup_{u^i\in U^i} f^i(x,u^i) \leq \epsilon$ holds for all $i\in[m]$ as a \emph{robust $\epsilon$-feasibility certificate}. Similarly, any realization of the uncertain parameters $\bar{u}\in U$ such that there exists no $x\in X$ satisfying $f^i(x,\bar{u}^i) \leq 0$ for all $i\in[m]$ is referred to as a \emph{robust infeasibility certificate}.
\subsection{Saddle Point Problems}\label{sec:SPprelim}
Saddle point (SP) problems play a vital role in our developments. In its most general form, a convex-concave SP problem is given by
\begin{equation}\label{eqn:SadVal}
\SadVal=\inf_{x \in X} \sup_{y \in Y} \phi(x,y), \tag{${\cal S}$}
\end{equation}
where the function $\phi(x,y)$ is convex in $x$ and concave in $y$ and the domains $X,Y$ are nonempty closed convex sets in Euclidean spaces ${\mathbb{E}}_x,{\mathbb{E}}_y$.
Any convex-concave SP problem \eqref{eqn:SadVal} gives rise to two convex optimization problems that are dual to each other:
\begin{equation*}\label{neq1}
\begin{array}{rclcr}
\Opt(P)&=&\inf_{x\in X}[ \overline{\phi}(x):=\sup_{y\in Y} \phi(x,y)]&&(P)\\
\Opt(D)&=&\sup_{y\in Y}[ \underline{\phi}(y):=\inf_{x\in X} \phi(x,y)] &&(D)\\
\end{array}
\end{equation*}
with $\Opt(P)=\Opt(D)=\SadVal$.
It is well-known that the solutions to \eqref{eqn:SadVal} --- the saddle points of $\phi$ on $X\times Y$ --- are exactly the pairs $[x;y]$ formed by optimal solutions to the problems $(P)$ and $(D)$.
We quantify the accuracy of a candidate solution $[\bar{x},\bar{y}]$ to SP problem \eqref{eqn:SadVal} with the \emph{saddle point gap} given by
\begin{equation}\label{eqn:SPgap}
\epsilon_{\sad}^\phi(\bar{x},\bar{y}):=\overline{\phi}(\bar{x})-\underline{\phi}(\bar{y})
=\underbrace{\left[\overline{\phi}(\bar{x})-\Opt(P)\right]}_{\geq0}+
\underbrace{\left[\Opt(D)-\underline{\phi}(\bar{y})\right]}_{\geq0}.
\end{equation}
Because convex-concave SP problems are simply convex optimization problems, they can in principle be solved by polynomial-time interior point methods (IPMs). However, the computational complexity of such methods depends heavily on the dimension of the problem. Thus, scalability of resulting algorithms becomes an issue in large-scale applications.
As a result, for large-scale SP problems, one has to resort to first-order subgradient-type methods.
On a positive note, there are many efficient first-order methods (FOMs) for convex-concave SP problems. These in particular include Nesterov's accelerated gradient descent algorithm \cite{Nesterov2005} and Nemirovski's Mirror-Prox algorithm \cite{Nemirovski2005}, both of which bound the saddle point gap at a rate of $\epsilon_{\sad}^\phi(\bar{x}_T,\bar{u}_T) \leq O\left({1\over T}\right)$ where $\bar{x}_T,\bar{u}_T$ are solutions obtained after $T$ iterations.
\input{OnlineProblemsSummary}
\section{Numerical Study}\label{sec:numerical}
In this section, we conduct a numerical study comparing the approaches discussed so far. We consider the following quadratic program inspired by mean-variance portfolio optimization problems with a factor model for the return vector (see, e.g., \cite{GoldfarbIyengar2003}):
\begin{equation}\label{eqn:port-opt}
\min_{x} \left\{ \| V x\|_2^2 + x^\top D x - \lambda \mu^\top x :~
x \in \Delta_n
\right\} ,
\end{equation}
where $\mu\in{\mathbb{R}}^n$ is the expected return vector, the term $x^\top (V^\top V+D) x$ captures the risk associated with the portfolio via a factor model, and $\lambda \geq 0$ represents the trade-off between the expected return of the portfolio and the risk associated with the portfolio.
In the robust formulation of \eqref{eqn:port-opt}, we consider the case where the true parameters $\mu \in \mathbb{R}^n$ and $V \in \mathbb{R}^{m \times n}$ belong to uncertainty sets ${\cal M}$ and ${\cal V}$ of form
\[
\mathcal{M} := \left\{ \mu :~ \mu_0-\gamma \leq \mu,~ \mu \leq \mu_0 + \gamma \right\},
\quad \mathcal{V} := \left\{ V = V_0 + \sum_{k=1}^K P_k u_k :~ \|u\|_2 \leq 1 \right\},
\]
where the nominal data $\mu_0 \in {\mathbb{R}}^n$, $\gamma\in{\mathbb{R}}^n$, and $V_0 \in {\mathbb{R}}^{m \times n}, \{P_k \in {\mathbb{R}}^{m \times n}\}_{k=1}^K$ are given to us. Then the robust problem is given by
\begin{equation}\label{eqn:robust-port-opt}
\min_{x} \left\{ \max_{V \in \mathcal{V}} \|Vx\|_2^2 + x^\top D x - \lambda \min_{\mu \in \mathcal{M}} \mu^\top x :~
x \in \Delta_n
\right\}.
\end{equation}
Our test instances are synthetically generated, largely following the random instance generation model from \cite{GoldfarbIyengar2003}. We begin by specifying three parameters: $n$, the number of variables; $m$, the number of factors (which controls the rank of $V$); and $\alpha \in (0,1)$, a parameter controlling the size of the uncertainty sets. For each instance, we randomly generate matrices $V \in \mathbb{R}^{m \times n}$ and $F \in \mathbb{R}^{m \times m}$, where we ensure $F$ is positive semidefinite, and define $D = 0.1 \Diag(V^\top F V)$. We then generate $p > m$ factor samples $f_{(l)} \in \mathbb{R}^m$, $l \in [p]$, where each $f_{(l)} \sim N(0,F)$, and we also generate $\mu \in \mathbb{R}^n$ where each entry $\mu_i \sim U(1,5)$. We then set $\mu_{(l)} = \mu + V^\top f_{(l)} + \epsilon_l$, where $\epsilon_{(l)} \sim N(0,D)$ are independent of the factor sample $f_{(l)}$. The matrices $\mu$ and $V$ are estimated via linear regression on $\mu_{(l)}$ and $f_{(l)}$, to obtain $\bar{\mu}, \bar{V}$. The nominal data for \eqref{eqn:port-opt} are set to be $\mu_0 = \bar{\mu}$, $V_0 = F^{1/2} \bar{V}$. To define the uncertainty sets, we first compute the scaled sum of squared errors for each $i \in [n]$, $s_i^2 = \frac{1}{p-m-1} \sum_{l=1}^{p} (\mu_{(l),i} - \mu_{0,i} - V_{0,i}^\top f_{(l)})^2$. Let $c_J(\alpha)$ be the $\alpha$-critical value of an $F$-distribution with $J$ degrees of freedom, and let $\nu$ be the top-left entry of $A^{-1}$, where $A \in \mathbb{R}^{(p+1) \times (p+1)}$ is the Gram matrix of the vectors $\bm{1}_m, \{f_{(l)}\}_{l=1}^p$. Then we set $\gamma_i = \sqrt{\nu c_1(\alpha) s_i^2}$ for $i\in[n]$, which defines the uncertainty set for $\mu$. The uncertainty set for $V$ is chosen by randomly generating matrices $P_k$, and then scaling them appropriately so that the norm of each column $i$ of $V-V_0$ is at most $\sqrt{m c_m(\alpha) s_i^2}$ for every $V \in \mathcal{V}$.
We set $p=90$ and $\alpha = 0.95$, while varying $m\in\{3,5,7,10,15,20,25\}$ and $n\in\{100,200,300,400,500,600,700\}$. We fix the underlying dimension of the uncertainty set ${\cal V}$ to be $K = \min\{2m,15\}$. We generate five instances for each combination of $m$ and $n$.
The four approaches we test are our OFO-based approach from Section \ref{sec:RegretforRO}, our FO-based pessimization approach from Section \ref{sec:pessimization-oracle} (see Theorem~\ref{thm:pessimization-oracle-approach}), the nominal oracle-based approach of \cite{BenTalHazan2015} from Section \ref{sec:oracle-based-approach}, and the full pessimization approach of \cite{MutapcicBoyd2009}, which requires both a pessimization and an extended nominal feasiblity oracle. Since \eqref{eqn:robust-port-opt} is an instance of a robust quadratic program, the form for nominal and pessimization oracles can be derived from Section \ref{sec:Application}. One-dimensional line search using Brent's algorithm \cite{Brent1973} was used to choose step sizes for each iteration of FO-based methods. An error tolerance of $\epsilon=0.002$ is used in all instances.
Experiments are performed on a Linux machine with 2.8GHz processor and 64GB memory using Python v3.5.2. Whenever the nominal (extended nominal) oracles and pessimization oracles do not have closed form solutions, they are implemented in Gurobi v7.0.2. We use standard Gurobi tolerances and parameter choices. We employ the implementation of Brent's algorithm in Python's \texttt{scipy.optimize} package.
Figure \ref{fig:average-solve-times} plots the average solve times in seconds against different $n$ for each of the approaches, averaging across all $m$. As we expect, for low dimensions $n$, the oracle-based approaches solve the instances very quickly compared to our first-order based approaches. However, as $n$ increases to $400,500,600,700$, we see that the solution times of our first-order based approaches beat the nominal oracle approach, and become comparable to the full pessimization approach for $n=600,700$. In particular, we observe that the FO-based pessimization approach (see Theorem~\ref{thm:pessimization-oracle-approach}) solves faster when $n=700$.
\begin{figure}[h]
\centering
\includegraphics[width=100mm]{avg_solve_times.pdf}
\vspace{-10pt}
\caption{Average solve times (seconds) for different $n$.
}
\label{fig:average-solve-times}
\end{figure}
The dimension $m$ influences the rank of the nominal matrix $V_0^\top V_0$ and controls the difficulty of the problems. Examining the results for different $m$ further highlights the benefits of utilizing the first-order based approaches. Figure \ref{fig:average-solve-times-plots} plots average solve times for different $m$ while fixing $n = 400,500,600,700$. For the oracle-based methods, the solution times increase with $m$, while the solution times for first-order based methods remains relatively constant with $m$. For $m \geq 20$, we observe that our first-order based approaches significantly outperforms the oracle-based methods which require a nominal solver.
\begin{figure}[h]
\centering
\subfigure[$n=400$]{
\includegraphics[width=78mm]{avg_solve_times_n400.pdf}
}
\subfigure[$n=500$]{
\includegraphics[width=78mm]{avg_solve_times_n500.pdf}
}
\subfigure[$n=600$]{
\includegraphics[width=78mm]{avg_solve_times_n600.pdf}
}
\subfigure[$n=700$]{
\includegraphics[width=78mm]{avg_solve_times_n700.pdf}
}
\caption{Average solve times (seconds) for different $n$ and $m$.}
\label{fig:average-solve-times-plots}
\end{figure}
Notice that, while we expect our OFO-based approach to outperform the FO-based pessimization approach due to the burden of solving an eigenvalue problem in each iteration for computing the pessimization oracle, our results indicate the opposite. This is because for small values of $K$, calling a pessimization oracle is faster than the line search performed in the FO-based noise update. However, we believe that as $K$ increases, one-dimensional line search will become more efficient.
Finally, we examine the number of iterations and cost per iteration of different approaches averaged across all instances in Table \ref{tab:method-iters-summary}. We observe that, contrary to their theoretical iteration guarantees, the oracle-based approaches of \cite{MutapcicBoyd2009,BenTalHazan2015} need very few iterations to find a solution. However, as expected, the average time per iteration is significantly higher for these methods due to their reliance on full nominal optimization solvers.
This further highlights the benefit of utilizing first-order methods for robust optimization when the deterministic version of the problem is already very expensive, and hence nominal oracles become expensive.
\begin{table}[h!tb]
\caption
{Average number of iterations and average time per iteration for each approach.
\label{tab:method-iters-summary}}
{\vspace{-10pt}
\begin{center}
\begin{tabular}{l|rc}
\toprule
& \# iterations & seconds per iteration \\ \hline
first-order & 961.487 & 0.015 \\
FO-based pessimization & 1009.054 & 0.013 \\
nominal & 3.708 & 4.841 \\
full pessimization & 1.975 & 4.875 \\
\bottomrule
\end{tabular}
\end{center}
}
{}
\end{table}
\section*{Acknowledgments}
This research is supported in part by NSF grant CMMI 1454548.
The authors would like to thank Arkadi Nemirovski for suggesting the convex-concave reformulation of the robust feasibility problem presented in Section~\ref{sec:concaveSP}, in particular Lemma~\ref{lem:concaveSPReformulation}. The authors also would like to thank to the review team for useful suggestions and feedback that improved the presentation of the material in this paper.
\input{OnlineFOMforRO_Web.bbl}
\newpage
\subsection{Online Convex Optimization Tools}\label{sec:OCO}
Our efficient framework for RO employs tools from the online convex optimization domain. We now briefly outline these and refer to \cite{CesaBianchiLugosi2006,Hazan2011,Shalev-Shwarz2011} for further details and applications of OCO.
OCO is used to capture decision making in dynamic environments. We are given a finite time horizon $T$, closed, bounded, and convex domain $Z$, and in each time period $t \in [T]$, a convex loss function $f_t:Z \to {\mathbb{R}}$ is revealed. At time periods $t \in [T]$ we must choose a decision $z_t \in Z$, and based on this we suffer a loss of $f_t(z_t)$ and receive some feedback typically in the form of first-order information on $f_t$. Our goal is to minimize the \emph{weighted regret}
\begin{equation}\label{eqn:weighted-regret-defn}
\sum_{t=1}^{T} \theta_t f_t(z_t) - \inf_{z\in Z} \sum_{t=1}^{T} \theta_t f_t(z),
\end{equation}
where $\theta \in \Delta_T$ is a vector of convex combination weights.\footnote{Note that in the OCO literature, regret is usually defined with uniform weights $\theta_t = 1/T$. Nonuniform weights introduce flexibility to our framework by allowing selection of specific customization of OCO algorithms for exploiting structural properties of the constraint functions $f^i$ to achieve better convergence rates. A prime example for this is when the functions are strongly convex.}
Most OCO algorithms are closely related to offline iterative FOMs. In this paper, we will make use of the proximal setup of \cite{JuditNem2012Pt1} to choose the sequence $\{z_t\}_{t=1}^T$ which ensures that the weighted regret \eqref{eqn:weighted-regret-defn} converges to $0$ as $T \to \infty$. Thus, we make the following assumption on $Z$ for the existence of a proximal setup.
\begin{assumption}\label{ass:proximal-setup-OFO}
Let $\mathbb{E}_z$ be the Euclidean space containing $Z$. There exists a norm $\|\cdot\|$ and its dual norm $\|\cdot\|_*$ on $\mathbb{E}_z$, a distance-generating function $\omega:Z \to {\mathbb{R}}$ which is $1$-strongly convex with respect to $\|\cdot\|$ and leads to an easy-to-compute prox function $\Prox_z(\xi):=\mathop{\rm argmin}_{w\in Z} \left\{\langle \xi,w\rangle+\omega(w)-\langle\omega'(z),w-z\rangle\right\}$ and set width $\Omega := \max_{z \in Z} \omega(z) - \min_{z \in Z} \omega(z)$ which is finite when $Z$ is bounded.
\end{assumption}
The proximal setup of Assumption~\ref{ass:proximal-setup-OFO} allows us to adjust to the geometry of domain $Z$. The standard basic domains satisfying Assumption~\ref{ass:proximal-setup-OFO} include simplex, Euclidean ball, and spectahedron; see \cite[Section 1.7]{JuditNem2012Pt1} for the standard proximal setups (i.e., Assumption~\ref{ass:proximal-setup-OFO}) for these basic domains in terms of selection of $\|\cdot\|$ and resulting $\omega$, $\Prox$ computation, and set width $\Omega$.
Under Assumption~\ref{ass:proximal-setup-OFO} and various structural properties, the straightforward extension of the standard online mirror descent algorithm (see, e.g., \cite{KakadeSSTewari2012}) from uniform weigths to weighted regret achieves the following convergence rate.
\begin{theorem}[{\cite[Theorem 5]{KakadeSSTewari2012}}]\label{thm:OCO-non-smooth}
Suppose there exists $G \in (0,\infty)$ such that $\|\ensuremath{\nabla} f_t(z)\|_* \leq G$ for all $z \in Z$, $t \in [T]$. Define $\gamma = \sqrt{\frac{2 \Omega}{G^2 \sum_{t=1}^T \theta_t^2}}$. Choose $z_1 = \mathop{\rm argmin}_{z \in Z} \omega(z)$ and $z_{t+1} = \Prox_{z_t} (\gamma \theta_t \ensuremath{\nabla} f_t(z_t))$ for $t\in[T]$. Then
\[ \sum_{t=1}^{T} \theta_t f_t(z_t) - \inf_{z\in Z} \sum_{t=1}^{T} \theta_t f_t(z) \leq \sqrt{2 \Omega G^2 \sum_{t=1}^T \theta_t^2}. \]
In particular, for uniform weights $\theta_t = 1/T$, the upper-bound becomes $O(1/\sqrt{T})$.
\end{theorem}
We refer to \cite{KakadeSSTewari2012} for details of the proof. When $\omega(z) = z^\top z/2$ and weights $\theta_t=1/T$ for $t\in[T]$, the update rule $z_{t+1} = \Prox_{z_t} (\gamma \ensuremath{\nabla} f_t(z_t))$ becomes simply gradient descent, and Theorem~\ref{thm:OCO-non-smooth} reduces to the standard bound of online gradient descent from \cite{Zinkevich2003}.
\section{General Framework for Robust Feasibility Problems}\label{sec:SPFormulation}
In this section, we build a general framework to solve the robust feasibility problem \eqref{eqn:approx-robust-feas} by working with its natural saddle point formulation.
Given constraint functions $f^i(x,u^i)$, $i \in [m]$, let us define $\Phi(x,u):= \max_{i\in[m]} f^i(x,u^i)$. Then $\Phi(x,u)$ is a convex function of $x$, but not necessarily concave in $u$. In addition, with this definition of $\Phi(\cdot)$, the robust approximate feasibility problem~\eqref{eqn:approx-robust-feas} is equivalent to simply verifying
\begin{equation}\label{eqn:convex-nonconcaveSP}
\text{either}\quad \inf_{x \in X} \sup_{u \in U} \Phi(x,u) = \inf_{x \in X} \max_{i\in[m]} \sup_{u^i \in U^i} f^i(x,u^i) \leq \epsilon \quad \text{or} \quad \inf_{x \in X} \sup_{u \in U} \Phi(x,u) > 0,
\end{equation}
which is nothing but solving a specific SP problem and checking its value. Analogous to the convex-concave SP gap \eqref{eqn:SPgap}, for a given solution $[\bar{x},\bar{u}]$, we define the SP gap of problem \eqref{eqn:convex-nonconcaveSP} as
\[
\epsilon_{\sad}^\Phi(\bar{x},\bar{u}):=\overline{\Phi}(\bar{x})-\underline{\Phi}(\bar{u})
=\sup_{u\in U} \Phi(\bar{x},u) - \inf_{x\in X} \Phi(x,\bar{u}).
\]
In general, solving a convex-nonconcave SP problem of form~\eqref{eqn:convex-nonconcaveSP}, i.e., finding a solution $[\bar{x},\bar{u}]$ such that $\epsilon_{\sad}^\Phi(\bar{x},\bar{u})\leq\epsilon$, can be difficult. That said, a bound on the SP gap $\epsilon_{\sad}^\Phi(\bar{x},\bar{u})$ along with the value of $\Phi(\bar{x},\bar{u})$ leads to robust feasibility certificates for \eqref{eqn:convex-nonconcaveSP} as follows.
\begin{theorem}\label{thm:saddlept-certificate}
Let $\Psi:X \times U \to {\mathbb{R}}$ be a given function associated with a SP (not necessarily admitting a convex-concave structure). Suppose we have $\bar{x} \in X$, $\bar{u} \in U$, and $\tau\in(0,1)$ such that $\epsilon_{\sad}^\Psi(\bar{x},\bar{u})\leq \tau\epsilon$.
Then if $\Psi(\bar{x},\bar{u}) \leq (1-\tau)\epsilon$, we have $\sup_{u \in U} \Psi(\bar{x},u) \leq \epsilon$. Moreover, if $\Psi(\bar{x},\bar{u}) > (1-\tau)\epsilon$ and $\tau\leq{1\over 2}$, we have $\inf_{x \in X} \Psi(x,\bar{u}) > 0$.
\end{theorem}
{\emph{#1}}{Proof.}
Suppose $\Psi(\bar{x},\bar{v}) \leq (1-\tau)\epsilon$. Because $\epsilon_{\sad}^\Psi(\bar{x},\bar{v}) =\sup_{u \in U} \Psi(\bar{x},u) - \inf_{x\in X} \Psi(x,\bar{u}) \leq \tau\epsilon$, we have $\sup_{u \in U} \Psi(\bar{x},u) \leq \inf_{x\in X} \Psi(x,\bar{u}) + \tau\epsilon \leq \Psi(\bar{x},\bar{u}) + \tau\epsilon \leq \epsilon$. On the other hand, when $\Psi(\bar{x},\bar{u}) > (1-\tau)\epsilon$, we have $(1-\tau)\epsilon < \Psi(\bar{x},\bar{u}) \leq \sup_{u \in U} \Psi(\bar{x},u) \leq \inf_{x \in X} \Psi(x,\bar{u}) + \tau\epsilon$, which implies $\inf_{x\in X} \sup_{u\in U} \Psi(x,u) \geq \inf_{x\in X} \Psi(x,\bar{u}) > (1-2\tau)\epsilon\geq 0$ when $\tau\leq{1\over 2}$.
\begin{remark}\label{rem:SPm=1}
When $m=1$, $\Phi(x,u)=f^1(x,u^1)$, and it is thus convex in $x$ and concave in $u$ due to Assumption~\ref{ass:f-concave-u}. Therefore, in the case of a single robust constraint, i.e., $m=1$, under Assumption~\ref{ass:f-concave-u} and assuming $U=U^1$ is a closed convex set, the optimization problem in \eqref{eqn:convex-nonconcaveSP} reduces to a standard convex-concave SP problem.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
While it is not very common, a few robust convex optimization problems come with a single robust constraint and convex uncertainty set $U$; see for example \cite{Ben-TalRobustSVM2012} for a robust version of a SVM problem with one constraint. In such cases, based on Remark~\ref{rem:SPm=1}, the resulting convex-concave SP problems can directly be solved via efficient FOMs. On the other hand, in the presence of multiple constraints, the function $\Phi(x,u)$ is not concave in $u = [u^1;\ldots;u^m]$ even under Assumption~\ref{ass:f-concave-u}. Nevertheless, when $m>1$, it is still possible to have a convex-concave SP reformulation of the optimization problem in \eqref{eqn:convex-nonconcaveSP} in an extended space via perspective transformations, which we present in Appendix~\ref{sec:concaveSP}. While this reformulation has the benefit of reducing the robust feasibility problem to a well-known and well-studied problem, it destroys the simplicity of the original domains and constraint functions and hence comes with some challenges. Therefore, we develop a framework where we work directly with the convex-nonconcave SP formulation in \eqref{eqn:convex-nonconcaveSP} in the space of original variables.
Moreover, because we work in the original space of variables, we simply utilize the first-order information on the original constraint functions $f^i$ and original domains $X$ and $U^i$. This direct approach in particular allows us to take greater advantage of the structure of the original formulation such as the availability of efficient projection (prox) computations over domains $X,\, U^i$, and/or better parameters for smoothness, Lipschitz continuity, etc., of the functions $f^i$.
Because $\Phi(x,u)$ is not concave in $u$, we cannot bound the SP gap $\epsilon_{\sad}^\Phi(\bar{x},\bar{u})$ by using traditional FOMs designed for solving convex-concave SP problems. However, we next show that by just \emph{partially} upper bounding $\epsilon_{\sad}^\Phi(\bar{x},\bar{u})$, we can derive a general iterative framework to obtain robust feasibility/infeasibility certificates. We describe the further specifics of this framework in Section~\ref{sec:OCOforRO}.
Henceforth we will no longer use the shorthand notation $\Phi(x,u) = \max_{i \in [m]} f^i(x,u^i)$, but we will denote the SP gap $\epsilon_{\sad}^\Phi(\bar{x},\bar{u})$ as
\begin{equation}\label{eqn:SVgap}
\epsilon(\bar{x},\bar{u}) := \epsilon_{\sad}^\Phi(\bar{x},\bar{u}) = \max_{i\in[m]} \sup_{u^i \in U^i} f^i(\bar{x},u^i) - \inf_{x\in X} \max_{i\in[m]} f^i(x,\bar{u}^i).
\end{equation}
The robust feasibility certificate result from Theorem~\ref{thm:saddlept-certificate} indicates the importance of bounding the SP gap $\epsilon(\bar{x},\bar{u})$. Often, FOMs achieve this by iteratively generating points $x_t \in X$, $u_t \in U$ for $t\in [T]$ and tracking the points $\bar{x}$ and $\bar{u}$ obtained from a convex combination of $\{x_t,u_t\}_{t=1}^T$. In order to simplify our notation, given convex combination weights $\theta\in\Delta_T$ and points $\{x_t,u_t\}_{t=1}^T$, we let
\[ \bar{x}_T := \sum_{t=1}^{T} \theta_t x_t \quad\text{and}\quad \bar{u}_T := \sum_{t=1}^{T} \theta_t u_t. \]
We now present an upper bound on $\epsilon(\bar{x}_T,\bar{u}_T)$ that follows naturally from the convex-concave structure of functions $f^i$.
To this end, given a set of vectors $y_t\in\Delta_m$ for $t\in[T]$, we also define
\begin{align*}
\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) &:= \max_{i\in[m]} \left\{\sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)\right\},\quad\text{and}\\
\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T) &:= \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) - \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i),
\end{align*}
together with
\[
\widehat{\epsilon}(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T) := \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i) - \inf_{x\in X} \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x,u_t^i).
\]
Our next result relates these quantities to the value of the SP gap function $\epsilon\left(\bar{x}_T, \bar{u}_T\right)$.
\begin{proposition}\label{prop:gap-nonconcave}
Let $x_t \in X$ and $u_t \in U$ for $t\in[T]$ be given a set of vectors. Then for any set of vectors $y_t\in\Delta_m$ for $t\in[T]$ and any $\theta\in\Delta_T$, we have
\begin{equation}\label{eqn:saddlept-gap-bound}
\epsilon\left(\sum_{t=1}^{T} \theta_t x_t, \sum_{t=1}^{T} \theta_t u_t\right) \leq \epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) +\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T) + \widehat{\epsilon}(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T).
\end{equation}
\end{proposition}
{\emph{#1}}{Proof.}
Given $y_t\in\Delta_m$ for $t\in[T]$ and $\theta\in\Delta_T$, let us define $\bar{x} := \sum_{t=1}^{T} \theta_t x_t$ and $\bar{u} := \sum_{t=1}^{T} \theta_t u_t$. We first partition $\epsilon(\bar{x},\bar{u})$ as $\epsilon(\bar{x},\bar{u})=\overline{\epsilon}(\bar{x})+\underline{\epsilon}(\bar{u})$ where
\begin{align*}
\overline{\epsilon}(\bar{x}) &:=\max_{i\in[m]} \sup_{u^i\in U^i} f^i(\bar{x},u^i) - \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) , \\
\underline{\epsilon}(\bar{u}) &:= \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) - \inf_{x\in X} \max_{i\in[m]} f^i(x,\bar{u}^i),
\end{align*}
and then derive upper bounds on $\overline{\epsilon}(\bar{x})$ and $\underline{\epsilon}(\bar{u})$.
We start with bounding $\overline{\epsilon}(\bar{x})$. Because the functions $f^i(x,u^i)$ are convex in $x$ for all $i$ and $\theta\in\Delta_T$, we have $\max_{i\in[m]} \sup_{u^i\in U^i} f^i(\bar{x},u^i) \leq \max_{i\in[m]} \sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) $. Therefore,
{\small
\begin{align}
\overline{\epsilon}(\bar{x}) &= \max_{i\in[m]} \sup_{u^i\in U^i} f^i(\bar{x},u^i)
- \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) \nonumber\\
&\leq \max_{i\in[m]} \sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) + \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) - \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i)\nonumber\\
&\leq \max_{i\in[m]} \left\{\sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)\right\} + \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) - \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) \label{eqn:epsOver},
\end{align}}
where the last inequality follows since $\max_{i\in[m]} \{\alpha_i-\beta_i\} \geq \max_{i\in[m]}\alpha_i - \max_{i\in[m]} \beta_i$ for any sequence of numbers $\alpha_i,\beta_i$, $i \in [m]$.
Note that $ \inf_{x\in X} \max_{i\in[m]} f^i(x,u^i) \geq \inf_{x\in X} \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x,u^i)$ because under Assumption~\ref{ass:f-concave-u} the functions $f^i(x,u^i)$ are concave in $u^i$ for all $i$.
Thus, we arrive at
\begin{align}
\underline{\epsilon}(\bar{u}) &= \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) - \inf_{x\in X} \max_{i\in[m]} f^i(x,\bar{u}^i) \nonumber\\
&\leq \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) - \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i) + \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i) \nonumber\\
&\qquad\qquad - \inf_{x\in X} \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x,u_t^i) \nonumber\\
&= \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i) - \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i) + \widehat{\epsilon}(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T). \label{eqn:epsUnder}
\end{align}
Then by summing \eqref{eqn:epsOver} and \eqref{eqn:epsUnder} and rearranging the terms, we deduce the result.
We are now ready to state our main result. This is analogous to Theorem~\ref{thm:saddlept-certificate} except that we do not need to bound all three terms in \eqref{eqn:saddlept-gap-bound}, but instead it suffices to guarantee that
\[ \epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) +\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T) \leq \epsilon. \]
We show that when the above condition holds, based on the value of $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)$ we can then obtain either a robust $\epsilon$-feasible solution, or an infeasibility certificate.
\begin{theorem}\label{thm:robust-feas-oracle}
Suppose we have sequences $\{x_t,u_t,y_t,\theta_t\}_{t=1}^T$ with $x_t\in X$, $u_t\in U$, $y_t\in\Delta_m$ for all $t\in[T]$, $\theta\in\Delta_T$. Let $\tau \in (0,1)$.
If $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \tau \epsilon$ and $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)\leq (1-\tau) \epsilon$, then the solution $\bar{x}_T := \sum_{t=1}^{T} \theta_t x_t$ is $\epsilon$-feasible with respect to \eqref{eqn:approx-robust-feas}.
If $\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T)\leq (1-\tau) \epsilon$ and $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$, then \eqref{eqn:approx-robust-feas} is infeasible.
\end{theorem}
{\emph{#1}}{Proof.}
First suppose there exists a $\tau\in(0,1)$ and corresponding vectors $\{x_t,u_t,y_t,\theta_t\}_{t=1}^T$ such that $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \tau \epsilon$ and $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$ holds as well. Note that
\begin{align}
\tau \epsilon &\geq \epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)
=\max_{i\in[m]} \left\{\sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)\right\}\nonumber\\
&\geq \max_{i\in[m]} \sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \label{eqn:thm:feas} ,
\end{align}
where the last inequality follows since $\max_{i\in[m]} \{\alpha_i-\beta_i\} \geq \max_{i\in[m]}\alpha_i - \max_{i\in[m]} \beta_i$ for any sequence of numbers $\alpha_i,\beta_i$, $i \in [m]$. Then $\bar{x}_T$ is an $\epsilon$-feasible solution for \eqref{eqn:approx-robust-feas} because
\begin{align*}
\max_{i\in[m]} \sup_{u^i\in U^i} f^i\left(\bar{x}_T,u^i\right) &= \max_{i\in[m]} \sup_{u^i\in U^i} f^i\left(\sum_{t=1}^{T} \theta_t x_t,u^i\right) \\
&\leq \max_{i\in[m]} \sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) \leq \tau \epsilon + \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \leq \epsilon,
\end{align*}
where the first inequality follows from the convexity of the functions $f^i$ and the fact that $\theta\in\Delta_T$, the second inequality from \eqref{eqn:thm:feas}, and the last inequality holds since $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$.
On the other hand, suppose $\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T)\leq (1-\tau) \epsilon$ and $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$. Note that
\begin{align}\label{eqn:epsPos}
\inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i) &\leq \inf_{x\in X} \sum_{t=1}^{T} \theta_t \max_{i\in[m]} f^i(x,u_t^i) \nonumber\\
&\leq \inf_{x\in X} \sum_{t=1}^{T} \theta_t \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i)
= \inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i),
\end{align}
where the first inequality follows since $y_t\in\Delta_m$ for all $t\in[T]$, the second inequality holds because $f^i(x,u_t^i) \leq\sup_{u^i\in U^i} f^i(x,u^i)$ for all $i\in[m]$ and $y_t^{(i)} \geq 0$ for $i \in [m]$, $t \in [T]$, and the last equation follows from $\theta\in\Delta_T$. Then using the bound
\begin{equation}\label{eqn:epsbullet-bound}
(1-\tau) \epsilon \geq \epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T)
=\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) - \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i),
\end{equation}
we arrive at
\[
\inf_{x\in X} \max_{i\in[m]} \sup_{u^i\in U^i} f^i(x,u^i)
\geq \inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} y^{(i)}_t f^i(x,u_t^i)
\geq \max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) - (1-\tau) \epsilon
>0,
\]
where the first inequality follows from inequality~\eqref{eqn:epsPos}, the second inequality from \eqref{eqn:epsbullet-bound} and the last inequality holds because $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$. This implies \eqref{eqn:approx-robust-feas} is infeasible.
In Section \ref{sec:RegretforRO} we will show that $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ can be interpreted as a weighted regret term \eqref{eqn:weighted-regret-defn}.
On the other hand, the term $\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T)$ has no such direct interpretation. In order to upper-bound it by a weighted regret term, we need the following result.
\begin{corollary}\label{cor:onlineSP-upperbound}
Given sequences $\{x_t,u_t,\theta_t\}_{t=1}^T$ with $x_t\in X$, $u_t\in U$, for all $t\in[T]$, $\theta\in\Delta_T$, there is an appropriate choice of sequence $\{\bar y_t\}_{t=1}^T$ where $\bar y_t\in\Delta_m$ for all $t\in[T]$, such that $\epsilon^\bullet(\{x_t,u_t,\bar y_t,\theta_t\}_{t=1}^T)$ is upper-bounded by
\begin{equation}\label{eqn:onlineSPgap-upperbound}
\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) := \sum_{t=1}^{T} \theta_t \max_{i \in [m]} f^i(x_t,u_t) - \inf_{x \in X} \sum_{t=1}^T \theta_t \max_{i \in [m]} f^i(x,u_t).
\end{equation}
Thus, if $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)\leq (1-\tau) \epsilon$ and $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$, then \eqref{eqn:approx-robust-feas} is infeasible.
\end{corollary}
{\emph{#1}}{Proof.}
Given $\{u_t\}_{t=1}^T$, let $x^* \in \mathop{\rm argmin}_{x \in X} \sum_{t=1}^T \theta_t \max_{i \in [m]} f^i(x,u_t^i)$ and for all $t\in[T]$ define $\bar y_t\in{\mathbb{R}}^m$ to be the $i$-th unit vector where $i$ is the smallest index satisfying $i \in {\mathop{\rm argmax}}_{i' \in [m]} f^i(x^*,u_t^{i'})$. Then $\inf_{x\in X} \sum_{t=1}^{T} \theta_t \sum_{i=1}^{m} \bar y^{(i)}_t f^i(x,u_t^i) = \inf_{x \in X} \sum_{t=1}^T \theta_t \max_{i \in [m]} f^i(x,u_t^i)$, and the bound follows from
$\max_{i \in [m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t) \leq \sum_{t=1}^{T} \theta_t \max_{i \in [m]} f^i(x_t,u_t)$. We deduce the last result from Theorem~\ref{thm:robust-feas-oracle}.
The following corollary demonstrates how we can choose $\tau$ in Theorem~\ref{thm:robust-feas-oracle}.
\begin{corollary}\label{cor:feasCertificates}
Suppose $\{x_t,u_t,y_t,\theta_t\}_{t=1}^T$ with $x_t\in X$, $u_t\in U$, $y_t\in\Delta_m$ for all $t\in[T]$, and $\theta\in\Delta_T$ is such that there exists $\kappa^\circ,\kappa^\bullet\in(0,1)$ satisfying $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \epsilon\,\kappa^\circ$ and $\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T)\leq \epsilon\,\kappa^\bullet$ with $\kappa^\circ+\kappa^\bullet\leq 1$. Let $\tau \in [\kappa^\circ,1-\kappa^\bullet]$.
Whenever $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)\leq (1-\tau) \epsilon$ as well, the solution $\bar{x}_T := \sum_{t=1}^{T} \theta_t x_t$ is $\epsilon$-feasible with respect to \eqref{eqn:approx-robust-feas}. Also, whenever $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$, then \eqref{eqn:approx-robust-feas} is infeasible.
\end{corollary}
{\emph{#1}}{Proof.}
Note that $\tau \in (0,1)$ follows from its definition, $\kappa^\circ,\kappa^\bullet\geq0$, and $\kappa^\circ+\kappa^\bullet\leq 1$. Furthermore, the interval $[\kappa^\circ,1-\kappa^\bullet]$ is well-defined since $\kappa^\circ \leq 1 - \kappa^\bullet$ always holds. Moreover, $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \epsilon\,\kappa^\circ \leq \epsilon\,\tau$ and $\epsilon^\bullet(\{x_t,u_t,y_t,\theta_t\}_{t=1}^T)\leq \epsilon\,\kappa^\bullet \leq \epsilon(1-\tau)$ holds from the definition of $\tau$. The result now follows from Theorem~\ref{thm:robust-feas-oracle}.
Theorem~\ref{thm:robust-feas-oracle} and Corollary~\ref{cor:onlineSP-upperbound} points to our general iterative framework for finding robust feasibility/infeasibility certificates of \eqref{eqn:approx-robust-feas}: generate sequences $\{x_t,u_t\}_{t=1}^T$ iteratively to bound $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ and $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$, and then evaluate the term $\max_{i \in [m]} \sum_{t=1}^T \theta_t f^i(x_t,u_t^i)$.
We provide a description of our framework in Algorithm \ref{alg:approx-robust-feas}. We assume that we have access to weights $\{\theta_t\}_{t=1}^T$ and update algorithms $\mathcal{A}_i$ and $\mathcal{A}_x$ for choosing $u_t^i \in U^i$ and $x_t \in X$ based on past observations $\{x_s,u_s\}_{s=1}^{t-1}$. We denote the updates by
\[
u_t^i = \mathcal{A}_i(\{x_s,u_s\}_{s=1}^{t-1}) \in U^i~\forall i\in[m], \quad x_t = \mathcal{A}_x(\{x_s,u_s\}_{s=1}^{t-1}) \in X, \]
and initializations $u_1^i = \mathcal{A}_i(\{\})~\forall i\in[m]$, $x_1 = \mathcal{A}_x(\{\})$. Moreover, we assume that these algorithms enjoy the following convergence guarantees: for any sequence $\{x_t\}_{t=1}^T$, let $u_t^i = \mathcal{A}_i(\{x_s,u_s\}_{s=1}^{t-1})~\forall i\in[m]$, then
\begin{equation}\label{eqn:noise-update-regret-bound}
\sup_{u^i\in U^i} \sum_{t=1}^T \theta_t f^i(x_t,u^i) - \sum_{t=1}^T \theta_t f^i(x_t,u_t^i) \leq \mathcal{R}_i(T);
\end{equation}
for any sequence $\{u_s\}_{s=1}^T$, let $x_t = \mathcal{A}_x(\{x_s,u_s\}_{s=1}^{t-1})$, then
\begin{equation}\label{eqn:solution-update-regret-bound}
\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) = \sum_{t=1}^T \theta_t \max_{i \in [m]} f^i(x_t,u_t^i) - \inf_{x\in X} \sum_{t=1}^T \theta_t \max_{i \in [m]} f^i(x,u_t^i) \leq \mathcal{R}_x(T).
\end{equation}
Explicit examples of $\mathcal{A}_i, \mathcal{A}_x$ and their bounds $\mathcal{R}_i, \mathcal{R}_x$ will be discussed in Section \ref{sec:OCOforRO}. Generally, we desire that the error bounds $\mathcal{R}_i(T), \mathcal{R}_x(T) \to 0$ as $T \to \infty$,
which can be achieved by using online mirror descent as in Theorem~\ref{thm:OCO-non-smooth}. That said, our OFO-based approach in Algorithm \ref{alg:approx-robust-feas} is quite flexible in terms of the selection of OFO algorithms $\mathcal{A}_i, \mathcal{A}_x$, and is certainly not restricted to only using online mirror descent.
\begin{algorithm}[ht!]
\caption{OFO-based approximate robust feasibility solver.}\label{alg:approx-robust-feas}
\begin{algorithmic}
{\small
\STATE {\bf input:~} update algorithms $\mathcal{A}_i$, $i \in [m]$, $\mathcal{A}_x$, tolerance level $\epsilon>0$, sufficiently large $T = T(\epsilon)$ such that $\max_{i \in [m]} \mathcal{R}_i(T) + \mathcal{R}_x(T) \leq \epsilon$, and convex combination weights $\theta_1,\ldots,\theta_T > 0$.
\STATE {\bf output:~} either $\bar{x} \in X$ such that $\sup_{u^i\in U^i} f^i(\bar{x},u^i) \leq \epsilon$ for all $i\in[m]$, or an infeasibility certificate for \eqref{eqn:approx-robust-feas}.
\STATE initialize $u_1^i = \mathcal{A}_i(\{\})$ for $i\in[m]$ and $x_1 = \mathcal{A}_x(\{\})$.
\FOR{$t=2,\ldots,T$}
\FOR{$i=1,\ldots,m$}
\STATE compute $u_t^i = \mathcal{A}_i(\{x_s,u_s\}_{s=1}^{t-1}) \in U^i$.
\ENDFOR
\STATE compute $x_t = \mathcal{A}_x(\{x_s,u_s\}_{s=1}^{t-1}) \in X$.
\STATE obtain upper bounds $\max_{i \in [m]} \mathcal{R}_i(t) \geq \epsilon^\circ(\{x_s,u_s,\theta_s\}_{s=1}^t)$ and $\mathcal{R}_x(t) \geq \epsilon^\bullet(\{x_s,u_s,\theta_s\}_{s=1}^t)$.
\STATE compute $\kappa_t^\circ = \max_{i \in [m]} \mathcal{R}_i(t)/\epsilon$, $\kappa_t^\bullet = \mathcal{R}_x(t)/\epsilon$.
\IF{$\kappa_t^\circ + \kappa_t^\bullet \leq 1$}
\STATE set $\vartheta_t := \max_{i\in[m]} \sum_{s=1}^{t} \theta_s f^i(x_s,u_s^i)$ and
$\tau_t := 1 - \kappa_t^\bullet$.
\LINEIF{$\vartheta_t > (1-\tau_t)\epsilon$}{\algorithmicreturn~ `infeasible'.}
\LINEIF{$\vartheta_t \leq (1-\tau_t) \epsilon$}{\algorithmicreturn~ $\bar{x}_t = \frac{1}{t} \sum_{s=1}^t x_s$ as a robust $\epsilon$-feasible solution to \eqref{eqn:approx-robust-feas}.}
\ENDIF
\ENDFOR
}
\end{algorithmic}
\end{algorithm}
\begin{remark}\label{rem:anticipatory-requirements}
Notice that in Algorithm \ref{alg:approx-robust-feas} we generate $u_t$ \emph{before} generating $x_t$. Thus, nothing stops us from choosing $x_t$ based on the knowledge of $u_t$, or vice versa. Indeed, the pessimization oracle approach of \cite{MutapcicBoyd2009} and the nominal feasibility oracle approach of \cite{BenTalHazan2015} fit within our framework if we rewrite Algorithm \ref{alg:approx-robust-feas} to reflect this, and it is a trivial matter to do so.
However, a conflict may arise if we encounter a situation where generating $x_t$ requires knowledge of $u_t$, and generating $u_t$ also requires knowledge of $x_t$. Thus, when selecting the update algorithms $\mathcal{A}_i$, $\mathcal{A}_x$, care must be taken to avoid such situations.
Our suggested OFO approach in Section \ref{sec:RegretforRO} utilizes Theorem~\ref{thm:OCO-non-smooth}.
Moreover, Theorem~\ref{thm:OCO-non-smooth} generates the current decision in a \emph{non-anticipatory manner} based on only the knowledge of $f_{t-1}$, and not of $f_t$.
That is, it ensures that we will only use $u_{t-1}$ to generate $x_t$, and similarly we only use $x_{t-1}$ to generate $u_t$, thus no conflicts will arise.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
\begin{remark}\label{rem:tauInfeas}
Note that Algorithm~\ref{alg:approx-robust-feas} chooses $\tau_t = 1 - \kappa_t^\bullet$, whereas Corollary~\ref{cor:feasCertificates} allows us to choose from a range $\tau_t \in [\kappa_t^\circ,1-\kappa_t^\bullet]$. This is because it is theoretically possible for \eqref{eqn:approx-robust-feas} to simultaneously be infeasible and robust $\epsilon$-feasible, but in practice we would like to discover infeasibility of \eqref{eqn:approx-robust-feas} rather than an approximately feasible solution. Then the best value for $\tau_t\in[\kappa_t^\circ,1-\kappa_t^\bullet]$ in detecting infeasibility of \eqref{eqn:approx-robust-feas} is given by $\tau_t = 1 - \kappa_t^\bullet$.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
In the next section, we describe some approaches to implement Algorithm \ref{alg:approx-robust-feas} in practice by providing explicit examples of $\mathcal{A}_i$ and $\mathcal{A}_x$.
\section{Customizations of the General Framework}\label{sec:OCOforRO}
In this section, we examine how to generate the sequences $\{x_t,u_t\}_{t=1}^T$ in practice. In Section~\ref{sec:RegretforRO}, we first interpret the terms $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ and $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ from Section~\ref{sec:SPFormulation} as weighted regret terms, which gives rise to our OFO-based approach. In Section \ref{sec:pessimization-oracle}, we modify the pessimization oracle-based approach of \cite{MutapcicBoyd2009} to solving \eqref{eqn:approx-robust-feas} within our framework. In Section \ref{sec:oracle-based-approach}, we examine the nominal feasibility oracle-based approach of \cite{BenTalHazan2015} within the context of our general framework. Finally, in Section~\ref{sec:RateDiscussion}, we summarize and compare the convergence rates achievable via various customizations of our framework using these different approaches.
\subsection{The OFO-based Approach}\label{sec:RegretforRO}
Let us first consider $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$. For any $i\in[m]$, given $x_t$, we define the function $f_t^i: U^i\to {\mathbb{R}}$ as $f_t^i(u^i) = -f^i(x_t,u^i)$. Then the function $f_t^i(u^i)$ is convex in $u^i$ under Assumption~\ref{ass:f-concave-u}, and the subterm of $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ given by
\begin{equation}\label{eqn:RO-weighted-regret}
\sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)
\end{equation}
is the weighted regret \eqref{eqn:weighted-regret-defn} corresponding to the sequence of functions $\{ f_t^i \}_{t=1}^T$. When the uncertainty sets $U^i$, $i \in [m]$ admit proximal setups as in Assumption~\ref{ass:proximal-setup-OFO}, Theorem~\ref{thm:OCO-non-smooth} from Section~\ref{sec:OCO} gives an efficient OFO algorithm for choosing $\{u_t^i\}_{i=1}^m$ to bound the regret subterms \eqref{eqn:RO-weighted-regret} with $O(1/\sqrt{T})$.
Therefore, by using the online mirror descent algorithm of Theorem \ref{thm:OCO-non-smooth} as $\mathcal{A}_i$ in the computation of our $u^i_t$, we guarantee that
\[ \epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) = \max_{i \in [m]} \left\{ \sup_{u^i\in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \right\} \leq \max_{i \in [m]} \mathcal{R}_i(T), \]
where $\mathcal{R}_i(T) = O(1/\sqrt{T})$ with uniform weights $\theta_t = 1/T$.
On the other hand, given $u_t^i\in U^i$ for $i\in[m]$, let us define $\varphi_t(x) := \max_{i \in [m]} f^i(x,u_t^i)$. Then $\varphi_t(x)$ is convex in $x$ over $X$ since the functions $f^i$ are convex in $x$ by Assumption~\ref{ass:f-concave-u}. We can then rewrite $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ as
\begin{align}\label{eqn:RO-onlineSPgap-upperbd}
\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) &= \sum_{t=1}^{T} \theta_t \max_{i \in [m]} f^i(x_t,u_t^i) - \inf_{x\in X} \sum_{t=1}^{T} \theta_t \max_{i \in [m]} f^i(x,u_t^i)\notag\\
&= \sum_{t=1}^{T} \theta_t \varphi_t(x_t) - \inf_{x\in X} \sum_{t=1}^{T} \theta_t \varphi_t(x).
\end{align}
Then $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ is also a weighted regret term \eqref{eqn:weighted-regret-defn} corresponding to the sequence of functions $\{\varphi_t\}_{t=1}^T$. When the domain $X$ admits a proximal setup as in Assumption~\ref{ass:proximal-setup-OFO}, Theorem~\ref{thm:OCO-non-smooth} again gives an efficient OFO algorithm for choosing $x_t$ to bound \eqref{eqn:RO-onlineSPgap-upperbd}.
Once again, we may choose $\mathcal{A}_x$ to be the online mirror descent, and get $\mathcal{R}_x(T) = O(1/\sqrt{T})$ with uniform weights $\theta_t = 1/T$.
Algorithm \ref{alg:approx-robust-feas} can now be employed, provided we choose $T = \Omega(1/\epsilon^2)$, to solve the robust feasibility problem \eqref{eqn:approx-robust-feas}. Since the online mirror descent algorithm of Theorem~\ref{thm:robust-feas-oracle} only uses first-order information, we can solve the robust feasibility problem \eqref{eqn:approx-robust-feas} while avoiding reliance on a pessimization oracle for $u$ as in \cite{MutapcicBoyd2009} or a nominal feasibility oracle for $x$ as in \cite{BenTalHazan2015}.
\subsection{The Pessimization Oracle-Based Approach}\label{sec:pessimization-oracle}
Mutapcic and Boyd \cite{MutapcicBoyd2009} generate solutions $x_t \in X$ at each iteration $t$ by solving an extended nominal problem
\begin{equation}\label{eqn:extended-nominal-solver}
\min_{x \in X} \left\{ f^0(x) :~ f^i(x,u^i) \leq 0, \ \forall u^i \in \hat{U}_{t-1}^i, \ i \in [m] \right\},
\end{equation}
where $\hat{U}_{t-1}^i \subset U^i$ are finite approximate uncertainty sets based on past noise realizations $\{u_{t'}^i\}_{i=1}^m$ for $t'\in[t-1]$. New noises $u_t$ are then generated by calling the \emph{pessimization oracles} on the current solution $x_t$. More precisely, given $x_t \in X$, the pessimization oracles solve $\sup_{u^i \in U^i} f^i(x_t,u^i)$ and return
\begin{equation}\label{eqn:pessimization-oracle}
u_t^i \in U^i \quad \text{s.t.} \quad f^i(x_t,u_t^i) \geq \sup_{u^i \in U^i} f^i(x_t,u^i) - \tau \epsilon.
\end{equation}
In terms of our framework of Algorithm~\ref{alg:approx-robust-feas}, the update policy of generating new noises $u_t$ in this approach of \cite{MutapcicBoyd2009} corresponds to selecting the algorithms $\mathcal{A}_x$ to be a extended nominal solver for \eqref{eqn:extended-nominal-solver} and the algorithms $\mathcal{A}_i$ to be pessimization oracles that solve \eqref{eqn:pessimization-oracle}. Note that computing $u_t^i$ requires knowledge of $x_t$ (see Remark \ref{rem:anticipatory-requirements}), and consequently the bound for the regret term \eqref{eqn:noise-update-regret-bound} is $\mathcal{R}_i(T) \leq \tau \epsilon$ for any $T$. We show this in the proof of Theorem \ref{thm:pessimization-oracle-approach}. If for all $i \in [m]$ we have $f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$, then we terminate and declare $x_t$ is a robust $\epsilon$-feasible and optimal solution; otherwise, we append $\hat{U}_t^i = \hat{U}_{t-1}^i \cup \{u_t^i\}$ and re-solve \eqref{eqn:extended-nominal-solver} with the new approximate sets $\hat{U}_t^i$. It is shown in \cite[Section 5.2]{MutapcicBoyd2009} that the number of iterations $T$ needed before termination with a robust $\epsilon$-feasible solution $x_T$ is upper bounded by $(1 + O(1/\epsilon))^n$ where $n$ is the dimension of $x$.
Suppose now that we are interested in robust feasibility \eqref{eqn:approx-robust-feas}. \cite[Section 5.3]{MutapcicBoyd2009} discusses a number of variations for generating $x_t$ by modifying \eqref{eqn:extended-nominal-solver}. In contrast, we propose the following modification: instead of solving \eqref{eqn:extended-nominal-solver}, generate $\{x_t\}_{t=1}^T$ via a non-anticipatory algorithm $\mathcal{A}_x$ (see Remark \ref{rem:anticipatory-requirements}) to bound $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \mathcal{R}_x(T) = (1-\tau)\epsilon$. We call our modification \emph{FO-based pessimization}. Then the FO-based pessimization approach fits within our framework as a special case.
\begin{theorem}\label{thm:pessimization-oracle-approach}
Let $\tau \in (0,1)$. Suppose $\{x_t\}_{t=1}^T$ are generated iteratively to guarantee that $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq (1-\tau)\epsilon$ for \emph{any} sequence $\{u_t\}_{t=1}^T$. Suppose $u_t^i$ are generated by pessimization oracles \eqref{eqn:pessimization-oracle} for $i \in [m]$. If there exists $t \in [T]$ such that for all $i \in [m]$ we have $f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$, then $x_t$ is a robust $\epsilon$-feasible solution to \eqref{eqn:approx-robust-feas}. If $\max_{i \in [m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$, then $\bar{x}_T = \sum_{t=1}^{T} \theta_t x_t$ is a robust $\epsilon$-feasible solution to \eqref{eqn:approx-robust-feas}. If $\max_{i \in [m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$, then we certify that \eqref{eqn:approx-robust-feas} is robust infeasible.
\end{theorem}
{\emph{#1}}{Proof.}
It is clear that if there exists $t \in [T]$ such that for all $i \in [m]$ we have $f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$, then $x_t$ is a robust $\epsilon$-feasible solution to \eqref{eqn:approx-robust-feas}. Furthermore, the fact that $\max_{i \in [m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) > (1-\tau) \epsilon$ implies robust infeasibility of \eqref{eqn:approx-robust-feas} follows from our assumption that $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq (1-\tau) \epsilon$ and Corollary \ref{cor:onlineSP-upperbound}. To show that $\max_{i \in [m]} \sum_{t=1}^{T} f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$ implies that $\bar{x}_T$ is robust $\epsilon$-feasible, we only need to show that $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \tau \epsilon$. Observe that by our definition of $u_t^i$ in \eqref{eqn:pessimization-oracle}, we have $f^i(x_t,u_t^i) \geq \sup_{u^i \in U^i} f^i(x_t,u^i) - \tau \epsilon$, hence the regret terms in $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ satisfy
\begin{align*}
\sup_{u^i \in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) &\leq \sup_{u^i \in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t \left(\sup_{u^i \in U^i} f^i(x_t,u^i) - \tau \epsilon\right)\\
&= \sup_{u^i \in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t \sup_{u^i \in U^i} f^i(x_t,u^i) + \tau \epsilon\\
&\leq \tau \epsilon.
\end{align*}
Then
\[ \epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) = \max_{i \in [m]} \left\{ \sup_{u^i \in U^i} \sum_{t=1}^{T} \theta_t f^i(x_t,u^i) - \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \right\} \leq \tau \epsilon, \]
and the result follows from Corollary \ref{cor:onlineSP-upperbound}.
Theorem~\ref{thm:pessimization-oracle-approach} can only be used to certify robust feasibility/infeasibility. Hence, to find a robust $\epsilon$-optimal solution in FO-based pessimization approach, we must perform a binary search and solve at most $O(\mathop{{\rm log}}(1/\epsilon))$ instances of robust feasibility problems. Despite this, in Section~\ref{sec:RateDiscussion}, we discuss how FO-based pessimization approach which uses OFO algorithms to generate $\{x_t\}_{t=1}^T$ to bound $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ results in much better complexity guarantees than using an extended nominal feasibility solver \eqref{eqn:extended-nominal-solver} as proposed by \cite{MutapcicBoyd2009}, even when taking into account the additional $O(\mathop{{\rm log}}(1/\epsilon))$ factor.
\begin{remark}\label{rem:anticipatory-pessimization-oracle}
In the pessimization oracle-based approach, the noises $u_t$ need to be generated with knowledge of $x_t$, because it is not possible to guarantee $f^i(x_t,u_t^i) \geq \sup_{u^i \in U^i} f^i(x_t,u^i) - \tau \epsilon$ if the vectors $u_t^i$ were chosen with only the knowledge of $x_1,\ldots,x_{t-1}$.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
\subsection{The Nominal Feasibility Oracle-Based Approach}\label{sec:oracle-based-approach}
The nominal feasibility oracle-based approach of Ben-Tal et al.\@ \cite{BenTalHazan2015} suggest using OFO algorithms to choose a sequence $\{u_t\}_{t=1}^T$ that guarantees $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ is small, in a non-anticipatory fashion, for \emph{any} sequence $\{x_t\}_{t=1}^T$. In this aspect, it essentially matches with our OFO-based approach outlined in Section~\ref{sec:RegretforRO} i.e., the choice of $\mathcal{A}_i$ is essentially the same. The key differentiating point between our OFO-based approach and that of \cite{BenTalHazan2015} lies in which algorithm is chosen for $\mathcal{A}_x$. At step $t$, \cite{BenTalHazan2015} utilizes a \emph{nominal feasibility oracle}. That is, given parameters $u_t$, they call a powerful, and potentially expensive, nominal feasibility oracle that solves the following feasibility problem to $\epsilon$-accuracy
\begin{equation}\label{eqn:nominal-feas-oracle}
\begin{cases}
\text{\emph{Either}: find}\ \ x \in X \quad\text{s.t.}\quad f^i(x,u_t^i) \leq (1-\tau) \epsilon \quad \forall i\in[m];\\
\text{\emph{or}: declare infeasibility, } \forall x \in X,\ \exists i \in [m] \quad \text{s.t.} \quad f^i(x,u_t^i) > 0.
\end{cases}
\end{equation}
We denote $x_t \in X$ to be the point returned by this oracle at step $t$, if it exists. For this approach, the outputs of a nominal feasibility oracle can be used to deduce a result similar to Corollary \ref{cor:onlineSP-upperbound}, except that we no longer need to evaluate $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$, we just need to bound $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$.
\begin{theorem}\label{thm:robust-feas-oracle-solver}
Given weights $\theta\in\Delta_T$, suppose that the sequence $\{u_t\}_{t=1}^T$ is generated in a non-anticipatory manner to guarantee $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \tau \epsilon$ for any sequence $\{x_t\}_{t=1}^T$. Also, suppose that at each step $t \in [T]$, $x_t$ is generated by the nominal feasibility oracle which solves \eqref{eqn:nominal-feas-oracle}. If there exists $t \in [T]$ such that \eqref{eqn:nominal-feas-oracle} declares infeasibility, then \eqref{eqn:approx-robust-feas} is infeasible. Otherwise, if $x_t$ satisfies $f^i(x_t,u_t^i) \leq (1-\tau)\epsilon$ for all $t \in [T]$ and $i \in [m]$, we have a robust $\epsilon$-feasibility certificate for \eqref{eqn:approx-robust-feas}.
\end{theorem}
{\emph{#1}}{Proof.}
If \eqref{eqn:nominal-feas-oracle} declares infeasibility, then it is obvious that the robust feasibility problem is infeasible. We focus on the latter case. By the premise of the theorem, we have $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \tau\epsilon$. Let us evaluate $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i)$. Because $\theta \in \Delta_T$ and from the definition of the nominal feasibility oracle we have $f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$ for all $t \in [T]$ and $i \in [m]$, we conclude $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$. The conclusion now follows from Theorem~\ref{thm:robust-feas-oracle}.
Thus, the approach of \cite{BenTalHazan2015}, which works with nominal feasibility oracles, fits within our framework of Algorithm~\ref{alg:approx-robust-feas} right away. We next make three important remarks.
\begin{remark}\label{rem:anticipatory-nom-feas-oracle}
Similar to Remark~\ref{rem:anticipatory-pessimization-oracle}, a critical property required in the approach of \cite{BenTalHazan2015} of the vectors $x_t$ is that $f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$. This is possible only if $x_t$ were chosen with the knowledge of $\{u_1^i\}_{i=1}^m,\ldots,\{u_{t}^i\}_{i=1}^m$.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
\begin{remark}\label{rem:oracle-based-optimal-sol}
Theorem~\ref{thm:robust-feas-oracle-solver} states that the nominal feasibility oracle-based approach can solve robust feasibility problems \eqref{eqn:approx-robust-feas}. This then recovers \cite[Theorems 1,2]{BenTalHazan2015}. In addition, we next make a nice and practical observation that was overlooked in \cite{BenTalHazan2015}. We show that slightly adjusting this oracle will let us directly solve the robust \emph{optimization} problem \eqref{eqn:robust-opt-intro}, i.e., optimize a convex objective function $f^0(x)$ instead of relying on a binary search over the optimal objective value. Recall that $\Opt$ is the optimal value of the RO problem (see \eqref{eqn:robust-opt-intro}). Naively, to solve for $\Opt$, we would embed $f^0$ into the constraint set, and then perform a binary search over the robust feasible set by repeatedly applying the oracle-based approach and Theorem~\ref{thm:robust-feas-oracle-solver} to check for robust feasibility. Suppose that now, instead of using a nominal feasibility oracle to solve \eqref{eqn:nominal-feas-oracle}, we work with a \emph{nominal optimization oracle}. That is, given fixed parameters $\{u_t^i\}_{i=1}^m$, we have access to an oracle that solves
\[ \Opt_t = \inf_x \left\{ f^0(x) : f^i(x,u_t^i) \leq 0, \ i \in [m],\ \ x \in X \right\}. \]
When solving for $\Opt_t$, most convex optimization solvers will either declare that the constraints are infeasible, or return a point $x_t \in X$ such that $f^i(x_t,u_t^i) \leq (1-\tau)\epsilon$ and $f^0(x_t) \leq \Opt_t + \epsilon$. It is clear that $f^0(x_t) \leq \Opt_t + \epsilon \leq \Opt + \epsilon$. Given such a sequence of points $\{x_t\}_{t=1}^T$, from Theorem~\ref{thm:robust-feas-oracle-solver} we deduced that $\bar{x}_T = \sum_{t=1}^{T} \theta_t x_t$ is a robust $\epsilon$-feasible solution. Moreover, convexity of $f^0$ implies
\[ f^0(\bar{x}_T) \leq \sum_{t=1}^{T} \theta_t f^0(x_t) \leq \sum_{t=1}^{T} \theta_t (\Opt + \epsilon) = \Opt + \epsilon. \]
Hence, not only do we claim that $\bar{x}_T$ is robust $\epsilon$-feasible, but that it is also $\epsilon$-optimal. Thus, when our oracle can return $\epsilon$-optimal solutions, which most solvers can, we eliminate the need to perform a binary search.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
Below we elaborate on the differences between Theorem~\ref{thm:robust-feas-oracle-solver} and Corollary \ref{cor:onlineSP-upperbound}.
\begin{remark}\label{rem:nominal-feas-oracle-inefficient-SV}
In contrast to Corollary \ref{cor:onlineSP-upperbound}, Theorem~\ref{thm:robust-feas-oracle-solver} does not need to control the term $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$. There are two reasons for this: $(i)$ due to \eqref{eqn:nominal-feas-oracle}, each point $x_t$ satisfies $f^i(x_t,u_t^i) \leq (1-\tau)\epsilon$, hence $\max_{i\in[m]} \sum_{t=1}^{T} \theta_t f^i(x_t,u_t^i) \leq (1-\tau)\epsilon$ always holds. Therefore, the infeasibility part of Corollary \ref{cor:onlineSP-upperbound} never becomes relevant, and $(ii)$ due to the oracle solving \eqref{eqn:nominal-feas-oracle}, infeasibility may be declared at any step $t \in [T]$ in Theorem~\ref{thm:robust-feas-oracle-solver}. This offers the possibility of stopping early rather than having to wait until all $T$ steps are completed.
Thus, the nominal feasibility oracle-based approach trades off using more effort at each iteration $t$ to solve \eqref{eqn:nominal-feas-oracle} for the ability to terminate early. In contrast, our OFO-based approach opts to keep the per-iteration cost cheap while giving up the ability to terminate early.
More formally, let us examine a particular way of solving \eqref{eqn:nominal-feas-oracle} within a nominal feasibility oracle. Note that \eqref{eqn:nominal-feas-oracle} is equivalent to checking $F_t\leq (1-\tau) \epsilon$ or $F_t>0$, where
\begin{equation}\label{eqn:SV-feas-points}
F_t := \inf_{x \in X} \left\{ \max_{i \in [m]} f^i(x,u_t^i) \right\}.
\end{equation}
Since each $f^i(x,u_t^i)$ is convex in $x$ for fixed $u_t^i$, $\max_{i \in [m]} f^i(x,u_t^i)$ is convex in $x$ also, hence standard convex optimization methods may be employed to find $x_t \in X$ such that
\[ F_t \leq \max_{i \in [m]} f^i(x_t,u_t^i) \leq F_t + (1-\tau) \epsilon. \]
Then, by checking whether $\max_{i \in [m]} f^i(x_t,u_t^i) \leq (1-\tau) \epsilon$ or $\max_{i \in [m]} f^i(x_t,u_t^i) > (1-\tau) \epsilon$, we can determine whether $F_t \leq (1-\tau) \epsilon$ or $F_t > 0$ respectively. In particular, if we find that $F_t \leq (1-\tau) \epsilon$, our point $x_t$ is feasible for \eqref{eqn:nominal-feas-oracle}.
Also, when all the vectors $x_t$ satisfy \eqref{eqn:SV-feas-points}, we have the bound
\begin{align*}
\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) &= \sum_{t=1}^{T} \theta_t \max_{i \in [m]} f^i(x_t,u_t^i) - \inf_{x \in X} \sum_{t=1}^{T} \theta_t \max_{i \in [m]} f^i(x,u_t^i)\\
&\leq \sum_{t=1}^{T} \theta_t \left[ \max_{i \in [m]} f^i(x_t,u_t^i) - \inf_{x \in X} \max_{i \in [m]} f^i(x,u_t^i) \right] \leq (1-\tau) \epsilon.
\end{align*}
Consequently, we deduce that the nominal feasibility oracle, implemented as a convex optimization problem, also naturally bounds $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ although this bound is not utilized in Theorem~\ref{thm:robust-feas-oracle-solver}. In terms of our framework of Algorithm~\ref{alg:approx-robust-feas}, the update policy of generating new solutions $x_t$ in this approach corresponds to selecting the algorithm $\mathcal{A}_x$ to be a convex optimization solver that solves $\Opt_t$. Then whenever the solver returns a feasible solution, the regret bound \eqref{eqn:solution-update-regret-bound} satisfies $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \mathcal{R}_x(T) = (1-\tau) \epsilon$ for any $T$. Note that the term $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ inherently includes the objective functions $\max_{i \in [m]} f^i(x,u_t^i)$ of each problem $F_t$. At each iteration $t$, instead of evaluating $F_t$ to $(1-\tau)\epsilon$ accuracy, our OFO-based approach performs only a simple update based on the first-order information, and it yields a bound on $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ from the overall collection of these simple updates.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
\subsection{Convergence Rates and Discussion}\label{sec:RateDiscussion}
We summarize the convergence rates achievable in our general RO framework for various cases.
We first examine the number of iterations required for each approach discussed, then proceed to analyze the per-iteration cost of each approach. A summary of our discussion is given in Table \ref{tab:rate-comparison}.
We use the notation $r_u(\epsilon)$ to denote the number of iterations
$T$ required for algorithms $\mathcal{A}_i$ to guarantee $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \max_{i \in [m]} \mathcal{R}_i(T) \leq \epsilon/2$. Similarly, we let $r_x(\epsilon)$ be the number of iterations $T$ required for algorithm $\mathcal{A}_x$ to guarantee that $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \mathcal{R}_x(T) \leq \epsilon/2$. Then the resulting worst-case number of iterations needed in Algorithm~\ref{alg:approx-robust-feas} to obtain robust $\epsilon$-feasibility/infeasibility certificates is $\max\{r_u(\epsilon),r_x(\epsilon)\}$.
As outlined in Section \ref{sec:OCOforRO}, employing standard OFO-based algorithms, i.e., Theorem \ref{thm:OCO-non-smooth}, on the terms $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T)$ and $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T)$ requires $r_u(\epsilon) = O(1/\epsilon^2)$ and $r_x(\epsilon) = O(1/\epsilon^2)$ iterations to ensure they are no larger than $\epsilon/2$. Thus, our OFO-based approach from Section \ref{sec:RegretforRO} requires $O(1/\epsilon^2)$ iterations to solve \eqref{eqn:robust-feas}. Since our OFO-based approach returns only robust $\epsilon$-feasible solutions, we need to perform a binary search and repeatedly invoke our method $O(\mathop{{\rm log}}(1/\epsilon))$ times to obtain $\epsilon$-optimal solutions, so the total number of iterations is $O(\mathop{{\rm log}}(1/\epsilon)/\epsilon^2)$.
Our FO-based pessimization approach, i.e., our modification of the pessimization oracle-based approach of \cite{MutapcicBoyd2009} outlined in Section \ref{sec:pessimization-oracle}, requires $r_x(\epsilon)$ iterations to solve \eqref{eqn:robust-feas} because by Theorem~\ref{thm:pessimization-oracle-approach} we only need to guarantee $\epsilon^\bullet(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \mathcal{R}_x(T) \leq \epsilon/2$. Taking into account the binary search factor $O(\mathop{{\rm log}}(1/\epsilon))$ to find a robust $\epsilon$-optimal solution, the total number of iterations required is $O(\mathop{{\rm log}}(1/\epsilon)/\epsilon^2)$, which is much better than the exponential $(1 + O(1/\epsilon))^n$ bound of \cite[Section 5.2]{MutapcicBoyd2009} that uses a full nominal solution oracle \eqref{eqn:extended-nominal-solver}. Similarly, the nominal feasibility/optimization oracle-based approach of \cite{BenTalHazan2015} outlined in Section \ref{sec:oracle-based-approach} requires $r_u(\epsilon) = O(1/\epsilon^2)$ iterations (or $r_u(\epsilon)\mathop{{\rm log}}(1/\epsilon) = O(\mathop{{\rm log}}(1/\epsilon)/\epsilon^2)$ iterations if only a feasibility oracle is used) to obtain robust $\epsilon$-optimal solutions because by Theorem \ref{thm:robust-feas-oracle-solver} we only need to bound $\epsilon^\circ(\{x_t,u_t,\theta_t\}_{t=1}^T) \leq \max_{i \in [m]} \mathcal{R}_i(T) \leq \epsilon/2$.
\begin{remark}\label{rem:FOMflexibility}
The flexibility of our general framework in terms of the selection of algorithms $\mathcal{A}
_i,{\cal A}_x$ extends beyond just using Theorem \ref{thm:OCO-non-smooth}. Depending on the structure of functions $f^i$ and uncertainty domains $U^i$, the algorithms ${\cal A}_i$ and ${\cal A}_x$ may be replaced by more appropriate OCO algorithms. For example, when $f^i$ are strongly convex, certain OCO algorithms achieve faster convergence rates. Moreover, unless explicitly required by the algorithms ${\cal A}_i$, we do not need to assume convexity of the sets $U^i$. As a result, the follow-the-leader or follow-the-perturbed-leader type algorithms from \cite{KalaiVempala2005} can be utilized as ${\cal A}_i$ in our framework even when $U^i$ are nonconvex but certain assumptions ensuring applicability of these algorithms are satisfied. Such assumptions are satisfied for example when $f^i(x,u^i)$ are linear in $u^i$ and the nonconvex sets $U^i$ admit a certain linear optimization oracle. This is for example the case in a certain lifted representation of the robust convex quadratic constraint discussed in \cite[Section 4.2]{BenTalHazan2015}.
Similarly, when the functions $f^i(x,u^i)$ are exp-concave in $u^i$, applying the online Newton step algorithm of \cite{HazanAgarwalKale2007} for ${\cal A}_i$ results in a weighted regret bound of at most $O\left( \mathop{{\rm log}}(T)/ T \right)$ in $T$ iterations. Such $f^i$ that are exp-concave in $u^i$ satisfying Assumption~\ref{ass:f-concave-u} arise in optimization under uncertainty problems where variance is used as a risk measure, e.g., mean-variance portfolio optimization problems, see for example \cite[Example 25]{BenTalDenHertog2015}. Essentially, the same flexibility for acceleration and/or working with nonconvex sets $U^i$ is present in \cite{BenTalHazan2015} as well.
In the presence of favorable problem structure, based on Table \ref{tab:rate-comparison}, if an accelerated algorithm to exploit problem structure is employed in the place of $\mathcal{A}_i$, the overall number of iterations of the nominal feasibility approach is immediately reduced accordingly. Analogous result holds for $\mathcal{A}_x$ and the FO-based pessimization approach. However, in the case of our OFO-based approach, we need to have favorable structure in \emph{both} $x$ and $u$ and utilize the corresponding accelerated algorithms $\mathcal{A}_x, \mathcal{A}_i$ to attain the acceleration of the overall approach.
\hfill\hbox{\hskip 4pt \vrule width 5pt height 6pt depth 1.5pt}\vspace{0.0cm}\par
\end{remark}
We now discuss the per-iteration cost for each approach. In order to discuss the total \emph{arithmetic complexity} of each approach, we let $k$ be the maximum dimension of the uncertain parameters $u^i$ for $i\in[m]$ and recall that $n$ denotes the dimension of the decision variables $x$. In the case where our domains $X,\{U^i\}_{i=1}^m$ have favorable geometry, such as Euclidean ball or simplex, the vectors $x_t, \{u_t^i\}_{i=1}^m$ are updated via simple closed-form prox operations, which cost $O(n)$ and $O(km)$ per iteration respectively. The cost of computing the subgradients $\ensuremath{\nabla}_x f^i(x,u^i), \ensuremath{\nabla}_u f^i(x,u^i)$ is at least $O(km + mn)$ each iteration. This cost is incurred in each iteration of all of the approaches we discuss. From this, we deduce that the per-iteration cost of our OFO-based approach is at most $O(km + mn)$.
The per-iteration cost of the pessimization oracle based approaches involve calling $m$ pessimization oracles \eqref{eqn:pessimization-oracle} and the costs related to updating $x_t$.
We denote by $\Pess(\epsilon,k)$ the complexity of a pessimization oracle with tolerance $\epsilon$ and $k$ variables. A summary of different possible implementations is given in Table \ref{tab:pess-complexity}.
If $\sup_{u^i \in U^i} f^i(x,u^i)$ has a simple closed form solution, then the resulting arithmetic cost for $\Pess(\epsilon,k)$ is $O(k)$ for each pessimization oracle. If we can use polynomial-time IPMs, this cost becomes $O(k^3 \mathop{{\rm log}}(1/\epsilon))$ (see \cite[Section 6.6]{BenTalNemirovski2001book}), and using FOMs has cost $O(k \mathop{{\rm log}}(1/\epsilon))$ in the best case when the functions $f^i$ are smooth \emph{and} strongly convex in $u^i$.
In the case of our FO-based pessimization approach, the update involving $x_t$ will be given by simple closed form formulas for prox operations when $X$ has favorable geometry, resulting in a cost of $O(mn)$. The full pessimization approach of \cite{MutapcicBoyd2009} incurs the cost of solving an extended nominal feasibility problem for the update of $x_t$.
The per-iteration cost of the nominal feasibility/optimization oracle-based approach of \cite{BenTalHazan2015}, as well as that of of \cite{MutapcicBoyd2009}, depends on the type of solver used to solve the nominal optimization/feasibility problem \eqref{eqn:nominal-feas-oracle}.
We denote by $\Nom(\epsilon,m,n)$ the complexity of a nominal oracle with tolerance $\epsilon$, $m$ constraints and $n$ variables. Note that nominal solvers can be either optimization or feasibility solvers. If it is the latter, an extra $\mathop{{\rm log}}(1/\epsilon)$ factor is incurred to perform binary search. A summary of different possible implementations for $\Nom(\epsilon,m,n)$ is given in Table \ref{tab:nom-complexity}. When applicable for $\Nom(\epsilon,m,n)$ implementation, polynomial-time IPMs are guaranteed to terminate in $O(\sqrt{m}\mathop{{\rm log}}(1/\epsilon))$ iterations with a solution to \eqref{eqn:nominal-feas-oracle} and thus offer the best rates in terms of their dependence on $\epsilon$. They also have the advantage that they can act as a nominal optimization oracle, and hence by Remark~\ref{rem:oracle-based-optimal-sol} there will be no need to perform an additional binary search to find an $\epsilon$-optimal solution. On the other hand, they demand significantly more memory, and their per-iteration cost is quite high in terms of the dimension, usually around the order of $O(n^3 + mn)$, see \cite[Chapter 6.6]{BenTalNemirovski2001book}.
In order to keep both the memory requirements and the per-iteration cost associated with implementing the nominal feasibility oracle $\Nom(\epsilon,m,n)$ low, one may opt for a FOM called the CoMirror algorithm that can work with functional constraints, see \cite{BeckBenTal2010} and \cite[Section 1.3]{JuditNem2012Pt1}. CoMirror algorithm is guaranteed to find a solution to the nominal $\epsilon$-feasibility problem within $O(1/\epsilon^2)$ iterations, with a much cheaper per-iteration cost of $O(mn)$. Because CoMirror method can optimize as well, it does not need binary search. However, to the best of our knowledge, its possibility to exploit further structural properties of the functions $f^i$, such as smoothness in $x$, to improve the dependence on $\epsilon$ are not known. In order to exploit such properties in the implementation of $\Nom(\epsilon,m,n)$, it is possible to cast \eqref{eqn:nominal-feas-oracle} as a convex-concave SP problem, and then apply efficient FOMs such as Nesterov's algorithm~\cite{Nesterov2005} or Nemirovski's Mirror Prox algorithm~\cite{Nemirovski2005} to achieve a convergence rate of $O(\mathop{{\rm log}}(m)/\epsilon)$ and per-iteration cost of $O(mn)$. This convex-concave SP approach can only be used as a nominal feasibility oracle, so we must repeat the process $\mathop{{\rm log}}(1/\epsilon)$ times to obtain an $\epsilon$-optimal solution.
\input{SaddlePointTables}
Recall that Table~\ref{tab:rate-comparison} summarizes the rates for the various approaches, together with rates for the full pessimization approach of \cite{MutapcicBoyd2009} and using the CoMirror with pessimization (discussed in Section \ref{sec:connections-other-FOM}). Note that the total \emph{overall arithmetic complexity} of each approach is obtained by multiplying the quantities in each row in Table~\ref{tab:rate-comparison}. The quantities $r_u(\epsilon), r_x(\epsilon)$ will generally be $O(1/\epsilon^2)$, with potential for application-specific acceleration when the functions $f^i$ exhibit favorable structure.
Table~\ref{tab:rate-comparison} indicates that our FO-based pessimization approach when it admits a closed form solution for the implementation of $\Pess(\epsilon,k)$ and the nominal feasibility oracle-based approach which uses a polynomial-time IPM solver to implement the nominal feasibility oracle $\Nom(\epsilon,m,n)$ give the best dependence on $\epsilon$ among all of the methods. These are better than our OFO-based approach by factors of $\max\{1,r_u(\epsilon)/r_x(\epsilon)\}$ and $\max\{1,r_x(\epsilon)/r_u(\epsilon)\}$ respectively. However, in many applications, we can expect that $r_u(\epsilon) \approx r_x(\epsilon)$, so these factors will be constant. In this case, our OFO-based approach becomes competitive with having a closed form pessimization oracle in our FO-based pessimization approach or using a nominal IPM solver in \cite{BenTalHazan2015}. That said, compared to IPMs, our OFO-based approach demands much less memory, and it is able to maintain a much lower dependence on the dimensions $m,n$ and thus is much more scalable, whereas the cost per iteration of such IPMs has a rather high dependence on the dimension. In addition, the memory requirements of IPMs are far more than OFO algorithms, posing a critical disadvantage to their use in large-scale applications. Similar comparisons of our OFO-based approach against pessimization or nominal feasibility oracle-based approaches utilizing other methods point out its advantage, which is at least an order of magnitude better in terms of its dependence on $\epsilon$. In fact, when $r_x(\epsilon)\approx r_u(\epsilon)$, our method can lead to savings over the approach of \cite{BenTalHazan2015} with CoMirror algorithm used in its oracle by a factor as large as $O(1/(\epsilon^2 \mathop{{\rm log}}(1/\epsilon)))$.
\subsection{Connections with Existing First-order Methods}\label{sec:connections-other-FOM}
Finally, we would like to discuss and contrast directly solving robust convex optimization problems \eqref{eqn:robust-opt-intro} via general first-order methods.
Many FOMs require domains that are simple so that the prox operations can be easily done. In that respect, domains defined by multiple functional constraints $g^i(x) \leq 0$ creates a challenge for directly applying many of these algorithms. We now discuss two existing classes of FOMs that are designed to handle such domains: primal-dual methods and the CoMirror approach. Applying these FOMs to the RO problem \eqref{eqn:robust-opt-intro} can be viewed as another alternative solution methodology to solve RO problems without using the robust counterpart.
A general technique to address the functional constraints in the domain is to embed these constraints into the objective through Lagrange multipliers, and then solve the associated dual problem via FOMs (see e.g., \cite{NedicOzdaglar2009a}). Such methods are known as primal-dual methods. For the RO problem \eqref{eqn:robust-opt-intro}, this corresponds to solving
\[ \max_{\lambda} \left\{ \mathcal{L}^*(\lambda) := \min_{x \in X} \left[ f^0(x) + \sum_{i=1}^m \lambda^{(i)} g^i(x) \right] : \lambda \geq 0 \right\}, \]
where we define $g^i(x) := \sup_{u^i \in U^i} f^i(x,u^i)$. Primal-dual methods (e.g., \cite{NedicOzdaglar2009a}) commonly require us to solve the inner minimization problem over $x \in X$ at each iteration. For RO, this means we must solve an expensive SP problem at each iteration. Our OFO-based approach aims to improve on this by reducing the per-iteration cost of each step to simple first-order updates. Two exceptions within the primal-dual methods are the work of Nedi\'{c} and Ozdaglar \cite{NedicOzdaglar2009b} and Yu and Neely \cite{YuNeely2016}, which have cheap per-iteration cost based on only gradient computations and projection operations in the Euclidean setup. Nedi\'{c} and Ozdaglar \cite{NedicOzdaglar2009b} provide a convergence rate of $O(1/\sqrt{T})$ in the non-smooth case. While using such a primal-dual method has the advantage that no binary search is needed, we note that this requires two assumptions to guarantee convergence: we have access to exact first-order information for the robust constraint functions $g^i(x) := \sup_{u^i \in U^i} f^i(x,u^i)$, and the standard Slater constraint qualification condition (i.e., strict feasibility) is satisfied. The first assumption is often not satisfied, since we may only be able to compute $g^i(x)$ up to accuracy $\epsilon$. While there exists some FOMs that work with inexact objective gradients over simple domains, see e.g., \cite{DevolderNesterovGlineur2014}, such methods have only been applied to specific max-type objectives, e.g., objectives obtained from smoothing. It is unclear how such methods can be extended for more general max-type functions which can arise in RO. Secondly, enforcing the Slater condition implicitly enforces feasibility of \eqref{eqn:robust-opt-intro}. In contrast, our framework directly uses the functions $f^i(x,u^i)$, so it does not need to take into account the inexact gradient information, and can certify infeasibility of \eqref{eqn:robust-opt-intro}.
Yu and Neely \cite{YuNeely2016} present a method that can guarantee $O(1/T)$ convergence when all functions are smooth. However, for RO problems, the constraint functions $g^i(x)$ are non-smooth due to the supremum operation, thus their results do not apply to RO.
The only FOM that we are aware of that can solve convex problems with functional constraints without assuming feasibility is the CoMirror algorithm \cite{BeckBenTal2010} and its earlier variations in the Euclidean setup \cite{NemYudin1983,Nesterov2004Book,Polyak1967}. The CoMirror\footnotemark[1] algorithm finds an $\epsilon$-optimal $\epsilon$-feasible solution in $O(1/\epsilon^2)$ iterations to a convex program $\min_{x \in X} \left\{ f^0(x) : g^i(x) \leq 0,\ i \in [m] \right\}$ or certifies its infeasibility by using (sub)gradient information of the objective $f^0$ as well as the constraint functions $g^i$.
\footnotetext[1]{Recall that the CoMirror algorithm is also discussed in Section~\ref{sec:RateDiscussion} as a method to implement the nominal feasibility solver; in that case we are given the noises $\bar{u}^i$ resulting in $g^i(x) := f^i(x,\bar{u}^i)$, and thus the subgradient of $g^i(x)$ is simply the subgradient of $f^i(x,\bar{u}^i)$.}
In the RO problem \eqref{eqn:robust-opt-intro} we defined $g^i(x) := \sup_{u^i \in U^i} f^i(x,u^i)$. As mentioned above, in many cases, we may only be able to compute $g^i(x)$ approximately, thus only have access to approximate/inexact gradient information. It is unknown to us whether or not techniques such as the ones from \cite{DevolderNesterovGlineur2014} can be applied to the CoMirror algorithm in the presence of this type of gradient information. While the CoMirror algorithm's complexity is $O(1/\epsilon^2)$ (see also Nesterov \cite[Chapter 3.2.4]{Nesterov2004Book} for a similar result in the Euclidean case), our iterative framework can exploit favorable structure on the functions $f^i$ that can improve on the iteration complexity $r_u(\epsilon), r_x(\epsilon)$. For the Euclidean case, Nesterov \cite[Chapters 2.3.4-2.3.5]{Nesterov2004Book} shows also that convergence can be obtained in $O(\mathop{{\rm log}}(1/\epsilon))$ iterations when the objective and all constraint functions are both smooth and strongly convex in $x$. However, such an improvement does not apply to the RO problem, since we cannot in general guarantee that $g^i(x)=\sup_{u^i \in U^i} f^i(x,u^i)$ is smooth in $x$. It is unknown whether the iteration complexity of CoMirror algorithm can be improved when only the underlying function $f^i(x,u^i)$ is strongly convex or smooth, or when $g^i(x)$ is strongly convex but non-smooth.
Finally, let us get back to the case when we have a robust feasibility problem with a \emph{single} constraint $m=1$ and a convex uncertainty set $U=U^1$. In such a case, as discussed in Remark~\ref{rem:SPm=1}, we have a direct convex-concave SP problem \eqref{eqn:SadVal} under Assumption~\ref{ass:f-concave-u}. The OFO-based approach then corresponds to bounding the whole SP gap \eqref{eqn:SPgap}, the FO-based pessimization corresponds to bounding the primal gap i.e., the first term in \eqref{eqn:SPgap}, and the nominal feasibility oracle approach corresponds to bounding the dual gap, i.e., the second term in \eqref{eqn:SPgap}. Without any further structural assumptions on $f^1$, convex-concave SP problems can be solved in $O(1/\epsilon^2)$ iterations. Our approaches also achieve this rate immediately, see Table~\ref{tab:rate-comparison}. Moreover, when $f^1$ is smooth in $x$ and strongly concave in $u^1$, our general framework can achieve a rate of $O(1/\epsilon)$.
However, for specific applications involving a single robust constraint, directly working with a specialized convex-concave SP formulation can improve this rate further. For example, such an improved rate of $O(1/\sqrt{\epsilon})$ is achieved in \cite{Ben-TalRobustSVM2012} for a robust support vector machine problem.
\section{Supplementary Numerical Results}\label{sec:supplementary-numerical}
\renewcommand{\thefootnote}{\fnsymbol{footnote}}
Below, we provide the exact numerical values corresponding to the data used to generate the figures in our numerical study in Section \ref{sec:numerical}.
\begin{table}[ht]
\caption
{Average solve time (seconds) of each approach for different $n$ (Figure \ref{fig:average-solve-times}).
\label{tab:method-time-summary}}
{
\begin{tabular}{ll|cccc}
\toprule
& & \multicolumn{4}{c}{Approach} \\ \cline{3-6}
& & OFO-based & FO-based pessimization & nominal & full pessimization \\ \hline
\multirow{7}{*}{$n$} & 100 & 6.24\footnotemark[1] & 8.03\footnotemark[2] & 1.18 & 0.47 \\
& 200 & 8.28\footnotemark[3] & 9.21\footnotemark[3] & 4.14 & 1.88 \\
& 300 & 10.67 & 10.62 & 8.57 & 4.13 \\
& 400 & 14.38\footnotemark[3] & 12.91\footnotemark[3] & 14.19 & 7.55 \\
& 500 & 16.80 & 14.63 & 22.27 & 12.10 \\
& 600 & 21.33 & 17.96 & 30.83 & 17.17 \\
& 700 & 24.52 & 20.14 & 39.76 & 23.72 \\
\bottomrule
\end{tabular}
}
{}
\end{table}
\footnotetext[1]{Three instances out of 35 did not solve due to numerical issues.}
\footnotetext[2]{Two instances out of 35 did not solve due to numerical issues.}
\footnotetext[3]{One instance out of 35 did not solve due to numerical issues.}
\footnotetext[4]{One instance out of five did not solve due to numerical issues.}
\vspace{-5pt}
\begin{table}[ht]
\caption
{Average solve time (seconds) of each approach for different $m$ and $n$ (Figure \ref{fig:average-solve-times-plots}).
\label{tab:method-factors-time-summary}
}
{
\begin{tabular}{rll|rrrrrrr}
\toprule
& & & \multicolumn{7}{c}{$m$} \\ \cline{4-10}
& & & 3 & 5 & 7 & 10 & 15 & 20 & 25 \\ \hline
\multirow{16}{*}{$n$} & \multirow{4}{*}{400} & OFO-based & 13.17\footnotemark[4] & 12.21 & 14.78 & 14.03 & 14.73 & 16.23 & 15.24 \\
& & FO-based pessimization & 12.47\footnotemark[4] & 11.04 & 11.92 & 12.97 & 14.27 & 14.81 & 12.77 \\
& & nominal & 2.16 & 4.23 & 6.77 & 11.31 & 15.82 & 24.86 & 31.77 \\
& & full pessimization & 1.61 & 2.69 & 3.86 & 6.03 & 8.87 & 12.53 & 16.05 \\
\cline{2-10}
& \multirow{4}{*}{500} & OFO-based & 15.96 & 15.55 & 15.88 & 17.22 & 17.41 & 17.52 & 18.04 \\
& & FO-based pessimization & 13.04 & 12.26 & 14.08 & 17.09 & 16.28 & 14.55 & 15.07 \\
& & nominal & 4.02 & 5.92 & 10.22 & 17.86 & 22.06 & 45.16 & 50.63 \\
& & full pessimization & 2.57 & 4.29 & 6.06 & 9.95 & 13.92 & 22.44 & 25.46 \\
\cline{2-10}
& \multirow{4}{*}{600} & OFO-based & 18.62 & 19.52 & 20.97 & 21.25 & 20.86 & 23.46 & 24.62 \\
& & FO-based pessimization & 15.75 & 16.48 & 18.11 & 16.55 & 17.16 & 20.89 & 20.80 \\
& & nominal & 3.59 & 8.68 & 13.01 & 23.71 & 38.05 & 55.65 & 73.11 \\
& & full pessimization & 2.91 & 6.16 & 8.68 & 13.25 & 21.25 & 30.91 & 37.05 \\
\cline{2-10}
& \multirow{4}{*}{700} & first-order & 19.49 & 21.86 & 24.46 & 25.04 & 25.12 & 28.48 & 27.18 \\
& & pessimization & 16.28 & 17.26 & 21.51 & 22.07 & 20.86 & 22.10 & 20.87 \\
& & nominal & 6.38 & 12.53 & 14.24 & 29.81 & 48.79 & 77.03 & 89.51 \\
& & full pessimization & 4.98 & 8.39 & 12.02 & 18.66 & 29.02 & 42.89 & 50.10 \\
\bottomrule
\end{tabular}
}
{}
\end{table}
|
\section{Radial Cauchy solutions as limits of exterior solutions}\label{rad_solns}
Consider the Cauchy problem for the 3-d linear wave equation
with radial initial data:
\[\text{(CP)}\qquad
\left\{\begin{array}{ll}
\square_{1+3}U=0 & \text{on $(0,T)\times\mathbb{R}^3$}\\
U(0,x)=\Phi(x) & \text{on $\mathbb{R}^3$}\\
U_t(0,x)=\Psi(x) & \text{on $\mathbb{R}^3$,}
\end{array} \right.\]
where
\[\Phi\in H^s_{rad}(\mathbb{R}^3), \qquad \Psi\in H^{s-1}_{rad}(\mathbb{R}^3),\]
with
\begin{equation}\label{rad_versns}
\Phi(x)=\varphi(|x|)\qquad \Psi(x)=\psi(|x|).
\end{equation}
Throughout we refer to the unique solution $U$ of (CP) as the {\em Cauchy
solution}.
In this work we consider how the radial Cauchy solution $U$ can be realized
as a limit of solutions to initial-boundary value problems posed on the
exterior of vanishing balls $B_\varepsilon$ ($\varepsilon\downarrow 0$) about the origin.
The precise issue will be formulated below.
We shall consider exterior solutions satisfying either a vanishing Neumann
condition or a vanishing Dirichlet condition along $|x|=\varepsilon$.
It is well known that the Sobolev spaces $H^s$ provide a natural
setting for the Cauchy problem for the wave equation; see \cite{ra} and
\eq{opt_reg_1}-\eq{opt_reg_2} below.
The choice of space dimension $3$ is for convenience: it is particularly
easy to generate radial solutions in this case. Next, both the choice of
spaces for the initial data for (CP), as well as the boundary condition
imposed on the exterior solutions, will influence the convergence of
exterior solutions toward the Cauchy solution.
For the wave equation in $\mathbb{R}^3$ the different convergence behavior of
exterior Neumann and exterior Dirichlet solutions is brought out by
considering $H^2$ vs.\ $H^1$ initial data; see Remark \ref{1st_rmk} below.
The scheme of generating radial solutions to Cauchy problems as limits of
exterior solutions has been applied to a variety of evolutionary PDE problems;
see \cite{jt1} for references and discussion.
In our earlier work \cite{jt1} we used the 3-d wave equation to gauge
the effectiveness of this general scheme in a case where
``everything is known.'' In order that the results be relevant
to other (possibly nonlinear) problems, the analysis in \cite{jt1} deliberately
avoided any use of explicit solution formulae. Based on energy arguments
and strong convergence alone, it was found that the exterior solutions do
converge to the Cauchy solution as the balls vanish. However,
the arguments did not yield optimal information about the regularity of the
limiting Cauchy solution. Specifically,
for $s=2$ we obtained the Cauchy solution as a limit only in
$H^1$ (via exterior Neumann solutions) or in $L^2$ (via exterior Dirichlet
solutions). This is strictly less regularity than what is known to be the case, see
\eq{opt_reg_2}.
Thus, in general, while limits of exterior solutions to evolutionary PDEs
may be used to establish existence for radial Cauchy problems, one should
not expect optimal regularity information about the Cauchy solution via this
approach.
On the other hand, for the particular case of the 3-d wave equation
with radial data, it is natural to ask what type of convergence we can
establish if we exploit solution formulae (for the Cauchy
solution as well as for the exterior solutions). The present work
addresses this question, and our findings are summarized
in Theorem \ref{main_result} below.
We stress that while \cite{jt1} dealt with the issue of using exterior
solutions as a stand-alone method for obtaining existence of radial
Cauchy problems, the setting for the present work is different.
We are now exploiting what is known about the solution of the
Cauchy solutions as well as exterior solutions for the radial 3-d
linear wave equation, and the only issue is how the former
solutions are approximated by the latter.
\begin{remark}\label{1st_rmk}
Before starting the detailed analysis we comment on a
slightly subtle point. As recorded in our main result (Theorem
\ref{main_result}), we
establish $H^2$-regularity of the limiting Cauchy solution $U$
when the initial data $(\Phi,\Psi)$ belong to $H^2\times H^1$,
and $H^1$-regularity when the data belong to $H^1\times L^2$.
This is as it should be according to
\eq{opt_reg_1}. Now, in the former case $U$
is obtained as a limit of exterior Neumann solutions, while in the
latter case it is obtained as a limit of exterior Dirichlet solutions.
This raises a natural question: what regularity is obtained for
$U$ in the case of $H^2\times H^1$-data, if we insist on
approximating by exterior Dirichlet solutions?
To answer this we need to specify how we compare the
everywhere defined Cauchy solution $U$ to exterior solutions
$U^\varepsilon$, which are defined only on the exterior domains
$\mathbb{R}^3_\varepsilon:=\mathbb{R}^3\setminus B_\varepsilon$.
There are at least two ways to do this\footnote{When using
exterior solutions to establish existence for (CP) (as in \cite{jt1}),
there is no such choice: one must produce approximations to
$U$ that are everywhere defined.}:
\begin{itemize}
\item[(a)] by calculating $\|U(t)-U^\varepsilon(t)\|_{H^s(\mathbb{R}^3_\varepsilon)}$;
\item[(b)] by first defining a suitable extension $\tilde U^\varepsilon$ of $U^\varepsilon$
to all of $\mathbb{R}^3$, and then calculating
$\|U(t)-\tilde U^\varepsilon(t)\|_{H^s(\mathbb{R}^3)}$.
\end{itemize}
With {\em (b)}, which is what we do in this paper, the natural choice
is to extend $U^\varepsilon(t)$ continuously as a constant on $B_\varepsilon$
at each time. I.e., for
exterior Dirichlet solutions, we let $\tilde U^\varepsilon(t,x)$ vanish identically on
$B_\varepsilon$, while for exterior Neumann solutions its value there
is that of $U^\varepsilon(t,x)$ along the $|x|=\varepsilon$.
It turns out that regardless of whether we use {\em (a)} or {\em (b)} to compare
the Cauchy solution to the exterior solutions, the answer to the question
above is that we obtain only $H^1$-convergence when
exterior Dirichlet solutions are used. In fact, for {\em (b)} this is immediate:
the exterior Dirichlet solution $\tilde U^\varepsilon$ will typically have a nonzero
radial derivative at $r=\varepsilon+$ so that its extension $\tilde U^\varepsilon$
contains a ``kink'' along $|x|=\varepsilon$. Thus, second derivatives of
$\tilde U^\varepsilon$ will typically contain a $\delta$-function along $|x|=\varepsilon$,
and $\tilde U^\varepsilon$ does not even belong to $H^2(\mathbb{R}^3)$ in this case.
For {\em (a)} it suffices to consider the situation at time zero.
With $\Phi$ as above we consider smooth cutoffs $\Phi^\varepsilon$
(see \eq{approx_dir_data_1} below).
A careful calculation, carried out in \cite{jt1}, shows that
$\|\Phi-\Phi^\varepsilon\|_{H^2(\mathbb{R}^3_\varepsilon)}$ blows up as
$\varepsilon\downarrow 0$.
These remarks highlight the unsurprising but relevant fact that
exterior Dirichlet solutions are more singular than exterior Neumann
solutions; see \cite{jt1} for a discussion.
\end{remark}
The goal is to show that the Cauchy solution $U$ of (CP) can be approximated,
uniformly on compact time intervals, in $H^2$-norm by suitably chosen
exterior Neumann solutions and in $H^1$-norm by
suitably chosen exterior Dirichlet solutions.
As indicated we shall use explicit solution formulae for both the Cauchy
problem (CP) as well as for the exterior Neumann and Dirichlet problems.
These formulae for radial solutions are readily available in 3 dimensions
and exploits the fact that radial solutions of $\square_{1+3}U=0$
admit the representation
\[U(t,x)=\frac{u(t,|x|)}{|x|}\]
where $u(t,r)$ solves $\square_{1+1}u=0$ on the half-line $\mathbb{R}^+$.
(Exterior Neumann solutions require a little work to write down
explicitly; see \eq{U_eps}.)
Of course, with the explicit formulae in place, it is a matter of
computation to analyze the required norm differences. However,
it is a rather involved computation since the formulae involve different
expressions in several different regions.
Also, the answers do not follow by appealing to well-posedness
for the wave equation (see \eq{opt_reg_2} below): the Cauchy solution
and the exterior solutions are defined on different domains.
As noted above we opt to extend the exterior solutions to the interior
of the balls $B_\varepsilon$, before comparing them to the Cauchy solution.
Instead of a direct comparison we prefer to estimate the
$H^2$- and $H^1$-differences in question by employing the
natural energies for the wave equation. These energies will majorize
the $L^2$-distances of the first and second derivatives, and will
also provide an estimate on the $L^2$-distance of the functions
themselves.
There are two advantages of this approach: first, it is straightforward
to calculate the exact rates of change of the energies in question, and second,
these rates depend only on what takes place at or within radius $r=\varepsilon$.
The upshot is that it suffices to analyze fewer terms than required by
a direct approach. Finally, to estimate the rates of change of the relevant
energies we make use of the explicit solution formulae.
\section{Setup and statement of main result}
\subsection{The Cauchy solution}
A standard result (see e.g.\ \cites{bjs,ra}) shows that the radial Cauchy solution $U$
of (CP) may be calculated explicitly by using the representation
\[U(t,x)=\frac{u(t,|x|)}{|x|},\]
where $u(t,r)$ solves the half-line problem
\begin{equation}\label{half_line_dir}
\text{(Half-line)}\qquad
\left\{\begin{array}{ll}
\square_{1+1}u =0 & \text{on $(0,T)\times\mathbb{R}^+$}\\
u(0,r)=r\varphi(r) & \text{for $r\in\mathbb{R}^+$}\\
u_t(0,r)=r\psi(r) & \text{for $r\in\mathbb{R}^+$}\\
u(t,0)\equiv 0 & \text{for $t>0$,}\end{array} \right.
\end{equation}
where $\varphi$ and $\psi$ are as in \eq{rad_versns}.
By using the d'Alembert formula for the half-line problem (see \cite{bjs})
we obtain that
\begin{equation}\label{U}
U(t,r)=\left\{
\begin{array}{ll}
0\leq r\leq ct: & \frac{1}{2r}\left[ (ct+r) \varphi(ct+r)-(ct-r)\varphi(ct-r)\right]\\
& +\frac{1}{2cr}\int_{ct-r}^{ct+r} s\psi(s)\, ds\\\\
r\geq ct: & \frac{1}{2r}\left[ (r+ct) \varphi(r+ct)+(r-ct)\varphi(r-ct)\right]\\
& +\frac{1}{2cr}\int_{r-ct}^{r+ct} s\psi(s)\, ds.
\end{array}
\right.
\end{equation}
In addition to the solution formula \eq{U} we shall also exploit the
following well-known stability property \cites{ra,sel}:
with data $\Phi\in H^s(\mathbb{R}^3)$ and $\Psi\in H^{s-1}(\mathbb{R}^3)$
(radial or not), the Cauchy problem (CP) admits a unique solution
$U$ which satisfies
\begin{equation}\label{opt_reg_1}
U\in C([0,T];H^s(\mathbb{R}^3))\cap C^1([0,T];H^{s-1}(\mathbb{R}^3))
\end{equation}
and
\begin{equation}\label{opt_reg_2}
\|U(t)\|_{H^s(\mathbb{R}^3)}+\|U_t(t)\|_{H^{s-1}(\mathbb{R}^3)}
\leq C_T\left( \|\Phi\|_{H^s(\mathbb{R}^3)}+\|\Psi\|_{H^{s-1}(\mathbb{R}^3)}\right),
\end{equation}
for each $T>0$, where $C_T$ is a number of the form
$C_T=\bar C\cdot(1+T)$, and $\bar C$ a universal constant.
\subsection{Exterior solutions and their extensions}\label{ext_solns}
With $\Phi\in H^s_{rad}(\mathbb{R}^3)$ and $\Psi\in H^{s-1}_{rad}(\mathbb{R}^3)$, $s=1$
or $2$, the goal is to show that the solution $U$ of (CP) can be ``realized as a limit of
exterior solutions'' defined outside of $B^\varepsilon$ as $\varepsilon\downarrow 0$.
To make this precise we need to specify:
\begin{enumerate}
\item precisely which exterior solutions we consider: which boundary
conditions do they satisfy along $\partial B_\varepsilon$, and how are their
initial data related to the given Cauchy data $\Phi$, $\Psi$;
\item how we compare the everywhere defined Cauchy solution $U$
with exterior solutions $U^\varepsilon$, that are defined only outside of
$B_\varepsilon$; and
\item which norm we use for comparing $U$ and $U^\varepsilon$.
\end{enumerate}
Concerning (1) we shall consider exterior solutions that satisfy either
vanishing Neumann or vanishing Dirichlet conditions along $\partial B^\varepsilon$.
In either case, the initial data for the exterior problem are generated
by a two-step procedure: we first approximate the original
Cauchy data by $C^\infty_{c,rad}(\mathbb{R}^3)$-functions,
and then apply an appropriate modification of these
smooth approximations near the origin. These modifications
use smooth cut-off functions and are made so that the result
satisfies vanishing Neumann or Dirichlet conditions along $|x|=\varepsilon$.
(See \eq{approx_neum_data_1}-\eq{approx_neum_data_2} and
\eq{approx_dir_data_1}-\eq{approx_dir_data_2} for details.)
In either case we denote the exterior, radial solutions corresponding to
the approximate, smooth data by $U^\varepsilon(t,x)\equiv U^\varepsilon(t,r)$;
they are given explicitly in \eq{U_eps} and \eq{U_eps_dir} below.
As mentioned in Remark \ref{1st_rmk}, for (2) we opt to compare
the Cauchy solution $U$ to the natural extensions $\tilde U^\varepsilon$ of
the smooth exterior solution $U^\varepsilon$: at each time $t$,
$\tilde U^\varepsilon(t,x)$ takes the constant value $U^\varepsilon(t,\varepsilon)$ on
$B_\varepsilon$, and coincides with $U^\varepsilon(t,x)$ for $|x|\geq \varepsilon$.
Thus, in the case of Dirichlet data, $\tilde U^\varepsilon(t,x)$ vanishes
identically on $B_\varepsilon$, while for Neumann data its value there
is that of $U^\varepsilon(t,x)$ along the boundary $|x|=\varepsilon$.
Finally, concerning (3), Remark \ref{1st_rmk} above also explains
the choice of $H^2$-norm for comparing the Cauchy solution $U$
to exterior Neumann solutions, and $H^1$-norm for comparison
to exterior Dirichlet solutions.
Our main result is as follows.
\begin{theorem}\label{main_result}
Let $T>0$ be given and let $U$ denote the solution of the
radial Cauchy problem $\mathrm{(CP)}$ for the linear wave
equation in three space dimensions with initial data $(\Phi,\Psi)$.
\begin{itemize}
\item[(i)] For initial data in $H^2_{rad}(\mathbb{R}^3)\times H^1_{rad}(\mathbb{R}^3)$ the
Cauchy solution $U$ can be realized as a $C([0,T];H^2(\mathbb{R}^3))$-limit
of suitable extended exterior Neumann solutions as
$\varepsilon\downarrow 0$.
\item[(ii)] For initial data in $H^1(\mathbb{R}^3)\times L^2(\mathbb{R}^3)$ the
Cauchy solution $U$ can be realized as a $C([0,T];H^1(\mathbb{R}^3))$-limit
of suitable extended exterior Dirichlet solutions as
$\varepsilon\downarrow 0$.
\end{itemize}
\end{theorem}
We point out that, e.g.\ in part (i), we do not claim that the
extended Neumann solutions $\tilde U^\varepsilon$ converge to $U$
in $H^2$-norm. In fact, we establish this latter property only for
the case with $C^\infty_c(\mathbb{R}^3)$ initial data. However, thanks
to the stability property \eq{opt_reg_2}, this is sufficient to obtain
(i); see Proposition \ref{smooth_case} below.
The rest of the paper is organized as follows. After reducing to
the case with smooth and compactly supported data in Section
\ref{smooth_c}, we treat $H^2$-convergence of exterior Neumann
solutions in Sections \ref{gen_neu_soln}-\ref{comp_neu}, while
$H^1$-convergence of exterior Dirichlet solutions is established
in Sections \ref{gen_dir_soln}-\ref{comp_dir}.
\subsection{Reduction to smooth case}\label{smooth_c}
The first step of the proof is to use well-posedness \eq{opt_reg_2}
for the Cauchy problem to reduce to the case of smooth initial data.
\begin{proposition}\label{smooth_case}
With the setup in Theorem \ref{main_result},
let $\tilde U^{N,\varepsilon}$ and $\tilde U^{D,\varepsilon}$ denote the
extensions of the exterior Neumann and Dirichlet solutions,
respectively, as described in Section \ref{ext_solns}.
Then, Theorem \ref{main_result} follows once it is established that
\begin{equation}\label{smooth_case1}
\sup_{0\leq t\leq T}\|U(t)- \tilde U^{N,\varepsilon}(t)\|_{H^2(\mathbb{R}^3)}\to 0
\qquad\text{as $\varepsilon\downarrow 0$}
\end{equation}
and
\begin{equation}\label{smooth_case2}
\sup_{0\leq t\leq T}\|U(t)- \tilde U^{D,\varepsilon}(t)\|_{H^1(\mathbb{R}^3)}\to 0
\qquad\text{as $\varepsilon\downarrow 0$,}
\end{equation}
for any initial data $\Phi,\Psi\in C_{c,rad}^\infty(\mathbb{R}^3)$.
\end{proposition}
\begin{proof}
For concreteness consider the case of exterior
Neumann solutions, and let arbitrary data $\Phi\in H^2_{rad}(\mathbb{R}^3)$,
$\Psi\in H^1_{rad}(\mathbb{R}^3)$ be given. Fix any $\delta>0$.
We first choose $\Phi_0$, $\Psi_0$ in $C_{c,rad}^\infty(\mathbb{R}^3)$ with
\[\|\Phi-\Phi_0\|_{H^2}+ \|\Psi-\Psi_0\|_{H^1}< \frac{\delta}{2C_T},\]
where $C_T$ is as in \eq{opt_reg_2}.
The existence of such $\Phi_0$, $\Psi_0$ may be established in a
standard manner via convolution (using a radial mollifier)
and smooth cutoff at large radii. Let $U_0$ denote the solution of
(CP) with data $\Phi_0$, $\Psi_0$. Also, for any $\varepsilon>0$
let $\tilde U^{N,\varepsilon}_0(t,x)$ denote the extension of the exterior Neumann
solution with data $\Phi_0^\varepsilon$, $\Psi_0^\varepsilon$, as described in
Section \ref{ext_solns}.
Then, assuming that \eq{smooth_case1} has been
established, we can choose $\varepsilon>0$
sufficiently small to guarantee that
\[\sup_{0\leq t\leq T}\|U_0(t)- \tilde U^{N,\varepsilon}_0(t)\|_{H^2}<\frac{\delta}{2}.\]
Hence, for any $t\in[0,T]$ we have
\begin{align*}
\|U(t)- \tilde U^{N,\varepsilon}_0(t)\|_{H^2}
&\leq \|U(t)- U_0(t)\|_{H^2}+\|U_0(t)- \tilde U^{N,\varepsilon}_0(t)\|_{H^2}\\
&\overset{\eq{opt_reg_2}}{\leq}
C_T\left(\|\Phi-\Phi_0\|_{H^2}+ \|\Psi-\Psi_0\|_{H^1}\right)+\frac{\delta}{2}
<\delta,
\end{align*}
by the choice of $\Phi_0$, $\Psi_0$.
\end{proof}
From now on we therefore consider an arbitrary but fixed pair of
functions $\Phi,\, \Psi\in C^\infty_{c,\, rad}(\mathbb{R}^3)$.
Note that we then have that the functions $\varphi$ and $\psi$
in \eq{rad_versns} are smooth on $\mathbb{R}_0^+$ and satisfy $\varphi'(0+)=\psi'(0+)=0$.
\section{Exterior Neumann solutions}\label{gen_neu_soln}
In this section and the next we consider the case of exterior Neumann
solutions.
For the fixed initial data $\Phi,\, \Psi\in C^\infty_{c,\, rad}(\mathbb{R}^3)$ and
any $\varepsilon>0$ we derive a formula for $U^\varepsilon(t,x)\equiv U^{N,\varepsilon}(t,x)$,
defined for
$|x|\geq\varepsilon$ and satisfying $\partial_rU^\varepsilon|_{r=\varepsilon}=0$. We refer
to $U^\varepsilon$ as the {\em exterior Neumann solution} corresponding
to the solution $U$ of (CP) with data $\Phi$, $\Psi$. In Section
\ref{comp_neu} we will then estimate how it (really, its extension
$\tilde U^\varepsilon(t,x)$ to all of $\mathbb{R}^3$) approximates the solution
$U(t)$ in $H^2(\mathbb{R}^3)$ at fixed times.
To generate the exterior Neumann solution $U^\varepsilon$ we
fix a smooth, nondecreasing function $\beta:\mathbb{R}_0^+\to\mathbb{R}_0^+$ with
\begin{equation}\label{beta_props_1}
\beta\equiv 1 \quad\text{on $[0,1]$,}\quad \beta(s)=s\quad\text{for $s\geq 2$.}
\end{equation}
Then, with $\varphi$ and $\psi$ as in \eq{rad_versns}, we define
\begin{equation}\label{approx_neum_data_1}
\Phi^\varepsilon(x)\equiv \varphi^\varepsilon(|x|):=\varphi\big(\varepsilon\beta
\big(\textstyle\frac{|x|}{\varepsilon}\big)\big)
\end{equation}
and
\begin{equation}\label{approx_neum_data_2}
\Psi^\varepsilon(x)\equiv \psi^\varepsilon(|x|):=\psi\big(\varepsilon\beta
\big(\textstyle\frac{|x|}{\varepsilon}\big)\big).
\end{equation}
We refer to $(\Phi^\varepsilon,\Psi^\varepsilon)$ as the {\em Neumann data}
corresponding to the original Cauchy data $(\Phi,\Psi)$ for (CP).
Note that the Neumann data are actually defined on all of $\mathbb{R}^3$,
that they are constant (equal to $\varphi(\varepsilon)$ and $\psi(\varepsilon)$,
respectively) on $B_\varepsilon$, and that their restrictions to the
exterior domain $\{x\in\mathbb{R}^3\,:\, |x|\geq\varepsilon\}$
satisfy homogeneous Neumann conditions along $|x|=\varepsilon$.
The exterior Neumann solution $U^\varepsilon$ is now defined as
the unique radial solution of the initial-boundary value problem
\[\left\{\begin{array}{ll}
\square_{1+3}V =0 & \text{on $(0,T)\times\{|x|>\varepsilon\}$}\\
V(0,x)=\Phi^\varepsilon(x) & \text{for $|x|>\varepsilon$}\\
V_t(0,x)=\Psi^\varepsilon(x) & \text{for $|x|>\varepsilon$}\\
\partial_r V(t,x)=0 & \text{along $|x|=\varepsilon$ for $t>0$.}
\end{array} \right.\]
To obtain a formula for $U^\varepsilon$ we exploit the fact that $V$ is a radial
solution of the 3-d wave equation if and only if $v=rV$ solves the 1-d wave equation.
Setting
\[u^\varepsilon(t,r):=rU^\varepsilon(t,r),\]
we obtain that $u^\varepsilon$ solves the corresponding 1-d problem on $\{r>\varepsilon\}$:
\[\text{($\varepsilon$-Half-line)}\qquad
\left\{\begin{array}{ll}
\square_{1+1}u=0 & \text{on $(0,T)\times\{r>\varepsilon\}$}\\
u(0,r)=r\varphi^\varepsilon(r) & \text{for $r>\varepsilon$}\\
u_t(0,r)=r\psi^\varepsilon(r) & \text{for $r>\varepsilon$}\\
u_r(t,\varepsilon)= \frac{1}{\varepsilon}u(t,\varepsilon) & \text{for $t>0$.}
\end{array} \right.\]
Note that the Neumann condition for the 3-d solution corresponds to a Robin
condition for the 1-d solution. (A direct calculation shows that the initial data
for $u^\varepsilon$ and $u_t^\varepsilon$ both satisfy this Robin condition.)
The solution $u^\varepsilon$ to the $\varepsilon$-Half-line problem is explicitly
given via d'Alembert's formula\footnote{One way to solve the 1-dimensional
Robin IBVP is to first solve the IBVP with general Dirichlet data
$u^\varepsilon(t,\varepsilon)=h(t)$ along $r=\varepsilon$, for which a d'Alembert formula is
readily available (see John \cite{bjs}, p.\ 8); one may then identify the $h$
which gives $u_r= \frac{1}{\varepsilon}u$ along $r=\varepsilon$.}:
\begin{equation}\label{u_eps}
u^\varepsilon(t,r)=\left\{
\begin{array}{ll}
\varepsilon\leq r\leq ct+\varepsilon: & \frac{1}{2}\left[ (ct+r) \varphi^\varepsilon(ct+r)
+(ct-r+2\varepsilon)\varphi^\varepsilon(ct-r+2\varepsilon)\right]\\
& +\frac{1}{2c}\int_{ct-r+2\varepsilon}^{ct+r} s\psi^\varepsilon(s)\, ds \\
&+e^{\frac{r-ct-2\varepsilon}{\varepsilon}}
\int_\varepsilon^{ct-r+2\varepsilon}\left[\frac{s\psi^\varepsilon(s)}{c}
-\frac{s\varphi^\varepsilon(s)}{\varepsilon}\right]e^{\frac{s}{\varepsilon}}\, ds\\\\
r\geq ct+\varepsilon: & \frac{1}{2}\left[ (r+ct) \varphi^\varepsilon(r+ct)
+(r-ct)\varphi^\varepsilon(r-ct)\right]\\
& +\frac{1}{2c}\int_{r-ct}^{r+ct} s\psi^\varepsilon(s)\, ds.
\end{array}\right.
\end{equation}
A direct calculation shows that $u^\varepsilon$ is a classical solution on
$\mathbb{R}_t\times \{r>\varepsilon\}$. From this we obtain the radial exterior
Neumann solution $U^\varepsilon(t,r):=\frac{u^\varepsilon(t,r)}{r}$:
\begin{equation}\label{U_eps}
U^\varepsilon(t,r)=\left\{
\begin{array}{ll}
\varepsilon\leq r\leq ct+\varepsilon: & \frac{1}{2r}\left[ (ct+r) \varphi^\varepsilon(ct+r)
+(ct-r+2\varepsilon)\varphi^\varepsilon(ct-r+2\varepsilon)\right]\\
& +\frac{1}{2cr}\int_{ct-r+2\varepsilon}^{ct+r} s\psi^\varepsilon(s)\, ds\\
&+\frac{1}{r}e^{\frac{r-ct-2\varepsilon}{\varepsilon}}
\int_\varepsilon^{ct-r+2\varepsilon}\left[\frac{s\psi^\varepsilon(s)}{c}
-\frac{s\varphi^\varepsilon(s)}{\varepsilon}\right]
e^{\frac{s}{\varepsilon}}\, ds\\\\
r\geq ct+\varepsilon: & \frac{1}{2r}\left[ (r+ct) \varphi^\varepsilon(r+ct)
+(r-ct)\varphi^\varepsilon(r-ct)\right]\\
& +\frac{1}{2cr}\int_{r-ct}^{r+ct} s\psi^\varepsilon(s)\, ds.
\end{array}
\right.
\end{equation}
We finally extend $U^\varepsilon$ at each time to obtain an everywhere
defined approximation of the Cauchy solution $U(t,x)$. As discussed
earlier we use the natural choice of extending $U^\varepsilon$ continuously
as a constant on $B_\varepsilon$ at each time:
\begin{equation}\label{tilde_U_eps}
\tilde U^\varepsilon(t,x)=\left\{
\begin{array}{ll}
U^\varepsilon(t,\varepsilon) & \text{for $0\leq |x|\leq \varepsilon$}\\\\
U^\varepsilon(t,x) & \text{for $|x|\geq \varepsilon$.}
\end{array}
\right.
\end{equation}
For later use we record that the value along the boundary is explicitly given as
\begin{equation}\label{U_eps_bndry_val}
U^\varepsilon(t,\varepsilon)=\frac{1}{\varepsilon}(ct+\varepsilon)\varphi^\varepsilon(ct+\varepsilon)
+\frac{1}{\varepsilon}e^{-\frac{ct+\varepsilon}{\varepsilon}}
\int_\varepsilon^{ct+\varepsilon}\left[\frac{s\psi^\varepsilon(s)}{c}-\frac{s\varphi^\varepsilon(s)}{\varepsilon}\right]
e^{\frac{s}{\varepsilon}}\, ds,
\end{equation}
and we also note that
\begin{equation}\label{init_approx_data}
\tilde U^\varepsilon(0,x)=\Phi^\varepsilon(x),\qquad
\tilde U_t^\varepsilon(0,x)=\Psi^\varepsilon(x)\qquad\text{for all $x\in\mathbb{R}^3$.}
\end{equation}
\section{Comparing Cauchy and exterior Neumann solutions}\label{comp_neu}
The issue now is to estimate the $H^2$-distance
\[\|U(t)-\tilde U^\varepsilon(t)\|_{H^2(\mathbb{R}^3)}\]
as $\varepsilon\downarrow 0$. As explained in Section \ref{rad_solns}
we prefer to estimate this $H^2$-difference by employing the
natural energies for the wave equation.
These energies will majorize the $L^2$-distances of the
first and second derivatives of $U(t)$ and $\tilde U^\varepsilon(t)$, and
also provide control of the $L^2$-distance of the functions themselves.
\subsection{Energies}\label{energies}
For any function $W(t,x)$ which is twice weakly differentiable on
$\mathbb{R}\times\mathbb{R}^3$ we define the following 1st and 2nd order energies
(note their domains of integration):
\[\mathcal E_{W}(t):={\textstyle\frac{1}{2}}\int_{\mathbb{R}^3}
|\partial_t W(t,x)|^2+c^2|\nabla W(t,x)|^2\, dx,\]
\[\mathcal E^\varepsilon_{W}(t):={\textstyle\frac{1}{2}}\int_{|x|\geq \varepsilon}
|\partial_t W(t,x)|^2+c^2|\nabla W(t,x)|^2\, dx,\]
and
\[\mathbb{E}_{W}(t):=\sum_{i=1}^3 \mathcal E_{\partial_i W}(t)
=\sum_{i=1}^3{\textstyle\frac{1}{2}}\int_{\mathbb{R}^3}
|\partial_t \partial_i W(t,x)|^2 +c^2|\nabla \partial_i W(t,x)|^2\, dx,\]
\[\mathbb{E}^\varepsilon_{W}(t):=\sum_{i=1}^3 \mathcal E^\varepsilon_{\partial_i W}(t)
=\sum_{i=1}^3{\textstyle\frac{1}{2}}\int_{|x|>\varepsilon}
|\partial_t \partial_i W(t,x)|^2 +c^2|\nabla \partial_i W(t,x)|^2\, dx.\]
The first goal is to estimate the energies
\begin{equation}\label{ult_energ_1}
\mathcal E^\varepsilon(t):=\mathcal E_{U-\tilde U^\varepsilon}(t)
\end{equation}
\begin{equation}\label{ult_energ_2}
\mathbb{E}^\varepsilon(t):=\mathbb{E}_{U-\tilde U^\varepsilon}(t),
\end{equation}
which majorizes the $L^2$-distances between the 1st and 2nd
derivatives of $U$ and $\tilde U^\varepsilon$, respectively.
As a first step we observe the following facts.
\begin{lemma}\label{energy_1}
With $U$ and $U^\varepsilon$ as defined above we have: each of the energies
\[\mathcal E_{U}(t),\quad \mathcal E^\varepsilon_{U^\varepsilon}(t),\quad
\mathcal E_{\partial_i U}(t),\quad\text{and}\quad\mathcal E^\varepsilon_{\partial_i U^\varepsilon}(t)\]
are constant in time.
\end{lemma}
\begin{proof}
The constancy of the first three energies is standard, while the
constancy of $\mathcal E^\varepsilon_{\partial_i U^\varepsilon}(t)$ is a consequence
of the fact that we consider radial solutions. Indeed, as $U^\varepsilon$
is radial and satisfies vanishing Neumann conditions along $|x|=\varepsilon$,
we have that $\nabla U^\varepsilon(t,x)\equiv 0$ along $|x|=\varepsilon$. Thus,
$U^\varepsilon_{x_it}\equiv 0$ for each $i=1,2,3$ along $|x|=\varepsilon$.
Differentiating in time, using that $U^\varepsilon$ is a solution of the wave equation,
and integrating by parts, we therefore have
\begin{align*}
\dot{\mathcal E}^\varepsilon_{\partial_i U^\varepsilon}(t)
&= \int_{|x|>\varepsilon} U^\varepsilon_{x_it}U^\varepsilon_{x_itt}
+c^2\nabla U^\varepsilon_{x_i}\cdot\nabla U^\varepsilon_{x_it}\, dx\\
&=c^2\int_{|x|>\varepsilon} U^\varepsilon_{x_it}\Delta U^\varepsilon_{x_i}
+\nabla U^\varepsilon_{x_i}\cdot\nabla U^\varepsilon_{x_it}\, dx
= c^2\int_{\partial\{|x|>\varepsilon\}}U^\varepsilon_{x_it}
\frac{\partial U^\varepsilon_{x_i}}{\partial\nu}\, dS=0.
\end{align*}
\end{proof}
Next, to estimate $\mathcal E^\varepsilon(t)$, we expand the integrand and use that
$\nabla \tilde U^\varepsilon$ vanishes on $B_\varepsilon$ (by our choice of extension),
to get
\begin{align*}
\mathcal E^\varepsilon(t) &= \mathcal E_{U- \tilde U^\varepsilon}(t)
=\frac{1}{2} \int_{\mathbb{R}^3} |U_t-\tilde U^\varepsilon_t|^2 + c^2|\nabla U-\nabla \tilde U^\varepsilon|^2\, dx\\
&=\mathcal E_U(t)+\mathcal E_{\tilde U^\varepsilon}(t)
-\int_{\mathbb{R}^3} U_t\tilde U^\varepsilon_t +c^2\nabla U\cdot\nabla \tilde U^\varepsilon\, dx\\
&= \mathcal E_U(t)+\mathcal E^\varepsilon_{U^\varepsilon}(t)
+\frac{\text{vol}(B_\varepsilon)}{2} |U^\varepsilon_t(t,\varepsilon)|^2
-U^\varepsilon_t(t,\varepsilon)\int_{|x|<\varepsilon} U_t(t,x)\, dx\\
&\quad-\int_{|x|>\varepsilon} U_tU^\varepsilon_t+c^2\nabla U\cdot\nabla U^\varepsilon\, dx.
\end{align*}
Differentiating in time, applying Lemma \ref{energy_1}, integrating by parts,
and using the boundary condition $\partial_r U^\varepsilon(t,\varepsilon)\equiv 0$,
then yield
\[\dot{\mathcal E}^\varepsilon(t) = \frac{d}{dt}\left[\frac{\text{vol}(B_\varepsilon)}{2} |U^\varepsilon_t(t,\varepsilon)|^2
-U^\varepsilon_t(t,\varepsilon)\int_{|x|<\varepsilon} U_t(t,x)\, dx\right]
+c^2\int_{|x|=\varepsilon} U^\varepsilon_t \partial_rU\, dS.\]
Integrating back up in time, and recalling that $U^\varepsilon$ and $U$ are radial, we obtain
\begin{align}\label{1_st_energy_diff}
\mathcal E^\varepsilon(T) &= \mathcal E^\varepsilon(0)
+\left[\frac{\text{vol}(B_\varepsilon)}{2} |U^\varepsilon_t(t,\varepsilon)|^2
-U^\varepsilon_t(t,\varepsilon)\int_{|x|<\varepsilon} U_t(t,x)\, dx\right]_{t=0}^{t=T}\nonumber\\
&\quad + c^2\text{area}(B_\varepsilon)\int_0^T U^\varepsilon_t(t,\varepsilon) \partial_rU(t,\varepsilon)\, dt.
\end{align}
Below we shall carefully estimate the terms on the RHS to show that
$\mathcal E^\varepsilon(T)\to 0$ as $\varepsilon\downarrow 0$.
Before carrying out a similar representation of the 2nd order energy
difference $\mathbb{E}^\varepsilon(t)$, we observe how $\mathcal E^\varepsilon(t)$
controls the $L^2$-distance between $U$ and $\tilde U^\varepsilon$.
Setting
\begin{equation}\label{L2_dist}
\mathcal D^\varepsilon(t):=\frac{1}{2}\int_{\mathbb{R}^3}|U(t,x)-\tilde U^\varepsilon(t,x)|^2\, dx,
\end{equation}
the Cauchy-Schwarz inequality gives
\[\dot{\mathcal D}^\varepsilon(t)\leq 2\mathcal D^\varepsilon(t)^\frac{1}{2}
\mathcal E^\varepsilon(t)^\frac{1}{2},\]
such that
\begin{equation}\label{1_vs_0_energy}
\mathcal D^\varepsilon(T)\lesssim \mathcal D^\varepsilon(0)+\int_0^T \mathcal E^\varepsilon(t)\, dt.
\end{equation}
We now consider how $\mathbb{E}^\varepsilon(t)$ changes in time. Arguing as above, using
Lemma \ref{energy_1} and the fact that $U^\varepsilon_{x_i}\equiv 0$
on $B_\varepsilon$, we have
\begin{align}
\mathcal E_{\partial_i U-\partial_i \tilde U^\varepsilon}(t) &=
\frac{1}{2}\int_{\mathbb{R}^3} |U_{x_it}-\tilde U^\varepsilon_{x_it}|^2
+c^2|\nabla U_{x_it}-\nabla \tilde U^\varepsilon_{x_it}|^2\, dx\nonumber\\
&=\mathcal E_{\partial_i U}(t)+\mathcal E_{\partial_i \tilde U^\varepsilon}(t)
-\int_{\mathbb{R}^3} U_{x_i,t}\tilde U^\varepsilon_{x_it}+c^2\nabla U_{x_i}
\cdot\nabla \tilde U^\varepsilon_{x_i}\, dx\nonumber\\
&=\mathcal E_{\partial_i U}(0)+\mathcal E^\varepsilon_{\partial_i U^\varepsilon}(0)
- \int_{|x|>\varepsilon}U_{x_i,t} U^\varepsilon_{x_it}+c^2\nabla U_{x_i}
\cdot\nabla U^\varepsilon_{x_i}\, dx.
\end{align}
Differentiating in time and integrating by parts in the last integral, give
\begin{equation}\label{indiv_term}
\dot{\mathcal E}_{\partial_i U-\partial_i \tilde U^\varepsilon}(t)
=c^2\int_{|x|=\varepsilon}\big(U_{x_it}\big)\big(\partial_r U^\varepsilon_{x_i}\big)\, dS.
\end{equation}
Observing that we have
\[\sum_{i=1}^3U_{x_it}\partial_r U^\varepsilon_{x_i}=\big(\partial_r U_t\big)\big(\partial_{rr}U^\varepsilon\big)\]
along $\{|x|=\varepsilon\}$ (recall that $\partial_r U^\varepsilon$ vanishes along $\{|x|=\varepsilon\}$),
we obtain from \eq{indiv_term} that
\begin{equation}\label{2_nd_energy_diff}
\mathbb{E}^\varepsilon(T)=\mathbb{E}^\varepsilon(0)+c^2\text{area}(B_\varepsilon)
\int_0^T \big(\partial_r U_t(t,\varepsilon)\big)\big(\partial_{rr}U^\varepsilon(t,\varepsilon)\big)\, dt.
\end{equation}
To estimate $\mathcal E^\varepsilon(T)$ and $\mathbb{E}^\varepsilon(T)$, and hence also
$\mathcal D^\varepsilon(T)$ according to \eq{1_vs_0_energy}, we employ
the solution formulae \eq{U} and \eq{U_eps}.
\subsection{Initial differences in energy }
The details of estimating the initial differences of the first and second
order energies, i.e.\ $\mathcal E^\varepsilon(0)$ and $\mathbb{E}^\varepsilon(0)$, were
carried out in Section 3.2 of \cite{jt1} (and makes use of
\eq{init_approx_data}). Translating to our present notation
we have that
\begin{equation}\label{initial_0th_energy_diff}
\mathcal D^\varepsilon(0)\lesssim \varepsilon^2\|\Phi\|_{H^1(B_{2\varepsilon})}^2,
\end{equation}
\begin{equation}\label{initial_1st_energy_diff}
\mathcal E^\varepsilon(0)\lesssim \varepsilon^2\|\Psi\|_{H^1(B_{2\varepsilon})}^2
+\|\Phi\|_{H^1(B_{2\varepsilon})}^2,
\end{equation}
and
\begin{equation}\label{initial_2nd_energy_diff}
\mathbb{E}^\varepsilon(0)\lesssim \|\Psi\|_{H^1(B_{2\varepsilon})}^2
+\|\Phi\|_{H^2(B_{2\varepsilon})}^2.
\end{equation}
\subsection{Estimating growth of first order energy difference}
According to \eq{1_st_energy_diff}, to estimate $\mathcal E^\varepsilon(T)$
we need to estimate the quantities $U^\varepsilon_t$ and $\partial_r U$ along $|x|=\varepsilon$.
For the remaining term involving $U_t(t,x)$ in \eq{1_st_energy_diff} (for $|x|\leq \varepsilon$),
it will suffice to employ an energy estimate that does not require formulae.
Before considering these terms in detail we record the following fact.
For any $k\in \mathbb{R}$ and for any $t>0$ let
\[Q_k^\varepsilon(t):=\frac{1}{\varepsilon^k}\left(e^{-\frac{ct+\varepsilon}{\varepsilon}}
\int_\varepsilon^{ct+\varepsilon}\left[\frac{s\psi^\varepsilon(s)}{c}
-\frac{s\varphi^\varepsilon(s)}{\varepsilon}\right]e^{\frac{s}{\varepsilon}}\, ds
+ (ct+\varepsilon)\varphi^\varepsilon(ct+\varepsilon)\right);\]
then
\begin{equation}\label{Q}
Q_k^\varepsilon(t)\to 0\qquad\text{as $\varepsilon\downarrow 0$}.
\end{equation}
To see this, integrate by parts in the $\varphi^\varepsilon$-term to get that
\begin{align*}
Q_k^\varepsilon(t)&=\frac{1}{c}
\int_\varepsilon^{ct+\varepsilon}\!\!\!\! s\psi^\varepsilon(s)\varepsilon^{-k}e^{\frac{s-ct-\varepsilon}{\varepsilon}}\, ds
+\varphi^\varepsilon(\varepsilon)\varepsilon^{1-k}e^{-\frac{ct}{\varepsilon}}
+\int_\varepsilon^{ct+\varepsilon}\!\!\!\!\left[\varphi^\varepsilon(s)
+s{\varphi^\varepsilon}'(s)\right]\varepsilon^{-k}e^{\frac{s-ct-\varepsilon}{\varepsilon}}\, ds.
\end{align*}
Recalling \eq{approx_neum_data_1}-\eq{approx_neum_data_2} and using that
$\varphi$ and $\psi$ are fixed, smooth functions, the Dominated Convergence Theorem
yields $Q_k^\varepsilon(t)\to 0$ as $\varepsilon\downarrow 0$.
\subsubsection{Estimating $U^\varepsilon_t(t,\varepsilon)$}
According to \eq{U_eps_bndry_val} we have
\[U^\varepsilon_t(t,\varepsilon)=\frac{c}{\varepsilon}\left(\varphi^\varepsilon(ct+\varepsilon)
+(ct+\varepsilon){\varphi^\varepsilon}'(ct+\varepsilon)
+\frac{(ct+\varepsilon)}{c}\psi^\varepsilon(ct+\varepsilon)\right)-cQ^\varepsilon_2(t).\]
As $Q^\varepsilon_2(t)$ tends to zero while $\varphi^\varepsilon$ and $\psi^\varepsilon$
remain bounded, we conclude that
\begin{equation}\label{U^veps_t_bound}
|U^\varepsilon_t(t,\varepsilon)|\lesssim \frac{1}{\varepsilon}\qquad\text{for all $t\in[0,T]$ as $\varepsilon\downarrow 0$.}
\end{equation}
\subsubsection{Estimating $\partial_rU(t,\varepsilon)$}
According to \eq{U} we have
\begin{equation}\label{U_r}
\partial_rU(t,\varepsilon)=\left\{
\begin{array}{ll}
t\geq \frac{\varepsilon}{c}: & -\frac{1}{2\varepsilon^2}\left[ (ct+\varepsilon) \varphi(ct+\varepsilon)-(ct-\varepsilon)\varphi(ct-\varepsilon)\right]\\\\
&+\frac{1}{2\varepsilon}\left[\varphi(ct+\varepsilon)+ (ct+\varepsilon)\varphi'(ct+\varepsilon)+\varphi(ct-\varepsilon)
+(ct-\varepsilon)\varphi'(ct-\varepsilon)\right]\\\\
&-\frac{1}{2c\varepsilon^2}\int_{ct-\varepsilon}^{ct+\varepsilon} s\psi(s)\, ds
+\frac{1}{2c\varepsilon}\left[(ct+\varepsilon) \psi(ct+\varepsilon)+(ct-\varepsilon)\psi(ct-\varepsilon)\right]\\\\\\
t\leq \frac{\varepsilon}{c}: & -\frac{1}{2\varepsilon^2}\left[ (\varepsilon+ct) \varphi(\varepsilon+ct)+(\varepsilon-ct)\varphi(\varepsilon-ct)\right]\\\\
&+\frac{1}{2\varepsilon}\left[ \varphi(\varepsilon+ct) +(\varepsilon+ct)\varphi'(\varepsilon+ct)+\varphi(\varepsilon-ct)
+(\varepsilon-ct)\varphi'(\varepsilon-ct)\right]\\\\
& -\frac{1}{2c\varepsilon^2}\int_{\varepsilon-ct}^{\varepsilon+ct} s\psi(s)\, ds
+\frac{1}{2c\varepsilon}\left[(\varepsilon+ct) \psi(\varepsilon+ct)-(\varepsilon-ct) \psi(\varepsilon-ct)\right].
\end{array}
\right.
\end{equation}
The terms for $t\geq\frac{\varepsilon}{c}$ are estimated by 2nd order
Taylor expansion of $\varphi(ct\pm\varepsilon)$ and $\psi(ct\pm\varepsilon)$
about $\varepsilon=0$. The terms for $t\leq \frac{\varepsilon}{c}$ are estimated
by 2nd order Taylor expansion of $\varphi$ and $\psi$ about zero,
and then using that $\varphi'(0)=\psi'(0)=0$. (As observed earlier, this
holds since $\varphi$ and $\psi$ are profile functions of the {\em smooth},
radial functions $\Phi$ and $\Psi$, respectively).
These expansions are straightforward and we omit them. The end
result is that the leading order terms in \eq{U_r} cancel, leaving terms
of size at most $O(\varepsilon)$. We thus have that
\begin{equation}\label{U_r_bound}
|\partial_r U(t,\varepsilon)|\lesssim \varepsilon
\qquad\text{for all $t\in[0,T]$ as $\varepsilon\downarrow 0$.}
\end{equation}
(Note: this is actually obvious since we know that $U$ is a smooth,
radial solution satisfying $\partial_r U(t,0)\equiv 0$ and with fixed data
independent of $\varepsilon$.)
Finally, the Cauchy-Schwarz inequality and Lemma \ref{energy_1}
give
\[\Big|\int_{|x|<\varepsilon} U_t(t,x)\, dx\Big|
\lesssim \varepsilon^\frac{3}{2}\mathcal E_U(0)^\frac{1}{2}. \]
Applying this together with \eq{U^veps_t_bound} and \eq{U_r_bound}
in \eq{1_st_energy_diff}, we conclude that
\begin{equation}\label{1_st_energy_diff_1}
\mathcal E^\varepsilon(T) \lesssim \mathcal E^\varepsilon(0) +\varepsilon^\frac{1}{2}.
\end{equation}
\subsection{Estimating growth of second order energy differences}
Next, according to \eq{2_nd_energy_diff}, to estimate $\mathbb{E}^\varepsilon(T)$,
we need to estimate the quantities $\partial_rU_t$ and $\partial_{rr} U^\varepsilon$
along $|x|=\varepsilon$.
\subsubsection{Estimating $\partial_rU_t(t,\varepsilon)$}
By taking the time derivative of \eq{U_r} and then Taylor expanding
the various terms as outlined above, we deduce that
\begin{equation}\label{U^_rt_bound}
|\partial_r U_t(t,\varepsilon)|\lesssim \varepsilon\qquad\text{for all $t\in[0,T]$ as $\varepsilon\downarrow 0$.}
\end{equation}
\subsubsection{Estimating $\partial_{rr}U^\varepsilon(t,\varepsilon)$}
This estimate again requires a direct, but rather long, calculation
(which we omit), followed by a careful analysis of the resulting expression.
The first step is to calculate $\partial_{rr}U^\varepsilon(t,r)$ for $\varepsilon\leq r\leq ct+\varepsilon$,
by using the first part of formula \eq{U_eps}. A number of cancellations occur
when the resulting expression is evaluated at $r=\varepsilon$, and we are left with
\begin{align*}
\partial_{rr}U^\varepsilon(t,\varepsilon) = Q^\varepsilon_3(t)
&-\frac{1}{\varepsilon^2}\left[\varphi^\varepsilon+(ct-\varepsilon){\varphi^\varepsilon}'
-\varepsilon(ct+\varepsilon){\varphi^\varepsilon}''\right]
-\frac{1}{c\varepsilon^2}\left[ct\psi^\varepsilon-\varepsilon(ct+\varepsilon){\psi^\varepsilon}'\right],
\end{align*}
where $\varphi^\varepsilon$, $\psi^\varepsilon$, and their derivatives are evaluated at $ct+\varepsilon$.
According to \eq{approx_neum_data_1}-\eq{approx_neum_data_2} we have that
$\varphi^\varepsilon$, $\psi^\varepsilon$, and their first derivatives remain bounded independently
of $\varepsilon$, while ${\varphi^\varepsilon}''$ is at most of order $\frac{1}{\varepsilon}$.
Since $Q^\varepsilon_3(t)\to 0$ as $\varepsilon\downarrow 0$ by \eq{Q}, we therefore have that
\begin{equation}\label{U^eps_rr_bound}
|\partial_{rr}U^\varepsilon(t,\varepsilon)|\lesssim \frac{1}{\varepsilon^2}
\qquad\text{for all $t\in[0,T]$ as $\varepsilon\downarrow 0$.}
\end{equation}
Finally, by using \eq{U^_rt_bound} and \eq{U^eps_rr_bound} in
\eq{2_nd_energy_diff}, we conclude that
\begin{equation}\label{2_nd_energy_diff_1}
\mathbb{E}^\varepsilon(T) \lesssim \mathbb{E}^\varepsilon(0) +\varepsilon.
\end{equation}
\subsection{Convergence of exterior Neumann solutions}
According to the definitions of $\mathcal D^\varepsilon(t)$, $\mathcal E^\varepsilon(t)$, and
$\mathbb{E}^\varepsilon(t)$, together with the estimates \eq{1_vs_0_energy},
\eq{1_st_energy_diff_1}, \eq{2_nd_energy_diff_1} we have
\begin{align*}
\|U(t)-\tilde U^\varepsilon(t)\|_{H^2(\mathbb{R}^3)}^2
&\lesssim \mathcal D^\varepsilon(t) + \mathcal E^\varepsilon(t)+\mathbb{E}^\varepsilon(t)\\
&\lesssim \mathcal D^\varepsilon(0)+\mathcal E^\varepsilon(0)+\mathbb{E}^\varepsilon(0)
+\varepsilon^\frac{1}{2},
\end{align*}
at any time $t\in[0,T]$.
Applying the bounds \eq{initial_0th_energy_diff}, and \eq{initial_1st_energy_diff},
\eq{initial_2nd_energy_diff}, we conclude that the (extended) Neumann solutions
$\tilde U^\varepsilon(t)$ converge to the Cauchy solution $U(t)$ in $H^2(\mathbb{R}^3)$,
uniformly on bounded time intervals, as $\varepsilon\downarrow 0$. Thanks to Proposition
\ref{smooth_case}, this concludes
the proof of part (i) of Theorem \ref{main_result}.
\section{Exterior Dirichlet solutions}\label{gen_dir_soln}
In this and the next section $U^\varepsilon$ refers to the exterior Dirichlet
solutions; similarly for their extensions $\tilde U^\varepsilon(t,x)$.
For fixed initial data $\Phi,\, \Psi\in C^\infty_{c,\, rad}(\mathbb{R}^3)$ and any
$\varepsilon>0$ we shall derive a formula for the exterior, radial Dirichlet
solution $U^\varepsilon(t,x)$, defined for $|x|\geq\varepsilon$ and satisfying
$U^\varepsilon|_{r=\varepsilon}=0$. We refer to $U^\varepsilon$ as the {\em exterior
Dirichlet solution} corresponding to the solution $U$ of (CP) with data
$\Phi$, $\Psi$. In Section \ref{comp_dir} we will then estimate how it
(really, its extension $\tilde U^\varepsilon(t,x)$ to all of $\mathbb{R}^3$) approximates
the solution $U(t)$ in $H^1(\mathbb{R}^3)$ at fixed times.
To generate the exterior Dirichlet solution $U^\varepsilon(t,x)$ and its
extension we use the following scheme.
To smoothly approximate the original data $(\Phi,\Psi)$ with exterior
Dirichlet data we fix a smooth, nondecreasing cutoff function
$\chi:\mathbb{R}_0^+\to\mathbb{R}_0^+$ with
\begin{equation}\label{eta_props}
\chi\equiv 0 \quad\text{on $[0,1]$,}
\quad \chi\equiv 1\quad\text{on $[2,\infty)$.}
\end{equation}
Then, with $\varphi$ and $\psi$ as in \eq{rad_versns} we define
\begin{equation}\label{approx_dir_data_1}
\Phi^\varepsilon(x)\equiv \varphi^\varepsilon(|x|)
:=\chi\big(\textstyle\frac{|x|}{\varepsilon}\big)\varphi(|x|)
\end{equation}
and
\begin{equation}\label{approx_dir_data_2}
\Psi^\varepsilon(x)\equiv \psi^\varepsilon(|x|)
:=\chi\big(\textstyle\frac{|x|}{\varepsilon}\big)\psi(|x|).
\end{equation}
We refer to $(\Phi^\varepsilon,\Psi^\varepsilon)$ as the {\em Dirichlet data}
corresponding to the original Cauchy data $(\Phi,\Psi)$ for (CP).
Note that the Dirichlet data are actually defined on all of $\mathbb{R}^3$,
that they vanish identically on $B_\varepsilon$, and hence their restrictions
to the exterior domain $\{x\in\mathbb{R}^3\,:\, |x|\geq\varepsilon\}$
satisfy homogeneous Dirichlet conditions along $|x|=\varepsilon$.
The exterior Dirichlet solution $U^\varepsilon(t,x)$ is then the unique radial
solution of the initial-boundary value problem
\[\left\{\begin{array}{ll}
\square_{1+3}V =0 & \text{on $(0,T)\times\{|x|>\varepsilon\}$}\\
V(0,x)=\Phi^\varepsilon(x) & \text{for $|x|>\varepsilon$}\\
V_t(0,x)=\Psi^\varepsilon(x) & \text{for $|x|>\varepsilon$}\\
V(t,x)=0 & \text{along $|x|=\varepsilon$ for $t>0$.}
\end{array} \right.\]
We next record the solution formula for the exterior, radial Dirichlet solution
$U^\varepsilon(t,r)$ (which is simpler to derive than the formula for the
exterior Neumann solution):
\begin{equation}\label{U_eps_dir}
U^\varepsilon(t,r)=\left\{
\begin{array}{ll}
\varepsilon\leq r\leq ct+\varepsilon: & \frac{1}{2r}\left[ (ct+r) \varphi^\varepsilon(ct+r)
-(ct-r+2\varepsilon)\varphi^\varepsilon(ct-r+2\varepsilon)\right]\\
& +\frac{1}{2cr}\int_{ct-r+2\varepsilon}^{ct+r} s\psi^\varepsilon(s)\, ds\\\\
r\geq ct+\varepsilon: & \frac{1}{2r}\left[ (r+ct) \varphi^\varepsilon(r+ct)+(r-ct)\varphi^\varepsilon(r-ct)\right]\\
& +\frac{1}{2cr}\int_{r-ct}^{r+ct} s\psi^\varepsilon(s)\, ds.
\end{array}
\right.
\end{equation}
We finally extend $U^\varepsilon$ at each time to obtain an everywhere defined approximation of the
Cauchy solution $U$. The natural choice is to extend $U^\varepsilon$ continuously as zero
on $B_\varepsilon$ at each time:
\begin{equation}\label{tilde_U_eps_dir}
\tilde U^\varepsilon(t,x)=\left\{
\begin{array}{ll}
0 & \text{for $0\leq |x|\leq \varepsilon$}\\\\
U^\varepsilon(t,x) & \text{for $|x|\geq \varepsilon$.}
\end{array}
\right.
\end{equation}
We note that
\begin{equation}\label{initial_approx_N_data}
\tilde U^\varepsilon(0,x)=\Phi^\varepsilon(x),\qquad \tilde U_t^\varepsilon(0,x)=\Psi^\varepsilon(x)
\qquad\text{for all $x\in\mathbb{R}^3$.}
\end{equation}
\section{Comparing the Cauchy and exterior Dirichlet solutions}\label{comp_dir}
We proceed to estimating the $H^1$-distance
\[\|U(t)-\tilde U^\varepsilon(t)\|_{H^1(\mathbb{R}^3)},\]
and show that it vanishes as $\varepsilon\downarrow 0$.
As for exterior Neumann solutions we prefer to estimate this
difference by estimating the first order energy
\[\mathcal E^\varepsilon(t)=\mathcal E_{U-\tilde U^\varepsilon}(t)\]
as defined in \eq{ult_energ_1}. This energy bounds the $L^2$-norm
of the gradient of the difference $U-\tilde U^\varepsilon$, and it also
controls the $L^2$-norm of $U-\tilde U^\varepsilon$ itself.
The calculations for these estimates are similar to the
ones for the Neumann case in Section \ref{energies}, and will only be outlined.
First, a direct calculation similar to what was done above (using that the
energies $\mathcal E_U(t)$ and $\mathcal E^\varepsilon_{U^\varepsilon}(t)$ are both
conserved in time), shows that
\[\mathcal E^\varepsilon(t)=\mathcal E^\varepsilon_{U^\varepsilon}(0)+\mathcal E_U(0)
-\int_{|x|>\varepsilon}U_tU^\varepsilon_t+c^2\nabla U^\varepsilon\cdot\nabla U\, dx.\]
Differentiating with respect to time, integrating by parts, and applying the
Dirichlet condition for $U^\varepsilon$ yield
\begin{equation}\label{E_dot}
\dot{\mathcal E}^\varepsilon(t)=c^2\int_{|x|=\varepsilon}
U_t(t,\varepsilon)\partial_r U^\varepsilon(t,\varepsilon)\, dS.
\end{equation}
Also, with $\mathcal D^\varepsilon(t)$ defined as in \eq{L2_dist}, we have that
\eq{1_vs_0_energy} holds also in the case of Dirichlet boundary conditions.
To estimate the $H^1$-distance between the Cauchy
solution $U(t)$ and the exterior Dirichlet solution $\tilde U^\varepsilon(t)$,
we proceed to provide bounds for the initial terms $\mathcal D^\varepsilon(0)$ and
$\mathcal E^\varepsilon(0)$, as well as for the surface integral in \eq{E_dot}.
It is immediate to verify that
\[\mathcal D^\varepsilon(0)\lesssim \|\Phi^\varepsilon
-\Phi\|_{L^2(\mathbb{R}^3)}^2\leq \|\Phi\|_{L^2(B_{2\varepsilon})}^2,\]
and similarly that
\[\mathcal E^\varepsilon(0) \lesssim \|\Psi\|_{L^2(B_{2\varepsilon})}^2
+\|\nabla \Phi^\varepsilon-\nabla\Phi\|_{L^2(\mathbb{R}^3)}^2.\]
To bound the last term we recall the definition of $\Phi^\varepsilon$
in \eq{approx_dir_data_1} to calculate that
\begin{align}
\|\nabla \Phi^\varepsilon-\nabla\Phi\|_{L^2(\mathbb{R}^3)}^2
&\lesssim \int_{|x|<2\varepsilon}|\nabla \Phi|^2\, dx
+ \frac{1}{\varepsilon^2}\int_{\varepsilon<|x|<2\varepsilon}|\Phi(x)|^2\, dx \nonumber\\
&\lesssim \|\nabla \Phi\|_{L^2(B_{2\varepsilon})}^2
+ \int_{\varepsilon<|x|<2\varepsilon}\frac{|\Phi(x)|^2}{|x|^2}\, dx.\label{intrmed_energy}
\end{align}
As $\Phi$ belongs to $H^1(\mathbb{R}^3)$, Hardy's inequality (as formulated in
Lemma 17.1 in \cite{tar}) shows that $\frac{|\Phi(x)|^2}{|x|^2}$ belongs to
$L^1(\mathbb{R}^3)$, so that the Dominated Convergence Theorem yields
\[\|\nabla \Phi^\varepsilon-\nabla\Phi\|_{L^2(\mathbb{R}^3)}^2 \to0 \qquad\text{as $\varepsilon\downarrow 0$.}\]
We have thus established that
\begin{equation}\label{init_enegies}
\mathcal D^\varepsilon(0)\to 0\qquad\text{and}\qquad \mathcal E^\varepsilon(0)
\to 0 \qquad\text{as $\varepsilon\downarrow 0$.}
\end{equation}
\subsubsection{Estimating $\partial_tU(t,\varepsilon)$}
According to \eq{U} we have
\begin{equation}\label{U_t}
\partial_tU(t,\varepsilon)=\left\{
\begin{array}{ll}
t\geq \frac{\varepsilon}{c}: &\frac{c}{2\varepsilon}\left[\varphi(ct+\varepsilon)+ (ct+\varepsilon)\varphi'(ct+\varepsilon)-\varphi(ct-\varepsilon)
-(ct-\varepsilon)\varphi'(ct-\varepsilon)\right]\\\\
&\frac{1}{2\varepsilon}\left[(ct+\varepsilon)\psi(ct+\varepsilon)-(ct-\varepsilon)\psi(ct-\varepsilon)\right]\\\\\\
t\leq \frac{\varepsilon}{c}: &\frac{c}{2\varepsilon}\left[ \varphi(\varepsilon+ct) +(\varepsilon+ct)\varphi'(\varepsilon+ct)-\varphi(\varepsilon-ct)
-(\varepsilon-ct)\varphi'(\varepsilon-ct)\right]\\\\
&+\frac{1}{2\varepsilon}\left[(\varepsilon+ct) \psi(\varepsilon+ct)+(\varepsilon-ct) \psi(\varepsilon-ct)\right].
\end{array}
\right.
\end{equation}
We estimate the terms for $t\geq\frac{\varepsilon}{c}$ by 2nd order
Taylor expansion of $\varphi(ct\pm\varepsilon)$ and $\psi(ct\pm\varepsilon)$
about $\varepsilon=0$. The terms for $t\leq \frac{\varepsilon}{c}$ are estimated
by 2nd order Taylor expansion of $\varphi$ and $\psi$ about zero,
and then using that $\varphi'(0)=\psi'(0)=0$.
These expansions are straightforward and are omitted. The
result is that the leading order term in \eq{U_t} for all times is $O(1)$.
We thus have that
\begin{equation}\label{U_t_bound}
|U_t(t,\varepsilon)|\lesssim 1
\qquad\text{for all $t\in[0,T]$ as $\varepsilon\downarrow 0$.}
\end{equation}
\subsubsection{Estimating $\partial_{r}U^\varepsilon(t,\varepsilon)$}
We first calculate $\partial_{r}U^\varepsilon(t,r)$ for $\varepsilon\leq r\leq ct+\varepsilon$
by using the first part of formula \eq{U_eps_dir}. Evaluating at $r=\varepsilon$
gives that
\begin{align*}
\partial_{r}U^\varepsilon(t,\varepsilon) =
\frac{1}{\varepsilon}\left[\varphi^\varepsilon(ct+\varepsilon)+(ct+\varepsilon){\varphi^\varepsilon}'(ct+\varepsilon)\right]
+\frac{(ct+\varepsilon)}{c\varepsilon}{\psi^\varepsilon}(ct+\varepsilon).
\end{align*}
Recalling the definitions of $\varphi^\varepsilon$ and $\psi^\varepsilon$ in
\eq{approx_dir_data_1}-\eq{approx_dir_data_2}, and
splitting the calculations into $t\gtrless \frac{\varepsilon}{c}$, we obtain that
\begin{equation}\label{U^eps_r_bound}
|\partial_{r}U^\varepsilon(t,\varepsilon)|\lesssim \frac{1}{\varepsilon}
\qquad\text{for all $t\in[0,T]$ as $\varepsilon\downarrow 0$.}
\end{equation}
\subsection{Convergence of exterior Dirichlet solutions}
By using \eq{U_t_bound} and \eq{U^eps_r_bound} in \eq{E_dot} we obtain that
\[|\dot{\mathcal E}^\varepsilon(t)|\lesssim \varepsilon \qquad\text{for all $t\in[0,T]$ as
$\varepsilon\downarrow 0$,}\]
such that \eq{init_enegies}${}_2$ gives
\[\mathcal E^\varepsilon(t)\to 0 \qquad\text{uniformly for $t\in[0,T]$ as
$\varepsilon\downarrow 0$.}\]
Finally, recalling that \eq{1_vs_0_energy} also holds in the Dirichlet case,
we have that \eq{init_enegies}${}_1$ yields
\[\mathcal D^\varepsilon(t)\to 0 \qquad\text{uniformly for $t\in[0,T]$ as
$\varepsilon\downarrow 0$,}\]
as well. We thus conclude that
\[\|U(t)-\tilde U^\varepsilon(t)\|_{H^1(\mathbb{R}^3)}\lesssim \mathcal D^\varepsilon(t)+
\mathcal E^\varepsilon(t)\to 0 \qquad\text{uniformly for $t\in[0,T]$ as
$\varepsilon\downarrow 0$.}\]
Thanks to Proposition \ref{smooth_case}, this concludes
the proof of part (ii) of Theorem \ref{main_result}.
\begin{bibdiv}
\begin{biblist}
\bib{bjs}{book}{
author={Bers, Lipman},
author={John, Fritz},
author={Schechter, Martin},
title={Partial differential equations},
note={With supplements by Lars G\.arding and A. N. Milgram;
With a preface by A. S. Householder;
Reprint of the 1964 original;
Lectures in Applied Mathematics, 3{\rm A}},
publisher={American Mathematical Society, Providence, R.I.},
date={1979},
pages={xiii+343},
isbn={0-8218-0049-3},
review={\MR{598466}},
}
\bib{jt1}{article}{
author={Jenssen, Helge Kristian},
author={Tsikkou, Charis},
title={Radial solutions to the Cauchy problem for $\square_{1+3}U=0$
as limits of exterior solutions},
journal={Submitted, available at https://arxiv.org/abs/1512.02297},
date={2015},
}
\bib{ra}{book}{
author={Rauch, Jeffrey},
title={Partial differential equations},
series={Graduate Texts in Mathematics},
volume={128},
publisher={Springer-Verlag, New York},
date={1991},
pages={x+263},
isbn={0-387-97472-5},
review={\MR{1223093 (94e:35002)}},
doi={10.1007/978-1-4612-0953-9},
}
\bib{sel}{book}{
author={Selberg, Sigmund},
title={Lecture Notes, Math 632, PDE},
publisher={Johns Hopkins University},
date={2001},
}
\bib{tar}{book}{
author={Tartar, Luc},
title={An introduction to Sobolev spaces and interpolation spaces},
series={Lecture Notes of the Unione Matematica Italiana},
volume={3},
publisher={Springer, Berlin; UMI, Bologna},
date={2007},
pages={xxvi+218},
isbn={978-3-540-71482-8},
isbn={3-540-71482-0},
review={\MR{2328004 (2008g:46055)}},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
Colour centers in diamond are widely studied because of their
exceptional properties as bright solid-state quantum emitters at room
temperature~\cite{Schroeder2016,Doherty2013}. Their dynamics is being
analysed in a wide variety of setups, which for instance can achieve
the strong coupling with high-finesse optical
resonators~\cite{Albrecht2013,Brouri2000,Kurtsiefer2000} and/or the
strain coupling with high-Q vibrating
structures~\cite{Brouri2000,Kurtsiefer2000,Teissier2014,Ovartchaiyapong2014,Kipfstuhl2014a,Li2015,Lee2016}
or standing mechanical
waves~\cite{MacQuarrie2013,MacQuarrie2015}. This experimental progress
makes NV centers promising candidates for realizing quantum hybrid
devices, namely, devices capable of interfacing photons, phonons, and
spin excitations in a controlled way, and can offer a wide range of
applications for quantum information
processing~\cite{Wrachtrup2006,Childress2013,Nemeto2014,Aspelmeyer2014}
and quantum
sensing~\cite{Wrachtrup2013,Hong2013,Schirhagl2014,Rondin2014,Aspelmeyer2014}.
It thus calls for identifying the perspectives for control of these
hybrid devices, which requires a systematic characterization of their
dynamics.
In this work, we theoretically analyse laser cooling of a high-Q
vibrating mode, which is strain coupled to the electronic transitions
of an NV-center in diamond and optomechanically coupled to an optical
cavity. This situation can be realised, for instance, when NV center,
high-Q mechanical mode, and optical resonators are assembled in a
monolithic diamond structure, as illustrated in Fig.~\ref{fig:1} and
recently discussed in Refs.~\cite{Kipfstuhl2014a,Li2015}. In this
setup the high-Q vibrating mode can be optomechanically cooled by the
coupling with the cavity and/or laser-cooled by the strain-coupling
with the NV-center transitions between the state
$\ket{g}\equiv\ket{{}^3A_{20}}$ and the levels
$\ket{E_x}\equiv\ket{x}$ and $\ket{E_y}\equiv\ket{y}$, sketched in
Fig.~\ref{fig:1}(b). The starting point of our study is the
theoretical model of Ref.~\cite{Kepesidis2013}, where the authors
investigated the effect of the NV multilevel structure on the dynamics
of a high-Q vibrational mode. We extend this model by including the
high-finesse mode of an optical cavity, which couples to the
electronic transitions of the NV center and to the mechanical
resonator by means of radiation pressure, and determine the laser
cooling dynamics. We focus in particular on the regime where the
linewidth of the resonances induced by the coupling with the cavity is
of the same order as the one of the electronic transitions of the NV
center. We further determine the effect of pure dephasing, which
tends to destroy the coherence of the NV-center excitations, on the
cooling dynamics. Surprisingly, we identify regimes where pure
dephasing can improve the cooling rate.
\begin{figure}[!htb]
\includegraphics[width=\linewidth]{fig1}
\caption{ \label{fig:1} (Color online) (a) A mechanical resonator,
which is also a photonic crystal, interacts with a NV center in a
diamond bulk via strain coupling. (b) The NV-center internal
level structure, including the photonic excitations: the ground
state $\ket{g}\equiv\ket{{}^3A_{20}}$ couples to the excited
states $\ket{x}\equiv\ket{E_x}$ and $\ket{y}\equiv\ket{E_y}$,
which radiatively decay at rate $\Gamma$. A mode of the
high-finesse optical cavity decays at rate $\kappa$ and drives
quasi-resonantly the transitions
$\ket{g,n+1}\to\ket{x,n},\ket{y,n}$ with $n$ the intracavity
photon number. Coefficients $g_x$ and $g_y$ denote the
corresponding vacuum Rabi frequencies. (c) Sketch of the relevant
frequencies $\delta_\text{L},\Delta_\text{c},\Delta$, as a
function of which the cooling efficiency is characterised, in
presence of a laser driving the transition $\ket{g}\to\ket{y}$
with Rabi frequency $\Omega$.}
\end{figure}
This article is organised as follows. In Sec.~\ref{sec:cav} we
review some general concepts ruling the cooling dynamics in presence
of the strong coupling with a resonator. In Sec.~\ref{sec:sys} the
theoretical model is introduced and in Sec.~\ref{sec:parameters} the
parameter regime is discussed with reference to existing experimental
realisation. In Sec.~\ref{sec:cooling} the rate equations for the
phonon dynamics are derived and in Sec.~\ref{sec:results} the cooling
rate, the asymptotic temperature, and the spectrum of resonance
fluorescence are determined and discussed in the presence and in the
absence of the coupling with the optical cavity mode. Moreover, the
cooling efficiency as a function of the dephasing rate is
analysed. The conclusions are drawn in Sec.~\ref{sec:conclusions}.
\section{General considerations}
\label{sec:cav}
Our study is motivated by an experimentally existing platform, like
the one sketched in Fig.~\ref{fig:1}. Our purpose is to investigate
whether the optomechanical coupling can help in achieving lower
temperatures than the ones that have been predicted by sideband
cooling using the strain-coupling with the NV center, see
Ref.~\cite{Kepesidis2013}. In fact, there can be an advantage by
resorting to the optical cavity if the final occupation of the
mechanical oscillator is lower than by just performing sideband
cooling with the NV center, and thus if (i) the cavity-assisted
cooling processes are sufficiently faster than the thermalization with
the external environment and yet (ii) the final occupation of the
oscillator is smaller than the one obtained by solely employing
sideband cooling, according to a protocol like the one described in
Ref.~\cite{Kepesidis2013}. This analysis draws from several works
where it was studied how the interplay between the mechanical effects
due to the coupling of an electronic transition with a laser and with
a cavity can increase the cooling efficiency of a mechanical
oscillator
\cite{Cirac:1995,Vuletic:2003,Zippilli:2005,Zippilli:2007}. There it
was found that ground state cooling can be achieved as long as the
mechanical oscillator frequency, here denoted by $\nu$, is larger than
either the linewidth of the electronic transition, $\Gamma$, or of the
optical resonator, $\kappa$. The minimal final mechanical oscillator
occupation one can achieve is then controlled by the ratio between the
linewidth of the narrower resonance, which we denote here by
$\Gamma_{\rm min} = \min{(\kappa,\Gamma)}$, and $\nu$. Accordingly,
the cooling rate $\tilde{\Gamma}$ is slower and scales with
$\Gamma_{\rm min}$.
These dynamics can be often illustrated by means of a set of rate
equations for the occupations $p_n$ of the oscillator's state with $n$
excitations ($n=0,1,2,\ldots$)~\cite{Stenholm:1986}:
\begin{equation}
\label{Eq:rate}
\dot p_n=-n(A_++A_-)p_n+(n+1)A_-p_{n+1}+nA_+p_{n-1}\,,
\end{equation}
with $\sum_np_n=1$ (see Sec.~\ref{sec:cooling} for details how this
equation is derived). Here, $nA_+$ and $nA_-$ are the rates with which
the oscillator in state $|n\rangle $ is heated and cooled,
respectively, by one phonon, and can have the form of lorentz
functions, whose linewidth is determined by the linewidth scattering
resonance.
Specific predictions for the relevant quantities, whose dynamics
Eq.~\eqref{Eq:rate} describes, can be directly extracted from the
equation for the mean phonon occupation number
$\avg{n} = \avg{a^\dag a}=\sum_nnp_n$~\cite{Stenholm:1986}:
\begin{equation}
\dot{\avg{n}} = -(\tilde{\Gamma} + \gamma)(\avg{n} - n_f).
\end{equation}
Here $\gamma$ is the thermalization rate and $n_f$ the final phonon
occupation of the mechanical mode. Finally
\begin{gather}
\label{eq:cooling_rate}
\tilde{\Gamma} = A_- - A_+
\end{gather}
is the cooling rate when $A_->A_+$, whose maximum amplitude scales as
$\tilde \Gamma \sim (\omega_r/\nu)\Gamma_{\rm min}$ with $\omega_r$
the frequency scaling the mechanical effects due to the coupling with
light (when these are due to the phase or intensity gradient of the
light wave, $\omega_r$ is the recoil frequency; Here,
$\omega_r\sim(\Lambda/\hbar)^2/\nu$, with $\Lambda$ the strength of
the strain coupling). In this regime and for $\gamma=0$ radiation
cools the vibrations to the asymptotic occupation $N_0$, which is
given by
\begin{gather}
\label{eq:mean_phonon}
N_0 = \frac{A_+}{A_- -A_+} = \frac{A_+}{\tilde{\Gamma}}\,,
\end{gather}
and whose minimum scales with $N_0\sim (\Gamma_{\rm min}/\nu)^2$.
In a solid-state environment, where the heating rate due to the
coupling with the external reservoir is not negligible, slowing down
the cooling dynamics can be detrimental. This is visible when
considering the final occupation:
\begin{equation}
\label{eq:n_f}
n_f = \frac{\tilde{\Gamma}}{\tilde{\Gamma}+\gamma} N_0 +\frac{\gamma}{\tilde{\Gamma}+\gamma} N_{\rm th}\,,
\end{equation}
where $N_{\rm th}$ is the mean phonon occupation at the temperature of
the external reservoir. Thus, maximizing the ratio
$\Gamma_{\rm min}/\gamma$ and minimizing the ratio
$\Gamma_{\rm min}/\nu$ is crucial and limits the parameter interval
where cavity-assisted cooling can improve the efficiency.
From these considerations one can generally identify the regime where
the coupling with a resonator can increase the sideband cooling
efficiency. In fact, a large cavity decay rate such that
$\kappa>\nu>\Gamma$ would increase the cooling rate
$\tilde{\Gamma}$. Yet it can also increase the asymptotic occupation
number of the mechanical mode $N_0$. On the other hand, a very good
cavity with $\kappa<\Gamma<\nu$ can allow one to achieve smaller
values of $N_0$, but at the price of decreasing $\tilde{\Gamma}$, so
that the final occupation number of the mechanical mode $n_f$ becomes
effectively larger.
The parameter regime to explore is quite large. However in general we
expect that, in the regime where laser sideband cooling is efficient,
the coupling to a resonator at linewidth $\kappa>\Gamma$ can be of
help only if it substantially increases the cooling rate keeping
$N_0<1$. The coupling to a resonator with $\kappa<\Gamma<\nu$ can help
in reaching ultralow temperatures, provided thermalization can be
neglected. In this article we limit our analysis by taking the
optimal parameters for sideband cooling of an NV center and adding the
coupling with a cavity with linewidth $\kappa\sim \Gamma$, in order to
search for possible effects which cannot be foreseen drawing from
these simple considerations. We refer the reader to
Sec.~\ref{sec:parameters} where the choice of the parameter regime is
discussed in relation to existing experimental implementations. The
cooling rate, the asymptotic temperature, and the spectrum of
resonance fluorescence are then determined and discussed in
Sec.~\ref{sec:results} in the presence and in the absence of the
coupling with the optical cavity mode. The reader who is solely
interested in the resulting cooling efficiency can skip
Sec.~\ref{sec:cooling} and jump directly to Sec.~\ref{sec:results}.
\section{\label{sec:sys}The system}
In this Section we introduce the theoretical model which is at the
basis of our study. We describe the interaction of a high-Q mechanical
resonator mode of a phononic crystal cavity, with a quantum emitter,
specifically, a NV center in diamond, and a high-finesse optical
resonator mode of a photonic crystal cavity. The NV center is
strain-coupled with the mechanical resonator and the electronic dipole
transitions strongly couple with the photonic mode. The mechanical
resonator, in turn, is optomechanically coupled to the photonic
cavity. The interactions in this system are expected to be strongly
enhanced by the co-localization in a single structure ensuring a
perfect spatial overlap between the different degrees of freedom,
which is achieved by assemblance in a monolithic diamond structure
sketched in Fig.~\ref{fig:1}(a). The system is intrinsically
dissipative due to radiative decay of the electronic excitations and
optical cavity losses. Additionally, the mechanical resonator couples
to an external thermal reservoir. We assume that it is continuously
driven by a laser field, which directly couples to an electric dipole
transition of the defect. In what follows we define the master
equation governing the dynamics of the density matrix $\rho$, which
describes the state of the composite system composed by the NV center,
and the photonic and phononic resonators.
\subsection{Basic equations}
The dynamics of the hybrid system's density operator $\rho$,
describing the state of the system composed by the internal degrees of
freedom of the NV-center, of the optical cavity mode and of the
mechanical oscillator, is governed by the master equation
$\partial_t\rho=\Lio \rho$, where superoperator $\Lio$ is defined as
($\hbar=1$):
\begin{equation}
\label{eq:mastereq}
\Lio \rho =-i \commutator{H}{\rho} + \mathcal{L}_{\rm dis}\rho\,,
\end{equation}
and which will be conveniently reported in the reference frame
rotating with the laser frequency $\omega_{\rm L}$. Below we provide
the detailed form of Hamiltonian $H$ and superoperator
$\mathcal{L}_{\rm dis}$.
\subsubsection{Unitary dynamics}
We first give the detailed form of the Hamiltonian $H$, which
generates the unitary part of the time evolution. For convenience, we
decompose it into the sum of Hermitian operators:
\begin{equation}
\label{eq:hamiltonian}
H = H_{\rm mec}+H_\text{I} + (a+a^\dag)V\,,
\end{equation}
where $a$ and $a^{\dagger}$ annihilate and create, respectively, a
mechanical vibration at frequency $\nu$, while $V$ acts on the cavity
and NV-center degrees of freedom and is specified later on. Operator
\begin{equation}
H_{\rm mec} = \nu a^\dag a
\end{equation}
is the internal energy of the mechanical resonator, while Hamiltonian
$H_\text{I}$ describes the coupled dynamics of the NV center and the
optical cavity:
\begin{eqnarray}
\label{H:I}
H_\text{I}&=&\br{ \omega_y-\omega_{\rm L}}\ketbra{y}{y} +
\br{\omega_x - \omega_{\rm L}}\ketbra{x}{x}+\br{\omega_{\rm c} - \omega_{\rm L}} c^\dag c\nonumber\\
& &+ \left[\frac{\Omega}{2}\ketbra{y}{g} + \br{g_x\ketbra{x}{g} + g_y\ketbra{y}{g}}c+ {\rm H.c.}\right]\,.
\end{eqnarray}
Here, $\omega_x$ ($\omega_y$) is the frequency splitting in the
laboratory frame between the excited state $\ket{x}$ ($\ket{y}$) and
the ground state $\ket{g}$; operators $c$ and $c^{\dagger}$ annihilate
and create, respectively, a cavity photon at frequency
$\omega_{\rm c}$ (in the laboratory frame). The splitting between
the $\ket{x}$ and $\ket{y}$ states is, for instance, due to a
non-zero strain coupling, which is not related to the mechanical
mode we consider. The frequencies appear shifted by
$\omega_{\rm L}$ since Hamiltonian $H_\text{I}$ is reported in the
reference frame rotating at the laser frequency. The second line of
Eq. \eqref{H:I} describes, from left to right, the external laser
driving the transition $\ket{g}\to\ket{y}$ with Rabi frequency
$\Omega$, while the optical mode drives the transitions
$\ket{g}\to\ket{x}$ and $\ket{g}\to\ket{y}$ with vacuum Rabi frequency
$g_x$ and $g_y$, respectively. We note that the laser polarization can
be chosen to selectively drive one electronic transition, as we do in
our model, while in general the cavity mode's polarization has a
finite projection to the dipole moment of both transitions, since this
depends on the preparation of the sample. Therefore, we generally
assume $g_x,g_y\neq 0$ unless otherwise stated. The relevant NV center
and cavity states are reported in Fig.~\ref{fig:1}(b)-(c) with the
relative detunings with respect to the laser frequencies. These are
defined as:
\begin{equation}
\label{eq:defdeltas}
\begin{aligned}
&\delta_{\rm L} = \omega_{\rm L} - \omega_{\rm y}, \\
&\Delta = \omega_{\rm x} - \omega_{\rm y}, \\
&\Delta_{\rm c} = \omega_{\rm c}- \omega_{\rm y}\,.
\end{aligned}
\end{equation}
Finally, operator $V$ is the sum of the strain and of the
optomechanical coupling of the mechanical resonator with NV center and
optical cavity, respectively. We decompose it hence into the sum
$V=V_{\rm strain}+V_{\rm om}$, where $V_{\rm strain}$ acts on the NV
degrees of freedom and reads~\cite{Fu2009}
\begin{equation}
V_{\rm strain} = \sum_{j=I,X,Z}\Lambda_j A_j \,,
\end{equation}
where $\Lambda_j$ are the strain coupling constants and the operators
$A_j$ are defined as:
\begin{equation}
\begin{aligned}
\label{eq:As}
A_I &= \ketbra{x}{x} + \ketbra{y}{y}\,, \\
A_X &= \ketbra{x}{y} + \ketbra{y}{x}\,, \\
A_Z &= \ketbra{x}{x} - \ketbra{y}{y}\,.
\end{aligned}
\end{equation}
The optomechanical coupling reads $V_{\rm om}=- \chi c^\dag c$ with
$\chi$ the optomechanical coupling constant \cite{Law1994,Law1995}.
\subsubsection{Dissipation}
The irreversible processes we consider in our theoretical description
are: (i) the radiative decay of the NV excitations and pure dephasing
of the electronic coherences, (ii) cavity losses, and (iv) the
mechanical damping rate due to the coupling of the mechanical
resonator with an external thermal reservoir. We model each of these
phenomena by a Born-Markov process described by the corresponding
superoperator, such that superoperator $\mathcal{L}_{\rm dis}$ in
Eq. \eqref{eq:mastereq} can be cast in the form
\begin{align}
\mathcal{L}_{\rm dis}=\Lio_{\Gamma} + \Lio_\kappa +\Lio_\gamma\,.
\end{align}
The individual terms read
\begin{align}
\label{eq:NVliouvillian}
\mathcal{L}_\Gamma &= \frac{\Gamma}{2} \sum_{\xi = x, y} \mathcal{D}[\ketbra{g}{\xi}]
+ \frac{\Gamma_\phi}{2} \sum_{\xi = x, y} \mathcal{D}[\ketbra{\xi}{\xi}]\,,\\
\mathcal{L}_\kappa &= \frac{\kappa}{2}\mathcal{D}[c]
\,,\\ \label{eq:mechanicalliouvillian}
\mathcal{L}_\gamma &= \frac{\gamma}{2}\left(N_{\rm th}+1 \right)\mathcal{D}[a] + \frac{\gamma}{2}N_{\rm th}\mathcal{D}[a^\dag]\,,
\end{align}
where we used the definition
\begin{equation}
\mathcal{D}[o]\rho = 2o\rho o^\dag -o^\dag o\rho -\rho o^\dag o\,,
\end{equation}
with $o=\ketbra{g}{\xi},\ketbra{\xi}{\xi},c,a,a^\dag$. The
coefficients are the radiative decay rate $\Gamma$ of the NV-center
excited states, the dephasing rate of the electronic coherences
$\Gamma_\phi$, cavity losses at rate $\kappa$, and the damping rate of
the mechanical oscillator $\gamma$. Finally,
$N_{\rm th} = \left(\exp(\nu / k_{\rm B} T) -1 \right)^{-1}$ is the
equilibrium phonon occupation number of the bath to which the
oscillator couples, with $T$ the bath's temperature.
\subsection{Spectrum of resonance fluorescence}
In what follows we will use the master equation,
Eq. \eqref{eq:mastereq}, in order to analyze the cooling efficiency of
the mechanical resonator and the spectrum of the light emitted by the
NV center at the steady state of the cooling dynamics. In order to
better characterize the parameter regime where cooling is efficient we
choose an analytical approach, which is based on a perturbative
expansion of the Lioville operator and allows us to determine the
cooling regime, the corresponding rate and the asymptotic
temperature. This approach is reported in the following Section.
Moreover, in the regimes of interest we determine the spectrum of the
scattered light, for the purpose of identifying the relevant features
in the photons which are emitted by the NV center outside of the
resonator. The spectrum of resonance fluorescence is, apart from a
constant proportionality factor, the Fourier transform of the
auto-correlation function of the electric field~\cite{Glauber2007}:
\begin{equation}
\label{eq:spectrum}
\mathcal{S}(\omega)\propto{\rm Re}\int_0^{\infty}{\rm d}\tau\,{\rm
e}^{-i\omega\tau} \langle E^{(-)}(\tau)E^{(+)}(0)\rangle_{\rm st}
\end{equation}
where $E^{(-)}(t)$ and $E^{(+)}(t)$ are the negative and positive
frequency component of the electric field at time $t$ and
$\langle\cdot\rangle_{\rm st}\equiv {\rm Tr}\{\cdot\rho_{\rm st}\}$
denotes the trace taken over the steady state density matrix
$\rho_{\rm st}$ which solves $\Lio\rho_{\rm st}=0$. The intensity of
the scattered field (away from the forward direction) is proportional
to the source field, hence in the far-field the electric field is
proportional to the sum of the operators
$\vec{d}_j\ketbra{g}{j}+{\rm H.c.}$, for $j=x,y$ where $\vec{d}_x$ and
$\vec{d}_y$ are the dipole moments of the transitions
$\ket{g}\to\ket{x}$ and $\ket{g}\to\ket{y}$, respectively (notice that
$\abs{\vec{d}_x} = \abs{\vec{d}_y}$). Since the dipole moments are
mutually orthogonal, the spectrum integrated over the full solid angle
$4\pi$ is the incoherent sum of the two components coming from the
$\ketbra{g}{x}$ and $\ketbra{g}{y}$ operators, i.e.\@\xspace the interference
term integrates to zero. With the help of the quantum regression
theorem~\cite{carmichael1} one can cast the spectrum into the form
\begin{equation}
\label{eq:spectyum1}
\mathcal{S}(\omega) \propto \sum_{j=x,y}\Re{\Tr{\brrr{\ketbra{j}{g}
\brr{i(\omega-\omega_{\rm L}) -\Lio}^{-1} \ketbra{g}{j} \rho_{\rm st}}}}.
\end{equation}
In this work we numerically determine the spectrum for the parameter
regimes of interest.
\section{Parameter regime}
\label{sec:parameters}
In order to justify the experimental relevance of the cooling dynamics
we discuss in the rest of this article, we now relate the theoretical
model to existing experimental realisations and identify the parameter
regime which we will consider in our analysis.
{\it Optical resonator}. A structure like the one discussed here can
be found for instance in a so-called phoxonic crystal (PxC), which
co-localizes confined optical and mechanical resonator
modes~\cite{Kipfstuhl2014a}. Photonic crystals are formed by a
periodic modulation of the refractive index (in this case air holes in
diamond), resulting in the formation of optical bands similar to
electronic band structures in solids. A local defect like e.g.\@\xspace a
variation of the hole diameters along the PxC structure perturbs the
perfect periodicity and gives rise to an optical cavity mode. So far,
fabrication imperfections limit experimental quality factors to $10^4$
at visible wavelengths suitable for the interaction with colour
centers in diamond and up to $10^5$ in the telecom band around
$1550\unit{nm}$~\cite{Hausmann2013a,Burek2014,Li2015,Burek2015}. Nevertheless,
simulations of one-dimensional photonic crystal cavities designed for
visible light predict quality factors up to $10^7$ and mode volumes
around $1$ cubic wavelength with cavity loss rate in the range
$\kappa\sim\! 10\unit{MHz}-1\unit{GHz}$~\cite{Kipfstuhl2014a}.
{\it Mechanical resonator}: In a PxC a periodic variation of the
elastic modulus creates a mechanical band structure and a suitable
variation of the regular pattern allows for a localized mode of the
mechanical resonator. Recent experiments with structures at
mechanical frequencies of $6\unit{GHz}$ with optical properties
suitable for telecom wavelengths show mechanical quality factors of
$10^3$~\cite{Burek2015}. Numerical modeling shows that modes with
frequencies in the range $10-20\unit{GHz}$ with quality factors
reaching $10^7$ can be achieved with an effective mass of
$10^{-16}\unit{kg}$ for structure dimensions matching visible
wavelengths with the confined optical mode~\cite{Kipfstuhl2014a}. The
parameters we choose are consistent with assuming mechanical
frequencies of the order of $1-10\unit{GHz}$ and a quality factor of
the order of $10^6-10^7$, giving a damping rate $\gamma$ of few
$\unit{kHz}$. The strain coupling constants are taken to be of the
order of
$1-10\unit{MHz}$~\cite{Kepesidis2013,Tamarat2006,Doherty2011}. The
optomechanical coupling constant $\chi$ is taken to be of the order of
few $\unit{MHz}$~\cite{Kipfstuhl2014a}.
{\it NV center}. Figure~\ref{fig:1}(b)-(c) reports the relevant level
structure of the NV center in diamond. In absence of strain coupling,
the $m_s=0$ ground state $\ket{{}^3A_{20}}$ can be selectively coupled
to the excited states $\ket{E_{x,y}}$, which have zero spin angular
momentum. While the ground state is much less sensitive to lattice
distortion, these excited states are highly susceptible to external
perturbations~\cite{Maze2011,Doherty2011,Doherty2013}. Axial strain
(parallel to the NV center axis, equivalent
$\left\langle 111 \right\rangle$ crystal direction) leads to an
additional splitting between ground and excited states as well as
between the $m_s=0$ and $m_s=\pm1$ levels in the ground state. Radial
strain (perpendicular to NV axis) mixes the excited state levels $E_x$
and $E_y$ and leads to a splitting of the new states $E_x^*$ and
$E_y^*$ ($m_s=+1^*$ and $m_s=-1^*$). The effect of strain coupling on
the excited states is several orders of magnitude larger than on the
ground state and hence dominates the strain-induced modification of
the NV's optical properties. Therefore, we restrict our model to the
interaction between the mechanical resonator mode and the transition
coupling the ground state $\ket{g}\equiv\ket{{}^3A_{20}}$ to the
excited states $\ket{x}\equiv\ket{E_x}$ and
$\ket{y}\equiv\ket{E_y}$. For the excited states we take the radiative
decay rate
$\Gamma\!\sim\!100\unit{MHz}$~\cite{Sipahigil2012,Bernien2012}. The
interaction between the NV transitions and the $71\unit{meV}$ lattice
phonon modes~\cite{Gali2011} changes the energy of the $\ket{x}$ and
$\ket{y}$ states and can thus give rise to a dephasing mechanism of
the electronic coherence~\cite{Albrecht2013,Betzholz2014}. In our
model we neglect the mixing between the states and consider only pure
dephasing with rates of the order of $\Gamma_\phi \sim 100\unit{MHz}$,
which can be achieved in bulk diamond at temperatures lower than
$10\unit{K}$~\cite{Abtew2011,Fu2009}. We restrict the frequency of
the mechanical resonator mode to $\nu = 2\pi\times 1\unit{GHz}$ in
order to avoid coupling to NV excited states other that $E_x$ and
$E_y$. As the optical cavity mode should still be near resonant on the
optical transition of the NV at $637\unit{nm}$ this doesn't correspond
to a real structure design for the full threefold
hybrid-system. However, we still model this artificial parameter set
in order to obtain qualitative results on the nature of the
interaction.
{\it Cooling regime}: The analysis of the cooling efficiency is
performed by determining the cooling rate $\tilde{\Gamma}$ and the
ideal asymptotic occupation number of the mechanical mode $N_0$ as a
function of the tunable parameters, which we take here to be the
frequency splitting of the electronic excited states and the laser
frequency, corresponding to changing $\delta_{\rm L}$,
$\Delta_{\rm c}$, and $\Delta$. The analysis is performed by
searching for the parameter regime where the asymptotic occupation
number $N_0<1$ and the cooling rate $\tilde{\Gamma}$ is maximized, in
order to realise regimes where the radiative cooling can overcome
thermalization by the external reservoir. This constrains the range of
parameters. A necessary condition for performing ground state cooling
is the presence of a resonance whose linewidth $L$ is smaller than the
trap frequency \cite{Eschner:2003}, which poses an upper bound to
$L$. Moreover, the cooling rate shall exceed the thermalization
rate. Since the cooling rate is proportional to the effective
linewidth of the cooling transition, this condition sets a lower bound
to $L$. If one performs optomechanical cooling by driving the optical
resonator, then $L=\kappa$. In absence of the resonator, the
mechanical oscillator can be cooled by driving the NV center
transitions with a laser and $L=\Gamma$. When the NV center
transitions also couple with the optical cavity, then $L$ is a linear
interpolation of the cavity linewidth $\kappa$ and of the NV
transition linewidth $\Gamma$, and varies between $\Gamma$ and
$\kappa$ \cite{Zippilli:2005} (smaller linewidths could be achieved by
coupling to other stable electronic transitions, which in our system
are not considered \cite{EIT:2000,CEIT:2012}).
In order to get a relatively small phonon occupation of the bath
$N_{\rm th}$ we take a large mechanical frequency,
$\nu\sim 2\pi\times 1\unit{GHz}$, and thus for our parameter choice
$\Gamma<\nu$. We then fix the cavity linewidth $\kappa\simeq\Gamma$.
\section{Effective dynamics of the mechanical resonator}
\label{sec:cooling}
For the parameter regime we consider all characteristic frequencies
characterizing the coupling of the mechanical resonator with NV center
and optical cavity are much smaller than the mechanical resonator
eigenfrequency
$\Lambda_{I},\Lambda_{X},\Lambda_{Z},\chi\bar n_{\rm c}\ll \nu$
($\bar n_{\rm c}$ being the mean intracavity photon occupation
number). This justifies a perturbative treatment, which allows us to
eliminate the degrees of freedom of NV and optical cavity from the
dynamics of the mechanical oscillator in second-order perturbation
theory. By means of this procedure we derive an effective master
equation for the mechanical resonator only, which allows us to
determine the parameter regime where the vibrations are cooled, the
corresponding cooling rate and the asymptotic vibrational state.
\subsection{\label{sec:pert}Perturbative expansion}
We derive a closed master equation for the mechanical oscillator
starting from Eq.~\eqref{eq:mastereq} and assuming that the coupling
frequencies, which scale the operator $a+a^\dag$, are much smaller
than $\nu$. This can be summarized by the inequality $\alpha\ll\nu$,
with $\alpha=\Lambda_{I},\Lambda_{X},\Lambda_{Z},\chi\bar n_{\rm c}$
and $\bar n_{\rm c}$ the mean intracavity photon occupation number. We
then perform perturbation theory in second order in the small
parameter $\alpha/\nu$. We further assume that the incoherent dynamics
of the oscillator due to the coupling with the environment is
sufficiently slow that the occurrence of these processes during a
scattering process can be discarded. This requires that
$\gamma N_{\rm th}\ll \alpha$, which for the parameters considered in
this work is valid also at room temperature, so that we treat it in
first order.
According to these considerations we split the Liouville operator as
$$\Lio=\Lio_0+\mathcal{V}+\Lio_\gamma\,,$$
with $\Lio_0= \Lio_{\rm E}+ \Lio_{\rm I}$, where $\Lio_{\rm E}$ and
$\Lio_{\rm I}$ are the Liouville operators that generate the uncoupled
mechanical oscillator and internal (NV center + optical cavity)
dynamics, respectively, while $\mathcal{V}$ describes the coupling
between mechanical and internal degrees of freedom. In detail,
\begin{align}
\label{eq:LiouvillianDecompose}
\Lio_{\rm E}\rho&=-i [H_{\rm mec},\rho],\\
\label{eq:LiouvillianDecompose2}
\Lio_{\rm I}\rho &=-i [H_{\rm I},\rho] + \Lio_{\Gamma}\rho + \Lio_\kappa \rho,\\
\label{eq:LiouvillianDecompose:L1}
\mathcal{V}\rho &= -i [V(a+a^\dagger), \rho]\,.
\end{align}
We formally eliminate the coupling between mechanical resonator and
internal degrees of freedom as done for instance in
Refs.~\cite{Javanainen1984,Cirac1992,Morigi2003a,Zippilli:2005}. We
first introduce the superoperators $\mathcal{P}_k$ such that
\begin{align}
\mathcal{P}_k\rho=\sigma_{\rm st}\sum_{n=0}^\infty|n\rangle\!\langle
n+k|\langle n|\mu|n+k\rangle\,,
\end{align}
with $\mu=\Tr_{\rm I}\{\rho(t)\}$ the reduced density matrix,
$\Tr_{\rm I}\{\cdot\}$ being the trace over the internal degrees of
freedom, $\ket{n}$ the eigenstates of the mechanical oscillator,
$k=0,\pm1,\pm2,...\,$ ($k\ge -n$) and $\sigma_{\rm st}$ the steady
state for the internal degrees of freedom:
$\Lio_{\rm I}\sigma_{\rm st}=0$. Applying $\P_k$ to the master
equation~\eqref{eq:mastereq}, with the definitions of the
superoperators~\eqref{eq:LiouvillianDecompose}-\eqref{eq:LiouvillianDecompose:L1},
in a second-order perturbative expansion in parameter $\alpha/\nu$ and
first order in $\gamma (N_{\rm th}+1)$, leads to the equation
\begin{align}
\label{eq:effeqwithprojector}
\pd{}{t}\mathcal{P}_k \rho&=\Big\{ik\nu + \P_k\mathcal{V}\big(ik\nu
- \Lio_0\big)^{-1}\mathcal{Q}_k\mathcal{V}\P_k\Big\}\P_k\rho+\Lio_{\gamma}\mathcal{P}_k \rho\,,
\end{align}
with $\mathcal{Q}_k=1-\P_k$ and $1$ is here the superoperator whose
action is the identity on both sides of the density matrix. The master
equation for the reduced density matrix $\mu$ is obtained after
tracing out the internal degrees of freedom in
Eq.~\eqref{eq:effeqwithprojector} and reads
\begin{equation}
\label{eq:effectivedynmecexplit}
\begin{aligned}
\dot\mu = -i\bar{\nu}\commutator{a^\dag a}{\mu}+
\frac{A_-}{2}\mathcal{D}[a]\mu +
\frac{A_+}{2}\mathcal{D}[a^\dag]\mu+\Lio_{\gamma}\mu\,.
\end{aligned}
\end{equation}
The rates $A_\pm$ are defined as
\begin{gather}
\label{eq:A+A-nubar}
A_\pm = 2\,\Re{s(\mp \nu)}, \\
\bar{\nu} = \nu+\Im{s(\nu)} + \Im{s(-\nu)}\,,
\end{gather}
with
\begin{equation}
\label{eq:sofv}
s(v) = \int_0^\infty dt\, e^{ivt}\avg{V\exp{\br{\Lio_{\rm I} t}}V}_{\rm st}\,,
\end{equation}
which is the Fourier component at frequency $\nu$ of the
autocorrelation function of operator $V$, where the average
$\avg{\cdot}_{\rm st}$ is taken in the steady state $\sigma_{\rm st}$.
The diagonal elements of Eq. \eqref{eq:effectivedynmecexplit} give a
set of rate equations for the occupation $p_n=\langle n|\mu|n\rangle$
of the phonon state $\ket{n}$, which are reported in
Eq. \eqref{Eq:rate}.
\section{Results}
\label{sec:results}
In this section we characterize the parameter regimes in which the
mechanical resonator is cooled by photon scattering process in the
setup of Fig.~\ref{fig:1}(a). We focus on the range of parameters
discussed in Sec. \ref{sec:parameters}. We consider laser cooling of
the mechanical resonator by strain coupling with the NV center and
analyse how the cooling dynamics is affected by the presence of the
optical resonator and of dephasing. The results we report are compared
to the predictions in absence of the optical resonator and for
vanishing dephasing. This latter case has been extensively discussed
in Ref. \cite{Kepesidis2013} and we refer the interested reader to it
for a detailed discussion of the predicted dynamics in that specific
limit.
\subsection{Cavity-assisted cooling}
We now analyse how laser cooling dynamics of the mechanical resonator
by strain coupling with the NV center is affected by the presence of
the optical resonator. In order to better understand the role of the
resonator, we first discard thermal effects and dephasing (setting
$\gamma=\Gamma_\phi=0$).
\begin{figure}[!htb]
\includegraphics[]{fig2}
\caption{ \label{fig:2} (Color online) Predictions on the cooling
efficiency extracted from the rate equation, Eq. \eqref{Eq:rate},
for laser cooling of the mechanical resonator by driving the NV
center with a laser (left panel) and by additionally coupling the
dipole transitions to an optical cavity (right panel). (a) and (b)
show the cooling rate $\tilde{\Gamma}$,
Eq.~\eqref{eq:cooling_rate} in units of $\Lambda^2/\nu$, (c) and
(d) the asymptotic occupation $N_0$ of the vibrational mode,
according to Eq. \eqref{eq:mean_phonon}, as a function of the
excited level splitting $\Delta$ and the laser detuning
$\delta_{\rm L}$ (in units of $\nu$). The white region are heating
regions ($\tilde{\Gamma}<0$) or where $N_0>1$. The parameters for
the left panel are $\Omega=0.1\nu$, $\Gamma=1.6\times 10^{-2}\nu$,
$\Gamma_\phi=0$,
$\Lambda_I=0,\ \Lambda_X=\Lambda_Z=\chi=\Lambda=0.1\Gamma$ and
$g_x=g_y=0$. In the right panel we take the same parameters except
for $g_x=g_y=\kappa=\Gamma$. The cavity frequency is fixed to the
value $\Delta_{\rm c}=8.5\times 10^{-2}\nu$ (see text).}
\end{figure}
\begin{figure}[!htb]
\includegraphics[]{fig3}
\caption{ \label{fig:3} (Color online) (a) Cooling rate
$\tilde{\Gamma}$ and (b) asymptotic occupation $N_0$ of the
vibrational mode as a function of $\delta_{\rm L}$ for the same
parameters as in Fig.~\ref{fig:2} and $\Delta=\nu$. The dashed
(solid) line corresponds to the predictions in absence (presence)
of the coupling to the cavity. The shaded region indicates the
regime where the resonator is heated by the radiative processes
($\tilde{\Gamma}<0$) or where $N_0>1$. }
\end{figure}
Figure~\ref{fig:2} displays the cooling rate $\tilde{\Gamma}$ and the
mean vibrational number at the asymptotics $N_0$ as a function of
$\delta_{\rm L}$ and $\Delta$ in absence (left panels) and in presence
of the optical cavity (right panel). Both plots show that the cooling
rate is maximum, and the final occupation minimum, along the lines
$\delta_{\rm L}=-\nu$ and $\delta_{\rm L}=\Delta-\nu$. In the first
case cooling is achieved by setting the laser frequency to the value
$\omega_{\rm L}=\omega_y-\nu$, hence resonantly driving the transition
$|g,n\rangle\to |y,n-1\rangle$ (red sideband). In the second case the
laser frequency is $\omega_{\rm L}=\omega_x-\nu$, so that the
transition $|g,n\rangle\to |x,n-1\rangle$ is resonantly driven by an
effective process, which combines the laser and the strain
coupling. For most values of the detuning $\Delta$ the excitation of
the intermediate state $|y\rangle$ is virtual, except for
$\Delta=\omega_x-\omega_y=\nu$. This latter case corresponds to the
vertical line visible in both figures, where cooling results to be
efficient. These properties have been identified and discussed in
Ref.~\cite{Kepesidis2013} and do not depend on the coupling with the
resonator. The curves in Fig.~\ref{fig:3} show the cooling rate and
the minimum phonon occupation as a function of $\delta_{\rm L}$ after
fixing the detuning $\Delta=\nu$. Some (relatively small) differences
are visible close to the values $\delta_{\rm L}=0$ and
$\delta_{\rm L}=-\nu$, which are due to the level splitting induced by
the strong coupling with the resonator: for this choice of $\Delta_c$,
in fact, the cavity drives almost resonantly the transition
$|g\rangle\to |y\rangle$.
We have tested that the value of the detuning $\Delta_{\rm c}$, and
thus of the cavity frequency, in Figs. \ref{fig:2} and \ref{fig:3},
leads to the best results by comparing cooling rate and final
temperature for different values of $\Delta_{\rm c}$. The results of
this analysis are summarized in Fig.~\ref{fig:4}, which displays (a)
the maximum cooling rate (maximized by varying $\Delta$ and
$\delta_{\rm L}$ by keeping $\Delta_c$ fixed). The mean phonon
occupation in (b) and the mean intracavity photon number in (c) are
reported for the corresponding values of $\Delta$ and
$\delta_{\rm L}$, at which $\tilde{\Gamma}$ is maximum. These plots
show that maximal cooling rates are found for $\Delta_c\simeq 0$. We
verified that the curves do not differ substantially if instead we
search for $\Delta_{\rm c}$ by minimizing the mean phonon
number. Therefore, the contour plots in Fig. \ref{fig:2}(b) and (d)
show the optimal cooling rate and temperature in presence of the
resonator. On the basis of the comparison with the plots on the left
panels, we can thus conclude that the coupling with the cavity does
not substantially improve the cooling efficiency for the chosen
parameter regime.
\begin{figure}[!htb]
\includegraphics[]{fig4}
\caption{ \label{fig:4} (Color online) (a) Maximum cooling rate
$\tilde{\Gamma}_{\rm max}$ in presence of the resonator as a
function of $\Delta_{\rm c}$. The value $\tilde{\Gamma}_{\rm max}$
has been calculated by varying $\delta_{\rm L}$ and $\Delta$ and
keeping $\Delta_{\rm c}$ fixed. Subplot (b) displays the
corresponding value of $N_0$ and (c) the mean intracavity photon
number. The parameters are: $\Omega=0.1\nu$, $\Gamma_\phi=0$,
$\Gamma=\kappa=g_x=g_y=1.6\times10^{-2}\nu$,
$\Lambda_I=0,\ \Lambda_X=\Lambda_Z=\chi=\Lambda=0.1\Gamma$. The
dashed lines in (a) and (b) indicate the maximum cooling rate and
corresponding value of $N_0$ in absence of the optical
resonator. In the latter case $\tilde{\Gamma}_{\rm max}$ is
maximum for $\Delta\approx0.93\nu$ and
$\delta_{\rm L}\approx-3.5\times 10^{-2}\nu$.}
\end{figure}
We now analyse how the spectrum of resonance fluorescence is modified
by the coupling with the resonator. We focus on the light emitted once
the system has reached the stationary state. Figure~\ref{fig:5}
displays the resonance fluorescence spectrum in absence and in
presence of the optical cavity for the parameters of Fig.~\ref{fig:3}
with $\delta_{\rm L}=0$. To better understand how the cavity modifies
the dynamics, we first discuss the spectrum in absence of the
cavity. In this case we observe the three broad resonances around
$\omega=\omega_{\rm L}$. These are due to inelastic processes in
which the motion is not involved and can be interpreted as a
Mollow-type triplet~\cite{Mollow1969}. We further observe the narrow
resonances at $\omega = \omega_{\rm L} \mp \nu$, which are the red and
the blue motional sidebands of the elastic peak.
Subplots~\ref{fig:5}(b) and~\ref{fig:5}(c) report the details of the
sidebands of the elastic peak. These spectral components correspond to
the photons emitted in the processes where a phonon is created
($\omega_{\rm L} - \nu$) or destroyed ($\omega_{\rm L} + \nu$) in the
mechanical resonator. The motional sideband has a width of the order
of $\propto \Lambda_X^2$, and appears on a broader background with
linewidth $\approx \Gamma$. Our analysis shows that this structure is
due to the fact that mechanical effects are dominated by the strain
coupling $A_X$, which mixes the two excited states. For our parameter
choice, where $\Delta = \nu$, this coupling is weak but resonant so
that the effect of the strain coupling is particularly enhanced.
Figure~\ref{fig:5}(d)-(f) displays corresponding spectra of resonance
fluorescence in presence of the cavity. The significantly
different features are due to the modified dressed state structure
because of the strong coupling between cavity and NV center, while
for both cases the cooling (heating) processes are dominated by
emission along the transition $|x\rangle\to |g\rangle$
($|y\rangle\to |g\rangle$).
\begin{figure*}[!htb]
\includegraphics[]{fig5}
\caption{ \label{fig:5} (Color online) Spectrum of light emitted by
the NV center at the asymptotics of the laser cooling
dynamics. The upper panels correspond to the parameter regime of
Fig.~\ref{fig:2}(a)(c) (no cavity), the lower panels to the
parameter regime of Fig.~\ref{fig:2}(b)(d) (cavity assisted
cooling). The dashed (dashed-dotted) line correspond to the
emission from the transition $g \leftrightarrow x$
($g \leftrightarrow y$), the solid line correspond to the sum of
these two contributions. Here, we took $\Delta=\nu$ and
$\delta_{\rm L}=0$. Panels (b), (c), (e) and (f) show the details
of the sidebands.}
\end{figure*}
The summary of this analysis is that the effect of the optical
resonator on the cooling dynamics can consist in a very small
improvement of the cooling efficiency. This result, which seems to
contrast with previous investigations where the effect of the cavity
on the cooling efficiency was relevant \cite{Zippilli:2005,CEIT:2012},
can be understood when considering that (i) the loss rate of the
resonator and the radiative decay rate of the electronic excitations
have been chosen to be of the same order of magnitude, and (ii) the
cooperativity $C=g^2/\kappa\Gamma\sim 1$, so that the level splitting
induced by the coupling with the resonator is of the order of the loss
rate $\kappa$. Because of (i) the coupling with the resonator gives
rise to an effective level structure where the linewidths of all
excited levels is of the same order of magnitude. Since for sideband
cooling the linewidth determines both the cooling rate as well as the
final temperature, the improvement of the cooling efficiency by
coupling this level structure to a resonator is incremental. Because
of (ii), the level splitting induced by the coupling with the cavity
does not exceed the linewidth of the resonances, so that the regime of
optimal detunings is essentially the same as without the cavity.
\subsection{Dephasing-assisted cooling}
We now analyse the effect of other noise sources on the cooling
efficiency, and consider in particular dephasing, which is an
important source of loss of coherence in solid-state systems. We here
discard the coupling with the optical resonator and calculate the
cooling efficiency when $\Gamma_\phi\neq 0$. Figure~\ref{fig:6}
compares the cooling rate and final occupation for $\Gamma_\phi=0$
(left panel) and $\Gamma_\phi=\Gamma$ (right panel). We observe that
pure dephasing decreases the cooling efficiency when cooling is
achieved by tuning the laser to the red sideband of the dressed
states. Nevertheless, the cooling region is larger and the dependence
on the exact values of the experimental parameters is less
pronounced. Moreover, the cooling performance is enhanced in most
parts of parameter landscape. Figures~\ref{fig:7}(a) and (b) compare
the cuts along the line $\Delta=\nu$: one clearly sees that the case
of $\Gamma_\phi=\Gamma$ outperforms the case when
$\Gamma_\phi=0$. This occurs over almost the full range of
$\delta_{\rm L}$ in terms of both cooling rate and minimal phonon
number. We have checked that the value $\Gamma_\phi=\Gamma$ is close
to the optimal dephasing rate. We also found the range of values in
which the dephasing has a beneficial effect on the cooling spans till
several $\Gamma$ (see dotted line, which shows the predictions for
$\Gamma_\phi=10\Gamma$).
\begin{figure}[!htb]
\includegraphics[]{fig6}
\caption{ \label{fig:6} (Color online) Predictions on the cooling
efficiency extracted from the rate equation, Eq. \eqref{Eq:rate},
for laser cooling of the mechanical resonator by driving the NV
center with a laser in absence (left panel) and in presence of
pure dephasing (right panel). (a) and (b) show the cooling rate
$\tilde{\Gamma}$, Eq. \eqref{eq:cooling_rate} in units of
$\Lambda^2/\nu$, (c) and (d) the asymptotic occupation $N_0$ of
the vibrational mode, according to Eq. \eqref{eq:mean_phonon}, as
a function of the excited level splitting $\Delta$ and the laser
detuning $\delta_{\rm L}$ (in units of $\nu$). The white area are
heating regions ($\tilde{\Gamma}<0$) or where $N_0>1$. The
parameters are $\Omega=0.1\nu$, $\Gamma=1.6\times10^{-2}\nu$,
$\Lambda_I=0,\ \Lambda_X=\Lambda_Z=\chi=\Lambda=0.1\Gamma$, and
(left panel) $\Gamma_{\phi}=0$, (right panel)
$\Gamma_{\phi}=\Gamma$.}
\end{figure}
\begin{figure}[!htb]
\includegraphics[]{fig7}
\caption{ \label{fig:7} (Color online) (a) Cooling rate
$\tilde{\Gamma}$ and (b) asymptotic occupation $N_0$ of the
vibrational mode as a function of $\delta_{\rm L}$ for the same
parameters as in Fig.~\ref{fig:6} and $\Delta=\nu$. The dashed
line corresponds to the predictions in absence of dephasing. The
solid (dotted) line corresponds to the predictions when the
dephasing rate is $\Gamma_\phi=\Gamma$
($\Gamma_\phi=10\Gamma$). The shaded region indicates the regime
where the resonator is heated by the radiative processes
($\tilde{\Gamma}<0$) or where $N_0>1$.}
\end{figure}
The effect of dephasing is also visible in the spectrum of resonance
fluorescence. We observe in Fig.~\ref{fig:8} for $\delta_{\rm L}=0$ a
broadening of the background at the motional sidebands, which now
scale with $\approx \Gamma + \Gamma_\phi$. This linewidth is indeed
the cooling rate, which results to be enhanced by the presence of pure
dephasing.
\begin{figure*}[!htb]
\includegraphics[]{fig8}
\caption{ \label{fig:8} (Color online) Spectrum of light emitted by
the NV center at the asymptotics of the laser cooling
dynamics. The parameters are the same as in Fig.~\ref{fig:6}(b)(d)
(dephasing assisted cooling with $\Gamma_\phi=\Gamma$). The dashed
(dashed-dotted) line correspond to the emission from the
transition $g \leftrightarrow x$ ($g \leftrightarrow y$), the
solid line correspond to the sum of these two contributions. Here,
we took $\Delta=\nu$ and $\delta_{\rm L}=0$. Panels (b) and (c)
show the details of the sidebands.}
\end{figure*}
We understand this behaviour since pure dephasing increases the width
of the excited states $\ket{x}$ and $\ket{y}$ without increasing their
decay rate. Thus it increases the excitation probability. Since this
cooling scheme is optimal when population is transferred to the
excited state, then pure dephasing leads to larger transition rates,
and thus larger cooling rate. This reasoning works within a certain
parameter interval: dephasing rates exceeding the Rabi frequency, in
fact, tend to suppress population transfer and thus are detrimental.
The beneficial role of pure dephasing on the cooling efficiency can be
best illustrated by analysing the final mean occupation for different
temperatures of the bath. Figure~\ref{fig:9} illustrates how dephasing
can improve the cooling efficiency over a large parameter regime,
flattening out the minimum of $n_f$ (Eq.~\eqref{eq:n_f}) as a function
of the frequency of the driving laser.
\begin{figure}[!htb]
\includegraphics[]{fig9}
\caption{\label{fig:9} (Color online) Final phonon number of the
mechanical resonator with $\nu=2\pi\times 1\unit{GHz}$ and a
quality factor $Q=\nu/\gamma=10^7$, Eq.~\eqref{eq:n_f}, for (a)
$T=0.1\unit{K}$ ($N_{\rm th}\approx1.6$) and (b) $T=4\unit{K}$
($N_{\rm th}\approx83$), for the same parameter regime of
Fig.~\ref{fig:7}(b). The dashed (solid) line corresponds to the
predictions in absence (presence) of pure dephasing. The black
dotted lines correspond to $n_f=N_{\rm th}$. The shaded region
indicates the regime where the resonator is heated by the
radiative processes.}
\end{figure}
\section{Discussion and Conclusions}
\label{sec:conclusions}
We have analysed the cooling efficiency of a mechanical resonator
which is laser cooled by the strain coupling with a NV center. The
cooling dynamics is essentially due to the strain coupling with the NV
center and the parameter regime is such that the resolved-sideband
cooling can be performed by driving the NV center electronic
resonances. In this regime we have analysed the effect of the coupling
to an optical resonator, and found that it does only incrementally
improve the cooling efficiency. We have further shown that pure
dephasing can make the cooling dynamics more robust against parameter
fluctuations, without affecting the overall efficiency, as long as the
dephasing rate does not exceed the driving strength of the laser.
In our analysis the optomechanical coupling was a small effect. It can
be increased in configurations where the cavity is driven: in this
case the optomechanical coupling would cool the resonator according to
the dynamics explored in
Refs.~\cite{WilsonRae2008,Marquardt2007}. Another interesting
possibility is to drive both optical cavity and NV center for large
cooperativity: In this situation phonon excitation or absorption can
be realised by means of two excitation paths, that can interfere. This
interference depends on the relative phase between the lasers and
could be a control parameter for realising multi-wave mixing.
\begin{acknowledgments}
We thank Peter Rabl, Stefan Sch\"utz, and Christoph Becher for
useful discussions and helpful comments. This work was partially
financially supported by the DFG ("Optomechanical cavity Quantum
Electrodynamics with colour centers in diamond"), the German
Ministry of Education and Research (BMBF, "Q.com"), the DFG
Forschergruppe FOR 1493, and the GradUS program of Saarland
University.
\end{acknowledgments}
|
\section{Introduction}
Numerous classical problems in extremal graph theory have highly structured extremal configurations. For example, Tur\'an~\cite{Tu} in 1941 proved that $\ex(n,K_k)$, the maximum number of edges in an $n$-vertex $K_k$-free graph, is attained only by the balanced complete $(k-1)$-partite graph, known now as the \emph{Tur\'an graph} $T_{n,k-1}$. Motivated by the fact that the Tur\'an graph is particularly symmetric, admitting a $(k-1)$-partition into linear-sized independent sets, Erd\H os and S\'os~\cite{ES} introduced \emph{Ramsey-Tur\'an} type questions, where they investigated the maximum size of a $K_k$-free graph $G$ with the additional condition that $\alpha(G)=o(|G|)$.
Denote by $\RT(n,K_k, o(n))$ the Ramsey-Tur\'an function for $K_k$, i.e.~the maximum size of an $n$-vertex $K_k$-free graph with independence number $o(n)$. In 1970, Erd\H{o} s and S\'os~\cite{ES} determined $\RT(n,K_k,o(n))$ for every odd $k$. The problem becomes much harder when an even clique is forbidden. For $k=4$, Szemer\'edi~\cite{Sz-k4}, using the regularity lemma, proved that $\RT(n,K_4,o(n))\le n^2/8+o(n^2)$. It had remained an open question whether $\RT(n,K_4,o(n))=\Omega(n^2)$. Bollob\'as and Erd\H{o} s, in their seminal work~\cite{BE}, constructed a dense, $K_4$-free graph with sub-linear independence number, matching the upper bound above (see Section~\ref{sec-pre} for more details). For all even $k$, the order of magnitude of $\RT(n,K_k,o(n))$ was finally determined by Erd\H{o} s, Hajnal, S\'os and Szemer\'edi~\cite{EHSSz} in 1983. See~\cite{SS} for a survey and~\cite{Jozsi-1, Jozsi-2} for more recent developments on this topic.
In this paper, we will study Ramsey-Tur\'an extensions of some classical results, whose extremal graphs are close to the Tur\'an graph. See e.g.~\cite{BMSh} for one such extension of a graph tiling problem.
\subsection{Edge-colorings forbidding monochromatic cliques}
Denote by $F(n,r,k)$ the maximum number of $r$-edge-colorings that an $n$-vertex graph can have
without a monochromatic copy of $K_k$. A trivial lower bound is given by $T_{n,k-1}$ as every $r$-edge-coloring of a $K_k$-free graph is monochromatic $K_k$-free: $F(n,r,k)\ge r^{\ex(n,K_k)}$. Erd\H{o} s and Rothschild~\cite{E(R)} in 1974 conjectured that, for sufficiently large $n$, the above obvious lower bound is optimal for 2-edge-colorings. This was verified for $k=3$ by Yuster~\cite{Y}. In 2004, Alon, Balogh, Keevash and Sudakov~\cite{ABKS} settled this conjecture in full, proving that, for all $k\ge 3$ and sufficiently large $n$, the Tur\'an graph $T_{n,k-1}$ maximizes the number of $2$-edge-colorings and $3$-edge-colorings with no monochromatic $K_k$ among all graphs:
\begin{eqnarray}\label{eq-ER-orig-lower}
F(n,2,k)=2^{\ex(n,K_k)} \quad \mbox{ and } \quad F(n,3,k)=3^{\ex(n,K_k)}.
\end{eqnarray}
For $4$-edge-colorings, the only two known cases are when $k=3,4$: an asymptotic result was given in~\cite{ABKS} for $k=3,4$; the exact result was proved by Pikhurko and Yilma~\cite{Oleg}, who showed that $T_{n,4}$ and $T_{n,9}$ maximize the number of $4$-edge-colorings with no monochromatic $K_3$ and $K_4$ respectively, see~\cite{Oleg-K} for more recent development.
Since the Tur\'an graph is extremal in the Erd\H{o} s-Rothschild problem for $r=2,3$, it is natural to consider its Ramsey-Tur\'an extension. Formally, given a function $f(n)$, we define $\RF(r,k,f(n))$ to be the maximum number of $r$-edge-colorings that an $n$-vertex graph with independence number at most $f(n)$ can have without a monochromatic copy of $K_k$. Similarly, the trivial lower bound on $\RF(r,k,o(n))$ is given by taking all edge-colorings on an extremal graph for Ramsey-Tur\'an problem:
\begin{eqnarray}\label{eq-ER-lower}
\RF(r,k,o(n))\ge r^{\RT(n,K_k,o(n))+o(n^2)}.
\end{eqnarray}
Considering~\eqref{eq-ER-orig-lower}, it is not inconceivable that the lower bound in~\eqref{eq-ER-lower} is optimal when $r$ is small. However, as shown in the following example, $\RF(r,k,o(n))$ exhibits rather different behavior than $F(n,r,k)$, even in the 2-edge-coloring case when $K_4$ is forbidden. Let $G$ be a graph obtained by putting a copy of $\Gamma$ in each partite set of $T_{n,2}$, where $\Gamma$ is a triangle-free graph with independence number $o(n)$.\footnote{The existence of $K_k$-free graph $\Gamma$ with $\alpha(\Gamma)=o(|\Gamma|)$ was proved by Erd\H{o} s~\cite{Erdos-girth}.} Since $\Gamma$ is triangle-free, the neighborhood of every vertex is an independent set. Therefore, the independence number of the graph $\Gamma$ is at least its maximum degree, which implies that $\Gamma$ has maximum degree $o(n)$. Consider the following set of $2$-edge-colorings of $G$. Color the edges inside one partite set red, the edges inside the other partite set blue, and color all the remaining cross-edges either red or blue. It is not hard to see that none of these colorings contain monochromatic $K_4$'s,
hence, $\RF(2,4,o(n))\ge 2^{n^2/4}$, while $\RT(n,K_4,o(n))=(1/8+o(1))n^2$.
The above example already suggests an obstacle in determining $\RF(2,k,o(n))$, that is, the subgraphs induced by each color could simultaneously have linear-sized independent set. Nonetheless, our first result reveals the asymptotic behavior of $\RF(2,k,o(n))$ for every integer $k$.
\begin{theorem}\label{thm-2coloringall}
$\RF(2,3,o(n))=2^{o(n^2)}$. For every integer $t\ge 1$ and $i\in\{1,2,3\}$,
\begin{eqnarray*}
\RF(2,3t+i,o(n))= 2^{\RT(n,K_{4t+i},o(n))+o(n^2)}.
\end{eqnarray*}
\end{theorem}
The following well-known theorem determines the asymptotic value of $\RT(n,K_k,o(n))$, for every $k\ge 3$. For odd $k$, this was proved by Erd\H os and S\'os~\cite{ES}. For $k=4$, the upper bound is due to Szemer\'edi~\cite{Sz-k4}. In~\cite{BE}, Bollob\'as and Erd\H os showed that this upper bound is asymptotically sharp. These results were extended by Erd\H os, Hajnal, S\'os, and Szemer\'edi~\cite{EHSSz} to every even $k$.
\begin{theorem}\label{thm-RT}
For every integer $k\ge 3$,
\begin{eqnarray*}
\RT(n,K_k,o(n))= (b_k+o(1))n^2,
\end{eqnarray*}
where
\begin{eqnarray}\label{eq-bk}
b_{k}=
\left\{
\begin{array}{ll}
\frac{1}{2}.\frac{k-3}{k-1}&\mbox{\quad\quad if } k\mbox{ is odd}, \\
\\
\frac{1}{2}.\frac{3k-10}{3k-4}&\mbox{\quad\quad if } k\mbox{ is even}.
\end{array}
\right.
\end{eqnarray}
\end{theorem}
The definition of $b_{k}$ comes from optimizing the number of edges in a construction that we will describe in Section~\ref{sec-pre} (Construction~\ref{const-oddeven}). By Theorem~\ref{thm-RT}, to prove Theorem~\ref{thm-2coloringall}, it suffices to prove the following theorem.
\begin{theorem}\label{thm-2coloringreduction}
$\RF(2,3,o(n))=2^{o(n^2)}$. For every integer $t\ge 1$ and $i\in\{1,2,3\}$,
\begin{eqnarray*}
\RF(2,3t+i,o(n))= 2^{(b_{4t+i}+o(1))n^2}.
\end{eqnarray*}
\end{theorem}
Note that since the value of the Ramsey-Tur\'an function is only known asymptotically, we will not try to determine the exact value of $\RF(2,k,o(n))$.
Our constructions for the lower bound in Theorem~\ref{thm-2coloringreduction} are based on the Bollob\'as-Erd\H os graph~\cite{BE}.
\subsection{A generalized Ramsey-Tur\'an problem}
The generalized Tur\'an-type problem, i.e.~for given graphs $F$ and $H$, determine $\ex(n,F,H)$, the maximum number of copies of $F$ in an $n$-vertex $H$-free graph, has been studied for various choices of $F$ and $H$. Erd\H os~\cite{Erd} determined $\ex(n,K_s,K_t)$ for all $t>s\ge 3$, showing that among all $K_t$-free graphs, $T_{n,t-1}$ has the maximum number of $K_s$'s. See also Bollob\'as and Gy\H ori~\cite{BG} for $\ex(n,K_3,C_5)$, and more recently, Alon and Shikhelman~\cite{AS} for the cases when $(F,H)$ are $(K_3,C_5)$, $(K_m,K_{s,t})$, and when both $F$ and $H$ are trees.
Our second result studies the general function $\RT(F,H,f(n))$, which is the maximum number of copies of $F$ in an $H$-free $n$-vertex graph $G$ with $\alpha(G)\le f(n)$. It is not hard to see that $\RT(K_s,K_{s+1},o(n))=o(n^s)$. We determine, in the following two theorems, $\RT(K_3,K_t, o(n))$ for every integer $t$.
\begin{theorem}\label{thm-k3}
Let $t\ge 6$ be an integer and $\ell=\left\lfloor\frac{t}{2}\right\rfloor$. Then as $n$ tends to infinity,
\begin{eqnarray*}
\RT(K_3,K_t, o(n))=a_{\ell}n^3(1+o(1)),
\end{eqnarray*}
where
\begin{eqnarray}\label{eq-al}
a_{\ell}=
\left\{
\begin{array}{ll}
\underset{0 \leq x \leq 1}\max {\ell-2\choose 3}\left(\frac{1-x}{\ell-2}\right)^3+
x{\ell-2\choose 2}\left(\frac{1-x}{\ell-2}\right)^2+
\frac{1}{2}\cdot\left(\frac{x}{2}\right)^2\left(1-x\right)
&\mbox{\quad\quad if } t=2\ell, \\
\\
\left(\frac{1}{\ell}\right)^3{\ell\choose 3} &\mbox{\quad\quad if } t=2\ell+1.
\end{array}
\right.
\end{eqnarray}
\end{theorem}
In fact, our proof shows that all the extremal graphs should have the structure as those in~Construction~\ref{const-oddeven}. The definition of $a_{\ell}$ comes from optimizing the number of $K_3$'s in these graphs.
For the general case $t>s\ge 3$, we present a construction in the concluding remark which we believe gives the right answer. Our next result verifies the first non-trivial case.
\begin{theorem}\label{thm-ks}
For every $s\ge 3$,
$$\RT(K_s,K_{s+2},o(n))=\left(2^{-{s\choose 2}}+o(1)\right)\left(\frac{n}{s}\right)^s.$$
\end{theorem}
\medskip
\noindent\textbf{Organization.} We first introduce some tools in Section~\ref{sec-pre}. Then in Section~\ref{sec-2coloring}, we prove Theorem~\ref{thm-2coloringreduction}, and in Section~\ref{sec-ks}, Theorems~\ref{thm-k3} and~\ref{thm-ks}.
\medskip
\noindent\textbf{Notation.} Let $G=(V,E)$ be an $n$-vertex graph and $e(G)=|E(G)|$. For every $v\in V$ and
$U,U'\subseteq V$, denote by $d_{U}(v)$ the degree of $v$ in $U$. Also, let $N_{U}(v)$ be the set of
vertices $u\in U$ such that $vu\in E(G)$. Denote by $G[U,U']$ the induced bipartite subgraph of $G$ on partite sets $U$ and $U'$. Let $k_s(G)$ be the number of $K_s$ in $G$. For every
$A\subseteq V(G)$ and an \mbox{$r$-coloring} of $E(G)$ with colors $\{c_1,\ldots, c_r\}$, let $G_{c_i}[A]$ be
the $c_i$-colored subgraph of $G$ induced by the vertex set $A$. We will write $G_{c_i}$ instead of
$G_{c_i}[V(G)]$. We fix throughout the paper a function $\omega(n)$ of $n$ such that $\omega(n)\rightarrow\infty$ arbitrary slowly. If we claim that a result holds whenever $0 < a \ll b \leq 1$, then this means that there is a non-decreasing
function $f : (0, 1] \rightarrow (0, 1]$ such that the result holds for all $0 < a, b\leq1 $ with $a \leq f(b)$.
\section{Preliminaries}\label{sec-pre}
We start with a formal definition for $\RT(n,H,o(n))$.
\begin{definition}
For a graph $H$ and a function $f(n)$, let
\begin{eqnarray*}
\RT(n,H,o(f(n)))=n^2\cdot
\lim\limits_{\varepsilon\rightarrow 0}\lim\limits_{n\rightarrow\infty}\frac{\RT(n,H,\varepsilon f(n))}{n^2}+o(n^2).
\end{eqnarray*}
\end{definition}
Bollob\'as and Erd\H{o} s~\cite{BE} constructed a family of $n$-vertex $K_4$-free graphs with independence number $o(n)$ and $(\frac{1}{8}+o(1))n^2$ edges. We follow the description in~\cite{SS} to present their construction. For a constant $\varepsilon>0$, and sufficiently large integers $d$ and $n_0$, assume $n>n_0$ is even and $\mu=\varepsilon/\sqrt d$. Next, partition the high-dimensional unit sphere $\ensuremath{\mathbb{S}}^d$ into $n/2$ domains, $D_1,\ldots,D_{n/2}$, of equal measure with diameter\footnote{The diameter is the maximum distance between any two points in each domain.} less than $\mu/2$. For every $1\le i\le n/2$, choose two points $x_i, y_i\in D_i$. Let $X=\{x_1,\ldots,x_{n/2}\}$ and $Y=\{y_1,\ldots,y_{n/2}\}$. Let $\BE(X,Y)$ be the graph with vertex set $X\cup Y$ and edge set as follows. For every $x,x'\in X$ and $y,y'\in Y$,
\begin{enumerate}
\item let $xy\in E(\BE(X,Y))$ if their distance is less than $\sqrt 2-\mu$,
\item let $xx'\in E(\BE(X,Y))$ if their distance is more than $2-\mu$,
\item let $yy'\in E(\BE(X,Y))$ if their distance is more than $2-\mu$.
\end{enumerate}
Note that the number of edges with both ends in $X$ or $Y$ is $o(n^2)$.
Next, for every integer $k\ge 3$, we will describe a family of $n$-vertex $K_k$-free graphs with independence number $o(n)$. As we mentioned earlier, the constant $b_k$ defined in~\eqref{eq-bk} comes from maximizing the number of edges in the construction below. In other words, some of these graphs are extremal graphs for Theorem~\ref{thm-RT}, i.e.~they have $(b_k+o(1))n^2$ edges.
\begin{construction}\label{const-oddeven}
Given $k\ge 3$, denote by $\mathcal{H}(n,k)$ the family of $n$-vertex graphs $G$ obtained as
follows. Let $\ell=\floor{\frac{k}{2}}$, and $\Gamma_n$ be an $n$-vertex triangle-free graph with
$\alpha(\Gamma_n)=o(|\Gamma_n|)$. If $k$ is odd, start
with a complete balanced $\ell$-partite graph on vertex set $V_1\cup\ldots\cup
V_{\ell}$. Then put a copy of $\Gamma_{|V_i|}$ in each $V_i$. If $k$ is even,
partition the vertex set into $\ell$ parts $\{V_1,\ldots,V_\ell\}$, such that $|V_1|=|V_2|$. First, let $G[V_1\cup V_2]$ be a copy of the Bollob\'as-Erd\H{o} s graph $\BE(V_1,V_2)$; then for every $i\in \{1,\ldots,\ell\}$ and $j\in\{3,\ldots,\ell\}\setminus\{i\}$, let $G[V_i,V_j]$ be a complete bipartite graph; next, for every $i\in\{3\ldots\ell\}$, put a copy of
$\Gamma_{|V_i|}$ in each $V_i$.
\end{construction}
\begin{rmk}
Note that in Construction~\ref{const-oddeven}, for even $k$, $|V_1|=|V_2|$ and $\{V_1,\ldots,V_\ell\}$ is not necessarily an equipartition.
\end{rmk}
We will also need the following definitions of regular partitions and weighted cluster graph.
\begin{definition}
Let $G$ be a graph and $A,B\subseteq V(G)$. Denote by $d(A,B):=\frac{e(G[A,B])}{|A||B|}$ the {\it density} of the pair $(A,B)$. Given $\varepsilon>0$, a pair $X,Y\subseteq V(G)$ is \emph{$\varepsilon$-regular} if for every $A\subseteq X$ and $B\subseteq Y$ with $|A|\ge \varepsilon |X|$ and $|B|\ge \varepsilon |Y|$, $|d(A,B)-d(X,Y)|\le \varepsilon$. A vertex partition of $G$, $V(G)=C_1\cup \ldots\cup C_m$ is $\varepsilon$-regular if all but $\varepsilon m^2$ pairs of $(C_i,C_j)$ are $\varepsilon$-regular.
\end{definition}
\begin{definition}
For every $\varepsilon>0$, positive integer $t$, and an $n$-vertex graph $G=(V,E)$, let $\mathcal{C}=\{C_1, \ldots, C_m\}$ be an $\varepsilon$-regular partition of $V(G)$ with $m\ge t$. Denote by $R$ the \emph{cluster graph} (with respect to $\varepsilon$) with vertex set $\mathcal{C}$, and $C_i$ and $C_j$ are adjacent if the pair $(C_i,C_j)$ is $\varepsilon$-regular with density at least $10\varepsilon$. We now define the \emph{weighted cluster graph}, $R=(\mathcal{C},w)$ (with respect to $\varepsilon$), on the vertex set $\mathcal{C}$ as follows. For an $\varepsilon$-regular pair $(C_i,C_j)$, we will define:
\begin{eqnarray*}
w(C_i,C_j):=\left\{
\begin{array}{ll}
0\mbox{\ \ \ \ if }d(C_i,C_j)\le 10\varepsilon \mbox{ or }(C_i,C_j)\mbox{ is an irregular pair},\\
\frac{1}{2}\mbox{\ \ \ \ if }10\varepsilon<d(C_i,C_j)\le 1/2+10\varepsilon, \\
1\mbox{\ \ \ \ if }d(C_i,C_j)>1/2+10\varepsilon.
\end{array}
\right.
\end{eqnarray*}
\end{definition}
\begin{definition}\label{def-weight}
A \emph{weighted graph} $G$ is an ordered triple $(V,E,w)$ where $E:={V\choose 2}$, set of all unordered pairs of vertices, and $w:E\rightarrow
\{0,1/2,1\}$. Define $G_{1/2}=(V,E_{1/2})$ where $E_{1/2}=\{e\in E: w(e)\ge 1/2\}$ and $G_{1}=(V,E_{1})$
where $E_{1}=\{e\in E: w(e)=1\}$. Denote by $e(G)=\sum_{e\in E(G)}w(e)$.
For two weighted graphs $G=(V,E,w)$ and $G'=(V,E,w')$, define $G\cap G'=(V,E,w'')$ where $w''(e)=\min\{w(e),w'(e)\}$. For $X\subseteq Y\subseteq V$, we call $(X,Y)$ a \emph{weighted $(|X|,|Y|)$-clique} or \emph{weighted complete subgraph} of size $\ell$ if ${X\choose 2}\subseteq E_1$ and ${Y\choose 2}\subseteq E_{1/2}$ and $|X|+|Y|=\ell$. Also, let the \emph{weighted clique number} of $G$ be the size of the largest weighted complete subgraph of $G$.
For a triangle $T=e_1e_2e_3$, let $w(T)=\prod_{i=1}^{3} w(e_i)$. Also, let
$T(G)=\sum_{T\in\mathcal{T}} w(T)$, where $\mathcal{T}$ is the set of all triangles in $G$. For $X\subseteq V(G)$, denote $T(X)=\sum_{T\in G[X]}w(T)$.
\end{definition}
We need the following two lemmas and theorem from~\cite{EHSSz}, the first one has been proved in the proof of Theorem~2 in~\cite{EHSSz}.
\begin{lemma}\label{lem-completefree2}
For every $\varepsilon>0$, there exist $\delta>0$ and $n_0$ such that for every $n$-vertex graph $G$ with $n\ge n_0$, if its weighted cluster graph $R(\mathcal{C},w)$ with respect to $\varepsilon$ contains a weighted clique $(X,Y)$ of size $\ell$ such that $\alpha(G[X])=\delta n$, then $G$ contains a copy of $K_\ell$.
\end{lemma}
\begin{lemma}\label{lem-weightedclique}
For every $\varepsilon>0$ and integer $k\ge 3$ there exists $n_0$ such that for every $n$-vertex weighted graph $G=(V,E,w)$ with $n\ge n_0$, if $G$ does not contain a weighted complete subgraph of size $k$, then
\begin{eqnarray*}
e(G)\le (b_{k}+\varepsilon)n^2,
\end{eqnarray*}
where $b_{k}$ is defined in~\eqref{eq-bk}.
\end{lemma}
We will use the following multicolored version of the Szemer\'edi regularity lemma (for example, see~\cite{KS}).
\begin{theorem}\label{thm-color}
For every $\varepsilon>0$ and integer $r$, there exists an $M$ such that for every $n>M$ and every $r$-coloring of the edges of an $n$-vertex graph $G$ with colors $\{c_1,\ldots,c_r\}$, there exists a partition of $V(G)$ into sets $V_1,\ldots, V_m$ with $||V_i|-|V_j||\le 1$, for some $1/\varepsilon<m<M$, which is $\varepsilon$-regular with respect to $G_{c_i}$ for every $1\le i\le r$.
\end{theorem}
\section{Proof of Theorem~\ref{thm-2coloringreduction}}\label{sec-2coloring}
To overcome the obstacle that all subgraphs induced by each color could have linear-sized independent sets, we need the following simple, but somewhat surprising observation.
\begin{lemma}\label{lem-partition}
For every $0<c<1$, $r\ge 2$, and $a\le a_r(c):=c^{3\cdot 2^{r-2}-1}$ the following holds. Let $G$ be an $n$-vertex graph with $\alpha(G)\le an$ and an $r$-edge-coloring $C:E(G)\rightarrow \{c_1,\ldots,c_r\}$. Then there exists a partition $V(G)=C_1\cup\ldots\cup C_r$ such that $\alpha(G_{c_i}[C_i])\le cn$ for every $1\le i\le r$.
\end{lemma}
\begin{proof}
We fix a $c>0$, and use induction on the number of colors $r$. For the base case when $r=2$ and $a\le c^2$, if $\alpha(G_{c_i})\le cn$, for some $i\in\{1,2\}$, then we can partition $V(G)$ into $C_i=V(G)$ and $C_j=\emptyset$, where $j=\{1,2\}\setminus\{i\}$, finishing the proof. Therefore, we may assume $\alpha(G_{c_1})> cn$ and $\alpha(G_{c_2})> cn$. Let $X_0=V(G)$, $Y_0=\emptyset$. We iterate the following operation for $i\ge 1$. At step $i$, if $\alpha(G_{{c_1}}[X_{i-1}])\le cn$ then we will stop. Otherwise, let $I$ be a maximum independent set in $G_{c_1}[X_{i-1}]$. Since $\alpha(G)\le an$, we have $\alpha(G_{c_2}[I])\le an$. We define $X_i:=X_{i-1}\setminus I$ and $Y_i:=Y_{i-1}\cup I$. Notice that $\alpha(G_{c_2}[Y_i])\le \alpha(G_{c_2}[Y_{i-1}])+an$. Suppose the iteration stops after $k$ steps, i.e.~$\alpha(G_{c_1}[X_k])\le cn$, then $k\le \frac{n}{cn}=1/c$, which implies that $\alpha(G_{c_2}[Y_k])\le k\cdot an\le cn$ as desired.
For the inductive step, let us assume that the lemma holds for $r-1$ colors, where $r\ge 3$. In particular, we assume that for every $a>0$, $n'$-vertex graph $H$ with $\alpha(H)\le an'$, and $(r-1)$-edge-coloring of $H$ with colors $c'_1,\ldots,c'_{r-1}$, there exists a partition of $V(H)=C'_1\cup\ldots\cup C'_{r-1}$ such that $\alpha(H_{c'_i}[C'_i])\le a^{1/(3\cdot 2^{r-3}-1)}n'$ for all $1\le i\le r-1$.
Now, we will prove the lemma for $r$ colors. Fix an arbitrary $r$-edge-coloring of $G$,
we can assume that $\alpha(G_{c_1})> cn$, otherwise $C_1=V(G)$ and $C_2=\ldots=C_r=\emptyset$. Let $C_{1,0}=V(G)$ and $C_{i,0}=\emptyset$ for all $2\le i\le r$.
We iterate the following operation. At step $k$, if $\alpha(G_{c_1}[C_{1,k-1}])\le cn$ then we will stop.
Otherwise, let $I$ be a maximum independent set of $G_{c_1}[C_{1,k-1}]$ with $n'> cn$ vertices. Since $\alpha(G)\le a\cdot n$, for some constant $a\le a_r(c)$, we have
$\alpha(G_{c_2}[I]\cup\ldots\cup G_{c_r}[I])\le an<an'/c$. We can apply the induction hypothesis to the graph $G[I]$. Therefore, there exists a partition of $I=I_2\cup\ldots\cup I_r$ such that for every $2\le i\le r$, we have
\begin{eqnarray}\label{eq-indnum}
\alpha(G_{c_i}[I_i])\le\left(\frac{a}{c}\right)^{\frac{1}{3\cdot 2^{r-3}-1}}n'\le c^{\frac{3\cdot 2^{r-2}-2}{3\cdot 2^{r-3}-1}}n'\le c^{\frac{3\cdot 2^{r-2}-2}{3\cdot 2^{r-3}-1}}n.
\end{eqnarray}
Then, we define $C_{1,k}:=C_{1,k-1}\setminus I$ and $C_{i, k}:=C_{i,k-1}\cup I_i$, for $2\le i\le r$. By~\eqref{eq-indnum},
\begin{eqnarray*}
\alpha(G_{c_i}[C_{i,k}])\le \alpha(G_{c_i}[C_{i,k-1}])+c^{\frac{3\cdot 2^{r-2}-2}{3\cdot 2^{r-3}-1}}n.
\end{eqnarray*}
Let us assume that the iteration stops after $l$ steps, i.e.~$\alpha(G_{c_1}[C_{1,l}])\le cn$. Note that $l\le \frac{n}{cn}=1/c$, which implies that for every $2\le i\le r$,
\begin{eqnarray*}
\alpha(G_{c_i}[C_{i,l}])\le l\cdot c^{\frac{3\cdot 2^{r-2}-2}{3\cdot 2^{r-3}-1}}n\le\frac{1}{c}\cdot c^{\frac{3\cdot 2^{r-2}-2}{3\cdot 2^{r-3}-1}}n=c^{\frac{3\cdot 2^{r-2}-2}{3\cdot 2^{r-3}-1}-1}n=cn.
\end{eqnarray*}
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm-2coloringreduction}]
(Lower bound) Fix an arbitrary $t\ge 1$. For each $i=1,2,3$, by Theorem~\ref{thm-RT}, there is a $\floor{\frac{4t+i}{2}}$-partite graph $G_i$, with partite sets $\{V_1,\ldots,V_{\lfloor(4t+i)/2\rfloor}\}$, in $\mathcal{H}(n,4t+i)$ from Construction~\ref{const-oddeven} such that $e(G_i)=\RT(n,4t+i,o(n))=(b_{4t+i}+o(1))n^2$. We take the following set of colorings.
\begin{itemize}
\item $i=1$: Set $V(G_1)=V_1\cup \ldots\cup V_{2t}$. Color $G_1[V_p]$, for $1\le p\le t$, red, color $G_1[V_q]$, for $t+1\le q\le 2t$, blue, and all cross-edges in $G_1[V_p,V_q]$, $1\le p<q\le 2t$, in either red or blue.
\item $i=2$: Set $V(G_2)=V_1\cup \ldots\cup V_{2t+1}$. Color $G_2[V_p]$, for $1\le p\le t+1$, red, color $G_2[V_q]$, for $t+2\le q\le 2t+1$, blue, and all cross-edges in $G_2[V_p,V_q]$, $1\le p<q\le 2t+1$, in either red or blue.
\item $i=3$: Set $V(G_3)=V_1\cup \ldots\cup V_{2t+1}$. Color $G_3[V_p]$, for $1\le p\le t$, red, color $G_3[V_q]$, for $t+1\le q\le 2t+1$, blue, and all cross-edges in $G_3[V_p,V_q]$, $1\le p<q\le 2t+1$, in either red or blue.
\end{itemize}
In all three cases, the total number of edges inside all $V_i$'s is $o(n^2)$. Therefore, the total number of cross-edges is $RT(n,4t+i,o(n))-o(n^2)$, which implies that we obtain $2^{\RT(n,4t+i,o(n))-o(n^2)}$ $2$-edge-colorings. Hence, we are left to show that all these colorings are monochromatic $K_{3t+i}$-free. For $i=1,3$, note that every blue (red~resp.) clique have at most one vertex from each $V_p$ ($V_q$~resp.), and at most two vertices from each $V_q$ ($V_p$~resp.). Hence, the size of the largest blue (red~resp.) clique is at most $t+2\cdot(t+\floor{\frac{i}{2}})<3t+i$ ($2t+t+\floor{\frac{i}{2}}<3t+i$~resp.). For the case when $i=2$, fix arbitrary $p,q$ such that $1\le p\le t+1$ and $t+2\le q\le 2t+1$. Note that to get a blue clique, we can have at most 1 vertex from each $V_p$ and 2 vertices from each $V_q$, hence, the largest blue clique has size at most $1\cdot(t+1)+2\cdot t=3t+1$. For a red clique, we can have at most 1 vertex from each $V_q$, at most a $K_3$ from $V_1\cup V_2$, and at most 2 vertices from each $V_p$, $p\neq 1,2$. Thus, the largest red clique is of size at most $3+2\cdot (t-1)+1\cdot t=3t+1$.
\medskip
(Upper bound) We will prove that for a given constant $\varepsilon>0$, there exists $\gamma>0$ such that, for sufficiently large $n$, the following holds. Let $G$ be an $n$-vertex graph with $\alpha(G)\le \gamma^2 n$, the number of $2$-edge-colorings of $G$ without a monochromatic $K_{3t+i}$ is at most $2^{(b_{4t+i}+\varepsilon)n^2}$, for $t\ge 1$ and $i=1,2,3$. Also, the number of $2$-edge-colorings with no monochromatic $K_3$ is at most $2^{\varepsilon n^2}$. Throughout the proof, constants are chosen from right to left according to the following hierarchy:
\begin{eqnarray}\label{eq-hierarchy}
0<\gamma\ll\varepsilon_2\ll\frac{1}{n_1}\ll\varepsilon_1\ll\varepsilon<1.
\end{eqnarray}
Let $n_0$ be the constant returned from Lemma~\ref{lem-weightedclique} with $\varepsilon_1$ playing the role of $\varepsilon$ and choose $n_1\ge n_0$. Let $\delta$ be the constant returned from Lemma~\ref{lem-completefree2} with $\varepsilon_2$ playing the role of $\varepsilon$ and choose $\gamma<\delta$.
For any fixed $2$-edge-coloring of $G$, $\phi: E(G)\rightarrow\{\phi_1,\phi_2\}$, apply Lemma~\ref{lem-partition} with $r=2$, $c=\gamma$, and $a_2(c)=\gamma^2$. Let $\{A,B\}$ be the resulting partition such that
\begin{eqnarray}\label{eq-part}
\alpha(G_{\phi_1}[A])\le \gamma n, \quad\mbox{ and }\quad \alpha(G_{\phi_2}[B])\le \gamma n.
\end{eqnarray}
We then apply Theorem~\ref{thm-color}, with $\varepsilon_2$ playing the role of $\varepsilon$, to $G$ with coloring $\phi$, and let $\mathcal{P}=\{P_1,\ldots,P_m\}$ be the resulting partition of $V(G)$, where $m\ge 1/\varepsilon_2$. Note that we may assume the regularity partition $\mathcal{P}$ refines the $\{A,B\}$-partition. Let $R_{\phi_1}$ and $R_{\phi_2}$ be the $\phi_1$-colored and $\phi_2$-colored weighted cluster graphs respectively, both on vertex set $\{p_1,\ldots,p_m\}$, where the vertex $p_i$ represents the vertex set $P_i$, for all $i\in [m]$. We have
\begin{eqnarray}\label{eq-partitioning}
\mbox{number of ways to fix an }\{A,B\}\mbox{-partition of }V(G) &\le&2^n,\nonumber\\
\mbox{number of ways to fix a }\mathcal{P}\mbox{-partition of }V(G) &\le&m^n,\nonumber\\
\mbox{number of ways to fix }R_{\phi_1}\mbox{ and } R_{\phi_2} &\le&\left(2^{m\choose 2}\right)^4.
\end{eqnarray}
Now, we will count the number of colorings with a fixed $\{A,B\}$-partition, $\mathcal{P}$-partition and weighted cluster graphs $R_{\phi_1}$ and $R_{\phi_2}$. First, note that the number of edges of the graph $G$ with both ends in one of the $P_i$'s, between irregular or sparse pairs is at most
\begin{eqnarray}\label{eq-removededges}
m\cdot\left(\frac{n}{m}\right)^2+\varepsilon_2\cdot m^2\cdot\left(\frac{n}{m}\right)^2+10\cdot\varepsilon_2\cdot m^2\cdot\left(\frac{n}{m}\right)^2\le \varepsilon_2n^2+\varepsilon_2n^2+10\cdot\varepsilon_2n^2.
\end{eqnarray}
Hence, the number of ways to color these edges is at most $2^{12\varepsilon_2n^2}$. From now on, we will only consider the rest of the edges of $G$, i.e.~the edges between pairs of clusters that are adjacent in $R_{\phi_1}\cup R_{\phi_2}$. Note that there is a unique way to color edges in $R_{\phi_1}\Delta R_{\phi_2}$. Thus the number of $2$-edge-colorings corresponding to the fixed $\{A,B\}$-partition, $\mathcal{P}$-partition and weighted cluster graphs $R_{\phi_1}$ and $R_{\phi_2}$ is at most
\begin{eqnarray}\label{eq-clusteredges}
2^{\left(\frac{n}{m}\right)^2e(R_{\phi_1}\cap R_{\phi_2})+12\varepsilon_2n^2}.
\end{eqnarray}
To complete the proof, it remains to show that
\begin{itemize}[noitemsep,topsep=0pt]
\item[(i)]\label{itm-3} when $\phi$ is monochromatic $K_3$-free, $e(R_{\phi_1}\cap R_{\phi_2})=0$, and
\item[(ii)]\label{itm-3t+i} when $\phi$ is monochromatic $K_{3t+i}$-free for $t\ge 1$ and $i=1,2,3$,
\end{itemize}
\begin{eqnarray}\label{eq-mainedges}
e(R_{\phi_1}\cap R_{\phi_2})\le (b_{4t+i}+\varepsilon_1)\cdot m^2,
\end{eqnarray}
where $b_{4t+i}$ is defined in~\eqref{eq-bk}. Indeed, since the choice of $G$ is arbitrary, (i) together with~\eqref{eq-partitioning} and~\eqref{eq-clusteredges}, implies
\begin{eqnarray*}
\RF(2,3,\gamma^2n)&\le& 2^n\cdot m^n\cdot 2^{4\cdot{m\choose 2}}\cdot 2^{12\varepsilon_2n^2}\le 2^{\varepsilon n^2},
\end{eqnarray*}
and~(ii), together with~\eqref{eq-partitioning} and~\eqref{eq-clusteredges}, implies
\begin{eqnarray*}
\RF(2,3t+i,\gamma^2n)&\le& 2^n\cdot m^n\cdot 2^{4\cdot{m\choose 2}}\cdot 2^{(b_{4t+i}+\varepsilon_1) n^2+12\varepsilon_2n^2}\le 2^{(b_{4t+i}+\varepsilon) n^2}.
\end{eqnarray*}
To see (i), notice that if there is an edge, say $uv\in E(R_{\phi_1}\cap R_{\phi_2})$, then, without loss of generality, we may assume that $u\in A$. Therefore, by setting $X=\{u\}$ and $Y=\{u,v\}$, it follows from~\eqref{eq-part} that we have a $\phi_1$-colored weighted $(1,2)$-clique $(X,Y)$ with $\alpha(G_{\phi_1}[X])\le \gamma n$, which by Lemma~\ref{lem-completefree2} yields a monochromatic $K_3$ in $\phi$, a contradiction.
For (ii), suppose that~\eqref{eq-mainedges} is not satisfied. Since $m>1/\varepsilon_2>n_1$, we can apply Lemma~\ref{lem-weightedclique} to the graph $R_{\phi_1}\cap R_{\phi_2}$, with $\varepsilon_1$ playing the role of $\varepsilon$. Hence, the graph $R_{\phi_1}\cap R_{\phi_2}$ has a weighted complete subgraph $(X,Y)$ of size $4t+i$, and we shall find a monochromatic $K_{3t+i}$ in $G$ using $(X,Y)$, which is a contradiction. Let $x=|X|$, $y=|Y|$ and $X=\{p_1\cup\ldots\cup p_{x}\}$. Without loss of generality, we may assume that $p_1\cup\ldots\cup p_{\ceiling{x/2}}:=X'\subseteq A$, i.e.~$\alpha(G_{\phi_1}[P_1\cup\ldots\cup P_{\ceiling{x/2}}])\le\gamma n$. We have thus found a weighted clique $(X',Y)$ in $G_{\phi_1}$ such that $\alpha(G_{\phi_1}[X'])\le\gamma n$. Hence, Lemma~\ref{lem-completefree2} shows that $G_{\phi_1}$ contains a copy of $K_{{\ceiling{x/2}}+y}$.
\begin{claim}\label{cl-x-small}
$\lfloor \frac{x}{2}\rfloor\le t$.
\end{claim}
\begin{proof}
Suppose that $\lfloor \frac{x}{2}\rfloor\ge t+1$. Recall from the definition of weighted clique that $X\subseteq Y$, i.e.~$x\le y$, and $x+y=4t+i$. Thus,
$$4t+3\ge 4t+i=x+y\ge 2x\ge 4\left\lfloor \frac{x}{2}\right\rfloor\ge 4(t+1),$$
a contradiction.
\end{proof}
Claim~\ref{cl-x-small} then implies that the monochromatic clique corresponding to $(X',Y)$ we found in $G_{\phi_1}$ is of order
\begin{eqnarray}\label{eq-cliquesize}
{\ceiling{\frac{x}{2}}}+y=x+y-\floor{\frac{x}{2}}\ge 4t+i-t=3t+i,
\end{eqnarray}
a contradiction.
\end{proof}
\section{Proof of Theorems~\ref{thm-k3} and~\ref{thm-ks}}\label{sec-ks}
\begin{proof}[Proof of Theorem~\ref{thm-ks}](Lower bound) Let $D_1,\ldots,D_{n/s}$ be a partition of the high-dimensional unit sphere of equal measure with small diameter as in the Bollob\'as-Erd\H os graph construction. Let $G$ be an $n$-vertex graph with a balanced vertex partition $V_1$, $\ldots$, $V_s$, where each $V_i$ consists of one point from each of the $n/s$ domains $D_1,\ldots,D_{n/s}$. For every pair of distinct integers $i,j\in [s]$, let $G[V_i\cup V_j]$ be a copy of $\BE(V_i,V_j)$. Note that each $G[V_i]$ is triangle-free and each $G[V_i\cup V_j]$ is $K_4$-free. We claim that $G$ is $K_{s+2}$-free. Indeed, let $F$ be a largest clique in $G$ and let $g_i=|V(F)\cap V_i|$. Since $G[V_i]$ is triangle-free, each $g_i\le 2$. If $|V(F)|\ge s+2$, then there exists at least two indices $p,q$ such that $g_p=g_q=2$, which contradicts to $G[V_p,V_q]$ being $K_4$-free. We will count the number of $K_s$ with exactly one vertex from each $V_i$. Fix a vertex $v_1\in V_1$, a uniformly at random chosen $v_2\in V_2$ is adjacent to $v_1$ if $v_2$ is in the cap (almost a hemisphere) centered at $v_1$ with measure $1/2-o(1)$, which happens with probability $1/2-o(1)$. Now we fix a clique on vertex set $\{v_1,\ldots, v_{\ell-1}\}$ with $\ell\ge 2$ and $v_i\in V_i$. The number of vertices in $V_{\ell}$ that are in $\bigcap_{i=1}^{\ell-1}N(v_i)$ is at least $2^{-(\ell-1)}n/s-o(n)$. Therefore, we have
\begin{eqnarray*}
k_s(G)\ge\prod_{i=1}^{s}\left[\left(2^{-(i-1)}-o(1)\right)\frac{n}{s}\right]=\left(2^{-{s\choose 2}}-o(1)\right)\left(\frac{n}{s}\right)^s.
\end{eqnarray*}
(Upper bound) We will prove that for a given $\varepsilon>0$ and integer $s\ge 3$, there exists $\gamma>0$ such that for any sufficiently large $n$ the following holds. Let $G$ be an $n$-vertex $K_{s+2}$-free graph with $\alpha(G)\le \gamma n$, the number of edges of copies of $K_s$ in $G$ is at most $(2^{-{s \choose 2}}+\varepsilon)(n/s)^s$. Throughout the proof, constants are chosen from right to left according to the following hierarchy,
\begin{eqnarray*}
0<\gamma\ll\varepsilon_1\ll\varepsilon<1.
\end{eqnarray*}
Let $\varepsilon_1$ play the role of $\varepsilon$ in Lemma~\ref{lem-completefree2}, and choose $\gamma$ such that it is smaller than the resulting $\delta$. First, we apply Theorem~\ref{thm-color}, with $\varepsilon_1$ playing the role of $\varepsilon$, to the graph $G$ and let $\mathcal{P}=\{P_1,\ldots,P_m\}$ be the resulting partition of $V(G)$, where $m\ge 1/\varepsilon_1$, and let $R$ be the weighted cluster graph with respect to $\varepsilon_1$. We call an edge in $R$ \emph{heavy} if it has weight $1$. We claim that the graph $R$ does not contain any weighted $(1,s+1)$- or $(2,s)$-clique. Otherwise, we apply Lemma~\ref{lem-completefree2} to the graph $R$ with $\varepsilon_1$ playing the role of $\varepsilon$. Since $\alpha(G[X])\le\alpha(G)\le \gamma(n)\le \delta_1(n)$, $G$ contains a copy of $K_{s+2}$, a contradiction. In other words, we have that $R$ is $K_{s+1}$-free and does not have a copy of $K_{s}$ with at least one heavy edge.
Now, we can count the total number of copies of $K_s$ in $G$. Note that similarly to~\eqref{eq-removededges}, the total number of edges inside all clusters, between irregular pairs, or sparse pairs is at most $12\cdot\varepsilon_1n^2$. Therefore, the total number of copies of $K_s$ with at least one such edge is at most $12\cdot\varepsilon_1n^2\cdot n^{s-2}$. Since $R$ is $K_{s+1}$-free, by the result of Erd\H os~\cite{Erd}, it has at most $(m/s)^s$ copies of $K_s$. Also, since $R$ does not have a copy of $K_s$ with a heavy edge, it implies that each $K_s$ in $R$ has weight at most $(1/2+10\cdot\varepsilon_1)^{s\choose 2}\le 2^{-{s\choose 2}}+\varepsilon/2$, where the last inequality holds because $\varepsilon_1$ is sufficiently small with respect to $\varepsilon$. Hence, we have that the number of $K_s$ in $G$ is at most
\begin{eqnarray*}
\left(2^{-{s\choose 2}}+\frac{\varepsilon}{2}\right)\cdot\left(\frac{n}{m}\right)^s\cdot\left(\frac{m}{s}\right)^s+12\cdot\varepsilon_1n^s\le
\left(2^{-{s\choose 2}}+\varepsilon\right)\left(\frac{n}{s}\right)^s.
\end{eqnarray*}
\end{proof}
The following lemma is the main step for proving Theorem~\ref{thm-k3}.
\begin{lemma}\label{lem-triangle}
For every integer $t\ge 4$ and $n$-vertex weighted graph $G=(V,E,w)$ (as in Definition~\ref{def-weight}) with no weighted complete subgraph of size $t$, we have
\begin{eqnarray*}
T(G)\le a_{t}n^3,
\end{eqnarray*}
where $a_t$ is as in~\eqref{eq-al}.
\end{lemma}
\begin{proof}
Let $G=(V,E,w)$ be an $n$-vertex weighted graph that satisfies the hypothesis and is extremal, i.e.~has the maximum number of triangles. First, we will apply two rounds of the so-called symmetrization method to the graph
$G$. For $v,v'\in V(G)$, denote $T_v(G)=\sum_{v\in T\in G}w(T)$, the number of weighted triangles containing $v$. Similarly, define $T_{vv'}(G)=\sum_{v,v'\in T\in G}w(T)$. Let $V(G)=\{v_1,\ldots,v_n\}$ such that $T_{v_1}(G)\ge\ldots\ge T_{v_n}(G)$. Define $S_1(i,j)$, for $i\neq j\in [n]$ to be the following operation: if $v_iv_j\notin G_{1/2}$ then we replace $v_j$
with a copy of $v_i$, i.e.~change $w(v_jv_k)$ to $w(v_i,v_k)$ for all $k\neq i,j$. If $i<j$, then the number of triangles changes by $T_{v_i}(G)-T_{v_j}(G)\ge 0$. Since $G$ is extremal, we have that $T_{v_i}(G)=T_{v_j}(G)$ for any $v_iv_j\notin G_{1/2}$. Consequently, the following process, denoted by $S_1$, is finite: apply $S_1(i,j)$ for every $1\le i<j\le n$ with $v_iv_j\notin G_{1/2}$. Note that $S_1$ will not increase the weighted clique number and keep the same number of triangles. After $S_1$, in the resulting graph
$v_iv_j\notin G_{1/2}$ is an equivalence relation. Denote by $\mathcal{A}=\{A_1,\ldots, A_m\}$ the equivalence
classes of this relation, i.e.~two vertices $u$ and $v$ are in the same class if and only if $uv\notin
G_{1/2}$. Therefore, for fixed $1\le i,j\le m$, all the edges between $A_i$ and $A_j$ have equal
weights, which we denote by $w(A_iA_j)$, and for all vertices $x,x'\in A_i$ and $y,y'\in A_j$, we have
\begin{eqnarray*}
T_x(G)=T_{x'}(G)\quad\mbox{and}\quad T_{xy}(G)=T_{x'y'}(G).
\end{eqnarray*}
Therefore, we can define $T_{A_i}(G)=T_x(G)$ and $T_{A_iA_j}(G)=T_{xy}(G)$.
Note that if $(X,Y)$ is one of the largest weighted complete
subgraphs of $G$, then $|Y|=m$.
We summarize the structure of $G$ as follows: Let $H$ be a weighted complete graph on vertex set $\{a_1,\ldots,a_m\}$ with all its edges having weight either $1$ or $1/2$, and $w(a_ia_j)=w(A_iA_j)$. The graph $G$ is a blow-up of $H$ where we replace each $a_i$ with a set of $|A_i|$ vertices, and inside each $A_i$ the weight of all edges is zero.
Our next goal is to show that a second round of symmetrization can be carried out in $G$, in other words, in $H$, $w(a_ia_j)=1/2$ is an equivalence relation. Without loss of generality we may assume $T_{A_1}(G)\ge\ldots\ge T_{A_m}(G)$. For every $1\le i,j\le m$, define $S_2(i,j)$ to be the following operation: Change $w(A_jA_k)$ to $w(A_iA_k)$ for all $k\neq i,j$, and denote $G_{A_i}$ the resulting graph. Define $G_{A_j}$ analogously as the graph obtained from applying $S_2(j,i)$ to $G$. The following claim states that when $w(A_iA_j)=1/2$, we can replace vertices in $A_i$ with copies of vertices in $A_j$, or the other way around, without decreasing the number of triangles.
\begin{claim}\label{cl-2-symm}
For every pair of integers $1\le i<j\le m$ with $w(A_iA_j)=1/2$,
(i) $T_{A_i}(G)=T_{A_j}(G)$;
(ii) $k_3(G_{A_i})= k_3(G_{A_j})= k_3(G)$.
\end{claim}
\begin{proof}
Define
\begin{eqnarray*}
T^o_{A_i}(G)=T_{A_i}(G)-T_{A_iA_j}(G)\quad\mbox{ and }\quad G'=G\setminus \{A_i\cup A_j\}.
\end{eqnarray*}
Since $T_{A_i}(G)\ge T_{A_j}(G)$, we have
\begin{eqnarray}\label{eq-To}
T^o_{A_i}(G)+T_{A_iA_j}(G)\ge T^o_{A_j}(G)+T_{A_iA_j}(G)\quad \Leftrightarrow\quad T^o_{A_i}(G)\ge T^o_{A_j}(G).
\end{eqnarray}
\noindent For (i), it suffices to show
\begin{eqnarray}\label{eq-o}
T^o_{A_i}(G)=T^o_{A_j}(G).
\end{eqnarray}
Note that
\begin{eqnarray}
k_3(G)&=&k_3(G')+|A_i|\cdot T^o_{A_i}(G)+|A_j|\cdot T^o_{A_j}(G)+|A_i|\cdot |A_j|\cdot T_{A_iA_j}(G),\label{eq-G}\\
k_3(G_{A_i})&=& k_3(G')+(|A_i|+|A_j|)\cdot T^o_{A_i}(G)+|A_i|\cdot |A_j|\cdot T_{A_iA_j}(G_{A_i}).\label{eq-GAi}
\end{eqnarray}
Then, since $G$ is extremal,
\begin{eqnarray*}
0\ge k_3(G_{A_i})-k_3(G)=|A_j|\cdot (T^o_{A_i}(G)-T^o_{A_j}(G))+|A_i|\cdot |A_j|\cdot (T_{A_iA_j}(G_{A_i})-T_{A_iA_j}(G)).
\end{eqnarray*}
Therefore, by~\eqref{eq-To}, we only need to show $T_{A_iA_j}(G_{A_i})\ge T_{A_iA_j}(G)$. Let
\begin{eqnarray*}
V_{1,1/2}&=&\left\{A_\ell: w(A_iA_\ell)=1\mbox{ and }w(A_jA_\ell)=\frac{1}{2}\right\},\\
V_{1/2,1}&=&\left\{A_\ell: w(A_iA_\ell)=\frac{1}{2}\mbox{ and }w(A_jA_\ell)=1\right\},\\
V_{1/2,1/2}&=&\left\{A_\ell: w(A_iA_\ell)=\frac{1}{2}\mbox{ and }w(A_jA_\ell)=\frac{1}{2}\right\},\\
V_{1,1}&=&\left\{A_\ell: w(A_iA_\ell)=1\mbox{ and }w(A_jA_\ell)=1\right\}.
\end{eqnarray*}
Denote by $|V_{p,q}|=\sum_{A_{\ell}\in V_{p,q}}|A_{\ell}|$ for $p,q\in\{1/2,1\}$. We have
\begin{eqnarray*}
T_{A_iA_j}(G_{A_i})-T_{A_iA_j}(G)=\left(\frac{1}{2}-\frac{1}{4}\right)|V_{1,1/2}|-\left(\frac{1}{4}-\frac{1}{8}\right)|V_{1/2,1}|=\frac{1}{4}|V_{1,1/2}|-\frac{1}{8}|V_{1/2,1}|.
\end{eqnarray*}
Therefore, it suffices to show $2|V_{1,1/2}|\ge |V_{1/2,1}|$. For the sake of contradiction, assume
\begin{eqnarray}\label{eq-cont}
|V_{1/2,1}|> 2|V_{1,1/2}|.
\end{eqnarray}
We will show that~\eqref{eq-cont} contradicts the extremality of $G$. Note that
\begin{eqnarray}\label{eq-GAj}
k_3(G_{A_j})&=& k_3(G')+(|A_i|+|A_j|)\cdot T^o_{A_j}(G)+|A_i|\cdot |A_j|\cdot T_{A_iA_j}(G_{A_j}).
\end{eqnarray}
By~\eqref{eq-G},~\eqref{eq-GAi},~\eqref{eq-GAj}, and the extremality of $G$ we have
\begin{eqnarray}
k_3(G_{A_i})\le k_3(G)\Leftrightarrow\left(\frac{1}{4}|V_{1,1/2}|-\frac{1}{8}|V_{1/2,1}|\right)|A_i|\cdot|A_j|+|A_j|\cdot T^o_{A_i}(G)\le |A_j|\cdot T^o_{A_j}(G),\label{eq-k3i}\\
k_3(G_{A_j})\le k_3(G)\Leftrightarrow\left(\frac{1}{4}|V_{1/2,1}|-\frac{1}{8}|V_{1,1/2}|\right)|A_i|\cdot|A_j|+|A_i|\cdot T^o_{A_j}(G)\le |A_i|\cdot T^o_{A_i}(G).\label{eq-k3j}
\end{eqnarray}
Then~\eqref{eq-k3i} and~\eqref{eq-k3j} imply
\begin{eqnarray*}
\left(\frac{1}{4}|V_{1,1/2}|-\frac{1}{8}|V_{1/2,1}|\right)|A_i|+T^o_{A_i}(G)\le T^o_{A_j}(G),\\
\left(\frac{1}{4}|V_{1/2,1}|-\frac{1}{8}|V_{1,1/2}|\right)|A_j|+ T^o_{A_j}(G)\le T^o_{A_i}(G).\end{eqnarray*}
Therefore,
\begin{eqnarray*}
&&\left(\frac{1}{4}|V_{1/2,1}|-\frac{1}{8}|V_{1,1/2}|\right)|A_j|\le T^o_{A_i}(G)-T^o_{A_j}(G)\le
\left(\frac{1}{8}|V_{1/2,1}|-\frac{1}{4}|V_{1,1/2}|\right)|A_i|\\
&\Rightarrow&\frac{1}{8}|V_{1/2,1}|(2|A_j|-|A_i|)\le \frac{1}{8}|V_{1,1/2}|(|A_j|-2|A_i|){\overset{\eqref{eq-cont}}{<}}\frac{1}{16}|V_{1/2,1}|(|A_j|-2|A_i|)\\
&\Rightarrow&4|A_j|-2|A_i|<|A_j|-2|A_i|\quad \Rightarrow\quad 4|A_j|<|A_j|,
\end{eqnarray*}
a contradiction.
For (ii), by the extremality of $G$, it suffices to show that $k_3(G_{A_i})+k_3(G_{A_j})\ge 2k_3(G)$. By~\eqref{eq-o},~\eqref{eq-G},~\eqref{eq-GAi} and~\eqref{eq-GAj}, we have
$$k_3(G_{A_i})+k_3(G_{A_j})-2k_3(G)=|A_i||A_j|\cdot ( T_{A_iA_j}(G_{A_i})+T_{A_iA_j}(G_{A_j})-2T_{A_iA_j}(G)).$$
It is left to show that $T_{A_iA_j}(G_{A_i})+T_{A_iA_j}(G_{A_j})-2T_{A_iA_j}(G)\ge 0$. Indeed,
\begin{eqnarray*}
&&T_{A_iA_j}(G_{A_i})+T_{A_iA_j}(G_{A_j})-2T_{A_iA_j}(G)\\
&&=\frac{1}{2}\sum_{\substack{1\le k\le m,\\ k\neq i,j}}w(A_iA_k)^2|A_k|+
\frac{1}{2}\sum_{\substack{1\le k\le m,\\ k\neq i,j}}w(A_jA_k)^2|A_k|-
2\cdot\frac{1}{2}\sum_{\substack{1\le k\le m,\\ k\neq i,j}}w(A_iA_k)w(A_jA_k)|A_k|\\
&&=\frac{1}{2}\sum_{\substack{1\le k\le m,\\ k\neq i,j}}\left(w(A_iA_k)-w(A_jA_k)\right)^2|A_k|\ge 0.
\end{eqnarray*}
\end{proof}
Denote by $S_2$ the following process: let $\sigma$ be the lexicographical ordering of ${[m]\choose 2}$ and apply $S_2(i,j)$, according to $\sigma$, for all pairs $(i,j)$ with $w(A_iA_j)=1/2$ . By Claim~\ref{cl-2-symm}, $S_2$ is finite and keeps the number of triangles.
\begin{claim}
The operation $S_2$ does not change the weighted clique number of $G$.
\end{claim}
\begin{proof}
Let $(X,Y)$ be one of the largest weighted complete subgraphs of $G$ of size $\ell$. Note that $|Y|$ is still $m$.
Also, since we only repeat this operation for vertices $x$ and $y$ with $w(xy)=1/2$, the operation is
not changing $|X|$ either. Hence, after repeated applications of this operation, the weighted clique number of $G$ will not change.
\end{proof}
After applying $S_2$, we have an equivalence relation on classes $A_1,\ldots, A_m$, which naturally extends to $V(G)$. To be precise,
denote by $\mathcal{B}=\{B_1, \ldots, B_{m'}\}$ the equivalence classes of this relation, i.e.~two vertices $u$ and $v$ are in the same class if and only if $uv\notin G_{1}$. Then, the $\mathcal{A}$-partition is a
refinement of the $\mathcal{B}$-partition. More importantly, the size of the largest weighted complete subgraph is $m+m'$.
We will next show that we can perform some transformations (Claims~\ref{claim-twoA} and~\ref{claim-oneB}) to get a more structured graph (as those in Construction~\ref{const-oddeven}) without increasing the weighted clique number and decreasing the number of triangles.
\begin{claim}\label{claim-twoA}
Each $B_i$ contains at most two $A_j$'s.
\end{claim}
\begin{proof}
Let us assume that $B_1$ contains $k$ $A_j$'s, $A_1,\ldots, A_k$, where $k\ge 3$. Denote by $U$ the vertex
set of $B_1$ and write $u=|U|$.
Note that the edges between two $B_i$'s always have weight $1$ and the edges inside an $A_i$ have
weight $0$ and all the other edges have weight $1/2$. We will divide the proof into three cases depending
on the value of $k$. In each case, we will modify $B_1$ by splitting it into multiple parts. This modification will only change the
weight of the edges with both ends in $U$ and also the equivalence classes $\mathcal{A}$ and $\mathcal{B}$. Then we
need to prove that the weighted clique number did not increase, and the number of triangles increases. For the latter, since the weight of the edges with at least one end in $V\setminus U$ remain the
same, we only need to show that the number of triangles with two or three vertices in $U$ did not
decrease. Therefore, it suffices to show that both $e(U)$ and $T(U)$ increase.
\textbf{Case 1:} Assume $k\ge 5$, which implies $u\ge 5$. We will split vertices in $U$ into three parts, $B_{11}$, $B_{12}$ and
$B_{13}$, such that $|B_{11}|\le|B_{12}|\le|B_{13}|\le |B_{11}|+1$. Also, define $A_i=B_{1i}$ for all
$1\le i\le 3$. For every $u\in U$ and $v\in V\setminus U$, we will not change $w(uv)$. For all
vertices $u,u'\in U$ if they belong to the same $B_{1i}$, let $w(uu')=0$, otherwise let $w(uu')=1$. The
equivalence classes $\mathcal{A}$ and $\mathcal{B}$ will change to $\{A_1,A_2,A_3,A_{k+1},\ldots,A_{m}\}$ and
$\{B_{11}, B_{12}, B_{13}, B_2,\ldots,B_{m'}\}$. Since $k\ge 5$, the number of classes in the $\mathcal{A}$ partition
decreased by at least two and the number of classes in the $\mathcal{B}$ partition increased by exactly $2$, hence,
the weighted clique number of $G$ will not increase. Now, we only need to show that the number of
triangles in the graph $G$ increases.
\begin{eqnarray*}
&\mbox{before:}&e(U)\le{k\choose 2}\frac{u^2}{k^2}\cdot\frac{1}{2}< \frac{u^2}{4},\\
&\mbox{after:}&e(U)=
\left\{
\begin{array}{ll}
3\cdot\frac{u^2}{9}=\frac{u^2}{3}& \mbox{if }u\equiv 0\text{ (mod } 3), \\
\frac{(u-1)^2}{9}+2\cdot\frac{(u-1)(u+2)}{9}=\frac{u^2-1}{3}& \mbox{if }u\equiv 1\text{ (mod } 3), \\
\frac{(u+1)^2}{9}+2\cdot\frac{(u-2)(u+1)}{9}=\frac{u^2-1}{3}& \mbox{if }u\equiv 2\text{ (mod } 3).
\end{array}
\right.
\end{eqnarray*}
Therefore $e(U)$ increases for $u\ge 2$. Now, for $T(U)$ we have
\begin{eqnarray*}
&\mbox{before:}&T(U)\le{k\choose 3}\frac{u^3}{k^3}\cdot\frac{1}{8}\le \frac{u^3}{48},\\
&\mbox{after:}&T(U)=
\left\{
\begin{array}{ll}
\frac{u^3}{27} & \mbox{if }u\equiv 0\text{ (mod } 3), \\
\frac{(u-1)(u-1)(u+2)}{27}& \mbox{if }u\equiv 1\text{ (mod } 3), \\
\frac{(u-2)(u+1)(u+1)}{27}& \mbox{if }u\equiv 2\text{ (mod } 3),
\end{array}
\right.
\end{eqnarray*}
which means that $T(U)$ increases if $u\ge 3$.
\textbf{Case 2:} Assume $k=4$ which implies $u\ge 4$. Let us split vertices in $U$ into three parts $A_{1}$, $A_{2}$ and $A_{3}$, such that $|A_{1}|\le|A_{2}|\le|A_3|\le |A_1|+1$. Also let $B_{11}=A_1\cup A_2$ and $B_{12}=A_3$. For all vertices $u,u'\in U$ if they are in different $B_{1i}$'s then $w(uu')=1$. If they are both in $B_{11}$ but in different $A_i$'s then $w(uu')=1/2$, and $w(uu')=0$ if they are in the same $A_i$. The equivalence classes $\mathcal{A}$ and $\mathcal{B}$ will change to $\{A_1,A_2,A_3,A_{5},\ldots,A_{m}\}$ and $\{B_{11}, B_{12}, B_2,\ldots,B_{m'}\}$. Notice that the number of classes in the $\mathcal{A}$ partition
decreased by one and the number of classes in $\mathcal{B}$ increased by one, hence,
the weighted clique number of $G$ will not change. For $e(U)$:
\begin{eqnarray*}
&\mbox{before:}&e(U)\le{4 \choose 2}\frac{u^2}{16}\cdot \frac{1}{2}=\frac{3u^2}{16},\\
&\mbox{after:}&e(U)=
\left\{
\begin{array}{ll}
\frac{1}{2}\cdot\frac{u^2}{9}+2\cdot\frac{u^2}{9} & \mbox{if }u\equiv 0\text{ (mod } 3), \\
\frac{1}{2}\cdot\frac{(u-1)(u-1)}{9}+2\cdot\frac{(u-1)(u+2)}{9}& \mbox{if }u\equiv 1\text{ (mod } 3), \\
\frac{1}{2}\cdot\frac{(u-2)(u+1)}{9}+\frac{(u-2)(u+1)}{9}+\frac{(u+1)(u+1)}{9}& \mbox{if }u\equiv 2\text{ (mod } 3).
\end{array}
\right.
\end{eqnarray*}
For $u\ge 2$, $e(U)$ increases. We also need to show that $T(U)$ increases:
\begin{eqnarray*}
&\mbox{before:}& T(U)\le 4\cdot\frac{u^3}{4^3}\cdot\frac{1}{8}=\frac{u^3}{4^3}\cdot\frac{1}{2},\\
&\mbox{after:}&T(U)=
\left\{
\begin{array}{ll}
\frac{1}{2}\cdot\frac{u^3}{27} & \mbox{if }u\equiv 0\text{ (mod } 3), \\
\frac{1}{2}\cdot\frac{(u-1)(u-1)(u+2)}{27}& \mbox{if }u\equiv 1\text{ (mod } 3), \\
\frac{1}{2}\cdot\frac{(u-2)(u+1)(u+1)}{27}& \mbox{if }u\equiv 2\text{ (mod } 3).
\end{array}
\right.
\end{eqnarray*}
Therefore $T(U)$ will increase for $u\ge 3$.
\textbf{Case 3:} Assume $k=3$, which implies $u\ge 3$. First, suppose that $u\le 12n/13$. Split vertices in $U$ into two equal parts $B_{11}$ and $B_{12}$. Also define $A_1=B_{11}$ and $A_2=B_{12}$. For all vertices $u,u'\in U$ if they are in different $B_{1i}$'s then set $w(uu')=1$, and \mbox{$w(uu')=0$} otherwise. The equivalence classes $\mathcal{A}$ and $\mathcal{B}$ will change to $\{A_1,A_2,A_{4},\ldots,A_{m}\}$ and $\{B_{11}, B_{12}, B_2,\ldots,B_{m'}\}$. Notice that the number of classes in the $\mathcal{A}$ partition
decreased by one and the number of classes in the $\mathcal{B}$ partition increased by one, hence,
the weighted clique number of $G$ will not change. For the change on the number of triangles, we have
\begin{eqnarray*}
&\mbox{before:}& T(U)+e(U)(n-u)\le \frac{u^3}{3^3}\cdot\frac{1}{2^3}+\frac{1}{2}\cdot{3\choose 2}\frac{u^2}{9}(n-u),\\
&\mbox{after:}&T(U)+e(U)(n-u)=
\left\{
\begin{array}{ll}
0+\frac{u^2}{4}\cdot(n-u) & \mbox{if }u\text{ is even},\\
0+\frac{(u-1)(u+1)}{4}\cdot(n-u)\ge\frac{u^2}{4.5}(n-u)& \mbox{if }u\text{ is odd}.
\end{array}
\right.
\end{eqnarray*}
Since $u\le 12n/13$, we have
\begin{gather*}
\frac{u^3}{3^3}\cdot\frac{1}{2^3}+\frac{3}{2}\cdot(n-u)\cdot\frac{u^2}{3^2}\le (n-u)\cdot\frac{u^2}{4.5}\quad\Leftrightarrow\quad\frac{u^3}{6^3}\le (n-u)\cdot\frac{u^2}{18}\quad\Leftrightarrow\\ \frac{u}{12}\le (n-u)\quad\Leftrightarrow\quad u\le \frac{12}{13}n.
\end{gather*}
We may now assume that $u>12n/13$. Let $U'$ be the vertex set of
$B_2$ and $u'=|B_2|$. Since $u\ge 12n/13$ and $u'\le n/13$, we may assume $B_2$ contains at most two $A_i$'s. Note that $u'\le u/12$. We split $U\cup U'$ into three classes of the same size, $B_{0}$, $B_{1}$ and $B_{2}$. Define $A_0=B_0$, $A_1=B_1$, and $A_2=B_2$. For two vertices $u,u'\in U\cup U'$, if they belong to the same $B_i$ then $w(uu')=0$, otherwise $w(uu')=1$. The equivalence classes $\mathcal{A}$ and $\mathcal{B}$ will change to $\{A_0,A_1,A_2, A_{5},\ldots,A_{m}\}$ and $\{B_{0}, B_{1}, B_2, B_3, \ldots,B_{m'}\}$. Notice that the number of classes in $\mathcal{A}$ decreased by one and the number of classes in $\mathcal{B}$ increased by one, which implies that the weighted clique number of $G$ will not change. We are left to
show that this operation will increase $e(U\cup U')$ and $T(U\cup U')$:
\begin{eqnarray*}
&\mbox{before:}& e(U\cup U')\le\frac{u^2}{3^2}\cdot\frac{3}{2}+uu'+\frac{u'^2}{8}\le\frac{u^2}{6}+\frac{u^2}{12}+\frac{u'^2}{8}=\frac{3u^2}{12}+\frac{u'^2}{8},\\
&\mbox{after:}&e(U\cup U')=
\left\{
\begin{array}{ll}
3\cdot\frac{(u+u')^2}{9} & \mbox{if }u+u'\equiv 0\text{ (mod } 3), \\
\frac{(u+u'-1)(u+u'-1)}{9}+2\cdot\frac{(u+u'-1)(u+u'+2)}{9}& \mbox{if }u+u'\equiv 1\text{ (mod } 3), \\
2\cdot\frac{(u+u'-2)(u+u'+1)}{9}+\frac{(u+u'+1)^2}{9}& \mbox{if }u+u'\equiv 2\text{ (mod } 3).
\end{array}
\right.
\end{eqnarray*}
\noindent Hence $e(U\cup U')$ is increasing for $u\ge 3$. We also have
\begin{eqnarray*}
&\mbox{before:}& T(U\cup U')\le\left(\frac{u}{3}\right)^3\cdot\frac{1}{8}+\frac{3}{2}\cdot\frac{u^2}{3^2}\cdot u'+u\cdot\frac{u'^2}{8}\le \frac{u^3}{6^3}+\frac{u^3}{6\cdot 12}+\frac{u^3}{8\cdot 12^2}\le\frac{u^3}{51.5},\\
&\mbox{after:}&T(U\cup U')=
\left\{
\begin{array}{ll}
\frac{(u+u')^3}{27} & \mbox{if }u+u'\equiv 0\text{ (mod } 3), \\
\frac{(u+u'-1)(u+u'-1)(u+u'+2)}{27}& \mbox{if }u+u'\equiv 1\text{ (mod } 3), \\
\frac{(u+u'-2)(u+u'+1)(u+u'+1)}{27}& \mbox{if }u+u'\equiv 2\text{ (mod } 3).
\end{array}
\right.
\end{eqnarray*}
Therefore $T(U\cup U')$ increases for $u+u'\ge 3$.
\end{proof}
\begin{claim}\label{claim-oneB}
There is at most one $B_i$ that contains two $A_j$'s.
\end{claim}
\begin{proof}
Now, we know that no $B_i$ contains three or more $A_i$'s. Let us assume that $B_1=A_1\cup A_2$ and $B_{2}=A_3\cup A_4$. Denote by $U$ the vertex set of $B_1\cup B_{2}$, and write $u=|U|$. Since each $A_i$ contains at least one vertex, we have that $u\ge 4$. We will split the vertices in $U$ into three equal pieces, $B_{11}$, $B_{12}$ and $B_{13}$, and redefine $A_1=B_{11}$, $A_2=B_{12}$, and $A_3=B_{13}$. For two vertices $u,u'\in U$ if they are in two different $B_{1i}$'s then $w(uu')=1$, otherwise $w(uu')=0$. This operation will change $\mathcal{A}$ and $\mathcal{B}$ to $\{A_1,A_2,A_3, A_{5},\ldots,A_{m}\}$ and $\{B_{11}, B_{12}, B_{13}, B_3, \ldots,B_{m'}\}$, therefore the weighted clique number does not change. We only need to show that $e(U)$ and $T(U)$ increase.
\begin{eqnarray*}
&\mbox{before:}&e(U)\le\frac{u^2}{4^2}+\frac{u^2}{4}=\frac{5u^2}{16},\\
&\mbox{after:}&e(U)=
\left\{
\begin{array}{ll}
3\cdot\frac{u^2}{9} & \mbox{if }u\equiv 0\text{ (mod } 3), \\
\frac{(u-1)^2}{9}+2\cdot\frac{(u-1)(u+2)}{9}& \mbox{if }u\equiv 1\text{ (mod } 3), \\
\frac{(u+1)^2}{9}+2\cdot\frac{(u-2)(u+1)}{9}& \mbox{if }u\equiv 2\text{ (mod } 3),
\end{array}
\right.\\
&\mbox{before:}& T(U)\le\frac{u^2}{16}\cdot\frac{u}{2}=\frac{u^3}{32},\\
&\mbox{after:}&T(U)=
\left\{
\begin{array}{ll}
\frac{u^3}{27} & \mbox{if }u\equiv 0\text{ (mod } 3), \\
\frac{(u-1)^2(u+2)}{27}& \mbox{if }u\equiv 1\text{ (mod } 3), \\
\frac{(u+1)^2(u-2)}{27}& \mbox{if }u\equiv 2\text{ (mod } 3).
\end{array}
\right.
\end{eqnarray*}
It can be easily checked that for $u\ge 4$, both $e(U)$ and $T(U)$ are not decreasing.
\end{proof}
Now, we will use the Claims~\ref{claim-twoA} and~\ref{claim-oneB} to complete the proof of Lemma~\ref{lem-triangle}. Let us assume that the extremal graph has partitions $\mathcal{A}=\{A_1,\ldots,A_m\}$ and $\mathcal{B}=\{B_1,\ldots,B_{m'}\}$. Also, since $\mathcal{A}$ is a refinement of $\mathcal{B}$ and also by Claims~\ref{claim-twoA} and~\ref{claim-oneB}, we have $m'\le m\le m'+1$. When $t=2\ell+1$, the graph does not contain a weighted clique of size $2\ell+1$, which implies $m+m'\le 2\ell$. Therefore $m'=m=\ell$ will maximize the number of triangles. In particular, the extremal graph is an $\ell$-partite graph with partite sets $B_1\cup\ldots\cup B_{\ell}$, where $||B_i|-|B_j||\le 1$ for all $1\le i<j\le \ell$. Define $A_i=B_i$ for all $1\le i\le \ell$, and for two vertices $u$ and $v$ if they belong to two different $B_i$'s then $w(uv)=1$, otherwise $w(uv)=0$.
When $t=2\ell$, the graph does not contain a weighted clique of size $2\ell$ which implies $m+m'\le 2\ell-1$. Therefore, in the extremal example, $m'=\ell-1$ and $m=\ell$. Hence, the extremal example is an $(\ell-1)$-partite graph, with partite sets $B_1\cup\ldots\cup B_{\ell-1}$, and let $B_1=A_1\cup A_2$. Simple optimization shows that $|A_1|=|A_2|$, and, for all $2\le i\le \ell-1$, all the $B_i$'s have the same size, i.e.~$|B_1|=x$ and $|B_i|=(n-x)/(\ell-2)$ for all $2\le i\le\ell-1$. Fix two vertices $u$ and $v$, if they belong to two different $B_i$'s then set $w(uv)=1$. Otherwise, if they both
belong to $B_1$ but to different $A_i$'s then set $w(uv)=1/2$, and $w(uv)=0$ in all other cases. Now, we only need to maximize the number of triangles with respect to $x$, which is exactly the optimization in~\eqref{eq-al}, showing that $T(G)\le a_t n^3$. This completes the proof of
Lemma~\ref{lem-triangle}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm-k3}]
For any given integer $t\ge 6$, let $\ell=\floor{\frac{t}{2}}$. The lower bound comes from $\mathcal{H}(n,k)$ with $k=t$ in Construction~\ref{const-oddeven}. In this case, we solve an optimization problem to find the size of $V_i$'s that maximizes the number of triangles, which is how $a_\ell$ is defined in~\eqref{eq-al}.
For the upper bound, we will show that for any $\varepsilon>0$, there exists $\delta>0$ such that the following holds for sufficiently large $n$. Let $G$ be an $n$-vertex $K_t$-free graph with $\alpha(G)\le \delta n$. Then $G$ has at most $(1+\varepsilon)a_{\ell}n^3$ triangles.
Choose constants $0\ll \delta \ll\varepsilon_1\ll\varepsilon<1$. Let $R=R(\mathcal{C},w)$ be the weighted cluster graph obtained from applying Theorem~\ref{thm-color} to $G$ with $\varepsilon_1$ playing the role of $\varepsilon$. By Lemma~\ref{lem-completefree2}, we have that $R(\mathcal{C},w)$ does not contain a weighted clique of size $t$. Then the upper bound follows from Lemma~\ref{lem-triangle} and that
$$k_3(G)\le T(R)\cdot\frac{n^3}{|R|^3}+\varepsilon n^3\le (1+\varepsilon)a_{\ell}n^3,$$
as desired, where the last term bounds the number of triangles in $G$ that do not correspond to a triangle in $R$.
\end{proof}
\section{Concluding remarks}
In this paper, we study the Ramsey-Tur\'an extensions of two special cases of classical problems. We determine $\RF(2,k,o(n))$, that is, the maximum number of $2$-edge-colorings an $n$-vertex graph with independence number $o(n)$ can have without a monochromatic $K_k$, and $\RT(K_3,K_t,o(n))$, the maximum number of triangles in an $n$-vertex $K_t$-free graph with $o(n)$ independence number.
\subsection{3-edge-colorings}
The Ramsey-Tur\'an extension of the Erd\H{o} s-Rothschild problem for more than 2 colors remains widely open. It is known~\cite{ABKS} that $F(n,3,k)=3^{\ex(n,K_k)}$. It will be interesting to study for $3$-edge-colorings, $\RF(3,k,o(n))$. The following determines the case when forbidding monochromatic triangles. We give here only a sketch of a proof.
\begin{theorem}
$\RF(3,3,o(n))=2^{n^2/4+o(n^2)}$.
\end{theorem}
\begin{proof}[Sketch of a proof.] (Lower bound) Let $G\in \mathcal{H}(n,5)$ with $|V_1|=|V_2|=n/2$. Consider the following 3-edge-colorings. Color edges in $G[V_i]$, $i=1,2$, red and color the cross-edges $G[V_1,V_2]$ either green or blue.
(Upper bound) Let $G$ be an extremal graph, $\phi: E(G)\rightarrow \{\phi_1,\phi_2,\phi_3\}$ be a $3$-edge-coloring with no monochromatic $K_3$, and $(A_1,A_2,A_3)$ be the partition obtained from Lemma~\ref{lem-partition} such that $\alpha(G_{\phi_i}[A_i])=o(n)$. Let $R^*$ be the multigraph by taking the union $\cup_{i=1}^3R_{\phi_i}$, where $R_{\phi_i}$ is the cluster graph in color $\phi_i$.
Denote by $\mu_i$, $i=1,2,3$, the edge-density of the subgraph of $R^*$ induced by edges with multiplicity $i$. Note first that $\mu_3=0$, since otherwise a multiplicity-3 edge results in a weighted clique of size 3 in $\cap_i R_{\phi_i}$, contradicting to $\phi$ containing no monochromatic $K_3$. It suffices then to show $\mu_2\le 1/2$. Notice that no $\phi_i$-colored edge can have an endpoint in $A_i$, otherwise we have a $\phi_i$-colored triangle. This implies that
(i) for every $i\neq j\in [3]$, all edges in $R^*[A_i,A_j]$ have multiplicity 1 with color $\phi_k$, $k\neq i,j$;
(ii) for $i\in [3]$, all edges in $R^*[A_i]$ have multiplicity at most 2, colored in $\{j,k\}=[3]\setminus\{i\}$.
By (i), we only need to consider edges in $\cup_i R^*[A_i]$. By (ii), inside $A_i$, there is no color $\phi_i$. This together with the observation that edges colored in $\{\phi_p,\phi_q\}$ for any $p\neq q\in [3]$ is triangle-free, we have $\mu_2\le 1/2$ as desired.
\end{proof}
\subsection{Generalized Ramsey-Tur\'an for larger cliques}
It seems plausible that for the general case $\RT(K_s,K_t,o(n))$, $t> s\ge 3$, some graph from the following construction has the maximum number of $K_s$.
\begin{construction}\label{const-Hst}
Given $3\le s<t\le 2s-1$, denote by $\mathcal{H}(n,s,t)$ the family of $n$-vertex graphs $G$ on vertex set $V_1\cup\ldots\cup V_s$ obtained as follows. Let $H$ be an extremal $K_{t-s}$-free graph on vertex set $[s]$. Make $[V_i,V_j]$ complete bipartite if $ij\in E(H)$; otherwise, put a copy of $\BE(V_i,V_j)$ if $ij\not\in E(H)$. For every $i\in V(H)$ with $d_H(i)=s-1$, put a $|V_i|$-vertex triangle-free graph with $o(|V_i|)$ independence number in $V_i$.
\end{construction}
Note that all graphs $G$ in the above construction are $K_t$-free and have $o(n)$ independence number. Indeed, since $G[V_i]$ is triangle-free for all $i$, in order to have a copy of $K_t$, there should be at least $t-s$ classes, $V_{j_1},\ldots,V_{j_{t-s}}$, each containing 2 vertices that form a $K_{2(t-s)}$. This would imply for every $1\le p<q\le t-s$, $G[V_{j_p},V_{j_q}]$ contains a $K_4$, which contradicts to $H$ being $K_{t-s}$-free. It should also be noted that the sizes of $V_i$'s need to be optimized.
\begin{conj}
Given integers $t> s\ge 3$, one of the extremal graphs for $\RT(K_s,K_t,o(n))$ lies in $\mathcal{H}(n,s,t)$ from Construction~\ref{const-Hst} when $t\le 2s-1$, and lies in $\mathcal{H}(n,k)$ with $k=t$ from Construction~\ref{const-oddeven} when $t\ge 2s$.
\end{conj}
\subsection{Phase transition}
For a given graph $H$ and two functions $f(n)\le g(n)$, we say that the Ramsey-Tur\'an function for $H$ exhibits a {\em jump} or has a {\em phase transition} from $g(n)$ to $f(n)$ if
\begin{eqnarray*}
\limsup\limits_{n\rightarrow\infty}\frac{\RT(n,H,f(n))}{n^2}
<\liminf\limits_{n\rightarrow\infty}\frac{\RT(n,H,g(n))}{n^2}.
\end{eqnarray*}
Let $g_r(n)=n2^{-\omega(n)\log^{1-1/r}n}$. Balogh, Hu and Simonovits~\cite{BHS} showed that the Ramsey-Tur\'an function for the even clique $K_{2r}$ exhibits a jump from $o(n)$ to $g_r(n)$. A similar phenomenon happens in the more general setup.
\begin{theorem}\label{thm-jump}
\begin{enumerate}
\item\label{itm-jump56} $\RT(K_3,K_5,g_3(n))=o(n^3)$ and $\RT(K_3,K_6,g_3(n))=o(n^3)$.
\item\label{itm-jumpodd} Odd cliques larger than 5 are stable: for every $\ell\ge 3$,
$$\RT(K_3,K_{2\ell+1},g_{\ell+1}(n))=(1+o(1))\RT(K_3,K_{2\ell+1},o(n)).$$
\item\label{itm-jumpeven} Even cliques always exhibit a jump: for every $\ell\ge 3$,
$$\RT(K_3,K_{2\ell+2},g_{\ell+1}(n))=(1+o(1))\RT(K_3,K_{2\ell+1},o(n)).$$
\end{enumerate}
\end{theorem}
We will need a lemma by Balogh-Hu-Simonovits (Claim 6.1 in~\cite{BHS}).
\begin{lemma}\label{lem-BHS}
Let $G$ be an $n$-vertex graph with $\alpha(G)=g_q(n)$, where $g_q(n)=n2^{-w(n)\log^{1-1/q}n}$ and $\omega(n)\rightarrow\infty$ arbitrary slowly. If there exists a $K_q$ in the cluster graph of $G$, then $K_{2q}\subseteq G$.
\end{lemma}
\begin{proof}[Proof of Theorem~\ref{thm-jump}]
For~(\ref{itm-jump56}), note that $\RT(K_3,K_5,g_3(n))\le\RT(K_3,K_6,g_3(n))$. Therefore, we only need to prove $\RT(K_3,K_6,g_3(n))=o(n^3)$. By Lemma~\ref{lem-BHS}, if $G$ is an $n$-vertex $K_6$-free graph with $\alpha(G)\le g_3(n)$ then the cluster graph of $G$ is $K_3$-free, which means that $k_3(G)=o(n^3)$.
For~(\ref{itm-jumpodd}), note that $\RT(K_3,K_{2\ell+1},g_{\ell+1}(n))\le \RT(K_3,K_{2\ell+1}, o(n))$, hence, by Theorem~\ref{thm-k3}, it is sufficient to prove $\RT(K_3,K_{2\ell+1},g_{\ell+1}(n))\ge(1+o(1)){\ell\choose 3}\left(\frac{n}{\ell}\right)^3$. Construction~\ref{const-oddeven} shows that this inequality holds.
For~(\ref{itm-jumpeven}), note that $\RT(K_3,K_{2\ell+1},g_{\ell+1}(n))\le \RT(K_3,K_{2\ell+2}, g_{\ell+1})$. Hence, using~(\ref{itm-jumpodd}), we only need to show that \mbox{$ \RT(K_3,K_{2\ell+2}, g_{\ell+1}(n))\le \RT(K_3,K_{2\ell+1},o(n))$}. By Lemma~\ref{lem-BHS}, if $G$ is an $n$-vertex $K_{2\ell+2}$-free graph with $\alpha(G)\le g_{\ell+1}(n)$ then the cluster graph of $G$ is $K_{\ell+1}$-free. Then, by the result of Erd\H{o} s~\cite{Erd}, among all $K_{\ell+1}$-free graphs the $\ell$-partite Tur\'an graph has the maximum number of triangles. Hence, we have
\begin{eqnarray*}
\RT(K_3,K_{2\ell+2},g_{\ell+1}(n))\le \left(1+o(1)\right){\ell\choose 3}\left(\frac{n}{\ell}\right)^3=RT(K_3,K_{2\ell+1}, o(n)),
\end{eqnarray*}
where the last equality is by Theorem~\ref{thm-k3}.
\end{proof}
\section{Acknowledgement}
The authors would like to thank the anonymous referees for their careful reading and helpful comments.
|
\section{Introduction}
The standard model (SM) exhibits features, such as the family repetition and the structure of
mixing matrices for quarks and leptons, that suggest an underlying structure.
Non--Abelian discrete flavor symmetries appear in many bottom--up models as a promising explanation
for these observations~\cite{Ishimori:2010au,Hirsch:2012ym,King:2013eh}.
A large set of Abelian and non--Abelian discrete symmetries has been successfully investigated
in this context~\cite{Aranda:1999kc,Aranda:2000tm,Aranda:2007dp,Babu:2004tn,Babu:2011mv,Babu:2002dz,Hirsch:2003dr,Ma:2001dn,Altarelli:2005yp,Altarelli:2010gt,deMedeirosVarzielas:2005qg,Grimus:2004rj,Grimus:2004hf,Morisi:2007ft,Bazzocchi:2008ej,Hirsch:2010ru,Campos:2014zaa,Hernandez:2013hea,King:2013eh,deMedeirosVarzielas:2006fc,King:2006np,Hernandez:2012ra,Ding:2009iy,Ding:2012wh}.
Particularly, the groups \Z3~\cite{Dev:2011jc,Peinado:2012tp,Sierra:2014kua,Aranda:2014lna},
$S_3$~\cite{Pakvasa:1977in,Frere:1978ds,Ma:1991eg,Fukugita:1998vn,Kubo:2003iw,Caravaglios:2005gw,Mondragon:2007af,Das:2015sca}
and $\Delta(27)$~\cite{deMedeirosVarzielas:2006fc,Luhn:2007uq,Morisi:2012hu,Ferreira:2012ri,Holthausen:2012dk,Aranda:2013gga,Varzielas:2015aua,Vien:2016tmh}
have shed some light on the structure of the quark and neutrino sectors, providing in some cases an explanation of proton stability
and dark matter~\cite{Morisi:2012hu,Chen:2013dpa,Hirsch:2010ru,Boucenna:2011tj,Lamprea:2016egz} or
an explanation of the {\it Dirac--ness} of neutrinos~\cite{Aranda:2013gga}.
These symmetries have in common that they are subgroups of $\Delta(54)$, which
however has been explored only aiming at a tri--bimaximal neutrino--mixing structure or
similar~\cite{Ishimori:2008uc,Ishimori:2009ew,Escobar:2011mq,Boucenna:2012qb}.
Since $\theta_{13}$ is now known to be non--zero, the potential of
$\Delta(54)$ as a flavor symmetry must be revisited. To pave the way to a
vast revision on this subject is one of the goals of this work.
On the other hand, despite their success, the origin of flavor symmetries remains
unexplained in bottom--up model building. Fortunately, non--Abelian flavor symmetries
emerge naturally in different compactification schemes
of string theory~\cite{Kobayashi:2004ud,Kobayashi:2006wq,Ko:2007dz,Nilles:2012cy,Beye:2014nxa,Beye:2015wka,Abe:2016eyh}
that enjoy the properties of the SM or its supersymmetric extension(s), yielding a promising
ultraviolet completion of flavor phenomenology.
Toroidal heterotic orbifolds~\cite{Dixon:1985jw,Dixon:1986jc} (see e.g.~\cite{Bailin:1999nk} for a comprehensive introduction)
lead to models which reproduce the gauge group and matter spectrum of the SM~\cite{Blaszczyk:2014qoa},
its minimal supersymmetric extension~\cite{Kobayashi:2004ya,Buchmuller:2005jr,Kim:2007mt,Blaszczyk:2009in}
and other non--minimal extensions~\cite{Lebedev:2009ag}, as well as many other observed and/or desirable properties
of particle physics~\cite{Buchmuller:2007zd,Kappl:2008ie,Choi:2009jt,Brummer:2010fr,Krippendorf:2012ir,Badziak:2012yg,Kim:2015mpa}.
As we discuss in section~\ref{sec:symmetries}, following previous findings of~\cite{Kobayashi:2006wq,Nilles:2012cy},
a $\Delta(54)$ flavor symmetry can emerge in these constructions as a result of dividing a $\mathbb T^2$ torus by
\Z3 in the compact dimensions. A paramount difference between the flavor theory emerging in
this context and one arbitrarily proposed is that all properties, including the flavor
representations and number of fields, are dictated by the string compactification itself, resulting
in interesting phenomenological consequences that we aim at studying in this paper.
Due to their geometrical structure, \Z3 or \Z3\x\Z2 heterotic orbifolds could
in principle yield a $\Delta(54)$ flavor symmetry,
but it is known that no promising model where this symmetry remains unbroken arises
in those cases~\cite{Lebedev:2006kn,Lebedev:2007hv,Lebedev:2008un}. Therefore, the simplest
complete string scenarios with SM--like physics and this flavor symmetry are \Z3\x\Z3 heterotic orbifolds.
In this paper, we explore the phenomenological viability of the $\Delta(54)$ flavor
symmetry from a top--down and a bottom--up perspective. After explaining in section~\ref{sec:symmetries}
how flavor symmetries relate to geometry in heterotic string compactifications, in section~\ref{sec:class}
we perform a search of semi--realistic \Z3\x\Z3 heterotic orbifold models, which turn out
to display $\Delta(54)$ as a flavor symmetry more naturally than other possibilities.
In section~\ref{sec:stringyDelta54} we inspect the flavor symmetries and spectrum
properties of one string sample model. Inspired by the features of the string models,
in section~\ref{sec:Delta54pheno} we propose a model that reproduces at some level
known flavor observations and provides predictions for the neutrino sector.
In section~\ref{sec:conclusions} we provide our concluding remarks.
\section{Origin of flavor symmetries in heterotic orbifolds}
\label{sec:symmetries}
We follow here the discussion of~\cite{Kobayashi:2006wq,Nilles:2012cy}, stressing some important aspects for our work.
In higher dimensional models, such as the string theories, flavor symmetries result from
the geometrical symmetries (and other properties) of the extra dimensions (see e.g.~\cite{Altarelli:2006kg}
for a field--theoretical proposal). Since in those models
the extra dimensions must be compactified in order to justify that we only perceive four dimensions,
the compact space adopts geometrical structures which are endowed with symmetries that
are passed down, as flavor symmetries, to the fields arising in those constructions.
Among all possibilities, orbifolds are perhaps the simplest compactifications.
A $d$--dimensional orbifold is defined as the quotient of $\mathbb R^d$ divided by a
discrete group. The resulting space is a compact solid, exhibiting typically
some curvature singularities (fixed points of the orbifold),
at which matter states may be localized. In the absence of local effects
at the singularities, the states attached to all singularities are
indistinguishable. The transformations (permutations, reflections, etc.) of those
identical states that leave the matter distribution invariant build a (non-Abelian)
symmetry of the compactified theory. Note that such transformations are equivalent to
field relabelings.
As a first example, let us suppose that an orbifold yields a compact space endowed with
two singularities at which two matter generations are chosen to be localized. Since these localized
matter generations are indistinguishable, i.e. have identical quantum numbers,
excepting of course for their localization properties, a permutation or relabeling of
the generations does not alter the system. That is, the system is invariant under
an $S_2$ permutation symmetry, leading to an effective model with
two generations related to each other under the non--trivial (flavor)
transformation of that group.
In string theory, the simplest and yet quite promising compactifications of this kind
are toroidal heterotic orbifolds~\cite{Dixon:1985jw,Dixon:1986jc}.
They are achieved by letting first the six extra
dimensions of a 10D heterotic string be compact by imposing the quotient
$\ensuremath{\mathbb{R}}^6/\Lambda_G$, where $\Lambda_G$ can be chosen as a 6D root lattice of a
Lie group $G$. The resulting 6D torus $\ensuremath{\mathbb{T}}^6=\ensuremath{\mathbb{R}}^6/\Lambda_G$ is then divided by
a discrete group of its isometries $P$, yielding the orbifold $\mathbb O = \ensuremath{\mathbb{T}}^6/P$.
$\mathbb O$ is Abelian when $P$ is Abelian. For simplicity,
we shall focus here only on Abelian orbifolds.
Not any arbitrary choice of $\mathbb T^6$ and $P$ is admissible. Requiring unbroken
supersymmetry in the effective 4D field theory as well as considering topological
equivalences between compactifications with different geometries
reduce greatly the number of allowed heterotic orbifolds. In fact,
all possible 6D orbifolds of this type have been exhaustively
classified~\cite{Fischer:2012qj}, resulting in a small number of Abelian orbifolds
and thus a small number of possible geometrical symmetries to be considered.
In contrast to a bottom--up approach, where matter fields are arbitrarily
localized at the singularities or let free in the bulk, in heterotic orbifolds
matter localization is restricted by the compactification rules.
All fields of the 4D effective field theories emerging from heterotic compactifications
arise from the (anomaly, tachyon and ghost free) spectrum of excitations
of {\it closed} strings that are not affected by the action of the orbifold.
In (supersymmetric) heterotic orbifolds, bulk or {\it untwisted fields}
correspond to the orbifold--invariant states arising directly from the 10D closed strings of
the uncompactified heterotic string, whose field limit is 10D $\ensuremath{\mathcal{N}}=1$ supergravity endowed
with an \E8\x\E8 or \SO{32} Yang--Mills theory. Thus, the 4D gauge (super)fields, generating
the unbroken 4D gauge group $\ensuremath{\mathcal{G}}_{4D}\subset\E8\x\E8$ or \SO{32}, and some 4D matter
states live in the bulk of a heterotic orbifold.
Additionally, there are the so--called {\it twisted fields}, which arise from strings that
are closed only due to the action of the orbifold.
Twisted fields are always localized at singularities of the orbifold and are thus
instrumental in the conception of a flavor theory with non--trivial representations from strings.
As long as there are no further compactification ingredients, such as Wilson lines~\cite{Ibanez:1986tp}
or discrete torsion~\cite{Vafa:1986wx,Sharpe:2000ki,Gaberdiel:2004vx,Ploger:2007iq}, that may lead to differences
in the states at the singular points, the {\it twisted} spectrum is degenerate, i.e. all
singularities carry identical twisted string states.
Couplings among string states are subject to a set of constraints called string
selection rules~\cite{Hamidi:1986vh,Dixon:1986qv,Casas:1991ac,Kobayashi:1991rp,Kobayashi:2011cw,Nilles:2013lda,Bizet:2013gf},
due to symmetries of the underlying conformal field theory of the compactified string theory.
These selection rules establish for which combination of string states there is a non--zero
correlation function, and thus a non--zero coupling for the associated effective fields.
In the 4D model emerging from an Abelian heterotic orbifold, the selection rules amount to including additional
(Abelian $\Z{N}\x\Z{M}\x\cdots$) symmetries and assign thus appropriate discrete
charges to each field in the model.
Thus, we notice that flavor symmetries in Abelian toroidal heterotic orbifolds
have two sources: the group of non--Abelian (relabeling) symmetries $G_{nA}$ from the geometrical
structure of the compactification space and the group of Abelian symmetries $G_A$
from the string selection rules.
In the case that the string selection rules provide a normal subgroup (invariant under
conjugation) of the full symmetry group, the resulting flavor symmetry is isomorphic to
the semi--direct product $G_{nA}\ltimes G_A$ (see e.g.~\cite{Ishimori:2012zz}).
Let us turn now to a relevant example for the present work. Suppose that two extended
dimensions are compactified in the orbifold $\ensuremath{\mathbb{T}}^2/\Z3$, where we choose the torus
to be defined by the root lattice $\Lambda_{\SU3}$ which is invariant under the \Z3
generator $\vartheta=e^{2\pi\mathrm{i}/3}$ in complex coordinates. That is, in the orbifold,
points $z_1$ and $z_2$ of $\mathbb C$ are equivalent if they can be related by
$z_1=\vartheta z_2 + \lambda$, $\lambda\in\Lambda_{\SU3}$. In this orbifold, there exist
three inequivalent fixed points or orbifold singularities\footnote{Analogous results are obtained for
the second non--trivial \Z3 group element, $\vartheta^2$.}
$z_{f,m},\,m=0,1,2$, such that $z_{f,m}=\vartheta z_{f,m} + \lambda_m$
for some lattice vectors $\lambda_m$. We can choose the inequivalent fixed points to
be $z_{f,0}=0$, $z_{f,1}=\frac13(2 e_1 + e_2)$ and $z_{f,2}=\frac13(e_1 + 2 e_2)$,
where $\{e_\alpha\}$ span $\Lambda_{\SU3}$, as depicted in fig.~\ref{fig:T2overZ3a}. The
gray region contains all inequivalent points in this orbifold.
\begin{figure}[!t!]
\begin{center}
\subfigure[Fixed points]{
\label{fig:T2overZ3a}
\input Z3fixedpoints.pdf_t
}
\hskip 1cm
\subfigure[Symmetries and charges]{
\label{fig:T2overZ3b}
\input Delta54inZ3.pdf_t
}
\caption{Geometrical origin of a $\Delta(54)$ flavor symmetry in a $\ensuremath{\mathbb{T}}^2/\Z3$ orbifold.
If the fixed points are not further affected by the compactification, there is an $S_3$ permutation
symmetry. Further, string selection rules impose additional a $\Z3\x\Z3$ symmetry based on the
localization charges $m$ and $q$ of twisted states. The resulting symmetry is $S_3\ltimes\Z3^2=\Delta(54)$.}
\label{fig:T2overZ3}
\end{center}
\end{figure}
Remarking in fig.~\ref{fig:T2overZ3a} that the upper tip is equivalent to $z_{f,0}$
and that the lines on both sides of $z_{f,i}$, $i\neq0$, are identified, the orbifold
becomes the triangular pillow--like object with
three apices displayed in fig.~\ref{fig:T2overZ3b}. This solid is clearly invariant under
all possible apex permutations, as symbolized by the arrows in that figure. Thus, we identify a
geometrical $S_3$ symmetry.
When the current example is applied to heterotic orbifolds, the string selection rules demand
additionally that any coupling of the form $\Phi_{m_1}\Phi_{m_2}\Phi_{m_3}\cdots$ among string states $\Phi_{m_i}$
(setting $m=0$ for untwisted states) satisfy first $\sum_i m_i = 0\mod 3$. Noting that this relation
corresponds to a \Z3 symmetry, it can be rewritten as $\prod_i \kappa^{m_i}=\mathbbm1$ in terms of
a \Z3 generator $\kappa=e^{2\pi\mathrm{i}/3}$. Furthermore, assigning a charge $q=1$ to $\vartheta$--twisted states
(and $q=2$ to $\vartheta^2$--twisted states and $q=0$ to untwisted states),
non--vanishing string couplings require that the couplings themselves be non--twisted,
i.e. $\prod_i\vartheta^{q_i}=\mathbbm1$, which can be rewritten as $\sum_i q_i = 0\mod3$.
Thus, we identify a \Z3\x\Z3 arising from the selection rules.
Finally, since the \Z3\x\Z3 obtained is a normal subgroup of the group generated by $S_3$ and $\Z3\x\Z3$, then the resulting
effective flavor symmetry of a $\ensuremath{\mathbb{T}}^2/\Z3$ orbifold can be written as $\Delta(54)=S_3\ltimes\Z3^2$.
In the absence of Wilson lines and discrete torsion, twisted string states replicate in all
orbifold singularities, thus appearing always with a multiplicity of three
and building triplet representations. Since $\vartheta^{-1} = \vartheta^2$, twisted
states located at the $\vartheta^2$ fixed points have the opposite geometrical
quantum numbers of the $\vartheta$--twisted states. That is, if we label as $\rep3_{11}$
the $\vartheta$--twisted states,\footnote{We follow here the notation of~\cite{Ishimori:2012zz}
for $\Delta(54)$ representations; see appendix A.} those generated at the $\vartheta^2$ singularities build then the
representation $\rep3_{12}$. Untwisted states and twisted states affected by Wilson lines or
discrete torsion are just $\Delta(54)$ trivial singlets $\rep1_0$. No other $\Delta(54)$
representations appear in this context, yielding a tight and useful string constraint for
flavor phenomenology.
This discussion has been explicitly developed for all possible sub--orbifolds (in less than six dimensions)
appearing in Abelian toroidal heterotic orbifolds~\cite{Kobayashi:2006wq}, resulting in a reduced number
of family symmetries. The findings include, besides $\Delta(54)$, only the symmetries\footnote{Note
though that, under certain conditions, other symmetries may appear, as in~\cite{Altarelli:2006kg}.}
$D_4$, $(D_4\x D_4)/\Z2$, $(D_4\x\Z4)/\Z2$, $(D_4\x\Z8)/\Z2$ and $S_7\ltimes\Z7^6$.
As we shall see, these symmetries are enlarged in the full 6D heterotic orbifold,
but can then be finally reduced back to these symmetries in phenomenologically viable models.
This may already be considered a phenomenologically relevant observation: not
any flavor symmetry is allowed in particle physics if it arises
from a compactified string theory.
\section{Classification of \Z3\x\Z3 heterotic orbifolds with $\Delta(54)$}
\label{sec:class}
The purpose of this section is to identify string models exhibiting a number of semi--realistic
properties and $\Delta(54)$ flavor symmetry in the simplest compactification scheme where such
models are present, \Z3\x\Z3 heterotic orbifolds.
\Z3\x\Z3 heterotic orbifolds are characterized by the quotient of
a so--called factorizable torus $\ensuremath{\mathbb{T}}^6=\ensuremath{\mathbb{T}}^2_1\x\ensuremath{\mathbb{T}}^2_2\x\ensuremath{\mathbb{T}}^2_3$ divided
by the joint action of two \Z3 isometries of $\ensuremath{\mathbb{T}}^6$
in the extra dimensions of a heterotic string. In the simplest case,\footnote{There
are 15 \Z3\x\Z3 choices, among which many include rototranslations~\cite{Fischer:2012qj}.}
the tori are described by the root lattice of $\SU3_1\x\SU3_2\x\SU3_3$
and the \Z3 generators act diagonally on the tori as
\begin{equation}
\vartheta = \diag\left(e^{2\pi\mathrm{i} v_1},e^{2\pi\mathrm{i} v_2},e^{2\pi\mathrm{i} v_3}\right)\,,\qquad
\omega = \diag\left(e^{2\pi\mathrm{i} w_1},e^{2\pi\mathrm{i} w_2},e^{2\pi\mathrm{i} w_3}\right)\,,
\label{eq:Z3xZ3twists}
\end{equation}
where $v$ and $w$ are the so--called twist vectors
\begin{equation}
v = \left(1/3,0,-1/3\right),\qquad w=\left(1/3,-1/3,0\right)\,.
\label{eq:Z3xZ3twistvectors}
\end{equation}
Consequently, each of the 2--tori are subject to a \Z3 orbifold.
According to our previous discussion, one may conjecture that these constructions
lead to a $\Delta(54)^3$ flavor symmetry, but this is wrong. In fact, in this
case the relabeling symmetry that naturally appears is $S_3\x S_3\x S_3$.
Further, concerning the symmetries due to string selection rules,
invariance under the two twists, $\vartheta$ and $\omega$, leads to
two \Z3 symmetries analogous to the one for the $q$ charge in the previous section.
In addition, localization selection rules introduce one extra \Z3 factor for each
2--torus. That is, the natural flavor symmetry in these heterotic orbifolds
is $(S_3\x S_3\x S_3)\ltimes\Z3^5$.
Note, however, that the relabeling symmetry can be further enhanced to
$S_{27}$ if the sizes of the tori $\ensuremath{\mathbb{T}}^2_a$, $a=1,2,3$,
are identical and no Wilson lines nor discrete torsion is invoked. As we shall shortly see,
phenomenologically viable models only arise if one introduces Wilson lines. In fact,
most promising models have two Wilson lines that distinguish the states located at the singularities
of two of the tori, retaining only the non-Abelian $S_3$ relabeling symmetry.
Further, there is no reason why all tori should have the same size; their sizes
(and also their shapes) are encoded in the values of (untwisted) moduli that can
{\it a priori} have arbitrary values.
Once the generic geometrical aspects of the compactification have been set, our
task is now to apply this compactification to a heterotic string. We restrict ourselves
here to the $\ensuremath{\mathcal{N}}=1$ \E8\x\E8 heterotic string, but expect similar results from
the $\ensuremath{\mathcal{N}}=1$ \SO{32} heterotic string.\footnote{We also expect promising non--supersymmetric
models arising from the $\ensuremath{\mathcal{N}}=0$ \SO{16}\x\SO{16} heterotic string, although the
presence of tachyons at some level of the theory would still be a worry.}
Modular invariance of the partition function demands the orbifold to be embedded
into the gauge group \E8\x\E8. This gauge embedding consists in choosing a 16D
(shift) vector for each of the twists performed in the six compact dimensions
and a so--called 16D Wilson--line vector $A_\alpha$, $\alpha=1,\ldots,6$,
encoding in the gauge degrees of freedom each $e_\alpha$ of $\ensuremath{\mathbb{T}}^6$.
The gauge embedding is subject to three constraints. First,
modular invariance additionally imposes in \Z3\x\Z3 heterotic orbifolds that~\cite{Ploger:2007iq}
\begin{eqnarray}
\label{eq:ModInv}
3\,(V^2-v^2) = 0\mod2\,, &\quad& 3\,(V\cdot A_\alpha) = 0\mod 2\,,\quad\alpha=1,\ldots,6\,,\\
3\,(W^2-w^2) = 0\mod2\,, &\quad& 3\,(W\cdot A_\alpha) = 0\mod 2\,, \nonumber\\
3\,(V\cdot W-v\cdot w) = 0\mod2\,, &\quad& 3\ A_\alpha^2 = 0\mod 2\,,\nonumber\\
&\quad& 3\,(A_\alpha\cdot A_\beta) = 0\mod 2\,,\quad \alpha\neq\beta\,,\nonumber
\end{eqnarray}
where $V$ and $W$ are the 16D vectors that denote respectively the gauge embeddings of the twists $v$ and $w$
of eq.~\eqref{eq:Z3xZ3twistvectors}. Secondly, both $V$ and $W$ must be consistent
with a \Z3\x\Z3 action. This amounts to requiring that three times these
vectors must be a trivial gauge transformation within \E8\x\E8, i.e.
for the shift vector $V$ (with entries $V^{(i)}$)~\footnote{These constraints
arise from the fact that the root lattice of each \E8 is even (and self--dual).
An arbitrary shift within the lattice does not alter the gauge degrees of freedom.}
\begin{equation}
3\sum_{i=1}^8 V^{(i)} = 0\mod 2\,,\qquad 3\sum_{i=9}^{16} V^{(i)} = 0\mod 2\,,
\label{eq:lattice-condition}
\end{equation}
demanding that the entries $V^{(i)}$ be all integer or half--integer, independently for $i=1,\ldots,8$
and $i=9,\ldots,16$. Analogous conditions must then be imposed to $W$.
The final constraint imposes that Wilson--line vectors must be consistent with
the choice of $\ensuremath{\mathbb{T}}^6$ lattice and the action of the orbifold on it.
The fact that the lattice vectors $e_\alpha$ are related by the action of $\vartheta$ and $\omega$
translates to relations among all $A_\alpha$. For instance, in \Z3\x\Z3 heterotic orbifolds
since $e_2 = \vartheta e_1$ (see e.g. fig.~\ref{fig:T2overZ3a}, valid in this case),
then $A_1 = A_2$ up to a trivial gauge transformation in \E8\x\E8.
One finds that these geometrical considerations lead to the conditions
\begin{eqnarray}
\label{eq:condWL}
A_\alpha = A_{\alpha+1}\,, &&\alpha=1,3,5\,,\\
3\sum_{i=1}^8 A_\alpha^{(i)} = 0\mod 2\,,&& 3\sum_{i=9}^{16} A_\alpha^{(i)} = 0\mod 2\,.\nonumber
\end{eqnarray}
A comment is in order. Notice that each 2--torus can be affected by up to one inequivalent, non--trivial
Wilson line. If one includes the Wilson line $A_{2a-1} = A_{2a}$ associated with the compactification
in the $\ensuremath{\mathbb{T}}^2_a$ torus, $a=1,2,3$, the relabeling symmetry $S_3$ of that torus disappears.
Thus, with one and two non--vanishing Wilson lines, the non--Abelian relabeling symmetry gets
broken down, respectively, to $S_3\x S_3$ and $S_3$, while no non--Abelian symmetry is left when
all three Wilson lines are non--trivial. Hence, it follows that only models with two non--trivial Wilson lines
can lead to a $\Delta(54)=S_3\ltimes\Z3^2$ flavor symmetry in \Z3\x\Z3 heterotic orbifolds.
After finding solutions to the constraints~\eqref{eq:ModInv}--\eqref{eq:condWL}, there are standard
techniques, discussed elsewhere in great detail (see e.g.~\cite{RamosSanchez:2008tn,Vaudrevange:2008sm}),
to determine the spectrum of massless string states, including their gauge quantum numbers, localization,
couplings and other properties of the {\it supersymmetric} effective field theory. Spectra obtained this way must then be
inspected from a phenomenological perspective, imposing criteria based on observable particle physics
(and/or cosmology) that may discriminate phenomenologically viable models from others.
Clearly, given the number of gauge--embedding parameters, the constraints~\eqref{eq:ModInv}--\eqref{eq:condWL}
can be satisfied for a large number of shift and Wilson--line vectors, making the task of identifying
phenomenologically viable heterotic orbifolds very time--consuming. Fortunately, this task becomes accessible
thanks to tools such as the \texttt{orbifolder}~\cite{Nilles:2011aj}, which automatizes the computation
of massless spectra, couplings and other important features of the models.
With the purpose of finding promising models endowed with a $\Delta(54)$ flavor symmetry,
we have used the \texttt{orbifolder} to randomly construct a large number of inequivalent \Z3\x\Z3
heterotic orbifold models. Models are considered to be equivalent by the software if no differences are
found when comparing the full gauge group, the non-Abelian gauge quantum numbers of the
resulting states and the number of non-Abelian gauge singlets in the massless spectrum.
From the created models, we have then selected the most promising ones. Here,
a promising model must yield the SM gauge group, such that the hypercharge generator be
non--anomalous and (with normalization) compatible with grand unification, three generations
of quarks and leptons, at least a couple of Higgs (super)fields, $H_u$ and $H_d$,
and only vectorlike exotics w.r.t. the SM gauge group.
Our results are as follows. We have obtained over $7\x10^{6}$
inequivalent \Z3\x\Z3 heterotic orbifold models~\footnote{Following the statistical approach
of~\cite[sec. 2.2]{Lebedev:2008un}, we estimate that the number of generated models represents
about 90\% of the total of possible models in this scenario.}, with up to (the maximum of) three
inequivalent Wilson lines. After applying our phenomenological constraints, only
789 models exhibit the required properties. We have verified that, considering
couplings\footnote{Given the persistent controversy about the selection rules in heterotic orbifolds,
we have considered only the so--called rule 4~\cite{Kobayashi:2011cw}, gauge and space--group invariance,
and $R$--charge conservation~\cite{Nilles:2013lda,Bizet:2013gf}.}
of the vectorlike exotics with up to six SM singlets, in a large number
of these models all exotics decouple once the SM singlets develop vacuum expectation values (VEVs).
Other models require higher dimensional operators to yield mass terms for all vectorlike exotics.
An interesting geometrical quality of the promising models regards the
effective family symmetry. Among the 789 selected models, most (696) of them
have two inequivalent non--vanishing Wilson lines. About 10\% of the
viable models (81 of them) require one non--trivial Wilson line,
and only 12 result from compactifications with three Wilson lines.
Therefore, we find that $\Delta(54)$ as a flavor symmetry of (MS)SM--like models
is favored in \Z3\x\Z3 heterotic orbifold models.
This outcome is compatible with previous results found in the
literature. Particularly, in ref.~\cite{Nilles:2014owa} the authors have found
445 \Z3\x\Z3 heterotic orbifold models with the properties we have required,
out of which 369 of them exhibit two non--trivial Wilson lines. In this
perspective, our search shows to be more exhaustive.
\section{A sample model with stringy $\Delta(54)$ flavor}
\label{sec:stringyDelta54}
With the purpose of exploring the flavor phenomenology produced by string compactifications,
let us now study the properties of one of the promising models
from our \Z3\x\Z3 heterotic orbifold scan, chosen due to its simplicity.
The parameters that define the model are the shift vectors
\begin{subequations}
\label{eqs:shifts_model9}
\begin{eqnarray}
3 V &=& \left(-\tfrac{1}{2}, -\tfrac{1}{2}, -\tfrac{1}{2}, -\tfrac{1}{2}, \tfrac{1}{2}, \tfrac{1}{2}, \tfrac{1}{2}, \tfrac{1}{2}; -2, 0, 0, 1, 1, 1, 1, 4\right)\,, \\
3 W &=& \left( 0, 1, 1, 4, 0, 0, 1, 1; 1, -1, 4, -4, -1, 0, 0, 1\right)\,,
\end{eqnarray}
\end{subequations}
and the Wilson lines
\begin{subequations}
\label{eqs:WL_model9}
\begin{eqnarray}
3 A_{1} = 3 A_2 & = & \left(-\tfrac{7}{2}, -\tfrac{3}{2}, \tfrac{9}{2}, \tfrac{7}{2}, -\tfrac{7}{2}, -\tfrac{3}{2}, \tfrac{5}{2}, \tfrac{7}{2}; -3, 0, -2, 0, -2, -4, 3, -2\right)\,,\\
3 A_{3} = 3 A_4 & = & \left( 3, 3, -3, -2, -1, 2, 4, -4; -3, 1, -1, -4, 1, 1, 4, 1\right)\,.
\end{eqnarray}
\end{subequations}
These parameters yield the unbroken gauge group $\SU3_C\x\SU2_L\x\U1_Y\x[\SU2\x\U1^{11}]$, where the
additional \SU2 factor is considered hidden because no SM--field carries a charge under that group.
However, all fields in the spectrum are charged under the additional \U1 factors.
\begin{table}[!t!]
\begin{center}
\begin{tabular}{|rlcc|rlcc|}
\hline
\# & irrep & $\Delta(54)$ & label & \# & anti-irrep & $\Delta(54)$ & label \\
\hline
3 & $\left(\rep{3}, \rep{2}\right)_{\frac{1}{6}}$ & $\rep3_{11}$ & $Q_i$ & & & & \\
3 & $\left(\crep{3}, \rep{1}\right)_{-\frac{2}{3}}$ & $\rep3_{11}$ & $\bar{u}_i$ & & & & \\
3 & $\left(\crep{3}, \rep{1}\right)_{\frac{1}{3}}$ & $\rep3_{11}$ & $\bar{d}_i$ & & & & \\
3 & $\left(\rep{1}, \rep{2}\right)_{-\frac{1}{2}}$ & $\rep3_{11}$ & $L_i$ & & & & \\
3 & $\left(\rep{1}, \rep{1}\right)_{1}$ & $\rep3_{11}$ & $\bar{e}_i$ & & & & \\
3 & $\left(\rep{1}, \rep{1}\right)_{0}$ & $\rep3_{12}$ & $\bar{\nu}_i$ & & & & \\
1 & $\left(\rep{1}, \rep{2}\right)_{-\frac{1}{2}}$ & $\rep1_0$ & $H_d$ & 1 & $\left(\rep{1}, \rep{2}\right)_{\frac{1}{2}}$ & $\rep1_0$ & $H_u$ \\
\hline
\multicolumn{8}{c}{Flavons}\\
\hline
3 & $\left(\rep{1}, \rep{1}\right)_{0}$ & $\rep3_{11}$ & $\phi^{u}_i$ & & & & \\
3 & $\left(\rep{1}, \rep{1}\right)_{0}$ & $\rep3_{11}$ & $\phi^{d,e}_i$ & & & & \\
3 & $\left(\rep{1}, \rep{1}\right)_{0}$ & $\rep3_{12}$ & $\bar\phi^{\nu}_i$ & & & & \\
2 & $\left(\rep{1}, \rep{1}\right)_{0}$ & $2\p\rep1_0$ & $s^{(d,e)},s^u$ & & & & \\
128 & $\left(\rep{1}, \rep{1}\right)_{0}$ & $77\p\rep1_0+16\p\rep3_{12}+\rep3_{11}$ & $N_i$ & & & & \\
\hline
\multicolumn{8}{c}{Exotic states}\\
\hline
16 & $\left(\rep{1}, \rep{2}\right)_{\frac{1}{6}}$ & $10\p\rep1_0+2\p\rep3_{12}$ & $v_i$ & 16 & $\left(\rep{1}, \rep{2}\right)_{-\frac{1}{6}}$ & $4\p\rep1_0+4\p\rep3_{12}$ & $\bar v_i$ \\
3 & $\left(\rep{3}, \rep{1}\right)_{0}$ & $\rep3_{12}$ & $y_i$ & 3 & $\left(\crep{3}, \rep{1}\right)_{0}$ & $3\x\rep1_0$ & $\bar y_i$ \\
1 & $\left(\crep{3}, \rep{1}\right)_{-\frac{1}{3}}$ & $\rep1_0$ & $z_i$ & 1 & $\left(\rep{3}, \rep{1}\right)_{\frac{1}{3}}$ & $\rep1_0$ & $\bar z_i$ \\
7 & $\left(\rep{1}, \rep{1}\right)_{-\frac{2}{3}}$ & $4\p\rep1_0+\rep3_{12}$ & $x_i$ & 7 & $\left(\rep{1}, \rep{1}\right)_{\frac{2}{3}}$ & $4\p\rep1_0+\rep3_{11}$ & $\bar x_i$ \\
51 & $\left(\rep{1}, \rep{1}\right)_{-\frac{1}{3}}$ & $30\p\rep1_0+7\p\rep3_{12}$ & $w_i$ & 51 & $\left(\rep{1}, \rep{1}\right)_{\frac{1}{3}}$ & $24\p\rep1_0+9\p\rep3_{12}$ & $\bar w_i$ \\
\hline
\end{tabular}
\caption{Massless spectrum. Representations w.r.t. $\SU3_C\times\SU2_L$
are given in bold face, the hypercharge is indicated by the subscript.
The 3rd and 7th columns display the $\Delta(54)$ flavor representations.}
\label{tab:spectrum_model9}
\end{center}
\end{table}
\begin{table}[!h!]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c||c|c||c|c|c||c|c|}
\hline
$\phantom{A^{A^{A^A}}}$& $Q_i$ & $\bar{d}_i^c$ & $\bar{u}_i^c$ & $L_i$ & $\bar{e}_i^c$ &$\bar{\nu}_i$& $H_u$ & $H_d$ & $\phi^u_i$ & $\phi^{(d,e)}_i$ & $\bar{\phi}^\nu_i$ & $s^u$ & $s^{(d,e)}$\\
\hline
$\Delta(54)$ &$\bs3_{11}$ & $\bs3_{11}$ & $\bs3_{11}$ & $\bs3_{11}$ & $\bs3_{11}$& $\bs3_{12}$ & $\bs1_0$ & $\bs1_0$ & $\bs3_{11}$ & $\bs3_{11}$ & $\bs3_{12}$ & $\bs1_0$ & $\bs1_0$ \\
\hline
$\Z3^{(1)}$ & $\omega$ & $1$ & $\omega$ & $1$ & $\omega$ & $1$ & $1$ & $1$ & $1$ & $1$ & $1$ & $\omega$ & $\omega^2$ \\
\hline
$\Z3^{(2)}$ & $1$ & $\omega^2$ & $1$ & $\omega^2$ & $1$ & $\omega$ & $1$ & $1$ & $1$ & $\omega$ & $\omega$ & $1$ & $1$ \\
\hline
$\Z3^{(3)}$ & $\omega$ & $1$ & $\omega$ & $1$ & $\omega$ & $1$ & $1$ & $1$ & $1$ & $1$ & $1$ & $\omega$ & $\omega^2$ \\
\hline
\end{tabular}
\caption{Flavor representations for the SM matter and flavon fields in a \Z3\x\Z3 sample model.
The \Z3 charges are defined in terms of the twist and field localizations in~\eqref{eq:Z3generators} with $\omega=e^{2\pi\mathrm{i}/3}$.}
\label{tab:matterfields}
\end{center}
\end{table}
Due to its two Wilson lines~\eqref{eqs:WL_model9}, the model has the flavor symmetry $S_3\ltimes\Z3^5\supset\Delta(54)$.
The $\Delta(54)$ quantum numbers are associated with the symmetries of the third torus, $\ensuremath{\mathbb{T}}^2_3$, whose
localized states are not affected by any Wilson line. If we allow for the spontaneous breakdown of the
three additional \Z3 symmetries by VEVs of appropriate SM singlets transforming as $\bs1_0$ under $\Delta(54)$,
the flavor symmetry in the vacuum is just $\Delta(54)$ and the extra $[\SU2\x\U1^{11}]$ gauge factors are broken too.
The gauge and $\Delta(54)$ representations of the massless matter spectrum of our sample model
are provided in table~\ref{tab:spectrum_model9}. As explained before, the only possible
$\Delta(54)$ representations are the trivial singlet $\rep1_0$ and the two triplets $\rep3_{11}$
and $\rep3_{12}$. One particular feature of the observable sector is that, directly from the string computation,
only three SM generations that build non--trivial flavor representations arise, while the
Higgs states are untwisted fields and thus uncharged under the flavor symmetry. On the other hand,
the exotic particles are vectorlike w.r.t. the SM gauge group, but not necessarily under the flavor group.
Despite this hurdle, there exist SM singlets $N_i$ in the appropriate flavor representations, so
that all exotics and the singlets $N_i$ themselves can acquire masses when $\vev{N_i}\neq0$.
To understand better the flavor phenomenology of the observable sector of this model,
we display in table~\ref{tab:matterfields} all flavor charges of the SM superfields and some gauge
singlets that shall serve as flavons. The \Z3 charges are given in terms of
\begin{equation}
\label{eq:Z3generators}
\omega^q\qquad\text{ and }\qquad \kappa_{a}^{\ m_a}\equiv (e^{2\pi\mathrm{i}/3})^{m_a}\,\qquad
\text{with}\quad a=1,2;\,q,m_a=0,1,2\,,
\end{equation}
where $\omega$ is the (eigenvalue of the) second twist in eq.~\eqref{eq:Z3xZ3twists}, $\kappa_a$ correspond to the \Z3
generators associated with the localization labels $m_a$ in the (first or second) torus $\ensuremath{\mathbb{T}}^2_a$,
as described in section~\ref{sec:symmetries}, and $q$ is the power of the twist that yields
the corresponding twisted states. Note that $\omega=\kappa_a=e^{2\pi\mathrm{i}/3}$.
Since the SM matter fields are charged under flavor symmetries, the presence of the properly charged
$s$ and $\phi$ flavon fields allows for Yukawa couplings in the (non--renormalizable) superpotential,
which in this case can be written as follows
\begin{eqnarray}
\label{eq:Wyuk}
W_{Y} &=& y_{ijk}^u Q_i H_u \bar{u}_j \phi^u_k s_u + y_{ijk}^d Q_i H_d \bar{d}_j \phi^{(d,e)}_k s^{(d,e)} + y_{ijk}^e L_i H_d \bar{e}_j \phi^{(d,e)}_k s^{(d,e)} \\
&+& y_{ijkl}^\nu L_i H_u \bar{\nu}_j+ \lambda_{ijk} \bar\nu_i \bar\nu_j \bar\phi^\nu_k\,, \qquad\qquad\qquad\qquad i,j,k=1,2,3, \nonumber
\end{eqnarray}
where the summation over repeated indices must follow the rules of the product of $\Delta(54)$ representations
that lead to invariant singlets (cf. appendix A),
$$\rep1_0\subset \rep3_{11}\x\rep3_{12},\quad \rep1_0\subset \rep3_{11}\x\rep3_{11}\x\rep3_{11}\,,\quad \rep1_0\subset \rep3_{12}\x\rep3_{12}\x\rep3_{12}.$$
In principle, all Yukawa--coupling coefficients, $y$ and $\lambda$, are computable by applying CFT techniques for the string model.
However, it is known that there are still some challenges to be solved for non--renormalizable couplings.
The best we can do here is to estimate that $y$ are order one (but with a suppression due non--renormalizability)
because they include the untwisted Higgs fields, whereas $\lambda$ must be somewhat suppressed because all involved
fields are twisted. We observe that the second row of $W_Y$ admits neutrino masses from a type I see--saw
mechanism with three right--handed (RH) neutrinos with proper $\phi$ flavon VEVs.
Similarly, the Dirac masses of charged leptons and quarks are determined by the VEVs of other
flavons $\phi$ and $s$. We point out that the structure of masses for down--quarks and charged leptons
is predicted in this model to be identical because the flavons involved in the corresponding couplings
are unavoidably the same. As we shall see, this enforces a more stringent sort of $b-\tau$ unification.
\section{Fermion masses from a $\Delta(54)$ flavor symmetry}
\label{sec:Delta54pheno}
The properties of the string--derived model presented before can be now studied
from a bottom--up perspective. Although our string sample model is supersymmetric and
all couplings are determined at the compactification scale, the general structure
of the Yukawa Lagrangian at low energies can be determined from $W_Y$, if we insist on retaining
the $\Delta(54)$ flavor symmetry in the soft--breaking sector. Besides, it
is known that Yukawa couplings do not receive large contributions
through the renormalization running~\cite{Xing:2007fb}. Similarly, threshold
corrections shall not alter the mass and mixing structure of quarks and leptons,
since it depends mainly on mass ratios.
Therefore, we can safely study the viability of the model by restricting ourselves
to the behavior of the appropriate non--supersymmetric fields.
In a compact notation, the effective Yukawa Lagrangian for quarks and charged leptons
that is obtained from $W_Y$ reads
\begin{eqnarray}
\label{eq:Lyuk}
{\cal L}_{Y}^f&=&y_1^f \left[F_1 H\bar{f}_1 \phi_1+F_2 H\bar{f}_2 \phi_2+F_3 H\bar{f}_3 \phi_3\right] \\
&+&y_2^f\left[(F_1 H\bar{f}_2+F_2 H\bar{f}_1)\phi_3+(F_3 H\bar{f}_1+F_1 H\bar{f}_3)\phi_2+(F_2 H \bar{f}_3+F_3 H\bar{f}_2)\phi_1\right]+ h.c.\,,\nonumber
\end{eqnarray}
where generically $F$ and $\bar f$ denote respectively the left--chiral and right-chiral components of SM fermions,
$H$ labels the Higgs associated with $\bar f$, and $\phi$ stands for flavon fields.
Further, we have let the VEVs of the $s$ flavons be absorbed in the Yukawas $y$, as
they do not alter the structure of the couplings.
From the Yukawa Lagrangian~\eqref{eq:Lyuk}, the Dirac mass matrices
for the charged fermions (namely, up and down quarks, and charged leptons)
generically take the form
\begin{equation}\label{diracmass}M_f^D=
\left(\begin{array}{ccc}
y_1^f \phi_1^f&y_2^f \phi_3^f&y_2^f \phi_2^f\\
y_2^f \phi_3^f&y_1^f \phi_2^f&y_2^f \phi_1^f\\
y_2^f \phi_2^f&y_2^f \phi_1^f&y_1^f \phi_3^f
\end{array}
\right)\,.
\end{equation}
Let us now make a phenomenological assumption on the flavon VEVs. Suppose the possibility of
a VEV alignment of the form $\langle \phi^f \rangle= v_{\phi}^f(0,r^f,1)$,
with $f=u,d,e$, for some real values $v_\phi^f$ and $r^f$.
This greatly simplifies the mass matrices to
\begin{equation}\label{mass:charged1}M_f^D=
\left(\begin{array}{ccc}
0&a^f&a^f r^f\\
a^f &b^f r^f&0\\
a^f r^f&0&b^f
\end{array}
\right)\,,
\end{equation}
where we define $a^f\equiv y_2^f v_{\phi}^f$ and $b^f\equiv y_1^f v_{\phi}^f$.
Using now the invariant traces and determinant of $M_f^D$ (we take a negative
$m_1^f$ to compensate the minus sign in the determinant),
\begin{eqnarray}
\label{eq:Minvariants}
\tr M^D_f &=& b^f\,(1+r^f) \hspace{27mm} \stackrel{!}{=} -m^f_1+m^f_2+m^f_3\,, \\
\tr(M^D_f)^2 &=& [2(a^f)^2+(b^f)^2][1+(r^f)^2]\, \stackrel{!}{=} \ (m_1^f)^2+(m_2^f)^2+(m_3^f)^2\,, \nonumber\\
\det M^D_f &=& -(a^f)^2 b^f\,[1+(r^f)^3] \hspace{11mm} \stackrel{!}{=} -m^f_1 m^f_2 m^f_3\,, \nonumber
\end{eqnarray}
it is straightforward to write down the Dirac mass matrices in terms of its eigenvalues, i.e.
the three (observable) fermion masses of type $f$, $m_i^f$.
Clearly, any solution to the invariants~\eqref{eq:Minvariants} provides the right masses
for quarks and charged leptons. If we take e.g. the hierarchical solution,
i.e. $r^f\ll1$ and $a^f\ll b^f$, the mass matrices take the form
\begin{equation}\label{mass:charged2}
M_f^D\approx
\left(\begin{array}{ccc}
0&\sqrt{m^f_1 m^f_2}& \frac{m^f_2-m^f_1}{m^f_3}\sqrt{m^f_1 m^f_2}\\
\sqrt{m^f_1 m^f_2}&m^f_2-m^f_1&0\\
\frac{m^f_2-m^f_1}{m^f_3}\sqrt{m^f_1 m^f_2}&0&m^f_3
\end{array}
\right),
\end{equation}
which corresponds to
\begin{equation}
\label{eq:hierarchysol}
r^f \approx (m_2^f-m_1^f) / m_3^f\,,\qquad (a^f)^2\approx m_1^f m_2^f\,,\qquad b^f\approx m_3^f\,.
\end{equation}
We notice that the hierarchical solution is compatible with the hierarchy of observed fermion masses.
In the down--quark sector, this structure gives the Gatto-Sartori-Tonin
formula for the Cabibbo angle, which is approximately the ratio
$(M_d^{D})_{12}/(M_d^{D})_{22}$,
\begin{equation}
\label{gst}
\lambda_C\approx \sqrt{\frac{m_d}{m_s}}\,,
\end{equation}
where we additionally used that $m_d/m_s\ll1$.\footnote{Eq.~\eqref{gst} has a small
correction of order $\sqrt{m_u/m_c}$ from the up--quark sector.}
The other two mixing angles are very small at leading order,
but could be generated if some of the vectorlike quarks mix with
the SM quarks, see for instance~\cite{Morisi:2013eca}.
For charged leptons, on the other hand, the same flavon VEV alignment must be imposed
because down--quarks and charged leptons share the same flavons. It follows that
the corresponding mass matrix is diagonalized by a rotation in the 1--2 entries with
the mixing angle of the order $\sqrt{m_e/m_\mu}$.
\begin{figure}[!b!]
\centering
\includegraphics[scale=0.5]{diagram1.pdf}
\includegraphics[scale=0.5]{diagram2.pdf}
\caption{Phenomenologically viable operators in the model presented that may alleviate the tension
observed by the predicted relation~\eqref{eq:massrelation}.}
\label{fig:operatorsvsbottom-tau}
\end{figure}
There is another consequence of the parallelism between the down--quarks and
charged leptons. Since $r^d = r^e$, it follows from eq.~\eqref{eq:hierarchysol}
that the following mass relation in our model is required
\begin{equation}
\label{eq:massrelation}
\frac{m_s-m_d}{m_b} \ \stackrel{!}{=} \ \frac{m_\mu-m_e}{m_\tau}.
\end{equation}
This relation does not match observations.
We find that some possibilities to amend eq.~\eqref{eq:massrelation} include either to abandon
the flavor structure in the soft-terms of the supersymmetry breaking sector
or that some (colored and uncolored) exotics acquire masses after the breakdown of $\Delta(54)$,
providing different suppression factors for down--quarks and charged leptons.
The latter can be achieved by allowed couplings as those represented in fig.~\ref{fig:operatorsvsbottom-tau},
which yield effective contributions to Yukawa couplings, such as
$$\frac{1}{m_vm_xm_z} Q H_d \bar d \langle N_i N_j N_k\rangle + \frac{1}{m_vm_xm_w} L H_d \bar e \langle N'_i N'_j N'_k \rangle\,,$$
where both $N_{i,j,k}$ and $N'_{i,j,k}$ denote some of the 128 flavons of table~\ref{tab:spectrum_model9}, and $m_{\chi}$ denotes
the effective mass of a given exotic field $\chi$. Realizing particularly that $m_z$ and $m_w$ differ in general and, moreover, that the
flavons in the couplings may be different, we find that the issue underlined by the constraint~\eqref{eq:massrelation} may be alleviated.
Unfortunately, even if this hurdle is tackled, we do not expect these effects to alter the smallness
of the remaining two quark mixing angles since that depends on the
hierarchical structure of the fermion masses.
\subsection{The neutrino sector}
For neutrinos, the major difference w.r.t. the other sectors is that, besides
the presence of Majorana mass terms, neutrinos build a conjugate $\Delta(54)$
triplet, $\rep3_{12}$. Therefore, renormalizable Yukawa couplings become possible.
As stated before, the neutrino masses arise from a type I see--saw according
to the second row of the superpotential~\eqref{eq:Wyuk}.
From there, we can read off the Yukawa Lagrangian for neutrinos:
\begin{eqnarray}
\label{eq:Lyuknu}
{\cal L}_{Y}^\nu&=&y_1^\nu \left[L_1 H_u\bar{\nu}_1+L_2 H_u\bar{\nu}_2+L_3 H_u\bar{\nu}_3\right] \\
&+&\lambda_1 \left[\bar{\nu}_1 \bar{\nu}_1 \bar\phi_1^\nu + \bar{\nu}_2 \bar{\nu}_2 \bar\phi_2^\nu + \bar{\nu}_3 \bar{\nu}_3\bar\phi_3^\nu\right] \nonumber\\
&+&\lambda_2\left[2\bar{\nu}_1 \bar{\nu}_2 \bar\phi_3^\nu + 2\bar{\nu}_1 \bar{\nu}_3\bar\phi_2^\nu + 2\bar{\nu}_2 \bar{\nu}_3\bar\phi_1^\nu\right]. \nonumber
\end{eqnarray}
Hence, the Dirac neutrino mass matrix is proportional
to the identity matrix, while RH neutrino masses are governed by a structure
similar to the one in eq.~\eqref{diracmass}, that is,
\begin{equation}M_{RH}=
\left(\begin{array}{ccc}
\lambda_1 \bar\phi_1^\nu&\lambda_2 \bar\phi_3^\nu&\lambda_2 \bar\phi_2^\nu\\
\lambda_2 \bar\phi_3^\nu&\lambda_1 \bar\phi_2^\nu&\lambda_2 \bar\phi_1^\nu\\
\lambda_2 \bar\phi_2^\nu&\lambda_2 \bar\phi_1^\nu&\lambda_1 \bar\phi_3^\nu
\end{array}
\right).
\end{equation}
We can now make a working assumption about the VEV of the neutrino flavon $\bar\phi^\nu$.
Considering the alignment $\vev{\bar\phi^\nu}=v_{\nu_3}\left(R_1,\delta,1\right)$,
the light neutrino mass matrix becomes
\begin{equation}
\label{numass}
M_\nu~=~\lambda \left(
\begin{array}{ccc}
\delta-R^2 R_1^2 & R(-1+R R_1 \delta) & R(-\delta^2+R R_1 )\\
R(-1+R R_1 \delta) & R_1-R^2\delta^2 & R(R \delta-R_1^2)\\
R(-\delta^2+R R_1 ) & R(R \delta-R_1^2) & R_1\delta-R^2
\end{array}
\right),
\end{equation}
where we used the definitions
\begin{equation}
R= {\lambda_2}/{\lambda_1}\,,\qquad
\lambda={y_1^2 \langle H_u\rangle^2}\Big/\left[\lambda_1v_{\nu_3}\left(R_1\delta+2R^3R_1\delta-R^2(1+R_1^3+\delta^3)\right)\right]\,.
\end{equation}
After performing a scan of our parameters, restricting the values of the computed
$\Delta m^2_{12}$, $\Delta m^2_{13}$ and neutrino mixing angles to lie within
the $3\sigma$ region of the global fits~\cite{Forero:2014bxa}, we find that
the mass matrix in eq.~\eqref{numass} is compatible only with a normal hierarchy
of neutrino masses, i.e. an inverted hierarchy is disfavored, coinciding with
recent preliminary results from the T2K collaboration~\cite{T2K}.
\begin{figure}[!t!]
\centering
\includegraphics{correlationatmreact.pdf}
\caption{Correlation between the atmospheric and reactor mixing angles for normal mass ordering in
a string--inspired $\Delta(54)$ flavor model. The correlation (blue) points in the upper--left part of
the plot result from a scan of our parameters $\lambda,\delta,R,R_1$, imposing consistency within
$3\sigma$ with measured values of $\Delta m^2_{12}$, $\Delta m^2_{13}$ and $\theta_{12}$.
The dark/light/lighter gray areas correspond to $1\sigma$/$2\sigma$/$3\sigma$ experimental precision
around the best fit value (denoted by the star) for the neutrino mixing angles~\cite{Forero:2014bxa}.
}
\label{fig:atmoreactor}
\end{figure}
Furthermore, we observe that our model leads to a correlation between the atmospheric
and the reactor mixing angles in normal ordering, as displayed by the blue
region in fig.~\ref{fig:atmoreactor}. Comparing with the precision intervals,
we see that the atmospheric mixing angle lies in the second octant, approximately
between $51.3$ and $53.1$ degrees, while the reactor mixing angle has values between
$7.8$ and $8.9$ degrees, in agreement with the oscillation global fits within $3\sigma$.
These values are crucial for the model since a better measurement of the neutrino
mixing angles could falsify it.
A final result from our parameter scan is that the lightest neutrino mass, $m_{\nu_1}$,
takes values in the region between $6$~meV and $6.8$~meV, and the sum of the light
neutrino masses, $\sum m_{\nu}$, lies in the interval between $65$~meV and $70$~meV,
in consistency with data.
\section{Final remarks}
\label{sec:conclusions}
Flavor symmetries arise naturally in string compactifications, which provide
a promising ultraviolet completion of usual bottom--up setups. Particularly,
we have shown that $\Delta(54)$, as a flavor symmetry, appears most naturally
in semi--realistic \Z3\x\Z3 heterotic orbifold compactifications. We have
identified almost 700 models with that flavor symmetry and other promising
particle--physics features, such as SM gauge group and three generations of
matter fields. By their nature, these constructions reduce the arbitrariness of
low--energy models by constraining the fields and their (flavor and gauge)
transformation properties and thereby providing useful guidelines
to inspect flavor phenomenology.
To test the viability of $\Delta(54)$ flavor scenarios arising from strings, we
have studied the phenomenology of one simple string model from our classification,
whose properties may differ from the other identified models.
In this model, SM fermion fields transform as triplets of the flavor symmetry
while the Higgs fields do not transform. As a result of the flavor quantum numbers,
the quarks and charged leptons acquire masses through dimension--6 operators,
and the Dirac neutrino masses as well as the RH Majorana neutrino masses are generated
at renormalizable level.
Furthermore, we observe that choosing some special flavon--VEV alignments results
in the following flavor phenomenology features:
\begin{itemize}
\item correct masses for quarks and charged leptons;
\item proper Gatto-Sartori-Tonin relation in the quark sector (although the other two mixing angles are very small);
\item a mass relation between the down--quark sector and the charged leptonic sector (see eq.~\eqref{eq:massrelation});
\item compatibility (only) with normal hierarchy of neutrino masses;
\item smallest neutrino mass of order $6-7$ meV; and
\item PMNS matrix compatible with current constraints (atmospheric and reactor
mixing angles are in the $3\sigma$ region of the global best fit),
with the atmospheric mixing angle greater than $45$ degrees.
\end{itemize}
Interestingly, an inverted hierarchy being disfavored as well as the atmospheric
mixing angle lying in the second octant, are features compatible with recent
preliminary findings of the T2K collaboration~\cite{T2K}.
This outcome lets us assert that \Z3\x\Z3 heterotic orbifolds and $\Delta(54)$ as a flavor
symmetry provide a fertile playground for useful phenomenology which should be further investigated.
The particular model we have studied here was chosen due to its neat simplicity:
it has only three SM generations, the extra gauge sector includes only a hidden \SU2 and Abelian symmetries,
and all SM fields build $\Delta(54)$ triplets. These properties are only shared by three more models in the
set of promising \Z3\x\Z3 compactifications. Other models include additional (exotic) vectorlike pairs of quarks and
leptons, larger Abelian and non--Abelian hidden gauge symmetries, and some SM fields may build only trivial
representations of $\Delta(54)$.
This does not imply that other models are more or less promising, but their analysis is somewhat more involved
and shall be the purpose of future studies.
Despite these encouraging features, there are still some challenges to overcome. First,
in heterotic orbifolds it is challenging to obtain the VEV alignments chosen in section~\ref{sec:Delta54pheno}
because VEVs must be settled by a moduli stabilization mechanism that is not fully understood.
Secondly, we found that two of the quark mixing angles in our model are too small and the mass
relation eq.~\eqref{eq:massrelation} is incorrect. To attempt to alleviate these issues,
one should study in detail the soft--terms and other corrections in this kind of models.
Another potential hurdle is the absence of a symmetry that forbids rapid proton decay.
However, it is conceivable that such symmetry does appear as one of the extra \Z3 symmetries
of another model where matter fields have the correct charges. Finally, as in most flavor
models, flavor--changing neutral currents pose a challenge that must and shall be studied
elsewhere in the context of our proposal.
\section*{Acknowledgments}
We would like to thank P.K.S. Vaudrevange for useful discussions.
B.~C-P. would like to thank the CLAF-ICyTDF for support during her posdoctoral fellowship.
E.~P. is partly supported by PAPIIT IA101516 and PAPIIT IN111115.
S.~R-S. would like to thank the ICTP for the kind hospitality and support received
through its Junior Associateship Scheme during the realization of this work.
\begin{appendix}
\section{$\Delta(54)$ tensor product for triplet representations}
\label{sec:appendix}
In this appendix, we provide the features of $\Delta(54)$ that are relevant for our proposal, following the
notation of ref.~\cite{Ishimori:2012zz}. The $\Delta(54)$ symmetry group has two one--dimensional, four
two--dimensional and four three--dimensional irreducible representations. These representations are denoted
as $\rep1_0$ (invariant under the group), $\rep1_{1}$, $\rep2_1$, $\rep2_2$, $\rep2_3$,
$\rep2_4$, $\rep3_{11}$, $\rep3_{12}$, $\rep3_{21}$ and $\rep3_{22}$.
Due to the matter content of our model, the only tensor products that are relevant in this work are those
among the three--dimensional representations $\rep3_{11}$ and $\rep3_{12}$, which are obtained as
\begin{eqnarray}
\begin{pmatrix}
x_1 \\ x_2 \\ x_3 \\
\end{pmatrix}_{{\bf 3}_{11}} \otimes
\begin{pmatrix}
y_1 \\ y_2 \\ y_3 \\
\end{pmatrix}_{{\bf 3}_{11}}
=\begin{pmatrix}
x_1y_1 \\ x_2y_2 \\ x_3y_3 \\
\end{pmatrix}_{{\bf 3}_{12}} \oplus
\begin{pmatrix}
x_2y_3+x_3y_2 \\ x_3y_1+x_1y_3 \\ x_1y_2+x_2y_1 \\
\end{pmatrix}_{{\bf 3}_{12}}
\oplus
\begin{pmatrix}
x_2y_3-x_3y_2 \\ x_3y_1-x_1y_3 \\ x_1y_2-x_2y_1 \\
\end{pmatrix}_{{\bf 3}_{22}} ,
\end{eqnarray}
\begin{eqnarray}
\begin{pmatrix}
x_1 \\ x_2 \\ x_3 \\
\end{pmatrix}_{{\bf 3}_{12}} \otimes
\begin{pmatrix}
y_1 \\ y_2 \\ y_3 \\
\end{pmatrix}_{{\bf 3}_{12}}
=\begin{pmatrix}
x_1y_1 \\ x_2y_2 \\ x_3y_3 \\
\end{pmatrix}_{{\bf 3}_{11}} \oplus
\begin{pmatrix}
x_2y_3+x_3y_2 \\ x_3y_1+x_1y_3 \\ x_1y_2+x_2y_1 \\
\end{pmatrix}_{{\bf 3}_{11}}
\oplus
\begin{pmatrix}
x_2y_3-x_3y_2 \\ x_3y_1-x_1y_3 \\ x_1y_2-x_2y_1 \\
\end{pmatrix}_{{\bf 3}_{21}} ,
\end{eqnarray}
and finally
\begin{eqnarray}
\begin{pmatrix}
x_1 \\ x_2 \\ x_3 \\
\end{pmatrix}_{{\bf 3}_{11}} \otimes
\begin{pmatrix}
y_1 \\ y_2 \\ y_3 \\
\end{pmatrix}_{{\bf 3}_{12}}
&=&
\begin{pmatrix}
x_1y_1 + x_2y_2 + x_3y_3
\end{pmatrix}_{{\bf 1}_{0}} \oplus
\begin{pmatrix}
x_1y_1+\omega ^2x_2y_2+\omega x_3y_3 \\ \omega x_1y_1+\omega
^2x_2y_2+x_3y_3
\end{pmatrix}_{{\bf 2}_1} \nonumber \\
& \oplus &
\begin{pmatrix}
x_1y_2+\omega ^2x_2y_3+\omega x_3y_1 \\ \omega x_1y_3+\omega
^2x_2y_1+x_3y_2
\end{pmatrix}_{{\bf 2}_2} \oplus
\begin{pmatrix}
x_1y_3+\omega ^2x_2y_1+\omega x_3y_2 \\ \omega x_1y_2+\omega
^2x_2y_3+x_3y_1
\end{pmatrix}_{{\bf 2}_3} \nonumber \\
& \oplus &
\begin{pmatrix}
x_1y_3+x_2y_1+x_3y_2 \\ x_1y_2+x_2y_3+x_3y_1
\end{pmatrix}_{{\bf 2}_4},
\end{eqnarray}
where $\omega=e^{2\pi\mathrm{i}/3}$. It follows that the only products of $\Delta(54)$ triplets
up to trilinear order that yield invariant combinations are $\rep3_{11}\otimes\rep3_{12}$,
$\rep3_{11}\otimes\rep3_{11}\otimes\rep3_{11}$ and $\rep3_{12}\otimes\rep3_{12}\otimes\rep3_{12}$.
The latter two products lead to two invariant singlets $\rep1_0$ each.
\end{appendix}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\frenchspacing
\newcommand{}{}
\let=\ttfamily
\renewcommand{\ttfamily}{ \hyphenchar\font=`\-}
|
\section{Introduction}
In astrophysics, the calculation of interstellar extinction is a
complex and recurring problem. For many objects, such as those buried
in star-forming regions, unreddening the photometries is a difficult
and demanding task. In the case of a star, the calculation of
interstellar extinction requires a detailed knowledge of its
luminosity class, spectral type, and intrinsic colors. That is a lot
of parameters, not always available, whose robustness is often
uncertain.This leads to the accumulation of errors, and
makes it nearly impossible to attempt any massive statistical
analysis.
We recently introduced the concept of
pseudomagnitude for the calculation of the apparent size of stars,
thus avoiding to deal with the problem of visual extinction
\citep{Chelli14,Chelli16}. This has allowed us to compile a catalogue
of $453\,000$ angular diameters, with an accuracy of the order of $1\%$
($2\%$ systematic). Pseudomagnitudes are linear combinations of
magnitudes constructed in such a way as to eliminate interstellar
extinction. They are purely observational quantities that are
unaffected by reddening effects, and can be applied to any type of
object. As in the case of magnitudes, pseudomagnitudes are distance
indicators, and absolute pseudomagnitudes, measured at a distance of
$10$\,pc, are luminosity indicators.
Knowledge of the pseudomagnitudes and absolute pseudomagnitudes of
stars allows their distance to be estimated.
In the present study, we use the parallax measurements of
Hipparcos~\citep{ESAHIP,2007A&A...474..653V} to calculate
the mean absolute pseudomagnitude of field
dwarf stars, as a
function of their spectral type. As an example, we use this technique
to determine the centroid of the distance distribution of 360 stars in
the Pleiades cluster.
In section~\ref{sec:pseudomagnitudes}, we explain the concept of
pseudomagnitudes. In section~\ref{sec:absolutepseudomagnitudes}, we use
distance filtered parallax measurements to calculate the mean absolute
pseudomagnitudes (V,J), (V,H) and (V,Ks) of dwarf stars, and the centroid of the distance
distribution of our Pleiades stars is calculated and discussed in
section~\ref{sec:pleiades}.
\section{Pseudomagnitudes}
\label{sec:pseudomagnitudes}
We define the pseudomagnitude $pm_{\{i,j\}}$ of an astrophysical
object as follows:
\begin{equation}
pm_{\{i,j\}} = \frac{c_im_j-c_jm_i}{c_i-c_j}
\label{eq:pseudomagnitude}
\end{equation}
where $m_i$ and $m_j$ are the magnitudes measured in the photometric
bands $i$ and $j$, $c_i$ (resp. $c_j$) is the ratio of the interstellar
extinction coefficients $R_i$ and $R_v$ between band
$i$ and the visible band. We note that when one of the coefficients $c_i$ or $c_j$
tends to zero, the pseudomagnitude tends to the magnitude $m_i$ or
$m_j$. The pseudomagnitude is by construction a reddening free
distance indicator. It can be written as:
\begin{equation}
pm_{\{i,j\}} = \frac{c_iM_j-c_jM_i}{c_i-c_j}+DM
\end{equation}
where $M_i$ and $M_j$ are absolute magnitudes and $DM$ is the distance
modulus. At this stage, we define the absolute pseudomagnitude
$PM_{\{i,j\}}$ as:
\begin{equation}
PM_{\{i,j\}} = \frac{c_iM_j-c_jM_i}{c_i-c_j}=pm_{\{i,j\}}-DM
\label{eq:absolute_pseudomagnitude}
\end{equation}
The absolute pseudomagnitude is a reddening free luminosity luminosity
indicator that can be computed very easily. This
requires the knowledge of two magnitudes and a distance. On the other
hand, once the mean absolute pseudomagnitude has been calculated for a
group of stars sharing the same physical properties, the distance
modulus of a star from the same group can be estimated with the
knowledge of just two magnitudes.
\section{Absolute pseudomagnitudes of dwarf stars}
\label{sec:absolutepseudomagnitudes}
For our calculations,
we use Eqs.\ref{eq:pseudomagnitude} and~\ref{eq:absolute_pseudomagnitude},
with the second reduction of Hipparcos parallaxes
\citep{2007A&A...474..653V}, the spectral type and the magnitude pairs
(V,J), (V,H) and (V,Ks) provided by SIMBAD. We adopt the interstellar
extinction coefficients determined by \cite{Fitzpatrick99},
thus leading to the following expressions for the pseudomagnitudes:
\begin{eqnarray}
pm_{\{V,J\}} & = & 1.389 \times m_{J}-0.389 \times m_V \nonumber \\
pm_{\{V,H\}} & = & 1.205 \times m_{H}-0.205 \times m_V \nonumber \\
pm_{\{V,Ks\}} & = & 1.136 \times m_{Ks}-0.136 \times m_V
\label{eq:pmvk}
\end{eqnarray}
\subsection{Hipparcos data}
A priori, the absolute pseudomagnitude of a group of stars with the
same spectral type and luminosity class should be constant as a
function of distance. Figure~\ref{fig:1}a
plots the pseudomagnitude (V,Ks) of Hipparcos class III and V stars
with a spectral type K0 (3747 objects), as a function of their
distance modulus. Figure~\ref{fig:1}b
shows the absolute pseudomagnitude (V,Ks), with the dwarfs lying at the
top and the giants lying at the bottom. For the
same class of stars it is firstly constant, to within the limits
resulting from noise, but beyond a certain distance it then
appears to decrease.
It is a
mere artifact, due to the fact that below 10\% noise, the inverse of the parallax
begins to be numerically biased. In this example, 75\% of the
dwarfs and only 26\% of the giants have a parallax noise smaller than
10\%.
\subsection{Practical absolute pseudomagnitude calculation}
In order to calculate the mean absolute pseudomagnitudes of
dwarf stars, we proceed as follows: a) we consider all of the stars in
the Hipparcos catalogue having the same spectral type, with or without
selecting their luminosity class, depending on the possible degree of
confusion; b) we place a limit on the distance of the sample in order
to minimize the influence of the numerical bias\footnote{For example, in the case
of the K0 stars of figure~\ref{fig:1}\,b, this limit would be around
$DM=4$
for dwarves and $DM=7$
for giants.}; c) since we do not control the
astrophysical biases (see below), we assume that all of the objects
are statistically equivalent, and adjust the fit of the absolute
pseudomagnitude distribution to one, or even ---in some cases--- to two
Gaussian functions.
This is a difficult operation because the absolute
pseudomagnitude distribution is not always strictly Gaussian.
In practice, stars from the same luminosity class and with the same
spectral type often have stratified luminosities as a function of
their distance. This phenomenon confirms what was already known,
i.e. that for any given spectral type and class of luminosity, there
are hidden sub-classes of stars with distinct physical
properties. Although the absolute pseudomagnitudes would permit a
detailed investigation of these physical properties, for the time
being we do not have sufficient statistical information to implement
such an analysis. This will become possible when the measurements
provided by GAIA~\citep{GAIA} become available.
Manual calculations were made for each spectral type, and were
repeated several times on various samples of stars. These were based
on the analysis of the pseudomagnitudes of approximately $6000$
dwarf stars, distributed over 56 spectral
sub-types. It corresponds to about 25\% of the Hipparcos stars
identified as dwarfs. 90\% of the selected data have a parallax with
less than 10\% noise, 98\% less than 20\%.
Figure~\ref{fig:2} shows the mean absolute pseudomagnitudes
(V,Ks) of these dwarf stars as a function of their spectral types,
ranging from O9 to M4.
\begin{figure}
\centering
\includegraphics[angle=90,width=8cm]{paper2_fig1.eps}
\caption{a) (V,Ks) pseudomagnitudes of the 3747
Hipparcos K0 class III and V stars as a function of their distance
modulus; b) Absolute (V,Ks) pseudomagnitudes of the same
stars, with the dwarfs lying at the top and the giants lying
at the bottom of this figure. The decrease of
the pseudomagnitudes beyond a certain distance is an artefact
due to numerical bias at low signal to noise ratio (see
text). }
\label{fig:1}
\end{figure}
\begin{figure}
\centering
\includegraphics[angle=90,width=8cm]{paper2_fig2}
\caption{ a) Mean (V,Ks) absolute pseudomagnitudes of field dwarf stars
as a function of
spectral type, b) Open circles: (V,Ks)
pseudomagnitudes of 280 Pleiades
stars located at less than 0.84 mag (3 times the Gaussian dispersion of
figure 3) from the Pleiades barycentric
distance modulus. Hipparcos stars are identified
with larger filled circles, superimposed our main sequence
model shifted at the Pleiades distance.}
\label{fig:2}
\end{figure}
The median statistical error on the mean absolute
pseudomagnitudes is equal to $0.03$ magnitudes, which corresponds to
an error of $1.5\%$ in terms of distance. For a given group of stars,
the observed dispersions can be
accounted for by the natural width of the group, which is increased
by the influence of multiplicity, errors of magnitude, distance and
classification. To a lesser extent, they also reflect the star's age
or metallicity. We estimate, to within a factor of 2, that the
systematic error on a correctly characterised single dwarf star is of
the order of $0.05$ magnitude.
Although pseudomagnitudes have many potential
applications, the most immediate of these is the determination of
the mean distance of a spatially concentrated group of stars, as
for example in the case of stellar clusters and
galaxies. In the following section we calculate the centroid of the
distance distribution of $360$ stars in the Pleiades cluster, and
whenever possible compare our results with those obtained by other
authors.
\section{Pseudomagnitude distance of the Pleiades}
\label{sec:pleiades}
The Pleiades is one of the most commonly observed young open clusters,
and the properties of its stars provide a \textit{de facto} definition
of the properties of main sequence stars at age zero. Numerous studies
continue to be published regarding the census of this cluster's coeval stars,
and the highest possible accuracy is needed in their distance
determinations in order to test the models of stellar structure and
evolution. The pseudomagnitude method can be applied to all of the
stars in this cluster, for which the spectral type and at least one
pair of magnitudes is known. It is perfectly adapted to the
calculation of the cluster's mean distance, and could even be
sufficient for the accurate evaluation of the individual distances of
these stars (see section~\ref{vlbi}).
\begin{table}
\tiny
\begin{tabular}{llll}
\hline\T
Refs & Method & N & DM / distance (pc) \B \\
\hline \hline \T
1 & Hipparcos first release & 54 & 5.32 (0.05) / 115.9 (2.7)\\
2 & Photometry & 55 & 5.60 (0.04) / 131.8 (2.4) \\
3 & Moving cluster & 65 & 5.58 (0.18) / 130.6 (11.) \\
4 & Ground parallax & 9 & 5.58 (0.12) / 130.6 (7.0) \\
5 & Photometry & 30 & 5.61 (0.03) / 132.4 (1.8) \\
6 & Hipparcos (Makarov) & 54 & 5.55 (0.06) / 129.0 (3.3) \\
7 & Binary & 1 & 5.60 (0.03) / 131.8 (1.8) \\
8 & Binary & 1 & 5.65 (0.03) / 134.9 (1.9) \\
9 & Binary & 1 & 5.60 (0.07) / 131.8 (4.2) \\
10 & HST parallax & 10 & 5.66 (0.06) / 135.5 (3.7) \\
11 & HST parallax & 3 & 5.65 (0.05) / 134.9 (3.1) \\
12 & Binary & 1 & 5.72 (0.05) / 139.3 (3.2) \\
13 & Hipparcos (van Leeuwen) & 54 & 5.40 (0.03) / 120.2 (1.7) \\
14 & VLBI & 5 & 5.67 (0.02) / 136.2 (1.2) \\
15 & Binary & 1 & 5.61 (0.08) / 132.4 (4.9) \\
16 & Photometry & 120 & 5.62 (0.03) / 132.7 (1.8) \\
This work & Pseudomagnitude & 360 & 5.715 (0.018) / 139.0 (1.2) \B\\
\hline
\end{tabular}
\caption{\tiny Measured distances of Pleiades stars, errors are
between parenthesis. 1: \cite{vanLeeuwen98},
2: \cite{Pinsonneault1998}, 3: \cite{Narayanan99}, 4: \cite{Gatewood00},
5: \cite{Stello2001}, 6: \cite{Makarov02}, 7: \cite{Munari04},
8: \cite{2004Natur.427..326P}, 9: \cite{Zwahlen04}, 10: \cite{Johns-Krull05},
11: \cite{Soderblom05}, 12: \cite{Southworth05}, 13: \cite{vanLeeuwen09},
14: \cite{Melis14}, 15: \cite{David16}, 16: \cite{Kim2016}; N:
target number; DM: distance modulus.
}
\label{pleiades:distances}
\end{table}
\subsection{On the Pleiades distance controversy }
Whereas an history of distance estimations of the Pleiades cluster can
be found in \cite{An07} and \cite{Melis14},
Table~\ref{pleiades:distances} provides a summary of the measurements
published in the last $20$
years. Various methods have been used. Excluding Hipparcos, the other
direct distance measurements (ground and spaceborne parallaxes,
binaries, VLBI) have relied on the analysis of a total of $\approx30$
stars, and position the Pleiades at a distance between 130 and
139\,pc. The indirect photometric methods were applied on a total of
$\approx 120$ stars and have positioned the cluster at a distance of
132\,pc. In contrast, the mean distance of $54$ Pleiades stars
of spectral types B, A and F by Hipparcos~\citep{vanLeeuwen09} lead to
the controversial distance of $120.2\pm1.7$\,pc, which is indeed
markedly lower (by 10\%) than all other measurements.
It should be recalled that the Pleiades cluster probably contains more
than one thousand stars. When projected onto the sky, it extends over
a distance of the order of 10 to 20\,pc, and it would be reasonable to
assume that the Pleiades has a similar size along its line of sight
when viewed from Earth. Under these conditions, the distances measured
on a few, or even a few tens of objects, with an accuracy much better
than the cluster's expected size, are representative of these objects
distances only. In view of the size of this cluster, it could well be
possible to find star concentrations at distances of the order of
15\,pc from one another. The controversy does not have as much to do
with the so-called distance of the Pleiades cluster\footnote{We
observe that, given the currently achievable precision on an
individual star distance and the size of the cluster compared to its
distance ($\approx10\%$),
the concept of ``Pleiades distance'' is bound to loose its intended
meaning.}, as with the mean distance of the 54-odd stars used in the
Hipparcos estimate.
It is difficult to
compare various distance measurements, as they are based on generally
small and generally disjoint samples of stars. The Hipparcos sample
was not used by other independent distance estimations, it was only
reused in new attempts to refine the Hipparcos reduction, first by
\cite{Makarov02} which led to a distance of $129.0\pm3.3$\,pc,
and then by \cite{vanLeeuwen09}, who determined a value of only
$120.2\pm1.7$\,pc.
We note that in view of their uncertainties, these two distance
estimations are only marginally ($2.4\,\sigma$) different.
\begin{figure}
\centering
\includegraphics[angle=90,width=8cm]{paper2_fig3}
\caption{ Distance moduli distribution for 360 Pleiades stars,
fitted by a Gaussian distribution plus second degree
polynomial. The Gaussian dispersion (0.28 mag) is dominated by
spectral classifications errors.}
\label{fig:3}
\end{figure}
Our absolute pseudomagnitude calibration allows us to evaluate the
distance of any sample of stars. In the following section, we
calculate the distance of $360$ Pleiades stars, as well as that of the
Hipparcos sample.
\subsection{Distance of 360 Pleiades stars}
In this section, we assume that Pleiades stars have, at the same
spectral types, the same pseudomagnitudes (V,J), (V,H) and (V,Ks) that
field dwarfs. Significant differences
occur for cool stars somewhere within the M spectral class. Our sample
of Pleiades stars was obtained from a
total of $3721$ stars associated with the ``M45'' identifier in the Simbad
database. After filtering (multiplicity, variability, etc.), a total
of $512$ stars remained, of which only $360$
had the required information for the calculation of their distance. As
the Pleiades cluster is very young, in order to increase the size of
our sample, we assumed all of the selected stars to be of luminosity
class V. As the pseudomagnitude is sensitive to the luminosity class,
any non-dwarf star will contribute to the broadening of the distance
distribution, or will get a distance very different to that of the
cluster and will thus be excluded from the analysis. We did not try to
perform filtering for membership, non members will
form a diffuse background that is taken into account in our
statistical modeling.
The adopted distance modulus of each object is the average of the
distance moduli computed from the photometric pairs (V,J),
(V,H) and (V,Ks), and its error is the dispersion of the three
estimates.
The (V,Ks) pseudomagnitudes
of our sample, outliers excluded, are shown in Figure~\ref{fig:2}b as
a function of the spectral type. It is not a classical color-magnitude
diagram. The observed dispersions per spectral type, 0.2 to 0.4 magnitude,
are not imposed by the physics of the cluster but by spectral
classifications errors, which is probably the limiting noise of our
present approach. We fit the resulting
distance modulus distribution by a Gaussian plus a second degree
polynomial, see Figure~\ref{fig:3}. The centre of the Gaussian
function provides the barycentre of the distance moduli of the $360$
stars studied, i.e. $5.715\pm0.018$, which corresponds to a distance
of $139.0\pm1.2$\,pc.
Although this comparison is somewhat risky, in view of the small
samples used previously, our distance calculation is globally in
agreement with most estimations, but tends to position the cluster at
the high end of measured ``distances''.
What of the stars measured by Hipparcos? We have all of the
information needed to characterise 44 of the 54 stars given in the
list of \cite{Makarov02}. The distribution of their distance moduli
exhibits two maxima, at approximately $5.4$ and $5.7$, a possible
indication of sub-clustering. A gaussian fit of this distribution
leads to a mean distance modulus of $5.66\pm0.06$, i.e. a distance of
$135.5\pm3.7$\,pc, respectively $1.3\sigma$ and $3.8\sigma$
above \cite{Makarov02} and \cite{vanLeeuwen09} estimates. Our result
tends to confirm that on average Hipparcos distances of these stars are
underestimated. Soon we will have the answer on who is right or who is
wrong. But the answer probably will not be as simple as yes or no.
However, the baby should not be thrown out with the bathwater, since
all of our distance moduli were obtained using absolute
pseudomagnitudes derived from correctly distance-filtered
Hipparcos parallax measurements. The fact that we obtain a barycentric
distance that is compatible (and probably more accurate in terms of
defining the cluster's centroid, as a consequence of the much greater
sample size) with distances measured from the ground, together
with the fact that we are able to apparently correct the same
controversial Hipparcos measurements, indicates that Hipparcos
parallaxes at large are robust.
\subsection{Distance to the VLBI stars}
\label{vlbi}
Among recent distance measurements, those of \cite{Melis14} determined
by VLBI are the most accurate. As they make it possible to test the
robustness of our pseudomagnitude estimations, we calculate the
distance of $6$ of the $10$ stars scheduled for VLBI observation by
\cite{Melis13} (the 4 others are either not single dwarfs or lacking
spectral type information). Table~\ref{tab:melis_distances} summarises
our predicted distances. For the two stars in common with
\cite{Melis14}, the agreement between VLBI and pseudomagnitude
distances is remarquable, with relative differences of 1\% ($0.5
\sigma$) and 4\% ($1.6\sigma$).
\section{Conclusion}
Pseudomagnitudes are remarkable distance indicators, since they are
free of interstellar reddening effects. We have calculated the mean
absolute pseudomagnitudes of field dwarfs from O9 to M4, based on the
Hipparcos parallax measurements of approximately 6000 stars, allowing
us to estimate the distance of 360 Pleiades stars. We position the
centroid of these stars at $139.0\pm1.2$\,pc, and we confirm that
the Pleiades stellar distances measured by Hipparcos are on average
underestimated by 10\%.
\begin{table}
\tiny
\begin{tabular}{llll}
\hline\T
HII & SpT & PMD (pc) & VLBI distance (pc) $^{(1)}$ \B\\
\hline \hline \T
75&G7 &136.2 (3.6)& \\
253&G1 &143.7 (2.1)& \\
625&G5 &137.0 (2.4)&138.4 (1.1) \\
1136&G7 &141.0 (3.3)&135.5 (0.6) \\
1883&K2 &139.0 (1.4)& \\
2244&K2 &145.1 (2.1)&\B\\
\hline\B
\end{tabular}
\caption{\tiny Pseusomagnitude distance (PMD) of 6 Pleiades stars of
the \cite{Melis13} list. (1) \cite{Melis14}}
\label{tab:melis_distances}
\end{table}
ESA's recently launched GAIA mission will make it possible to
accurately determine the fine structure of absolute pseudomagnitudes,
their natural width, and the influence of various parameters such as
age and metallicity. It will be possible to calibrate these very
accurately, in several different optical bands. But already, our initial
results obtained with the Pleiades cluster, together with their
comparison with VLBI measurements, are very encouraging. This
technique is purely observational, direct and simple to implement,
since it needs the knowledge of only the spectral type, two magnitudes
and the corresponding absolute pseudomagnitude.
\begin{acknowledgements}
This research has made use of NASA's Astrophysics Data System, of the SIMBAD database
\citep{2000A&AS..143....9W}, and of the VizieR catalog access
tool \citep{2000A&AS..143...23O}, CDS, Strasbourg, France. We used the
TOPCAT tool\footnote{available at
\url{http://www.starlink.ac.uk/topcat/}}\citep{2005ASPC..347...29T}
to easily manipulate the star databases used.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{introduction}\label{sec:intro}
A well-known feature of the solar magnetic field is the observation of
filament channels at photospheric polarity inversion lines
(PILs). These magnetic structures, situated in the upper chromosphere and lower
corona, underlie and support the cool plasma that comprises
prominences and filaments \citep{Martin98,Gaizauskas00}. Filament
channels are regions of highly sheared magnetic field, containing large amounts
of free energy that ultimately is converted into kinetic and
thermal energy of the plasma, as well as nonthermal particle energies
when filament channels erupt and drive coronal mass ejections (CMEs). The
shear inherent in the filament channels is a form of magnetic
helicity, and filament channels are known as dextral if they have
negative helicity and sinistral if they have positive
helicity. Observations indicate that dextral (sinistral) filament
channels dominate in the northern (southern) hemisphere
\citep[e.g.][]{Martin92, Rust94b, Zirker97, Pevtsov03}. This
hemispheric helicity rule has also been observed in quiet-Sun magnetic
fields \citep{Pevtsov01b}, sigmoids \citep{Rust96}, active-region
magnetic fields \citep{Seehafer90}, coronal mass ejections (CMEs), and
sunspot whorls \citep{Pevtsov14}. The strength of the preference
ranges from about 55\% in active-region filaments \citep{Martin94}, to
over 80\% in quiescent filaments \citep{Pevtsov03} and does not seem
to change with solar cycle \citep{Hale27, Martin94,Hagino02}.\par
A second, seemingly unrelated, feature of the solar magnetic field is
the observation of loops in the closed-field corona that
appear to be near their minimum energy, current-free state. These
loops have been observed at high resolution in \emph{Transition Region
and Coronal Explorer (TRACE)} XUV and X-ray images, such as the one
in Figure \ref{fig:obs}, where they are seen to be very smooth and
laminar, with little to no tangling \citep{Schrijver99}. In other
words, there appears to be very little magnetic helicity associated
with the topology of these loops. The picture of the corona that
emerges, therefore, is one in which magnetic helicity manifests itself
at specific locations, namely above PILs, while leaving the rest of
the corona generally smooth and quasi-potential.\par
Since the corona has very high Lundquist number, its magnetic helicity
is believed to originate by injection from the photosphere, either by
flux emergence or footpoint motions after emergence, and to be lost to the
heliosphere by flux opening in CMEs and streamer blowouts.
\emph{Solar Dynamics Observatory (SDO)} measurements of helicity
injection into the coronal field indicate that shearing and twisting
motions by the photosphere dominate the flux emergence \citep{Liu12},
meaning that the helicity budget of the corona is primarily due to the
jostling of existing flux at the surface, rather than due to new flux
emerging into the corona. Observations of photospheric convection show
that these footpoint motions are highly complex \citep{Schmieder14},
with convective cells appearing randomly throughout the photosphere,
and occur over a broad range of scales, of order minutes for granules and
days for supergranules \citep{Hirzberger08}. From the standpoint of
helicity injection, however, the important flows are those that
twist the field. Compression of the field caused by converging flows
is not expected to impart any helicity into the field, so these
flows cannot be responsible for the shear observed in filament
channels. The flows that twist up the field, in contrast, do inject a
net helicity into the corona. Therefore, we conclude that it is
sufficient to model the helicity injection into the corona with simple
twisting motions as has been done by many authors
\citep[e.g.][]{WilmotSmith10, Rappazzo13}. Such flows have routinely
been observed in helioseismic measurements
\citep{Duvall00,Gizon03,Komm07,Seligman14}. Vortical flows on the
scale of granules \citep[e.g.][]{Bonet08, Bonet10, VD11, VD15} and
supergranules \citep{Brandt88,Attie09} have also been observed.\par
These considerations make it very challenging to understand the
simultaneous presence in the corona of both filament channels and
coronal loops. Magnetic stress is, apparently, injected
throughout the solar photosphere, yet is almost nowhere to be found in
the corona, except in filament channels. \citet{Antiochos13}
presented a new model for the formation of filament channels, magnetic
helicity condensation, based on the well-known inverse cascade of
magnetic helicity in turbulent systems. In the helicity condensation
model, photospheric convection imparts helicity into the coronal field,
and this helicity is then transported throughout the corona by
magnetic reconnection, which is well-known to conserve helicity
\citep{Woltjer58,Taylor74,Taylor86,Berger84b}.
Surface convection imparts the same sense
of twist to adjacent flux tubes, which are then able to undergo
component reconnection at their contact point. This component
reconnection produces a single flux tube with an axial flux equal to
the sum of the two original axial fluxes, but encircled by the same
twist field present on each of the two original flux tubes. In this
way the helicity, in the form of twist, inverse-cascades to larger and
larger scales. The PIL forms a natural boundary of the flux system, so
that when the twist reaches this boundary, it cannot proceed further,
since all of the flux has already reconnected. The end result of this
process is a mostly axial (untwisted) internal field, and a highly
sheared (twisted) field at the PIL, precisely what is observed as a
filament channel. At the same time, the untwisted internal field
corresponds to the laminar coronal loops. In this way, the helicity
condensation model provides a natural mechanism for the simultaneous
formation of both highly sheared filament channels and relatively
untwisted coronal loops. In this model, these two seemingly unrelated
features of the solar atmosphere are actually created by the same
process \citep{Antiochos13}.\par
The helicity condensation model was initially simulated by
\citet{Zhao15}, who injected magnetic helicity into a plane-parallel
Parker corona \citep{Parker72}. These authors found that photospheric
motions that inject the same helicity everywhere form filament
channels at the PIL. Furthermore, randomizing the photospheric motions
while keeping the same helicity injection rate did not qualitatively
affect the accumulation of twist flux at the PILs. \citet{Zhao15} also
tested the effect of injecting helicity of opposite signs on adjacent
flux tubes, and found that their fields could not reconnect due to the
twist components being co-aligned.\par
In subsequent work \citep[][hereafter KAD15]{Knizhnik15}, we
rigorously tested the helicity condensation model. We found that
it not only qualitatively produced results consistent with the
properties of filament channels, but that the inverse cascade of
magnetic helicity due to reconnection produces a twist flux at
the PIL that agrees quantitatively with the predictions of the
helicity condensation model. Based on this result, we estimated that
with the helicity preference observed on the Sun, filament channels
will form in about a day or so, in agreement with observations of filament
channel formation \citep{Martin98, Gaizauskas00}. We showed that
helicity condensation agreed both qualitatively and quantitatively
with observed properties of filament channels, and that the process
produced relatively untwisted coronal loops everywhere except at the
PIL. These results, however, were obtained for a $100\%$ helicity
rule, meaning that all of the helicity injected into the corona was of
the same sign. An obvious question to be raised is: what happens
if a fraction of the injected helicity has the opposite sign? Indeed,
this is a more realistic scenario since, as described above, the
corona has a hemispheric helicity preference, rather than a rule, so
that some helicity of the non-preferred sign is injected into the
corona at all times.\par
It is reasonable to expect that injecting helicity of the opposite sign
into the corona would simply slow down the helicity condensation process.
However, this result is not as straightforward as it may seem. The
simulations of \citet{Zhao15} demonstrated that adjacent flux tubes have
difficulty reconnecting if they are twisted in opposite senses. As
a result, the twist was unable to inverse-cascade to larger scales,
as is required to form the sheared filament channels and smooth
coronal loops. Even if reconnection between the adjacent flux tubes is
eventually achieved as a result of some instability -- such as ideal
kinking -- driving the interaction, the twist-flux cancellation is
expected to be far from perfect. Substantial residual twist could
remain in what otherwise would have been smooth, untwisted coronal
loops. Therefore, it is important to test whether sheared filament
channels and smooth coronal loops form when both signs of magnetic
helicity are injected into the corona.\par
In this paper, we investigate the effect of varying helicity preference
on the structure of the closed-field corona. We performed helicity-conserving
numerical simulations that inject helicity into a plane-parallel
Parker corona, as we did in KAD15. Extending our previous work, the
fraction of helicity of each sign that is injected into the corona
was varied. We report on three cases: 1) $100\%$ of the
injected helicity is positive; 2) $75\%$/$25$\% of the injected
helicity is positive/negative; and 3) $50$\%/$50$\% of the injected
helicity is positive/negative. To make the simulations as realistic as
possible, we also randomized the pattern of helicity injection. We
compare the simulations with a fixed pattern of helicity injection to
those with a randomized pattern of helicity injection.\par
The paper is organized as follows. In \S \ref{sec:model} we discuss
the setup and initialization of our numerical simulations.
In \S \ref{sec:Hinjection} we describe how magnetic helicity
is injected into the domain, and how the helicity preference is
employed. In \S \ref{sec:Results} we discuss the results of our
simulations, exploring the formation of filament channels and the
smoothness of coronal loops for various helicity preferences, and
compare the simulations with fixed and randomized patterns of
helicity injection. We discuss the implications for understanding
coronal magnetic structure in \S \ref{sec:implications}. \par
\section{Numerical Model}\label{sec:model}
The numerical model used in this study was described previously in KAD15. We solve the equations of magnetohydrodynamics (MHD) using the Adaptively Refined Magnetohydrodynamics Solver \citep[ARMS; ][]{DeVore08} in three Cartesian dimensions. The equations have the form
\beg{cont}
\pd{\rho}{t}+\divv{\rho\textbf{v}}=0,
\end{equation}
\beg{momentum}
\pd{\rho\textbf{v}}{t} + \divv{\left( \rho\textbf{v}\vecv \right)} = - \nabla P + \frac{1}{4\pi} \left( \curl{\textbf{B}} \right) \times \textbf{B},
\end{equation}
\beg{energy}
\pd{T}{t} + \divv{\left( T\textbf{v} \right)} = \left( 2 - \gamma \right) T \divv{\textbf{v}},
\end{equation}
\beg{induction}
\pd{\textbf{B}}{t} = \curl{ \left( \textbf{v} \times \textbf{B} \right)}.
\end{equation}
Here $\rho$ is mass density, $T$ is temperature, $P$ is thermal pressure, $\gamma$ is the ratio of specific heats, $\textbf{v}$ is velocity, $\textbf{B}$ is magnetic field, and $t$ is time. We close the equations via the ideal gas equation,
\beg{ideal}
P = \rho RT,
\end{equation}
where $R$ is the gas constant.\par
ARMS uses finite-volume representations of the variables to solve the system of equations. Its Flux Corrected Transport algorithms \citep{DeVore91} provide minimal, though finite, numerical dissipation, which allows reconnection to occur. As a result, to a very good approximation, ARMS conserves the magnetic helicity in the system.\par
We set up a model coronal field that is initially straight and uniform between two plates, as shown in Figure \ref{fig:init}. In this model, straight flux tubes represent coronal loops whose apex is located in the center of the domain and with the boundaries representing the photosphere. Our domain size is $[0,L_x] \times [-L_y,L_y] \times [-L_z,L_z]$, where $x$ is taken normal to the photosphere (the vertical direction) and we set $L_x=1$, and $L_y=L_z=1.75$. At all six sides, we use zero-gradient conditions, and the four side walls have open boundary conditions. Closed boundary conditions are employed at the top and bottom, where the magnetic field is line tied at the high-$\beta$ photosphere. The footpoints of the field lines do not move in response to magnetic forces, but do respond to imposed boundary flows to mimic driving at the plasma-dominated photosphere.\par
As in our previous work, we set the initial, uniform values in our dimensionless simulations to $\rho_0=1$, $T_0=1$, $P_0=0.05$, and $B_0=\sqrt{4\pi}$. These choices set the gas constant, $R=0.05$, the Alfv\'en speed, $c_{A0}=B_0 / \sqrt{4\pi\rho_0} = 1$, and the plasma beta, $\beta_0=8\pi P_0/B_0^2=0.1$. $\beta\ll1$ corresponds to a magnetically dominated plasma, which is generally true of the corona. The results discussed below will be given in simulation time, which is normalized to the time required for an Alfv\'en wave at unit speed ($c_{A0}=1$) to travel the distance separating the top and bottom plates ($L_x=1$).\par
As before, we set up our convective cells in a hexagonal pattern, with 84 cells on the top and bottom plates. Each individual cell has the same spatial and temporal profiles described in KAD15. The angular velocity of each cell is given by
\beg{angular}
\Omega(r,t)=
\begin{cases}
-\Omega_0g(r)f(t) & \quad \text{if } r\le a_0 \\
0 & \quad \text{if } r > a_0 \\
\end{cases}
\end{equation}
where
\beg{gofr}
g(r)=\left(\frac{r}{a_0}\right)^4-\left(\frac{r}{a_0}\right)^8
\end{equation}
and
\beg{foft}
f(t)=\frac{1}{2}\left[1-\cos\left(\frac{2\pi t}{\tau}\right)\right]
\end{equation}
with $\Omega_0=7.5$ and $r$ the cylindrical radial coordinate with respect to the center of the convective cell. The flow is confined to a circle of radius $a_0=0.125$, and the magnitude of the flow is ramped up and back down over a period $\tau=3.35$. As demonstrated in KAD15, this velocity profile for the convective cells conserves the normal magnetic field distribution, $\textbf{B}_x$, on the photospheric boundaries.\par
The simulation mesh for this study is specified, as in KAD15, such that we resolve very finely that part of the domain where these photospheric flows are imposed and the coronal magnetic field is influenced by the surface stresses. We use $4\times14\times14$ elemental blocks to span the simulation domain, each containing $8\times8\times8$ uniform, cubic grid cells. In the highly resolved portion of our simulation volume, which included the convective cells, the lanes between them, the untwisted region in the interior, and a buffer region around the outer perimeter of the pattern, we applied two additional levels of refinement, such that each rotation was covered by 32 grid points across its diameter. Closer to the side walls, the grid was allowed to coarsen by two levels, such that the grid spacing near the walls was a factor of four larger than in the interior. With this grid distribution, the ratio of the smallest grid spacing to the height of the box was about 0.001, resulting in a magnetic Reynolds number $R_m \sim 10^3$.
The key difference from our previous work, described in detail below, is that the sense of rotation of each cell, as well as the angular orientation of the entire pattern of cells, is changed randomly during the course of the simulations. The top plate mirrors the bottom plate at all times. These variations are meant to model more faithfully the random nature of the Sun's surface convection. Below, we describe how the relative fraction of cells rotating in the opposite sense affects the amount of magnetic helicity injected into our simulations.\par
\section{Magnetic Helicity Injection}\label{sec:Hinjection}
Magnetic helicity is a topological quantity describing linkages in the magnetic field, such as twist, shear, and writhe. In a volume $V$ bounded by a surface $S$, which need not be a magnetic flux surface, the relative magnetic helicity is given by \citep{Finn85}
\beg{Hdef}
H = \int_V{\Big(\textbf{A}+\textbf{A}_P\Big)\cdot\Big(\textbf{B}-\textbf{B}_P\Big) \;dV}.
\end{equation}
Here $\textbf{B}=\nabla\times\textbf{A}$ is the magnetic field in the volume $V$, generated by the vector potential $\textbf{A}$, and $\textbf{B}_P = \nabla \times \textbf{A}_P$ is a current-free field ($\nabla\times\textbf{B}_P = 0$) satisfying $\textbf{B}_P\cdot\unit{n}|_S = \textbf{B}\cdot\unit{n}|_S$. The rate of change of the helicity in Equation (\ref{Hdef}) is given in ideal MHD by
\beg{dHdtdef}
\frac{dH}{dt} = 2 \oint_S{\Big[\Big(\textbf{A}_P\cdot\textbf{v}\Big)\textbf{B} - \Big(\textbf{A}_P\cdot\textbf{B}\Big)\textbf{v}\Big]\cdot d\textbf{S}}.
\end{equation}
The first term represents the effects of motions on the boundary with velocity $\textbf{v}$, while the second term represents the emergence or submergence of magnetic field through the boundary. As a result, the magnetic helicity in our simulation changes only due to motions on or through the boundary. In highly conducting ($R_m \gg 1$) plasmas, such as the corona, magnetic helicity is conserved even in the presence of a small localized resistivity that enables magnetic reconnection \citep{Woltjer58, Taylor74, Taylor86, Berger84b}. Since no new flux is being injected at our top and bottom boundaries, and the side boundaries are sufficiently far away that no flux leaves the system, the rate of change of magnetic helicity in our simulation reduces to
\beg{dHdtsim}
\frac{dH}{dt} = 2 \oint_S{\Big(\textbf{A}_P\cdot\textbf{v}\Big)\textbf{B}\cdot d\textbf{S}}.
\end{equation}
The magnetic helicity $H_0$ injected into a single flux tube (i.e. one top/bottom pair of rotation cells) over one cycle is obtained by integrating Equation \ref{dHdtsim} from $t=0$ to $t=\tau$, employing Equations \ref{angular}--\ref{foft} and the vector potential for the uniform, current-free initial field,
\beg{Ap}
\textbf{A}_P = \frac{B_0}{2}\Big(y\unit{z}-z\unit{y}\Big).
\end{equation}
If the sense of rotation is clockwise, the resulting positive helicity injected is
\beg{helicity1cell}
H_0 = 2\times10^{-2}.
\end{equation}
Trivially, the net helicity $H_\Sigma$ injected into $N$ such flux tubes all twisted in the clockwise sense, as in KAD15, is
\beg{NH1KAD15}
H_\Sigma = N H_0.
\end{equation}
In this paper, we generalize to cases in which $N_+$/$N_-$ cells rotate clockwise/counter-clockwise and inject positive/negative helicity, with $N_+ + N_- = N$. The net helicity injected into the corona then becomes
\beg{helicityallN}
H_\Sigma = \left( N_+ - N_- \right) H_0.
\end{equation}
The case studied in KAD15 has $N_+ = N$ and $N_- = 0$, so Equation (\ref{helicityallN}) reduces to Equation (\ref{NH1KAD15}). If, on the other hand, $N_+ = N_-$, equal numbers of cells rotate in each sense and the net injected helicity vanishes, $H_\Sigma = 0$. In our simulations described below, we allowed $N_+$ and $N_-$ to vary from cycle to cycle. The net helicity injected into the corona after $M$ cycles therefore is
\beg{Htotal}
H_\Sigma = \sum_{i=1}^{M} \left( N_{+,i} - N_{-,i} \right) H_0,
\end{equation}
where $N_{+,i}/N_{-,i}$ is the number of cells that rotate clockwise/counter-clockwise during cycle $i$.\par
In each simulation, we assign a probability $k$ that any individual top/bottom pair of cells injects positive helicity and $1-k$ that the pair injects negative helicity. A random number $\kappa_j \in [0,1]$ is generated for each pair of cells $j \in [1,N]$ during each cycle, and $\kappa_j$ is compared with $k$ to determine whether the sense of rotation is clockwise or counter-clockwise over that cycle. The helicity $H_j$ injected by the $j$th pair of cells is
\beg{Hj}
H_j = H_0 \times
\begin{cases}
+1 & \text{if } \kappa_j \le k; \\
-1 & \text{if } \kappa_j > k. \\
\end{cases}
\end{equation}
On average, the expectation is that during each cycle, a fraction $2k-1$ of the maximum positive helicity $N H_0$ will be injected into the corona,
\beg{Hk}
\langle H_\Sigma \rangle = (2k-1) N H_0 = f N H_0.
\end{equation}
Throughout the paper, we will refer to $k$ as the {\rm helicity preference} of each simulation, and to $f = 2k-1$ as the {\rm net fractional helicity} associated with $k$.\par
For this paper, the cases $k=0.75$ and $k=0.5$ were simulated to complement the $k=1$ case previously presented in KAD15. For reference, the expectation values of the net helicity injected per cycle are
\beg{helicityk}
\langle H_\Sigma \rangle = f N H_0 = N H_0 \times
\begin{cases}
1.0 & \text{if } k=1.0; \\
0.5 & \text{if } k=0.75; \\
0.0 & \text{if } k=0.5. \\
\end{cases}
\end{equation}
Below, we use the precise number of cells injecting each sign of helicity during each cycle to evaluate $H_\Sigma$ in Equation (\ref{Htotal}). That prediction is compared to the instantaneous value $H(t)$ calculated directly from the \citet{Finn85} volume integral for the relative magnetic helicity in the simulation, as described in KAD15.\par
The helicity preference $k$ introduces one aspect of randomness into our simulations through the assignment of a clockwise/counter-clockwise sense of rotation (and positive/negative helicity) to each pair of rotation cells during each cycle of rotation. In order to emulate the stochastically shifting spatial pattern of convection on the solar surface and investigate its effect on coronal structure, we also introduce a second aspect of randomness into a separate set of simulations. After each cycle of rotations, we displace the entire hexagonal cellular pattern shown in Figure \ref{fig:init} by a randomly chosen angle $\theta \in [0^\circ,60^\circ]$ about its central vertical axis $(y,z) = (0,0)$. For simplicity, the same angular displacement is applied to both the top and bottom plates, so that the top/bottom pairs of rotation cells remain aligned as before. Now, however, the random displacement means that the rotation cells will, in general, encompass parts of multiple neighboring flux tubes that were twisted during the previous cycle of rotations. The ensuing cycle therefore introduces braiding, as well as twisting, into the coronal magnetic field between the plates. As we will show, however, this displacement-induced braiding has no effect on the rate of helicity accumulation in the corona, and has only a minor influence on the smoothness of the induced magnetic structure. This result concurs with the qualitative conclusions of \citet{Zhao15} from a much simpler simulation setup and is analyzed quantitatively here for the first time.\par
\section{Results}\label{sec:Results}
In this section, we first describe the results of our simulations with both fixed and randomized patterns and for the various helicity preferences. Then, we analyze those results in the context of filament-channel formation and the smoothness of coronal loops.\par
\subsection{Fixed Cellular Pattern}
The first set of simulations holds the cellular pattern fixed in the orientation shown in Figure \ref{fig:init}, randomizing only the sense of rotation of the individual cells as described in \S \ref{sec:model}. To compare the $k=0.75$ simulation most consistently with the $k=1.0$ case presented in KAD15, we ran it for twice as many cycles (42 vs.\ 21). As shown by Equation (\ref{Htotal}), the expectation value for the net injected helicity is the same (50\% of $42 \times N$ vs.\ 100\% of $21 \times N$). The $k=0.5$ case, in contrast, accumulates a net helicity only due to statistical fluctuations away from its average value of zero. For that case, therefore, we simply ran the simulation for $21$ cycles. All three simulations then were extended for $5$ additional cycles without imposing any rotational motions, to allow transients to die down and the system to relax toward a quasi-equilibrium final state.\par
As an example, Figure \ref{fig:vphi_f} shows the azimuthal component of the velocity on the bottom plate, $V_\phi(x=0,y,z)$, during the first (left) and second (right) cycles of the $k=0.75$ case. For both the $k=0.75$ and $k=0.5$ cases, each individual cell does not necessarily preserve its sense of rotation from one cycle to the next. The sense of rotation is assigned randomly at each cycle, as given above in Equation (\ref{Hj}). For $k=1.0$, all cells rotate in the same sense in the first and second -- indeed, throughout all -- cycles. \par
Figure \ref{fig:helicityf} shows the analytically expected (solid line) and numerically calculated (dashed line) helicities for each simulation. The orange ($k=1.0$) curve is the same as that presented in KAD15. After $21$ cycles, the rotation cells have injected $H=36$ units of helicity. The red ($k=0.75$) curve shows that approximately the same $H=36$ units are injected over twice the time (cf.\ Equation \ref{helicityk}) for the 75\% preference. The blue ($k=0.5$) curve shows, as expected, that almost no net helicity is accumulated in the simulation with 50\% preference over its first $21$ cycles. All of the numerically calculated curves match very well the analytical values at each cycle, demonstrating that our simulations conserve helicity to a very high degree of accuracy. Therefore, the evolution of the magnetic field in our simulations is due predominantly to convection and reconnection, rather than to numerical diffusion that would dissipate helicity.\par
\subsection{Randomly Displaced Patterns}
The second set of simulations is identical to the first, except that we also displace the entire hexagonal cellular pattern through a random angle after each cycle as described in \S \ref{sec:model}. In these setups, different flux tubes wrap around each other, creating a braided field, in addition to being twisted by the rotation cells. We ran each simulation for the same number of cycles as in the fixed-pattern cases and for the same values of $k$.\par
As an example, Figure \ref{fig:vphi_r} shows $V_\phi(x=0,y,z)$ during the first (left) and second (right) cycles of the $k=0.75$ case. Like the corresponding fixed pattern cases, in the random $k=0.75$ and $k=0.5$ cases, each individual cell does not necessarily maintain its sense of rotation. Unlike the fixed pattern cases, however, the pattern itself is displaced by a random angle after each cycle. Except for the $k=1.0$ case, the randomized patterns (Fig.\ \ref{fig:vphi_r}) exhibit different distributions of color than the fixed patterns (Fig.\ \ref{fig:vphi_f}). We used different sequences of random numbers $\kappa_j$ to set the clockwise/counter-clockwise sense of rotation of the individual rotation cells in the two sets of simulations. The random angular displacements of the cellular pattern between the first and second cycles are evident by comparing the left and right columns for each helicity preference.\par
Figure \ref{fig:helicityr} shows the analytically expected (solid line) and numerically calculated (dashed line) helicities for the various cases. Although the average helicities (Eq.\ \ref{helicityk}) injected into the corona are identical for each value of $k$, the precise helicities (Eq.\ \ref{Htotal}) actually injected differ between the fixed and randomized cases due to statistical fluctuations. Thus, the curves in Figures \ref{fig:helicityf} and \ref{fig:helicityr} are slightly different for $k \ne 1.0$. The orange ($k=1.0$) curve shows that the helicity injected for the 100\% preference is identical for the fixed and randomized patterns, as expected. The red ($k=0.75$) curve shows that the 75\% preference injects slightly more helicity in twice the time. The blue ($k=0.50$) curve shows that, as before, almost no net helicity is injected in this case. In this randomized-pattern simulation, the residual net helicity for the 50\% preference is small and positive, whereas in the fixed-pattern simulation, it is negative. In all cases, we again find excellent agreement between the numerically calculated and analytically expected helicities.\par
\subsection{Formation of Filament Channels}
Figure \ref{fig:Bphi_beg} shows the azimuthal component of the magnetic field in the horizontal mid-plane, $B_\phi(x=0.5,y,z)$, halfway through the first cycle of twist for each simulation. At this early stage, each case exhibits the characteristic hexagonal pattern of rotation cells. For $k=1.0$, every cell injects the same sign of $B_\phi$, so adjacent flux tubes always have oppositely directed twist fields and are able to reconnect readily. For the $k=0.75$ and $k=0.5$ cases, in contrast, adjacent twist fields sometimes are in the same direction. On average, this is true half the time in the $k=0.5$ case, suppressing reconnection between adjacent flux tubes whose twist fields are parallel rather than anti-parallel.\par
The effect of the helicity preference on the formation of filament channels can be seen clearly in the final-time $B_\phi$ maps in Figure \ref{fig:Bphi_end}. The $k=1.0$ case has accumulated oppositely signed bands of twist flux at the outer and inner boundaries of the hexagonal pattern, as described previously in KAD15. These bands result from the inverse-cascade of twist flux from small to large scales due to reconnection, collecting at the boundaries to form filament channels according to the helicity-condensation model \citep{Antiochos13}. The $k=0.75$ case has been run out twice as long in order to accumulate roughly the same helicity, and it has acquired similar bands of twist at the outer and inner boundaries of the hexagonal pattern. Thus, despite the one-third (25\%/75\%) of twist fields on neighboring flux tubes that are parallel rather than anti-parallel in this simulation, sufficient reconnection has occurred to enable the helicity to condense at the flux-system boundaries here, as well. The shapes of the filament channels differ slightly in the $k=1.0$ and $k=0.75$ cases. The contrast is most evident in the randomized-pattern simulations, where the twist flux has a very uniform, circular appearance in the $k=1.0$ case, while the structure is more ragged, especially at the inner boundary, in the $k=0.75$ case. At the largest scales, however, these two cases yield qualitatively identical outcomes: the twist flux forms two bands of opposite polarity at the boundaries of the hexagonal pattern of rotations. In the corona, such bands would manifest themselves as extended, sheared filament channels.\par
The sharply contrasting $k=0.5$ case, on the other hand, displays a very different final-time appearance. No long, coherent bands of twist flux have accumulated at either the outer or inner boundaries of the hexagonal pattern. Instead, there are localized concentrations of twist flux dispersed across the interior of the pattern, as well as at its boundaries. Because zero net helicity is injected into this system, on average, zero net twist flux is available to be transported by reconnection to the hexagonal boundaries where it can accumulate. Turning this argument around, if the net condensed twist flux were finite, then the net helicity would be finite as well. We demonstrated this result analytically in KAD15. Consequently, the helicity-condensation process does not form filament channels in the case of a 50\% helicity preference.\par
These examples demonstrate that the helicity preference plays a major role in the organization of the twist flux and the formation of filament channels. The $k=0.75$ case forms similarly strong, although rather more structured, bands of twist flux over twice the time as the $k=1.0$ case. As is argued below, the time scale for filament-channel formation is inversely proportional to the average net fractional helicity injected, i.e.\ to $f$. This dependence is supported further by the absence of filament-channel structure in the $k=0.5$ case, whose predicted time scale for channel formation is infinite.\par
\subsection{Accumulation of Twist Flux}\label{sec:twistflux}
The results above demonstrate that there are major qualitative and quantitative contrasts between the results for different helicity preferences, but more minor differences between the fixed and random patterns for a given helicity preference. We begin the quantitative analysis of our simulations by calculating the positive twist flux $\Phi_{tw}^+$ through the $z=0$ plane,
\beg{Phitw}
\Phi_{tw}^+ = \int_0^{L_x}{dx} \int_0^{L_y}{dy \; B_{tw}^+(x,y,z=0)},
\end{equation}
where the corresponding positive twist field $B_{tw}^+$ is
\beg{Btw}
B_{tw}^+ = \frac{1}{2} \left( B_\phi + \left\vert B_\phi \right\vert \right) \ge 0.
\end{equation}
The twist flux $\Phi_{tw}^+$ is plotted in Figure \ref{fig:tw_v_time} for both the fixed (solid curves) and random (dashed curves) patterns. All six simulations exhibit a brief initial phase of ideal evolution, of about one rotation cycle in duration, in which twist flux is injected into and stored in individual, non-interacting flux tubes. At this stage, the sense of rotation of adjacent cells is irrelevant to the accumulation of twist flux. As the twisting continues, however, the flux tubes expand laterally to compress the volume between them. This forms and strengthens electric current sheets between neighboring tubes that have anti-parallel twist fields. Reconnection between such tubes commences during subsequent twist cycles. This process, together with the randomization of the sense of rotation of individual cells (for $k \ne 1.0$) and of the orientation of the cellular pattern (for the random cases), causes the curves to deviate increasingly from one another at later times.
The two cases with nonzero net fractional helicities, $k=1.0$ (orange) and $k=0.75$ (red), show relatively small differences between the fixed and random patterns for fixed $k$. Over the full duration of the simulations, each preference accumulates essentially the same twist flux. The slightly larger values for the $k = 0.75$ case reflect the slightly larger magnetic helicities accumulated in those simulations (Figs.\ \ref{fig:helicityf} and \ref{fig:helicityr}) compared to the $k = 1.0$ case. All four of these simulations eventually accumulate twist flux at a rate per cycle that is in good agreement with the calculation by KAD15 (their Equation 4.13 and Figure 11),
\beg{DPhitw}
\Delta \Phi_{tw} = \frac{1}{2} \frac{\Delta \langle H_\Sigma \rangle}{\Phi_N} = \frac{f}{2} \frac{H_0}{\Phi_0},
\end{equation}
where each of the $N$ twisted flux tubes contains $\Phi_0$ units of magnetic flux. The expression in Equation (\ref{DPhitw}) assumes that the twist flux $\Phi_{tw}$ condenses at the outer boundary of the flux system, which occurs in our simulations with $k = 1.0$ and $k = 0.75$. In a spirit similar to KAD15 (their Equation 4.28), we calculate the filament-channel formation time $\tau_{fc}$ over which a critical amount of twist magnetic flux $\Delta \Phi_{fc}$ accumulates,
\beg{Taufc}
\tau_{fc} = \frac{\Delta \Phi_{fc}}{\Delta \Phi_{tw}} \tau_0 = \frac{2}{f} \left( \frac{\Phi_0}{H_0} \tau_0 \right) \Delta \Phi_{fc},
\end{equation}
where $\tau_0$ is the duration of one twist cycle. Equations (\ref{DPhitw}) and (\ref{Taufc}) quantitatively express the observed factor-of-two differences in twist-flux accumulation rates and filament-channel formation times between our $k = 1.0$ ($f = 1.0$) and $k = 0.75$ ($f = 0.5$) cases. They also predict how these quantities should change for other helicity preferences $k$.\par
Our last two simulations, with $k=0.5$, have zero net fractional helicity, $f = 0.0$. For this case, the predicted accumulated twist flux $\Delta \Phi_{tw}$ vanishes and the filament-channel formation time $\tau_{fc}$ is infinite. We observe in these simulations (Figure \ref{fig:tw_v_time}) that the fluctuations in the twist flux are relatively large, and the average amount of flux saturates after about 10 cycles have elapsed. Thereafter, the average seems to be statistically quasi-steady, increasing or decreasing randomly according to the cycle-to-cycle variations of the sign of twist in individual rotation cells (in both simulations) and of the orientation of the cellular pattern (in the random-pattern simulation only). Evidently, these simulations have reached a roughly steady-state balance between the rates of twist-flux injection by the twisting motions and extraction via a combination of untwisting motions and reconnection between anti-parallel twist fields.\par
\subsection{Smoothness of Coronal Loops}
We have seen that, when it is effective, the helicity-condensation
process transports twist via reconnection to the boundaries of the
flux system, where it condenses. This leaves the interior of the
system relatively smooth and untwisted. The final configuration then
corresponds to a corona with strong shear concentrated at its PILs and
laminar coronal loops in interior regions away from its PILs. This result can
be seen clearly in Figure \ref{fig:Bphi_end}. In the $k=1.0$ and $k=0.75$
cases with nonzero net fractional helicities, at a glance, the interior
of each flux system seems very smooth, with little twist evident. A
careful comparison of the two cases reveals that the annular region
between the filament channels is somewhat more structured for
$k=0.75$, with localized, small-amplitude twists of both signs
accumulating in the interior. As might be anticipated, this structure
is somewhat less noticeable for the simulations with randomly
displaced patterns compared to their fixed-pattern counterparts.\par
The appearance of the $k=0.5$
is strikingly different from the $k=1.0$ and $k=0.75$
cases. For the fixed pattern especially (bottom row, left column of Figure
\ref{fig:Bphi_end}), small-scale, coherent concentrations
of twisted field are present throughout the interior of the hexagonal
flow region. In addition, the magnitude of the accumulated twist is
significantly smaller than in the $k = 1.0$ and $k = 0.75$ cases. This
is due to both the random untwisting of previously twisted field lines
in successive cycles and the zero net twist flux that can accumulate
globally and be transported to the flux-system boundaries. The local
twist concentrations that are formed appear and disappear transiently
as the system evolves. Each such concentration has a lifetime on the
order of one rotation period of the convection cells.
Taking the rotation period to be of order a day, or the lifetime of
a typical supergranule, these concentrations of twist should easily
survive for timescales long enough to be detected remotely. The lack
of such observations indicates that the photosphere likely injects
helicity with a significant preference.\par
To demonstrate the stark difference in the amount of structure in the different helicity preference cases we plot in Fig. \ref{fig:structure_f} a set of magnetic field lines from the same set of fixed points for the fixed-pattern $k=1$, $k=0.75$, and $k=0.5$ cases. All of the field lines are chosen from the interior of the hexagonal region, which represents the `loop' portion of the corona. The bottom plate represents magnetic field magnitude, which shows a structure similar to that shown in Fig. \ref{fig:Bphi_end}. Although only a sample of field lines is chosen, they are representative of the field lines in the rest of the `loop' portion of the corona. The striking difference in the amount of structure in the corona is immediately evident by comparing the $k=1$ and $k=0.75$ cases with the $k=0.5$ case. In the latter simulation, field lines are twisted and braided around each other in a complicated fashion. In the first two cases, although the field lines are traced from the same points, the field lines themselves are quite smooth and laminar. There may appear to be some structure due to field lines passing behind, or in front of, each other, but there is almost no significant twisting or braiding around each other. In fact, the top two figures very closely resemble the initial, uniform field configuration, and, importantly, the smooth, laminar structure observed in Fig. \ref{fig:obs}, albeit in a plane-parallel geometry. In this sense, the $k=1$ and $k=0.75$ coronal loops are quasi-potential, while the $k=0.5$ coronal loops clearly deviate quite strongly from quasi-potentiality. In Fig. \ref{fig:structure_r}, a different set of field lines is plotted for the random-pattern $k=1$, $k=0.75$ and $k=0.5$ cases. These sample field lines are traced from the same footpoints in each case, and again demonstrate the quasi-potentiality of the corona in the former two cases, and the large amount of structure in the latter case. \par
\subsection{Fluctuations of Twist Field}
To quantify the amount of small-scale structure in the various cases, we calculated the angle-averaged azimuthal magnetic field $\langle B_\phi(r) \rangle$ and its root mean square deviation $\delta B_\phi(r)$ in the mid plane $(x = 0.5)$. Specifically, we evaluated (for $m = 1,2$)
\begin{align}
\label{Bm}
\langle B_\phi^m(r) \rangle &= \frac{1}{2\pi} \int_0^{2\pi} B_\phi^m(x=0.5,y,z) d\phi,\\
\label{dB}
\delta B_\phi(r) &= \sqrt{\langle B_\phi^2(r) \rangle - \langle B_\phi(r) \rangle^2},
\end{align}
where $r = \sqrt{y^2+z^2}$. A discrete $\rho$ grid, with the same
spacing as the $y$ and $z$ grids, was adopted, and all cell-center
positions $(y,z)$ were grouped into corresponding $r$ intervals to
calculate the integrals in Equation (\ref{Bm}).
Figure \ref{fig:width} shows $\delta B_\phi(r)$ at the end of the
random pattern simulations for each value of $k$ (color-coded). The
rms deviations $\delta B_\phi(r)$ shown in Figure \ref{fig:width}
all exhibit small-scale statistical fluctuations. The average
amplitude of the fluctuations is smallest for the 100\% helicity
preference ($k = 1.0$, orange) and largest for the 50\% preference
($k = 0.5$, blue), where they are nearly three times as large
compared to the 100\% helicity preference. A very similar trend was
seen in the fixed-pattern simulations. As evident from Figure 11,
these localized fluctuations in twist field for the zero helicity
case correspond to field line tangling that would easily be
observed if it were present in the real corona.\par
\subsection{Length of Field Lines}
Further insight into the overall magnetic structure of the corona is gleaned by examining the lengths of field lines throughout the domain. In Figure \ref{fig:Blength}, we plot the length of magnetic field lines for the various helicity preferences. Here the differences between the filament channels and coronal loops is greatest in the $k = 1.0$ simulations, where substantially longer field lines reside in the filament channels. The coronal loops in the interior are quite short. Indeed, their lengths are very close to those of the untwisted field lines exterior to the hexagonal region. The two $k = 0.75$ simulations, meanwhile, display a slightly more mixed character in the interior. The coronal-loop field lines in these cases are slightly longer than the exterior untwisted field, although not nearly as long as the filament-channel field lines, which themselves are somewhat shorter than for $k = 1.0$.\par
As we found for several diagnostics described previously, the $k = 0.5$ case displays a strikingly different appearance from those with nonzero net fractional helicities $f$. As can be seen in Figure \ref{fig:Blength} (bottom row), the hexagonal region of rotation cells hosts a rather homogeneous mixture of relatively short field lines, structured at small scales. This is not dissimilar to the interior of the hexagonal region in the $k = 0.75$ case (middle row). However, no large-scale organization of the field-line length is evident, beyond the exclusion of twist from the center of the domain and the region beyond the perimeter of the hexagonal region. This is in sharp contrast to the cases $k = 1.0$ and $0.75$, where much longer field lines accrue at both the inner and outer boundaries of the flux system.\par
\par
\section{Implications for Coronal Structure}\label{sec:implications}
The results described in the preceding section have important implications for the global structure of the solar corona. Our findings demonstrate that the magnetic helicity preference $k$ plays key roles in determining how the corona is structured and the time scale over which that structure develops. The contrast is particularly strong between the cases with 100\% ($k = 1.0$) and 50\% ($k = 0.5$) preferences and randomly displaced cellular patterns shown at the top and bottom right, respectively, in Figures \ref{fig:Bphi_end} and \ref{fig:Blength}. For the 100\% preference, the twist flux condenses into two primary bands with opposite senses of twist at the inner and outer boundaries of the hexagonal region of rotation cells, with very little twist in the interior. These concentrations and dilution are reflected in the lengths of the associated magnetic field lines, which are very long near the two boundaries but minimally short in the interior. For the 50\% preference, on the other hand, the twist flux does not condense into any recognizable global-scale structure, and the field lines have an essentially homogeneous distribution of intermediate lengths. These two cases have the largest and smallest (in magnitude) net fractional helicities, $f = 1.0$ and $f = 0.0$, respectively.\par
Our intermediate case with 75\% preference ($k = 0.75$, $f = 0.5$) exhibits some features of both of the previous limiting cases but, importantly, qualitatively resembles more closely the results for the 100\% preference. The bands of condensed twist flux still form, albeit twice as slowly and with significant intrusions of twisted structures between them, and the field-line lengths are correspondingly longer at the hexagonal-region boundaries than in its interior, although with less contrast. Extrapolating to other cases with even smaller preferences but nonzero fractional helicities -- say, $k = 0.625$ and $f = 0.25$ -- we would expect these trends to continue, with a further increase in the filament-channel formation time (another doubling for $f = 0.25$) and in the amount and homogeneity of small-scale structure in the interior of the hexagonal region.\par
Perhaps the clearest example of the effect of helicity preference on the structure of the closed loop corona is evident in Figures \ref{fig:structure_f} and \ref{fig:structure_r}. The smoothness of the $k=1$ and $k=0.75$ coronal loops is manifestly different than the complexity of the $k=0.5$ coronal loops. Observations of the coronal magnetic field, meanwhile, invariably reveal smooth, laminar loops that closely resemble those observed in the $k=1$ and $k=0.75$ cases (cf. Figure \ref{fig:obs}), rather than those observed in the $k=0.5$ case. Our simulations indicate, therefore, that the photosphere must inject a significant net helicity so that structures such as those seen in the $k=0.5$ are not observed.
The simulation setups assumed in this paper are simplified
compared to the complex photospheric polarity structure exhibited by the Sun, illustrated by
Figure \ref{fig:obs}. Nevertheless, the quasi-random structure that we
obtained for a 50\% helicity preference, shown in the bottom panels of
Figures \ref{fig:Bphi_end} and \ref{fig:Blength}, obviously does not
resemble the clean, bimodal structure observed for the
corona. In contrast, our results for both 100\% and 75\% preferences,
shown in the other panels of those figures, do exhibit the bimodal
characteristics of the corona: concentrations of twist at PILs in the
form of highly nonpotential magnetic shear in filament channels, and
generally smooth, quasi-potential fields free of twist away from PILs
in arcades of coronal loops. Therefore, a principal conclusion of our
work is that the Sun must inject helicity into the corona with a
significant hemispheric preference, favoring negative helicity forming
left-handed structures in the north, and positive helicity forming
right-handed structures in the south. These preferences are reflected
in the observed statistics of solar filaments, sigmoids, and sunspot
whorls. They also have been detected directly in the photospheric
convection, although that measurement is very challenging, near the
limits of observational resolution.\par
Our simulation is simplified in another important way compared to the
Sun: there is no source of new, weakly sheared or unsheared magnetic
flux in our domain, nor is there a sink of the strongly sheared flux
condensing in the filament channels. Flux emergence from below the
photosphere constantly injects fresh magnetic field into the corona,
and coronal mass ejections regularly eject sheared magnetic field and
its entrained magnetic helicity away from the Sun into the
heliosphere. The characteristic time scales for these phenomena
compete directly with the filament-channel formation time $\tau_{fc}$
to establish a quasi-steady balance among these processes and the
coronal magnetic structure that is observed. Such a calculation is
well beyond the scope of this paper, but a first attack on the problem
could be taken using global force-free modeling of the corona
\citep[e.g.\ ][]{Mackay14}. We point out that $\tau_{fc}$ in Equation
(\ref{Taufc}) is inversely proportional to the product of the net
fractional helicity $f$ and the angular rotation rate $\omega_0$ of
the twisting motions ($H_0 \propto \omega_0 \tau_0$; KAD15). If this
product $f \omega_0$ is too small or too large, then the
filament-channel formation time will be too long or too short compared
to the emergence and ejection time scales, and the model is unlikely
to replicate the Sun's observed appearance. We anticipate that global
modeling of the combined processes could provide rigorous bounds on
the rotation rate $\omega_0$, to complement the narrowly constrained
range of values available to the net fractional helicity,
$0.5 < f < 1.0$.\par
Our simulations show that random displacements of the pattern of
photospheric convection have only a secondary effect on the resulting
coronal structure. This also is evident in Figures \ref{fig:Bphi_end}
and \ref{fig:Blength} by comparing the left (fixed-pattern) and right
(random-pattern) columns for each helicity preference. The latter
structures are somewhat smoother than the former, especially at small
scales, but the large-scale organization is no different between
them. This conclusion agrees with that reached by \citet{Zhao15}, who
used a much simpler setup with far fewer rotation cells. The
fundamental reason that the random convection pattern shows a very
similar coronal end state to that of the fixed pattern is that, as shown by
numerous simulations, magnetic reconnection is efficient at
destroying all higher order topological features such as braiding,
leaving only the global helicity \citep[e.g.,][]{WilmotSmith10}. Our
results are fully in agreement with this hypothesis. \par
In summary, this paper presents the first simulations of the evolution
of the coronal magnetic field driven by photospheric motions with
varying helicity preference. Our results agree well with the helicity
condensation model of \citet{Antiochos13}, which accounts for both the
formation of sheared filament channels adjacent to PILs and the
quasi-potential, smooth character of coronal loops away from PILs. By
transferring the magnetic twist injected by photospheric motions to
ever larger scales, reconnection concentrates the twist at the
boundaries of flux systems (i.e., at the PILs) while diluting it
throughout their interiors. The remarkable implication of the model is
that the global organization of the magnetic shear in the solar
atmosphere is a direct consequence of local twisting of the footpoints
of coronal flux tubes by surface convection. Even more remarkable and
somewhat counter-intuitive, is our finding that in order for the hot
closed corona -- the loops -- to exhibit no structure such as tangling
or twisting, then a great deal of structure needs to be injected! The
photospheric driving motions must have a clear helicity preference, in
which case all the injected structure ends up localized around
PILs. The corona, therefore, is a striking example on cosmic scales of
a strongly self-organized system. \par
\acknowledgments{
K.J.K acknowledges the use of post-processing codes originally written by Benjamin Lynch and Peter Wyper. K.J.K received funding for this work through a NASA Earth and Space Science Fellowship. The numerical simulations were performed under a grant of High-End Computing resources to C.R.D. at NASA's Center for Climate Simulation. S.K.A. and C.R.D. were supported, in part, by grants from NASA's Living With a Star and Heliophysics Supporting Research programs.
}
|
\section{Introduction}
Cataclysmic variables (CVs) are interacting binaries in which a white dwarf (WD) is accreting matter from a late--type donor. In the classical picture of a CV a Roche lobe filling star transfer mass through an accretion disc to the WD. In some CVs the material accreted onto a WD reaches a pressure and temperature sufficient to trigger a thermonuclear reaction, which gives rise to a classical nova outburst. If more than one outburst is observed, then it is classified as a recurrent nova (RN; see e.g. \citealt{2003cvs..book.....W}). If the donor star is also a red giant (RG), it is classified as a symbiotic recurrent nova (SyRN; see e.g. \citealt{1999A&A...344..177A}; \citealt{2011A&A...527A..98S}), with the system also belonging to the family of symbiotic stars (SySt; see \citealt{2012BaltA..21....5M} for a recent review).
T~CrB is a SyRN with two outbursts recorded to date (Nova CrB 1866, 1946). The RG in the system has a spectral type M4III \citep{1999A&AS..137..473M} and fills its Roche Lobe \citep{1998MNRAS.296...77B}. The masses of the components are M$_{\mathrm{WD}}=1.2\pm0.2$M$_\odot$ and M$_{\mathrm{RG}}=0.8\pm0.2$M$_\odot$ \citep{1998MNRAS.296...77B,2004A&A...415..609S}. The average quiescent luminosity of the WD is close to 40~L$_\odot$ \citep{1992ApJ...393..289S}. T CrB was discovered as an X--ray source by \citet{1981ApJ...245..609C} and the X--ray emission has a relatively hard spectrum for a symbiotic star \citep{2008ASPC..401..342L}.
Occasionally between the nova eruptions, T~CrB and other SyRNe (V745~Sco, V3890~Sgr and RS~Oph) show enhanced activity of the hot component, manifested by an increase in the emission line fluxes and appearance of a blue continuum. In particular, the \mbox{He\,{\sc i}} and \mbox{He\,{\sc ii}} emission lines became clearly visible in the spectrum whereas there are very weak or absent during most of the quiescent phase \citep{1990JAVSO..19...28I}. The line variability is correlated with the changes in the UV continuum and there is no correlation with the orbital period \citep[e.g.][]{1985ESASP.236..213C}. Similar activity was reported in the episodic appearance of the \mbox{He\,{\sc ii}\,4686} emission line and a hot continuum in the optical spectrum before the 1946 outburst (e.g. \mbox{\citealt{1939ZA.....17..246H}}; \citealt{1943PASP...55..101M}). The nature of these active phases is yet fully understood.
One of the most mysterious phenomena observed in T~CrB is flickering. Flickering is a stochastic variation in the light curve of a star with an amplitude of few tenths of a magnitude on time-scales of seconds or minutes. It is well known that flickering in CVs is associated with accretion onto the WD, but the exact physical process is unknown. One of the first models of flickering included unsteady accretion through a bright spot \citep{1971MNRAS.152..219W}. The first to conduct a systematic study was \citet{1992A&A...266..237B} who estimated theoretical energies and time-scales of flickering. The author considered unstable mass transfer and interaction of matter with the disc edge, dissipation of magnetic loops, turbulence in the accretion disc and unstable accretion in the boundary layer. As a result the most promising model contained unstable mass transfer through a boundary layer and turbulence in the accretion disc (see also \citealt{1993A&A...275..219B}). Different models of flickering include magnetohydrodynamic turbulence transporting the angular momentum outward \citep{1998RvMP...70....1B}, turbulent transport of angular momentum in the whole disc \citep{2010MNRAS.402.2567D}, occasional flare-like events and subsequent avalanche flow in the accretion disk atmospheres \citep{1997ApJ...486..388Y}, discrete flares in the accretion disc \citep{2006Ap&SS.304..291R}, or a more recent model by \citet{2014MNRAS.438.1233S} which is mainly based on a fluctuating accretion disc.
The flickering in T~CrB was observed by a number of authors (see \citealt{2006AcA....56...97G} and references therein). The colours of the flickering source in T~CrB indicate a temperature of $\sim$9000~K, which is significantly lower than the average temperature of flickering sources in classical CVs ($\sim$20000~K; \citealt{2015AN....336..189Z}). Flickering in T~CrB seems to originate in the vicinity of the WD \citep{1998A&A...338..988Z}. The ratio of the amplitude of flickering in T~CrB to the average flux remains constant \citep{2004MNRAS.350.1477Z}, which according to a model by \citet{1993A&A...275..219B}, means that the size of the boundary layer remains roughly constant. This is not necessarily the case in all CVs \citep{1998A&A...332..586F}. Flickering was also observed in the X--ray observations of T~CrB \citep{2008ASPC..401..342L} and in the emission lines \citep{2005PASP..117..268Z}.
In this paper we analyse optical and X--ray observations of T~CrB in order to shed light on the nature of active phases and flickering observed in this system. The collected observations are presented in section~\ref{obs_sec}. In section~\ref{active_phase_section} we present results concerning the active phases of T~CrB. Section~\ref{flickering_section} is dedicated to the optical and X--ray observations of flickering in the system. A brief discussion about the nature of the observed phenomena is presented in section~\ref{discussion_sec} and a summary of the results is given in section~\ref{sumarysec}.
\section{Observations}\label{obs_sec}
\subsection{Spectroscopy}
Spectroscopic observations were obtained from the Astronomical Ring for Access to Spectroscopy database\footnote{http://www.astrosurf.com/aras/} (ARAS). A log of observations is presented in Table~\ref{logspec}. The data include mainly low resolution spectra covering a wide spectral range. Basic spectrophotometric calibration of these spectra was obtained by using a standard star spectrum. This procedure was suitable for obtaining the spectral energy distribution (SED). However, absolute flux calibration could not be obtained this way due to light losses on the slit, variable weather conditions and other effects. Therefore, we scaled the spectra using the known $V$ magnitudes (Sect. \ref{sec:phot}). The emission line intensities were measured by Gaussian profile fitting. When a Gaussian did not fit a line, we measured the local continuum level and integrated all of the signal above that level. The main source of uncertainty was the selection of the local continuum level. The uncertainty was estimated by measuring the same emission lines by different people. This procedure resulted in an estimated precision of 15\% and 30\% in the case of strong and faint lines, respectively. From our analysis we excluded two spectra in the ARAS database. The spectra from 01.04.2012 and 27.04.2015 had unusual SEDs that suggested problems with their flux calibration. The calibrated low resolution spectra, measured equivalent widths and fluxes of emission lines are presented in the Appendix \ref{sect:appendix} and Fig.~\ref{all_fluxes}.
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics{./figs/TCrB_fluxes}}
\end{center}
\caption{Variability of the equivalent width of H$\alpha$, other selected emission line fluxes and the H$\alpha$ to H$\beta$ flux ratio.}
\label{all_fluxes}
\end{figure}
\subsection{Photometry}
\label{sec:phot}
On the nights of 2013 August 3 and 4, we carried out $U$ band observations of the flickering in T~CrB with the 60cm Cassegrain telescope at Rozhen National Astronomical Observatory, Bulgaria. The telescope was equipped with a FLI~PL09000 CCD camera with 3056x3056 pixels and 27'x27' field of view. Single exposure times of 120~s were used for a total of about 6~hours of observations. Data reduction was carried out using standard IRAF procedures.\footnote{IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.} The typical error of each observation was 0.006~mag. A summary of the observations is presented in Table~\ref{tableflickering}. The light curves of the flickering are presented in Fig.~\ref{flickering_fig}.
\begin{figure*}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/TCrB_U1}\includegraphics{./figs/TCrB_U2}}
\caption{Observations of flickering in T CrB.}
\label{flickering_fig}
\end{figure*}
In order to analyse the long term variability we adapted photometry from publicly available catalogues. Namely we used $BVR$ observations from the American Association of Variable Star Observers (AAVSO) database\footnote{https://www.aavso.org/} and $V$ observations from The All Sky Automated Survey (ASAS, \citealt{1997AcA....47..467P}). The typical error of observations in the ASAS and AAVSO data are 0.03~mag and 0.05~mag, respectively. The light curve is presented in Fig.~\ref{TCrB_all}. We searched for periodicities in the data using the discrete Fourier transform method of the $Period04$ program \citep{2005CoAst.146...53L}. For estimating the errors we used the \citet{1999DSSN...13...28M} method. The analysis of the AAVSO $B$ data gave the highest signal at the period of 114d, i.e. half of the orbital period, associated with ellipsoidal variability of the system. Besides that the highest signal was at 1077$\pm$19d. In the case of the AAVSO $V$ data the $\sim$1077d variability seems to be only marginally detected. In $VRI$ observations there is an additional signal present at P=87d and P=164d, both of which are one year aliases of the orbital period (Fig.~\ref{TCrB_powerspec}). We did not detect the $\sim$55d variability discovered by \citet{1988AJ.....95.1505L} that the authors attributed to pulsations of the red giant. The reality of the $\sim$55d variability was already discussed by \citet{1998MNRAS.296...77B} and the authors concluded that it was unlikely to be caused by red giant pulsations.
\begin{figure*}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/TCrB_all}}
\caption{ a) $Swift$ observations of T~CrB from \citet{2009ApJ...701.1992K}; b) X--ray observations from RXTE; c) H$_\alpha$ equivalent width; d) Ratio of H$_\alpha$ to H$_\beta$ flux; e) and f) Photometric observations from AAVSO and ASAS. }
\label{TCrB_all}
\end{figure*}
\begin{figure}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/TCrB_AVVSO_power_spec}}
\caption{Power spectra of the AAVSO observations.}
\label{TCrB_powerspec}
\end{figure}
\subsubsection{Detection of flickering}
In order to detect flickering one must compare the measured scatter of the observations with expected uncertainty from CCD photometry. Usually this is done by using multiple observations to derive the parameters of the CCD camera (see e.g. \citealt{2001MNRAS.326..553S}; \citealt{2006AcA....56...97G}). This method was not useful for our purposes since we did not have a sufficient number of observations collected with the same instrumental setup. Another method is constructing a magnitude--standard deviation diagram for a large number of stars within the same exposure (see e.g. \citealt{1996AJ....111..414D}). This method was not fit for our observations since in our T~CrB field there were only two comparison stars. Therefore we derive our own method of estimating the expected uncertainty in observations of a star with a given magnitude, that requires only two comparison stars.
The error of an observed flux with a CCD camera is given by the equation
\begin{equation}\label{sigma1}
\sigma_i^2=\sigma_{CCD,i}^2+s^2\times N_i^2,
\end{equation}
where $i$ denotes the observed star, $\sigma_{CCD,i}$ is a standard uncertainty in CCD aperture photometry, $s$ is the ratio of rms flux variation to the mean flux due to turbulence in the atmosphere (\citealt{1963AJ.....68..395R}; \citealt{1967AJ.....72..747Y}) and $N_i$ is the number of photons received from the star. The parameter $s$ is a function of air mass, height of the observatory, aperture of the telescope and exposure time. Therefore it should have approximately the same value for every star in the exposure. The uncertainty of the CCD photometry is given by the standard CCD equation (see e.g. \citealt{2000hccd.book.....H})
\begin{equation}\label{ccdeq1}
\sigma_{CCD,i}^2= N_i+n_{bins}\left( 1+\frac{n_{bins}}{n_{sky}}\right)(N_S+N_R^2+N_D),
\end{equation}
where $n_{bins}$ is the number of pixels used to estimate the flux of the star, $n_{sky}$ is the number of pixels used to estimate the level of background, $N_S$ is the number of photons coming from the background, $N_R^2$ is the number of electrons from the readout noise and $N_D$ is the number of electrons from the dark current. We carried out our photometry in such a way that $n_{bins}$ and $n_{sky}$ was the same for every star. Moreover, we assume that $N_R^2$ and $N_D$ are equal for every star observed, which is a good approximation for a modern CCD camera. Similarly, we assume that $N_S$ is the same for every star observed. We stress that this approximation is valid only for small field of view with negligible vignetting. We can rewrite equation (\ref{ccdeq1}) in a simpler form
\begin{equation}
\sigma_{CCD,i}^2= N_i+\xi.
\end{equation}
Equation \ref{sigma1} can now be rewritten as
\begin{equation}\label{ccdeq2}
\sigma_i^2=N_i+\xi+s^2\times N_i^2
\end{equation}
and rearranged to calculate $\xi$:
\begin{equation}\label{xieq}
\xi=\sigma_i^2-N_i-s^2\times N_i^2
\end{equation}
In order to estimate $\xi$ and $s^2$ we will use two standard stars, therefore we will change the $i$ index to 1 and 2 accordingly. We rewrite the equation~(\ref{xieq}) for the first and second standard star and compare the right sides of the equations
\begin{equation}
\sigma_1^2-N_1-s^2\times N_1^2=\sigma_2^2-N_2-s^2\times N_2^2
\end{equation}
After some algebra this gives
\begin{equation}\label{s2eq}
s^2=\frac{(\sigma_2^2-\sigma_1^1)-(N_2-N_1)}{N_2^2-N_1^2}
\end{equation}
combining equations (\ref{s2eq}), (\ref{xieq}) and (\ref{ccdeq2}), and substituting $i$ in the equation (\ref{ccdeq2}) to $v$, for the variable star, we get
\begin{equation}\label{verr}
\sigma_v^2=N_v-N_1+\sigma_1^2 + (N_v^2 -N_1^2 ) \frac{(\sigma_2^2-\sigma_1^2)-(N_2-N_1)}{N_2^2-N_1^2}
\end{equation}
This equation can be used to compare expected uncertainty of flux to the measured one in order to detect flickering.
\begin{table}
\centering
\caption{Summary of the flickering observations. The $\sigma_{\mathrm{measured}}/\sigma_{\mathrm{expected}}$ ratio is the ratio of measured error of mean photon count to the one calculated from Eq.~(\ref{verr}). }\label{tableflickering}
\begin{tabular}{|ccc|}
\hline
Date & 03.08.2013 & 04.08.2013\\
MJD--mid & 56507.813 & 56508.839 \\
UT start--end & 19:07:36--21:31:30 & 18:41:10--20:20:30\\
mean U [mag] & 12.01 & 11.90 \\
U max--min [mag] & 11.839--12.132 & 11.729--12.036 \\
$\sigma_{\mathrm{measured}}/\sigma_{\mathrm{expected}}$ & 1.2 & 2.4\\
\hline
\end{tabular}
\end{table}
\subsection{X--ray observations}
Observations of T~CrB in the X--ray domain were carried out using the Rossi X--ray Timing Explorer (\citealt{1993A&AS...97..355B}; RXTE) Proportional Counter Array (\citealt{1996SPIE.2808...59J}; PCA). The PCA consists of five proportional counter units (PCUs). PCU0 and PCU1 are known to have lower quality observations at the end of the RXTE mission, therefore we used data from PCU 2, 3 and 4. Moreover we analyse data only from layer 1 of the PCUs, because the accuracy of background modelling for layers 2 and 3 is much lower. The observations were downloaded from the High-Energy Astrophysics Virtually ENlightened Sky database\footnote{www.isdc.unige.ch/heavens} \citep[HEAVENS;][]{2010int..workE.162W}. We extracted a light curve with a bin size equal to the exposure time of an individual observation for studying the long-term variability. Also extracted was a light curve with a bin size of 120s that we used for searching for short time variability. The bin size was selected in such a way that the lowest signal to noise ratio in 2-5~keV range at an individual poiting was hihger than 4.
A log of observations is presented in Table~\ref{xraylog}. Additionally, we used $Swift$ observations that were published by \citet{2009ApJ...701.1992K}. The long term variability of T~CrB in X--rays compared to the optical observations is presented in Fig.~\ref{TCrB_all}. A phase plot of RXTE data is presented in Fig~\ref{TCrB_rxte_phase}. The light curve with 120s bins from each individual pointing is presented in Figs.~\ref{rxte_flickering_individual}~and~\ref{rxte_flickering_individual2}.
\begin{table}
\centering
\caption{Log of RXTE PCA observations.}\label{xraylog}
\begin{tabular}{|cccc|}
\hline
Observation ID & Date & MJD & Exposure [ks] \\
\hline
93007-03-01-00 & 02.07.2007 & 54283 & 3.045 \\
93007-03-02-00 & 16.08.2007 & 54329 & 2.980 \\
93007-03-03-00 & 23.09.2007 & 54367 & 3.345 \\
93007-03-04-00 & 02.11.2007 & 54407 & 3.105 \\
93007-03-05-00 & 14.12.2007 & 54449 & 2.965 \\
93007-03-06-00 & 25.01.2008 & 54490 & 2.926 \\
93007-03-07-00 & 07.03.2008 & 54533 & 3.212 \\
93007-03-08-00 & 20.04.2008 & 54576 & 3.010 \\
93007-03-09-00 & 30.05.2008 & 54617 & 3.059 \\
93007-03-10-00 & 11.07.2008 & 54658 & 3.694 \\
93007-03-11-00 & 22.08.2008 & 54700 & 2.914 \\
93007-03-12-00 & 03.10.2008 & 54742 & 2.975 \\
93007-03-13-00 & 14.11.2008 & 54785 & 2.805 \\
94007-03-01-00 & 26.12.2008 & 54826 & 2.970 \\
94007-03-02-00 & 09.02.2009 & 54871 & 3.073 \\
94007-03-03-00 & 23.03.2009 & 54914 & 2.915 \\
94007-03-04-00 & 02.05.2009 & 54954 & 2.859 \\
94007-03-05-00 & 13.06.2009 & 54995 & 3.414 \\
94007-03-06-00 & 25.07.2009 & 55038 & 3.338 \\
94007-03-07-00 & 06.09.2009 & 55080 & 2.819 \\
94007-03-08-00 & 16.10.2009 & 55120 & 2.820 \\
94007-03-09-00 & 28.11.2009 & 55163 & 3.602 \\
\hline
\end{tabular}
\end{table}
\begin{figure}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/TCrB_rxte_phase}}
\caption{Phase plot of RXTE PCA observations. The orbital phase was calculated using the ephemeris from \citet{1988AJ.....95.1505L}. }
\label{TCrB_rxte_phase}
\end{figure}
\begin{figure}\centering
\includegraphics[width=.5\textwidth]{./figs/rxte_flickering/54490}\\
\includegraphics[width=.5\textwidth]{./figs/rxte_flickering/54616}
\caption{Selected observations of flickering in T~CrB in X--rays. The colours are the same as in Fig.~\ref{TCrB_all}.}
\label{rxte_flickering_individual}
\end{figure}
We attempted spectral analysis of the data using XSPEC version 12.9.0. Following the work of \citet{2009ApJ...701.1992K} and \citet{2014MNRAS.437..857E} a single-temperature bremsstrahlung emission together with a Gaussian line to account for the Fe~K$\alpha$ fluorescence line subject to full-covering and partial-covering absorption were employed. The obtained fits were very good (reduced $\chi^2\simeq0.95$), but physical parameters from our fit had errors too large for an in-depth analysis. Nonetheless, in order to investigate the spectral evolution of the star we show spectra from two epochs, i.e., the first epoch was set before MJD~54650, in which T~CrB showed flares in the soft band (see Fig.~\ref{TCrB_all}) of the spectrum; the second epoch was set after MJD 54650 where a flare in the hard band (see Fig.~\ref{TCrB_all}) of the spectrum occurred, while the soft component remained quiet.
The results of the spectral fitting are listed in Table~\ref{xrayfitparams} and shown in Fig.~\ref{rxte_fitplot}, indicating a possible tail of a soft excess in the spectrum at energies below 3~keV. This excess was not detected in previous observations of T CrB (\citealt{2008ASPC..401..342L}; \citealt{2009ApJ...701.1992K}; \citealt{2013A&A...559A...6L}; \citealt{2014MNRAS.437..857E}). If confirmed, it would change the classification of this system to $\beta/\delta$ type according to \citet{1997A&A...319..201M} and extended by \citet{2013A&A...559A...6L}.
\begin{table*}
\centering
\caption{Spectral analysis from the RXTE/PCA spectra of T CrB using a single-temperature, absorbed bremsstrahlung spectral model with a Gaussian line to account for the Fe~K$\alpha$ fluorescence line.}\label{xrayfitparams}
\begin{tabular}{|cccccccccc|}
\hline
Epoch & $N_H$(FC)$^{\rm a}$ & $N_H$(PC)$^{\rm b}$ & PCF$^{\rm c}$ & kT & Line Peak & Line Width & $\chi^2$/(d.o.f.) \\
& [$\times10^{22}$cm$^2$]& [$\times10^{22}$cm$^2$] & & [keV] & [keV]& [keV] \\
\hline
MJD<54650 & <1.7 & 30.5$^{+12.7}_{-11.4}$ & 0.80$_{-0.08}^{+0.06}$ & 31$_{-15}^{+84}$ & 6.53$_{-0.14}^{+0.15}$ & 0.28$_{-0.28}^{+0.35}$ & 0.98\\
MJD>54650 & <3.2 & 41.7$^{+17.1}_{-14.4}$ & 0.88$_{-0.05}^{+0.05}$ & 52$_{-34}^{+52}$ & 6.53$_{-0.15}^{+0.15}$ & 0.21$_{-0.21}^{+0.38}$ & 0.93\\
\hline
\end{tabular}
\raggedright
\textbf{Notes}: $^{\rm a}$~Column density of the material covering the whole source; $^{\rm b}$~Column density of the material partially covering the source; $^{\rm c}$~Partial covering fraction.
\end{table*}
\begin{figure}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/spec_fit}}
\caption{Top panel: Folded model with an absorbed bremsstrahlung plus a Gaussian (solid lines) and observed spectra (points) for MJD<54650 (black) and MJD>54650 (red). Bottom panel: the corresponding residuals.}
\label{rxte_fitplot}
\end{figure}
\section{Results}
\subsection{Active phases}\label{active_phase_section}
\subsubsection{Photometric variability}\label{photvar_sec}
The active phases of the hot component in T~CrB have been studied for many years (see e.g. \citealt{2004A&A...415..609S} and references therein). However, the recurrent time-scale of the active phases is not clear. On the basis of 40 years of visual observations, \citet{1997MNRAS.287..634L} estimated that the active phases occur with a period of $\sim$9840d. However, the observations analysed by \citet{1997MNRAS.287..634L} were done by eye and therefore may have been more sensitive to variability of the RG, which dominates the visual spectral range. Moreover, the $\sim$9840d period was not confirmed by more recent observations, most likely due to the observations covering an insufficient time baseline. Thus, the $\sim$9840d period remains uncertain.
Observations collected with the International Ultraviolet Explorer (IUE) during 1978--1990 revealed two minima in the UV continuum separated by $\sim$5000d \citep{1992ApJ...393..289S}. Similarly, the $U$ light curve in 1982--2001 shows two maxima with an amplitude of $\sim$2~mag separated by $\sim$5000d \citep{2004A&A...415..609S}. In addition, variability in the emission line fluxes with a similar period was reported by \citet{1997ppsb.conf..117A}. Therefore, we conclude that the most prominent active phases in T~CrB occur with a period of $\sim$5000d (hereafter `big active phases').
Additional activity of the hot component was detected with a period of $\sim$1000d by \citet{1997IBVS.4461....1Z}, where the authors observed three brightenings in the $U$ filter with an amplitude of $\sim$1~mag separated by $\sim$1000d at MJD 48100, 49100 and 50200. The variability reported by \citet{1997IBVS.4461....1Z} was however during a minimum of the big active phase \citep[see e.g. fig.~4 in][]{2004A&A...415..609S}. Variability of a similar time-scale was detected in the emission lines \citep{1997ppsb.conf..117A}. Therefore active phases with a smaller amplitude and a time-scale of $\sim$1000d are present as well (hereafter `small active phases'). The small active phases have had been observed only during a minimum of the big active phases (e.g. in the UV continuum in observations presented by \citealt{1992ApJ...393..289S} or in $U$ photometric observations presented by \citealt{2004A&A...415..609S}). This suggest that both the big and small active phases are connected.
The variability with a period of $\sim$1000d is also present in our observations (Fig. \ref{TCrB_all}; Fig~\ref{TCrB_powerspec}). This variability is particularly prominent in $B$ as expected for activity associated with the hot component. The amplitude and the time-scale are consistent with the small active phases.
During most of the period of our observations the magnitude of T~CrB, $B \ga 11$, i.e. the system has remained in a minimum of the big active phase \citep[see e.g. fig.~4 in][]{2004A&A...415..609S}, and only after MJD~57100 did the system begin to brighten. This suggests that our observations in the $B$ filter mainly covered a period of small active phases and ended at the beginning of a rise to the maximum of a big active phase. This rise is corroborated by archival Swift UV photometry that also show T~CrB brightening during the same period (see Appendix \ref{sec:uvot}).
\subsubsection{Emission line variability}\label{specsection}
The correlation between H$_\alpha$ fluxes and the active phases was discussed by \citet{1990JAVSO..19...28I}, \citet{1991MNRAS.253..605A}, \citet{1999A&A...344..177A} and \citet{2004A&A...415..609S}. The authors reached a conclusion that the line strengths are correlated with the brightness of the hot component. Most notable was the variability of emission lines with a period of 3640d and 906d \citep{1997ppsb.conf..117A}.
In the past, the most thorough studies of emission line variability concentrated on their equivalent widths (EW) \citep[e.g.][]{1997ppsb.conf..117A}, while the studies of emission line fluxes were much more sporadic \citep[e.g.][]{1999A&A...344..177A}. The downside of using EW is that the continuum can be influenced by hot component variability, especially at short wavelengths (e.g. around \mbox{He\,{\sc ii}\,4686}). In this study, we pay more attention to the variability of emission line fluxes.
From Fig.~\ref{all_fluxes} and Fig~\ref{TCrB_all} it is clear that the fluxes of all of the measured emission lines are correlated with the photometric observations, which is consistent with the previous studies. The EW of H$_\alpha$ during the first part of our observations was $<15\AA$, which points to the minimum of the big active phase \citep[see e.g.][]{2004A&A...415..609S}. The increase of the emission line fluxes and their EW at the last observations suggest we are approaching the maximum of a big active phase, which was already evident in photometric observations.
The ratio of F(H$_\alpha$)/F(H$_\beta$) changes from $\sim 3$ to 13. The reddening free ratio F(H$_\alpha$)/F(H$_\beta$) $\sim 3-10$ is typical for SySt, and it is due to self-absorption effects \citep{1996ApJ...471..930P,1997A&A...327..191M}. In particular, F(H$_\alpha$)/F(H$_\beta$) showed a short maximum around MJD 56750. Simultaneously there was a minimum in both photometric observations and line fluxes. Moreover, the observed increase of the F(Ha)/F(Hb) ratio was during a minimum of the small active phase as well. These changes are most likely due to variable self-absorption and optical depth effects as indicated by the Ha and Hb profiles being heavily affected by an absorption component seen in most SySt (see e.g. \citealt{1993A&AS..102..401V}).
\subsubsection{X--ray variability}\label{xraysection}
In the case of X--ray emission the variability in the soft range (2--5 keV) is not correlated with variability in the hard range (10--20 keV; Fig.~\ref{TCrB_hardness}). Furthermore, the variability in the intermediate range (5--10 keV) seems to be a superposition of the variability in the hard and soft range. This indicates two separate sources of the X--ray variability. In line with this interpretation is the variability in both hard and soft X--rays of T CrB reported by \citet{2009ApJ...701.1992K}. However, their observations in both spectral ranges were not contemporaneous, so they cannot directly confirm our interpretation.
The variability in the soft range is possibly related to changes in the column density of absorbing material or variable partial-covering fraction, as was observed in RT~Cru and suggested for T~CrB by \citet{2009ApJ...701.1992K}. This kind of variability would be observed in the soft and intermediate ranges, but would not affect the hard part of the spectrum, which is consistent with the X--ray light curve. On the other hand, variability in the hard range would reflect variability of the source of X--ray radiation. The X--ray photons in T~CrB are thought to originate in the boundary layer between the accretion disc and the WD \citep{2008ASPC..401..342L}, so this band may be tracing variable accretion.
It is clear that the X--ray variability is unrelated to the orbital phase (Fig~\ref{TCrB_rxte_phase}) and to UV changes. Most probably the X--ray variability is related to the active phases discovered in the visible bands. Due to the short time coverage of the RXTE observations it is impossible to study the periodicity of the changes in X--rays over the duration of the big and small active phases.
In the RXTE observations there are two maxima in the soft band at MJD~54329 and MJD~54533, followed by a period of relative quiescence (Fig.~\ref{TCrB_all}). The hard part of the spectrum showed a shallow minimum at MJD~54407 between these two maxima of the soft component. Moreover, in the hard band around MJD~54742 there was a minimum followed by a maximum around MJD 54871. In the time during the maximum in the hard band there were no significant changes in the soft band. Generally it seems that the maxima in the soft band are during a quiescence in the hard band and vice versa, but the number of observed maxima is too low for a meaningful conclusion to be drawn. If this hypothesis holds, then this could mean that the variability in both bands is a different phase of the same phenomenon, such as e.g. disc instability or increased mass loss from the RG.
In order to investigate the possible periodicity in the X--ray variability we utilised Swift observations from \citet{2009ApJ...701.1992K}. These observations are in the band 14--24~keV and therefore the changes in the count rate are associated with the hard band of RXTE. The count rate during the Swift observations showed similar behaviour as in the RXTE observations around MJD 54800. Namely it showed a period of relative quiescence with $7\times10^{-5}$ counts/s/detector followed by a short minimum with $5\times10^{-5}$ counts/s/detector succeeded by an immediate rise to the maximum reaching $12\times10^{-5}$ counts/s/detector.
The similarity of behaviour in the Swift and RXTE observations could point to a periodic or quasi-periodic nature of the variability. The two maxima in the Swift and RXTE observations are separated by $\sim$1200d. Moreover there is only one maximum in the hard band during the period of 880d covered by RXTE. This hints that the observed variability is associated with small active phases discussed in section~\ref{photvar_sec}. On the other hand, two maxima in the soft band took place around the time of a minimum of small active phases. Therefore the soft and hard X--ray variability could have the same period, but with a phase shift. The whole period covered by Swift and RXTE was during a quiescent state of the big active phases.
\begin{figure}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/TCrB_hardness}}
\caption{Relation between the count rate in the hard and soft parts of the X--ray spectrum in RXTE observations.}
\label{TCrB_hardness}
\end{figure}
\subsection{Flickering}\label{flickering_section}
\subsubsection{Optical observations}
\citet{2004MNRAS.350.1477Z} showed that the amplitude of flickering in T~CrB is correlated with the flux of the blue part of the spectrum. A similar relation was discovered in other symbiotic stars -- CH~Cyg \citep{1990AcA....40..129M}, MWC~560 \citep{1996A&AS..116....1T} and RS~Oph \citep{2015MNRAS.450.3958Z}. Most of the observations used by \citet{2004MNRAS.350.1477Z} were obtained during a period of maxima of two big active phases and only two from the quiescent epoch (section~\ref{photvar_sec}) and only two from the time of quiescence (see their fig.~2). Therefore, it is clear that the flickering amplitude changes are connected with big active phases, but it is not clear how a small active phase can influence the flickering.
Interestingly, the observations reported by \citet{2010ATel.2586....1Z} from MJD 54851 and by \citet{2015AN....336..189Z} from MJD 55316 are from the time of a minimum of the big active phase variability, but the amplitude of the flickering was $\sim0.4$~mag in $U$, which is larger than the mean amplitude of the flickering reported by \citet{2004MNRAS.350.1477Z} observations ($\sim$0.35~mag in $U$), but significantly lower than their maximum amplitude (0.6~mag in $U$). At these times the system was close to the maximum of a small active phase. Moreover, \citet{2015AN....336..189Z} reported a flickering with a very low amplitude of 0.17~mag in $U$ at MJD 55603, which was during the minimum of a small active phase. This amplitude is lower than any of the reported by \citet{2004MNRAS.350.1477Z}, where the lowest amplitude was 0.22~mag. Similarly, the amplitudes of the flickering of $\sim$0.3~mag we determined in section~\ref{sec:phot} are also from the time of another quiescent state of a big active phase and is close to the lower values reported by \citet{2004MNRAS.350.1477Z}. While the scarcity of the flickering observations during the minimum of big active phases is making the conclusions uncertain, the observations suggest that the small active phases follow the same flux--amplitude of the flickering relation as during the big active phases. This points to the same physical process behind the big and small active phases.
\subsubsection{X--ray observations}
X--ray flickering in T~CrB was discovered by \citet{2008ASPC..401..342L}, who reported variability on time scales of minutes. We discovered similar flickering in most of the RXTE observations. From the individual lightcurves it is clear that the amplitude of the flickering is larger in the hard band than in the soft band, where it is marginally detected (Fig.~\ref{rxte_flickering_individual},\ref{rxte_flickering_individual2}). This is consistent with the hypothesis that the hard band traces variability of accretion through the boundary layer, where the flickering is expected to originate.
\citet{2015A&A...579A..50B} discussed a time lag in the observations of flickering at different wavelengths in CVs on the basis of optical observations. The author discovered that in most systems the flares reach their maxima slightly earlier at shorter wavelengths. Our observations had too short a timebase for a meaningful cross--correlation analysis. At some epochs the flickering in the 5--10~keV range seems to be correlated with the flickering in 10--20~keV range within our time resolution. The best examples are observations at MJD 54490 and MJD 55163 with correlation coefficients between the 10--20~keV and 5--10~keV flux of r=0.83 and r=0.76, respectively. On the other hand, on MJD 54616 there was a flare in the 10--20~keV range which reached a maximum $\sim$12~min after the start of the observation and the corresponding flare in the 5--10~keV range reached its maximum $\sim$2~min later. On MJD 54407, a similar lag of $\sim$2~min was observed in a flare that reached maximum in the 10--20~keV range $\sim$42~min after the start of observation.
At other times the variations in the hard range are not correlated at all with the flickering in 5--10~keV range (see e.g. observations at MJD 54826). Most interestingly at some times the variations in the 5--10~keV range and the 10--20~keV range are anti--correlated. This can be seen e.g. between 30 and 35~min of observations at MJD 54913 (r=0.25) or in two brightenings at MJD 54616 (r=0.02) between 20 and 40~min. This suggests a change in the hardness ratio of the flickering rather than the total amplitude.
Overall the flickering seems to be most prominent in the hard band, but the hardness ratio of flares can change during the observing run. Some flares can have very hard spectra and are only observed in the 10--20~keV range (e.g. at MJD 54826 with r=0.27) and in some rare cases they can have softer spectra and are observed in the 2--10~keV range and not in the 10--20~keV range (see e.g. a flare in the first 5~min of observations at MJD 54995). This seems to be consistent with the fact that in the optical observations the source of the flickering has a different colour at different times \citep[see e.g. tab~2 in][]{2015AN....336..189Z}. Nevertheless, the source of the flickering in optical observations and in the X--rays could not be the same since in the optical the source typically has a temperature of $\sim$9000~K in symbiotic recurrent novae \citep{2015AN....336..189Z}, which seems consistent with reprocessing in the surrounding \mbox{H\,{\sc ii}} region.
We studied the total flux--flickering amplitude relation in the X--ray observations. For this we plotted an average count rate (F$_{\mathrm{AV}}$) versus the average amplitude of the flickering (F$_{\mathrm{FL}}$=F$_{\mathrm{max}}$--F$_{\mathrm{min}}$). We calculated the Pearson's correlation coefficient $r$ and probability $p$ for a hypothesis test whose null hypothesis is that the slope of the relation is zero. The calculated values are listed in Tab.~\ref{X--ray_flickering_tab} and the data together with the fit are shown in Fig.~\ref{flickering_amplitude}. From the analysis it is clear that the correlation is present in the 5--10~keV and 10--20 keV range and absent or marginally present in the 2--5~keV range. The correlation is stronger in the 5--10 keV range rather than in the 10--20 keV range. We attribute this to the fact that the fit was carried out to data spanning over a larger range of F$_{\mathrm{AV}}$ in the 5--10 keV range rather than in the 10--20 keV range. The best fit is obtained using data points from all of the spectral ranges simultaneously. This suggests that the total flux--flickering amplitude relation is the same over the whole spectral range. Since the count rate is proportional to the physical flux we can compare the slope of the fit to a similar fit in the optical range. The slope in the $U$ band for the F$_{\mathrm{AV}}$--F$_{\mathrm{FL}}$ relation is 0.156$\pm$0.042 for T~CrB \citet{2004MNRAS.350.1477Z}. Since the values for optical and X--ray observations are not consistent this is another indication that the flickering in those two wavelength bands originates from a different place. However, we note that if the relation is logarithmic rather than linear, then the optical and X--ray observations may follow the same relation \citep{2016MNRAS.457L..10Z}
There are only two observations of flickering in the optical range during the period covered by X--ray observations and unfortunately none of them were simultaneous with the X--ray observations. \citet{2010ATel.2586....1Z} observed flickering at MJD 54851 and MJD 54889 with amplitudes in $U$ 0.404~mag and 0.303~mag respectively, which are rather high for flickering outside of the big active phase in T~CrB. These observations were taken close to the maximum of the hard X--ray emission, which seems to confirm that the hard X--ray and optical flickering are connected.
\begin{figure}\centering
\resizebox{\hsize}{!}{\includegraphics{./figs/flickering_amplitude}}
\caption{The relation between brightness and flickering amplitude in the X--ray observations. The error is the mean error in count rate measure during one pointing of the telescope. The solid line is a linear fit to the data from a spectral range represented by the same colour. Black dashed is a linear fit to data points from all spectral ranges simultaneously. The colours are the same as in Fig.~\ref{TCrB_all}.}
\label{flickering_amplitude}
\end{figure}
\begin{table}
\caption{Quantities of the linear fit to the F$_{\mathrm{av}}$--F$_{\mathrm{fl}}$ relation in the X--ray observations. The combined fit is a simultaneous fit to data points from all of the spectral ranges.}
\label{X--ray_flickering_tab}
\begin{tabular}{|ccccc|}
\hline
Range & 2--5 kev & 5--10 keV & 10--20 keV & Combined \\
slope & 0.08$\pm$0.08 & 0.6$\pm$0.1 & 0.8$\pm$0.3 & 0.90$\pm$ 0.04\\
$r$ & 0.22 & 0.80 & 0.55 & 0.93 \\
$p$ & 0.32 & 8.0$\times$10$^{-6}$ & 7.9$\times$10$^{-3}$ & 3.7$\times$10$^{-30}$\\
\hline
\end{tabular}
\end{table}
\section{Discussion}\label{discussion_sec}
\subsection{Active phases}
Similar active phases to those observed in T~CrB have been discovered in the other SyRN RS~Oph \citep{2008ASPC..401..219G}. Namely, the authors observed brightenings of the system on time-scales of 1200--1800d with an amplitude of $\sim$1~mag. \citet{2008ASPC..401..219G} pointed out that the behaviour of RS~Oph outside of the nova outbursts resembles the classical Z~And variability, which is thought to be a result of an instability in the accretion disc (\citealt{2003ASPC..303....9M}; \citealt{2006ApJ...636.1002S}). The active phases in T~CrB and RS~Oph could have the same origin as in the Z~And type symbiotic systems with the difference that in the Z~And type systems there is steady hydrogen burning on the surface of the WD.
If the active phases in T~CrB and RS~Oph are driven by the same mechanism as the multiple-outburst activity of Z~And type SySt, similar time-scales should be expected. In fact, outbursts in many Z~And type systems recur with quasi-periods of one to tens of years.
Namely, the major outburst in AG~Dra occurs every 12--15 years, while minor outbursts are observed roughly once a year \citep{2016MNRAS.456.2558L}. Z~And showed a time-scale of $\sim$8400d of the outburst activity \citep{1994A&A...292..534F} and the interval between outbursts in BF~Cyg is $\sim$6376d \citep{2006MNRAS.366..675L}. Given the differences between the systems these time-scales are in rough agreement with the ones in T~CrB and RS~Oph.
A useful comparison can also be made based on similar physical parameters of these systems. In the prototype star Z~And, after assuming M$_{\mathrm{WD}}\sim$0.65M$_\odot$ \citep{1997A&A...327..219S}, a hot component mass function of f(m)=0.0240M$_\odot$ \citep{2000AJ....120.3255F}, an inclination of i=$41^\circ$ \citep{2010AJ....140..235I}, and orbital period of 759d \citep{2000AJ....120.3255F}, the radius of the Roche Lobe around the WD is $\sim$140R$_\odot$. Similarly, for T~CrB, assuming M$_{\mathrm{WD}}=1.2$M$_\odot$ and M$_{\mathrm{RG}}=0.8$M$_\odot$ \citep{1998MNRAS.296...77B,2004A&A...415..609S}, and an orbital period of 227.67d \citep{1988AJ.....95.1505L}, the radius of the Roche Lobe around the WD is $\sim$80R$_\odot$. In the case of RS~Oph, taking M$_{\mathrm{WD}}=1.3$M$_\odot$, M$_{\mathrm{RG}}=0.74$M$_\odot$, and an orbital period of 453.6d \citep{2009A&A...497..815B}, the radius of the Roche Lobe around the WD is $\sim$135R$_\odot$. The estimated mass transfer rate in T~CrB is 2.5$\times 10^{-8}$~M$_\odot$/yr \citep{1992ApJ...393..289S}, whereas in Z~And the mass transfer rate varies between 4.5$\times 10^{-9}$~M$_\odot$/yr during quiescence \citep{1988ApJ...324.1016F} to 3.2$\times 10^{-7}$~M$_\odot$/yr during the active phase \citep{2004A&A...428..985T}. Therefore, the mean mass transfer rates are of the same order of magnitude in both systems.
The ratio of Roche Lobe radii around the WDs in Z~And and T~CrB is $\sim$1.5. It is interesting to note that this value is similar to the ratio of $\sim$1.7 between time-scales of Z~And outburst activity and big active phases in T~CrB. On the other hand, the ratio of Roche Lobe radii around the WDs in RS~Oph and T~CrB is $\sim$1.7, which is similar to the ratio of $\sim$1.5 between time-scales of active phases in RS~Oph and small active phases in T~CrB. This is to be expected, because the time-scale of a disc instability scales with the size of the accretion disc, which in turn is proportional to the radius of the Roche Lobe around the WD. This is more consistent with the longer time-scale of outburst activity in Z~And than the time-scale of active phases in T~CrB.
Recently, disc instability in RS~Oph has been modeled by \citet{2011MNRAS.418.2576A}, who found recurring outbursts in the system on 10--20 yr time-scales, which is comparable to the time-scales of active phases actually observed in RS~Oph and Z~And \citep{2008ASPC..401..219G}. In some models alternating small and big outbursts take place. This could be associated with big and small active phases in T~CrB.
\citet{1997ppsb.conf..117A} proposed a solar--type cycle resulting in a variable mass transfer as an explanation of the active phases in T~CrB. In order to obtain solar-like cycles one must have a high magnetic field. In the case of T~CrB this is probable since it is argued that in SySt the RG is more magnetically active than a single RG \citep{2002MNRAS.337.1038S}. In general, the hypothesis that the RG is a source of the observed activity is particularly interesting since variability with a time-scale of $\sim$1000d was observed in a single semi--regular variable WZ~Cas \citep{2005A&A...440..295L}.
Overall it seems to be clear that at least the big active phases are related to a change in mass transfer rate. Therefore it is important to determine the exact nature of the T~CrB variability, especially since SySt are thought to be promising SNIa progenitors \citep{2013IAUS..281..162M}.
Recently, \citet{2016NewA...47....7M} have claimed that during 2015 T~CrB displayed ''super-active conditions never seen before''. They have also suggested that this super-active state is a new form of activity and it is different from the active phases observed e.g. by \citet{1990JAVSO..19...28I}. We argue that this is not true, and the activity observed by \citet{2016NewA...47....7M} is just a new maximum of the big active phases (see Sec.~\ref{active_phase_section}) and similar activity has been observed in the past.
\citet{2016NewA...47....7M} based their claim on a large increase in emission line fluxes and optical magnitudes of the system. However, as the authors noted, it is difficult to study the history of emission line variability, since the observations are sparse and the authors rarely have given line fluxes. Here we expand the analysis of \citet{2016NewA...47....7M} using EWs. From Fig.~\ref{all_fluxes} one can see that variability of emission line fluxes is well represented by variability of EW(H$\alpha$). In particular, the $H\alpha$ emission line flux, estimated on the spectrum from MJD~57447, is F($H\alpha$)=4.5$\times$10$^{-11}$~erg~cm$^{-2}$~s$^{-1}$ , and the corresponding corresponding EW($H\alpha$)=48.8. This is much higher than the maximum flux F($H\alpha$)=3.9$\times$10$^{-11}$~erg~cm$^{-2}$~s$^{-1}$ observed by \citet{2016NewA...47....7M}. Therefore we adopt EW($H\alpha$)=48.8 as a maximum EW observed during the supposed super-active conditions.
During the past big active phases, the EW($H\alpha$) varied from $\sim$5 to $\sim$40 \citep[see fig.~4 in][]{2004A&A...415..609S}. Moreover, the maximum EW observed during the 1989 big active phase, EW(H$\alpha$)$\sim$40, was bigger than that observed during the 1997 big active phase (EW(H$\alpha$)$\sim$30). This shows that the amplitude of emission line variability changes, which is consistent with the quasi-periodic nature of the known phenomenon rather than a new form of activity. This is further supported by the photometric variability observations reported by \citet{2016NewA...47....7M} with timescales and amplitudes consistent with previous big active phases \citep[see fig.~4 in][]{2004A&A...415..609S}.
\subsection{Flickering}
Since the flickering in T~CrB is mainly observed in the hard X--rays, which are thought to originate in the accretion disc boundary layer, our study seems to confirm the model of flickering originating from an unstable mass transfer through the boundary layer. However, it seems that more than one mechanism is responsible for the observed variability. The events in which there is brightening in the 5--10~keV range and fading in the 10--20~keV range seem to indicate Compton cooling. Moreover, it is obvious that the flickering observed in the optical range is not the same as that observed in the X--rays since the optical colours of the flickering source in T~CrB are consistent with a blackbody of $\sim$9000~K \citep{2015AN....336..189Z}. Nevertheless, it seems that flickering in the hard X--rays is related to the optical flickering since the maximum in hard X--rays was observed close to the maximum of the optical $\sim$1000d active phase, and for both wavelengths the amplitude of the flickering--mean flux relation holds. Our results seem to be best reproduced by a model proposed by \citet{2014MNRAS.438.1233S} in which the flickering observed in the visible range comes from X--ray reprocessing by a geometrically thick disc. This model provides a natural explanation for the difference between the flickering source colours in SySt and classical CVs \citep{2015AN....336..189Z} as the accretion disc in SySt is thought to be bigger than in CVs. Reprocessing of the X--ray radiation by the nebula is in principle not excluded, but reprocessing by a thick accretion disc seems likely since the UV to IR SED of the system was successfully modelled by \citet{2004A&A...415..609S} using a model containing an optically thick accretion disc. On the other hand, reprocessing of the radiation by a nebula would be consistent with flickering observed in emission lines of SySt \citep{2005PASP..117..268Z,2014MNRAS.442.2637W}.
\section{Summary}\label{sumarysec}
We analysed optical, UV and X--ray observations of active phases and flickering in T~CrB. The main results of our work include:
\begin{itemize}
\item We found that there are big and small active phases of the hot component. Big active phases occur with a period of $\sim$5000d, while small active phases occur with a period of $\sim$1000d during a quiescence of the big active phases.
\item Variability in the soft part and hard part of the X--ray spectrum of T~CrB is not correlated. Variability in the soft band is consistent with a variable amount of absorbing material in the line of sight. Variability of the hard part of the spectrum most probably traces variable accretion through the boundary layer. Variability in both spectral ranges is not correlated with orbital phase, but seems to be correlated with active phases seen in the optical.
\item The flickering seems to be present only in the hard X--ray component. The flickering in this spectral range follows a similar mean flux--amplitude of the flickering relation as the one observed in the optical.
\item The different behaviour of flickering in X--rays observed at different times seems to point at more than one physical process behind flickering. The most likely scenario suggested by our observations is formation of the flickering in the boundary layer, which is then reprocessed from X--rays to the visible light by the thick accretion disc or a nebula around the system.
\item Time-scale of active phases in T~CrB is similar to time intervals between outbursts in Z~And type SySt. This suggests that the mechanism driving the activity in these systems is the same, presumably unstable disc accretion.
\end{itemize}
To understand the nature of the system further it would be most beneficial to acquire simultaneous X--ray, UV and optical observations of the flickering behaviour. This would help connect the X--ray and optical flickering. Moreover, the long--term X--ray monitoring would help to understand the X--ray behaviour of the system during the active phases which would lead to understanding the nature of the mass transfer variability in the system. This, in turn, could help understand the variability of Z~And type systems.
\section*{Acknowledgements}
We are grateful to all of the amateur astronomers that contributed their observations to this paper. In particular, we are thankful to members of the ARAS group for their wonderful work. We acknowledge with thanks the variable star observations from the AAVSO International Database contributed by observers worldwide and used in this research. We would like to thank the anonymous referee for the constructive comments that improved the quality of our manuscript significantly.
KI has been financed by the Polish Ministry of Science and Higher Education Diamond Grant Programme via grant 0136/DIA/2014/43. This study has been partially financed by Polish National Science Centre grant 2015/18/A/ST9/00746. KS gratefully acknowledge observing grant support from the Institute of Astronomy and Rozhen National Astronomical Observatory, Bulgarian Academy of Sciences.
\bibliographystyle{mnras}
|
\section{Introduction}
Given the importance of star
formation in the evolution of galaxies, understanding the regulation
of star formation is crucial.
Early work on star formation on galaxy
scales relied on empirical star formation laws
\citep{Schmidt:1959,Kennicutt:1989},
with little connection
to the detailed studies of star formation in our own Galaxy.
Recently there has been more focus on integrating the understanding
of the process of star formation from the scale of galaxies to the
much smaller scales of regions within molecular clouds
\citep{Kennicutt:2012,Kruijssen:2014,Krumholz:2014}.
While large scale
studies provide essential information on the relation between
large scale properties of galaxies and star formation, the process of
converting gas into stars takes place on a smaller scale. Since molecular clouds (MCs)
are the sites of
star formation in galaxies, it is essential to establish the key processes and
sequences within molecular
clouds that regulate the production of newborn stars in order to gain a deeper
understanding of processes at galactic scales.
While there are some recent high-spatial
resolution studies of nearby galaxies that can resolve regions of MCs
(e.g., \citealt{2007ApJ...661..830R, 2013ApJ...779...46H}),
the Milky Way offers the highest resolution view
to investigate the connection
between star formation and the local gas properties.
There are several recent surveys of dust and molecular line emission in the
Milky Way that provide information on the distributions and
properties of \gmc s.
Ideally, star formation in \gmc s can be directly evaluated by identifying
stars or young-stellar objects inside the clouds, which along with the
information on their mass and lifetime provide a good estimate of star
formation rate (SFR) for the clouds. This direct method of estimating
SFR has been applied for nearby ($d < 830$ pc) molecular clouds
\citep{Heiderman:2010,Gutermuth:2011,Lada:2010,Evans:2014},
but these have a limited range in properties, making it difficult to test
theories for the importance of cloud properties in controlling star
formation rates. Furthermore, they are primarily low-mass
($\mean{\mbox{$M_{\rm cloud}$}} \sim 3000$ \msun) clouds
\citep{Heiderman:2010}
whose star formation does not fully sample the IMF.
Their star formation activity
would be almost entirely invisible to observers in other galaxies.
The goal of this paper is to extend this effort to larger clouds
where massive stars are formed, both to sample a larger range of
cloud properties and to examine regions more comparable to those that
can be observed in other galaxies.
The challenge is that the more massive clouds with more fully sampled
IMFs are all quite distant; even Orion does not sufficiently sample the
IMF to use the extragalactic indicators of star formation rate
\citep{Kennicutt:2012}.
For those distant clouds, counting YSOs is very difficult, both because
of sensitivity limits and because of background source confusion
\citep{Dunham:2011}, although recent work has been more successful
\citep{Heyer:2016}.
To study star formation in a larger sample of Galactic \gmc s,
we resort to indirect tracers of
SFR, such as those commonly used in extragalactic studies including
H$\alpha$, UV continuum, total infrared luminosity, mid-infrared
emission, and radio continuum emission.
The shorter wavelength tracers (H$\alpha$, UV continuum) cannot be
used in the plane of the Galaxy because of dust obscuration, and
the total far-infrared luminosity awaits full release of surveys
with {\it Herschel}. In this study, we use mid-infrared and radio continuum
emission.
It is known that these indirect tracers derived from extragalactic
data are problematic when applied to smaller regions such as \gmc s.
The problem arises mostly from the assumptions of a fully-sampled
IMF and a star formation history that is constant over a long
timescale
\citep{Kruijssen:2014,Krumholz:2015}.
Several recent studies of SFR tracers
in regions with different properties suggest that some tracers
offer reasonable measures of SFR (although still with large scatter) in regions
above a certain minimum SFR
\citep{Wu:2005,Vutisalchavakul:2013}.
In this paper, we collect data from surveys of radio recombination lines,
radio continuum, and mid-infrared emission to measure SFR,
\mbox{$^{13}$CO}\ spectroscopy to evaluate MC properties, and
millimeter dust continuum emission to trace the dense gas component.
We describe these data
sets in \S \ref{data} and summarize our selection of star forming
regions and their association with gas in \S \ref{analysis}.
Various models for star formation prediction are tested with these data
in \S \ref{tests}. In \S \ref{lowcomp}, we
compare our results to similar studies of nearby clouds, and in
\S \ref{exgal}, we put our results into the context of studies of other
galaxies.
\section{Data}\label{data}
\subsection{Radio Recombination Lines and \mbox{\ion{H}{2}}\ Region Catalog}
\citet{Anderson:2014}, hereafter A14, compiled a catalog of \mbox{\ion{H}{2}}\ regions
within the Milky Way.
The A14 catalog comprises
over 8000 Galactic \mbox{\ion{H}{2}}\ regions and \mbox{\ion{H}{2}}\ region candidates identified
by mid-infrared (MIR) emission morphology using WISE 12 $\mu$m and 22 $\mu$m
data. Many of these sources are associated with radio recombination lines
(hereafter RRLs)
H86$\alpha$ through H96$\alpha$ \citep{Anderson:2011}
or H$\alpha$ emission, confirming their association with \mbox{\ion{H}{2}}\ regions.
\subsection{Radio Continuum Data}
Radio continuum emission closely associated with recent star formation
in the Milky Way
comes from free-free emission of ionized gas around high-mass
stars.
The Very Large Array Galactic Plane Survey (VGPS),
observed the \ion{H}{1} spectral line and 21~cm
radio continuum emission from the
first Galactic quadrant covering the Galactic
longitude between $18^\circ < l < 67^\circ$ with varying Galactic
latitude range from $|b| < 1.3^\circ$ to $2.3^\circ$
at the resolution of 1\arcmin\ \citep{Stil:2006}. To provide sensitivity on
larger scales, short spacing data obtained with the Green Bank Telescope were
added to the interferometric continuum data.
\subsection{Mid-Infrared Data}
The Wide-field Infrared Survey Explorer (WISE)
mapped the entire sky in four IR bands at
3.4, 4.6, 12, and 22 \mbox{$\mu$m}\ \citep{Wright:2010}.
The MIR band at 22 \mbox{$\mu$m}, which has a resolution of 12\arcsec,
provides a measure of star formation rate.
The WISE Image Atlas provides image tiles covering 1.564$^\circ \times $
x 1.564$^\circ$ in area with 1.375\arcsec\ per pixel.
Larger 22 \mbox{$\mu$m}\ image mosaics
covering the segment of the Galactic Plane in this study were generated
from the set of image tiles
using the MONTAGE mosaic software \citep{Jacob:2010}. The
characteristic saturation level for the point sources, defined to be the level
at which 50\% of
the sources have some saturated pixels, in the 22 \mbox{$\mu$m}\ band is
about $12$ Jy.
For a few bright sources that are saturated in the WISE 22 \mbox{$\mu$m}\ band, we
use the data from the Infrared Astronomical Satellite (IRAS)
25 \mbox{$\mu$m}\ band. The IRAS Improved Reprocessing of the IRAS Survey
(IRIS) provides images of the sky at 25 \mbox{$\mu$m}\ band at the resolution
of $\approx 4$\arcmin\ \citep{Miville:2005}.
The resolution of IRAS 25 \mbox{$\mu$m}\ data is considerably
lower than that of WISE 22 \mbox{$\mu$m}\ band; therefore, we use IRAS 25
\mbox{$\mu$m}\ images for sources saturated in WISE 22 \mbox{$\mu$m}\ only if
the source is resolved with no confusion with other nearby
sources.
\subsection{\textsuperscript{13}CO\ \jj10\ Emission}
Data from the Boston University-FCRAO Galactic Ring Survey (GRS) and the Exeter-FCRAO
Galactic Plane Survey (EXFC) provide data on
molecular clouds.
The GRS
surveyed the \textsuperscript{13}CO\ \jj10\ emission between
Galactic longitudes
$18^\circ < l < 55.7^\circ$ and Galactic latitudes, $|b|<1^\circ$
using the Five College Radio Astronomy Observatory
14-m telescope \citep{Jackson:2006}.
The LSR
velocity (\mbox{$v_{LSR}$}) coverage is from $-5$ to 135 km s$^{-1}$ for $l \le 40^\circ$
and from $-5$ to 85 km s$^{-1}$ for $l > 40^\circ$. The final data cubes
are gridded at 22\arcsec\ and at velocity resolution of 0.21 km
s$^{-1}$. The median main beam sensitivity of the GRS is 0.21~K.
The EXFC survey covered \textsuperscript{12}CO\ and \mbox{$^{13}$CO}\ \jj10\ emission within two areas:
$55.7^\circ < l < 102.5^\circ$,
$|b|< 1^\circ$ and $141^\circ < l < 192^\circ$, $-3.5^\circ < b < 5.5^\circ$
with the FCRAO 14~m telescope
\citep{2016ApJ...818..144R}.
The median main beam sensitivity at 110~GHz in the area of this study is 0.46~K.
For both the GRS and EXFC surveys, the half power beam width of the telescope at 110 GHz is 48\arcsec.
All of the \textsuperscript{13}CO\ data used in this study (GRS and EXFC) have been post-processed to remove
contributions from the antenna error beam.
\subsection{Millimeter Dust Continuum Emission}
The dense gas within molecular clouds is examined using
1.1 mm dust emission from the
Bolocam Galactic Plane
Survey (BGPS). The BGPS observed part of the Galactic Plane
at the effective resolution of 33\arcsec\
and $1\sigma$ sensitivity ranging from 30 to 100 mJy per beam \citep{Aguirre:2011, Ginsburg:2013}.
The BGPS provides contiguous
observational data over the ranges of $|b| < 0.5^\circ$ for
$-10.5^\circ < l < 90.5^\circ$, $|b| < 1.5^\circ$ for $75.5^\circ < l < 87.5^\circ$,
and additional coverages over selected regions
described in \citet{Aguirre:2011} and \citet{Ginsburg:2013}.
Owing to the removal of atmospheric signal by spatial filtering,
the data from BGPS recovers most of the astrophysical emission out to
the scale of 80\arcsec\ and partially recovers emission out to the scale of
300\arcsec\ \citep{Ginsburg:2013}. The spatial filtering subtracts signal from
molecular clouds that subtend solid angles greater than 300\arcsec.
For these reasons, the data are sensitive to compact regions within clouds with
enhanced
column densities, and likely, volume densities.
The BGPS version 2.1 data release includes a source catalog extracted from 1.1 mm
maps using a seeded watershed algorithm \citep{Rosolowsky:2010, Ginsburg:2013}.
The Bolocat V2.1 consists of
8594 sources with 1.1 mm flux from photometry
(background-subtracted) for the aperture radii of 20\arcsec,
40\arcsec, and 60\arcsec, as well as flux integrated over the source area.
All the Bolocat sources in version 1 of the BGPS catalog within the Galactic
longitude range of $7.5^\circ < l < 194^\circ$ were followed up with
spectroscopic observations of dense gas tracers, using the
HCO$^+$ and N$_2$H$^+$ 3-2
transitions with the Arizona Radio Observatory Submillimeter
Telescope \citep{Schlingman:2011,Shirley:2013}
with 51\% of the sources detected in at least
one of the molecular lines. Additional molecular line observations
of NH$_3$ (1,1), (2,2), and (3,3) inversion lines are available for
sources in the inner Galaxy from the Green Bank Telescope
\citep{Dunham:2011}.
The molecular line observations of the Bolocat
sources provide kinematic information to link the dust continuum source to its
parent molecular cloud and the \mbox{\ion{H}{2}}\ regions.
\section{ANALYSIS}\label{analysis}
In this section we describe the selection criteria for the sample of
star forming regions (\S 3.1), the calculations of star formation rates
(\S 3.2), and the extraction and measurement of the properties
of MCs and dense gas (\S 3.3).
\subsection{The Samples}
Studies of star formation in the plane of the Galaxy are challenging
because it is non-trivial to clearly associate a star forming event, identified by broad-band
photometry, with
a particular molecular cloud.
In this study, the detection of a hydrogen radio recombination line
(RRL) arising
within an \mbox{\ion{H}{2}}\ region is required as a marker of star formation
because we need velocity information to tie the star formation to
the molecular gas.
Target sources
with well defined distances
are selected from the compilation of \mbox{\ion{H}{2}}\ regions with a RRL by A14.
First, we restrict our source list to Galactic longitudes, $30^\circ < l < 63^\circ$ to
sample a volume of the disk that is external to the stellar bar but crosses the
Scutum and Sagitarius spiral arms and extends to the Vulpecula Rift.
To ensure coverage by the GRS, whose minimum velocity is $-5$ km s$^{-1}$,
the LSR velocities of the radio recombination lines
(\mbox{$v_{\rm RRL}$}) are restricted to be greater than 0~km s$^{-1}$.
A consequence of
excluding negative velocities in this longitude range is the exclusion of star forming
regions on the far side of the Galaxy with Galactic radii greater than 8.5~kpc.
The primary sample \added{of 66 sources}
comprises the RRLs that could be associated with \gmc s.
The positions, velocities, and distances of the selected sample are
listed in Table~\ref{table1}. \added{The sample size is reduced to 51
sources after a minimum star formation rate is imposed (\S 3.2.3).}
When testing relations between dense gas and star
formation properties, we further restricted the sample to
sources that can be linked to one or more sources from the
BGPS survey of 1.1 mm dust continuum emission sources
\citep{Ginsburg:2013}
and follow-up spectral line observations in \mbox{{\rm HCO}$^+$}\ and \mbox{{\rm N}$_2${\rm H}$^+$}\
\citep{Schlingman:2011,Shirley:2013}.
To augment the sample, we include a subset of BGPS sources
studied by \citet{Battisti:2014} that are also associated with
RRL \mbox{\ion{H}{2}}\ regions.
\added{The resulting sample size for dense gas relations is 44.}
\input newtable1.tex
The choice of requiring \mbox{\ion{H}{2}}\ regions means that all the star-forming
regions associated with molecular clouds in this study are forming high-mass
stars, making the use of infrared and radio continuum tracers less
problematic.
The sample is clearly biased against starless
clouds or even clouds with low-level star formation similar to that in
solar-neighborhood clouds.
\subsubsection{Assigning Distances }
Distances to most of the target RRL regions are derived by A14 using
trigonometric parallax of associated masers or more typically, kinematic distances
assuming the rotation curve of the Milky Way derived by \citet{Brand:1993}
and resolving the distance ambiguity for the inner Galaxy.
For 9 sources where distances for the RRL \mbox{\ion{H}{2}}\ regions are not defined by A14,
the cloud distance is assigned to the mean of the
distances from the associated Bolocat sources, where distances were
obtained from \citet{Ellsworth:2015b}, who adopted the
rotation curve of \citet{Reid:2009}. \added{For the 9 sources taken from \citet{Battisti:2014},
we adopt the distance used in their study that assumed the rotation curve of \citet{Clemens:1985}. }
The difference in distances using the \replaced{two}{three}
rotation curves is small compared to other uncertainties with the kinematic distance method.
\added{The mean fractional distance uncertainties, $\sigma(D)/D$, for objects in our sample
are 0.15 for A14, 0.19 for \citet{Ellsworth:2015b}, and 0.20 for \citet{Battisti:2014}. }
The
molecular clouds are assigned distances of the \mbox{\ion{H}{2}}\ regions with which these are associated.
For sources without
a spectroscopic dense gas tracer, velocities were obtained by
connecting molecular gas observations from the GRS \textsuperscript{13}CO\ data
to Bolocat sources \citep{Ellsworth:2015b}. The combination of
dense gas tracers and \textsuperscript{13}CO\ data resulted in 45\% of the Bolocat
V2.1 sources with assigned \mbox{$v_{LSR}$}.
The distance uncertainties are
estimated from the combined uncertainties in the choice of Galactic
rotation curve, the streaming motions of 7 km s$^{-1}$, and the solar
circular rotation speed \citep{Anderson:2012, Anderson:2014}.
The sources span a large range of heliocentric distances of
1.7 to 13.2~kpc, with an average distance of $8\pm3$ kpc and a median
of 8.1 kpc. The span of Galactocentric radii ($R_{\rm GAL}$) is less, ranging
from 4.6 to 8~kpc with an average of $6.3\pm 0.8$~kpc and a median of 6.1~kpc.
They should be more representative of star formation
activity in the Galaxy than nearby, well-studied targets such as those
in the Gould Belt
\citep{2015ApJS..220...11D}.
\subsection{Star Formation Rates}
The radio continuum at 21 cm and MIR emission at 22 or 25 \mbox{$\mu$m}\ are the star
formation tracers used in this study.
We refer to them as ``radio" and ``MIR" in what follows.
Using continuum emission as a tracer creates a problem of source
confusion between different emitting regions along the line of
sight. If more than one emitting region lies within the
angular vicinity of a MC, one cannot distinguish which regions are associated
with the MC.
To mitigate the problem of confusion along the line of
sight, the RRL data are used to
associate radio continuum emission and MIR emission with molecular gas
\added{
by requiring velocity agreement with the molecular gas (see \S 3.2.1).
}
\input newtable2.tex
\subsubsection{Star Formation Rates Derived from Radio Continuum Emission}
For each of the RRL locations,
a region of associated radio continuum
emission is defined using the following steps.
First, the radio emission is fitted to a
2D-Gaussian profile, centered on the RRL position. If the emission is reasonably well fitted
with the 2D-Gaussian, the source is labeled as a compact
source with sizes ($\sigma_x, \sigma_y)$
of the fitted Gaussian. The radio flux is calculated from aperture photometry
with an aperture radius $3(\sigma_x+\sigma_y)/2$ and
an appropriate sky annulus. If the radio
emission could not be fitted well with a 2D Gaussian (which is the
case if the source is extended or in a crowded region), then a
polygon is used to define the region of the source emission and another
region for estimating background emission.
The RRL sources with no significant radio continuum emission above the
background are excluded from further analysis.
If more than one RRL source is found within the solid angle of the radio
aperture and
RRL velocities are within $\pm 5$ km s$^{-1}$\ of each other, then the RRL
sources are considered as a single star-forming region.
If the velocities differ by more than 5 km s$^{-1}$\ and the
radio emission cannot be separated, then the sources are excluded
from our target list.
Radio emission from galaxies is comprised of both synchotron and
free-free emission
components with synchotron emission dominating at 21~cm \citep{Condon:1992}.
In contrast, the higher spatial resolution in the Milky Way surveys favors
localized regions of free-free emission.
Since our targets are selected by the detection of the hydrogen
recombination line, we assume the origin of the 21~cm signal is
free-free emission from an \mbox{\ion{H}{2}}\ region
excited by ionizing far-UV radiation from massive stars.
Therefore, we do not use the extragalactic relation between radio emission
and star formation, which is primarily a relation between synchrotron
emission on large scales and star formation averaged over about 100 Myr
\citep{Kennicutt:2012}.
Instead, we use the compact free-free radiation from the \mbox{\ion{H}{2}}\ region, which,
like H$\alpha$, averages over about 3 Myr
\citep{2011ApJ...741..124H,Kennicutt:2012}.
The free-free
emission is related to the ionizing luminosity by the expression
\begin{eqnarray}
\frac{N_{\rm UV}}{\text{phot s}^{-1}} = 6.3 \times 10^{52}
\left(\frac{T_e}{10^4 \text{K}}\right)^{-0.45}
\left(\frac{\nu}{\text{GHz}}\right)^{0.1}
\nonumber \\
\times
\left(\frac{L_{\rm T}}{10^{20}\text{W Hz}^{-1}}\right),
\end{eqnarray}
where $N_{\rm UV}$ is the number of Lyman continuum photons per second,
$T_e$ is the electron temperature, $\nu$ is the frequency, and $L_{\rm T}$
is the thermal emission luminosity assuming optically thin gas \citep{Condon:1992}.
For an electron temperature of $10^4$ K, and an IMF described by
\added{\citep{Chomiuk:2011}}, the SFR is calculated from
\begin{equation}
\text{SFR(radio)}
= 0.47 \times 10^{-14}
\left( \frac{\nu}{\text{GHz}} \right)^{0.1} \times
\left ( \frac{L_{\rm T}}{\text{W Hz}^{-1}} \right) \;\;{\msun \text{Myr}^{-1}}
\end{equation}
\citep{Chomiuk:2011,Vutisalchavakul:2013}.
\explain{Reference corrected}
This expression may underestimate the SFR if UV photons are not
absorbed, as in the case of a density bounded \mbox{\ion{H}{2}}\ region.
\added{If a Salpeter slope is continued down to 0.1 \msun, the
derived SFRs would be 1.44 times higher \citep{Chomiuk:2011}.}
The star formation rates derived from radio emission for our
sample of RRL regions are listed in Table~\ref{table2}.
\subsubsection{Star Formation Rates Derived from MIR Emission}
The process used to define the
MIR-emitting region associated with a RRL source is similar to the
method used for radio continuum images described in the previous section.
The MIR emission at the location of the radio continuum region is
examined. If the distribution of MIR emission is compact and isolated,
it is fitted
to a 2D Gaussian profile. Otherwise we define the emitting
region with a polygon. Photometry for all of the MIR emitting
regions is performed similarly to the radio continuum regions.
Most of the sources that are compact in radio images are also compact in
MIR emission.
In the case of G033.643-00.229, the background subtracted flux is less than zero.
This target is excluded from subsequent analyses.
The SFR is calculated from the extragalactic
relation \citep{Calzetti:2007}:
\begin{equation} \label{eq:ssfr}
\mbox{\rm SFR(MIR)}
= 1.27{\times}10^{-32} \times
[L_{24} (\text{ergs s}^{-1})]^{0.885} \;\;(M_\odot \ \text{Myr}^{-1})
\end{equation}
where $L_{24}$ is the 24\mbox{$\mu$m}\ luminosity.
\added{The IMF assumed by \citet{Calzetti:2007} is consistent with
that used for the SFR from the radio continuum emission.}
Here, we substituted the WISE 22 \mbox{$\mu$m}\ or the IRAS 25 \micron\
luminosity for the
24 \mbox{$\mu$m}\ luminosity.
The corresponding star formation rates derived from the mid-infrared luminosity
are listed in Table~\ref{table2}.
\begin{figure*}[h]
\includegraphics[scale=0.5]{sfrcomp.eps}
\caption{The comparison between SFR(MIR) and SFR(radio)\added{, measured
in \msunmyr}. The solid
black line represents the 1-1 line where the two SFRs are equal. The
blue line is a fit to the data, including uncertainties in both axes.
The red line is fit to a smaller sample of sources from
\citet{Vutisalchavakul:2013}.
}
\label{fig:compare_sfr}
\end{figure*}
\subsubsection{Comparing SFR from radio continuum and MIR}
The two different estimations of SFR are compared in
Figure~\ref{fig:compare_sfr}. Because the luminosities of both tracers
are proportional to a flux times the square of the distance, the uncertainties
are computed by propagating a nominal 10\% uncertainty in the flux measurement
and the distance uncertainty in Table~\ref{table1} (see the Appendix for a
description of error propagation).
The distance uncertainty is usually, but not always, dominant.
The SFRs from the two tracers are
comparable, with the SFR from MIR usually being somewhat
higher, especially at low SFRs. The black solid line in
Figure~\ref{fig:compare_sfr} is a line of equality, while the blue solid
line is a fit to the data, using $\chi^2$ minimization with errors in both
variables. The fit is
\begin{equation}
\log[{\rm SFR(MIR)}] = a + b\ \log[{\rm SFR(radio)}]
\end{equation}
with $a = 0.38\pm0.04$ and $b = 0.85\pm0.02$. The slope is nearly
identical to that found by analysis of a more limited sample,
while the coefficient is a bit less, but within the uncertainties:
($a = 0.53\pm0.17$ and $b = 0.83\pm0.08$)
\citep{Vutisalchavakul:2013};
that fit is shown as a red line in Figure~\ref{fig:compare_sfr}.
The slightly sub-linear fits and the larger discrepancies at low
SFR can result from the fact that the radio emission is more
sensitive to the upper end of the IMF (as is also true for H$\alpha$
or any diagnostic requiring ionized gas), as can be seen from Table 1 of
\citet{Kennicutt:2012}.
For this reason, we use SFR(MIR) in further analysis. This tracer also
begins to underestimate the SFR for SFRs less than about 5 \msunmyr,
corresponding to a total far-infrared luminosity of $10^{4.5}$ \lsun\
(\citealt{Vutisalchavakul:2013, Wu:2005}). Consequently,
we limit further investigation to those sources with SFR above this
value\added{, leaving 51 sources}.
Even above the threshold of 5 \msunmyr, SFRs may be poorly estimated. The
dominant source of uncertainty in SFR is the conversion from the
star-formation tracers to SFR. The common method used to calculate the
conversion factors is to use a stellar population synthesis
model, stellar evolutionary models, and atmospheric models to connect
stellar populations to their photometric output. Then assumptions of
a fully-sampled IMF and a star-formation history (SFH)
connect the SFR of a region to the light output for each tracer
\citep{Leitherer:1999}.
The effect of the stochastic sampling of the IMF and SFH on
the reliability of the SFR tracers has been a topic of many recent
studies \citep{Fumagalli:2011, daSilva:2012,daSilva:2014, Kruijssen:2014,Krumholz:2015}.
\citet{daSilva:2014} used the
Stochastically Lighting Up Galaxies (SLUG) simulations to study the
effect of IMF and SFH sampling on the conversion from photometric
observations of the star-formation tracers to SFR. Instead of a unique
SFR for a given luminosity (in the tracers), the effect of IMF and SFH
sampling results in probability distributions of SFR, which
can have a large bias and scatter for regions with low SFR.
The result from SLUG, given an input of a constant SFR on the timescale
of 500 Myr, for the bolometric luminosity as a SFR tracer gives a
scatter in the log(SFR) of $\approx 0.6$ dex in the SFR range of $10 -
100$ \msunmyr, assuming uncertainties in the flux of 0.25
dex \citep{daSilva:2014}.
Consistent with that analysis,
the SFR(MIR) appears to underestimate SFR for the Milky Way
by a factor of $\approx 2-3$ \citep{Chomiuk:2011}.
Because the effect is usually to underestimate the SFR,
we treat this as a systematic uncertainty,
rather than including it in the individual SFRs.
\subsection{Gas Structures}
\subsubsection{Molecular Cloud Properties}
The linking of a molecular cloud to the radio recombination line \mbox{\ion{H}{2}}\ regions is similar to the method
employed by \citet{Battisti:2014} to connect larger clouds to 1.1 mm continuum sources.
A 100$\times$100~pc$^2\times$40 km s$^{-1}$ data cube centered on the $l$,
$b$, \mbox{$v_{\rm RRL}$}\
coordinates of the \mbox{\ion{H}{2}}\ region is extracted from the larger CO survey. For a handful of nearby sources, the 100~pc
edge may lie outside the Galactic latitude boundary of the survey, in which case the area is truncated to the largest available
square area.
The segmentation program, CPROPS \citep{Rosolowsky:2006},
is applied to these extracted data to identify sets (islands) of contiguous voxels
with CO brightness temperatures above a given
threshold value. This threshold is varied depending on the complexity of the local background emission generated by unrelated
foreground and background molecular
clouds with comparable radial velocities with the intent to determine the lowest threshold value that distinguishes a
structure from this background emission. We do not examine substructure within these islands of
emission.
For a given threshold, hundreds of structures may be identified by the segmentation algorithm distributed
over the 10$^4$ pc$^2$ area and 40~km s$^{-1}$\ range. To narrow the search for the molecular cloud associated with the \mbox{\ion{H}{2}}\ region, we
select a subset of this list with the condition,
$|v_{\rm CO}-v_{\rm RRL}| < 10$ km s$^{-1}$, where $v_{\rm CO}$ is the
velocity centroid of the CO structure.
For each structure in this subset, the distribution of velocity-integrated CO emission and its spatial
relationship to the \mbox{\ion{H}{2}}\ region are examined.
A CO structure is assigned to the \mbox{\ion{H}{2}}\ region if its boundaries enclose the \mbox{\ion{H}{2}}\ region or
\replaced{its distribution is strongly
peaked towards the \mbox{\ion{H}{2}}\ region position.}
{if there is a local intensity maximum coincident with
the \mbox{\ion{H}{2}}\ region position.}
\replaced{If these conditions are not satisfied, a structure may also be linked if
its CO boundary,
is contiguous to the \mbox{\ion{H}{2}}\ region}
{A structure may also be linked if the solid angle of the detected
radio continuum emission from the
HII region partially overlaps with the solid angle of the CO emitting region or if the
respective boundaries of the radio continuum and CO emissions are within 2\arcmin\ (2 times the FWHM beam width of the radio continuum data) at some
point along their perimeters.
}
In several cases, no structure can be confidently assigned to the \mbox{\ion{H}{2}}\
region position, so we select
\replaced{a nearby structure that could be reasonably associated with the \mbox{\ion{H}{2}}\ region. }
{the structure with the most comparable velocity to the RRL velocity
that is within 5 \mbox{$^{13}$CO}\ FWHM beam widths of the HII region position. }
Two examples of the cloud extraction are shown in Figure~\ref{co_clouds}.
\begin{figure*}[h]
\includegraphics[scale=0.75]{G032p582+00p001_figure.eps}
\includegraphics[scale=0.75]{G033p941-00p039_figure.eps}
\caption{Images of integrated \textsuperscript{13}CO\ \jj10\ emission (contours) overlayed on
21~cm continuum emission for RRL targets
(left) G032.582+00.001 and (right) G033.941-00.039.
Contours range from 1 to 15 K km s$^{-1}$\ spaced by 2 K km s$^{-1}$\
for
G032.582+00.001 and 1 to 5 K km s$^{-1}$\ spaced by 1 K km s$^{-1}$\ for G033.941-00.039.
The radio continuum source to the southwest of the cloud is the associated source.
The solid, red circles mark the locations of associated Bolocam sources of dust
continuum emission.
\deleted{These are examples of Group~1 (left) and Group~2 (right) clouds.}
}
\label{co_clouds}
\end{figure*}
Once a cloud is linked to the \mbox{\ion{H}{2}}\ region by the above criteria, its properties are determined from the set of voxels
that comprise the structure. These properties include intensity-weighted positional moments and
\mbox{$^{13}$CO}\ luminosity.
While the voxels that comprise the cloud are defined by the applied threshold, the cloud properties such as mass, size,
and velocity dispersion, are
extrapolated from this brightness temperature threshold to a hypothetical cloud edge at a brightness temperature of 0~K.
The properties and uncertainties of each MC linked to an RRL source are summarized in Table~\ref{table3}.
\input newtable3.tex
For \textsuperscript{13}CO\ data, column densities are derived following the expressions in \citet{Pineda:2010}. Here, we assume
optically thin emission, so column density in the upper ($J=1$) rotational energy level is
\begin{equation}
N_{1,{\rm thin}}=
\frac{8{\pi}k_B\nu_{10}^2}{hc^3A_{10}C(\mbox{$T_{\rm ex}$})}\int T_{\rm B}(v)dv
\end{equation}
where $k_B$ is the Boltzmann constant, $h$ is the Planck constant,
$c$ is the speed of light, and
$\nu_{10}$, $A_{10}$, $T_{B}$, \mbox{$T_{\rm ex}$}\
are the frequency, Einstein A coefficient,
brightness temperature, and excitation temperature
for the \mbox{$^{13}$CO}\ \jj10\ transition.
The factor, $C(\mbox{$T_{\rm ex}$})$, contains temperature dependent terms,
\begin{equation}
C(\mbox{$T_{\rm ex}$})=\left(1-\frac{exp(T_{\rm L}/\mbox{$T_{\rm ex}$})-1}{exp(T_L/T_{\rm bg})-1}\right)
\end{equation}
where $T_{\rm L}=h\nu_{10}/k_B$ and $T_{\rm bg}$=2.725~K.
The total column density of \mbox{$^{13}$CO}\ is
\begin{equation}
N(\mbox{$^{13}$CO})=N_{1,thin}Z exp(h{B_\circ}J(J+1)/{k_B}\mbox{$T_{\rm ex}$})/(2J+1)
\end{equation}
where $B_\circ$ is the rotational constant for \mbox{$^{13}$CO}, and $Z$ is the partition function, which accounts for population in other states
\begin{equation}
Z = \sum_{J=0}^\infty (2J+1) exp(-h{B_\circ}J(J+1)/k_B\mbox{$T_{\rm ex}$})
\end{equation}
Finally, to convert to molecular hydrogen column density,
$N(\mbox{{\rm H}$_2$})=N(\mbox{$^{13}$CO})[\mbox{{\rm H}$_2$}/\mbox{$^{13}$CO}]$,
we assume a constant excitation temperature of 8~K and a H$_2$
to $^{13}$CO abundance ratio of 4.1$\times$10$^{5}$ for all clouds.
This value for \mbox{$T_{\rm ex}$}\ is based on studies by \citet{Heyer:2009} and \citet{Roman-Duval:2010} who
found mean \mbox{$T_{\rm ex}$}\ values of 7-8~K derived from \textsuperscript{12}CO\ data for large samples of clouds.
The assigned abundance value is based on mean abundance values for
nearby clouds using infrared-derived extinction as a proxy for
\mbox{{\rm H}$_2$}\ column density \citep{Lada:1994,Pineda:2008, Pineda:2010, Ripple:2013}.
This leads to a conversion between \textsuperscript{13}CO\ integrated intensity and molecular hydrogen column density of
3.9$\times$10$^{20}$ cm$^{-2}$/(K km s$^{-1}$) that is a required argument in the CPROPS
program to calculate mass. The cloud mass is then
\begin{equation}
\mbox{$M_{\rm cloud}$} = 3.9{\times}10^{20} \mu m_H D^2 \int d\Omega \int dv T_B(l,b,v)
\end{equation}
where $D$ is the distance and the integrals are over the solid angle, $\Omega_{cl}$, and velocity range for the
set of cloud voxels.
There are several sources of
uncertainty in the MC mass.
These include
distance,
abundance and $T_{ex}$ variations, and random errors owing to
the observations. This last component is small compared to the other sources and is ignored.
For \textsuperscript{13}CO, we
estimate the uncertainty of the conversion factor from integrated intensity
to $N(\mbox{$^{13}$CO})$
by varying the assumed excitation temperature from 4 to 16~K in steps of 2~K. The fractional root mean square of
values
about the adopted value for 8~K is 27\%. The variations of \textsuperscript{13}CO\ abundance,
$[\mbox{{\rm H}$_2$}/\mbox{$^{13}$CO}]$ both within clouds and from cloud-to-cloud in the
solar neighborhood is $\sim$30\% \citep{Ripple:2013}.
These uncertainties are included in the values in Table \ref{table3},
but distance errors are not (see Appendix).
Assuming a spherical cloud and a power law density profile with index of 1,
the CPROPS package calculates the virial mass, \mbox{$M_{\rm vir}$}, from the expression,
\begin{equation}
\mbox{$M_{\rm vir}$} = 1040{\delta}v^2 \mbox{$r_{\rm cloud}$} ~M_\odot
\end{equation}
where ${\delta}v$ is the 1-dimensional velocity dispersion in km s$^{-1}$\ and \mbox{$r_{\rm cloud}$}\ is the effective
radius of the cloud in parsecs.
\added{The coefficient can be found in
\citet{1988ApJ...333..821M} after converting from FWHM to $\delta v$.}
The virial parameter, $\alpha =\mbox{$M_{\rm vir}$}/\mbox{$M_{\rm cloud}$}$, offers a coarse measure
of the boundedness of the cloud, assuming there is no significant external pressure.
The mean volume density of the cloud, $\mbox{$n_{\rm cloud}$}$, is derived from the
\mbox{$M_{\rm cloud}$}\ and \mbox{$r_{\rm cloud}$}, assuming a spherical cloud geometry,
$\mbox{$n_{\rm cloud}$} = 3\mbox{$M_{\rm cloud}$}/4{\pi}{\mu}m_H \mbox{$r_{\rm cloud}$}^3$.
\input newtable4.tex
\subsubsection{Dense Gas Mass}
Dense gas properties are estimated for each MC using source fluxes and sizes from
the BGPS Version 2.1 source catalog.
The relation between gas mass and dust continuum flux is
given by the expression
\begin{equation}
\mbox{$M_{\text{dense}}$} = \frac{S_{1.1} D^2 (\rho_g/\rho_d)}{B_\nu (T_{\rm dust})
\kappa_{\rm dust, 1.1}},
\end{equation}
where $S_{1.1}$ is the 1.1 mm flux density, $D$ is the distance,
$\kappa_{\text{dust},1.1}$
is the dust opacity at 1.1 mm per dust mass, assumed to be
1.14 cm$^2$ g$^{-1}$\citep{Ossenkopf:1994},
and $\rho_g/\rho_d$ is the gas-to-dust mass ratio, taken to be
100 \citep{Hildebrand:1983}.
Using this expression, we derive the mass and associated uncertainties using a Monte Carlo
simulation that assumes Gaussian errors on the flux density and a
Gaussian distribution of dust temperatures with a mean of 20~K and
a standard deviation of 8~K \citep{Battersby:2011, Ellsworth:2015b}.
The total dense gas mass for each MC is estimated from the sum
of the masses for all Bolocam sources
within the cloud mask with and without \mbox{$v_{LSR}$}\ information.
We also derive the dense gas mass fraction (DGMF) corresponding to the ratio of the mass
traced by BGPS 1.1 mm sources to the mass of the
MC as traced by \textsuperscript{13}CO\ emission,
\begin{equation}
DGMF = \frac{\mbox{$M_{\text{dense}}$}}{\mbox{$M_{\rm cloud}$}}. \nonumber
\end{equation}
Properties of the dense gas component are given in Table \ref{table4}.
\section{Classification of MCs}\label{classification}
We classified our sample of star forming MCs into three distinct groups
based on the locations of the associated \mbox{\ion{H}{2}}\ regions with respect to the
molecular cloud.
Group 1 MCs have associated radio continuum emission whose solid angle is
mostly circumscribed by the boundary of \textsuperscript{13}CO\ emission.
There are 37 MCs in Group 1.
For Group 2 MCs,
the radio continuum emission partially
overlaps or is contiguous with \textsuperscript{13}CO\ boundaries but the RRL sources lie
outside the clouds.
\added{The left panel of figure \ref{co_clouds} shows a Group 1 cloud, while the right
panel shows a Group 2 cloud.}
There are 21 MCs in Group 2.
Finally,
Group 3 clouds are MCs in which the radio continuum emission is well displaced
from the \textsuperscript{13}CO\ boundaries.
The association of the MC with the \mbox{\ion{H}{2}}\ region is most uncertain for
Group 3 sources.
With no
angular overlap with the MC, the association is solely based on
the chosen spatial and velocity offsets. There are only 8 MCs in this
group. Due to the small number and the uncertainty of associating
gas and star formation in this group, we
exclude Group 3 MCs from all further analysis.
The classification into groups in this study has some
overlap with the classification used by \citet{Fukui:1999}
and \citet{Kawamura:2009} (K09),
but our spatial resolution is higher than that of K09 so we examine
associations between MCs and star formation on smaller
spatial scales but also require similar velocities to account for line of sight confusion.
Nevertheless, our classifications likely correspond to a similar
evolutionary
sequence in which Group 1 is the earliest stage of massive star formation
(Type II of K09), Group 2 describes the initial feedback from
massive stars, and Group 3 is a very late stage in
which the original cloud has been mostly photoionized or dispersed, leaving only small fragments, similar to Type III clouds in the K09 classification.
To explore whether Group 1 and Group 2 sources differ in the main
properties we are considering, we computed averages and standard
deviations (in the log for quantities with large ranges) for the two
groups; these are shown in Table \ref{grouptab2}. The two groups have very
similar averages for \mbox{$M_{\rm cloud}$}, \mbox{$M_{\rm dense}$}, \mbox{$n_{\rm cloud}$}, \mbox{$r_{\rm cloud}$}, and
SFE (SFR/\mbox{$M_{\rm cloud}$}).
Group 2 clouds may have somewhat larger mean values for $\alpha$ and somewhat smaller values for DGMF. Because the differences are not statistically
significant, we do not distinguish the two groups in the rest of the analysis.
\input tablegroups.tex
\section{Testing Star Formation Relations}\label{tests}
The question we address in this section is the following:
which properties of the ISM best predict the SFR for the clouds
in this sample?
The first proposed relation \citep{Schmidt:1959,Schmidt:1963}
suggested that the SFR would be a function of the density of gas.
For extragalactic studies, the surface densities of
star formation rate (\mbox{$\Sigma_{\text{SFR}}$}) and gas ($\Sigma_{gas}$) are
easier to measure than are volume densities.
Studies of other galaxies found that $\mbox{$\Sigma_{\text{SFR}}$} \propto \Sigma_{gas}^{1.4}$
provided a good fit for a wide range of galaxies \citep{Kennicutt:1998}.
When only molecular gas is considered, linear relations are generally
found ($\mbox{$\Sigma_{\text{SFR}}$} \propto \Sigma_{gas}$)
(\citealt{2008AJ....136.2846B,2013AJ....146...19L}),
even for the outer regions of galaxies, where the gas is predominantly
atomic
\citep{2011AJ....142...37S}.
Studies of galaxies on sub-kiloparsec scales have shown that
star formation is associated with molecular gas
(e.g., \citealt{2013AJ....146...19L}).
Within the Milky Way, studies with sub-parsec resolution show
that star formation is further concentrated within relatively
dense or opaque parts of molecular clouds
(\citealt{Lada:2010,Heiderman:2010}).
In this section, we use our sample of star forming regions
to test some of the proposed models for how SFRs are determined in
\gmc s. In doing so, we consider integrated properties, such
as cloud mass, mass of dense gas, or cloud mass per free fall time
as variables in the following equation, where $X$ represents the variable in
question.
\begin{equation}
\mbox{\rm SFR} = A X^n
\end{equation}
Because of the large range of parameters, and to be consistent with
previous studies, we generally fit the logarithmic version of these relations.
\begin{equation}
\log \mbox{\rm SFR} = \log A +n \log X
\label{eq:sfl_sk2}
\end{equation}
We examine the various relations
that are used in extragalactic work or that have been proposed by
theorists for clouds in the Milky Way selected to have substantial
star formation.
We use the SFR
from the mid-infrared data and consider only clouds with a SFR
of at least 5 \msunmyr, as discussed earlier.
The basic sample thus includes 51 sources. The samples for other
variables are somewhat smaller depending on which data are available.
For the dense gas relations, only 44 sources are available because not
all clouds had BGPS sources.
We calculate the Pearson correlation coefficient to evaluate the probability
that
the measured relationship could emerge from a population of random numbers.
If the correlation is significant \added{(using Pearson $r$, a $3 \sigma$
correlation requires $|r| > 3/\sqrt{N_s - 1}$, where $N_s$ is the sample
size)}, we fit the data to the model
in equation \ref{eq:sfl_sk2} to determine the parameters,
$\log(A)$ and $n$, considering errors in both variables using the MPFITEXY routine
\citep{Williams:2010}, which
depends on the MPFIT package \citep{Markwardt:2009}.
In this model-fitting program, the intrinsic scatter of the data about the model is
included in the error-weighting of the data. If necessary, this intrinsic
scatter is iteratively modified so that the reduced $\chi^2$ value of the fit is about unity.
The scatter of points is characterized by the root mean square of the displacement
of log(SFR) values from the best fit line for each value of X.
The parameters from the fits are given in Table \ref{tab:fits}.
The column labeled DOF shows the sample size minus the number (2) of
fit parameters.
\input tablefits.tex
\subsection{Star Formation Rate versus Cloud Mass }
\begin{figure*}[h]
\center
\includegraphics[scale=0.4]{sfrvsmass.eps}
\caption{The relation between SFR and \mbox{$M_{\rm cloud}$}.
The line is the linear least-square fit to the data. }
\label{fig:sfl_sk}
\end{figure*}
The first and the simplest proposition is that SFR depends solely on
the total mass of the MC (\mbox{$M_{\rm cloud}$}).
This is the extragalactic star formation relation
applied on the cloud scale.
In this case $X = \mbox{$M_{\rm cloud}$}$, as determined by the \mbox{$^{13}$CO}\ emission.
Figure~\ref{fig:sfl_sk} shows the SFR versus \mbox{$M_{\rm cloud}$}.
The dotted lines represent the linear least-square fits to the log of the data.
The fit parameters to Equation~\ref{eq:sfl_sk2} gives the slope
\added{$n = 0.66\pm0.12$, sub-linear by nearly 3 $\sigma$}
(see Table \ref{tab:fits}).
The relation for the Galactic Plane clouds is thus considerably
flatter than the linear relation found for other galaxies,
\added{averaging over kpc scales}
(\citealt{2008AJ....136.2846B,2013AJ....146...19L,2011AJ....142...37S}).
\subsection{Star Formation Rate versus Dense Gas Mass}
A tighter relation has been observed in other galaxies
between SFR and dense gas mass. \citet{Gao:2004}
observed a tight linear relationship between $L(\rm IR)$ and
$L$(HCN) in a sample of infrared galaxies. \citet{Wu:2005}
found that the relationship is extended to the scale of
Galactic dense clumps. These studies and the study of nearby
molecular clouds led to \replaced{a proposed model}{the
proposal} that molecular gas above some density can
form stars efficiently \citep{Goldsmith:2008, Lada:2010, Heiderman:2010}.
The star-formation threshold model \citep{Lada:2012, Evans:2014}
states that the SFR in \gmc s is \replaced{determined}{best predicted}
by the amount of dense gas above some
threshold density. In this case, $X = \mbox{$M_{\rm dense}$}$, as measured
by the sum of masses of the BGPS sources within the cloud,
with the caveat that BGPS sources are lower density and more like
clouds at large distances owing to spatial filtering to remove atmospheric
emission
\citep{Dunham:2011}.
Comparing SFR and \mbox{$M_{\text{dense}}$}\ in our sample yields a nearly linear relation
($n = 1.09\pm 0.18$) with a Pearson correlation coefficient of 0.58.
Figure~\ref{fig:sfl_bgps} shows the result of the relation
between SFR and \mbox{$M_{\text{dense}}$}.
\begin{figure*}[h]
\center
\includegraphics[scale=0.4]{sfrvsdense.eps}
\caption{The relation between SFR and \mbox{$M_{\text{dense}}$}.
The line is the linear least-square fit to
the data. }
\label{fig:sfl_bgps}
\end{figure*}
\explain{Removed subsection on DGMF}
\subsection{SFR versus \mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$} }
The free-fall model states that the SFR depends on
the molecular mass
over the free-fall time \citep{Krumholz:2005,Krumholz:2012}:
\begin{equation}
SFR = f_{H_2} \, \mbox{$\epsilon_{\rm ff}$} \times \frac{M_{gas}}{\mbox{$t_{\rm ff}$}},
\end{equation}
where $f_{H_2}$ is the fraction of molecular gas compared to the total gas
mass, $\mbox{$\epsilon_{\rm ff}$} = \mbox{$t_{\rm ff}$}/\mbox{$t_{\rm{dep}}$}$ is the star formation efficiency per
free-fall time, the gas depletion time is \mbox{$t_{\rm{dep}}$},
and the free-fall time is
\begin{equation}
\mbox{$t_{\rm ff}$} = \sqrt{\frac{3\pi}{32G\rho}}.
\end{equation}
For \gmc s, where only molecular gas is concerned, the relation becomes
\begin{equation}
SFR = \mbox{$\epsilon_{\rm ff}$} \times \frac{\mbox{$M_{\rm cloud}$}}{\mbox{$t_{\rm ff}$}}.
\label{eq:sflwtff}
\end{equation}
Comparing data from nearby clouds to other galaxies, \citet{Krumholz:2012} estimated
an approximately constant $\mbox{$\epsilon_{\rm ff}$}$ of 0.01. This value is aligned with theoretical
predictions \citep{Krumholz:2005,2012ApJ...759L..27P}.
We tested the free-fall
model with our cloud
sample by using the total \gmc\ mass from \textsuperscript{13}CO\ and using the
average density for \mbox{$t_{\rm ff}$}.
The result is shown in Figure~\ref{fig:sfl_tff}.
The red line
represents the relation in Equation~\ref{eq:sflwtff} with
$\mbox{$\epsilon_{\rm ff}$} = 0.01$.
The fit to the data shows a \added{super-linear relation with $n = 1.80\pm0.39$.}
\deleted{very similar to that found for \mbox{$M_{\rm cloud}$}.}
The Pearson correlation coefficient is 0.55.
Generally speaking, including the free-fall time did not improve
the relation compared to using \mbox{$M_{\rm cloud}$}, and the \replaced{sublinear}{super-linear} fit is
not consistent with the hypothesis of
\citet{Krumholz:2012}.
\begin{figure*}[h]
\center
\includegraphics[scale=0.5]{sfrvsmtff.eps}
\caption{The relation between SFR and mass over free-fall time.
The blue line is the linear least-square fit to
the data,
and the red line represents the free-fall model
with $\mbox{$\epsilon_{\rm ff}$} = 0.01$.
}
\label{fig:sfl_tff}
\end{figure*}
\subsection{\mbox{$\epsilon_{\rm ff}$}\ versus \mbox{$t_{\rm ff}$}/\mbox{$t_{\rm dyn}$}}
As summarized by
\citet{2014prpl.conf...77P},
simulations of turbulent molecular clouds indicate that the
speed of star formation, measured by \mbox{$\epsilon_{\rm ff}$}, should depend on
the virial parameter. In particular
\citet{2012ApJ...759L..27P} show that turbulent simulations
closely follow the following relation.
\begin{equation}
\mbox{$\epsilon_{\rm ff}$} = \epsilon_{\rm wind} e^{-1.6 \mbox{$t_{\rm ff}$}/\mbox{$t_{\rm dyn}$}}
\end{equation}
where $\epsilon_{\rm wind}$ is the core to star efficiency, taken
to be about 0.5 due to winds removing material, and
$\mbox{$t_{\rm dyn}$} = \mbox{$r_{\rm cloud}$}/\sigma_{\rm v, 3D}$, and
$\sigma_{\rm v, 3D} = \sqrt{3} \sigma_{\rm v, 1D}$. The variable
$\mbox{$t_{\rm ff}$}/\mbox{$t_{\rm dyn}$}$ is simply related to the virial parameter by
\begin{equation}
\mbox{$t_{\rm ff}$}/\mbox{$t_{\rm dyn}$} = 0.86 \sqrt{\alpha}
\end{equation}
where $\alpha$ is the virial parameter.
The data are plotted in Figure \ref{fig:padoanmod},
along with the prediction of \citet{2012ApJ...759L..27P}.
The data show no significant correlation (Pearson correlation
coefficient of $0.13$)
and lie on average a factor of 55 below the model
predictions.
\added{The mean efficiency \mbox{$\epsilon_{\rm ff}$}\ for \mbox{$M_{\rm cloud}$}\ is 0.008, comparable
to the \citet{Krumholz:2012} value of 0.01. For this sample,
$\mean{\alpha} = 1.6\pm 1.2$; if we use that value in equations 17 and 18,
they would predict $\mbox{$\epsilon_{\rm ff}$} = 0.09$, more than 10 times the observed
mean value.
}
\begin{figure*}[h]
\center
\includegraphics[scale=0.5]{padoanmod.eps}
\caption{The relation between \mbox{$\epsilon_{\rm ff}$}\ and
free-fall time over dynamical (crossing) time.
The blue line is the prediction of the model of
\citet{2012ApJ...759L..27P}.
The red vertical line indicates $\alpha = 1$.
}
\label{fig:padoanmod}
\end{figure*}
\subsection{Summary of Star Formation Relationships}
Comparison of the values in Table~\ref{tab:fits} shows that all
three variables (\mbox{$M_{\rm cloud}$}, \mbox{$M_{\text{dense}}$}, \mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$})
have similar correlation coefficients.
\deleted{Using the rule of thumb that a $3 \sigma$ correlation requires
$r = 3/\sqrt{N-1}$, where $N$ is the number in the sample,} All the
correlations are significant at \replaced{that}{the $3 \sigma$} level. Equivalently, we can
exclude the null hypthosis that the correlations arise from a random
distribution of variables with a high level of probability.
Only the correlation with \mbox{$M_{\rm dense}$}\ is consistent with a linear
relation. The fit to SFR versus \mbox{$M_{\rm cloud}$}\ shows a slope much flatter
($0.66\pm0.12$)
than the usual values of $1.0$ to $1.5$.
Theory predicts a linear relation between SFE and \mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$},
but the slope we find ($1.80\pm 0.39$) differs by more than
$2 \sigma$ from theoretical expectation.
Another theoretical prediction is that the star formation
efficiency per free-fall time should decrease exponentially with
the ratio of the free-fall time to the dynamical time.
As shown in Figure~\ref{fig:sfl_tff},
there is no signficant correlation in our data and the values
lie well below the predictions.
Comparing SFR and \mbox{$M_{\rm cloud}$}\, we found considerable scatter.
This scatter has been observed and quantitatively explained in
several studies \citep{Onodera:2010, Schruba:2010, Kruijssen:2014}.
One explanation is that the strong
correlation observed when looking at the scale of galaxies is due to
averaging over variations in different star forming regions'
properties. When looking at the scale of \gmc s, the properties of \gmc\
such as the evolutionary stages contribute to the scatter in the
relation between mass and SFR. Even in our sample where only sources
with overlap between molecular gas and star formation are selected
(thus leaving out the earlier stages such as IRDC and later stages
where stellar feedback disrupts the clouds), the scatter in SFR
and \mbox{$M_{\rm cloud}$}\ relations is large. In previous work, we have shown that
averaging over larger regions in our Galaxy reduces the scatter
\citep{Vutisalchavakul:2014}.
\section{Comparison to Nearby Clouds}\label{lowcomp}
Now that we have extended the study of star formation relations to
more distant regions forming more massive stars in the Milky Way, we can compare
the results to those for nearby clouds
\citep{Evans:2014,Heiderman:2010}.
\added{
Those authors considered star formation relations {\bf within}
the nearby clouds, but we lack the resolution to do that for the
Galactic Plane clouds. In addition,
\citet{2013ApJ...778..133L}
highlighted the differences between star formation relations {\bf within}
and {\bf between} clouds, and
\citet{2016arXiv160507623P} has summarized the issues in comparing
surface densities of SFR and gas between nearby and more distant
clouds.
Furthermore, plotting SFR versus mass introduces a correlation
because both are proportional to the square of distance.
Cognizant of these issues,
we focus on SFE (distance cancels out)
and deal with the masses contained within
the entire cloud or a dense clump within the cloud, rather than
using surface density {\it within} a cloud or clump.
}
First, we consider whether comparable regions have been
selected. For the nearby clouds, a region above a threshold extinction,
generally $\mbox{$A_V$} = 2$ mag, was selected for surveys of YSOs
\citep{2009ApJS..181..321E}.
For this paper, the clouds were defined by maps of \mbox{$^{13}$CO}\ \jj10\ emission, followed by
extrapolation to zero emission. Our estimates of an effective threshold
for the Galactic Plane clouds range from $\mbox{$A_V$} = 1$ to 3 mag.
This may be in addition to extinction of several magnitudes associated with layers
of atomic and molecular gas exterior to the \mbox{$^{13}$CO}\ boundary limited by photodissociation.
The mean density of the selected clouds for the Galactic Plane
sample is about 300 \cmv, lower than the average for the nearby
clouds of about 800 \cmv.
The measure of ``dense" gas in the nearby
clouds was an extinction threshold of $\mbox{$A_V$} = 8$ mag. Using the
typical rms noise in the BGPS survey of 0.1 mJy/beam
\citep{Ginsburg:2013},
a dust temperature of 20 K, and OH5 dust opacities
\citep{Ossenkopf:1994},
a 3 $\sigma$ detection limit translates to $\mbox{$A_V$} = 7$ mag,
comparable to that used to define dense regions in the nearby clouds.
A second analysis, using the actual distribution of rms noise
in the BGPS yields a slightly smaller threshold, $\mbox{$A_V$} = 4.6$ mag
at 90\% completeness (B. Svoboda, pers. comm.).
However, the mean volume density of the BGPS sources in the
Galactic Plane sample is
only \eten3 \cmv, substantially lower than the mean density of
the regions above $\mbox{$A_V$} = 8$ for the nearby clouds ($\mean{n} = 6\ee3$
\cmv). Thus, the galactic plane ``dense" gas is not as dense
as that defined in the nearby clouds. For more distant targets, where the
molecular clouds may subtend angles less than 3\arcmin, the mass and densities
may reflect the full cloud rather than high density fragments \citep{Dunham:2011, Battisti:2014}.
A common feature of all observational studies of star formation is
a low star formation efficiency. Only a small fraction of
the gas in galaxies is actively forming stars. The fractional mass of
young stars in molecular clouds is around 5\%
(\citealt{2009ApJS..181..321E,2015ApJS..220...11D}).
In the extragalactic context, star formation efficiency (SFE) refers in
fact to the slow {\it pace} of star formation, the star formation
rate per unit mass, as reflected in \mbox{$t_{\rm{dep}}$}, the inverse of SFE, being
1-2 Gyr.
Parallel to our equations for SFR, we consider relations for the SFE
and \mbox{$t_{\rm{dep}}$}\
\begin{equation}
\mbox{\rm SFE} = \mbox{\rm SFR}/X = A X^{n-1}
\end{equation}
\begin{equation}
\mbox{$t_{\rm{dep}}$} = X/\mbox{\rm SFR} = A^{-1} X^{1-n}
\end{equation}
and their logarithmic versions.
\begin{figure*}
\center
\includegraphics[scale=0.6, angle=0]{sfevsmass.eps}
\caption{
Plot of the logarithm of SFR per mass of molecular gas vs. the logarithm of
the mass of molecular gas. The red points are from this paper and the black
points are for nearby clouds, taken from
\citet{Evans:2014}.
The black horizontal line is the mean value for the nearby clouds and
the red horizontal line is the mean value for the data in this paper.
The error bars at the far left are the standard deviations of $\log(\mbox{\rm SFE})$.
}
\label{sfevsmass}
\end{figure*}
\begin{figure*}
\center
\includegraphics[scale=0.6, angle=0]{sfevsdense.eps}
\caption{
Plot of the logarithm of SFR per mass of dense gas vs. the logarithm of
the mass of dense gas. The red points are from this paper and the black
points are for nearby clouds, taken from
\citet{Evans:2014}.
The black horizontal line is the mean value for the nearby clouds and
the red horizontal line is the mean value for the data in this paper.
The error bars at the far right are the standard deviations of $\log(\mbox{\rm SFE})$.
}
\label{sfevsdense}
\end{figure*}
\begin{figure*}
\center
\includegraphics[scale=0.6, angle=0]{sfevsmtff.eps}
\caption{
Plot of the logarithm of SFR divided by the cloud mass over
the free-fall time vs. the logarithm of the cloud mass over the
free-fall time. The red points are from this paper and the black
points are for nearby clouds, taken from
\citet{Evans:2014}.
The black horizontal line is the mean value for the nearby clouds and
the red horizontal line is the mean value for the data in this paper.
The error bars at the far right of these lines
are the standard deviation of the $\log(\mbox{\rm SFR}/(\mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$}))$ for the
two data sets.
}
\label{sfevsmtff}
\end{figure*}
\input tablesfe.tex
We plot the SFR per entity for the three main entities used to
predict SFR in Figures \ref{sfevsmass} to \ref{sfevsmtff}.
The three entities are cloud mass (\mbox{$M_{\rm cloud}$}), dense gas
mass (\mbox{$M_{\text{dense}}$}), and cloud mass per free-fall time
(\mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$}). In extragalactic parlance, the first two are
``star formation efficiencies", or the reciprocal of the depletion time
\added{, measured in Myr}.
The last is unitless \added{and equal to \mbox{$\epsilon_{\rm ff}$}}.
We will refer to these generically as the SFE.
The black points are from the nearby
clouds, while the red points are from this paper.
The means and standard deviations (both in the log) are given in Table \ref{sfetab}.
The following facts are apparent.
The standard deviations in the SFE for Galactic Plane clouds
are comparable (about 0.5 dex)
for all the relations
(the figures are plotted on the same scales so the
eyeball estimate of the scatter is meaningful).
\added{The mean of the SFE for the Galactic Plane clouds
is lower by 0.8 in the log than that for the nearby clouds when cloud mass is used, but very similar (lower by 0.13 in the log) when dense gas mass is used.}
The agreement of the SFE for the dense gas relation is striking because
the method to estimate the SFR for the Galactic Plane sample is very
different (MIR emission) from that used for the nearby clouds (YSO counts)
so disagreement could be expected. Indeed, we know that MIR emission
seriously underestimates the SFR for the nearby clouds, and we
\replaced{could}{might}
expect an underestimate of 0.3 to 0.5 in the log for the Galactic Plane
sample, based on the discussion in \S 3.2.3. The agreement \deleted{here}
\added{for the dense gas}
encourages us that the Galactic Plane sources have SFR sufficiently high
that the MIR tracer is not greatly underestimating the SFR.
\added{The alternative explanation, that the SFR and the dense gas
mass are both underestimated by the same amount, is implausible because
the mean density of the Galactic Plane clumps is less than that of the
dense clumps in the nearby clouds, as noted above. If anything, the masses
of gas as dense as those in the nearby clouds is over-estimated in the
Galactic Plane clouds.}
\deleted{The mean values of SFE are lower for the Galactic Plane sample, but least
so for the dense gas, for which they agree to within 0.13 in the log. }
The SFE for the Galactic Plane clouds is about \eten{-3}
\added{Myr$^{-1}$}, which is
consistent with a depletion time of 1 Gyr, similar to that for galaxies
as a whole.
\added{
\citet{2011ApJ...729..133M} used WMAP data to find the most luminous \mbox{\ion{H}{2}}\
regions in the Galaxy and to compute star formation rates.
His sample is clearly biased to the regions with the highest star formation rates.
After connecting these \mbox{\ion{H}{2}}\ regions to molecular gas using catalogs
of GMCs, he computed star formation efficiences. For consistency, we used
our equation to recompute his star formation rates, but they are only slightly
lower than the rates he gave.
His data are added in Figure \ref{murray} as the blue points;
they have a mean
SFE similar to that of the nearby clouds, and they further extend the
dispersion among Galactic Plane clouds.
We cannot do a full analysis with Murray's sources as no uncertainties
are given, the identification with molecular gas is not clearly described,
and there is no information on the dense gas. However, a simple mean value
of SFE including his sample and ours yields $\mean{\log {\rm SFE}} = -2.52
\pm 0.72$ where SFE is measured in Myr$^{-1}$ as usual, bringing the mean
for the Galactic Plane clouds closer to that for the nearby clouds.
\begin{figure*}
\center
\includegraphics[scale=0.6, angle=0]{murray.eps}
\caption{
Plot of the logarithm of SFR per mass of molecular gas vs. the logarithm of
the mass of molecular gas. The red points and black points are as
in the previous figure. The blue points are from
\citet{2011ApJ...729..133M}.
The blue horizontal line is the mean value and error bar at the far right is the standard deviation for the data from
\citet{2011ApJ...729..133M}.
}
\label{murray}
\end{figure*}
}
\deleted{This also
implies that the low SFE for the Galactic Plane clouds in Fig. \ref{sfevsmass} is real.}
\deleted{Far from being inefficient star formers, the nearby clouds
are efficient relative to the Galactic Plane clouds, although
this statement carries the caveat that the regions appear to be
denser in the nearby clouds as discussed above.}
\replaced{A possible cause for the discrepancy of SFE found in nearby clouds and the Galactic Plane sample
is radiative and mechanical feedback from massive stars. }
{The main feature that emerges from the data is a large dispersion
in SFR per cloud mass for both the nearby and the Galactic Plane clouds.
The mean SFE depends on sample selection. When the dense gas mass is
used to determine SFE, the dispersion is less and the mean values
for nearby and Galactic Plane clouds agree. Differences between the
nearby clouds and the Galactic Plane clouds may
be related to different feedback effects.}
Other than the Orion cloud, the local regions are primarily generating low mass stars \added{because the IMF is not fully sampled}.
In such low mass star forming regions, protostellar outflows are the primary feedback process.
Such outflows
can perturb the cloud structure
over scales up to several parsecs over a restricted volume set by the jet opening angle.
However, their effect on the overall cloud structure is limited and
incapable of suppressing the SFE. Massive young stars provide a much stronger energy input to the
cloud primarily through far-UV radiation fields that drive expanding \mbox{\ion{H}{2}}\ regions \citep{Matzner:2002,
Dale:2014}.
These processes can impact the SFE by photoionizing part of the cloud or simply
modifying the conditions \replaced{that are}{to make them} less suitable for star formation.
Some support for this might be found in the apparently higher mean
value of $\alpha$ in the Group 2 clouds (Table \ref{grouptab2})
but the dispersion is quite large. In addition, the mean values for
$\alpha$ for the nearby clouds and the Galactic Plane clouds are
indistinguishable.
\added{The most likely effect of the massive stars is to confuse
observational measures by removing or ionizing the molecular gas,
leading to Group 2 or 3 sources, configurations which are not seen in the nearby clouds.}
\section{Comparison to Extragalactic Results}\label{exgal}
Studies of other galaxies have explored the same questions
as have been addressed in this paper. In this section, we compare
those results to those in the Milky Way. We continue to use the SFE
to avoid the strong apparent correlation introduced by the fact that
both mass and SFR generally scale as size, hence distance, squared.
For the extragalactic studies, clouds are generally not resolved, so
we use \mbox{$M_{\rm mol}$}\ to represent the aggregate mass in molecular gas, while
\mbox{$M_{\rm dense}$}\ continues to represent the aggregate mass of dense gas.
\begin{figure*}
\center
\includegraphics[scale=0.6, angle=0]{galmcloud.eps}
\caption{
Plot of the logarithm of SFR per mass of molecular gas vs. the logarithm of
the mass of molecular gas, showing averages and standard deviations (in
the logs). The red filled circle is from this paper and the open black
circle is for the nearby clouds, taken from
\citet{Evans:2014}.
Three green points are plotted for resolved studies
in nearby galaxies. The lowest point (square) is from
\citet{2013AJ....146...19L}
who observed with resolutions from 0.2 to 1.4 kpc.
The other two points (triangles) are from
\cite{2015ApJ...810..140C},
who observed M51 with 1 kpc resolution; the higher point is
for the outer galaxy and the lower point is for inner
($r < 1.66$ kpc) part of M51.
The blue point (pentagon) represents the normal galaxies
while the cyan points (hexagons) represent the (U)LIRGs from
\citet{2015ApJ...805...31L}.
Two values are plotted,
the higher one using the conversion from CO to molecular gas of
$\alpha_{\rm CO} = 0.8$ \msun (K km s$^{-1}$ pc$^2$)$^{-1}$,
while the lower one uses the same value as for the normal galaxies,
$\alpha_{\rm CO} = 4.6$ \msun (K km s$^{-1}$ pc$^2$)$^{-1}$.
}
\label{galmcloud}
\end{figure*}
\begin{figure*}
\center
\includegraphics[scale=0.6, angle=0]{galmdense.eps}
\caption{
Plot of the logarithm of SFR per mass of dense gas vs. the logarithm of
the mass of dense gas, showing averages and standard deviations (in
the logs). The red point (filled circle) is from this paper and the black
point (open circle) is for the nearby clouds, taken from
\citet{Evans:2014}.
The magenta point (heptagon) comes from
\citet{Heyer:2016}.
The two green points (triangles) are from
\cite{2015ApJ...810..140C},
who observed M51 with 1 kpc resolution; the higher point is
for the outer galaxy and the lower point is for inner
($r < 1.66$ kpc) part of M51.
The blue point (pentagon) represents the whole, normal galaxies
while the cyan point (hexagon) represents the whole (U)LIRGs, both from
\citet{2015ApJ...805...31L}.
}
\label{galmdense}
\end{figure*}
The first comparison is for overall molecular gas. Figure \ref{galmcloud}
plots the mean and standard deviation (in log space) of the
SFE for the nearby clouds and the Galactic Plane clouds, along
with data from other studies.
The green points show results from spatially resolved studies of nearby
galaxies. The lowest green point is from
\citet{2013AJ....146...19L},
who summarized the HERACLES data on 30 ``nearby" ($d < 22$ Mpc)
disk galaxies. The SFE plotted for the galaxies is actually the log
of $1/\mean{\mbox{$t_{\rm{dep}}$}}$, but calculation of \mean{SFE}\ from their
data tables produces a nearly identical result. The mass scale is
that of their resolution, obtained from the surface density and
resolution in their Table 1, converted to mass, converted to logs,
and averaged, resulting in a characteristic log(\mbox{$M_{\rm mol}$}) of
$6.63\pm0.59$.
The other two green points are from the study of M51 with 1 kpc
resolution as presented by
\citet{2015ApJ...810..140C};
the highest green point is for the outer parts of M51, while
the lower one at higher masses is from the inner 1.66 kpc.
Finally, whole galaxies can be plotted, based on a sample of 115 disk
galaxies in
\citet{2015ApJ...805...31L}.
The normal galaxies are plotted in blue and the
\ulirg s are plotted in cyan. The higher cyan point uses the
\ulirg\ conversion from CO to molecular gas while the lower
point uses the same conversion as used for the normal galaxies.
The plot shows a decline in SFE from the nearby
clouds to the Galactic Plane clouds to the extragalactic regions.
The standard deviation in the SFE also decreases, presumably due
to averaging over more and more star formation regions
(cf \citealt{Kruijssen:2014,Vutisalchavakul:2014}).
A rise in SFE is seen for the \ulirg s, especially if the lower,
\added{(U)LIRG,} value
for the conversion factor is used, for which the SFE exceeds even
the value in the nearby clouds.
The same plot can be made for the dense gas, again using
averages and standard deviations in logs, using masses based
on HCN emission and
\begin{equation}
\mbox{$M_{\rm dense}$}(\msun) = 10 L_{\rm HCN}
\end{equation}
where $L_{\rm HCN}$ is the observed luminosity of HCN \jj10\ emission
in K km s$^{-1}$\ pc$^2$
\citep{2015ApJ...805...31L}.
The \ulirg s in \citet{2015ApJ...805...31L} are again plotted in cyan,
but only one value of conversion from HCN emission to dense gas
mass was used by \citet{2015ApJ...805...31L}.
Observations of HCN in M51 with a resolution of 1 kpc bridge the gap
between the Milky Way clouds and whole galaxies. The green points
are based on data in
\citet{2015ApJ...810..140C};
the higher point is the average for the outer part, while the
lower point is for the inner part of M51.
A magenta point comes from the study of Milky Way clouds by
\citet{Heyer:2016}
who derived star formation rates from mid-infrared luminosities of Class I protostars in a sample of dense
clumps identified in the Atlasgal survey of submillimeter dust emission
\citep{Heyer:2016}.
Figure \ref{galmdense} shows that the SFE measured for dense gas
varies little from local clouds to whole galaxies.
These results are very consistent with those found by
\citet{2015ApJ...805...31L},
but extend them down to the scales of individual clouds in the Galaxy,
and even down to the scales of nearby clouds, where the methods of measuring
both SFR and \mbox{$M_{\rm dense}$}\ are quite different. The basic conclusion is that
the mass of dense gas is the most stable predictor of SFE across
seven orders of magnitude in dense gas mass.
The grand average of the data points in Figure \ref{galmcloud} yields
$\mean{{\rm log\ SFE(Myr^{-1})}} = -2.83\pm0.42$
\explain{corrected to add the units and that the mean is in the log}
if the \ulirg s with the Milky Way
value for $\alpha_{\rm CO}$ are included, or $-2.73\pm 0.59$ if the
point using the \ulirg\ value for $\alpha_{\rm CO}$ is included.
In contrast, the grand average for the dense gas data in Figure
\ref{galmdense} is $\mean{{\rm log\ SFE(Myr^{-1})}} = -1.82\pm 0.19$,
with a standard deviation
one-third that when all molecular gas is used.
\replaced{However,}{Despite the remarkable consistency shown in
Figure \ref{galmdense},}
\mbox{$M_{\rm dense}$}, at least as measured by HCN, is not perfect;
studies of our Galactic Center and nearby galaxies
using HCN \jj10\ as the tracer of dense gas
\citep{Usero:2015,2015ApJ...810..140C,2013MNRAS.429..987L}
have shown that the SFE for dense gas also depends on environment.
\added{This dependence can be seen in Figure \ref{galmdense} where the
lowest point is for the inner part of M51. Recent studies show that this
decrease in SFE is a function of various galactic properties, such
as molecular fraction \citep{2016ApJ...822L..26B}. The most plausible
explanation for lower SFE for gas probed by HCN emission in regions
of high density is that the criterion for rapid star formation changes.
\citet{2014MNRAS.440.3370K} provide an exhaustive examination of
both global and local mechanisms to explain the low SFE for gas
probed by \mbox{{\rm NH}$_3$}\ emission in the central
molecular zone of our Galaxy. They favor episodic star formation among
global mechanisms and a much higher threshold density for star formation
among local mechanisms. The higher threshold is caused by the greatly increased
turbulence in the central molecular zone, pushing the density threshold
to values as high as $n \approx \eten7$ \cmv, far above the density needed
to produce strong emission from HCN \jj10\
\citep{1999ARA&A..37..311E,2015PASP..127..299S}.
In these environments, probes of higher densities, such as higher $J$ transitions of HCN, may
be more diagnostic of the mass of gas above the threshold density.}
Further studies of central regions of galaxies, including our own,
are needed to refine the criteria for star formation.
\added{
There are many reasons why measures of SFE in the range of regions
plotted in figures \ref{galmcloud} and \ref{galmdense} might differ.
The methods used to determine cloud mass and dense gas mass differ and
the meaning of ``dense'' is not the same for all. Extragalactic studies,
especially whole galaxy observations, average over a huge range of
physical conditions and sample regions at different stages of
evolution, while studies of individual clouds are snapshots \citep{Kruijssen:2014}. The fact that the differences in SFE
across this range are so small when measured against dense gas is
an important clue for our understanding of what controls star formation.
The consistency suggests that simulations of star formation and galaxy
evolution that require higher densities to initiate star formation
are on the right track.
A picture in which most molecular clouds are unbound and only small, dense parts of the clouds are sites of star formation
\citep{2011MNRAS.413.2935D}
or a picture in which dense clump formation \replaced{is delayed}{from more
diffuse molecular gas is a continuous process \citep{2013ApJ...773...48B}} can potentially explain our results.
Pictures in which a high density
{\it contrast} is needed for star formation
may be able to incorporate the lower SFE for dense gas in galaxy centers
\citep{2016ApJ...822L..26B}.
A combination of these ideas may produce a more unified picture.
Future work on other galaxies with high
spatial resolution, using dense gas tracers, will test these relations
further.
}
\section{Summary}
We compiled a sample of Galactic \gmc s that are associated with \mbox{\ion{H}{2}}\
regions and estimated their properties
and SFR. The analysis of \gmc s, \mbox{\ion{H}{2}}\ regions,
and SF tracers (both radio continuum and MIR emission) shows
different degrees of associations between molecular gas and
star formation. We classified the \gmc s into different groups:
\gmc s with embedded \mbox{\ion{H}{2}}\ regions, \gmc s
with overlapping \mbox{\ion{H}{2}}\ regions, and \gmc s with separated \mbox{\ion{H}{2}}\ regions.
We did not use the last group because association between molecular gas
and star formation was too uncertain.
The sample was used to test relations between SFR and properties
of \gmc s. We tested \replaced{five}{four} different models of star formation.
No significant correlation was found between
\mbox{$\epsilon_{\rm ff}$}\ and $\mbox{$t_{\rm ff}$}/\mbox{$t_{\rm dyn}$}$.
Significant correlations exist between SFR and \mbox{$M_{\rm cloud}$}, \mbox{$M_{\text{dense}}$},
and $\mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$}$.
The relation between SFR and \mbox{$M_{\text{dense}}$}\ is consistent with linear, while
the other two are significantly \replaced{sub-linear}{non-linear}, unlike extragalactic
relations or the theoretical model by \citet{Krumholz:2012}
for $\mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$}$.
Combining the data
\added{on Galactic Plane clouds presented in this paper}
with that on
nearby clouds shows that the star
formation efficiency of the nearby clouds is higher when efficiency
is measured versus \mbox{$M_{\rm cloud}$}\ or \mbox{$M_{\rm cloud}$}/\mbox{$t_{\rm ff}$}. The efficiency per
mass of dense gas is very similar for the nearby clouds and the Galactic
plane clouds.
Adding extragalactic studies, we can extend the range of relevant
mass scales over 7 orders of magnitude. The star formation efficiency
for dense gas shows remarkable stability over this range, varying
over a factor of 4, while that
for total molecular gas varies by a factor of 40.
\added{The standard deviation in the log of the SFE(Myr$^{-1}$)
decreases by about a factor
of 3 to a value of 0.19 when dense gas mass, rather than molecular mass, is used.}
\medskip
\added{We thank the anonymous referee for a careful reading and excellent
suggestions which have improved the paper. We also thank C. McKee,
C. Federrath, G. Parmentier, M. Fall, and B. Ochsendorf for comments.}
We are grateful to the BGPS team for sharing ideas and information
over many years. We particularly thank B. Svoboda for calculations
to characterize the extinction threshold for the BGPS sample.
H. Chen kindly provided data on galaxies.
This work was supported by NSF
grant AST-1109116 to the University of Texas at Austin.
MH acknowledges support from NASA ADAP grant NNX13AF08G
to the University of Massachusetts.
|
\section{Introduction}
\justify
While modeling and optimizing the energy consumption of different applications and mobile systems have been the active research interest for a decade~\cite{6157576,hoquecsur2015}, the overall performance of the smartphone power management systems has not received significant attention yet. The diverse set of smartphone models available today are powered with batteries of different capacity volumes and technologies, such as Lithium-Ion and Lithium-Polymer. They employ different charging mechanisms to charge their batteries and rely on different state of charge (SOC) estimation techniques.
The growth of smartphone battery size has been linear with time. Charging large batteries with traditional charging techniques may take very long time. In addition, the context may not allow a user to charge long enough time. Therefore, it is necessary that the battery should be charged to some reasonable amount, e.g., 30-50\%, within a short amount of time. Consequently, users are increasingly relying on a number of Fast charging techniques from Qualcomm (Quick~\cite{qulcomm}) and Samsung (Fast). Nevertheless, the quality of charging plays an important role in the longevity of smartphone batteries. For example, if a battery is charged over the maximum battery voltage, the resulting chemical reactions may reduce the capacity significantly~\cite{Choi2002130} and increase the battery temperature beyond the safety limit. Understanding the inefficiency of the energy source, and other related contributing factors can enable better optimization of the applications, systems, and more accurate power consumption modeling.
\marginpar{%
\vspace{-21pc} \fbox{%
\begin{minipage}{0.925\marginparwidth}
\underline{\textbf{Findings and Contributions}}
{\small
\vspace{2mm}
\underline{\textbf{First:}} Most of the devices apply Constant Current-Constant Voltage (CC-CV) charging. A number of devices of latest models use Fast charging. Quick charging utilizes higher voltage and constant current during the CC phase of charging, whereas Samsung's Fast charging employs pulse charging.
\vspace{1mm}\underline{\textbf{Second:}} From SOC updates, we compute the charging time curves, and derive the SOC estimation techniques used by the devices. The SOC update time of voltage-based fuel gauges fluctuates during the CC-phase of charging, whereas Coulomb counter-based devices provide updates at periodic intervals.
\vspace{1mm} \underline{\textbf{Third:}} From battery voltage, we estimate that 85\% devices had their battery capacity reduced by 1-10\%. We further demonstrate that the battery health information provided by the system may not be appropriate.
\vspace{1mm} \underline{\textbf{Fourth:}} Both devices and users contribute to inefficient charging of the batteries. Quick charging charges a battery to over voltage. There are two forms of inefficient charging by the users; charging and actively using devices at the same time and longer over night charging.
}
\end{minipage}}\label{sec:sidebar} }
Given the number of smart mobile devices available on the market, it is not feasible to investigate their charging and battery properties, and the performance of the charging methods on batteries in a laboratory environment. Although there are studies on users' charging behavior~\cite{Ferreira:2011,Banerjee:2007}, it is not well understood how this battery and charging information could be presented in more meaningful ways to the users and mobile vendors other than just the battery level.
In this article, we explore a large battery analytic dataset comprising various battery sensor information from 30K devices of 1.5K unique smartphone models collected by the Carat~\cite{Oliner:2013} application. We explore their battery voltage behavior, charging rate and charging time, and demonstrate how these properties can be used to expose the characteristics of their power management systems. We identify their charging mechanisms, SOC estimation techniques and battery properties, and the distribution of these properties among the devices. To the best of our knowledge such comprehensive study on a large smartphone battery dataset has not been presented earlier. Our findings and contributions are listed in the sidebar.
The rest of the paper is organized as follows. Next section provides an overview on smartphone's power management system and describes the crowdsourced dataset. The subsequent sections explore the dataset and identify the characteristics of various power management techniques used by the smartphones and properties of smartphone batteries while charging. Before concluding the paper, we also discuss user behavior in charging their smartphones.
\section{Smartphone Power Management \& {Dataset}}
\label{sec:two}
The charger and three different ICs, a fuel gauge, a charging controller, and a protection IC, manage the charging of a mobile device. The charging controller is hosted in the device and the protection IC resides in the battery. The fuel gauge functionality may be distributed between the device and the battery. The fuel gauge determines the runtime battery capacity, i.e., SOC or battery level, using open circuit voltage, coulomb counter, or a combined mechanism of these two~\cite{Rezvanizaniani2014110}. It senses battery voltage, temperature, and charge or discharge current to/from the battery pack. At the same time, it also provides feedback to the charging IC. The charging controller applies the charging algorithm, such as CC-CV, and uses the fuel gauge provided information to control the charging current, voltage, and to terminate the charging. Finally, the protection IC protects the battery from over voltage or current from the device.
\subsection{Dataset and Pre-processing}
The Android Battery Manager collects charging and battery information from the fuel gauge (see Table~\ref{tab:charger_battery_info}) and broadcasts as events.
Carat collects information from mobile devices as samples with a broadcast receiver. A sample structure can be defined as $S = (t, (a_{1} : v_{1}), (a_{2} : v_{2}), (a_{3} : v_{3})...(a_{n} : v_{n}))$, where $t$ is the epoch timestamp of a SOC update event and $(a_{i} : v_{i})$ are the attribute and value pairs. From all the information collected in a sample, we consider the timestamp, SOC, battery voltage, battery health, battery temperature, charging status, charger type, and the screen status attributes.
\begin{margintable}[1pc]
\begin{minipage}{\marginparwidth}
\vspace{22mm}
\begin{tabular}{|p{12mm}|p{24mm}|}
\hline
\textbf{Attribute} & \textbf{Value}\\\hline
charging status & plugged/\break unplugged\\\hline
charger & ac main / usb\\\hline
battery voltage & 4.2 V\\\hline
battery temp & 29$^\circ$C \\\hline
battery health & good,overheat, over~voltage\\\hline
battery level & 0.99 (99\%)\\\hline
\end{tabular}
\caption{Charging and battery information of Android devices.}
\vspace{1mm}
\label{tab:charger_battery_info}
\vspace{7mm}
\end{minipage}
\end{margintable}
\marginpar{%
\fbox{%
\begin{minipage}{0.925\marginparwidth}
\begin{equation}
\label{eq:one}
C = \frac{36\times \Delta SOC}{{t}_{2}-{t}_{1}}
\end{equation}
\end{minipage}}\label{sec:sidebar2} }
\begin{algorithm}[t]
\caption{Smartphone Battery Analytics}\label{alg:cevent}
\begin{algorithmic}[1]
\Procedure{ChargingEvents~}{ rdd[reducedSample] }
\State userSamps$\leftarrow$ groupById (reducedSample)
\For{\textbf{each} \texttt{user} $\in$ \texttt{userSamps.\_1}}
\State filSamp $\leftarrow$ Filter (``ac'', ``soff'', userSamp)
\State srtSamp $\leftarrow$ SortByTime (filSamp)
\State samPairs $\leftarrow$ Pair (srtSamp)
\State chargEvent = 1
\For{\textbf{each} \texttt{pair} $\in$ \texttt{samPairs}}
\State $\Delta$SOC = onePair.\_1.soc - onePair.\_2.soc
\State $\Delta$t = onePair.\_2.time - onePair.\_1.time
\State $C$ = $\frac{36\times\Delta SOC}{\Delta t}$
\If{\texttt{($C<=0.03$)}}
\State Label (chargEvent, $\Delta$t, C, onePair.\_1)
\State chargEvent += 1
\State $\Delta$t = 0
\State $C$ = 0
\State Label (chargEvent, $\Delta$t, C, onePair.\_2)
\Else
\State Label (chargEvent, 0, 0, onePair.\_1)
\State Label (chargEvent, $\Delta$t, C, onePair.\_2)
\EndIf
\EndFor
\EndFor
\EndProcedure
\end{algorithmic}
\label{alg:one}
\end{algorithm}
\vspace{-2mm}
\subsection{Charging Events}
We analyze a subset of Carat dataset collected over ten months of size more than 200GB, in Spark platform~\cite{spark}. From the reduced charging samples, we need to construct SOC vs. battery voltage and SOC vs. charging rate curves. In order to generate such curves, it is essential that all the samples of a curve belong to the same charging event. We generate charging events for every user as described in Algorithm \ref{alg:one}. The reduced samples in a Resilient Distributed Dataset allows distributed computation on the samples in a cluster of 7 machines each having 8 CPU cores and 30GB of RAM.
\vspace{-1mm}
First, the reduced samples are grouped according to the user ID. We next sort the samples of a user according to timestamp and construct pairs of two consecutive samples. From these pairs, we compute one percent charging time and corresponding C\footnote{\footnotesize{A battery with 2000 mAh capacity will be charged with 2000 mA current at 1.0C rate and it will take 1 hour to complete the charging.}} rate. If the charging rate is 1C, then a mobile device spends 36 s to charge one percent. As a result, the charging rate used to charge 1\% of the battery can be presented as \eqref{eq:one}, where $t_{1},~t_{2}$ are the timestamps of two consecutive samples and $\Delta$SOC is the difference of the battery levels reported in those samples.
Ideally, a charging event begins by connecting a device with the charger and ends when the device is disconnected from the charger. Constructing charging events in this way is difficult from the dataset, as a user may turn on/off the phone while charging and turn on when the battery is charged to a reasonable capacity. The charging algorithms terminate charging when the charging rate is 0.07C and a mobile device spends 514 seconds maximum to charge one percent. However, the tablets may take even longer time during the CV phase of charging and therefore, we consider 0.03C as the terminating charging rate as stated in the algorithm. Finally, we label the samples with incremental numeric charging events, and update them with C rate and one percent charging time.
\begin{table}[t]
\begin{center}
{\small
\begin{tabular}{|p{17mm}|p{12mm}|p{9mm}|p{11mm}|p{13mm}|}
\hline
\textbf{Battery\break Health} &\textbf{Samples} & \textbf{Users}&\textbf{Models}&\textbf{Charging\break Events}\\\hline
Good & 3.3M&30K & 1.5K&180K \\\hline
Over voltage& 1554&90 & 15 & 100\\\hline
Overheat & 665 &165 & 40 & 200\\\hline
\end{tabular}}
\caption{Number of charging samples, corresponding users, device models, and charging events from Carat dataset.}
\label{tab:sample_stat}
\end{center}
\end{table}
\begin{marginfigure}[1pc]
\begin{minipage}{\marginparwidth}
\vspace{-25mm}
\subfigure[CC-CV \& DLC]{\label{fig:battery_voltage}\includegraphics[width=1.0\linewidth,height = 1.0\linewidth]{figures/voltage_compare_goodall.pdf}}
\subfigure[DLC \& Fast Charging]{\label{fig:charge_volt_compare}\includegraphics[width=1.02\linewidth,height = 1.1\linewidth]{figures/fast_charging_compare.pdf}}
\caption{Battery voltage curves while charging via AC with different charging techniques.}
\vspace{-3mm}
\label{fig:battery_voltage_all}
\end{minipage}
\end{marginfigure}
\section{Charging Techniques}
\label{sec:controller}
We first analyze battery voltage behavior as charging proceeds and derive the charging techniques used by the smartphones. Next, we analyze the charging rates. In this section, we consider the charging events of good samples and further consider screen off samples from the events to reduce device usage bias in the analysis.
\subsection{Battery Voltage}
From the charging events, we construct model specific charging voltage curves containing battery voltage information for every battery level update as shown in Figure~\ref{fig:battery_voltage}. The figure shows the \textit{initial} and \textit{final} voltages for each unique user. The initial voltage is the minimum voltage required to power up a device (when the battery level is 1\%) and final voltage is the maximum voltage when the battery is charged to 100\%. A battery should not be charged to more than this voltage.
The final voltages observed in the dataset broadly can be classified into two categories; 4.2$\pm$0.05 and 4.35$\pm$0.05 V. However, using CC-CV, a battery is charged to a maximum 4.2 V. It turns out that this voltage behavior is because of a different charging mechanism used by the devices. Although CC-CV is the well-known charging mechanism, some devices charge batteries to an extra 0.15 V to reduce charging hardware implementation complexity \cite{Thanh:2012}. This mechanism is also called Double Loop Control (DLC). We compare the final voltage of multiple devices of same model and find that 38\% of the devices use CC-CV and 59\% use DLC.
Figure~\ref{fig:battery_voltage} presents the distribution of voltage, and the relationship between battery voltage and battery level for the CC-CV and DLC methods. From the charging events, we take the median battery voltage for each battery level of the devices of two categories and plot them. It is shown that battery voltage increases almost linearly until the battery voltage reaches to the maximum, i.e., 4.2/4.35 V. Figure~\ref{fig:battery_voltage} further shows that the final voltage can be different for different users. Later, we demonstrate how this final voltage can be used to estimate capacity loss of the battery.
We next explore model specific voltage curves. Figure~\ref{fig:charge_volt_compare} compares the median voltage curves of a number of devices. We notice that Galaxy S3 and S4 curves are similar as shown in Figure~\ref{fig:battery_voltage}, whereas the voltage curves of Galaxy S6 Edge, and Nexus 6 devices have unique characteristics. Unlike S3/S4, the battery voltage of Nexus 6 increases to a maximum 4.48 V and then the CV phase begins. On the other hand, the battery voltage of Edge 6 increases sharply until the battery is charged to 30\%. After that battery voltage increases and decreases alternatively until reaches to the maximum 4.35 V. This hints that Edge 6 applies Fast pulse charging~\cite{Thanh:2012}. In other words, the charging current alternates between two rates. Among 30K, only 3\% of the devices use Quick and Fast charging.
\vspace{-1mm}
\subsection{Charging Rate}
\label{subsec:three_two}
Charging algorithms may apply different charging currents to charge the batteries. Android operating systems, however, do not expose charging rate or current through any API. We use the charging C rates estimated with Algorithm~\ref{alg:one}. In the earlier section, we have identified that batteries are charged in two phases, in general. In this section, we investigate the behavior of charging rates and present two additional charging mechanisms.
\noindent\textbf{CC-Phase: }
During the first phase, the batteries are charged to 50-90\% of total capacity, depending on the models. Figure~\ref{fig:charging_rates} illustrates the charging rates of 30K users. Mobile devices are charged mostly at rates smaller than 0.7C via AC and higher than the rate USB provides. The rates are typically constant during the first phase and vary among different models. There are a number of takeaways from this figure.
\begin{marginfigure}[1pc]
\begin{minipage}{\marginparwidth}
\vspace{-2mm}
\subfigure[CC-phase rates]{\label{fig:charging_rates}\includegraphics[width=1.0\linewidth,height = 0.9\linewidth]{figures/charging_rates_all.eps}}
\vspace{-3mm}
\subfigure[CV-phase rates]{\label{fig:cccv_dlc_rates}\includegraphics[width=1.02\linewidth,height = 0.9\linewidth]{figures/trickle_rates.eps}}
\caption{Charging rates of the devices.}
\label{fig:charging_all}
\end{minipage}
\includegraphics[width=1.0\textwidth,height = 0.7\linewidth]{figures/onePertime.eps}
\caption{Charging time curves of five smartphone models.}
\label{fig:onePercurve}
\end{marginfigure}
\begin{itemize}
\item Figure~\ref{fig:charging_rates} shows that a few numbers of rates are higher than 1C for smartphones. By examining the charging events, we have identified that 1\% of the devices use charging rates higher than 1C. These are the Fast charging enabled devices.
\item The tablets are charged at smaller rates than the smartphones, even when the AC main or wall chargers are used. One possible explanation is that tablets have larger batteries than the smartphones and their charging rates do not scale up according to the capacity.
\item Figure~\ref{fig:charging_rates} shows that 30\% and 20\% of the charging rates via USB are below 0.1C for the tablets and smartphones respectively. In the case of AC charging, 20\% of the rates are below 0.1C for the tablets, whereas such rates are less than 5\% for the smartphones. This is because a small number of devices apply the CV charging to charge the first 10\% and then the CC charging begins. This helps to restore the charge of deeply depleted cells inside the battery~\cite{Dearborn:2012}. Therefore, it took a longer time for an actual 1\% increment.
\end{itemize}
\noindent\textbf{CV-Phase:}
Figure~\ref{fig:cccv_dlc_rates} shows the charging rates during the second phase of charging. Both CC-CV and DLC methods trickle down the charging current gradually to less than or equal to 0.1C. Figure~\ref{fig:cccv_dlc_rates} also shows that charging rates can be around 1.0C. It turns out out that some device models use CC charging after the CV phase and this third phase begins after the battery is charged to 95\% and the remaining 5\% is charged at a higher constant rate than the first phase.
\subsection{Summary}
Finally, other than CC-CV and DLC, we have identified two more variants of these two, one of them applies CV at the beginning and the other uses CC at the end of charging. We have also identified two kinds of Fast charging technique. Their charging rates vary within 0.7-1.1C. Nexus 6 takes 35 minutes, where as Galaxy 6 Edge takes 30 minutes to charge the first 50\% of the battery.
\section{SOC Estimation Techniques}
\label{sec:fuelgauge}
As mentioned in Section 2 that fuel gauge chips estimate SOC and smartphones basically employ either a voltage-based or Coulomb counter-based fuel gauge. The first kind depends on a number of voltage look up tables to estimate SOC and the latter one uses current sense resistors to measure the charging/discharging current. It is difficult to identify the presence of such mechanisms without any explicit knowledge about the chipset model or name. In this section, we attempt to distinguish these two from their SOC reporting behavior.
From the charging time calculated in Section~2, we compute the median charging time curves of different models. Figure~\ref{fig:onePercurve} illustrates such curves for five models. We notice that SOC update times of Nexus 6 are almost constant until the battery is charged to 50\%. Given a device has a Coulomb-counter based fuel gauge, it can accurately measure SOC of the battery while charging. If the device is not utilized, the battery would receive the maximum constant charging current from the charger during CC period. Therefore, the amount of time required to charge one percent should be equal for every SOC update within the CC-phase. The plot for Nexus 6 suggests that this model has Coulomb counter-based fuel gauge and we verified so.
On the other hand, the curves of \texttt{I337M, I9500, I9505}, and \texttt{ISGH337} are almost identical, however, they are different from Nexus 6. Their charging time vary at the same SOC. This hints that these devices use similar SOC estimation method. Later we find that these models are different variants of Samsung Galaxy S4 manufactured for different operators. We further verified that Galaxy S4 devices use voltage-based fuel gauges. Most of the devices in our dataset use voltage-based fuel gauges.
\section{Battery Properties}
\label{sec:battery}
We next investigate the battery properties, such as capacity, temperature behavior while charging, and health. Unlike the studies in other sections, we consider all the samples in this section.
\begin{marginfigure}[1pc]
\begin{minipage}{\marginparwidth}
\vspace{-30mm}
\subfigure[Final Voltage vs. Capacity]{\label{fig:battery_capacity_rel}\includegraphics[width=1.0\linewidth,height = 1.0\linewidth]{figures/battery_capacity_short.pdf}}
\vspace{-3mm}
\subfigure[Battery Capacity]{\label{fig:battery_capacity_red}\includegraphics[width=1.0\linewidth,height = 0.9\linewidth]{figures/batter_capacity.pdf}}
\caption{Battery final voltage and capacity.}
\label{fig:battery_capacity}
\end{minipage}
\includegraphics[width=1.0\textwidth,height = 0.7\linewidth]{figures/charge_varone.eps}
\caption{Final voltage over time.}
\label{fig:charge_varone}
\end{marginfigure}
\subsection{Battery Capacity Loss}
The effect of charging with over voltage, and higher charging rate is degraded battery life. Choi and Lim~\cite{Choi2002130} studied the performance of Lithium-Ion batteries by charging them with different charging rates and by limiting the final voltage. They showed that charging at higher rates (1.4C, 1.2C) and voltage (4.35 V) reduces battery capacity faster and significantly more than those with smaller magnitudes (1.0C, 4.2 V). However, it is not possible to estimate the effect of an individual variable without measurements. A recent study has shown that batteries with different capacity exhibit different voltage behavior while charging~\cite{Hoque:2015}. Although Android APIs do not expose the actual capacity in mAh, we can determine the relative remaining capacity or capacity loss from the final voltage. Figure~\ref{fig:battery_capacity_rel} shows that the relation between these two is linear~\cite{kestersection}. Every 10 mV reduction in the final voltage is equivalent to 1\% capacity loss. The capacity loss can be computed as
\begin{equation}
Capacity_{Loss} (\%) = (V_{f} - V_{rf})/10,
\end{equation}
\noindent where the value of $V_{f}$ is 4.2 or 4.35 V and $V_{rf}$ is the reported final voltage in the sample. Another observation is that the final voltage for the DLC or 4.35V models fluctuates frequently (see Figure~\ref{fig:charge_varone}). Therefore, we take the average of all final voltages of an individual device and estimate the capacity loss. Figure~\ref{fig:battery_capacity_red} shows that 85\% of the devices have lost their capacity by 1-10\%. A number of devices have significant capacity loss. This information can be used as an input to the self-constructive power modeling approaches, such as PowerBooter~\cite{Zhang:2010} depends on relative battery capacity.
\subsection{Battery Temperature}
In order to understand the pattern of battery temperature while charging, we first group the battery temperature for each reported battery level and plot them in Figure~\ref{fig:cccv_temperature}. We also plot the median temperature for CC-CV and DLC models. The battery temperature varies as the battery level increases. At the beginning, the temperature decreases till the battery is charged to 20\%. After that temperature slowly decreases or remains almost constant until the battery is charged to 70 or 80\%. However, at the end the temperature begins to decrease again. The obvious reason is very low current charging during the CV phase. Although both CC-CV and DLC exhibit similar temperature variation pattern, the average temperature of CC-CV models is higher than the DLC models. Figure~\ref{fig:fast_temperature} compares battery temperature between DLC and Fast charging techniques. We notice that Fast charging increases battery temperature by 8-10$^{\circ}$C than that of the DLC model devices.
\subsection{Battery Health}
Android battery APIs provide battery health information, such as good, over voltage, and overheat. Table~\ref{tab:battery_health_stat} shows the ranges of battery voltage and temperature for good, over voltage, and overheat samples. We notice that such ranges for over voltage and overheat samples also vary within the similar range as good samples. After examining these samples, we have found that over voltage samples do not give any hint about temperature and similarly overheat samples do not hint whether over voltage results higher temperature.
\begin{marginfigure}[1pc]
\begin{minipage}{\marginparwidth}
\vspace{-5mm}
\subfigure[CC-CV \& DLC]{\label{fig:cccv_temperature}\includegraphics[width=1.0\linewidth,height = 1.0\linewidth]{figures/charging_temp_newall.eps}}
\subfigure[DLC \& Fast Charging]{\label{fig:fast_temperature}\includegraphics[width=1.04\linewidth,height = 1.0\linewidth]{figures/fastcharge_temp.eps}}
\caption{{Battery temperature with different charging techniques via AC.}}
\label{fig:battery_temperature}
\end{minipage}
\end{marginfigure}
\begin{margintable}[1pc]
\begin{minipage}{\marginparwidth}
{\footnotesize
\begin{tabular}{lll}
\hline
\textbf{Heal.} & \textbf{Volt.}& \textbf{Temp.}\\\hline
Good & 3.2-4.4V&10-57$^\circ$C\\\hline
Overheat & 3.2-4.32V &24-72$^\circ$C\\\hline
Over volt. & 3.2-4.2V &16-44$^\circ$C\\\hline
\end{tabular}}
\caption{Battery health, voltage, and temperature information.}
\label{tab:battery_health_stat}
\end{minipage}
\end{margintable}
\vspace{-2mm}
\section{User Behavior in Charging}
\label{sec:user}
We consider all the charging events and the corresponding samples in this section as well. A charging event should contain one sample for each battery level or SOC update. Therefore, the number of samples for a specific battery level should be unique in an event. However, we have found more than one sample for a single battery level update in the form of SOC fluctuation (e.g. battery level = 5$\uparrow$6$\downarrow$5$\uparrow$6). Nevertheless, such SOC fluctuations are not uniformly distributed, rather left skewed with respect to the battery level.
2\% of the charging events reside at the tail of the distribution, which contain fluctuation between two consecutive levels. The screen status of the corresponding samples suggests that the devices were being actively used. Therefore, it took a longer time for an actual 1\% increment.
Other than charging and actively using their devices at the same time, users may keep their devices connected with the chargers even when the batteries are completely charged. From the dataset, we have identified 3\% of such charging events. The duration of such events can be a few to thousands of seconds. In this case, the phone stops charging the battery and begins recharging whenever 1-2\% has been discharged. However, we have measured that the extra energy spent during a over night charging for 10 hours can be used to charge a iPhone 6 to its full capacity (1810mAh).
\section{Conclusions and Future Work}
\label{sec:conclusion}
In this study based on data gathered from in-the-wild devices, we have shown that a few thousand devices use inefficient charging mechanisms that can significantly reduce battery life. We have found that 2\% of the devices charge their batteries well above the maximum battery voltage. This charging method deteriorates battery capacity faster than normal. There has been very active discussion in various online forums identifying battery SOC anomalies and such SOC error is due to the capacity loss~\cite{Hoque:2015}. A small number of devices had a charging current higher than 1.0C, which also degrades battery performance quickly. We have also observed that 85\% of the devices suffered from 1-10\% capacity loss. Moreover, user behavior and interaction with the device during charging also contribute to energy waste. Our future research includes investigating the performance of different charging algorithms with a larger dataset and developing a battery analytics API based on Spark so that users and vendors can investigate the performance of their batteries and power management techniques.
\section{Acknowledgements}
This work was funded by the Academy of Finland CUBIC project with grant number 277498.
\balance{}
|
\section{Introduction}
\IEEEPARstart{W}{ired} data transmission networks like telephone and digital subscriber line (DSL) networks are well established communication technologies that allow the exchange of enormous amount of data among all the users therein connected. In recent years also power lines have been extensively investigated for data transmission, enabling any user connected to a power grid to exchange information \cite{lampe2016power}. This new technology contributed in fostering the advent of smart grids \cite{gungor2011sgt}: not only mere infrastructures to distribute energy to users, but also intelligent networks that exchange information in order to efficiently satisfy the user demands and manage bidirectional power flows.
In this paper, we address the problem of identifying the topology of such wired data transmission networks, considering but not restricting to PLC in the context of distribution grids and SMG as a possible application. The term topology herein refers to the network graph that describes the nodes relative displacement, and the length of the wired connections.
The identification of the network topology is important in many respects. In the context of DSL networks, it is part of the line qualification procedures that are used to assess the ability of a specific wired network to support different DSL services before the actual deployment. A proper knowledge of the network topology allows to compute the channel transfer functions and can also be used for support engineering and maintenance operations \cite{1007375,5256192,5953512}.
In the context of SMG, not only communication is involved, but also power transmission. PLC are used herein as a mean of controlling and monitoring of the grid. The knowledge of the network topology is a fundamental requirement to develop routing strategies for both power and data information, as well as coordination algorithms for distributed computation \cite{1430477}.
Some recent proposals \cite{ahmed2012topology2,7431885,erseghe2013topology,lampe2013tomography} aim to estimate the PLC network topology as plug-and-play solutions (i.e., no historical data is considered), by using a two-step procedure. First, the channel response is sensed at different frequencies to estimate the distances between the nodes. Subsequently, different algorithms are applied to infer the network topology. This two-step procedure can be repeated over time, so that the topology is updated when the state of the network changes. In \cite{ahmed2012topology2} Frequency Domain Reflectometry (FDR) is used to perform a single-end distance measurement between one node and all the others. However, the FDR reliability is limited by the maximum observable distance and the number of branches. Some better performance is achieved using Time Frequency Domain Reflectometry \cite{7431885}. In \cite{erseghe2013topology}, it is assumed that all the nodes of the network are equipped with a PLC modem. The distance between nodes is estimated via Time of Arrival (ToA), i.e., from the propagation delay of the transmitted signal using an energy detector. ToA estimation deploying PLC modems is also used in \cite{lampe2013tomography}, where the energy detector is compared to a sub-space estimation method that is generally more efficient.
In this paper, we present a novel technique for wired network topology estimation that does not rely on historical data and uses admittance measurements operated at all the network nodes. This last requirement is envisioned for a future SMG scenario in which all the nodes will be equipped with PLC modems \cite{erseghe2013topology}. To our knowledge, admittance measurements have already been used in power networks in order to implement efficient fault detection strategies \cite{6338332},\cite{6954542}, but never for topology estimation. On the other hand, equivalent S-parameter and impedance measurement have been used in DSL to identify the DSL topology \cite{5256192} and the channel transfer function \cite{5953512} under ideal conditions. The purpose of this paper is to discuss the role of admittance measurements, the fundamental aspects, the conditions and the assumptions that have to be made to allow the derivation of a wireline network topology via admittance measurements. The focal points are the application of the Transmission Line (TL) theory \cite{Pozar} and the measurement of the network admittances at all nodes of the network. Based on this, we formulate an analytical result and a related algorithm which allow us to derive the topology exactly and independently from the size and complexity of the network, when no noise or measurement error is present. Hence, in this paper we use the terminology topology derivation instead of identification or estimation. An analytical formulation of the line background noise effect on the admittance measurement is then derived. Furthermore, an approach to derive the topology in the presence of noisy measurements is described, and its performance is assessed.
\subsection{Relation with existing solutions and contribution}
The main aim of this paper is to introduce and thoroughly explain the theoretical aspects of a novel technique for topology derivation in wired networks, as an alternative or complement to the present techniques. Moreover, a section is devoted to expose the open issues and the possible directions that can be pursued to improve the topology derivation method presented in this paper, and the topology inference methods in general.
The proposed technique can tackle some limitations of the existing plug-and-play topology estimation techniques for PLC networks, namely \cite{ahmed2012topology2,7431885,erseghe2013topology,lampe2013tomography}, but at the same time introduces some challenges, as discussed below.
\emph{Meter}: the existing techniques require PLC modems or reflectometers. Our approach relies on the use of just impedance or voltage meters, which can also be embedded in PLC modems.
\emph{Channel model}: the existing techniques use a phenomenological channel model that requires some assumptions about the channel (propagation velocity equal for each cable, constant reflection coefficients over frequency, propagation constant be linear function of frequency). If some of the assumptions do not hold true, this might deteriorate the accuracy of the topology estimation. Our technique uses a physical channel model based on TL theory \cite{tonello2011bottomup}, which is more strictly related to the physics of propagation. On the other hand it requires information about all the loads and cable parameters, which might not always be available.
\emph{Operating frequency}: High frequencies and large bandwidths are needed to obtain reasonable performance of ToA techniques and good resolution in FDR or TFDR. Our approach operates at a single frequency that can also be in the range of the narrow band PLC spectrum, more commonly used in SMG.
\emph{Dimension of the network}: a problem of the existing two step techniques \cite{ahmed2012topology2,7431885,erseghe2013topology,lampe2013tomography}, is that the topology is inferred only if each node (or the main node in the case of FDR and TFDR) knows the distance between itself and any other node of the network. Moreover, the multipath propagation and the strong attenuation of the high frequency signals limit the maximum distance that can be sensed. This problem can be solved by splitting the network in many overlapped subsections (see \cite{lampe2013tomography2}). In our approach each node finds the distance only to its neighbor and can contextually infer its topological position. Hence, our approach is not limited by the dimension of the network, but by the maximum distance between two neighbor nodes.
At the same time our work shares some similarities with \cite{5256192}. Both of them use admittance or equivalent scattering parameter measurements and assume all the line and load parameters to be ideal. They however differ on the requirements, the algorithms and the final results. The work in \cite{5256192} applies a genetic algorithm to derive the network topology and the number of nodes starting from single or double end measurements. However the algorithm is tested on simple networks, and no parameter error or noise is taken into account. The work presented in this paper relies on measurements performed at every node of the network, but it can derive the topology for any network, when no noise is considered. Furthermore, an analysis of the impact of network noise on the derivation algorithm is performed.
The reminder of this paper is organized as follows. In Section~\ref{sec:ana}, a brief review of the basic equations of the TL theory used for distance computation is given. In Section~\ref{sec:sc2}, the main system of equations to solve the topology derivation problem is derived. In Section \ref{sec:noise}, the influence of the line network noise on admittance measurements is discussed. The topology derivation algorithm is then presented in Section~\ref{sec:topdev}. Numerical results are also reported in Section~\ref{sec:results} to study the robustness of the proposed algorithm to noise. Further remarks and open problems are also discussed in Section~\ref{sec:remarks}. Finally, the conclusion follows.
\section{Analytical evidence from TL theory}
\label{sec:ana}
In this section, we examine the TL theory in order to derive an equation that relates the length of the line that connects two nodes with the admittance at one end. We start by considering the simplest case of an unbranched transmission line of length $d$ that connects a Thevenin generator to a load $L$. The network admittance $Y(d)$ seen by the generator can be written as \cite{Pozar}
\begin{equation}
Y(d) = Y_C \frac{1-\rho_L e^{-2\Gamma d}}{1+\rho_L e^{-2\Gamma d}},
\label{eq:adm_TL}
\end{equation}
We refer to this relation as \emph{carry-back} equation since the load admittance is carried back to the input of the line to obtain $Y(d)$.
$Y_C$ is the characteristic admittance of the line and $\Gamma = \alpha + j\beta $ is the propagation constant of the line, where $\beta = 2\pi/\lambda$ and $\lambda$ is the wavelength used to perform the measurement. $\rho_L$ is the load reflection coefficient written as
\begin{equation}
\rho_L = \frac{Y_C - Y_L}{Y_C + Y_L},
\label{eq:rho_TL}
\end{equation}
where $Y_L$ is the load admittance. All the aforementioned quantities (except $d$) depend on the frequency. Herein and in the following, this dependency is implicit to ease the notation. However, as it will be discussed, the choice of the frequency influences the algorithm.
From \eqref{eq:adm_TL}, under the assumption that we know the load reflection coefficient \eqref{eq:rho_TL}, an equation that relate $d$ to the measured network admittance can be found\footnote{Actually the equations are two. The second one is
$
d = \frac{1}{2\beta}\left(\phase{\rho(d)}-\phase{\rho_L}\right)
$,
but it limits the estimation of $d$ due to the intrinsic periodicity of the phase.}:
\begin{equation}
d = \frac{1}{2\alpha}\left(\log{\left|\rho(d)\right|}-\log{\left|\rho_L\right|}\right)
\label{eq:l_TL}
\end{equation}
where
\begin{equation}
\rho(d) = (Y_C - Y(d))/(Y_C + Y(d))
\label{eq:rho_simple}
\end{equation}
and $k$ is an integer number. \eqref{eq:l_TL} is bijective since in real lines $\alpha \ne 0$, so that $\left|\rho(d)\right|$ is a monotonically decreasing function of $d$. We also remark that in two cases it is impossible to find the length of the transmission line using \eqref{eq:l_TL}:
\begin{enumerate}
\item if the transmission line is ideal, i.e. $\alpha = 0$. In fact in this case $Y(d)$ would be a periodic function of $d$ with period $\lambda/2$.
\item if $\rho_L=0$, i.e. when the load is matched to the transmission line impedance. In fact in this case $Y(d)$ would simply be a constant.
\end{enumerate}
When branches attached to the main transmission line are considered, the problem of finding the length of each line becomes more difficult. In the following sections, we consider a complex network made of $N$ nodes. We refer to the nodes as branch nodes (identified by line intersections) and termination nodes (leaves). The known parameters in such a network are the cable parameters $\Gamma_i$, $Y_{C_i}$ and the loads $Y_{L_i}$ or equivalently the reflection coefficients $\rho_{L_i}$ $\forall i \in [1,\dots N]$, and network admittances $Y_{i}$ at every node of the network.
\section{Derivation of lines' length with unknown topology graph}
\label{sec:sc2}
In this section, we derive the analytical formulas and the theorem that allow us to derive the network topology.
To understand how to proceed, we initially consider the simple network depicted in Fig. \ref{fig:two_loads}.
\begin{figure}
\centering
\includegraphics[width=0.30\textwidth]{two_loads.eps}
\caption{Sketch of a simple 2-loads network}
\label{fig:two_loads}
\end{figure}
Herein, two loads are connected by a line with length $d$, and the network admittances $Y_1$ and $Y_2$ are measured at the line ends. Using \eqref{eq:adm_TL} and \eqref{eq:rho_TL} it is possible to relate the network admittance $Y_1$ as a function of $d$ as follows:
\begin{equation}
\begin{cases}
\mathlarger{Y_1 = Y_C \frac{1-\rho_{L_2} e^{-2\Gamma d}}{1+\rho_{L_2} e^{-2\Gamma d}} + Y_{L_1}}\\[10pt]
\mathlarger{\rho_{L_2} = \frac{Y_C - Y_{L_2}}{Y_C + Y_{L_2}} = \frac{Y_C-Y_2+Y_C \frac{1-\rho_{L_1} e^{-2\Gamma d}}{1+\rho_{L_1} e^{-2 \Gamma d}}}{Y_C+Y_2-Y_C \frac{1-\rho_{L_1} e^{-2\Gamma d}}{1+\rho_{L_1} e^{-2\Gamma d}}}}
\end{cases}
\label{eq:two_loads}
\end{equation}
A similar system of equations can be written for the network admittance $Y_2$ measured at node $2$. In this system, exploiting the second relation, we can write that
\begin{equation}
Y_1 = f(Y_2,Y_{L_1},Y_C,\Gamma,d),
\label{eq:fund}
\end{equation}
so that the admittance measured at node $1$ is a function, in particular, of the network admittance $Y_2$ while the knowledge of the load admittance $Y_{L_2}$ is not explicitly required. It is important to point out that in general there can be two values of $d$ that are admissible, i.e., two possible solutions. This is because the term (load reflection coefficient)
\begin{equation}
\rho_2 = \rho_{L_2}e^{-2\Gamma d}
\label{eq:rhomale}
\end{equation}
that appears in the first equation of \eqref{eq:two_loads} may be such that
\begin{equation*}
\rho_2(d_\alpha) = \rho_2(d_\beta)
\end{equation*}
for $d_\alpha < \lambda/4$ and $\lambda/4 < d_\beta < \lambda/2$, depending on the parameters;
in fact, $\rho_2$ is the product of an exponential function with $\rho_{L_2}$, which also has an exponential trend. Using the Smith chart, an example of $\rho_2$ as a function of $d$ is plotted in Fig. \ref{fig:rho_example}. It can therein be clearly seen that there are two distances in correspondence of which the reflection coefficient takes the same value. Given this fact, $\rho_2$ is not a bijective function of $d$, conversely from \eqref{eq:rho_simple} that exploits the load admittance $Y_{L_2}$ instead of the network admittance $Y_2$. It follows that to grant a unique solution for the variable $d$, we must assume that measurements have to be taken at a wavelength $\lambda \geq 4d$.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{rho_example.eps}
\caption{Smith chart with a possible realization of $\rho_2$ according to equation \eqref{eq:rhomale}.}
\label{fig:rho_example}
\end{figure}
When the results shown above are extended to a more complex network, as for example the one depicted in Fig. \ref{fig:three_loads},
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{three_loads.eps}
\caption{Sketch of a simple 3-loads network}
\label{fig:three_loads}
\end{figure}
it is possible to write a system of equations similar to \eqref{eq:two_loads}. Now, the second equation of the system changes into
\begin{equation}
\rho_{L_2} = \frac{Y_C - Y_{L2eq}}{Y_C + Y_{L2eq}} = \frac{Y_C - (Y_{L_2}+Y_{cb_{3}})}{Y_C + (Y_{L_2}+Y_{cb_{3}})}
\label{eq:rho_three}
\end{equation}
where $Y_{cb_{3}}$ represents the load admittance $Y_{L_3}$ carried back to node 2. One can also write
\begin{equation}
Y_{L_2}= Y_2 - Y_C \left(\frac{1-\rho_{L_1} e^{-2\Gamma_1 d_1}}{1+\rho_{L_1} e^{-2\Gamma_1 d_1}} +
\frac{1-\rho_{L_3} e^{-2\Gamma_2 d_2}}{1+\rho_{L_3} e^{-2\Gamma_2 d_2}}\right)
\label{eq:Y2}
\end{equation}
so that with some simple algebraic manipulation, \eqref{eq:rho_three} finally becomes
\begin{equation}
\rho_{L_2} = \frac{Y_C-Y_2+Y_C \frac{1-\rho_{L_1} e^{-2\Gamma_1 d_1}}{1+\rho_{L_1} e^{-2\Gamma_1 d_1}}}
{Y_C+Y_2-Y_C \frac{1-\rho_{L_1} e^{-2\Gamma_1 d_1}}{1+\rho_{L_1} e^{-2\Gamma_1 d_1}}}
\label{eq:rho2_long}
\end{equation}
that is equal to the second equation in the system \eqref{eq:two_loads}. Hence, this proves that equation \eqref{eq:fund} is still valid when another part of the network is branched to node 2 so that to find the distance $d_1$ we need a network admittance measurement at nodes $1$ and $2$ and apply the first equation in \eqref{eq:two_loads}. Then, to obtain $d_2$ we can proceed with a similar reasoning so that we relate according to a similar set of equations $Y_2$ with $Y_3$.
Furthermore, the system \eqref{eq:two_loads} of equations can be written as a single complex equation in the unknown $d$. The resulting equation has the following quadratic form
\begin{equation*}
\gamma e^{-2\Gamma d} + \delta e^{-4\Gamma d} + \epsilon = 0,
\end{equation*}
where $\gamma$, $\delta$ and $\epsilon$ are polynomial functions of $Y_1$, $Y_2$, $Y_C$ and $Y_{L_1}$. The solutions can be written in closed form as
\begin{subequations}
\begin{align}
d_1 = \frac{1}{2\Gamma}\log{\left[-\frac{a\left(\sqrt{-b(cY^2_2+lY_2+e)}+fY_2+g\right)}{h}\right]}\\[20pt]
d_2 = \frac{1}{2\Gamma}\log{\left[-\frac{a\left(-\sqrt{-b(cY^2_2+lY_2+e)}+fY_2+g\right)}{h}\right]}
\end{align}
\label{eq:dist}
\end{subequations}
where $a$, $b$, $c$, $e$, $f$, $g$, $h$, $k$ and $l$ are polynomial functions of $Y_1$, $Y_C$, $Y_{L_1}$ and $\Gamma$. We wrote \eqref{eq:dist} as an explicit function of the sole $Y_2$, which is the network admittance of node 2 that is adjacent to node 1 with measured network admittance $Y_1$.
It should be observed that the admittance measured at one termination node (leaf) of the PL network depends only on the physical parameters of the cable to whom it is branched, on the length of the cable connecting it to the nearest node, and on the network admittance measured at this second node. However, although we have measured all the network admittances, we still do not know the topology graph and therefore we do not know the association between the admittances, i.e., we do not know what nodes/admittances are adjacent and what nodes are not directly connected by a line. It may be believed, at a first glance, that the system of equations \eqref{eq:two_loads}, and therefore \eqref{eq:dist}, applies to any pair of admittances measured in the network so that we always get a physically meaningful (although wrong) distance. In reality, since all the terms inside the logarithm in \eqref{eq:dist} are complex, $d_1$ and $d_2$ can be complex depending on $Y_2$. Of course the distance we are looking for must be a real number and this is the case when $Y_2$ is the true admittance of a node that is adjacent to the node 1 with admittance $Y_1$. A fundamental result is then given by the following theorem which turns out to be instrumental to obtain a topology derivation algorithm. \newline
\begin{theorem}
\label{th:1}
Considering a wired network made by $N$ nodes, the distance between any leaf $i$ and another node $j$ can be found by applying \eqref{eq:dist}. The result is the correct value with probability 1 either for $d_1$ or $d_2$, if and only if:
\begin{enumerate}
\item the admittance $Y_2$ used in \eqref{eq:dist} is the one measured at node $j$, to which node $i$ is directly connected;
\item the actual length of the line connecting $i$ and $j$ is less than $\lambda/4$, where $\lambda$ is the wavelength used to perform the admittance measurements.
\end{enumerate}
\end{theorem}
\vspace{5pt}
\begin{corollary}
\label{cor:1}
When no parameter or measurement error exists, the topology (graph and branch lengths) of any line network in which
\begin{equation*}
\max_{i,j \in N} d_{i,j} \le \frac{\lambda}{4},
\end{equation*}
can be exactly derived by means of a recursive use of \eqref{eq:dist}, with the exploitation of the measured network admittances and the available cable parameters and loads.
\end{corollary}
\vspace{17pt}
\textit{Proof:}
A sufficient condition to obtain a unique and real solution for the distance is when the pair of nodes are a leaf and the adjacent node. An this is obvious from the physical construction of the problem and associated TL equations.
To prove the necessity, i.e., that there exists a unique $Y_2$ for which the solution $d$ to \eqref{eq:dist} is real, we follow a probabilistic reasoning. Firstly, let's consider the plane where the impedance $Y_2$ can possibly lay and let's define an arbitrary value of it with $y_2$. Then, we note that the locus of points for which the imaginary part of $d$ is zero is a line, because $d$ is the logarithm of a polynomial function. Since any complex polynomial is holomorphic \cite{9781139171915}, it cannot be locally constant, so its imaginary part can assume one value only along a line, or a sequence of lines (and the logarithm does not influence the function in this sense).
As an example, $\Im(d_1)$ is plotted as a function of $y_2$ in Fig. \ref{fig:d1imag}. The bold-dashed line in this figure highlights the locus of points where the imaginary part of the distances is zero.
Secondly, let's assume to randomly pick $y_2$, i.e., the real and imaginary parts of it are independent, continuous random variables. Then, the probability that $y_2$ lays on a line is zero and consequently the probability that $\Im(d_2) = 0$ or $\Im(d_1) = 0$ is also zero.
We can therefore state that with probability 1, the only case for which $d_1$ or $d_2$ are real valued is the case corresponding to the sufficient condition, i.e., when $Y_1$ and $Y_2$ are the network admittances of the leaf and the adjacent node. Furthermore, only one among $d_1$ and $d_2$ will be real valued. This is because the problem of finding where $\Im(d_1(y_2)) = \Im(d_2(y_2)) = 0$ implies to find the intersection between two lines. Observing the form of \eqref{eq:dist}, the number of intersection points is finite, so that $P[\Im(d_2(y_2))=\Im(d_1(y_2))= 0] = 0$.
Finally, the assumption that the distance among the two nodes is less or equal to $\lambda/4$ is a prerequisite to invert equation \eqref{eq:two_loads} since, as we have already explained, that equation is bijective only when $d < \lambda/4$.
The corollary is an immediate consequence of the theorem. It will be constructively used in an algorithm to derive the topology graph and branch lengths in the next section.
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{d1imag_new.eps}
\caption{Example of one possible $\Im(d_1)$ as a function of the sole $y_2$. }
\label{fig:d1imag}
\end{figure}
\section{Admittance noise}
\label{sec:noise}
In real scenarios admittance measurements are perturbed by noise. In this section, a mathematical derivation of the admittance noise is performed. This noise will be related to the signal-to-noise ratio (SNR) normally defined in communications.
An admittance meter can be represented as its Thevenin equivalent, being $V_S$ the equivalent generated voltage phasor and $Y_S$ the equivalent output admittance. When an admittance measurement is performed, then a voltage divider is created between $Y_S$ and $Y_m$, the unknown network admittance at node $m$ (see Fig.~\ref{fig:partitore}). Using the voltage divider equation, $Y_m$ can be derived as $Y_{m} = Y_SV_S/V_m-Y_S$, where $V_m$ is the phasor of the voltage drop across $Y_m$. $V_m$ is affected by the background noise present in the network, so that it can be written as $V_m = V_{m_0} + V_{m_N}$, where $V_{m_N} \sim \mathcal{CN} (0,\sigma^2_N)$ and $V_{m_0} = \mathbb{E}\left[V_m\right]$. $\mathcal{CN} (0,\sigma^2_N)$ denotes a complex Gaussian variable with zero mean and variance $\sigma^2$; $\mathbb{E}[\cdot{}]$ denotes the expectation operator. The real and imaginary noise are assumed to be independent and with the same variance $\sigma^2_N/2$.
If we assume $V_S$ and $Y_S$ to be ideal, the noisy load admittance $Y_m$ can be written as
\begin{align}
Y_m &= \frac{Y_S\left(V_S - V_{m_0} - V_{m_N}\right)}{V_{m_0} + V_{m_N}} \\
&= \frac{Y_S\left(V_S - V_{m_0}\right)}{V_{m_0} + V_{m_N}} - \frac{Y_S V_{m_N}}{V_{m_0} + V_{m_N}}.
\label{eq:ylnoise}
\end{align}
\begin{figure}
\centering
\includegraphics[width=0.25\textwidth]{partitore.eps}
\caption{Sketch of a voltage measurement device. $V_S$ and $Y_S$ are the equivalent voltage generator and admittance, $Y_m$ is the network admittance at node $m$ to be measured.}
\label{fig:partitore}
\end{figure}
In particular, if the SNR at the node $m$, i.e. $\mathbb{E}\left[\left|V_{m_0}\right|^2\right]/\mathbb{E}\left[\left|V_{m_N}\right|^2\right]$, is sufficiently high, then \eqref{eq:ylnoise} can be simplified as
\begin{equation}
Y_m = \frac{Y_S\left(V_S - V_{m_0}\right)}{V_{m_0}} - \frac{Y_S V_{m_N}}{V_{m_0}} = Y_{m_0} + Y_{m_N},
\label{eq:ylnoise_sim}
\end{equation}
thus $Y_m$ can also be considered as a perfect measurement $Y_{m_0}$ corrupted by the Gaussian noise $Y_{m_N}$. To experimentally prove it, we generated thousands of realizations of $Y_m$ with different parameters and SNRs. By applying the standard Kolmogorow-Smirnow Test \cite{daniel1990applied}, we discovered that the simplification introduced in \eqref{eq:ylnoise_sim} is valid when SNR $>$ 35~dB. Such a value of SNR is easily exceeded for example when the measure is done according to the PLC standards, where in the worst condition at few kHz the background noise can reach -70~dBm, while the corresponding transmit power is around -15~dBm \cite{7037264}.
When \eqref{eq:ylnoise_sim} holds, we can define the Admittance to Noise Ratio (ANR) as
\begin{equation}
ANR = \frac{\mathbb{E}\left[\left|Y_{m_0}\right|^2\right]}{\mathbb{E}\left[\left|Y_{m_N}\right|^2\right]}
\simeq \frac{\left|Y_{m_0}\right|^2}{\sigma^2_{N}} = \frac{\left|V_S-V_{m_0}\right|^2}{\sigma^2_{N}},
\label{eq:ANR}
\end{equation}
where $\sigma^2_N$ is the variance of $V_{m_N}$ and the second equivalence holds when $Y_{m_0}$ can be considered static over time, i.e. within the coherence time of the channel.
\begin{figure}
\centering
\includegraphics[width=0.43\textwidth]{anr_vs_vl.eps}
\caption{Normalized plot of \eqref{eq:ANR} for different values of $V_{m_0}$}
\label{fig:anr_vs_vl}
\end{figure}
As we can see in Fig.~\ref{fig:anr_vs_vl}, when $V_{m_0}$ is close to $V_S$, the admittance noise is amplified, leading to an ANR lower than the SNR. Vice versa, when a high value of ANR is wanted, then $V_{m_0}$ has to be as little as possible. A little value of $V_{m_0}$ can be easily achieved by using $|Y_S| \ll |Y_{m_0}|$.
\section{Topology Derivation}
\label{sec:topdev}
In this section, we present an algorithm that relies on Theorem \ref{th:1} and that allows to derive the topology of a general tree-structured wired network, together with the length of all the lines connecting the nodes.
Let $\mathcal{T}=(\mathcal{N},\mathcal{L})$ denote the topology of a network, where $\mathcal{N}$ is the set of all the $N$ nodes of the network and $\mathcal{L}$ is the set made of all the physical connections between two network elements. The terminal nodes, i.e. those nodes that are connected to the rest of the network with a single branch, are referred to as leafs.
To each node of the network $i \in \mathcal{N}$ we associate a load admittance $Y_{L_i}$, that is characteristic of the device plugged to the network, and a network admittance $Y_{N_i}$ that is the admittance measured at node $i$ comprising the load at that node. The cable parameters for each line $l \in \mathcal{L}$ departing from node $i$ are assumed to be known. Such cable parameters are the propagation constant $\Gamma_l$ and the characteristic admittance of the line $Y_{C_l}$. Clearly, in a uniform network all cables have identical parameters.
Algorithm \ref{alg:mine} offers a method to derive $\mathcal{T}$ taking as inputs the parameters $Y_{L}$, $\Gamma$, $Y_{C}$, $Y_{N}$, that are known for each node and branch of the network. The last parameter needed is the ANR that is assumed to be the same for every node and can be sensed by the modem during the calibration period of each data transmission. The core idea of the algorithm is that if the imaginary part of one of the two computed cable lengths is small enough, then also the error on the real part is small, and the two nodes considered are with high probability connected.
The algorithm firstly considers the full set of nodes $\mathcal{N}$, and it assumes them to be, potentially, leafs. With this assumption, \eqref{eq:dist} is applied to every pair of nodes $i$ and $k$. If the result provides a solution whose imaginary part is greater than a certain threshold, then node $i$ cannot be a leaf and $k$ cannot be a directly connected node. If instead the imaginary part of one solution is lower than the threshold, the algorithm states that $i$ is a leaf and the two nodes $i$ and $k$ are directly connected with a branch having real length $\Re[d]$.
After having found all actual leafs, their load admittances are carried back to the connected internal nodes, to form a reduced network. The information about the connections and line lengths is contextually stored. The algorithm can also detect false positives and false negatives; in case of detection, the algorithm is interrupted.
The procedure described in the previous paragraph iterates until the whole network is reduced to a single node. The final outputs are the complete topology $\mathcal{T}=(\mathcal{N},\mathcal{L})$ and the complete set of lengths for the $\mathcal{L}$ links.
\begin{algorithm}
\caption{Topology derivation}
\label{alg:mine}
\begin{algorithmic}[1]
\Require $Y_m$ and $Y_L$ for each node, $\Gamma$ and $Y_C$ for each cable.
\Statex
\Procedure {$[\textbf{d},\mathcal{T}]$ = Derivation(${\bf y}_L$,${\bf y}_m$,${\bf y}_C$,$\boldsymbol\gamma$,ANR)}{}
\State $\textbf{gl} \gets \text{1: length of $\textbf{y}_L$}$
\State $\text{thr} \gets \text{f(ANR)}$
\While{$\text{length of ($\textbf{gl}$)} > 1$}
\For{$i \in \textbf{gl}$}
\State $j = 0$
\For{$k \in (\textbf{gl} \setminus i)$}
\State $\textbf{x}_{(\textbf{gl}(i),\textbf{gl}(k))} \gets \text{equations \ref{eq:dist}}$
\If{$(\Im\text[d] \subset \textbf{x}_{(\textbf{gl}(i),\textbf{gl}(k))}) < \text{thr}$}
\State $j \gets k$
\EndIf
\EndFor
\State $\textbf{d}_{(\textbf{gl}(i),\textbf{gl}(j))} \gets \textit{d}$
\State $\mathcal{N} \gets \mathcal{N} \cup \textbf{gl}(i)$
\State $\mathcal{L} \gets \mathcal{L} \cup (\textbf{gl}(i),\textbf{gl}(j))$
\State ${\bf y}_L (j) \gets {\bf y}_L (j) + \textit{carryback (${\bf y}_L (i), d$)}$
\EndFor
\State $\textbf{gl} \gets \textbf{gl}\setminus\textbf{gl}(i)$
\EndWhile
\EndProcedure
\Statex
\Ensure Network topology and length of all the branches.
\end{algorithmic}
\end{algorithm}
\section{Results}
\label{sec:results}
In this section we firstly introduce the physical layer simulator that we developed in order to test Algorithm \ref{alg:mine}; then we present and comment the results, and finally we discuss the open issues.
To test Algorithm \ref{alg:mine}, we developed a network simulator that creates a random tree network (see Fig. \ref{fig:example_topology}) and computes the network admittance at each node. The network simulator is based on the TL theory and exploits the concentrated parameter model used in \cite[Sec. III.C]{tonello2011bottomup} to describe the cable parameters.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{example_topology.eps}
\caption{Sketch of a network realized by the simulator.}
\label{fig:example_topology}
\end{figure}
Then, complex Gaussian noise is added to each network admittance according to the ANR specified by the user. Algorithm \ref{alg:mine} is finally applied to derive the topology. The simulator outputs whether a topology has been found or not, and in the positive case it outputs also $[\textbf{d},\mathcal{T}]$.
As an empirical proof of Theorem \ref{th:1} and Corollary \ref{cor:1}, we found that when no noise is added to the measurements, $[\textbf{d},\mathcal{T}]$ is correctly derived in 100$\%$ of the cases, independently from the size of the network and the number of nodes.
\subsection{Results with background noise}
As explained in Section \ref{sec:noise}, the network background noise causes the network admittance measurements to be affected by error. This noise deteriorates the performance of Algorithm \ref{alg:mine}, as shown in Fig.~\ref{fig:thresholding_effect_all_fit}. For the depicted test, an adaptive threshold has been used in order to get the best performance for each ANR. The results confirm that the performance of the algorithm decreases with the increasing number of nodes. Moreover lower measurement frequencies give better results: in fact, using a noisy $Y_2$ in \eqref{eq:two_loads}, we see that the error in $Y_1$ grows as a function of the measurement frequency and the length of the cable. We remark that if a topology is correctly identified, it means that also all the branch lengths are identified with a negligible error. In fact, a consistent error in the computation of a single branch length would deeply affect the subsequent iterations of Algorithm \ref{alg:mine}, thus leading to a topology identification error.
Fig.~\ref{fig:correctly_detected_when_bad} shows the percentage of correctly detected topology elements, i.e. branches, when the topology is not completely derived. Since the full topology is finally not derived, we infer that correctly detected topology elements of Fig.~\ref{fig:correctly_detected_when_bad} are affected by an error in the branch length derivation that is not negligible. However two aspects arise: the percentage does not change with the ANR, and it increases with the number of nodes. Both these aspects are a consequence of the fact that when the ANR decreases, also the number of correctly derived topologies decreases, so more topologies are taken into account for the computation here considered. The steep increments at low ANR for 20 and 30 nodes are due to the fact that for low ANRs almost no topology is completely derived. Nevertheless, Fig.~\ref{fig:correctly_detected_when_bad} shows that a consistent part of the topological information of the network can be retrieved even when the topology is not completely derived.
\begin{figure}
\centering
\subfloat[][$f_0$ = 10~kHz]
{\includegraphics[width=0.24\textwidth]{thresholding_adapt_10khz.eps}}
\subfloat[][$f_0$ = 30~kHz]
{\includegraphics[width=0.24\textwidth]{thresholding_adapt_30khz.eps}}
\caption{Percentage of correctly derived topologies as function of the ANR. Test run with a maximum cable length of 1.4~km. }
\label{fig:thresholding_effect_all_fit}
\end{figure}
\begin{figure}
\centering
\subfloat[][$f_0$ = 10~kHz]
{\includegraphics[width=0.24\textwidth]{thresholding_adapt_10khz_bad.eps}}
\subfloat[][$f_0$ = 30~kHz]
{\includegraphics[width=0.24\textwidth]{thresholding_adapt_30khz_bad.eps}}
\caption{Percentage of correctly detected junctions when the topology has not been completely derived. The same test conditions of Fig.\ref{fig:thresholding_effect_all_fit} are used.}
\label{fig:correctly_detected_when_bad}
\end{figure}
Table~\ref{tab:BondsOfTheValidity} provides some information about typical values of ANR that we could encounter by operating in the PLC band and fulfilling PLC norms. According to Table~\ref{tab:BondsOfTheValidity}, if the measurement is performed at $f_0 =$ 10~kHz, the ANR is approximately 100~dB, so that from Fig.~\ref{fig:thresholding_effect_all_fit} we see that for 10 nodes more than 90\% of the topologies is correctly detected. Moreover, about 60\% of the connections within the remaining topologies is correctly identified (see Fig.~\ref{fig:correctly_detected_when_bad}).
\begin{table*}[t]
\centering
\caption{Typical parameter values used to evaluate the proposed algorithm.}
\begin{tabular}{l|l|l|l}
\toprule
& Cenelec (3--150~kHz) & FCC (150--500~kHz) & Broad-Band (2--30~MHz) \\
\midrule
Maximum cable length & 16.6~km & 330~m & 25~m \\
Average Ground noise \cite{7037264} & -70 -- -90~dBm/Hz & -90 -- -110~dBm/Hz & $<$ -110~dBm/Hz \\
Transmitted power & $\sim$-15~dBm/Hz & -22~dBm/Hz & -55~dBm/Hz \\
ANR & 99 -- 135~dB & 122 -- 158~dB & $>$99~dB \\
\bottomrule
\end{tabular}
\vspace{0.2cm}
\caption*{Remarks: the maximum cable length depends on its propagation constant (herein light velocity $v_c$ = 2e8~m/s). The data about the transmitted power has been retrieved from \cite{lampe2016power} and from the standard IEEE 1901.2a-2015. As for the ANR, here we consider the condition in which the measured network voltage has a small absolute value compared to the voltage provided by the voltmeter, so that ANR = 1.8 SNR (see Fig.~\ref{fig:anr_vs_vl}).}
\label{tab:BondsOfTheValidity}
\vspace{-0.5cm}
\end{table*}
\section{Remarks and Open Issues}
\label{sec:remarks}
The analysis reported so far has shown that admittance measurements can be exploited to gain knowledge about the topology of a wired network. In particular, admittance measurements can be used to provide a real-time solution that can partially or completely derive the network topology, as well as the length of the branches connecting the nodes, all in a unified algorithm. The algorithm in presence of noise has been tested using typical values from PLC in the context of SMG.
The approach that we have discussed opens further questions on some open issues as we discuss in the following.
\begin{enumerate}
\item \label{item1}The frequency at which measurement have to be made, depends on the node distances and cannot be freely chosen. However, this is not very restrictive
\item \label{item2}The speed of the measurement is also important since the status of the network may change over time. In fact, although the length and the node connections can be considered time-invariant, the loads may change. Therefore, the coherence time of the topology must be larger than the measurement time.
\item \label{item3}Synchronization and coordination of the admittance measurements can also be required in a network with a time variant status.
This could be done for example with a GPS coordination system or with communication modules
\item \label{item4}An important aspect is the presence of measurement errors and uncertainty in the required parameters, i.e., loads and cable parameters. Proper estimation techniques can be used to tackle such a problem and also to enhance the performance of Algorithm \ref{alg:mine} in the presence of noise. Interestingly, the proposed topology derivation technique may be used inversely to track the cables deterioration, by sensing the increment of the error in the derivation of the branch lengths over time.
\end{enumerate}
We point out that the open issues \ref{item2} and \ref{item3} mentioned in this section have not yet been fully considered in the literature of wired network topology estimation, while open issue \ref{item4} is shared with \cite{5256192}. All these open issues provide stimulus for further research endeavors.
\section{Conclusion}
In this paper, we have addressed the question: ``Can we exploit admittance measurements to derive the topology of a wired network ?''
The approach differs from others presented in the literature, which use reflectometry or ToA estimation followed by topology inference algorithms. We have shown that the admittance measurement based approach can allow the derivation of the topology and the length of all branches by performing admittance measurements at all nodes. It is the direct application of the derived Theorem 1, which states that it is possible to identify whether a pair of nodes in a network are respectively a leaf and a directly connected node, and if so, a solution to the derivation of the length of the line connecting them can be found. We have further shown that the admittance measurements are perturbed by complex Gaussian noise in the presence of line background noise when the SNR at the receiver is greater than 35~dB. Moreover, the ANR is greater than the SNR when the internal impedance of the measurement device is greater than the measured impedance.
These findings have been used to develop an algorithm that derives the topology of a wired network provided that admittance measurements are done at all the network nodes and that the cable parameters and the loads are known. Future research directions in this topic have also been discussed and include: robust topology derivation in the presence of parameter errors; application of the admittance based topology derivation algorithm to track the cables deterioration.
\bibliographystyle{IEEEtran}
|
\section{\label{sec:motivation}Introduction}
The coupling of spin and orbital degrees of freedom lies at the heart of modern spintronics device concepts,
which aim at an ultrafast and low-power-consumption information processing beyond the reach of present-day electronics.\cite{Sinova, Barnes}
Such a coupling is realized by the Rashba effect,\cite{Rashba60} which generally results from a large atomic spin-orbit interaction (SOI) and the lack of inversion symmetry. Traditionally, the Rashba spin splitting (RSS) has mainly been
observed in two-dimensional systems like surfaces or interfaces between different materials.\cite{LaShell96, Nitta97, Ast07, 2012giant} By contrast, the recently observed {\it giant} RSS in the polar semiconductor BiTeI even turned out to be a {\itshape bulk} material property.\cite{2011giant, Crepaldi}
In particular, this implies the possibility of observing optical transitions between the spin-split energy bands, \cite{Lee} an unconventional orbital paramagnetism \cite{Schober} as well as an enhanced magneto-optical response in the infrared regime.\cite{Demko} Furthermore, the theoretical prediction of a pressure-induced topological phase transition towards a noncentrosymmetric topological insulating phase of BiTeI \cite{BiTeITop} has led to several (as of yet still controversial) experimental investigations.\cite{Xi13, Tran14, Park15}
Shortly after the discovery of giant bulk RSS in the semicondutor BiTeI, general conditions for its appearance have been formulated,\cite{BiTeI_WIEN2k} and a number of related compounds have been investigated such as BiTeCl and BiTeBr. \cite{Eremeev, Sakano, Akrap} These materials have a band structure different from BiTeI, and in particular display a smaller RSS. Nevertheless, they have attracted much interest in materials science due to their unique electronic structures and properties. For example, topological surface states were predicted to appear in BiTeCl at ambient pressure.\cite{Chen} Furthermore, the optical properties and Raman spectra of BiTeBr and BiTeCl have been investigated experimentally by \citet{Akrap}. There, it turned out that the optical properties of these two compounds are very similar despite their different space groups ($P3m1$ for BiTeBr, $P6_3mc$ for BiTeCl).\cite{Akrap} Thus, the bulk Rashba materials BiTeX (X = I, Cl, Br) do not only realize a tabletop laboratory for investigating relativistic electron dynamics,\cite{Lee} but are also regarded as promising candidates \cite{Rashba12} for the future application in spintronics devices such as the Datta-Das spin transistor.\cite{Datta89, Koo09}
Despite this tremendous theoretical and experimental progress, much remains to be done to fully characterize the electronic structure and properties of the bismuth tellurohalides. For example, the band gap of BiTeBr and BiTeCl is a matter of ongoing discussion.\cite{Sakano, Sasagawa} Furthermore, in the important experimental work of \citet{Akrap} the question was raised of how the RSS in BiTeCl and BiTeBr influences the interband electronic transitions. In fact, the optical conductivity is one of the most fundamental physical quantities for characterizing the spin and orbital states of matter.
Already in the first studies of the relativistic electron dynamics in BiTeI,\cite{Lee} the optical spectra served as a fingerprint to identify transitions between the Rashba-split energy bands. In these early works,\cite{Lee}
selected elements of the optical conductivity tensor for BiTeI were calculated by applying the Kubo formula on top of an 18-band tight binding model constructed from an {\itshape ab initio} Hamiltonian using maximally localized Wannier functions.\cite{BiTeI_WIEN2k,TB_Wannier1,TB_Wannier2,TB_Wannier3,TB_Wannier4} More recently, the optical conductivity of BiTeX has been measured independently by \citet{Makhnev} in a wide energy range up to $5$ eV. For an unambiguous deduction of the microscopic electronic structure and dynamics, it is therefore desirable to systematically calculate the optical conductivity of the bismuth tellurohalides from first principles.
Generally, the optical conductivity can also be used for the calculation
of the dielectric tensor and hence the refractive index, which has indeed been done for the case of BiTeI in the work of \citet{Rusinov15}.
In this article, we resume this line of research. In particular, we present the entire optical conductivity and dielectric tensors of the bismuth tellurohalides (BiTeI,~BiTeCl,~BiTeBr)
calculated {\itshape ab initio} from density functional theory (DFT).
In order to prove in the first place that the bismuth tellurohalides can be treated reliably within DFT, we will first compute several completely relaxed electronic and structural properties,
which we will later compare to experimental data. Then, we will identify those elements which mainly contribute to the Rashba effect
in the atom-species resolved projected density of states (PDOS). Furthermore, we will show that the electron localization function (ELF) \cite{ELF}
displays a layered structure. After that, we will report the frequency-dependent optical conductivity of BiTeI for different values of the Fermi energy. This will in turn enable us to identify the Rashba-specific peaks in the optical conductivity, which correpond to intra- and interband transitions between the Rashba-split bands. Finally, we will provide a comparison of our {\itshape ab initio} results for the optical conductivity spectra, dielectric constants and refractive indices with the recent experiments performed by \citet{Akrap}, \citet{Makhnev} and \citet{Rusinov15}.
\section{\label{sec:theory}Theoretical details}
As mentioned in the introduction, we will report calculations of the dielectric tensor and the refractive index of BiTeX (X = I, Cl, Br).
The basic quantity computed by the ELK code is, however, the conductivity, and thus we have to clarify its relation to the aforementioned material properties (for details see \citet{Giuliani}). For simplicity, we restrict ourselves to scalar relations of wavevector-independent quantities.
As a matter of principle, the (direct) conductivity relates the induced current to the external electric field by means of
$
\vec j_{\rm ind}=\sigma\vec E_{\rm ext}.
$
By contrast, the proper conductivity relates the induced current to the total electric field,
$
\vec E_{\rm tot}=\vec E_{\rm ext} + \vec E_{\rm ind},
$
by means of $\vec j_{\rm ind} = \widetilde\sigma \vec E_{\rm tot}$.
Finally, the dielectric function mediates between the external and the total electric field in the sense of
$
\vec E_{\rm ext}=\varepsilon_{\rm r} \hspace{1pt} \vec E_{\rm tot}.
$
These quantities are related by the well-known equations
\begin{align}
\varepsilon_{\rm r}^{-1}(\omega) &= 1+\frac{\sigma(\omega)}{\mathrm i\omega\varepsilon_0}\,,\label{eq_notuse}\\[3pt]
\varepsilon_{\rm r}(\omega) &= 1-\frac{\widetilde\sigma(\omega)}{\mathrm i\omega\varepsilon_0} \label{eq_use}\,.
\end{align}
However, although our calculations of the conductivity are based on the Kubo formalism and hence yield the direct conductivity
, we actually use Eq.~\eqref{eq_use} to perform the transition to the dielectric function.
In other words, we interpret the conductivity calculated by the ELK code as the {\it proper} conductivity.
This, of course, requires a certain justification, for which we will provide in the following. For this purpose, we start from the standard relations
\begin{align}
\varepsilon_{\rm r}^{-1}(\omega) &= 1+v\hspace{1pt}\chi(\omega)\,, \label{eq_worse}\\[3pt]
\varepsilon_{\rm r}(\omega) &= 1-v\hspace{1pt}\widetilde\chi(\omega)\,, \label{eq_better}
\end{align}
between the dielectric function and the (direct and proper) density response functions respectively defined by
$\chi = \delta\rho_{\rm ind}/\delta\varphi_{\rm ext}$ and $\widetilde\chi = \delta\rho_{\rm ind}/\delta\varphi_{\rm tot}$, where $\varphi$ is the scalar potential while $v$ denotes the Coulomb interaction kernel.
Using the functional chain rule, one shows directly that the direct density response function is related to its proper counterpart
by the self-consistent equation
\begin{equation}
\chi = \widetilde\chi + \widetilde\chi \hspace{1pt} v \hspace{1pt} \chi\,.
\end{equation}
Suppose now that we consider a many-body system and we are given the density response function $\chi_0$ in a noninteracting approximation,
as it is indeed the case for DFT.
In that case, $\chi_0$ describes the density response function under the assumption that the constituents of the system do not interact with each other.
We now want to approximate the true, i.e. interacting response function $\chi$ of the system by means of its noninteracting counterpart $\chi_0$.
It is plausible that we can do this, if we simply take $\chi_0$ as the response to both the external field {\it and} the induced field generated by the electrons
themselves. This approach takes the interactions into account by simply assuming that the electrons ``feel'' their own induced field in addition to the external field.
Concretely, this means to re-interpret $\chi_0$ as an approximation for the {\it proper} response function, such that the desired
approximation for the interacting response function is given by
\begin{equation}
\chi=\chi_0+\chi_0 \hspace{1pt} v \hspace{1pt} \chi\,.
\end{equation}
This equation constitutes the {\it random phase approximation} (see Chap.~5.3.1.1 in \citet{Giuliani}). Thus, in order to calculate
the dielectric function $\varepsilon_{\rm r}(\omega)$ from $\chi_0(\omega)$ one has to use Eq.~\eqref{eq_better} rather than Eq.~\eqref{eq_worse}.
Correspondingly, as both in the direct and proper case the conductivity and density response function are related by
\begin{equation}
\sigma(\omega)=\mathrm i\omega \hspace{1pt} \varepsilon_0 \hspace{1pt} v \hspace{1pt} \chi(\omega) \,,
\end{equation}
the dielectric function should be calculated by means of Eq.~\eqref{eq_use} rather than Eq.~\eqref{eq_notuse}. This concludes our discussion of the relation between the conductivity and the dielectric function. Finally, the frequency-dependent refractive index
can be directly evaluated as
$n(\omega)=\sqrt{\varepsilon_{\rm r}(\omega)}$. In the case of a wavevector dependence, the relation between the dielectric tensor and the refractive index may become more complicated,\cite{Refr} which is however not considered in this article.
\section{\label{sec:computation}Computational details}
Our calculations have been performed with the ELK \cite{ELK} code, which relies on a full potential (FP), linear augmented plane-wave (LAPW) basis.\cite{LAPW1,LAPW2,LAPW3,LAPW4}
Concretely, we have used a dense $k$-grid of $20\times 20 \times 20$ $k$-points and a cutoff value of $R_{\rm MT}K_{\rm max} = 7$. For the electronic structure calculations we have employed the PBE-GGA \cite{pbe} exchange-correlation functional.
Finally, in order to obtain a Rashba splitting in the band structure, spin-polarization and SOI have been taken into account.\cite{SpinSplitting1,SpinSplitting2}
\begin{figure}[h]
\centering
\includegraphics[width=1.0\linewidth]{PDP_BiTeX.png}
\caption{Calculated powder diffraction pattern for BiTeX (X = I,~Cl,~Br). The black lines in each subplot visualize the experimental structures, whereas the red dots correspond to the {\itshape ab initio} unit cells, where atomic positions and cell parameters have been relaxed.}
\label{PDP_FIG}
\end{figure}
We have used experimental lattice structures \cite{XRAY_STRUCTURES} as an input to
subsequently optimize the lattice constants and volumes to a minimal absolute force value of $0.5 \times 10^{-3}$ Hartree/Bohr.
In order to ascertain that our relaxed crystal structures really correspond to the experiment, we have also calculated
the corresponding powder diffraction patterns (PDP) for experimental as well as theoretically optimized structures (see Fig.~\ref{PDP_FIG}; cf.~also \citet{BiTeI_WIEN2k}). We have verified that the optimized DFT structures produce a similar powder diffraction pattern as compared to their experimental counterparts.
For these computations we have used the open source FullProf program suite.\cite{Fullprof}
In particular, we have meticulously simulated X-ray patterns in the Bragg-Brentano geometry.
As in the work of \citet{Lee}, charge carrier concentrations have been taken into account for the calculation of the
optical conductivity by adjusting the Fermi level manually,
for which purpose the ELK code had to be modified appropriately (see Fig.~\ref{BiTeI_Ef}(b)).
We note that the Fermi level had to be manipulated only for this concrete calculation.
The resulting effect has then been used to show the Rashba-specific features in the entries of the optical conductivity tensor.
Finally, the visualizations of the ELF have been plotted with VESTA, \cite{momma_vesta_2011} using the calculated ELF with an iso-level of 0.55 (min=0, max=0.80).
\section{\label{sec:results}Results and discussion}
For all bismuth tellurohalides, the projected density of states (PDOS) shows that the Fermi level is determined by the p-states of all atom species,
whereas the unoccupied states are dominated by the bismuth p-states (see Fig.~\ref{BiTeX_PDOS}). Consequently, the s-states are far below the Fermi level (approximately 10 eV). Neither do d-states play any r\^{o}le in the DOS around the Fermi level.
By contrast, the states in the valence band are dominated by the halide orbitals.
Correspondingly, the reason for the differences between the bismuth tellurohalides structures
can be ascribed to the contribution of the respective halogen atoms to the valence states. While for BiTeCl and BiTeBr three clearly separated s-bands show up, in the case of BiTeI the corresponding lowest s-bands \mbox{overlap.}
We have also calculated the electronic band gap $E_{\text{G},\text{DFT}}$, which we compare to the available experimental data in Table \ref{tab_E_F}.
While our predicted band gap is quite generally in remarkable agreement with the experiment for BiTeI (see \citet{2011giant}), previous theoretical predictions for BiTeCl and BiTeBr differ from it.\cite{2011giant,Akrap,Rusinov13,BiTeBr_optics} However, our calculations for BiTeCl and BiTeBr yield smaller band gaps as compared to the experimental values. Nevertheless, the overall agreement is satisfactory in our calculations (compare \citet{Rusinov13}, \citet{BiTeBr_optics}, and \citet{Guo16}).
\begin{table}[h!]
\centering
\caption{Fermi energies $E_{\text{F}}$ and band gaps $E_{\text{G}}$ for the three bismuth tellurohalides BiTeX (X = I, Cl, Br)}
\begin{tabular}{c|c|c|c}
\toprule[1pt]
System & $E_{\text{F}}$ (eV) & $E_{\text{G},\text{DFT}}$ (eV) & $E_{\text{G},\text{exp}}$ (eV) \\
\midrule[0.5pt]
BiTeI & 4.58 & 0.37 & 0.38\cite{2011giant}\\
BiTeCl & 4.26 & 0.50 & 0.77\cite{Akrap} \\
BiTeBr & 4.03 & 0.55 & 0.62\cite{Akrap}\\
\bottomrule[1pt]
\end{tabular}
\label{tab_E_F}
\end{table}
\begin{figure}[t!]
\centering
\subfigure[~BiTeI]{\includegraphics[width=\linewidth]{BiTeI_PDOS.png}}
\subfigure[~BiTeCl]{\includegraphics[width=\linewidth]{BiTeCl_PDOS.png}}
\subfigure[~BiTeBr]{\includegraphics[width=\linewidth]{BiTeBr_PDOS.png}}
\caption{Total density of states (TDOS) and its projections (PDOS) to s- and p-states of the constituent atoms for each halide. The corresponding Fermi energies $E_{\text{F}}$ and band gaps $E_{\text{G}}$ are given in Table \ref{tab_E_F}. All Fermi levels are exclusively determined by p-states, where the tellurium and halogen atoms show the largest share. The underlying s-states contribute to the DOS not until 10 meV below the Fermi level. In contrast, unoccupied states are mainly determined by p-states of Bi atoms.}
\label{BiTeX_PDOS}
\end{figure}
Turning to the ELF (see Fig.~\ref{BiTeX_ELF}), we first observe that in the case of BiTeI the electrons are localized in two layers (\enquote{2d-electron gas}),
one layer around the Te atom and the other layer around the I atom. This 2d-electron gas can also be observed in BiTeCl and BiTeBr. In the case of BiTeBr,
the electron density is mainly localized at the Br and the Te atoms in the form of two seperate layers. Similarly, the localization layers
of BiTeCl are centered at the Cl and the Te atom. However, all bismuth tellurohalides have in common that one electron localization layer is formed by the Te atom and the other by the halogen atom.
This layered structure is well known for the case of BiTeI,\cite{Kilic} but to our knowledge not generally discussed
for the whole class of bismuth tellurohalides.
Moreover, we have calculated the optical conductivity tensor of BiTeX (X = I, Cl, Br) in a wide energy range up to 12 eV.
In comparing its diagonal elements, one observes similarly shaped spectra with coinciding orders of magnitude
in $\sigma_{xx}$ and $\sigma_{yy}$, whereas $\sigma_{zz}$ displays a completely different behaviour. This is in fact the case
for all bismuth tellurohalides (see Fig.~\ref{BiTeX_sigma}). Interestingly, the lower frequency branches of the conductivity are generally peaked in the $(110)$-plane. The high-frequency spectra of BiTeCl and BiTeBr are very similar despite their different point groups, which is in accord with the experimental findings of \citet{Akrap}.
Furthermore, our calculated optical conductivity spectrum of BiTeI agrees well with the measurements of \citet{Makhnev} (see Fig.~3 there) over a wide frequency range.
\begin{figure}[ht!]
\centering
\subfigure[~BiTeI]{\includegraphics[width=0.45\linewidth]{BiTeI_ELF_110.png}
\hspace{2em} \includegraphics[width=0.05\linewidth]{ELF_scale.png}}
\subfigure[~BiTeCl]{\includegraphics[width=0.45\linewidth]{BiTeCl_ELF_110.png}
\hspace{2em} \includegraphics[width=0.05\linewidth]{ELF_scale.png}}
\subfigure[~BiTeBr]{\includegraphics[width=0.45\linewidth]{BiTeBr_ELF_110.png}
\hspace{2em} \includegraphics[width=0.05\linewidth]{ELF_scale.png}}
\caption{Electron localization function (ELF) for the three bismuth tellurohalides BiTeX (X = I,~Cl,~Br).
Red parts correspond to areas with highly localized electron density, whereas blue parts symbolize less strong localization.
Two layers of higher localized density are formed, one centered at the tellurium atom and the other one centered at the halogen atom.}
\label{BiTeX_ELF}
\end{figure}
\begin{table}[h!!!!!!]
\centering
\caption{Relative dielectric constants $\varepsilon_{{\rm r},ii} (\omega = 0)$ and refractive indices $n (\omega = 0 )$ for BiTeX (X = I, Cl, Br).}
\begin{tabular}{c|c|c|c|c}
\toprule[1pt]
System & $\varepsilon_{{\rm r},xx} (\omega = 0)$& $\varepsilon_{{\rm r},yy} (\omega = 0)$ & $\varepsilon_{{\rm r},zz} (\omega = 0)$ & $n (\omega = 0 )$ \\
\midrule[0.5pt]
BiTeI & 21.64 & 21.52 & 15.54 & 4.65\\
BiTeCl & 15.11 & 15.08 & 10.38 & 3.89 \\
BiTeBr & 16.59 & 16.60 & 11.64 & 4.07 \\
\bottomrule[1pt]
\end{tabular}
\label{tab_n}
\end{table}
\FloatBarrier
\onecolumngrid
\begin{figure}[hb!]
\centering
\subfigure[~BiTeI: low-frequency]{\includegraphics[width=0.329\linewidth]{BiTeI_sigma_small.png}}
\subfigure[~BiTeCl: low-frequency]{\includegraphics[width=0.329\linewidth]{BiTeCl_sigma_small.png}}
\subfigure[~BiTeBr: low frequency]{\includegraphics[width=0.329\linewidth]{BiTeBr_sigma_small.png}}
\subfigure[~BiTeI: high-frequency]{\includegraphics[width=0.329\linewidth]{BiTeI_sigma_large.png}}
\subfigure[~BiTeCl: high-frequency]{\includegraphics[width=0.329\linewidth]{BiTeCl_sigma_large.png}}
\subfigure[~BiTeBr: high-frequency]{\includegraphics[width=0.329\linewidth]{BiTeBr_sigma_large.png}}
\caption{Real parts of the optical conductivities for BiTeX (X = I,~Cl,~Br). The low-frequency regions are plotted separately in (a), (b) and (c). The
characteristic energies marked with $\gamma$ and $\delta$ correspond to interband transitions (cf.~Fig.~\ref{BiTeI_Ef}(b)).}
\label{BiTeX_sigma}
\end{figure}
\begin{figure}[ht!]
\centering
\subfigure[~BiTeI]{\includegraphics[width=0.329\linewidth]{BiTeI_epsi.png}}
\subfigure[~BiTeCl]{\includegraphics[width=0.329\linewidth]{BiTeCl_epsi.png}}
\subfigure[~BiTeBr]{\includegraphics[width=0.329\linewidth]{BiTeBr_epsi.png}}
\caption{Dielectric function for BiTeX (X = I,~Cl,~Br). Red lines represent real and blue lines imaginary parts of the dielectric function.}
\label{BiTeX_EPSI}
\end{figure}
\twocolumngrid
\FloatBarrier
Similarly as in the works of \citet{Lee} and \citet{Demko},
we have determined the {\itshape intra}band ($\alpha$,\,$\beta$) transitions (i.e.~transitions within the spin-split conduction bands) and {\itshape inter}band ($\gamma$,\,$\delta$) transitions (i.e.~transitions between valence and conduction bands) in the optical conductivity tensor. The spectral peaks at 0.4 and 0.6~eV are identical for BiTeI, which precisely corresponds to the said interband transitions ($\gamma$,\,$\delta$).\cite{Lee}
The corresponding peaks appear also in BiTeCl and BiTeBr, but their magnitudes and positions differ among these two bismuth tellurohalides. These intra- and interband transitions have also been detected experimentally by \citet{Akrap} (see Fig.~4 there). We remark that the optical transitions between bands with different spin polarization are theoretically expected to occur as a consequence of the SOI.\cite{Lee, Sakano}
\begin{figure}[ht!]
\centering
\subfigure[~BiTeI bandstruture]{\includegraphics[width=0.9\linewidth]{BiTeI_transitions.png}}
\subfigure[~BiTeI optical conductivity]{\includegraphics[width=0.9\linewidth]{BiTeI_sig_Ef.png}}
\caption{Real parts of the optical conductivity for BiTeI in dependency of the Fermi level $E_{\text{F}}$. For illustration we have plotted in (a) the different Fermi levels in the band structure of BiTeI,
where the same colors (black: $E_{\text{F},0}$ = 4.57 eV, blue: $E_{\text{F},1}$ = 4.80 eV, red: $E_{\text{F},2}$ = 5.00 eV) for the Fermi level are used in (b) for the corresponding optical conductivity results. In subfigure (b) the characteristic energies correspond to intraband transitions ($\alpha$,\,$\beta$) and interband transitions ($\gamma$,\,$\delta$) as shown schematically in subfigure (a). \label{BiTeI_Ef}}
\end{figure}
Specifically in the case of BiTeI, we have also calculated the optical conductivity (see Fig.~\ref{BiTeI_Ef}(b)) for three different Fermi levels ($E_{\text{F},0}$, $E_{\text{F},1}$, $E_{\text{F},2}$) to simulate the effect of different doping levels. For the highest Fermi level $E_{\text{F},2}$ = 5.0 eV we obtain intraband transitions ($\alpha$, $\beta$) at about 0.2 and 0.35 eV.
By contrast, in the case of the lowest Fermi level,
the first peak in the optical conductivity (see Fig.~\ref{BiTeI_Ef}(b)) near 0.39 eV corresponds to the electronic band gap of 0.37 eV,
while the second peak localized near 0.59 eV corresponds to the transition at the A point (see Fig.~\ref{BiTeI_Ef}(a)).
Generally, by inspection of Table~\ref{tab_E_F}, we observe a rapprochement of the $\gamma$ and $\delta$ peaks with decreasing Fermi level.
Furthermore, we read off that the $\gamma$-peak roughly coincides with the electronic band gap (compare Table~\ref{tab_E_F} with Figs.~\ref{BiTeX_sigma}(a)--(c)).
Note that BiTeCl displays only one peak at 0.57 eV, which is slightly different from the electronic gap whose value is 0.5 eV.
Fittingly, the transition at the $\Gamma$ point has a value of 0.61 eV.
In addition to the conductivity, we have calculated by means of Eq.~\eqref{eq_use} the dielectric function for all bismuth tellurohalides (see Fig.~\ref{BiTeX_EPSI})
and the resulting dielectric constant (see Table~\ref{tab_n}). Our results for BiTeI and BiTeCl turn out to be in good agreement with the values given in the work of \citet{Rusinov15}.
Correspondingly, we have also calculated the refractive index by means of the relation $n(\omega) = \sqrt{\varepsilon_{\rm r}(\omega)}$ (see Table~\ref{tab_n}) for BiTeX (X = I,~Cl,~Br). For BiTeI and BiTeCl our results are in good agreement with
the experimental values reported by \citet{Rusinov15}. Moreover, our results for BiTeBr agree with the theoretical values predicted by \citet{BiTeBr_optics}.
\section{\label{sec:conclusion}Conclusion}
We have shown that the bismuth tellurohalides BiTeI, BiTeCl and BiTeBr, which are highly anisotropic materials with spin-split energy bands, can be reliably treated within DFT (see Fig.~\ref{PDP_FIG}). Moreover, we have shown that all bismuth tellurohalides display a layered electronic localization (see Fig.~\ref{BiTeX_ELF}). Correspondingly, while two of the diagonal elements of the optical conductivity ($\sigma_{xx}$ and $\sigma_{yy}$) display a similar behaviour as a function of the frequency,
the contribution transverse to the electron localization layer, i.e.~$\sigma_{zz}$, is significantly smaller in the low-frequency region. A further central result of this work is the identification of Rashba-specific transitions (see Fig.~\ref{BiTeX_sigma} and Fig.~\ref{BiTeI_Ef}) within the low frequency branch of the optical conductivity of all three bismuth tellurohalides.
Our calculations of the optical conductivity complement and extend the theoretical results of \citet{Lee} in the following respects: (i) The optical conductivity has been calculated from first principles using all-electron DFT as implemented in the ELK \cite{ELK} code, (ii) we have calculated the whole conductivity tensor (including $\sigma_{zz}$) in both the low-frequency and high-frequency range, and (iii) we have extended the calculation to the compounds BiTeBr and BiTeCl. On the other hand, our results confirm the experimental findings of \citet{Akrap}, where optical transitions were observed even for BiTeBr and BiTeCl, which have a smaller RSS as compared to BiTeI. Furthermore, they confirm that the high-frequency optical spectra of BiTeBr and BiTeCl are similar despite their different space groups.\cite{Akrap} Finally, the optical conductivity of BiTeI as calculated from DFT agrees well with the recent experimental results of \citet{Makhnev} over a wide frequency range. Thus, this work contributes to the understanding of the electron dynamics in the Rashba semiconductors BiTeX (X = I, Cl, Br). Moreover, as optical transitions between Rashba-split bands may be relevant for the resonant dynamical magnetoelectric effect and the spin Hall effect, \cite{Lee} this work also confirms the bismuth tellurohalides as promising candidates for future spintronic applications.
\section*{\label{sec:acknowledgement}Acknowledgements}
The authors thank the group of theoretical physics at the TU Bergakademie Freiberg for seminal discussions.
Furthermore, we thank the ZIH in Dresden for computational time support. S.\,S.~thanks the
SPP 1473 (WeNDeLIB -- Werkstoffe mit neuem Design f\"ur verbesserte Lithium-Ionen-Batterien) for funding. G.\,S.~was supported by the DFG Research Unit FOR723.
|
\section{Exclusive longitudinal phase space analysis and its variables}
One of our common interests at that time was the so-called Longitudinal Phase Space (LPS) Analysis of 3- and 4- or even 5-particle final states and I will first try to recall the ideas behind that.
The most complete way to study a so-called exclusive reaction of multiplicity $n$
\begin{equation}
A + B \longrightarrow C_1 + C_2 + \dots + C_n
\end{equation}
is to look at the differential distribution of its matrix element in full phase space. This, however, requires a $3n\mi4$ dimensional analysis ($(3n\mi5)$-dimensional if the incident particles are unpolarized) and becomes increasingly impossible with increasing $n$.
Nature helps: at low cms energies, the vast majority of collisions is "soft", i.e. leads to low transverse (with respect to the collision axis) momenta of final state particles, largely independent of the nature of the particle, the multiplicity $n$ and the cms energy $s^{^1\hs-0.8truemm/\hs-0.5truemm_2}$. On the other hand, longitudinal (along the collision axis) momenta are unlimited (i.e. limited only by phase space) and depend strongly on the nature of the particle, the multiplicity and the energy.
In elastic and other two-particle production collisions, one is used to distinguish between forward and backward scattering. An extension of this classification to multiparticle final states is an analysis in just longitudinal phase space (LPS) \cite{1,2}. Then, each individual reaction of type (1) is represented by a point with coordinates ($p_{\| 1}, \dots , p_{\| n}$) in a now only
$n$-dimensional euclidean space $S_n$. Conservation of longitudinal momentum in the cms,
\begin{equation}
\sum^n p^{*}_{\| i}=0\ ,
\end{equation}
defines LPS as an $(n\mi1)$ dimensional hyperplane $L_{n-1}$. Furthermore, because of conservation of cms energy $s^{^1\hs-0.8truemm/\hs-0.5truemm_2}$
\begin{equation}
\sum^n_{i=1}(m^2_i+ \ifmath{{\mathrm{p}}}^2_{\ifmath{{\mathrm{T}}} i}+p^{*2}_{\| i})^{^1\hs-0.8truemm/\hs-0.5truemm_2}=s^{^1\hs-0.8truemm/\hs-0.5truemm_2}\ .
\end{equation}
All points with equal transverse momentum $|\ifmath{{\mathrm{p}}}_{\ifmath{{\mathrm{T}}} i}|$ lie on an $(n\mi2)$ dimensional hypersurface $K_{n-2}$
defined by (3). For the case of a transverse mass $m_{\ifmath{{\mathrm{T}}} i }= (m_{i}^2+\ifmath{{\mathrm{p}}}_{\ifmath{{\mathrm{T}}} i}^2)^{^1\hs-0.8truemm/\hs-0.5truemm_2} = 0$, (3) reduces to
\begin{equation}
{\sum^n_{i=1}} |p^*_{\| i}| = s^{^1\hs-0.8truemm/\hs-0.5truemm_2}
\end{equation}
and defines a regular polyhedron $H_{n-2}$. For $n=3$, this is the Van Hove Hexagon shown in Fig.~1 together with the
one-dimensional manifold $K_1$. For $n=4$, the polyhedron $H_2$ is the cuboctrahedron celebrated in Fig.~2.
A typical three-particle distribution in LPS for the final state of reaction
\begin{equation}
\pi^-\ifmath{{\mathrm{p}}} \to \ifmath{{\mathrm{p}}}\pi^-\pi^0
\end{equation}
at an incident lab momentum of 16 GeV/$c$ is given in Fig.~3 \cite{3}.
\begin{figure}[htb]
\begin{minipage}[t]{6cm}
\includegraphics[width=6cm]{fig1.pdf}
\caption{Longitudinal phase space plot (Van Hove Hexagon) for the final state $\pi\p$N\ at cms energy of
$s^{^1\hs-0.8truemm/\hs-0.5truemm_2}=4$ GeV. The innermost full line is $K_1$ for transverse
momenta of 0.4, 0.4, and 0.5 GeV/$c$, respectively,
while the outer one is $K_1$ for
vanishing transverse momenta. The dashed line represents the hexagon $H_1$ \cite{1}.
}
\end{minipage}
\hskip 5mm
\begin{minipage}[t]{6cm}
\includegraphics[width=6cm]{fig2.pdf}
\caption{The polyhedron $H_2$ for a four-particle final state (cartoon by R.Sosnowski).}
\end{minipage}
\end{figure}
\begin{figure}[htb]
\begin{minipage}[c]{6cm}
\includegraphics[width=6cm]{fig3.pdf}
\caption{Distribution of final state points for the reaction $\pi^-\to \pi^-\pi^0\ifmath{{\mathrm{p}}}$ at incident lab momentum
of 16 GeV/$c$ \cite{3}.}
\end{minipage}
\hskip4mm
\begin{minipage}[c]{6cm}
\includegraphics[width=5cm]{fig4.pdf}
\end{minipage}
\caption{(a) Distribution against the angle $\omega$ of final state points for the reaction $\pi^-\ifmath{{\mathrm{p}}} \to \pi^-\pi^0\ifmath{{\mathrm{p}}}$ at incident lab momentum of 16 GeV/$c$ \cite{3}. (b) same after correcting for non-constant phase space effects.
The solid lines are the distributions according to the C\L A model \cite{4} normalized to the data after exclusion of the sharp
$\rho^-$-resonance. The individual C\L A exchange graphs considered and their contributions to the total distribution are given
in sub-figure (c).}
\end{figure}
The distribution of (5) against the angle $\omega$ is given in Fig.~4, before and after correcting for phase space effects (sub-figure (a) and (b), respectively). According to the definition in Fig.~1,
the $\omega$ region considered ($60^\circ<\omega<120^\circ)$ corresponds\ to the hemisphere of LPS in which the proton is backward in the cms. The $\pi^0$\ is taken to be longitudinally at "rest" at $\omega=120^\circ$, the $\pi^-$ at $\omega=180^\circ$. Peaks in the\ (model-independent!) experimental data (histograms) indicate strong correlations between particles in the final state, in particular in the region $60<\omega<120^\circ$.
As a demonstration of how one can use LPS to test the success of theoretical models, the so-called C\L A model \cite{4}
of that time was used. It is a Reggeized form of a multiperipheral model, in which the amplitude is treated as an incoherent sum of contributions from various multiperipheral graphs. For the reaction studied in Fig.~4, they are given in sub-figure (c) together with their contributions. The full line in (a) and (b) corresponds to their incoherent sum.
After exclusion of a sharp $\rho$-resonance, the model can describe the overall distribution of Fig.~4 surprisingly well. From the contribution of the graphs in sub-figure (c) we can see that vacuum exchange, commonly called I$\!$P(omeron) exchange, on the upper vertex essentially determines the shape of the distribution.
Turning back to a model independent data analysis, we investigate the energy dependence of the distribution and its shape in
Fig.~5 according to its parametrization $\sigma(p_{\ifmath{{\mathrm{lab}}}}) \propto p_{\ifmath{{\mathrm{lab}}}}^{-N}$. As shown in sub-fibgure (b), $N$ is indeed close to zero for $60^\circ<\omega<120^\circ$,\ in agreement with I$\!$P\ exchange (diffraction dissociation) in that region \cite{5}.
\begin{figure}[htb]
\centerline{%
\includegraphics[width=7cm]{fig5.pdf}}
\caption{(a) $\omega$-distribution for the reaction $\pi^+\ifmath{{\mathrm{p}}} \to \pi^+\pi^0\ifmath{{\mathrm{p}}}$ at incident lab momentum of 4, 5
and 8 GeV/$c$ . (b) Exponent $N$ as a function of $\omega$ for the same reaction \cite{5}.}
\end{figure}
Where, however, is the $\Delta$\ resonance? Unlike the incident proton, it has isospin $I=\/^3\hs-0.9truemm/\hs-0.5truemm_2$ and can, therefore, not be produced via vacuum exchange.
One way to look for it, is the so-called prism plot [6] ingeniously combining the advantages of the angle $\omega$ along the z-axis with the subsystem masses given in a Fabri-Dalitz plot (triangle) at the basis (Figs.~6 and 7). The separation into individual mechanisms, each corresponding to a straight section of the tube within the prism, is better than in its projection onto the
z-axis or the basis in the xy-plane. The mass of the ($\ifmath{{\mathrm{p}}}\pi^+$)-subsystem is plotted in Fig.~8, for all events (sub-figure (a)) and for events in the corresponding section of the tube (sub-figure (b)). In the latter, the $\Delta^{++}$ is well separated from the background still present in (a).
\begin{figure}[htb]
\begin{minipage}[t]{6cm}
\centerline{%
\includegraphics[width=6cm]{fig6.pdf}}
\caption{The prism plot as constructed by pulling a Fabri-Dalitz plot out in the direction of the Van Hove angle $\omega$ (cartoon by Suzy Smile).}
\end{minipage}
\hskip4mm
\begin{minipage}[t]{6cm}
\centerline{%
\includegraphics[width=6cm]{fig7.pdf}}
\caption{Prism plot for $\pi^+\ifmath{{\mathrm{p}}} \to \pi^+\pi^0\ifmath{{\mathrm{p}}}$ at 3.9 GeV/$c$. (a) invariant phase space and (b) experimental
data \cite{6}.}
\end{minipage}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=6.5cm]{fig8a.pdf} \hskip 1mm \includegraphics[width=6.5cm]{fig8b.pdf}
\caption{Effective mass of the $(\ifmath{{\mathrm{p}}}\pi^+)$-subsystem for the reaction $\pi^+\ifmath{{\mathrm{p}}}\to \pi^+\pi^0\ifmath{{\mathrm{p}}}$ at incident lab
momentum of 3.9 GeV/$c$, for (a) all events and (b) for events in the corresponding section of the tube in the prism plot \cite{6}.}
\end{figure}
\begin{figure}[htb]
\centerline{%
\includegraphics[width=6cm]{fig9.pdf}}
\caption{The three diagrams corresponding to the amplitudes specified by the exchanged isospin $I_\ifmath{{\mathrm{E}}}$
and the isospin $I$ of the $(\ifmath{{\mathrm{N}}}\pi)$ system. (Note that the combination $I_\ifmath{{\mathrm{E}}}=0, I=\/^3\hs-0.9truemm/\hs-0.5truemm_2$ is excluded from isospin conservation.)}
\end{figure}
\begin{figure}[htb]
\centerline{%
\includegraphics[width=10cm]{fig10.pdf}}
\caption{Squared amplitudes $|M_I^{I_\ifmath{{\mathrm{E}}}}|^2$ and their interference terms as functions of the $(\ifmath{{\mathrm{N}}}\pi)$ mass obtained from the reactions $\pi^\pm \ifmath{{\mathrm{p}}} \to \pi\p\ifmath{{\mathrm{N}}}$ at 16 GeV/$c$ \cite{7}.}
\end{figure}
Another way of extracting a pure $\Delta$ signal is a separation of the
isospin matrix element according to the graph in
Fig.~9. Since isospin exchange $I_\ifmath{{\mathrm{E}}}=0$ is excluded for the production of the $I=\/^3\hs-0.9truemm/\hs-0.5truemm_2\ (\ifmath{{\mathrm{N}}}\pi)$-system, we are left with three matrix elements and their interferences. Their squares, respectively real parts, can readily be extracted model independently
from combinations of the 6 measurable (of the 7 possible) final states of $\pi^\pm\ifmath{{\mathrm{p}}}$ reactions. They are given in Fig.~10 as a function of the $(\ifmath{{\mathrm{N}}}\pi)$ effective mass. While a wide diffractive shoulder is observed in sub-figure (a), a sharp and well
separated $\Delta$ can be seen in subfigure (c).
Overlap between different sub-systems is a problem, in particular in the determination of spin-parity of a particular sub-system (partial wave analysis). However, interference also provides a unique possibility to study the relative phase between overlapping amplitudes. In such a study, all mechanisms contributing to a particular few-body final state have to be treated simultaneously in an iterative and interactive computer analysis. A beautiful method allowing that is the so-called Analytical Multichannel Analysis \cite{8}.
The method has successfully been applied to 30 000 events of the final state of $\ifmath{{\mathrm{K}}}^-\ifmath{{\mathrm{p}}}\to \overline\ifmath{{\mathrm{K}}}^0\pi^-\ifmath{{\mathrm{p}}}$
at 4,2 GeV/$c$ \cite{9}. As the four variables needed to describe a three-particle final state, the effective mass
$M$ has been used for the sub-system considered, the invariant four-momentum $t'$ in its production and the two angular variables $\Theta$ and $\phi$ of its decay.
\begin{figure}[t]
\centerline{%
\includegraphics[width=11cm]{fig11.pdf}}
\caption{Effective mass distribution of the $(\overline\ifmath{{\mathrm{K}}}^0\pi^-)$ system, four-momentum transfer $t'$ from
initial to final state proton, decay angles of the $(\overline\ifmath{{\mathrm{K}}}^0\pi^-)$ system, and effective mass of the $(\ifmath{{\mathrm{p}}}\pi^-)$- and
$(\ifmath{{\mathrm{p}}} \overline\ifmath{{\mathrm{K}}}^0)$-systems for the
$0^+0-$ and $1^-0-$ $(\overline\ifmath{{\mathrm{K}}}^0\pi^-)$ samples after iteration 9 \cite{9}.}
\end{figure}
As examples for the results after 9 iterations, Figs. 11 and 12 correspond to the $(\overline\ifmath{{\mathrm{K}}}^0\pi^-)$ S-wave and its P-waves
$1^-0-$, $1^-1-$ and $1^-1+$. Except for the S-wave which is not yet flat, the angular distributions correspond to the particular wave and are as expected. Of particular interest are the differences in the four-momentum $t'$ distributions for the three P-waves, typical for pseudo-scalar and vector exchange, respectively.
Striking is the difference in the reflection into the $(\ifmath{{\mathrm{p}}}\pi^-)$ and $(\ifmath{{\mathrm{p}}}\overline\ifmath{{\mathrm{K}}}^0)$ systems in the lowest row of Fig.~12.
The Monte Carlo curve superimposed on the $(\ifmath{{\mathrm{p}}}\pi^-)$ mass distribution shows a two-peak reflection from the $1^-1-$
wave. Just mind the enormous error introduced by the simple smooth hand-drawn background as used in earlier conventional analysis!
Very similar results are obtained \cite{9} for the three D-and even F-waves and for other sub-channels down to the $^o\hs-0.8truemm/\hs-0.8truemm_{oo}$ level of their contribution to the total final state, not detectable in earlier analysis.
\begin{figure}[t]
\centerline
\includegraphics[width=11cm]{fig12.pdf}}
\caption{Same as Fig. 11, but for the $1^-1-$ and $1^-1+$ samples \cite{9}.}
\end{figure}
In conclusion from this section: With the help of model independent data analysis we have moved from analysis in Longitudinal Phase Space to a complete Multichannel Analysis. While the LPS analysis has demonstrated strong correlation of final state particles, the increased number of variables of the prism plot could show overlap of these mechanisms in full phase space (not just in projections of it). These mechanisms can be separated by means of quantum numbers as isospin, spin and angular momentum, and their interferences can be studied when all channels contributing to a final state are treated simultaneously. This analysis is particularly useful for the isolation of channels at the permille level of cross section or branching ratio.
What had we learned for the future? Experiments have to be {\it complete} in the sense that acceptance losses should be minimal and the four-vectors of all particles should be known. Furthermore, the analysis has to be done {\it iteratively and interactively}, i.e has to be guided by computer graphics.
\section{Momentum correlations and density fluctuations}
\subsection{The formalism}
We start by defining symmetrized inclusive $q$-particle distributions
\begin{equation}
\rho_q (p_1,\dots,p_q)= \frac{1}{\sigma_\ifmath{{\mathrm{tot}}}} \frac{\ifmath{{\mathrm{d}}} \sigma_q(p_1,\dots,p_q)}{
\prod\limits^q_1 \ifmath{{\mathrm{d}}} p_q}\ \ ,
\end{equation}
\vskip-3mm
\noindent
where $\sigma_q(p_1,\dots,p_q)$ is the inclusive cross section for $q$
particles to be at $p_1,\dots,p_q$, irrespective of the presence and location
of any further particles, $p_i$ is the (four-) momentum of particle $i$ and
$\sigma_\ifmath{{\mathrm{tot}}}$ is the total hadronic cross section of the collision under
study.
For the case of identical particles, integration over an interval
$\Omega$ in $p$-space yields
\begin{eqnarray}
&~&\int_\Omega \rho_1(p) \ifmath{{\mathrm{d}}} p = \langle n\rangle\ , \ \ \ \ \
\int_\Omega \int_\Omega \rho_2(p_1,p_2)\ifmath{{\mathrm{d}}} p_1\ifmath{{\mathrm{d}}} p_2 =
\langle n(n-1)\rangle \ ,\nonumber \\
&~&\int_\Omega \ifmath{{\mathrm{d}}} p_1 \dots \int_\Omega \ifmath{{\mathrm{d}}} p_q \rho_q (p_1,\dots,p_q) =
\langle n(n-1)\dots (n-q+1)\rangle \ ,
\end{eqnarray}
where $n$ is the multiplicity of identical particles within $\Omega$ in a
given event and the angular brackets imply the average over the event ensemble.
Besides the interparticle {\it correlations} we are looking for,
the inclusive $q$-particle number densities $\rho_q(p_1,\dots,p_q)$ in general
contain ``trivial'' contributions from lower-order densities.
It is, therefore, advantageous to consider a new sequence of functions
$C_q(p_1,\dots,p_q)$ as those statistical quantities which vanish whenever one
of their arguments becomes statistically independent of the others
\cite{kahn:uhlenbeck,huang,Mue71}:
\vskip-3mm
\begin{eqnarray}
C_2(1,2)&=&\rho_2(1,2) -\rho_1(1)\rho_1(2)\ ,\\
C_3(1,2,3)&=&\rho_3(1,2,3)
-\sum_{(3)}\rho_1(1)\rho_2(2,3)+2\rho_1(1)\rho_1(2)\rho_1(3)\ ,
\label{a:4b}
\end{eqnarray}
\vskip-3mm
\noindent
etc. In the above relations, we have abbreviated $C_q(p_1,\dots,p_q)$ to\break
$C_q(1,2,\dots,q)$; the summations indicate that all possible permutations
must be taken. Expressions for higher orders can be derived from the related
formulae given in~\cite{kendall}.
Deviations of these functions from zero shall be addressed as {\it genuine}
correlations.
It is often convenient to divide the functions
$\rho_q$ and $C_q$ by the product of $q$ one-particle densities, which leads to
the definition of the normalized inclusive densities and correlations:
\begin{eqnarray}
R_q(p_1,\dots,p_q) &=& \rho_q(p_q,\dots,p_q)/\rho_1(p_1)\ldots
\rho_1(p_q),\label{3.8}\\
K_q(p_1,\ldots,p_q)& =& C_q(p_1,\ldots,p_q)/\rho_1(p_1)\ldots
\rho_1(p_q).\label{3.9}
\end{eqnarray}
In terms of these functions, correlations have been studied extensively for
$q=2$. Results also exist for $q=3$, but usually the statistics (i.e. number
of events available for analysis) are too small to isolate genuine
correlations. To be able to do that for $q\geq 3$, one must apply factorial
moments $F_q$
defined via the integrals Eq.~(7), but in limited phase-space cells
\cite{bialas,wolf96}.
\subsection{Density spikes}
To see whether it is worth the effort, we first look for
density fluctuations in single events, signalling high-order correlations.
A notorious JACEE event~\cite{Burn83} (Fig.~13a) at a pseudo-rapidity
resolution (binning) of $\delta\eta=0.1$ has
local fluctuations up to $\ifmath{{\mathrm{d}}} n/\ifmath{{\mathrm{d}}}\eta\approx300$
with a signal-to-background ratio of about 1:1. An NA22 event
\cite{AdamPL185-87} (Fig.~13b) contains a ``spike" at a rapidity
resolution $\delta y=0.1$ of $\ifmath{{\mathrm{d}}} n/\ifmath{{\mathrm{d}}} y=100$, as much
as 60 times the average density in this experiment.
\begin{figure}[htb]
\begin{minipage}[t]{6cm}
\includegraphics[height=6.9cm]{fig10-1anew.pdf}
\end{minipage}
\hskip -4mm
\begin{minipage}[t]{6cm}
\includegraphics[height=6.9cm]{fig10-1bnew.pdf}
\end{minipage}
\caption{a) The JACEE event \cite{Burn83}; b) The NA22 event \cite{AdamPL185-87}.}
\end{figure}
{Bia\l as} and Peschanski \cite{bialas} suggested that this type of spikes
could be a
manifestation of ``intermittency'', a phenomenon well known in fluid dynamics
\cite{Zeld90}. The authors argued that if
intermittency indeed occurs in particle production, large density
fluctuations are not only expected, but should also
exhibit self-similarity with respect to the size of the
phase-space volume.
In multiparticle experiments, the number of hadrons produced in a single
collision is small and subject to considerable noise. To
exploit the techniques employed in complex-system
theory, a method had to be devised to separate fluctuations of purely
statistical (Poisson) origin, due to finite particle numbers, from
possibly self-similar dynamical fluctuations of the underlying
particle densities.
A solution, already used in quantum optics
\cite{Bedard:67:1} and suggested for
multiparticle production in~\cite{bialas}, consists in measuring
$F_q(\delta y)$ in given
phase-space volumes (resolution) $\delta y$ of ever decreasing size.
\subsection{Power-law scaling}
Besides the property
of noise-suppression, high-order factorial moments act as a filter
and resolve the large-multiplicity tail of the multiplicity distribution.
They are thus particularly sensitive to large density fluctuations
at the various scales $\delta y$ used in the analysis.
As shown in~\cite{bialas}, a smooth density distribution, which does
not show any fluctuations except for the statistical ones, has the property
of normalized factorial moments $F_q(\delta y)$ being independent of the
resolution $\delta y$ in the limit $\delta y\to 0$. On the other hand, if
self-similar dynamical fluctuations exist, the $F_q$ obey the power law
\begin{equation}
F_q(\delta y) \propto (\delta y)^{-\phi_q}\ , \ \ (\delta y\to 0).
\label{12}
\end{equation}
Equation (\ref{12}) is a scaling law
since the ratio of the factorial moments at resolutions $L$ and $\ell$
\begin{equation}
R = \frac{F_q(\ell)}{F_q(L)} = \left(\frac{L}{\ell}\right)^{\phi_q}
\end{equation}
only depends on the ratio $L/\ell$, but not on $L$ and $\ell$,
themselves.
\begin{figure}[h
\centerline{%
\includegraphics[width=7cm]{fig10-2n.pdf}}
\caption{ log$F_5$ as a function of $-\log\delta\eta$ for the JACEE
event~\cite{Burn83} (full circles) compared to independent
emission (small crosses) \cite{bialas}.}
\end{figure}
In Fig.~14, log$F_5$ is plotted~\cite{bialas} as a function of -log
$\delta\eta$ ($\eta$ is the pseudorapidity) for the JACEE
event. It is compared with an independent-emission
Monte-Carlo
model tuned to reproduce the average $\eta$ distribution of
Fig.~13 a) and the global multiplicity
distribution, but has no short-range correlations.
While the Monte-Carlo model indeed predicts constant $F_5$, the
JACEE event shows a first indication for a linear increase, i.e. a possible
sign of intermittency.
This observation was the trigger for a tremendous outburst of
experimental research on all types of collisions from e$^+$e$^-$ to heavy
nuclei, all showing (approximate) power law scaling. An 118 page
summary including more than 300 references is given as chapters
7 and 10 in \cite{kitwolf2005}.
The powers $\phi_q$ (slopes in a double-log plot) are related~\cite{LiBu89}
to the anomalous (or co-) dimensions $d_q=\phi_q/(q-1)$, a measure for the
deviation from an integer dimension.
Anomalous dimensions $d_q$ fitted over the (one-dimensional)
range $0.1<\delta y<1.0$ are compiled in Fig.~15~\cite{Bia1991}. They typically
range from $d_q=0.01$ to $0.1$, which means that the
fractal (R\'enyi) dimensions $D_q=1-d_q$ are close to one.
The $d_q$ are larger and grow faster with increasing order $q$ in $\mu$p and
${\rm e}^+{\rm e}^-$ (Fig.~15a) than in hh collisions (Fig.~15b) and are small and almost
independent of $q$ in heavy-ion collisions (Fig.~15c). For hh collisions,
the $q$-dependence is considerably stronger for NA22 ($\sqrt{s}=22$ GeV,
all $p_\ifmath{{\mathrm{T}}}$) than for UA1 ($\sqrt{s}=630$ GeV, $p_\ifmath{{\mathrm{T}}}>0.15$ GeV/$c$).
\begin{figure}[htb]
\centerline
\includegraphics[width=12.5cm]{fig6-3.pdf}}
\caption{
Anomalous dimension $d_q$ as a
function of the order $q$, for a) $\mu$p and ${\rm e}^+{\rm e}^-$ collisions, b) NA22 and UA1,
c) KLM~\cite{Bia1991}.}
\end{figure}
\subsection{Factorial Cumulants}
One further has to stress the advantages of normalized factorial
cumulants $K_q$
compared to factorial moments, since the former measure {\it genuine}
correlation patterns.
As an example, high statistics data of the OPAL experiment \cite{Sar00}
are given in Fig.~16 in terms of $K_q$, as a function of the number
$M\propto 1/\delta y$ of phase
space partitions for $q=3$ to 5. In the leftmost column, the one-dimensional
rapidity variable $y$ is used for the analysis. The data (black dots) show an
increase of $K_q$ with increasing $M$ for small $M$, but a saturation at
larger $M$. Even though weaker, some saturation still persists when the
analysis is done in the two-dimensional plane of rapidity $y$ and azimuthal
angle $\varphi$ (middle column), but approximate power-law scaling is indeed
observed for the analysis in three-dimensional momentum space (right column).
Thus, in high-energy collisions, fractal behavior is fully
developed in three dimensions, while projection
effects lead to saturation in lower dimension.
\begin{figure}[htb]
\centerline
\includegraphics[width=9cm]{fig16.pdf}}
\caption{
Cumulants of order $q=3$ to 5 as a function of $M^{1/D}$
in comparison with the predictions of various multiplicity parametrizations
and two Monte Carlo models \cite{Sar00}. }
\end{figure}
In Fig.~16, the data are also compared to a number of parametrizations
of the multiplicity distributions, as well as to the Monte Carlo models
JETSET and HERWIG. One can see that the fluctuations given by the
negative binomial (NB) (dashed line) are weaker than observed in the data.
Contrary to the NB, the log-normal (LN) distribution (dotted line)
overestimates the cumulants, while these expected for a pure birth (PB)
process (dash-dotted) underestimate the data even more significantly than
the NB. Among the distributions shown, a modified NB (MNB) gives the best
results, even though significant underestimation is observed
also there. The Monte Carlo models do surprisingly well.
\subsection{Transverse-momentum dependence}
An interesting question is whether semi-hard effects
\cite{OchWo88-89}, observed to play a role in the transverse-momentum behavior
even at NA22 energies~\cite{Ajin87-2}, or low-$p_\ifmath{{\mathrm{T}}}$
effects ~\cite{VHov89,BialPL89} are
at the origin of intermittency. A first indication for
the latter comes from the most prominent NA22 spike event
(Fig. 13b), where 5 out of 10 tracks in the spike have
$p_\ifmath{{\mathrm{T}}}<0.15$ GeV/$c$.
In Fig.~17, NA22 data \cite{Ajin89-90} on $\ln F_q$ versus $-\ln\delta y$
are given for
particles with transverse momentum $p_\ifmath{{\mathrm{T}}}$ below and above 0.15 GeV/$c$, and
with $p_\ifmath{{\mathrm{T}}}$ below and above 0.3 GeV/$c$. For particles with $p_\ifmath{{\mathrm{T}}}$ below the
cut (left), the $F_q$ exhibit a far stronger $\delta y$ dependence than for
particles with $p_\ifmath{{\mathrm{T}}}$ above the cut (right).
\begin{figure}[b]
\centerline
\includegraphics[width=7cm]{fig17new.pdf}}
\caption{$\ln F_q$ as a function of $-\ln \delta y$ for various $p_\ifmath{{\mathrm{T}}}$
cuts as indicated~\cite{Ajin89-90}.}
\end{figure}
UA1 has a bias against $p_\ifmath{{\mathrm{T}}} < 0.15$ GeV/$c$ and the anomalous
dimension is indeed smaller in UA1 than in NA22 in Fig. 15.
We conclude that intermittency in hh collisions is not dominated
by semi-hard effects.
\subsection{Energy and multiplicity dependence}
As seen in Fig.~18, a strong multiplicity dependence of the intermittency
strength is observed for hh collisions by
UA1~\cite{Alba90}. The trend is opposite to the predictions of the models
used by this collaboration. This decrease of the intermittency strength with
increasing multiplicity is usually explained as a consequence of mixing of
independent sources of particles~\cite{LiBu89}.
\begin{figure}[htb]
\includegraphics[width=6cm]{fig10-25a.pdf}
\hskip-3mm
\includegraphics[width=7cm]{fig10-25b.pdf}
\caption{ a) Multiplicity dependence of the slope
$\phi_3$, compared to that expected from a number of models, the crosses
correspond to a combination of independent events~\cite{Alba90},
b) slope $\phi_2$ extrapolated $(\propto \rho^{-1})$ as a function of particle
density from NA22 (hp at 250 GeV) (solid line) and heavy-ion
collisions as indicated~\cite{Adamo90}.}
\end{figure}
Mixing of emission sources leads to a roughly linear
decrease of the slopes $\phi_q$ with increasing particle density $\langle \rho\rangle$
in rapidity~\cite{CapFia89,BiaFest89,Seib90}: $\phi_q\propto \langle\rho\rangle^{-1}$.
This is indeed observed by UA1~\cite{Alba90}.
Fig. 18a helps in explaining why intermittency is so weak in heavy-ion
collisions (cfr.~Fig.~15): the density (and
mixing of sources) is particularly high there. In Fig.~18b, EMU01
\cite{Adamo90}, therefore, compares $\phi_2$ for NA22 (hp at 250 GeV) and
heavy-ion collisions at similar beam momentum per nucleon, as a function of
the particle density. Whereas slopes averaged over multiplicity are smaller
for AA collisions than for NA22 in Fig.~15, at fixed
$\langle \rho\rangle$ they are actually higher than expected from an extrapolation
of hh collisions to high density and may even grow with
increasing size of the nuclei. The trend is confirmed by KLMM \cite{cher1}
for intermittency in azimuthal angle $\varphi$ and for
slopes up to order 5.
This may be evidence for re-scattering (see~\cite{Verlu90}) or another
(collective) effect, but, as shown by HELIOS
\cite{AAke90} and confirmed by EMU-01~\cite{Adamo90}, one has to be very sure
about the exclusion of $\gamma$-conversions before drawing definite conclusions.
\subsection{Density and correlation integrals}
\begin{figure}[b]
\centerline
\includegraphics[width=9cm]{abcde.pdf}}
\caption{
a) The integration domain $\Omega_\ifmath{{\mathrm{B}}}=\Sigma_m\Omega_m$ of $\rho_2(y_1,y_2)$
for the bin-averaged factorial moments, b) the
corresponding integration domain $\Omega_\ifmath{{\mathrm{S}}}$ for the density integral,
c) illustration of a $q$-tuple in snake
topology, d) GHP topology, e) star topology \cite{Char94}.}
\end{figure}
A fruitful development in the study of density fluctuations is
the density and correlation strip-integral method \cite{Hen83} illustrated in Fig.~19 \cite{Char94}.
By means of
integrals of the inclusive density over a strip domain in $y_1, y_2$ space,
rather than a sum of box domains, one not only avoids unwanted side-effects
such as splitting of density spikes, but also drastically increases the
integration volume (and therefore the statistical significance) at given
resolution.
In terms of the strips (or hyper-tubes for $q>2$), the density
integrals can be evaluated directly from the data after selection
of a proper distance measure, as e.g. the four-momentum difference
$Q^2_{ij} = -(p_i-p_j)^2$,
and after definition of a proper multiparticle topology (snake integral \cite{CaSa89},
GHP integral \cite{Hen83}, star integral \cite{Egger93}).
Similarly, {\it correlation} integrals can be defined by replacing the
density $\rho_q$ in the integral by the correlation function $C_q$.
\begin{figure}[htb]
\centerline
\includegraphics[width=6cm]{fig15.pdf}}
\caption{
Comparison of density integrals for $q=2$ in
their differential form (in intervals $Q^2, Q^2+\ifmath{{\mathrm{d}}} Q^2$)
as a function of $_2\log(1/Q^2)$ for e$^+$e$^-$ (DELPHI)
and hadron-hadron collisions (UA1).\cite{Abreu94}}
\end{figure}
Of particular interest is a comparison of hadron-hadron to e$^+$e$^-$ results
in terms of same and opposite charges of the particles involved. Such a
comparison is
shown in Fig.~20 for $q=2$ \cite{Abreu94}. An important difference between UA1
and DELPHI can be observed in a comparison of the two sub-figures:
For relatively large $Q^2(>0.03$ GeV$^2$), where Bose-Einstein effects do
not play a major role, the e$^+$e$^-$ data increase much faster with
increasing $-_2\log Q^2$ than the hadron-hadron results. For e$^+$e$^-$,
the increase in this $Q^2$ region is very similar for same and for
opposite-sign charges.
At small $Q^2$, however, the e$^+$e$^-$ results approach the hh results.
For e$^+$e$^-$ annihilation at LEP at least
two processes are considered to be responsible for the power-law behavior: Bose-Einstein
correlation at small $Q^2$ following the evolution of jets at larger $Q^2$ , but what is
remarkable is the smooth transition between the two domains (if at all present) (see Sect. 3).
\subsection{Genuine higher-order correlations}
The correlation integral method turns out
particularly useful for the unambiguous establishment of genuine higher-order
correlations in terms of the normalized cumulants
$K_q(Q^2)$, when using the star integration \cite{Egger93}.
Non-zero values of (star integral) $K^*_q(Q^2)$ increasing according to a power
law with decreasing $Q^2$ were first observed in NA22 up to
fifth order \cite{genuine} (see Fig.~21) and in E665 for third order \cite{Adams94}.
Again note the difference between all charged and like charged particles
and the smooth transition betwen larger and smaller $Q^2$.
\begin{figure}[htb]
\centerline
\includegraphics[width=8cm]{fig10-53.pdf}}
\caption{
$\ln K^*_q(Q^2)$ as a function of $-\ln Q^2$ for all charged
particles as well as for like-charged particles \cite{genuine}.}
\end{figure}
\subsection{Functional form}
The exact functional form of $F^\ifmath{{\mathrm{S}}}_2$ is derived from the data of
UA1~\cite{Alba90} and NA22~\cite{Char94}\footnote{In fact in this
form $F^\ifmath{{\mathrm{S}}}_2(Q^2)$ is identical
to $R(Q^2)$ usually used in Bose-Einstein analysis. The only difference is
that it is plotted on a double-logarithmic plot, here.} in Fig.~22.
Clearly, the data favour a power law in $Q$ over an
exponential, double-exponential or Gaussian law.
\begin{figure}[t]
\centerline
\includegraphics[width=6cm]{fig10-43a.pdf}\hskip1mm
\includegraphics[width=6cm]{fig10-43b.pdf}}
\caption{Density integrals $F^\ifmath{{\mathrm{S}}}_2$ (in
their differential form) as a function of $Q^2$ for like-charged
pairs in UA1~\cite{Alba90} and NA22~\cite{Char94},
compared to power-law, exponential,
double-exponential and Gaussian fits, as indicated. }
\end{figure}
If the observed effect is real, it supports a view developed
in~\cite{Bial92}. There, intermittency is explained from
Bose-Einstein correlations between
(like-charged) pions. As such, Bose-Einstein correlations from a static
source\index{source!static} are not power behaved. A power law
is obtained i) if the size of the interaction region is allowed to fluctuate,
and/or ii) if the interaction region itself is assumed to be a
self-similar object extending over a large volume.
Condition ii) would be realized if parton avalanches were to arrange
themselves into self-organized critical states \cite{Bak87}.
Though quite speculative at this moment, it is an
interesting new idea with possibly far-reaching implications.
We should mention also that in such a scheme intermittency is viewed as
a final-state effect and is, therefore, not troubled by
hadronization effects.
So, in conclusion of this section, (approximate) intermittency
is found to be all-present in hadron production and is evidence
for genuine correlations to high orders, but it seems dominated by
Bose-Einstein correlations. However, what we have learned
is that we have been fooled for more than half a century by
an assumed Gaussian behavior of the BE correlations, while
an approximate power law is required. This highly non-trivial
lesson we have learned indeed throws a completely new
light on the topic of femtoscopy.
\section{Bose-Einstein Correlations (or what?)}
\subsection{Early results}
Whether derived as Fourier transform of a (static and chaotic)
pion source distribution, a covariant Wigner-transform of the
(momentum dependent) source density matrix, or from the string model,
identical-pion correlation leads to a positive, non-zero two-particle
correlator $K_2(Q)$, i.e. to
\begin{equation}
R_2(Q)=1+K_2(Q)>1
\end{equation}
at small four-momentum difference $Q$.
These so-called Bose-Einstein Correlations, by now, are a well-established
effect in all types of collisions,
even in hadronic Z$^0$ decay (for reviews see \cite{kitwolf2005,Kit01,Ale03})
originally expected, however, to be too coherent to show an effect.
Other important observations are given in abstract form below.
1. When evaluated in two (or better three) dimensions in the Bertsch-Pratt
system, a small elongation of the emission region (better region of
homogeneity \cite{Sin} is observed along the event axis in all types of
collisions (hadron-hadron \cite{NA22}, all four LEP experiments \cite{LEPEL},
ZEUS \cite{ZEUS}, RHIC \cite{Mag}).
However, it is important to note that the longitudinal radius of
homogeneity is much shorter than the length of the sting (of order 1\%).
The observation that the out-radius does not grow beyond
the side-radius at RHIC \cite{Mag} points
to a short duration of emission and
causes a problem for some hydrodynamical models, but not for
e.g. the Buda-Lund hydro model. The latter, in fact gives a beautifully
consistent description of single-particle spectra and BEC in hadron-hadron
and heavy-ion collisions at SPS and RHIC \cite{Ster}.
The emission function resembles a Gaussian shaped fire-ball for AA
collisions, but a fire-tube for hh collisions.
2. The form of the correlator at small $Q$ is steeper than Gaussian,
in fact consistent with a power law as would be expected from the
intermittency phenomenon described above. Unifying progress is
reported in \cite{Cso}.
3. An approximate $m^{-1/2}_\ifmath{{\mathrm{T}}}$ scaling first observed in heavy-ion
collisions at the SPS \cite{SPS} and usually blamed on collective flow, is
observed at RHIC \cite{PHEN-PHO}, but also in e$^+$e$^-$ collisions \cite{MT}.
Quite generally, it follows from a strong position momentum correlation
\cite{Bia}, be it due to collective flow or to string fragmentation.
4. {\em Genuine three-pion correlations} exist in all types of collisions
and, in principle, allow a phase to be extracted from
\begin{equation}
\cos\phi \equiv \omega(Q_3)= K_3(Q_3)/2\sqrt{K_2(Q_3)} .
\end{equation}
At small $Q$, this $\omega$ is near unity (as expected from incoherence)
for hh \cite{NA22-2} and e$^+$e$^-$ \cite{L3} collisions, as well as for
PbPb \cite{NA44,WA98} and AuAu \cite{STAR03} collisions at SPS and RHIC, while
it is near zero (compatible with full coherence) in collisions of
light nuclei \cite{NA44}. This contradiction can be solved \cite{Kit01,Br}
if $\omega$ is interpreted as a ratio of normalized cumulants.
Since $K^{(N)}_q$ of $N$ independent overlapping sources gets diluted
like $1/N^{q-1}$, $\omega$ would be reduced if strings produced by light ions
do not interact. If, in {\em heavy} ion collisions,
the string density gets high enough for them to coalesce, some kind
of percolation sets in and full inter-string BEC gets restored.
5. Azimuthal anisotropy is observed in configuration space of
non-central heavy-ion collisions at AGS energies \cite{ANIS}, but
also at RHIC \cite{STAR2}. Contrary to elliptic flow, it is directed out
of the event plane, but consistent with the elliptic nuclear
overlap in a non-central collision. Due to larger pressure in the event
plane, the anisotropy gets reduced but not destroyed at RHIC.
Also this is evidence for a short duration of pion emission.
\subsection{The $\tau$ model}
In e$^+$e$^-$, BEC depend, at least approximately, only on $Q$ and not on its components separately, in
the sense that e$^+$e$^-$ BEC is large if Q is small even when any of its components are large.
Further, $R_2$ shows anti-correlations in the region 0.6--1.5 GeV as observed by L3 at LEP \cite{L3-Achard}
as well as by CMS \cite{CMS2011} and ATLAS (preliminary) \cite{Asta} at LHC (see Fig. 23).
A model which predicts such $Q$-dependence, as well as the absence of dependence on the components of $Q$ separately,
is the so-called $\tau$-model~\cite{Tamas;Zimanji:1990}.
Further it incorporates the Bjorken-Gottfried condition \cite{Bialas:1999} whereby
the four-momentum of a produced particle and the space-time position at which it is produced are linearly
related.
\begin{figure}
\begin{minipage}{7cm}
\includegraphics[width=6.9cm]{run_fitQ1mtmt-1.pdf}
\end{minipage}
\hskip-3mm
\begin{minipage}{6cm}\vskip-2mm
\includegraphics[width=6cm]{fig17bb.pdf}\vskip-2mm\hskip-1mm
\includegraphics[width=6cm]{fig17c.pdf}
\end{minipage}
\caption{The Bose-Einstein correlation function $R_2$ for a) L3 \cite{L3-Achard}, b) CMS \cite{CMS2011}
and c) ATLAS \cite{Asta}.
The curve in (a), the dashed line in (b) and the best fit in (c)
correspond to the fit of the $\tau$ model.
The results of the L3 fit are given in Table 1.
Also plotted in (a) is $\Delta$, the difference between the fit and the data.
The dashed line represents the long-range part of the fit, i.e., $\gamma(1+\epsilon Q)$.
The full line in (b) is an exponential fit. The lines in (c) correspond to Gaussian, exponential, and $\tau$ model fits.}
\end{figure}
In this model, it is assumed that the average production point in the overall center-of-mass system,
$\overline{x}=(\overline{t},\overline{r}_\mathrm{x},\overline{r}_\mathrm{y},\overline{r}_\mathrm{z})$,
of particles with a given four-momentum $p=(E,\ensuremath{p_\mathrm{x}},\ensuremath{p_\mathrm{y}},\ensuremath{p_\mathrm{z}})$ is given by
\begin{equation} \label{eq:tau-corr}
\overline{x}^\mu (p^\mu) = a\tau p^\mu \;.
\end{equation}
In the case of two-jet events,
$a=1/m_\ifmath{{\mathrm{T}}}$
where
$m_\ifmath{{\mathrm{T}}}$ is the transverse mass
and
$\tau = \sqrt{\overline{t}^2 - \overline{r}_{\kern -0.14em \mathrm{z}}^2}$ is the longitudinal proper time.
For isotropically distributed particle production, the transverse mass is replaced by the
mass in the definition of $a$ and $\tau$ is the proper time,
$\sqrt{\overline{t}^2 - \overline{r}_{\kern -0.14em \mathrm{x}}^2
- \overline{r}_{\kern -0.14em \mathrm{y}}^2
- \overline{r}_{\kern -0.14em \mathrm{z}}^2}$.
The second assumption is that the distribution of $x^\mu (p^\mu)$ about its average,
$\delta_\Delta ( x^\mu(p^\mu) - \overline{x}^\mu (p^\mu) )$, is narrower than the
proper-time distribution, $H(\tau)$.
Then the two-particle Bose-Einstein correlation function is indeed found to depend on the invariant
relative momentum $Q$, rather than on its separate components, as well as on the values of $a$ of the two particles \cite{ourTauModel}:
\begin{equation} \label{eq:levyR2}
R_2(p_1,p_2) = 1 + \mathrm{Re} \widetilde{H}\left(\frac{a_1 Q^2}{2}\right)
\widetilde{H}\left(\frac{a_2 Q^2}{2}\right) \;,
\end{equation}
where $\widetilde{H}(\omega) = \int \mathrm{d} \tau H(\tau) \exp(i \omega \tau)$
is the Fourier transform (characteristic function) of $H(\tau)$.
(Note that $H(\tau)$ is normalized to unity.)
Since there is no particle production before the onset of the collision,
$H(\tau)$ should be a one-sided distribution.
In the leading log approximation of QCD the parton shower
is a fractal~\cite{Dahlqvist:1989yc}.
Further, a L\'evy distribution arises naturally from a fractal~\cite{metzler}.
One is thus led to choose a one-sided L\'evy distribution for $H(\tau)$~\cite{ourTauModel}.
The characteristic function of $H(\tau)$ can then be written~\cite{Tamas:Levy2004} (for $\alpha\ne1$)
as
\begin{equation} \label{eq:levy1sidecharf}
\widetilde{H}(\omega) = \exp\left[ -\frac{1}{2}\left(\Delta\tau|\omega|\strut\right)^\alpha
\left( 1 - i\, \mathrm{sign}(\omega) \tan\left(\frac{\alpha\pi}{2}\right) \strut \right)
+ i\,\omega\tau_0\right]
\; ,
\end{equation}
where the parameter $\tau_0$ is the proper time of the onset of particle production
and $\Delta \tau$ is a measure of the width of the proper-time distribution.
$0<\alpha<2$ is the so-called index of stability \cite{Nolan:2010} of the L\'evy distribution.
Using this characteristic function in (17),
and incorporating the usual strength factor $\lambda$ and the long-range parametrization,
yields
\begin{equation}
\begin{split}
R_2(Q,a_1,a_2) &= {} \gamma \left\{ 1 +
\lambda\cos\left[\frac{\tau_0 Q^2 (a_1+a_2)}{2} +
\tan\left(\frac{\alpha\pi}{2}\right)\left(\frac{\Delta\tau {Q^2}}{2}\right)^{\!\alpha}\frac{a_1^\alpha+a_2^\alpha}{2} \right]
\right.
\\
&
\left.
\quad \cdot \exp \left[-\left(\frac{\Delta\tau {Q^2}}{2}\right)^{\!\alpha}\frac{a_1^\alpha+a_2^\alpha}{2} \right]
\right\} \left(1+\epsilon Q\right)
\; .
\end{split}
\end{equation}
It is the cosine factor which generates oscillations corresponding to alternating correlated and anti-correlated regions mentioned above.
Note also that since $a=1/m_\ifmath{{\mathrm{T}}}$ for two-jet events, the $\tau$-model\ predicts a decrease of the effective source size with increasing $m_\ifmath{{\mathrm{T}}}$.
\begin{table}[h]
\caption{Results of the fit of (19)\ for two-jet events,
as shown in Fig.~23 a) \cite{L3-Achard}.
The parameter $\tau_0$ is fixed to zero.
The first uncertainty is statistical, the second systematic.}
\begin{center}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{ l r@{$\;\pm\;$}l
}
\hline
parameter & \multicolumn{2}{c }{\ } \\
\hline
$\lambda$ & 0.58 & $0.03^{+0.08}_{-0.24}$ \\
\rule{0pt}{11pt}$\alpha$ & 0.47 & $0.01^{+0.04}_{-0.02}$ \\
$\Delta\tau$ (fm) & 1.56 & $0.12^{+0.32}_{-0.45}$ \\
$\epsilon$ (GeV${-2}$) & 0.001 & $0.001\pm0.003$ \\
$\gamma$ & 0.988 & $0.002^{+0.006}_{-0.002}$ \\
\hline
\chisq/DoF & \multicolumn{2}{c }{90/95} \\
confidence level & \multicolumn{2}{c }{62\%} \\
\hline
\end{tabular}
\end{center}
\end{table}
For each bin in $Q$ the average values of $m_{\mathrm{T}1}$ and $m_{\mathrm{T}2}$ are calculated,
where $m_{\mathrm{T}1}$ and $m_{\mathrm{T}2}$ are the transverse masses of the
two particles making up a pair, requiring $m_{\mathrm{T}1} > m_{\mathrm{T}2}$.
Using these averages, (19) is fit to $R_2(Q)$ by the L3 Coll. \cite{L3-Achard} .
The fit results in $\tau_0=0.00\pm0.02$~fm, and the results of a re-fit with $\tau_0$ fixed to zero
are shown in Table 1.
Note that no significant long-range correlation is observed: $\epsilon=0$ well within one
standard deviation and $\gamma$ is close to unity. Obviously, the $\tau$-model by itself
can reproduce the (smooth) shape of the $Q$-distribution over the full range considered, the anticorrelation near
$Q$=0.6 GeV included.
In the $\tau$-model, the basic assumption is the Bjorken-Gottfried condition \cite{Bialas:1999}
leading to (16). Recently, it has been demonstrated by the same authors \cite{bia;zal2013}, however, that already the
compositeness of pions can most naturally lead to an anti-correlation. At small distances the constituents
mix and there are no separate pions to interfere.
\subsection{The emission function}
The $\tau$ model results for BEC can be used together with the single-particle inclusive
spectra to reconstruct the space-time evolution of hadronization. The emission function in
configuration space, $S_\mathrm{x}(x)$, is the proper time derivative of the
integral over $p$ of $S(x,p)$ \cite{ourTauModel}.
Approximating $\delta_\Delta$ by a Dirac delta function again, gives
\begin{equation} \label{eq:Sspace}
S_\mathrm{x}(x) = \frac{1}{\bar{n}} \frac{\mathrm{d}^4 n}{\mathrm{d}\tau\mathrm{d}^{3}x}
= \left(\frac{m_\ifmath{{\mathrm{T}}}}{\tau}\right)^3 H(\tau) \rho_1\left( p=\frac{m_\ifmath{{\mathrm{T}}} x}{\tau} \right) \;,
\end{equation}
where $n$ and $\bar{n}$ are the number and average number of pions produced, respectively,
and $\rho_1(p)$ is the experimentally measurable single-particle spectrum.
Given the symmetry of two-jet events, $S_\mathrm{x}$ does not depend on the azimuthal angle, and one can
write it in cylindrical coordinates as
\begin{equation} \label{eq:Srzt}
S_\mathrm{x}(r,z,t) = P(r,\eta) H(\tau) \;,
\end{equation}
where $\eta$ is the space-time rapidity.
With the strongly correlated phase-space of the $\tau$-model,
$\eta$ is equal to the momentum-energy rapidity $y$ and $r=\pT\tau/\mT$.
Consequently,
\begin{equation} \label{eq:Preta}
P(r,\eta) = \left(\frac{\mT}{\tau}\right)^{\!3} \rho_\mathrm{\pT,y}(r\mT/\tau, \eta) \;,
\end{equation}
where $\rho_\mathrm{\pT,y}$ is the joint single-particle distribution of \pT\ and $y$.
The reconstruction of $S_\mathrm{x}$ is simplified if $\rho_\mathrm{\pT,y}$
can be factorized into the product of the single-particle \pT\ and rapidity distributions, \ie,
$ \rho_\mathrm{\pT,y} = \rho_\mathrm{\pT}(\pT) \rho_\mathrm{y}(y)$.
Then (22) becomes
\begin{equation} \label{eq:fact}
P(r,\eta) = \left(\frac{\mT}{\tau}\right)^{\!3} \rho_\mathrm{\pT}(r\mT/\tau) \rho_\mathrm{y}(\eta) \;,
\end{equation}
\begin{figure}[htb]
\centerline
\includegraphics[width=8cm]{S-zt-final_v4.pdf}
}
\caption{The temporal-longitudinal part of the emission function normalized to unity \cite{L3-Achard}.}
\end{figure}
\begin{figure}[htb]
\centerline
\includegraphics[width=12cm]{fig11-26.pdf}
}
\caption{
The reconstructed emission function
$S(t,z)$ in arbitrary vertical units, as a function of time $t$ and longitudinal
coordinate $z$ (left diagrams), as well as the reconstructed emission
function $S(x,y)$ in arbitrary vertical units, as a function of the
transverse coordinates $x$ and $y$ (right pictures), for hh
(upper pictures) and PbPb (lower pictures) collisions, respectively \cite{Agab1998,Ster2006,Ster1999}.}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=5cm]{Sxytau-01-small.pdf} \includegraphics[width=5cm]{Sxytau-04-small.pdf} \\
\includegraphics[width=5cm]{Sxytau-07-small.pdf} \includegraphics[width=5cm]{Sxytau-10-small.pdf} \\
\includegraphics[width=5cm]{Sxytau-15-small.pdf} \includegraphics[width=5cm]{plot_rtrans2.pdf}
\caption{The transverse emission function normalized to unity, and its transverse profile for various
proper times \cite{L3-Achard}. An animated gif file covering the first 0.15 fm=0.5$\times$10$^{-24}$ sec is available \cite{novak}.}
\end{figure}
The integral over the transverse distribution is shown
in Fig. 24. It exhibits a ''boomerang'' shape with a maximum at low $t$ and $z$,
but with tails reaching out to very large values of $t$ and $z$, a feature already
observed for hadron-hadron \cite{Agab1998,Ster2006} and heavy ion collisions \cite{Ster1999} (Fig. 25 a) and c))
in the framework of a hydrodynamical model \cite{Csor/9809011}.
The transverse part of the emission function is obtained by integrating over $z$ as well as azimuthal angle.
Fig. 26 shows
the transverse part of the emission function for various proper times. Particle
production starts immediately, increases rapidly and decreases slowly. In the
transverse direction, a ring-like structure is observed similar to the expanding,
ring-like wave created by a pebble in a pond. This ring like structure was also
observed in hadron-hadron collisions \cite{Agab1998} (Fig. 25 b)), where it was
interpreted as due to the production of a fire-ring. Despite this similarity the
physical process is different. Reflecting a non-thermal nature of e$^+$e$^-$ annihilation,
the proper-time distribution and space-time structure are reconstructed here
without any reference to a temperature.
Interpolating and extrapolating Fig. 26, the proper-time dependence of the transverse
expansion of the emission function can be best shown in a movie that ends in
0.15 fm (0.5$\times$10$^{-24}$ sec), making it the shortest movie ever made of a process in
nature \cite{novak}.
In conclusion, I find it absolutely amazing how the combination of experimental results on single particle spectra and two-particle correlations with some theoretical interpretation can allow us to construct a ''femtoscope'' and actually watch particle production at a scale below one fm taking place in less than
10$^{-24}$ sec !
However, one basic puzzle remains: why in the world should pions thought to be produced {\em coherently} in a flux tube
(at least in e$^+$e$^-$ and TeV pp collisions) be subject to {\em incoherent} Bose-Einstein correlations?
Have we fooled ourselves for the past half century or more? Perhaps! What Todorova-Nov\'a is trying
to tell us in a series of papers \cite{todo14} based on the Lund Helix model \cite{anguha98} is just that. Bose-Einstein correlations may not be needed to explain the charge asymmetry of pion pair production, a helix shaped flux tube would not only generate transverse momenta and hadronic masses, but a sharp correlation peak for like charged pion pairs at low values of four-momentum difference $Q$.
I think, it should be a fruitful challenge for younger ones among us to help sort that out in detail in the future.
\vskip 2mm
\underline{Acknowledgements.}
I would like to thank first of all Andrzej, himself,
for almost half a century of direct and indirect encouragement and
guidance in an attempt to understand multihadron dyamics, and I would
like to thank Michal Prasza\l owicz and the organizers of this very special
Symposium for the honor of being invited to contribute.
\newpage
|
\section{Introduction}
Molecular gas observations of galaxies throughout cosmic time are fundamental to understand the cosmic history of formation and evolution of galaxies \citep[see reviews by][]{kennicutt12,carilli13}. The molecular gas provides the fuel for star formation, thus by characterizing its properties we place quantitative constraints on the physical processes that lead to the stellar mass growth of galaxies. This has been a demanding task in terms of telescope time. To date, only a couple hundred sources at $z>1$ have been detected in a molecular gas tracer \citep[typically the rotational transitions of the carbon monoxide $^{12}$CO molecule; e.g.,][]{carilli13}. This sample is dominated by `extreme' sources, such as QSO host galaxies \citep[e.g.,][]{walter03,bertoldi03,weiss07,wang13} or sub-mm galaxies \citep[e.g.,][]{frayer98,neri03,greve05,bothwell13,riechers13,aravena16}, with IR luminosities $L_{\rm IR}\gg 10^{12}$\,L$_\odot${} and Star Formation Rates (SFR) $\gg 100$\,M$_\odot${}\,yr$^{-1}$. These extreme sources might contribute significantly to the star formation budget in the Universe at $z>4$, but their role declines with cosmic time \citep{casey14}. Indeed, the bulk of star formation up to $z\sim2$ is observed in galaxies along the so-called `main sequence' \citep{noeske07, elbaz07, elbaz11, daddi10a,daddi10b, genzel10, wuyts11, whitaker12}, a tight (scatter rms $\sim 0.3$\,dex) relation between the SFR and the stellar mass, $M_*$. Addressing the molecular gas content of main sequence galaxies beyond the local universe has become feasible only in recent years.
The first step in the characterization of the molecular gas content of galaxies is the measure of the molecular gas mass, $M_{\rm H2}$. The $^{12}$CO molecule (hereafter, CO) is the second most abundant molecule in the Universe, and it is relatively easy to target thanks to its bright rotational transitions. The use of CO as a tracer for the molecular gas mass requires assuming a conversion factor, $\alpha_{\rm CO}$, to pass from CO(1-0) luminosities to H$_2$ masses. At $z\sim0$, the conversion factor that is typically used is $\sim4$\,M$_\odot${}(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$ for ``normal'' $M_*>10^9$\,M$_\odot${} star-forming galaxies with metallicities close to solar \citep[see][for a recent review]{bolatto13}. If other CO transitions are observed instead of the J=1$\rightarrow$0 ground state one, a further factor is required to account for the CO excitation \citep[see, e.g.,][]{weiss07,daddi15}. \citet{tacconi10} and \citet{daddi10a} investigated the molecular gas content of highly star-forming galaxies at $z\sim1.2$ and $z\sim2.3$ via the CO(3-2) transition. They found large reservoirs of gas, yielding molecular--to--stellar mass ratios $M_{\rm H2}/M_*\sim1$. These values are significantly higher than those observed in local galaxies \citep[$\sim 0.1$, see e.g.][]{leroy08}, suggesting a strong evolution of $M_{\rm H2}/M_*$ with redshift \citep[see also][]{riechers10, geach11, casey11, magnelli12, aravena12, aravena16, bothwell13, tacconi13, saintonge13, chapman15a, genzel15}.
An alternative approach to estimate gas masses is via dust emission. The dust mass in a galaxy can be retrieved via the study of its rest-frame sub-mm spectral energy distribution (SED) \citep[e.g.,][]{magdis11, magdis12, magnelli12, santini14, bethermin15, berta16}, in particular via the Rayleigh-Jeans tail, which is less sensitive to the dust temperature \citep[see, e.g.,][]{scoville14,groves15}. Using the dust as a proxy of the molecular gas does not require assumptions on CO excitation and on $\alpha_{\rm CO}$. However, this approach relies on the assumption of a dust-to-gas mass ratio (DGR), which typically depends on the gas metallicity \citep{sandstrom13,bolatto13,genzel15}. Recent ALMA results report substantially lower values of $M_{\rm gas}$ than typically obtained in CO--based studies \citep{scoville14,scoville15}.
In the present paper, we study the molecular gas properties of galaxies in ASPECS, the ALMA Spectroscopic Survey in the Hubble Ultra Deep Field (UDF). This is a blind search for CO emission using the Atacama Large Millimeter/sub-millimeter Array (ALMA). The goal is to constrain the molecular gas content of an unbiased sample of galaxies. The targeted region is one of the best studied areas of the sky, with exquisitely deep photometry in $>25$ X-ray--to--far-infrared (IR) bands, photometric redshifts, and dozens of spectroscopic redshifts. This provides us with an exquisite wealth of ancillary data, which is instrumental to put our CO measurements in the context of galaxy properties. Thanks to the deep field nature of our approach, we avoid potential biases related to the pre-selection of targets, and include both detections and non-detections in our analysis. Our dataset combines 3mm and 1mm observations of the same galaxies, thus providing constraints on the CO excitation. Furthermore, the combination of the spectral line survey and the 1mm continuum image allows us to compare CO- and dust-based estimates of the gas mass. In other papers of this series, we present the dataset and the catalog of blindly-selected CO emitters (Paper I, Walter et al.), we study the properties of 1.2mm-detected sources (Paper II, Aravena et al.), we discuss the inferred constraints on the luminosity functions of CO (Paper III, Decarli et al.) and we search for [C\,{\sc ii}]{} emission in $z$=6--8 galaxies (Paper V, Aravena et al.). Paper VI (Bouwens et al.) places our findings in the context of the dust extinction law for $z>2$ galaxies, and Paper VII (Carilli et al.) uses ASPECS to put first direct constraints on intensity mapping experiments. Here we put the CO detections in the context of the properties of the associated galaxies. In sec.~\par\noindent\hangindent 20 pt{sec_obs} we summarize the observational dataset; in sec.~\par\noindent\hangindent 20 pt{sec_sample} we describe our sample; in sec.~\par\noindent\hangindent 20 pt{sec_results} we present CO-based measurements, which are discussed in the context of galaxy properties in sec.~\par\noindent\hangindent 20 pt{sec_discussion}. We summarize our findings in sec.~\par\noindent\hangindent 20 pt{sec_conclusions}.
Throughout the paper we assume a standard $\Lambda$CDM cosmology with $H_0=70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\rm m}=0.3$ and $\Omega_{\Lambda}=0.7$ \citep[broadly consistent with the measurements in][]{planck15}, and a \citet{chabrier03} stellar initial mass function. Magnitudes are expressed in the AB photometric system \citep{oke83}.
\section{Observations}\label{sec_obs}
\subsection{The ALMA dataset}
The details of the ALMA dataset (observations and data reduction) are presented in Paper~I of this series. Here we briefly summarize the observational details that are relevant for the present study. The dataset consists of two frequency scans in ALMA band 3 (3\,mm, 84--115 GHz) and in band~6 (1\,mm, 212--272 GHz). In the case of the 3mm observations, we obtained a single pointing centered at RA = 03:32:37.900, Dec = --27:46:25.00 (J2000.0), close to the Northern corner of the Hubble eXtremely Deep Field \citep[XDF,][]{illingworth13}. The primary beam has a diameter of $\approx 65''$ at the central frequency of the band ($99.5$\,GHz). The typical noise rms is $\sim 0.18$\,mJy\,beam$^{-1}$ per 50\,km\,s$^{-1}${} channel. The 1\,mm observations consist of a 7-pointing mosaic covering approximately the same area as the 3\,mm observations. The typical noise rms is 0.44\,mJy\,beam$^{-1}$ per 50\,km\,s$^{-1}${} channel. The resulting 1mm continuum image reaches a noise rms of $12.7$\,$\mu$Jy\,beam$^{-1}$ at the center of the mosaic (see Paper II).
\subsection{Ancillary Data}
We complement the ALMA data with X-ray--to--far-IR photometry from public catalogs of this field, as well as optical/near-IR spectroscopic information where available. The main sources for the photometry are the compilations by \citet{coe06} and \citet{skelton14}. The former includes optical photometry based on the original {\em HST} Advanced Camera for Survey (ACS) images of the {\em Hubble} Ultra Deep Field \citep[UDF,][]{beckwith06} and near-IR images obtained with {\em HST} NICMOS. The latter compiles also optical/near-IR observations with {\em HST} ACS and Wide Field Camera 3 (WFC3) from the {\em Hubble} XDF \citep{illingworth13}, {\em Spitzer} IRAC, as well as a wealth of ground-based optical/near-IR observations. {\em Spitzer} MIPS data at 24$\mu$m and 70$\mu$m, as well as {\em Herschel} PACS and SPIRE data come from the work by \citet{elbaz11}. X-ray data are taken from the Extended {\em Chandra} Deep Field South Survey \citep{lehmer05} and from the {\em Chandra} Source Catalogue \citep{evans10}.
Photometric redshifts ($z_{\rm photo}$) are available for all of the optically-selected sources in the field. At a limiting magnitude $i$=28\,mag, the median uncertainty is $\delta z_{\rm photo}\sim 0.5$, and it reaches $\delta z_{\rm photo}\sim 1$ at $i\approx 30$ mag \citep{coe06}. The compilation of \citet{skelton14} provides even more robust photometric redshifts, thanks to the expanded photometric dataset. The agreement with available spectroscopic redshifts is typically very good in these cases, with a standard deviation on $\Delta z/(1+z)$ of $\approx0.01$ for \citep{skelton14}. In addition, the 3D--HST survey provides {\em HST} ACS and WFC3 grism observations of the field, yielding grism redshifts for tens of sources in our pointing \citep{momcheva16}. Slit spectroscopy for 74 (mostly bright) galaxies in the field is also available \citep{lefevre05,skelton14,morris15}. Finally, integral field spectroscopy of this field has been secured with ESO VLT/MUSE. These data are part of a Guaranteed Time observing program targeting the UDF. In particular, a single (1 arcmin$^2$) deep (21\,hr on source) pointing overlaps with $\sim70$\% of the ASPECS coverage. The cubes have been processed and analyzed with the improved MUSE GTO pipeline. These observations will be presented in Bacon et al. (in prep).
Within a radius of $34''$ from the ALMA 3mm pointing center (approximately the size of the primary beam of our 3mm observations), there are $1302$\,galaxies from the combination of all the available photometric catalogs. We use the high-$z$ extension of MAGPHYS \citep{dacunha08,dacunha15} to fit the Spectral Energy Distributions (SEDs) of all of them. The input photometry includes 26 broad and medium filters ranging from observed $U$ band to {\em Spitzer} IRAC 8$\mu$m. Additionally, we include the ASPECS 1mm continuum photometry for those sources where $>2$-$\sigma$ emission is reported in our 1mm continuum image. We do not include any {\em Spitzer} MIPS or {\em Herschel} PACS photometry because the angular resolution of those instruments is not sufficient to accurately pinpoint the emission\footnote{For instance, including MIPS and PACS in the SED fits yields to an overestimate (by a factor $\sim3$) of the SFR in the brightest source in our sample, but with a poor SED fit quality, because of the contamination of foreground sources; on the other hand, the second brightest galaxy, which appears isolated, shows consistent results if the fit is performed with or without MIPS and PACS photometry.}. Our MAGPHYS analysis provides us with a posterior probability distribution of the stellar mass ($M_{\rm *}$), the star formation rate, the specific SFR (sSFR=SFR/$M_{\rm *}$), the dust mass ($M_{\rm dust}$) and the IR luminosity for each galaxy in the field. We take the 14\% and 86\% quartiles of the posterior distributions as the uncertainties in the parameters, and we account for an additional fiducial 10\% uncertainty due to systematics (subtleties in the photometric analysis adopted in the input catalogs, such as aperture corrections and deblending assumptions; zero point uncertainties; etc). Fig.~\par\noindent\hangindent 20 pt{fig_ms} shows the SFR as a function of $M_*$ for all the 1302 galaxies in our field.
\begin{figure}\begin{center}
\includegraphics[width=0.99\columnwidth]{fig_ms_co_bright.pdf}
\end{center}
\caption{Star formation rates and stellar masses of all the galaxies in our field, color-coded by redshift. Inferred parameters are derived using the high-$z$ extension of MAGPHYS \citep{dacunha08,dacunha15}. The sample discussed here is highlighted with big symbols: Diamonds refer to the CO detections, while galaxies in the present sample that are not detected in CO are marked with triangles. We stress that only galaxies with secure spectroscopic redshifts are considered in the present analysis. The loci of the main sequence in various redshift bins are shown as dotted lines \citep[from][]{whitaker12} and dashed lines \citep[from][]{schreiber15}. Half of the galaxies in our sample lie along the main sequence at their respective redshifts. ID.5, 7 and 11 occupy the `starburst' region above the main sequence, while ID.3 and 6 exhibit a SFR $\sim3 \times$ lower than what is typically observed in main sequence galaxies at the same redshifts and stellar masses.}\label{fig_ms}
\end{figure}
\begin{figure*}\begin{center}
\includegraphics[width=0.89\textwidth]{brightCO_i1_ps.png}
\includegraphics[width=0.99\columnwidth]{fig_spc_co_bright_1.pdf}
\includegraphics[width=0.99\columnwidth]{fig_sed_co_bright_1.pdf}\\
\end{center}
\caption{{\em Top left:} {\em HST} F105W/F775W/F435W RGB image of ID.1. The postage stamp is $20''\times20''$. {\em Top center:} {\em HST} F125W image of the same field. The map of the lowest-J accessible CO transition (in this case, CO[3-2]) is shown as contours ($\pm 2,3,$\ldots,$20$-$\sigma$ [$\sigma$(ID.1)=0.78\,mJy\,beam$^{-1}$]; solid black lines for the positive isophotes, dashed blue lines for the negative). The synthesized beam is shown as a black ellipse. {\em Top right:} Same as in the center, showing the 1.2mm dust continuum. {\em Bottom left:} Spectra of the CO lines encompassed in our spectral scan. {\em Bottom right:} Spectral Energy Distribution. The red line shows the best MAGPHYS fit of the available photometry (black points), while the blue line shows the corresponding model for the unobscured stellar component. The main output parameters are quoted. Similar plots for all the sources in our sample are available in the appendix.
}\label{fig_id1}
\end{figure*}
\section{The sample}\label{sec_sample}
We focus our discussion on those galaxies in the field covered by ASPECS that we originally expected to detect in CO emission. Our expectations are based on the MAGPHYS predictions discussed in the previous section. Fig.~\par\noindent\hangindent 20 pt{fig_ms} shows the stellar masses and star formation rates of all $1302$ galaxies in the field (color--coded by redshift). Out of these, 56 galaxies have secure spectroscopic redshifts within our CO redshift coverage, and are brighter than 27.5\,mag in the filters F850LW or F105W ($z$ and $Y$ band respectively)\footnote{This flux cut allows us to reject sources that are too faint for a reliable SED analysis.}. We further restrict our analysis to the redshift windows for which ASPECS covers at least one of the following low--J CO transitions: J=2$\rightarrow$1, J=3$\rightarrow$2, or J=4$\rightarrow$3.
From these galaxies, we select the 11 galaxies for which the MAGPHYS SED analysis yields an IR luminosity $L_{\rm IR}>10^{11}$\,L$_\odot${} at $>1$-$\sigma$ significance. These sources are marked by symbols in Fig.~\par\noindent\hangindent 20 pt{fig_ms}, and spectroscopic redshifts are available for all of these sources. The IR luminosity of a galaxy (derived from the SED fitting) has been found to correlate with the CO luminosity (see also Sec.~\par\noindent\hangindent 20 pt{sec_co_lum}). Following the best fit of the relation in \citet{carilli13}, the IR--luminosity cut above corresponds to $L'_{\rm CO(1-0)}>3\times 10^9$\,K~km\,s$^{-1}$\,pc$^2${}, i.e. similar to the line luminosity limit of our survey (see Paper I, Walter et al.\ 2016). Consequently, we should be able to detect CO, or at least place meaningful limits, on these 11 galaxies.
Tab.~\par\noindent\hangindent 20 pt{tab_sample} summarizes the main optical/near-IR properties of the galaxies considered in this paper. In Fig.~\par\noindent\hangindent 20 pt{fig_id1}, we show, for one of the sources, the {\em HST} image compared with the CO and dust continuum maps; the CO spectra; and the SED data and modeling. Similar plots are presented for all sample galaxies in the Appendix.
Four of the galaxies in our sample match some of the CO lines identified in our blind search (see Paper I). The ASPECS name for these sources is also reported in Tab.~\par\noindent\hangindent 20 pt{tab_sample}. Three additional galaxies show CO emission, although at lower significance. Finally, four sources remain undetected in CO.
\begin{figure*}
\includegraphics[width=0.99\textwidth]{fig_muse_dspiral.pdf}\\
\includegraphics[width=0.49\textwidth]{fig_muse_spc3.pdf}
\includegraphics[width=0.49\textwidth]{fig_muse_spc4.pdf}\\
\caption{MUSE optical spectroscopy of the counterparts of ID.3 and ID.4. {\em Top panels ---}(a) {\em HST} RGB image (see Fig.~\par\noindent\hangindent 20 pt{fig_id1} for details). (b) MUSE channel map at 8876 Ang, integrated over $\sim 5$ \AA{}, showing the [O\,{\sc ii}]{} emission of the background component plus the starlight continuum from both the galaxies. The {\em HST}/F160W contours are overplotted in black to guide the eye on the position of the sources. (c) MUSE channel map at 8894 \AA{}, i.e., a few \AA{} off the [O\,{\sc ii}]{} line, showing the continuum emission only. The map is integrated over $\sim10$ \AA{}. (d) Difference between panels (b) and (d). The continuum emission is effectively subtracted, as confirmed by the disappearance of all the field sources. The residual emission is the [O\,{\sc ii}]{} line emission from the background object, which thus resides at $z=1.382$ (consistent with the CO redshift of ASPECS 3mm.3). (e) Same as panel (d), but this time centered at 7780 \AA{}, thus highlighting the [O\,{\sc ii}]{} emission of the foreground galaxy at $z=1.088$. {\em Bottom panels ---} MUSE optical spectra of the [O\,{\sc ii}]{} lines of the counterparts of ASPECS 3mm.3 (left) and 3mm.5 (right). The vertical, dotted lines mark the wavelengths corresponding to the [O\,{\sc ii}]{} doublet at $z$=$z_{\rm CO}$. The gray shading shows the noise in the spectra (which, at these wavelengths, is dominated by sky emission lines). The [O\,{\sc ii}]{} emission is clearly seen in both sources.}\label{fig_muse_dspiral}
\end{figure*}
\subsection{Notes on individual galaxies}
\begin{table*}
\caption{\rm The sample of galaxies examined in this work, and their optical/near-IR global properties. The sorting is based on the significance of the CO detection. (1) Source ID. (2) ASPECS name for blind CO detections (3mm.X, see Paper I) and for the blind 1.2mm continuum detections (CX, see Paper II). (3-4) Optical coordinates in \citet{skelton14}. (5) Redshift. (6) J$_{\rm up}$ of the CO transitions encompassed in our ASPECS data. (7-10) MAGPHYS-derived stellar mass ($M_*$), star formation rate (SFR), specific star formation rate (sSFR), IR luminosity ($L_{\rm IR}$). (11) Effective radius from the near-IR analysis by \citet{vanderwel12}.} \label{tab_sample}
\begin{center}
\begin{tabular}{ccccccccccc}
\hline
ID & ASPECS& Optical RA & Optical Dec & $z$ & Obs.CO & $M_*$ & SFR & sSFR & $L_{\rm IR}$ & $R_{\rm e}$ \\
& name & & & & trans. &[$\times 10^{9}$\,M$_\odot$]& [M$_\odot$\,yr$^{-1}$] & [Gyr$^{-1}$] &[$\times 10^{11}$\,L$_\odot${}]& [kpc] \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) \\
\hline
1 & 3mm.1$^*$,C1 & 03:32:38.54 & -27:46:34.0 & 2.543 & 3,7,8 & $17.8_{-1.7}^{+1.8}$ & $63_{-6}^{+ 6}$ & $3.4_{-0.31}^{+0.34}$ & $12.3_{-1.1}^{+1.2}$ & 1.7 \\
2 & 3mm.2,C2 & 03:32:39.74 & -27:46:11.2 & 1.551 & 2,5,6 & $275_{-40}^{+70}$ & $74_{-30}^{+60}$ & $0.27_{-0.14}^{+0.27}$ & $12.0_{-4.3}^{+8.6}$ & 8.3 \\
3 & 3mm.3 & 03:32:35.55 & -27:46:25.5 & 1.382 & 2,5 & $52_{-10}^{+12}$ & $18_{-7}^{+ 9}$ & $0.42_{-0.25}^{+0.13}$ & $ 1.9_{-0.9}^{+1.3}$ & 8.3 \\
4 & 3mm.5,C6 & 03:32:35.48 & -27:46:26.5 & 1.088 & 2,4 & $28_{-5}^{+ 7}$ & $23_{-9}^{+20}$ & $0.9_{-0.4}^{+0.9}$ & $ 2.8_{-1.2}^{+2.4}$ & 5.8 \\
5 & & 03:32:36.43 & -27:46:31.8 & 1.098 & 2,4 & $5.8_{-0.5}^{+0.6}$ & $44_{-4}^{+ 4}$ & $7.41_{-0.7}^{+0.7}$ & $15.5_{-1.4}^{+1.5}$ & 6.0 \\
6 & C7 & 03:32:35.78 & -27:46:27.5 & 1.094 & 2,4 & $75_{-13}^{+12}$ & $16_{-6}^{+11}$ & $0.21_{-0.08}^{+0.17}$ & $ 3.1_{-1.1}^{+1.5}$ & 3.8 \\
7 & & 03:32:39.08 & -27:46:01.8 & 1.221 & 2,5 & $15.1_{-1.4}^{+1.5}$ & $148_{-13}^{+15}$ & $9.3_{-0.8}^{+0.9}$ & $49.0_{-4.4}^{+4.9}$ & 0.7 \\
8 & & 03:32:36.66 & -27:46:31.0 & 0.999 & 4 & $70_{-17}^{+11}$ & $40_{-9}^{+14}$ & $0.54_{-0.05}^{+0.40}$ & $ 7.1_{-2.5}^{+1.5}$ & 6.6 \\
9 & & 03:32:39.41 & -27:46:22.4 & 2.447 & 3,7,8 & $2.6_{-0.2}^{+ 0.3}$ & $11.8_{-1.1}^{+1.2}$ & $4.2_{-0.4}^{+0.4}$ & $ 1.35_{-0.12}^{+0.13}$ & 5.8 \\
10 & & 03:32:37.07 & -27:46:17.2 & 2.224 & 3,6,7 & $12.0_{-1.2}^{+ 1.2}$ & $22_{-2}^{+41}$ & $1.86_{-0.17}^{+3.53}$ & $ 1.95_{-0.18}^{+5.3}$ & 2.7 \\
11 & & 03:32:36.33 & -27:46:00.1 & 0.895 & 4 & $15.9_{-1.4}^{+ 9.0}$ & $42_{-12}^{+4}$ & $2.7_{-1.5}^{+0.27}$ & $ 5.8_{-1.6}^{+0.6}$ & 1.2 \\
\hline
\end{tabular}
\begin{tablenotes}
\small
\item $^*$ Also 1mm.1 and 1mm.2, see Paper I.
\end{tablenotes}
\end{center}
\end{table*}
{\it ID.1} (Tab.~\par\noindent\hangindent 20 pt{tab_sample}) is a compact galaxy at $z\approx 2.5$. \citet{momcheva16} report a grism redshift $z=2.561$, based on the detection of the [O\,{\sc ii}]{} line in the 3D-HST data. This redshift is improved by our blind detection of three CO transitions (ASPECS 3mm.1, 1mm.1, 1mm.2; see Paper I), clearly pininng down the redshift to $z=2.543$. The {\em HST} images show a blue component in the North and a red component in the South (or possibly a single, relatively blue component, partially reddened in the South by a thick dust lane). A group of bright galaxies is present a few arcsec North of this galaxy, but their spectroscopic redshifts show that the group is in the foreground, with only one other source lying at $z\sim2.5$ (the galaxy $\sim2''$ West of ID.1). The starlight emission coincident with the CO detection is compact, with a scale radius $R_e\approx 1.7$\,kpc \citep{vanderwel12}. Chandra reveals X-ray emission associated with this galaxy. The measured X-ray flux is $F_{\rm X}=5.7\times10^{-17}$\,erg\,s$^{-1}$\,cm$^{-2}$, yielding an X-ray luminosity $L_{\rm X}=3.0\times10^{42}$\,erg\,s$^{-1}$ \citep{xue11}.
{\it ID.2} has a {\em HST} morphology consistent with a large disk galaxy at $z$=1.552. Its slit redshift \citep[$z$=$1.552$,][]{kurk13} matches well our CO line detection (ASPECS 3mm.2), assuming CO(2-1). The disk has an inclination of $\sim 60^\circ$ (based on the aspect ratio, \citealt{vanderwel12}), with an effective radius of $8.3$\,kpc. The galaxy is detected with Chandra. \citet{xue11} report a flux of $F_{\rm X}=3.6\times10^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$ \citep[but $2.6\times10^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$ in][]{lehmer05}, yielding an X-ray luminosity $L_{\rm X}=5.5\times10^{43}$\,erg\,s$^{-1}$, suggesting that ID.2 hosts an AGN.
{\it ID.3} and {\it ID.4} are the two components of an apparent pair of overlapping spiral galaxies. The southern component exibits bright [O\,{\sc ii}]{} emission at $\sim7784$\,\AA{} (see Fig.~\par\noindent\hangindent 20 pt{fig_muse_dspiral}), clearly placing it at $z$=$1.088$ (in agreement with the CO redshift of ASPECS 3mm.5); the northern component shows bright CO emission (ASPECS 3mm.3) which could be interpreted as CO(2-1) at $z$=$1.382$. Our careful analysis of the MUSE data around 8880\,\AA{} reveals faint [O\,{\sc ii}]{} emission (although contaminated by sky line emission), supporting the CO identification (see Fig.~\par\noindent\hangindent 20 pt{fig_muse_dspiral}). The disk of ID.4 has a scale radius of $5.8$\,kpc based on HST imaging \citep{vanderwel12}; for ID.3, the estimated radius is $8.3$\,kpc (but the overlap with the southern component may partially affect this estimate). ID.4 appears as an upper limit in the X--ray catalog by \citet{xue11} ($F_{\rm X}<6.7\times10^{-17}$\,erg\,s$^{-1}$\,cm$^{-2}$, $L_{\rm X}<4.3\times10^{41}$\,erg\,s$^{-1}$).
{\it ID.5} and {\it ID.8} lie in a crowded region of our field. \citet{skelton14} report a spectroscopic redshift $z$=$1.047$ for ID.5. However, the inspection of the MUSE data reveals two, clearly distinguished line sets of the [O\,{\sc ii}]{} doublet, at $z$=$1.038$ and $z$=$1.098$. The latter matches the redshift of two CO lines which are slightly too faint to be selected in our blind search (S/N$\approx$4.8, see Paper I). ID.8 on the other hand is found at another redshift ($z$=$0.999$). No CO emission is found at this position and frequency, although the lowest-J transition that we encompass is CO(4-3) at 1mm. ID.8 is detected in the X--rays \citep{xue11}. Its faintness ($F_{\rm X}=8.2\times 10^{-17}$\,erg\,s$^{-1}$\,cm$^{-2}$, $L_{\rm X}=4.3\times10^{41}$\,erg\,s$^{-1}$) seems consistent with a starburst rather than an AGN \citep{ranalli03}.
{\it ID.6} is located $\sim 4''$ East of ID.3, and probably belongs to a common physical structure (together with other galaxies with a spectroscopic $z\approx1.09$). It is detected in the 1mm continuum, and its CO spectrum shows a $\sim 3$-$\sigma$ excess at the frequency of the expected CO(2-1) line. The CO(4-3) transition is also detected with similar significance, although the best gaussian fit of the line suggests a velocity shift of $\sim200$\,km\,s$^{-1}${} between the two transition. This is likely due to the poor S/N of the two lines.
{\it ID.7} appears as a very compact source ($R_{\rm e}=0.7$\,kpc) at $z$=$1.221$. Its {\em Chandra} image reveals the presence of a bright AGN ($F_{\rm X}=1.01\times10^{-14}$\,erg\,s$^{-1}$\,cm$^{-2}$ in \citealt{lehmer05}; $8.3\times10^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$ in \citealt{evans10}; $6.3\times10^{-15}$\,erg\,s$^{-1}$\,cm$^{-2}$ in \citealt{xue11}), yielding an X-ray luminosity of $L_{\rm X}=5.4\times10^{43}$\,erg\,s$^{-1}$). It is not detected in the 1mm dust continuum. A 3-$\sigma$ excess is measured at the expected frequency of the CO(2-1) transition.
{\it ID.9} and {\it ID.10} are both at $z\sim2.3$. They are among the faintest galaxies in our sample in terms of $L_{\rm IR}$, just above the $10^{11}$\,L$_\odot${} cut. ID.9 appears as a compact bulge. ID.10 appears as a spiral galaxy with disturbed morphology. ASPECS data cover 3 CO transitions in these galaxies: CO(3-2), CO(7-6), and CO(8-7). None of these lines is detected.
{\it ID.11} is a compact ($R_{\rm e}$=$1.2$\,kpc) galaxy at $z=0.895$. As for ID.8, the lowest-J CO transition in the ASPECS coverage is the CO(4-3), which remains undetected.
\section{CO--based measurements}\label{sec_results}
\subsection{CO luminosities and associated H$_2$ masses}\label{sec_MH2}
We measure the line fluxes (or place limits) for all the CO transitions covered in both the 3mm and 1mm line scans. We extract the CO spectra at the position of the optical coordinates of the sources in our sample. We fit the lowest-J transitions accessible with ASPECS data with a gaussian profile; in the case of a detection, we fit the higher-J lines imposing the same line width. We consider detections cases where the flux obtained in the gaussian fit is $>$3$\times$ its uncertainty. If the line is not detected, we assume a fiducial line width of 300\,km\,s$^{-1}${} and we use the upper boundary of the 3-$\sigma$ confidence range on the flux as upper limit. Tab.~\par\noindent\hangindent 20 pt{tab_lines1} reports the CO line fluxes, shifts compared with the nominal redshift, and the line width. The detected sources in our sample have a median CO flux of 0.19\,Jy\,km\,s$^{-1}${} (considering only the lowest-J transition observed in each object). For a comparison, the detected main-sequence galaxies in \citet{tacconi13} have a median CO flux of 0.57\,Jy\,km\,s$^{-1}${}, i.e., $3\times$ higher than the median flux of our detections.
The luminosity of the lowest-J transitions observed in our molecular scans is transformed into the equivalent ground state luminosity $L'_{\rm CO(1-0)}$ using $L'_{\rm CO(J-[J-1])}/r_{J1}$, where we adopt the recent CO excitation ladder of main--sequence galaxies derived by \citet{daddi15}: $r_{21}$=$0.76\pm0.09$, $r_{31}=0.42\pm0.07$, and $r_{41}=0.31\pm0.06$\footnote{\citet{daddi15} do not measure CO(4-3) in the galaxies in their sample. The value of $r_{41}$ adopted here is extrapolated from their measurements of $r_{31}$ and $r_{51}$, in the case of a CO ladder that peaks around J$\approx$5 (see their Fig.~10, left). As uncertainty, we adopt conservative 20\% error.}. The uncertainty in $L'_{\rm CO(1-0)}$ accounts for both the measured flux uncertainty and the standard deviation in the $r_{J1}$ values in the sample studied by \citet{daddi15}. The molecular gas masses are then derived as:
\begin{equation}\label{eq_MH2}
\frac{M_{\rm H2}}{\rm M_\odot} = \frac{\alpha_{\rm CO}}{r_{J1}}\,\frac{L'_{\rm (J-[J-1])}}{\rm K\,km\,s^{-1}\,pc^2}
\end{equation}
We adopt $\alpha_{\rm CO}$=3.6\,M$_\odot$\,(K km\,s$^{-1}${} pc$^{2}$)$^{-1}$ \citep{daddi10b}. This conversion factor has been demonstrated to be appropriate for main--sequence galaxies, through comparisons with dynamical masses \citep{daddi10b}, CO line SED--fitting \citep{daddi15} and detailed dust-SED modeling \citep{genzel15}. In Sec.~\par\noindent\hangindent 20 pt{sec_co_vs_dust} we further discuss the implications of our $\alpha_{\rm CO}$ assumption. Tab.~\par\noindent\hangindent 20 pt{tab_lines2} lists the values of molecular gas masses that we derive for each source. We then combine these measurements or limits on the molecular gas mass with properties of the galaxies inferred from the SED fitting (in particular, the stellar mass $M_*$ and the SFR), to compute the molecular--to--stellar mass ratio $M_{\rm H2}/M_*$ and the depletion time scale $t_{\rm dep}=M_{\rm H2}/{\rm SFR}$ (see Tab.~\par\noindent\hangindent 20 pt{tab_lines2}).
\begin{table}
\caption{\rm CO lines in the galaxies of our sample. (1) Source ID. (2) Upper J of the CO transition. (3) Velocity shift, compared with the redshift quoted in Tab.~\par\noindent\hangindent 20 pt{tab_sample}. (4) Line flux. (5) Line width, expressed as full width at half maximum (FWHM) from the gaussian fit.} \label{tab_lines1}
\begin{center}
\begin{tabular}{ccccc}
\hline
ID & J$_{\rm up}$ & $\Delta v$ & $F_{\rm line}$ & FWHM \\
& & [km\,s$^{-1}$] & [Jy\,km\,s$^{-1}$] & [km\,s$^{-1}${}] \\
(1) & (2) & (3) & (4) & (5) \\
\hline
1 &3&$ -45\pm 8$ & $0.723_{-0.003}^{+0.003}$ & $504\pm12$ \\
1 &7&$ -150\pm120$ & $0.786_{-0.006}^{+0.006}$ & $504^* $ \\
1 &8&$ -45\pm70$ & $1.098_{-0.005}^{+0.005}$ & $504^* $ \\
\hline
2 &2&$ 135\pm 9$ & $0.443_{-0.007}^{+0.007}$ & $538\pm13$ \\
2 &5&$ 135\pm45$ & $0.502_{-0.090}^{+0.090}$ & $538^* $ \\
2 &6&$ -45\pm45$ & $0.820_{-0.100}^{+0.100}$ & $538^* $ \\
\hline
3 &2&$ -37\pm 8$ & $0.135_{-0.003}^{+0.003}$ & $ 57\pm12$ \\
3 &5& --- & $<0.021$ & $ 57^* $ \\
\hline
4 &2&$ 52\pm 7$ & $0.180_{-0.006}^{+0.006}$ & $ 82\pm11$ \\
4 &4& --- & $<0.121$ & $ 82^* $ \\
\hline
5 &2&$ 220\pm 35$ & $0.190_{-0.040}^{+0.040}$ & $352\pm11$ \\
5 &4&$ -28\pm40$ & $0.390_{-0.065}^{+0.065}$ & $352^* $ \\
\hline
6 &2&$ -160\pm 70$ & $0.340_{-0.070}^{+0.060}$ & $530\pm11$ \\
6 &4&$ 230\pm 70$ & $0.370_{-0.090}^{+0.090}$ & $182^* $ \\
\hline
7 &2&$ 150\pm17$ & $0.104_{-0.029}^{+0.019}$ & $150\pm11$ \\
7 &5& --- & $<0.106$ & $150^* $ \\
\hline
8 &4& --- & $<0.059$ & --- \\
\hline
9 &3& --- & $<0.076$ & --- \\
9 &7& --- & $<0.012$ & --- \\
9 &8& --- & $<0.230$ & --- \\
\hline
10 &3& --- & $<0.048$ & --- \\
10 &6& --- & $<0.144$ & --- \\
10 &7& --- & $<0.465$ & --- \\
\hline
11 &4& --- & $<0.015$ & --- \\
\hline
\end{tabular}
\begin{tablenotes}
\small
\item $^*$ Fixed from the fit of a lower J line.
\end{tablenotes}
\end{center}
\end{table}
\begin{table*}
\caption{\rm CO luminosities and CO-based galaxy parameters. (1) Source ID. (2) Redshift. (3) Observed transition. (4) Line luminosity. (5) Equivalent CO(1-0) luminosity, assuming the $r_{J1}$ ratios in \citet{daddi15}. (6) Molecular gas mass, assuming $\alpha_{\rm CO}$=$3.6$ M$_\odot${} (K~km\,s$^{-1}$\,pc$^2$)$^{-1}$. (7) Molecular--to--stellar mass ratio, $M_{\rm H2}/M_*$. (8) Depletion time, $t_{\rm dep}=M_{\rm H2}/{\rm SFR}$.} \label{tab_lines2}
\begin{center}
\begin{tabular}{cccccccc}
\hline
ID & $z$ & J$_{\rm up}$ & $L'$ & $L'_{\rm CO(1-0)}$ & $M_{\rm H2}$ & $M_{\rm H2}/M_{*}$ & $t_{\rm depl}$ \\
& & & [$\times10^9$\,K~km\,s$^{-1}$\,pc$^2$] & [$\times10^9$\,K~km\,s$^{-1}$\,pc$^2$] & [$\times 10^9$ M$_\odot$] & & [Gyr] \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) \\
\hline
1&2.543&3& $24.03_{-0.10}^{+0.10}$ & $57_{-10}^{+10}$ & $206_{-34}^{+34}$ & $12_{-2}^{+2}$ & $3.3_{-0.6}^{+0.7}$ \\
2&1.551&2& $13.71_{-0.27}^{+0.21}$ & $18_{-2}^{+2}$ & $65_{-8}^{+8}$ & $0.24_{-0.05}^{+0.05}$ & $0.9_{-0.4}^{+0.6}$ \\
3&1.382&2& $3.364_{-0.08}^{+0.07}$ & $4.4_{-0.5}^{+0.5}$ & $15.9_{-1.9}^{+1.9}$ & $0.30_{-0.07}^{+0.08}$ & $0.9_{-0.3}^{+0.6}$ \\
4&1.088&2& $2.831_{-0.09}^{+0.09}$ & $3.7_{-0.5}^{+0.5}$ & $13.4_{-1.7}^{+1.7}$ & $0.48_{-0.11}^{+0.13}$ & $0.6_{-0.3}^{+0.4}$ \\
5&1.098&2& $3.089_{-0.66}^{+0.70}$ & $4.1_{-1.0}^{+1.0}$ & $15_{-4}^{+4}$ & $2.5_{-0.6}^{+0.7}$ & $0.33_{-0.09}^{+0.09}$ \\
6&1.094&2& $5.388_{-1.16}^{+0.91}$ & $7.1_{-1.7}^{+1.5}$ & $25_{-6}^{+5}$ & $0.34_{-0.09}^{+0.10}$ & $1.6_{-0.7}^{+0.9}$ \\
7&1.221&2& $2.047_{-0.57}^{+0.37}$ & $2.7_{-0.8}^{+0.6}$ & $10_{-3}^{+2}$ & $0.6_{-0.2}^{+0.16}$ & $0.066_{-0.020}^{+0.016}$ \\
8&0.999&4& $<0.20$ & $<0.63$ & $<2.3$ & $<0.03$ & $<0.06$ \\
9&2.447&3& $<2.4$ & $<5.6$ & $<21$ & $<8$ & $<1.8$ \\
10&2.224&3& $<2.2$ & $<5.3$ & $<19$ & $<1.6$ & $<0.9$ \\
11&0.895&4& $<0.53$ & $<1.7$ & $<6.2$ & $<0.4$ & $<0.15$ \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{figure*}
\includegraphics[width=0.49\textwidth]{fig_pv_mdyn.pdf}
\includegraphics[width=0.49\textwidth]{fig_mdyn_models.pdf}
\caption{{\em Left ---} Position-velocity diagram of the CO(2-1) emission in ID.2, extracted along the major axis of the galaxy. A velocity gradient is apparent. {\em Right ---} Simulated velocity maps of ID.2, assuming that the gas is emitted in a disk geometry and that the CO(2-1) emission traces the mass distribution. The three models refer to different radial scaling of the dynamical mass: (a) $M_{\rm dyn}\propto R$ (thus $v_{\rm rot}$={\it const}); (b) $M_{\rm dyn}\propto R^2$ (i.e., constant surface mass density in the disk; $v_{\rm rot}\propto \sqrt{R}$); (c) $M_{\rm dyn}\propto R^3$ (i.e., constant volume mass density; $v_{\rm rot}\propto R$, i.e., solid rotator). All models assume a dynamical mass $M_{\rm dyn}=2\times10^{11}$\,M$_\odot${} at $R$=8.3\,kpc. The expected line profiles (red histograms) are compared with the observed one (black dots). The flat rotation curve model seems to best reproduce the observed line profile.}\label{fig_mdyn}
\end{figure*}
\subsection{Size of the CO-emitting region in ID.2}\label{sec_mdyn}
In the case of ID.2, our ALMA observations spatially resolve the CO(2-1) emission over $>15$\,kpc, despite the relatively coarse spatial resolution of the 3mm data. A clear velocity gradient is observed in the line emission, as shown in Fig.~\par\noindent\hangindent 20 pt{fig_mdyn}. While the resolution and the signal-to-noise are too poor for an accurate modeling of the gas dynamics, we obtain an estimate of the dynamical mass assuming that the gas is rotating in a disk with the inclination derived from the {\em HST} near-IR imaging (PA=-55$^\circ$, inclination=60$^\circ$ with respect to the line of sight). We then assume a radial distribution of the mass that scales as $M_{\rm dyn}\propto R^\gamma$, where $\gamma$=1 yields the flat rotation curves typically observed in galaxies; $\gamma$=2 implies a constant surface density of mass in the disk, and yields $v_{\rm rot}\propto \sqrt{R}$; and $\gamma$=3 corresponds to a solid rotator ($v_{\rm rot}\propto R$). We then generated mock velocity maps for these three cases, assuming that the CO light traces the mass distribution; and we inferred expected line profiles (see Fig.~\par\noindent\hangindent 20 pt{fig_mdyn}). The $\gamma$=1 case shows the typical ``double-horned'' profile observed in local spiral galaxies. This seems to provide a better description of the observed CO(2-1) line than the other two models, which fail to reproduce the extension of the blue wing of the line. The implied dynamical mass is $M_{\rm dyn}\approx 2\times 10^{11}$\,M$_\odot${} at $R$=8.3\,kpc (= the effective radius). We stress however that this estimate is highly dependent on the model assumptions. A firmer estimate of the dynamical mass in this galaxy requires deeper data at higher spatial resolution.
ID.2 also appears in the SINFONI Integral field spectroscopy survey in the Near-IR \citep[SINS;][]{foerster09} as GMASS-1084 \citep[see also][]{kurk13}. SINS investigated the morphology and kinematics of ionized gas (as traced by the H$\alpha$ Hydrogen line) in a sample of galaxies at $z=1-3.5$. The H$\alpha$ line in ID.2 is emitted on a smaller region (half-light radius $R_{1/2}=3.1\pm1.0$\,kpc) than the CO. The observed H$\alpha$ circular velocity is $67\pm9$\,km\,s$^{-1}${}, which is corrected into 230\,km\,s$^{-1}${} by assuming a low inclination angle ($\sim 20^\circ$). This yields a dynamical mass of $1.2\times 10^{11}$\,M$_\odot${}, roughly consistent with our estimate, especially if one considers that the high level of dust reddening ($A_V=2.4$\,mag from our global MAGPHYS fit) in this source may be responsible of suppressing H$\alpha$ in parts of this galaxy. We note however that the SED fit of this source in \citet{foerster09} yields a stellar mass of only $M_*=3.61_{-0.60}^{+0.34}\times10^{10}$\,M$_\odot${} and a large SFR=$490_{-31}^{+190}$\,M$_\odot${}\,yr$^{-1}$ (i.e., $L_{\rm IR}\approx 5.7\times10^{12}$\,L$_\odot$). This last estimate disagrees with our dust continuum measurements: e.g., assuming a modified black body template with $\beta$=1.6 and $T_{\rm dust}$=25\,K, such a high SFR would imply a dust continuum flux density of 11\,mJy at 1.2mm (observed: $0.22\pm0.02$\,mJy) and of 32\,mJy at 160\,$\mu$m (observed: $6.9\pm0.3$\,mJy).
As seen from Fig.~\par\noindent\hangindent 20 pt{fig_mdyn}, the molecular gas, as traced through CO emission, is extended on scales of $>15$\,kpc ($> 2''$ at $z$=1.552), i.e., comparable to that of the stellar disk. On the other hand, the 1\,mm dust continuum is unresolved at $1.5''\times1.0''$ resolution, i.e., it is significantly smaller than that of the CO (see Fig.~\par\noindent\hangindent 20 pt{fig_id1_a}). This is not an effect of interferometric filtering or sensitivity of the 1mm data. If we convolve the 1mm continuum data to the synthesized beam of the 3mm data, we do not recover the size seen in CO emission. This serves as a cautionary note that CO and dust sizes may not be the same. As a consequence, the masses deduced from these measurements may trace different regions or components in the galaxy \citep[for other examples of mismatch between CO and dust morphology in high-redshift galaxies, see][]{riechers11,hodge15,spilker15}.
This may explain some of the differences between gas mass estimates derived from CO and dust imaging, with the gas masses derived from dust emission being typically smaller than those derived from CO (see Sec.~\par\noindent\hangindent 20 pt{sec_co_vs_dust}): At the observed wavelength (1.2 mm), dust is optically thin (with the only exception of ID.1, all the sources in our sample globally have $\Sigma_{\rm gas}\ll 10^4$\,M$_\odot${}\,pc$^{-2}$, i.e., $N_{\rm H2}\ll 10^{24}$\,cm$^{-2}$; this yields $\tau$[242\,GHz]$\ll$0.1 for solar metallicities, adopting the \citealt{draine84} formalism). The CO low-J emission, on the other hand, is optically thick practically everywhere in galaxies.
\subsection{CO excitation}
As shown in Tab.~\par\noindent\hangindent 20 pt{tab_sample}, ASPECS cover 2--3 different CO transitions in 9 out of 11 galaxies in our sample. Fig.~\par\noindent\hangindent 20 pt{fig_co_excit} shows the inferred constraints on the CO excitation ladder. In ID.1, all three observed transitions [CO(3-2), CO(7-6), and CO(8-7)] were detected in our blind search for line emission (ASPECS 3mm.1, 1mm.1, and 1mm.2, respectively). In ID.2, the CO(2-1) line appears in the results of our blind search (ASPECS 3mm.2). The CO(5-4) and CO(6-5) lines are also observed, but because of their lower significance, they were not detected in our blind search. In particular, the CO(6-5) line is very noisy as it is found at the high-frequency end of the 1mm spectral scan, and it is spatially located at the edge of our mosaic. The CO(2-1) transitions in ID.3 and ID.4 are also identified in our blind search (ASPECS 3mm.3 and 3mm.5, respectively). However, the CO(5-4) line in ID.3 and the CO(4-3) line in ID.4 are not detected. In particular for ID.3, this places very strong limits on the CO excitation of this galaxy, significantly below the average CO ladder of the Milky Way disk (see Fig.~\par\noindent\hangindent 20 pt{fig_co_excit}). In ID.5 and ID.6, we detect both CO(2-1) and CO(4-3). Finally, in ID.7 we only have a tentative detection of CO(2-1), while the CO(5-4) transition remains undetected. No other line is detected in the remainder of our sample.
In Fig.~\par\noindent\hangindent 20 pt{fig_co_excit}, we compare our measurements and limits with the CO excitation templates of the Milky Way disk, and of the starburst in M82 \citep{weiss07}. Additionally, we compare with the average template for high-$z$ main sequence galaxies by \citet{daddi15} and with the theoretical predictions based on the SFR surface density by \citet{narayanan14} (see Tab.~\par\noindent\hangindent 20 pt{tab_sample} and the discussion in sec.~\par\noindent\hangindent 20 pt{sec_ks}). In no case do we find starburst-like CO excitation, comparable with the center of M82 \citep{weiss07} or with what is typically observed in high-$z$ SMGs \citep[e.g.,][]{bothwell13,spilker14}. ID.1 shows a CO(7-6)/CO(3-2) ratio $r_{73}$=$0.2$, consistent with a high-density photon-dominated region \citep{meijerink07}. On the other hand, the CO(8-7) transition appears brighter, implying that a high-excitation component of the ISM might be in place. Interestingly, ID.2 shows a lower excitation (in particular in the CO[5-4]/CO[2-1] ratio, which is consistent with Milky Way excitation). This difference in CO excitation is remarkable if one considers that ID.1 is not detected with Chandra ($L_{\rm X}<6\times10^{42}$\,erg\,s$^{-1}$), whereas ID.2 shows a bright X--ray detection, indicative of the presence of a central AGN. The X-ray emission from the AGN can boost the emission of high-J CO transitions. The CO(7-6)/CO(3-2) ratio $r_{73}$ is typically $0.16-0.63$ in high-density photon-dominated regions powered by star formation (as in ID.2), but it can reach values as high as $r_{73}$=30 in presence of intense X-ray illumination \citep{meijerink07}. This might explain the higher CO excitation observed in high-$z$ QSOs with respect to sub-mm galaxies \citep{carilli13}. The lack of such high excitation feature suggests that the central AGN activity in ID.2 has no major impact on its global CO properties. We attribute the higher excitation in ID.1 to the much more compact emission in this galaxy. As shown in Tab.~\par\noindent\hangindent 20 pt{tab_sample}, ID.1 has a radius that is only $\sim$1/5 of that of ID.2, which translates into a difference in surface area of $\sim$24. Our MAGPHYS-based SFR estimates are comparable ($\sim 70$\,M$_\odot${}\,yr$^{-1}$), thus the surface density of star formation ($\Sigma_{\rm SFR}$) is much higher in ID.1. This is also discussed in Sec.~\par\noindent\hangindent 20 pt{sec_sfr_law} below. The increased radiation field intensity caused by the high star formation rate surface density and/or the higher gas density are likely the reason for the increased CO excitation (see \citealt{narayanan14}).
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_co_bright_excit.pdf}
\caption{CO ladder for the galaxies of our sample detected in CO. Filled symbols mark the transitions detected in our blind search (see Paper I), while empty symbols mark lines that do not match the blind detection requirements. Upper limits, marked with triangles, correspond to 3-$\sigma$ limits. The excitation templates of the Milky Way and M82 are taken from \citet{weiss07}, while the main sequence galaxy template is from \citet{daddi15} (D15). Finally, the theoretical predictions based on the SFR surface density are based on \citet{narayanan14} (NK14). All templates are scaled to match the observed CO flux of the lowest J transition detected in ASPECS. The galaxies in our sample typically show a modest to very low CO excitation. ID.1 (=ASPECS 3mm.1, 1mm.1,2) and ID.5 show slightly higher CO excitation than the template by \citet{daddi15}, although still well below the high-excitation case of the M82 starburst template.}\label{fig_co_excit}
\end{figure}
\section{Discussion}\label{sec_discussion}
In the following we discuss the sources of our sample in the broad context of gas properties in high--redshift galaxies.
\subsection{Location in the galaxy `main sequence' plot}
The stellar masses of the galaxies in our sample range between $(2.8-275)\times 10^{9}$\,M$_\odot${} (two orders of magnitude). The $L_{\rm IR}>10^{11}$\,L$_\odot${} cut in our sample definition selects sources with SFR$>$10\,M$_\odot$\,yr$^{-1}$. The measured SFRs range between 12--150\,M$_\odot${}\,yr$^{-1}$ \footnote{We note that the FAST analysis by \citet{skelton14} yields consistent SFRs for ID.1, ID.3, and ID.10 but different values (by a factor $2\times$ or more) for ID.2 ($6$\,M$_\odot$\,yr$^{-1}$), ID.4 (50\,M$_\odot$\,yr$^{-1}$), ID.5 (21\,M$_\odot$\,yr$^{-1}$), ID.6 ($3.7$\,M$_\odot${}\,yr$^{-1}$), ID.7 (230\,M$_\odot$\,yr$^{-1}$), ID.8 ($0.01$\,M$_\odot$\,yr$^{-1}$), ID.11 ($2.6$\,M$_\odot$\,yr$^{-1}$). No FAST-based SFR estimate is available for ID.9. These differences are likely due to 1) different assumptions on the source redshifts; 2) different coverage of the SED photometry, in particular thanks to the addition of the 1mm continuum constraint in our MAGPHYS analysis; 3) different working assumptions in the two codes. In particular, FAST relies on relatively limited prescriptions for the dust attenuation and star formation history, and does not model the dust emission.}.
Fig.~\par\noindent\hangindent 20 pt{fig_ms} shows the location of our galaxies in the $M_{\rm *}$--SFR (`main sequence') plane. We plot all the galaxies in the field with a F850LP or F160W magnitude brighter than $27.5$\,mag (this cut allows us to remove sources with highly uncertain SED fits). The galaxies in the present sample are highlighted with large symbols.
The different redshifts of the sources are indicated by different colors. As expected from the known evolution of the `main sequence' of star-forming galaxies \citep[e.g.,][]{whitaker12,schreiber15}, sources at higher redshifts tend to have higher SFR per unit stellar mass. Comparing with the {\em Herschel}--based results by \citet{schreiber15}, we find that half of the galaxies in our sample (ID.1, 2, 4, 8, 9, 10) lie on the main sequence (within a factor $3\times$) at their redshift. Three galaxies (ID.5, 7, 11) are above the main sequence (in the `starburst' region), and the remaining two galaxies (ID.3 and ID.6) show a SFR $\sim 3\times$ lower than main sequence galaxies at those redshifts and stellar masses. Similar conclusions are reached if we compare our results with the main sequence fits by \citet{whitaker12} (see Fig.~\par\noindent\hangindent 20 pt{fig_ms}).
\subsection{Star formation law}\label{sec_sfr_law}
The relationship between the total infrared luminosity ($L_{\rm IR}$, a proxy for the star formation rate) and the total CO luminosity ($L'_{\rm CO}$, a proxy for the available gas mass) of galaxy samples is typically referred to as the `integrated Schmidt--Kennicutt' law \citep{schmidt59,kennicutt98,kennicutt12}, or, more generally, the `star formation' law. Sometimes average surface density values are derived from these quantities, resulting in average surface star formation rate densities ($\Sigma_{\rm SFR}$) and gas densities ($\Sigma_{\rm gas}$). We here explore both relations.
\subsubsection{Global star formation law: IR vs.\ CO luminosities}\label{sec_co_lum}
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_co_lum.pdf}
\caption{IR luminosity as a function of the CO(1-0) luminosity for both local galaxies (grey open symbols) and high-redshift sources ($z>1$, grey filled symbols) from the compilation in \citet{carilli13}. The sources in our sample are shown with big symbols, using the same coding as in Fig.~\par\noindent\hangindent 20 pt{fig_ms}. In addition, we also plot the x-axis position of the remaining CO lines found in our 3mm blind search (down-ward triangles; see Paper I). The two parallel sequences of `normal' and `starburst' galaxies \citep{daddi10b,genzel10} are shown as dashed lines (in grey and red, respectively). Our sources cover a wide range of luminosities, both in the CO line and in the IR continuum. Most of the sources in our sample lie along the sequence of `main sequence' galaxies. Four sources lie above the relation: ID.5, which still falls close to the high-$z$ starburst region; ID.7, in which the AGN contamination may lead to an excess of IR luminosity; and ID.8 and ID.11, which are undetected in CO, and that could be shifted towards the relation if one assumes very low CO excitation (as observed in other galaxies of our sample). Conversely, most of the sources detected in CO in our blind search (see Paper I) which lack of an optical/IR counterpart lie significantly below the observed relation.}\label{fig_co_lum}
\end{figure}
In Fig.~\par\noindent\hangindent 20 pt{fig_co_lum} we compare the IR and CO(1-0) luminosities of our sources with respect to a compilation of galaxies both at low and high redshift from the review by \citet{carilli13}, and with the secure blind detections in \citet{decarli14}. For galaxies in our sample that are undetected in CO, we plot the corresponding 3-$\sigma$ limit on the line luminosities. The IR--CO luminosity empirical relation motivates the $L_{\rm IR}$ cut in our sample selection, as galaxies with $L_{\rm IR}>10^{11}$\,L$_\odot${} should have CO emission brighter than $L'\approx 3\times10^9$\,K~km\,s$^{-1}$\,pc$^2${} (i.e., our typical sensitivity limit in ASPECS; see Paper I). {\em All} the galaxies in our sample should therefore be detected in CO.
We find that most of the CO--detected galaxies in our sample lie along the 1-to-1 relation followed by local spiral galaxies as well as color-selected main sequence galaxies at $1<z<3$ \citep{daddi10b,genzel10,genzel15,tacconi13}. Only two galaxies significantly deviate: ID.5, which appears on the upper envelope of the IR--CO relation, close to high-redshift starburst galaxies; and ID.7, which is largely underluminous in CO for its bright IR emission. As discussed in the previous section, these two galaxies appear as starbursts in Fig.~\par\noindent\hangindent 20 pt{fig_ms}. Moreover, ID.7 hosts a bright AGN. If the AGN contamination at optical wavelengths is significant, our MAGPHYS-based SFR estimate is likely in excess (since MAGPHYS would associate some of the AGN light at rest-frame optical and UV wavelengths to a young stellar population), thus explaining the big vertical offset of this galaxy with respect to the `star formation law' shown in Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}. Notably, out of the 4 CO non-detections in our sample, ID.9 and ID.10 are still consistent with the relation, while ID.8 and ID.11 are not. These two galaxies are located at $z=0.999$ and $z=0.895$ respectively. The lowest-J transition sampled in our study is CO(4-3). Their non-detections might be explained if the excitation in these two sources was much lower than what we assumed to infer $L'_{\rm CO(1-0)}$ ($r_{41}=0.31$; see Sec.~\par\noindent\hangindent 20 pt{sec_MH2}).
The sources that are also detected in the blind search for CO (ID.1, 2, 3, 4) tend to lie on the lower `envelope' of the plot. This is expected, as these galaxies have been selected based on their CO luminosity (x--axis).
Fig.~\par\noindent\hangindent 20 pt{fig_co_lum} also shows the x-axis position of the remaining CO blind detections from the 3mm search in Paper I. The CO luminosities of these lines are uncertain (the line identification is ambiguous in many cases, and a fraction of these lines is expected to be a false-positive; see Paper I); however, it is interesting to note that these sources typically populate ranges of line luminosities that were previously unexplored at $z>1$ \citep[see similar examples in][]{chapman08,chapman15b,casey11}, and comparable with or even lower than the typical dust luminosities of local spiral galaxies. We emphasize that a significant fraction of these lines is expected to be real (see Paper I). Deeper data are required to better characterize these candidates.
\subsubsection{Average surface densities: SFR vs.\ gas mass}\label{sec_ks}
We infer average estimates of $\Sigma_{\rm SFR}$ and $\Sigma_{\rm gas}$ by dividing the global SFR and $M_{\rm H2}$ of the galaxy by a fiducial area set by the size of the stellar component, as CO and optical radii are typically comparable \citep{schruba11,tacconi13}. We thus use the information from the stellar morphology derived by \citet{vanderwel12} and reported in Tab.~\par\noindent\hangindent 20 pt{tab_sample} to infer $\Sigma_{\rm SFR}$=SFR/($2\,\pi \, R_e^2$) and $\Sigma_{\rm H2}$=$M_{\rm H2}$/($2\,\pi \, R_e^2$), where $M_{\rm H2}$ is our CO-based measurement of the molecular gas mass, and the factor 2 is due to the fact that the $R_e$ includes only half of the light of the galaxy \citep[see a similar approach in][]{tacconi13}.
In Fig.~\par\noindent\hangindent 20 pt{fig_ks} we show the star-formation law for average surface densities. Global measurements of local spiral galaxies and starbursts are taken from \citet{kennicutt98}, and corrected for the updated SFR calibration following \citet{kennicutt12} and to the $\alpha_{\rm CO}$ value adopted in this paper. We also plot the galaxies in the IRAM Plateau de Bure HIgh-z Blue Sequence Survey \citep[PHIBSS;][]{tacconi13}, again corrected to match the same $\alpha_{\rm CO}$ assumption used in this work, and the secure detections in \citet{decarli14}. Interestingly, the two CO-brightest galaxies in our sample, ID.1 and ID.2 appear to populate opposite extremes of the density ranges observed in high-$z$ galaxies: ID.1 appears very compact, thus reaching the top-right corner of the plot ($\Sigma_{\rm gas}\approx 10000$\,M$_\odot${}\,pc$^{-2}$). On the other hand, in ID.2 the vast gas reservoir is spread over a large area (as apparent in Fig.~\par\noindent\hangindent 20 pt{fig_mdyn}), thus yielding a globally low $\Sigma_{\rm gas}$. We also find that most of the sources in our sample lie along the $t_{\rm depl}\approx 1$\,Gyr line, in agreement with local spiral galaxies and the PHIBSS main sequence galaxies. Only ID.7 and ID.8 lie closer to the $t_{\rm depl}\approx0.1$\,Gyr line. In particular, the offset of ID.7 with respect to the bulk of the sample in the context of the global star-formation law (Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}) is combined here with the very compact size of the emitting region, thus isolating the source in the top-left corner of the plot (see Fig.~\par\noindent\hangindent 20 pt{fig_ks}). Once again, a significant AGN contamination in the estimates of both the rest-frame optical/UV luminosity and in the size of the emitting region could explain such outlier. We also caution that, in some of these galaxies, optical and CO radii might differ.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_ks.pdf}
\caption{The `global' star-formation law relates the average star formation rate surface density ($\Sigma_{\rm SFR}$) with the average gas density in galaxies. Here we consider only the molecular gas phase ($\Sigma_{\rm H2}$). Each point in the plot refers to a different galaxy. We plot the reference samples from \citet{kennicutt98} \citep[corrected for the updated SFR calibration in][]{kennicutt12}, as well as the PHIBSS galaxies from \citet{tacconi13}. Data from the literature have been corrected to match the same $\alpha_{\rm CO}=3.6$\,M$_\odot${}(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$ assumed in this work. The symbol code is the same as in Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}. The galaxies in our sample align along the $t_{\rm depl}\approx 1$\,Gyr, with the only exception of ID.7 and ID.8 which show a short depletion time. It is interesting to note that the two CO-brightest galaxies in our sample, ID.1 and ID.2, populate opposite extremes of the high-$z$ galaxy distribution, with the former being very compact (thus displaying higher SFR and gas densities), and the latter being very extended (thus showing lower SFR and gas densities).}\label{fig_ks}
\end{figure}
\subsection{Depletion times}
Fig.~\par\noindent\hangindent 20 pt{fig_tdepl_ssfr} shows the depletion time, $t_{\rm depl}=M_{\rm H2}/{\rm SFR}$, as a function of the specific star formation rate. This timescale sets how quickly the gas is depleted in a galaxy given the currently observed SFR (ignoring any gas repleneshing). Our data are compared again with the secure blind detections in \citet{decarli14}, with the PHIBSS sample, and with the sample of starburst galaxies studied by \citet{silverman15} (in the latter case, we do not change the adopted value of $\alpha_{\rm CO}=1.1$\,M$_\odot$(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$, as these are not main sequence galaxies). Starburst galaxies tend to reside in the bottom-right corner of the plot (they are highly star-forming given their stellar mass, and they are using up their gaseous reservoir fast). Galaxies with large gas reservoirs and mild star-formation populate the top-left corner of the plot. Since the IR luminosity is proportional to the SFR, and the CO luminosity is used to infer $M_{\rm H2}$, the y-axis of this plot conceptually corresponds to a diagonal line (top-left to bottom-right) in Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}. Also, diagonal lines in Fig.~\par\noindent\hangindent 20 pt{fig_tdepl_ssfr} mark the loci of constant molecular--to--stellar mass ratio $M_{\rm H2}/M_*$.
The sources in our sample range over almost 2 dex in sSFR and $t_{\rm depl}$. Noticeably, ID.1 is highly star-forming (it resides slightly above the main sequence of star forming galaxies at $z\sim 2.5$, see Fig.~\par\noindent\hangindent 20 pt{fig_ms}), so we would expect it to reside in the bottom-right corner of Fig.~\par\noindent\hangindent 20 pt{fig_tdepl_ssfr}; however, its gaseous reservoir is very large for its IR luminosity (see also Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}), thus placing ID.1 in the top-right corner of the plot ($M_{\rm H2}/M_*=12$). On the other hand, ID.2 hosts an enormous reservoir of molecular gas, but because of its even larger stellar mass (yielding low sSFR), it resides on the left side of the plot ($M_{\rm H2}/M_*=0.24$). Their depletion time scales however are comparable (1--3 Gyr). We stress that these results are based on very high-S/N CO line detections, and on very solid descriptions of the galaxy SEDs (see Fig.~\par\noindent\hangindent 20 pt{fig_id1_a}). The sources that populate the starburst region in Fig.~\par\noindent\hangindent 20 pt{fig_ms} and reside in the top-left part of Fig.~\par\noindent\hangindent 20 pt{fig_co_lum} (in particular, ID.5 and ID.7) consistently appear in the bottom-right corner of Fig.~\par\noindent\hangindent 20 pt{fig_tdepl_ssfr}, among starbursts.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_tdepl_ssfr.pdf}
\caption{The depletion time $t_{\rm depl}=M_{\rm H2}/{\rm SFR}$ as a function of the specific star formation rate sSFR=SFR/$M_{\rm *}$ for the galaxies in our sample, the secure blind detections in \citet{decarli14}, the PHIBSS sample by \citet{tacconi13}, and the starburst sample in \citet{silverman15}. The symbol code is the same as in Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}. Starburst galaxies typically reside in the bottom-right corner of the plot. Our ASPECS sources cover a wide range in parameter space, highlighting the diverse properties of these galaxies. }\label{fig_tdepl_ssfr}
\end{figure}
\subsection{Gas to stellar mass ratios}
A useful parameter to investigate the molecular gas content in high-$z$ galaxies is the molecular gas to stellar mass ratio, $M_{\rm H2}/M_{\rm *}$. We prefer this parameter rather than the molecular gas fraction, $f_{\rm gas}=M_{\rm H2}/(M_*+M_{\rm H2})$, as the two involved quantities ($M_{\rm H2}$ and $M_*$) appear independently at the numerator and denominator of the fraction, so that the parameter is well defined even if we only have upper limits on $M_{\rm H2}$. Fig.~\par\noindent\hangindent 20 pt{fig_fgas_z} shows the dependence of $M_{\rm H2}/M_*$ on redshift in the galaxies of our samples, and in galaxies from the literature. This plot informs us on the typical gas content as a function of cosmic time, and can help us shed light on the origin of the cosmic star-formation history (see, e.g., \citealt{geach11}, \citealt{magdis12}, and Paper III of this series). Color-selected star-forming galaxies close to the epoch of galaxy assembly are claimed to show large $M_{\rm H2}/M_*$, with reservoirs of gas as big as (or even larger than) the stellar mass (i.e., $M_{\rm H2}/M_*\sim1$; see, e.g., \citealt{daddi10a}, \citealt{tacconi10,tacconi13}). Indeed, we find examples of very high gas fractions: ID.1 ($M_{\rm H2}/M_*=12$) and the starburst galaxy ID.5 ($M_{\rm H2}/M_*=2.5$) are the most extreme cases. However, it is interesting to note that we also find galaxies with very modest gas fractions, such as ID.2 ($M_{\rm H2}/M_*=0.24$). The CO-detected galaxies at $1.0<z<1.7$ in our sample show an average $M_{\rm H2}/M_*$ ratio that is $\sim 2\times$ lower than the average value for the PHIBSS sample at the same redshift, and closer to the global trend established in \citet{geach11} and \citet{magdis12}. The non-detection of CO in ID.8 places particularly strict limits ($M_{\rm H2}/M_*<0.03$). If the lack of detection is attributed to the very low CO excitation in this galaxy, it would take a 10$\times$ lower $r_{41}$ (i.e., $r_{41}\approx0.03$) to shift ID.8 on the average trend reported by \citet{geach11}.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_fgas_z.png}
\caption{Gas mass fraction (defined as $M_{\rm H2}/M_*$) as a function of redshift from various samples of galaxies in the literature (grey, from the compilation in \citealt{carilli13}), compared with the secure CO detections in \citet{decarli14}, the PHIBSS sample \citep{tacconi13}, the starburst sample in \citet{silverman15}, and our results from this work. The symbol coding is the same as in Fig.~\par\noindent\hangindent 20 pt{fig_ms}. Our data seem to support the picture of a generally increasing $M_{\rm H2}/M_*$ ratio in main sequence galaxies as a function of redshift, as highlighted by the $f_{\rm gas}=0.1 \times (1+z)^2$ green line \citep{geach11,magdis12}. In particular, ID.2 appears as a starburst with respect to its position above the `main sequence' in Fig.~\par\noindent\hangindent 20 pt{fig_ms}, and shows a large $M_{\rm H2}/M_*$ ratio. On the other hand, we also point out that significant upper limits are present (triangles).}\label{fig_fgas_z}
\end{figure}
\subsection{CO vs. Dust-based ISM masses}\label{sec_co_vs_dust}
In addition to the CO line measurements, six of the 11 galaxies in our sample also have detections in the 1\,mm dust continuum. We can thus estimate the mass of the molecular gas independently of the CO data. The Rayleigh-Jeans part of the dust emission is only weakly dependent on the dust temperature, thus it can be used to trace the mass of dust. Using the dust-to-gas scaling \citep[see, e.g.,][]{sandstrom13}, it is possible to infer the gas mass via the dust mass.
\citet{groves15} compare CO-based gas masses with the monochromatic luminosity of the dust continuum in the Rayleigh-Jeans tail. Their analysis relies on a detailed study of 37 local spiral galaxies in the KINGFISH sample \citep{kennicutt11}. The galaxy luminosity in the {\em Herschel}/SPIRE 500$\mu$m band is found to scale almost linearly with the gas mass, yielding:
\begin{equation}\label{eq_groves}
\frac{M_{\rm gas}}{10^{10}\,{\rm M_\odot}}=28.5 \, \frac{\nu L_\nu (500\mu{\rm m})}{10^{10}\,\rm L_\odot}
\end{equation}
We compute the rest-frame luminosity $\nu L_\nu$(500$\mu$m) from the observed 1mm continuum of the galaxies in our sample. For the $k$-correction, we adopt a modified black body with $T_{\rm dust}$=25\,K and $\beta$=$1.6$ \citep[see, e.g.,][]{beelen06}, shifted at the redshift of each source. Since the observing frequency (242\,GHz) falls close to the rest-frame 500$\mu$m (as most of our sources reside at $z\sim1.2$), and we are sampling the Rayleigh-Jeans tail (which is almost insensitive to the dust temperature), the differences in the corrections due the adopted templates are negligible for the purposes of this analysis. The adopted values for the $k$ correction, as well as the resulting gas masses, are listed in Tab.~\par\noindent\hangindent 20 pt{tab_Mism}.
A similar approach was presented by \citet{scoville14,scoville15}. This calibration is tuned on a set of relatively massive [$(0.2-4)\times10^{11}$\,M$_\odot$] star-forming galaxies (30 local star-forming galaxies, 12 low-redshift ULIRGs, and 30 SMGs at $z$=1.4--3.0), all having literature observations of the CO(1-0) transition. The tight relation observed between CO(1-0) luminosity and the rest-frame 850$\mu$m monochromatic luminosity \citep[see Fig.~1 in][]{scoville15} suggests that a simple conversion factor can be used to derive gas masses from monochromatic dust continuum observations. Setting the dust temperature to $T_{\rm dust}=25$\,K \citep[following][]{scoville14}, from eq.~12 in their paper we derive $M_{\rm ISM}$ from our 1\,mm flux densities as follows:
\begin{equation}\label{eq_scoville}
\frac{M_{\rm ISM}}{10^{10}\,{\rm M_\odot}}=\frac{1.20}{(1+z)^{4.8}}\,\,\frac{F_\nu}{\rm mJy} \, \left(\frac{\nu}{\rm 350\,GHz}\right)^{-3.8} \, \frac{\Gamma_0}{\Gamma_{\rm RJ}} \, \left(\frac{D_{\rm L}}{\rm Gpc}\right)^2
\end{equation}
where $F_\nu$ is the observed dust continuum flux density at the observing frequency $\nu$ ($242$\,GHz in our case), $D_{\rm L}$ is the luminosity distance, and $\Gamma_{\rm RJ}$ is a unitless correction factor that accounts for the deviation from the $\nu^2$ scaling of the Rayleigh-Jeans tail. In the reference sample of local galaxies, low-redshift ULIRGs and high-$z$ SMGs that \citet{scoville14} used to calibrate eq.~\par\noindent\hangindent 20 pt{eq_scoville}, $\Gamma_{\rm RJ}=\Gamma_0=0.71$. The resulting ISM masses are listed in Tab.~\par\noindent\hangindent 20 pt{tab_Mism}.
\begin{table*}
\caption{\rm Gas mass estimates based on the dust continuum. Only sources detected at 1mm in ASPECS are considered. (1) Source ID. (2) Redshift. (3) Observed 242\,GHz = 1.2\,mm continuum flux density (see Paper II). (4) $k$ correction, expressed as the ratio between the flux density computed at $\lambda_{\rm rest\,frame}=500$\,$\mu$m and the one at $\lambda_{\rm obs}=1.2$\,mm, assuming a modified black body template for the dust emission with $\beta$=1.6 and $T_{\rm dust}=25$\,K. (5) Gas mass based on the 1mm flux density, derived following eq.~\par\noindent\hangindent 20 pt{eq_groves} \citep{groves15}. (6) Gas mass based on the 1mm flux density, derived following eq.~\par\noindent\hangindent 20 pt{eq_scoville} \citep{scoville14,scoville15}. (7) Gas mass derived from the dust mass estimate resulting from MAGPHYS SED fitting, assuming a dust-to-gas ratio DGR=1/100.} \label{tab_Mism}
\begin{center}
\begin{tabular}{ccccccc}
\hline
ID & $z$ & $F_{\nu}$(1.2mm) & $k$-corr & log\,$M_{\rm gas,\,Groves}$ & log\,$M_{\rm ISM,\,Scoville}$ & log\,$M_{\rm gas,\,MAGPHYS}$ \\
& & [$\mu$Jy] & & [M$_\odot$] & [M$_\odot$] & [M$_\odot$] \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) \\
\hline
1 & 2.543 & $552.7\pm13.8$ & 0.374 & $11.02_{-0.011}^{+0.011}$ & $10.69_{-0.011}^{+0.011}$ & $10.53_{-0.17}^{+0.17}$ \\
2 & 1.551 & $223.1\pm21.6$ & 0.919 & $10.63_{-0.04}^{+0.04}$ & $10.33_{-0.04}^{+0.04}$ & $10.09_{-0.14}^{+0.13}$ \\
4 & 1.088 & $ 96.5\pm24.7$ & 1.665 & $10.24_{-0.10}^{+0.13}$ & $9.95_{-0.13}^{+0.10}$ & $ 9.78_{-0.20}^{+0.18}$ \\
5 & 1.098 & $ 46.4\pm14.9$ & 1.641 & $ 9.93_{-0.12}^{+0.17}$ & $9.63_{-0.17}^{+0.12}$ & $ 9.25_{-0.19}^{+0.14}$ \\
6 & 1.094 & $ 69.6\pm18.9$ & 1.650 & $10.10_{-0.10}^{+0.14}$ & $9.81_{-0.14}^{+0.10}$ & $ 9.58_{-0.18}^{+0.17}$ \\
10 & 2.224 & $ 36.7\pm13.8$ & 0.478 & $ 9.86_{-0.14}^{+0.20}$ & $9.53_{-0.20}^{+0.14}$ & $ 9.25_{-0.20}^{+0.17}$ \\
\hline
\end{tabular}
\end{center}
\end{table*}
Finally, we can infer an estimate of $M_{\rm gas}$ from the estimate of the dust mass, $M_{\rm dust}$, that we obtain via our MAGPHYS fit of the available SED, simply scaled by a fixed dust-to-gas mass ratio (DGR). \citet{sandstrom13} investigate the dust and gas content in a sample of local spiral galaxies, and find DGR$\approx$1/70. \citet{genzel15} and \citet{berta16} perform a detailed analysis of both gas and dust mass estimates in galaxies at $0.9<z<3.2$ observed with {\em Herschel}, and find a lower value of DGR$\approx$1/100, which is the value we adopt here. We stress that there is a factor $>2\times$ scatter in the estimates of DGR due to its dependence on $M_*$ and metallicity \citep{sandstrom13,berta16}. Following the fundamental metallicity relation in \citet{mannucci10}, we estimate that galaxies in our sample typically have solar metallicities (the lowest metallicity estimates are for ID.9: $Z$=$0.6$\,Z$_\odot$; and ID.5: $Z$=$0.7$\,Z$_\odot$), so we do not foresee large intra-sample variations of DGR. For simplicity, in our analysis we thus assume a fixed DGR=1/100. While SED fits are available for all the galaxies in our sample, we consider here only those with a 1mm detection, in order to best anchor the Rayleigh-Jeans tail of the dust emission. The resulting masses are listed in Tab.~\par\noindent\hangindent 20 pt{tab_Mism}.
Fig.~\par\noindent\hangindent 20 pt{fig_mass_comparison} compares the gas estimates based on Eq.~\par\noindent\hangindent 20 pt{eq_groves}, following \citet{groves15}; the ones obtained via Eq.~\par\noindent\hangindent 20 pt{eq_scoville}, following \citet{scoville14}; and the estimates based on dust from the MAGPHYS SED fits, with our CO-based estimate (assuming $\alpha_{\rm CO}=3.6$\,M$_\odot${}[K~km\,s$^{-1}$\,pc$^2$]$^{-1}$). The dust--based gas estimates obtained with different approaches are strongly correlated with each other, as expected because they all scale (almost linearly) with $F_\nu$(1mm). They also correlate well with the CO-based H$_2$ mass estimates over one and a half dex of dynamic range. However, systematic offsets are observed. The \citet{groves15} estimates are on average $1.5\times$ lower than those based on CO. The estimates based on \citet{scoville15} are another $2\times$ lower, and the masses based on MAGPHYS are a factor $1.7\times$ lower than those obtained following \citet{scoville15}.
What causes the discrepancies between these mass estimates? The CO masses might be overestimated because of our assumptions in terms of CO excitation and $\alpha_{\rm CO}$. A higher CO excitation would imply higher $r_{J1}$, thus lower CO(1-0) luminosity (see eq.~\par\noindent\hangindent 20 pt{eq_MH2}). If we assume the M82 excitation template by \citet{weiss07} (see Fig.~\par\noindent\hangindent 20 pt{fig_co_excit}), the inferred $M_{\rm H2}$ masses would be $1.3\times$ lower for ID.2--6, and $2.2\times$ lower for ID.1 and ID.10. This would solve the discrepancy with respect to the estimates based on the Groves recipe, and it would mitigate, but not solve, the discrepancy with the other gas mass estimates. However, such a high CO excitation scenario is ruled out by our 1mm line observations (see Fig.~\par\noindent\hangindent 20 pt{fig_co_excit}). A lower value of $\alpha_{\rm CO}$ could also help. If we adopt the classical value for ULIRGs, $\alpha_{\rm CO}=0.8$\,M$_\odot${}(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$ \citep{bolatto13}, the CO-based gas masses would be a factor 4.5 smaller, thus in good agreement with the ones from the dust. Fig.~\par\noindent\hangindent 20 pt{fig_co_lum} shows that the majority of our sources lie along the relation of main sequence galaxies / local spiral galaxies in the $L_{\rm IR}$--$L'_{\rm CO}$ plot. This is irrespective of the choice of $\alpha_{\rm CO}$. Thorough studies of galaxies along this sequence support our choice for a larger value of $\alpha_{\rm CO}$ \citep[e.g.][]{daddi10a,genzel10,genzel15,sargent14}. Further support to our choice comes from the position of our sources along the `main sequence' of galaxies (Fig.~\par\noindent\hangindent 20 pt{fig_ms}). Among the sources listed in Tab.~\par\noindent\hangindent 20 pt{tab_Mism} and appearing in Fig.~\par\noindent\hangindent 20 pt{fig_mass_comparison}, only ID.5 could be considered a starburst in this respect. Adopting a lower $\alpha_{\rm CO}$ for only this source would lower its molecular gas mass by a factor $\sim 4.5$, thus bringing it close to the bulk of the `main sequence' galaxies in terms of gas fraction (Fig.~\par\noindent\hangindent 20 pt{fig_fgas_z}), but pushing it away from the sequence in the star formation law plot (Fig.~\par\noindent\hangindent 20 pt{fig_ks}). It would also reduce its depletion time scale (Fig.~\par\noindent\hangindent 20 pt{fig_tdepl_ssfr}) and bring the CO-based gas mass closer to the dust-based estimates (Fig.~\par\noindent\hangindent 20 pt{fig_mass_comparison}). Similar considerations could also apply for ID.1, the CO--brightest galaxy in our sample. The compact morphology and the small separation from a companion galaxy, the rising CO emission at high J, the high values of $\Sigma_{\rm SFR}$ and $\Sigma_{\rm H2}$ and the very large $M_{\rm H2}$/$M_*$ all point toward a starburst scenario for this source; however, it is located along the main sequence of galaxies at $z\sim2.5$ in Fig.~\par\noindent\hangindent 20 pt{fig_ms}, and the $L_{\rm IR}$--$L'_{\rm CO(1-0)}$ plot (Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}) shows that this source is located along the sequence of local spirals and main sequence galaxies (not along the sequence of starbursts), irrespective of the choice of $\alpha_{\rm CO}$, even if we assume the extreme case of thermalized CO(3-2) emission in order to derive $L'_{\rm CO(1-0)}$. Because of this, and because of the lack of any starburst signature (justifying a low $\alpha_{\rm CO}$) in all the other sources, the discrepancies between different gas mass estimates shown in Fig.~\par\noindent\hangindent 20 pt{fig_mass_comparison} cannot be mended only by tuning our assumptions on the CO-based mass estimates.
The dust-based gas mass estimates could also be affected by systematic uncertainties. The offset between the estimates based on Eq.~\par\noindent\hangindent 20 pt{eq_groves} and Eq.~\par\noindent\hangindent 20 pt{eq_scoville} suggests a systematic in the calibration of the two recipes. E.g., the luminosity range used in \citet{groves15} to derive Eq.~\par\noindent\hangindent 20 pt{eq_groves} does not cover the $>10^{10}$\,L$_\odot${} range, where our galaxies are found. Eq.~\par\noindent\hangindent 20 pt{eq_scoville}, based on \citet{scoville14}, is pinned down to a longer wavelength than what observed in ASPECS (850\,$\mu$m in the rest-frame, instead of $\sim 500$\,$\mu$m), thus the $k$ correction is significant and dependent on the adopted dust template. In particular, Eq.~\par\noindent\hangindent 20 pt{eq_scoville} explicitly assumes $\beta$=$1.8$, which might not be universally valid (see discussion in Paper II). Our dust SED is only poorly sampled. Most remarkably, the comparison between maps of the CO and dust emission in ID.2 suggests that the gas is optically-thick over a large area, while the dust is not. We might be missing part of the dust emission due to surface brightness limits, thus affecting our estimates of the total ISM mass. In ID.3, the dust continuum emission is not detected at all, despite the bright CO emission. Since we do not detect any significant 1mm continuum associated with the extended disk of ID.2, and no dust emission at all in ID.3, it is hard to assess how big a correction one should consider. It is possible that a similar issue is present in other sources, in particular in galaxies that we see as CO emitters but for which we do not recover any 1mm continuum emission (see, e.g., Fig.~\par\noindent\hangindent 20 pt{fig_co_lum}). Finally, the underlying assumption in the dust--based gas estimates is the dust-to-gas ratio. This can change significantly as a function of metallicity and other parameters in the galaxy \citep[see][for a detailed discussion]{sandstrom13}. A lower value of DGR (e.g., DGR$\sim$1/200) would halve the discrepancy between the MAGPHYS-based estimates and the CO-based ones. While this is a possibility at the low-mass end of Fig.~\par\noindent\hangindent 20 pt{fig_mass_comparison}, we point out that the relatively large stellar mass of the galaxies at the bright end support metallicity values close to solar, thus disfavoring the large DGR values needed to reconcile the two gas mass estimates.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_mass_comparison.pdf}
\caption{Comparison between the H$_2$ masses that we derive from CO for the sources in our sample (x--axis), and the gas masses inferred from the 1mm continuum, following eq.~\par\noindent\hangindent 20 pt{eq_groves} \citep{groves15}, eq.~\par\noindent\hangindent 20 pt{eq_scoville} \citep{scoville14,scoville15}, and based on the MAGPHYS-based estimates of $M_{\rm dust}$, assuming a dust-to-gas ratio of 1/100 \citep{genzel15} (y--axis). The dashed line shows the 1-to-1 case. Only sources with a 1mm continuum detection are shown. The dust-based estimates are correlated with each other, due to the strong dependence on the 1mm continuum emission. The various mass estimates are also correlated with the CO-based ones over 1.5 dex. There are however systematic offsets among the various gas mass recipes, with dust-based masses that appear lower than the ones inferred from CO.}\label{fig_mass_comparison}
\end{figure}
\section{Summary}\label{sec_conclusions}
We present a study of the molecular gas properties as derived from CO observations of high--redshift galaxies in ASPECS, the ALMA Spectroscopic Survey in the {\em Hubble} Ultra Deep Field (UDF). This dataset consists of a blind survey of molecular gas in the ALMA 3mm and 1mm bands targeting a region with the deepest {\em HST} imaging available, the so--called XDF. Our observations cover hundreds of high--redshift galaxies with well--characterized SEDs, i.e. we can test our expectations in terms of molecular gas content in galaxies without any prior selection through their optical or near-IR properties. This allows us to analyse our CO measurements and limits in the context of the global properties of the associated galaxies, thanks to an exquisite wealth of ancillary multi-wavelength information.
We focus on the galaxies for which a secure redshift is available, either via our ASPECS CO observations or from optical/near-IR spectroscopy reported in the literature. In particular, we consider those sources for which our sophisticated fit of the optical--to--mid-IR SED implies high IR luminosity ($L_{\rm IR}>10^{11}$\,L$_\odot${}). These galaxies are expected to be detected in CO based on the empirical relation between CO and dust luminosity. We also restrict our analysis to those galaxies with a redshift such that a CO transition with J$_{\rm up}<5$ is covered in our ASPECS observations.
{\em Success of CO detection ---} Out of 11 sources selected in this way, 4 are also identified in our CO blind search (see Paper I of this series). Three additional galaxies are detected in CO, although with lower significance. The faintest galaxy detected in CO (at $\sim 3$-$\sigma$ level) harbors an AGN. This likely leads to an overestimate of the IR luminosity in our analysis (if the AGN component contributes significantly to the rest-frame optical/UV emission). Finally, four sources remain undetected in CO. In two of them, the lack of CO detection might be attributed to CO excitation, as the lowest J transition that we targeted in these sources is the CO(4-3) line. This however would point toward a very low-excitation scenario for these two sources. The other two undetected galaxies are just above our IR luminosity cut. They reside at relatively high redshift ($z=2.0-2.5$). In these cases, the lack of a detection might be attributed to insufficient depth of our ASPECS observations, and/or modest CO excitation.
{\em CO excitation ---} As we cover CO emission in two separate ALMA bands, we constrain the CO excitation in all of our CO--detected galaxies. In no case do we find evidence of high excitation, as observed in the center of M82 or in IR--luminous SMGs or QSOs at high redshift. The galaxy that has the highest excitation is a bright, compact galaxy, showing high star formation rate surface density (ID.1). We attribute the high CO excitation to either the increased radiation density (due to the locally intense star formation) or to the high density of the gas. A second source (ID.5) shows CO excitation slightly higher than the average `main sequence' galaxy; this is consistent with this galaxy being a starburst, as suggested by other indicators (sSFR with respect to the `main sequence' at that redshift; IR--to--CO luminosity ratio; depletion time; etc). On the other hand, CO excitation is typically very low, often consistent or even lower than Milky Way excitation at least up to J=5. In one case, a $r_{52}$ value as low as $<0.025$ was measured (for a comparison, in the Milky Way we have $r_{52}=0.16$). An X-ray bright AGN with an extended gas reservoir (ID.2) also show modest CO excitation; in this case, any effect that the AGN may have in the center is diminished by the extended molecular gas emission in the disk. Interestingly, also the CO--brightest galaxy in our sample, ID.1, is detected in the X--rays. Its X--ray luminosity is modest however, and roughly consistent with the extrapolation of the SFR--$L_{\rm X}$ relation observed in local starbursts \citep{ranalli03}.
{\em Location with respect to the `main sequence' ---} We discuss our findings in context of previous molecular gas observations at high redshift (star formation law, gas depletion times, gas mass fractions), based on sophisticated SED modeling of their multi--wavelength properties using the high--redshift extension of the MAGPHYS code. Half of the galaxies in our sample reside on the `main sequence' of star-forming galaxies at their redshift. Three galaxies are found in the starburst region (although in one of them the SFR might be over-estimated due to the contamination from an optically-bright AGN). Finally, two sources are found below the main sequence, suggesting that they are more quiescent systems.
{\em The `star formation law' ---} To first order, the CO--detected galaxies in the UDF cover the same parameter space as previous galaxy samples in the $L_{\rm IR}$--$L'_{\rm CO}$ diagram, although they preferentially reside towards the low-IR luminosity envelope of the relation, along the same sequence of local spiral galaxies and close to color-selected galaxies at $z>1$. Only two CO--detected sources lie on the opposite side, closer to the locus of high-$z$ starbursts. Two of the CO non--detections are found to be inconsistent with the $L_{\rm IR}$--$L'_{\rm CO}$ relation, suggesting that CO excitation in these sources must be low. Using {\em HST} imaging to derive the scale radii of the galaxies in our sample, we discuss their location in the `star formation law' diagram: on average, the sources agree with a depletion time of $\sim$1\,Gyr, as found in previous studies, but outliers (up to 1 dex) exist.
{\em Gas fractions ---} With only two remarkable exceptions, the gas fractions observed in our study are slightly lower than what found in targeted observations of main sequence galaxies at similar redshift, but still significantly higher than what typically observed in the local universe.
{\em CO- vs. Dust-based estimates of gas mass ---} In a few cases, we have gas mass estimates derived from CO as well as the dust continuum, via different recipes involving the dust--to--gas ratio. The dust--based estimates are a factor of $\sim$2--5 smaller than those based on CO. This is consistent with recent reports in the literature that dust--based estimates of ISM masses of main sequence galaxies give significantly lower values than using the CO emission. All these methods use a number of assumptions: {\em CO:} extrapolation to a CO(1-0) flux from a higher--J transition + choice of CO--to--H$_2$ conversion factor, {\em dust:} assumption of temperature, dust SED template, optical depth, and dust--to--gas ratio. A larger sample of galaxies with well--defined dust SEDs is required to ultimately decide which gas--mass estimator is preferred.
In summary, accounting for detections as well as non--detections, we find large variations in the molecular gas properties of high--redshift galaxies. This might reflect the large variations in gas content of disk galaxies seen in semi-analytical models (see, e.g., \citealt{lagos11} and Fig.~9 in \citealt{popping14}). Perhaps not unexpectedly, global scaling relations cannot account for the large variations in gas content in individual high redshift galaxies. Our approach through blind frequency scans of well--characterized cosmological deep fields adds additional constraints to the studies of the molecular gas content in distant galaxies, and are thus complementary to dedicated studies of single galaxies that are pre--selected by their optical properties (e.g. SFR and stellar mass). Our study demonstrates that such studies are now feasible, even with early--cycle ALMA observations. Now that ALMA has reached completion, similar studies of larger fields will result in large, statistical samples, which are required to fully understand and beat down systematics and small number statistics. This will provide us with an entirely new approach to characterize the molecular gas content in distant galaxies.
\section*{Acknowledgments}
We thank the anonymous referee for her/his positive feedback and useful comments. RD thanks Laura Zschaechner for insightful discussions. FW, IRS, and RJI acknowledge support through ERC grants COSMIC--DAWN, DUSTYGAL, and COSMICISM, respectively. M.A. acknowledges partial support from FONDECYT through grant 1140099. DR acknowledges support from the National Science Foundation under grant number AST-1614213 to Cornell University. FEB and LI acknowledge Conicyt grants Basal-CATA PFB--06/2007 and Anilo ACT1417. FEB also acknowledge support from FONDECYT Regular 1141218 (FEB), and the Ministry of Economy, Development, and Tourism's Millennium Science Initiative through grant IC120009, awarded to The Millennium Institute of Astrophysics, MAS. IRS also acknowledges support from STFC (ST/L00075X/1) and a Royal Society / Wolfson Merit award. Support for RD and BM was provided by the DFG priority program 1573 `The physics of the interstellar medium'. AK and FB acknowledge support by the Collaborative Research Council 956, sub-project A1, funded by the Deutsche Forschungsgemeinschaft (DFG). PI acknowledges Conict grants Basal-CATA PFB--06/2007 and Anilo ACT1417. RJA was supported by FONDECYT grant number 1151408.
This paper makes use of the following ALMA data: \dataset[ADS/JAO.ALMA\# 2013.1.00146.S and 2013.1.00718.S.]{https://almascience.nrao.edu/aq}. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The 3mm-part of ALMA project had been supported by the German ARC.
|
\section{Introduction}
\IEEEPARstart{S}ocial networking applications, e.g., Facebook, have become one of the most important web services that provide Internet-based platforms for users to interact with other people that are socially-relevant to them. With the advances in location determination technologies, the flourishing of GPS-equipped mobile devices, and the development of wireless Internet connectivity during the last decade; location-based social networks (e.g., Foursquare, Facebook Places, etc) started to emerge. Such LBSNs allow users to share their location information with other people in their social structure \cite{zheng2011location}. In addition, they provide services with spatial relevance to the users such as finding interesting places within a certain geographical area. Moreover, LBSNs provide businesses opportunities for better user reach, including location-based ads and location-based business analytics. This leads to a wide interest in such networks both from academia and industry with companies such as Foursquare reporting nearly 55 million users with over 7 billions check-ins and millions more check-ins every day\footnote{\url{https://foursquare.com/about} (last accessed Sep. 2015.)}.
The main interaction among users in LBSNs is location sharing through the notion of check-in where users voluntarily share their locations with their peers. During the check-in operation, the user is presented with a ranked list of nearby venues to choose her current location. With
the limited screen size of mobile phones, accurate ranking
of location-based query results becomes crucial as the user
would find it hard to scroll beyond the top few results. A number of approaches have been proposed in literature to tackle the venues ranking problem in LBSNs. These approaches either rely on experts to evaluate places, rely on the review of all users that visited these places previously, rank places based on the closest distance to the estimated user location, or based on places popularity \cite{shankar2012crowds,yelp}. Regardless of the ranking algorithm used, places ranking usually depends on the accurate localization of the phone user for better efficiency and accuracy in location queries. However, traditional LBSNs depend on the GPS and/or network-based localization techniques. Consequently, current LBSNs provide reasonable accuracy \textbf{\emph{only}} for outdoor environments or entire buildings.
On the other hand, in indoor environments, GPS is unreliable and the accuracy of cellular-based approaches range from a few hundred meters to kilometers.
Even when WiFi is turned on (e.g., using Google MyLocation), our experiments below show that the median distance error in estimating the actual venue location is 84m, which is still coarse-grained for indoor environments. Such inaccuracy leads to a worse user experience, which in turn is reflected on the accuracy of the collected data and the business value. With the fact that users spend about 89\% of their time indoors \cite{klepeis2001national}, this sparks the need for a new LBSN that can work well in indoor environments.
\setlength{\belowcaptionskip}{-6pt}
Directly extending current LBSNs to use an accurate indoor location determination technique from literature does not solve the problem (as we quantify in the evaluation section) since there are a number of challenges that need to be addressed in order to have a truly indoor LBSN. Specifically, all indoor localization techniques that leverage smartphones sensors, including WiFi, have an average localization error in the range of few meters. This error in localization can lead to placing the user on the other side of the wall in a completely different venue \cite{azizyan2009surroundsense}.
Moreover, as LBSNs are organic systems which are based on users' contribution, their data are susceptible to some noise in the form of incorrect check-ins. These errors lead to problems in venue ranking and labelling. Furthermore, the system needs to be energy-efficient to avoid phone battery drainage. Finally, and \textbf{most importantly}, an indoor LBSN should learn the labels of indoor locations automatically to answer nearest-location queries efficiently and accurately. This cannot be done manually for scalability reasons and due to the inaccuracies of user check-ins and location.
In this paper, we introduce CheckInside: a fine-grained indoor LBSN that combines physical and logical localization techniques to address the above challenges and identify the user actual place accurately. The core idea is to link crowd-sensed data collected from users' smart phones during the check-in operation or opportunistically; with the available venues information retrieved from the traditional LBSNs. When the user performs a check-in operation, multi-modal sensor information (e.g., inaccurate indoor location, opportunistic images and audio samples, etc), are processed by the CheckInside server to construct a sensor-based fingerprint for the current user location. This fingerprint is filtered and matched against different venues fingerprints stored in the CheckInside venues database, which is constructed from the current information in traditional LBSNs and information extracted from previous check-ins in a crowd-sensing approach. The candidate venues achieving the closest matches with the user current place are then returned and displayed as a ranked list to the user. The venue selected by the user, to check-in at, is implicitly used to label the location, update the venues fingerprint database, as well as provide a dynamic feedback on the quality of the different sensors. All sensors used by CheckInside either have a low-energy profile, are already used for other purposes, or are explicitly used by the user. Hence, CheckInside is energy-efficient.
To further address the inherent inaccuracy in indoor localization and fake check-ins taking place outside the actual venue, CheckInside employs a novel outlier detection technique to distinguish fake check-ins from correct ones. This allows CheckInside to determine the true fingerprint of a particular venue, and consequently using only correct check-ins for floorplan semantic labelling and for assessing the weights of the different sensors in the feedback module. In addition, CheckInside can extend the coverage of current LBSNs using a coverage extender module that can predict the names of uncovered venues. \\
We implemented CheckInside on Android phones and evaluated it in four malls with 711 stores over six weeks with 20 different users. Our results show that it can provide the actual venue within the top five list in 99\% of the cases as compared to 17\% only in Foursquare. In addition, CheckInside can accurately detect new venues, increasing the coverage of current LBSN by more than 37\%.
Our main contributions are summarized as follow:
\begin{itemize}
\item We conduct a study to assess the performance of current LBSNs in indoor areas. Our study reveals interesting findings regarding the limitations of current LBSNs in terms of coverage and quality of the ranked venues list (Sec.~\ref{sec:study}).
\item We present the architecture and details of the CheckInside system as a fine-grained indoor LBSN that can address the limitations of the current LBSNs, provide semantic-rich floorplans, as well as increase the venues coverage of current LBSNs (Sec.~\ref{sec:InCheckIn} and \ref{sec:CheckInsidearc}).
\item We implement the CheckInside system and thoroughly evaluate its performance (Sec.~\ref{sec:eval}).
\end{itemize}
Finally, sections~\ref{sec:related}, \ref{sec:disc}, and \ref{sec:conclude} discuss related work, highlight system limitations, and conclude the paper respectively.
\begin{table}
\begin{center}
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{|l|p{6cm}|}
\hline
Category & Sub-categories\\ \cline{1-2}
\bf Food \& Restaurants & restaurant, cafe, dessert shop, ice-cream shop, bakery\\ \cline{1-2}
\bf Clothing \& Fashion &clothing store, accessories store, shoe store, cosmetic store, jewelry store \\ \cline{1-2}
\bf Entertainment \& Arts & cinema, theater, gym, gaming room, pool hall\\ \cline{1-2}
\bf Others &book store, bar, salon, high-tech outlet, grocery store, department store, supermarket\\\cline{1-2}
\end{tabular}
}
\end{center}
\caption{Venues categories.}
\label{tab:cat}
\end{table}
\section{Study of the Limitations of Current LBSNs for Indoor Environments}
\label{sec:study}
To motivate our work, we conducted a study to quantify the limitations of traditional LBSNs for indoor location based services as well as illustrate the characteristics of the check-in data at these LBSNs. We use Foursquare in our study. We investigate the limitations of Foursquare in terms of two main factors: coverage (the number of indoor venues covered by a LBSN to the total number of venues available) and quality of location information (ranking and distance error of the actual venue in the list of nearby venues). Moreover, we quantify how frequently certain stores exist in more than one shopping mall (i.e., chain of stores belonging to the same brand) which can be leveraged to increase LBSNs venues coverage ratio. To perform this study, we developed an Android application that uses the Foursquare API to perform check-ins at venues visited by participants. When a contributor issues a check-in query, the application consults Foursquare to retrieve the list of nearby venues that satisfy the user query. The retrieved list along with the ground truth venue, selected by the contributor or manually entered if the actual venue is not presented in the list, are stored on the phone for later analysis.
We surveyed 711 stores in four different malls by 20 people. Venues covered by the study are divided into four categories as shown in Table~\ref{tab:cat}. We had two modes of operation for WiFi: on and off. When \emph{\textbf{WiFi was turned on}}, as recommended by Foursquare application, this leads to higher localization accuracy as compared to using cellular localization as quantified in the next subsections.
\subsection{Coverage Study}
Coverage refers to the percentage of places that are included in the Foursquare database. Fig.~\ref{fig:studty_all} gives the overall coverage statistics and Table~\ref{tab:coverage} gives the category details. Our study shows that there are three main issues: missed venues, granularity mismatch, and duplicate entries.
First, Foursquare misses about 39\% of venues in the four malls included in the study (not registered at all or registered at a different granularity as discussed below).
Additionally, there is a mismatch between the users' expectations of a place name and the label returned by Foursquare (\textbf{granularity mismatch issue}). For example, for some restaurants, Foursquare reports "food court" as the name of the venue (the actual venue name is not registered in the Foursquare database), which is not expressive enough for participants about their current place, as the food court area contains a large number of venues. This contributes to 4\% of the venues in the study.
Finally, we noticed also that 8\% of the venues were registered more than once with slightly different names. We believe that the reason for this redundancy is that some users failed to find their current venue in the returned list from Foursquare due to inefficiencies in its ranking function and opted to register a new name.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.35]{coverage.pdf}
\caption{Venues coverage in Foursquare. "Granularity mismatch" refers to identifying a venue (e.g., a specific restaurant) by a coarse-grained label (e.g., food court).}
\label{fig:studty_all}
\end{figure}
Moreover, the coverage and the granularity mismatch problems of Foursquare are much worse in non-business buildings such as educational and residential venues. For example, in our university campus, only the university name and the names of buildings are covered (e.g., no lectures halls names or department names).
\begin{table}[!t]
\label{tab:coverage}
\begin{center}
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{ |l| l | l | l |l}
\hline
\bf Category &\parbox[t]{1.2cm}{\#actual \\ venues}& \parbox[t]{1.3cm}{\#covered\\ venues} &\parbox[t]{1.3cm}{\ \% of \\ coverage}\\ \cline{1-4}
\bf Food \& Restaurants & 101 & 91& 90.0\% \\ \cline{1-4}
\bf Clothing \& Fashion & 374& 235 &62.8\% \\ \cline{1-4}
\bf Entertainment \& Arts & 23 & 14 & 60.8\% \\ \cline{1-4}
\bf Others & 213& 96& 45.0\%\\ \hline\hline
\bf Total &711& 436& 61.3\%\\\cline{1-4}
\end{tabular}
}
\end{center}
\caption{Summary of indoor coverage for each category.}
\label{tab:coverage}
\end{table}
\begin{table}[!t]
\begin{minipage}[t]{0.485\linewidth}
\resizebox{1\textwidth}{!}{
\begin{tabular}{|p{2.5cm}|c|}
\hline
\bf Avg. check-ins/venue &122.5 \\ \cline{1-2}
\bf Avg. users/venue &57.2 \\ \cline{1-2}
\bf Avg. tips/venue&5.4 \\ \cline{1-2}
\bf \ \% of check-ins done by the same user/venue&53.4\% \\ \cline{1-2}
\end{tabular}
}
\vspace{0.3cm} \caption{Average check-ins statistics per venue.}
\label{tab:stat}
\end{minipage}
\begin{minipage}[t]{0.485\linewidth}
\tiny
\resizebox{1\textwidth}{!}{
\begin{tabular}{|l|c|}
\hline
Category& Percentage\\\hline
\bf Food \& Restaurants &88.1\% \\\hline
\bf Clothing \& Fashion &87.2\% \\\hline
\bf Arts \& Entertainment&56.5\% \\\hline
\bf Others&73.7\%\\\hline
\hline \bf Overall&82.3\%\\\hline
\end{tabular}}
\vspace{0.3cm}
\caption{Percentage of brand stores in the four malls.}
\label{tab:catcov}
\end{minipage}
\end{table}
The coverage problem differs by category as shown in Table~\ref{tab:coverage}. It is observed that most covered venues are those where users spend a considerable amount of time like food venues. In contrast, other venues like clothing stores have high miss ratios as users in these venues are busy browsing items and may not have enough time to perform check-ins.
Table~\ref{tab:stat} provides statistics from Foursquare check-ins data including: the average number of check-ins performed at each venue, the average number of tips and the average number of users at each venue, and how frequently users repeat check-ins at the same venue. As evident from the table, venues have sufficient number of tips (5.4 in our case), which motivate us to leverage words extracted from user tips to infer the user place. Moreover, given that on average 53.4\% of total check-ins at all venues are repeated by the same users, the user-venue familiarity can be harnessed as a feature in the place inference.
Finally as illustrated in Table~\ref{tab:catcov}, the majority (about 82\%) of shops in the four malls contained in this study are local and international brand venues that have many branches (chain of stores) distributed across different geographical locations. This information is valuable for CheckInside to increase the venues coverage (as we discuss in Sec.~\ref{sec:covext}).
\subsection{Quality Study}
To assess the quality of location information provided by Foursquare in indoor places, our study answers two questions:
(1) What is the average error in distance between the actual venue and the top venue in the ranked list of nearby venues provided by Foursquare? and (2) What is the rank of the actual venue in the list of nearby venues?
To calculate the inter-venues distance, we have used the shortest door-to-door walking distance. Our study comprises two cases of the WiFi connectivity: on and off.
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth,height=2.5cm]{distance}
\caption{Distance error between the actual venue
and first venue in the list.}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth,height=2.5cm]{rank}
\caption{The rank of the actual venue in the list of nearby venues.}
\end{subfigure}
\caption{Quantifying the quality of the default Foursquare ranking function (WiFi on).}
\label{fig:rankdis}
\end{figure}
\textbf{WiFi turned on:}
Fig.~\ref{fig:rankdis}(a) shows that the median distance error is about 84m, which is not suitable for indoor environments. Similarly, for the second question regarding venues ranking, Fig.~\ref{fig:rankdis}(b) shows that more than 47\% of actual venues has a rank that is higher than 30 in the list returned by Foursquare. The actual places that were not provided in the returned list are either not included in the Foursquare databases (74\% of cases) or covered venues that are ranked beyond the default list size of Foursquare (i.e., 30). In addition, we observed from the collected data that there are about 6\% of the reported venues that are outdoor venues (even though the user was indoors).
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth,height=2.5cm]{distancenowifi}
\caption{Distance error between the actual venue
and first venue in the list.}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth,height=2.5cm]{ranknowifi}
\caption{The rank of the actual venue in the list of nearby venues.}
\end{subfigure}
\caption{Quantifying the quality of the default Foursquare ranking function (WiFi off).}
\label{fig:rankdisnowifi}
\end{figure}
\textbf{WiFi turned off:}
Fig.~\ref{fig:rankdisnowifi}(a) shows that the error in distance is larger than 200m in 45\% of cases and even worse it reaches 500m, which happens when the top venue is outside the actual user building. Regarding the venues rank accuracy, Fig.~\ref{fig:rankdisnowifi}(b) shows that about 82\% of actual venues do not appear on the list of nearby venues (contains at most 30 venues) returned by Foursquare. This is due in part to the coarse grained accuracy of cellular based localization. More specifically, we observed that while performing check-ins at a set of neighboring venues (same block of a building), the venues lists returned from Foursquare are very similar. The most prevalent reason is that the phone serving cell (from the cellular service provider) is the same in this block, making it difficult to identify the user location. In addition, the venues that are always returned on the top of the list are the most popular (having the largest number of check-ins) venues in the nearby area. Finally, about 22\% of the reported venues are outdoor venues.
\subsection{Summary of Findings}
In summary, \emph{\textbf{our study highlights}} that a user will find a difficulty in finding her venue in the list and will either add a duplicate venue or not check-in at all, reducing both the system coverage and the user experience as well as missing business opportunities. This means that there is a potential to enhance the venues rankings of LBSNs for better user experience as well as reducing duplicates in the LBSN database. In addition, automatic prediction of uncovered venues names has the potential of increasing venues coverage and reducing the granularity mismatch.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{arch.pdf}
\caption{CheckInside system architecture.}
\label{fig:arch}
\end{figure}
\section{System overview}
\label{sec:InCheckIn}
In this section, we present a typical scenario of how CheckInside works to illustrate the high level flow of information through the system architecture (Fig.~\ref{fig:arch}).
The CheckInside client installed on the user's phone triggers sensor data collection when a user is stable in a certain venue for some time (detected by the \textbf{Fixed Venue Determination} module). The sampled sensors are only those which are enabled according to the data collection policy configured in the \textbf{User Privacy Profile}. Once a user issues a check-in request, the CheckInside client forwards the collected sensors information to the CheckInside cloud server. Sensors used are either low-energy sensors (e.g., inertial sensors), sensors that are already used for other purposes (e.g., cellular information), and/or sensors that are used opportunistically if the user turned them on for other purposes (e.g., WiFi, camera, mic). At the heart of CheckInside is an indoor localization technology. We use the Unloc system~\cite{wang2012no} due to its high accuracy, low-energy consumption, and its reliance only on the phone sensors.
Using the reported phone location, even with a coarse-grained accuracy, the \textbf{Venues Database Manager} contacts traditional LBSNs, e.g., Foursquare, to obtain a list of nearby venues and their associated information (e.g., pictures, user tips, and other check-ins data). These candidate venues are combined with the list of nearby venues already stored in the CheckInside database and the merged list is annotated with the multi-sensor fingerprint of each venue stored in the CheckInside venues database.
The \textbf{Features Extraction Module} creates a test fingerprint of the current user location based on the collected sensors information.\\
The \textbf{Venues Ranking Module} performs a series of accept/reject filtering operations on the returned venues from the \emph{Venues Database Manager} to reduce the candidate set based on the location and WiFi fingerprint by computing the pairwise similarity among fingerprints of the test and candidate places. It then performs a set of ranking operations, based on the different sensors employed, to rank the candidate locations. The different rankings are then aggregated using the \textbf{Rank Aggregation Module} to produce a final ranked list of candidate locations. This list is returned to the user to select the check-in venue.
Once the user selects her current venue, the \textbf{Incorrect Check-ins Detector Module} runs to
handle outliers and noise in the user location check-in operations performed far from the actual venue. After the incorrect check-ins are removed, the \textbf{User Feedback Module} uses the correct check-ins to update the weights of the different ranking modules to enhance the future system performance. Concurrently, the selected user place and the test fingerprint are passed to the \textbf{Semantic Floorplan Labelling Module} to label the venue location on the map.
Finally, if the test place is not suitably matched to any of the candidate venues in the venues database, it will be marked as a new place and the \textbf{Coverage Extender Module} will try to predict its name.
\section{The CheckInside System}
\label{sec:CheckInsidearc}
In this section, we present the details of CheckInside modules depicted in Fig. \ref{fig:arch}. Without loss of generality, we take Foursquare as an example of traditional LBSNs for the rest of the paper.
\subsection{Sensor Sampling Module}
This module is responsible for collecting the sensor features from the user's phone including the accelerometer, microphone, camera, gyroscope, magnetometer, and the received WiFi signal strength values from the available access points. The GPS is also queried with a low duty cycle to detect the user's transition from outdoors to indoors. The sensor measurements are piggy-backed to the cloud server when the user performs a check-in operation.
\subsection{Privacy Controller}
Privacy is an important issue in the design of mobile sensing applications. People are sensitive to data captured by their phone, particularly multimedia data, and how this data is used by the application. Given this fact, CheckInside gives users full control over the sensed data by means of a personalized privacy configuration. Specifically, CheckInside has different modes of operations (full sensors, partial sensors, privacy insensitive data only, etc) that tailor the amount of data collected based on the user's preferences.
There is a trade-off between the performance of the system and privacy. However, according to recent studies \cite{chon2013understanding}, most sensors harnessed by CheckInside (inertial sensors and WiFi) are enabled by most users and even the privacy-sensitive sensors (i.e., camera and microphone) are enabled by about 78\% of users according to the same study. Finally, we process most of the collected sensors data locally on the user's device, further enhancing the user privacy.
\subsection{Energy Consumption}
As sensors sampling (to capture a place fingerprint) needs several seconds, initiating it after the user starts a check-in process will incur a high delay. The phone sensing, therefore need to run in the background to have the place fingerprint ready when the user wants to check-in. However, continuous sensing without duty cycling leads to faster battery depletion. To save the phone battery, we apply the adaptive sensor scheduling scheme \textit{triggered sensing} \cite{mohan2008nericell}. The key idea behind triggered sensing is that sensors that are relatively inexpensive in energy consumption (e.g., accelerometer) is used to trigger the operation of more expensive sensors (e.g., camera and mic). In addition, all sensors are sampled at low rates (compass, accelerometer and gyroscope at 24Hz; WiFi at 1Hz; audio at 32Hz). We quantify the energy consumption of {CheckInside}{} in Sec.~\ref{sec:perf_comp}.
\subsection{Fingerprint Preparation}
This module is responsible for preparing the test fingerprint for the venue the user is currently located at as well as retrieving the fingerprints for candidate venues from the venues database. It consists of three main modules: Fixed Venue Determination, Venues Database Manager and Feature Extraction (green modules in Fig.~\ref{fig:arch}):
\subsubsection{Fixed Venue Determination}
\label{sec:venue_event}
To reduce energy consumption and enhance users' privacy, this module determines if the user is stationary at the same venue for certain amount of time to start data collection. Since the estimated indoor location may have inherent errors that may place the user at the wrong side of a wall, i.e., another venue, we revert to using WiFi similarity for determining the stationarity within a venue, which gives better performance~\cite{wang2012no}.
In particular, the system considers that a user is staying at the same venue if the similarity of consequently received signal strength from WiFi APs is larger than a certain threshold. We experimented with different similarity functions \cite{jaccard1912distribution,park2010growing,azizyan2009surroundsense} and found that a modified version of \cite{azizyan2009surroundsense} gives the best performance. Specifically, given two lists of APs at two locations ($\textrm{APs}_1$) and ($\textrm{APs}_2$), the similarity is given as:
\begin{equation}
S=\frac{1}{|\textrm{APs}_u|}\sum_{a\in \textrm{APs}_u} (f_{1}(a)+f_{2}(a))\dfrac{\min(f_{1}(a),f_{2}(a))}{\max(f_{1}(a),f_{2}(a))}
\label{eq:wifi_sim}
\end{equation}
where $\textrm{APs}_u$ is the union of the MAC addresses of the APs in the two locations,
$f_{1}(a)$ and $ f_{2}(a) $ are the fraction of times each unique MAC address $a$ was
observed over all recordings in the two locations respectively. Note that this metric has the advantage of not depending on the signal strength (which varies by different devices) and, different from \cite{azizyan2009surroundsense}, is normalized to be independent of the number of APs at a particular location (it ranges from 0 and 2).
Once the user is detected to be stationary, sensors data as well as the stay duration are collected. When the user performs a check-in operation, sensor features are piggy-backed with the check-in request to the CheckInside server. Otherwise, if the user leaves the venue without performing a check-in, all venue related data are discarded.
\subsubsection{Venues Database Manager}
This module prepares a list of the candidate venues that will be further filtered out and ranked by the Venues Ranking Module to identify the user location. It first consults the Foursquare database to retrieve the list of nearby venues given the current user location. Other data retrieved from the Foursquare database include the pictures associated with the venue, tips, check-in history, and location\footnote{The venues' location in the Foursquare database are not accurate as they are based on the outdoor GPS location or the network-based location.}. It then stores/updates this data in the CheckInside local database and retrieves the associated multi-sensor fingerprint of the retrieved list as well as the location of the venues as estimated by CheckInside, if the venue already exists in our database. It also builds an index for brand venues (having branches in different buildings) that will be used by the Coverage Extender Module to increase venues coverage.
Since 8\% of the venues
were registered in the Foursquare database more than once with
slightly different names (Sec.~\ref{sec:study}), to mitigate this problem for brand
venues (constituting 82\% of shops in the four malls), we
compare the list of brand names with all names registered on the venue database
based on the edit distance using the Levenshtein algorithm
for the string similarity calculation \cite{guo2014shopprofiler}. The venue name
encountering an edit distance from a brand name less than
a certain threshold is updated to the brand name.
On the other hand,
for the non-brand venues, we compare the names of all
registered venues on the same building against each other
using the edit distance. Names having low edit
distances will be clustered together as representatives of the same venue.
\subsubsection{Feature Extraction Module}
\label{sec:features}
This module extracts the features used to characterize a certain venue to generate the test fingerprint of the location the user is currently at, and is used later by the Venues Ranking Module. Features extracted cover both the user's behavior as well as surrounding environment. Specifically, we use the following features:
\textbf{\textit{Location:} }This is based on the Unloc system \cite{wang2012no} that performs dead-reckoning to provide a rough estimate of the phone location. To reset the dead-reckoning accumulated error, it leverages points in the environment with unique sensors signatures (e.g., elevators, turns, etc). Unloc has the advantages of not requiring any calibration or infrastructure, high accuracy, and low energy consumption.\\
\textbf{\textit{Mobility data:}} This group of features captures users' behavior while visiting different venues as it is a key indicator of place category.
For example, people are stationary for a longer time in restaurants and they mostly visit them during a certain time of day (i.e., meals time). Similarly, users may go to certain shops more frequently based on the season (e.g. ice cream shops in summer).
On the other hand, users are more mobile in clothing shops and there is no fixed pattern for the visiting time of this category. CheckInside uses three mobility features to characterize the nature of venue: (1) the user activity in the venue, (2) the timestamp (time within day)
this type of venue is usually visited, and (3) the time the user spends in this venue.\\
The first feature, the user activity, is define as the ratio ($r$) between the user mobility time to user stationary time within a certain period. User mobility is retrieved from Android activity recognition service, which can distinguish whether the user is stationary or moving. This is quantized into three levels: stationary (e.g., sitting in a restaurant, if $r <0.2$), browsing (e.g., in a clothing shop, if $0.2<r\le 2$), and walking (e.g., in a grocery store, if $r> 2$)\cite{lu2009soundsense}.\\
On the other hand, visiting time is quantized into different periods within the day: early morning, late morning, early afternoon, late afternoon, early evening, and late evening\footnote{generalization to other granualities, e.g. over a week or a year, is left to future work.}.
Finally, stay duration is quantized into 30 minutes intervals. The fingerprint associated with the three mobility features is the histogram of the feature samples collected at this particular venue from different check-ins.\\
\textbf{\textit{ WiFi Fingerprint:}}
Due to the limited range of WiFi in indoor environments, it can be used to characterize venues indoors. CheckInside stores the fraction of times each unique MAC address was observed in the venue over all check-ins as the WiFi fingerprint for that venue. \\
\textbf{\textit{Sound Fingerprint:}}
Sound captured by a mobile device's microphone is a rich source of information that can be used to make accurate inference about the surrounding environment. For example, some venues (e.g., a music store) play music in the background while others (e.g., a library) are quieter. To recognize venues using ambient sound, CheckInside fingerprinting is based on the signal amplitude to capture the loudness of the sound \cite{azizyan2009surroundsense}. Specifically, the amplitude is divided into 100 equal intervals and the number of samples per-interval is normalized by the total number of samples in the recording. The 100 normalized values are considered to be features of the ambient environment. Since sound from the same venue can vary over time, we divide the day into 24 1-hour bins and use a separate sound fingerprint for each bin.\\
\textbf{\textit{Image Fingerprint:}}
There are many features used in literature to represent images including the Scale-Invariant Features Transform (SIFT) \cite{lowe2004distinctive} and the gist features \cite{oliva2001modeling} which captures local and scene features in images respectively. While these features capture the essential characteristics of images, they are not directly appropriate for our system due to their large size. For instance, each SIFT feature is a 128 dimensional vector and there are several hundred of such SIFT vectors for an image. The large size makes it inefficient for image matching, which is not suitable for the real-time operation required by CheckInside.
To resolve this problem, we leverage the visterms compact features \cite{yan2010crowdsearch} which reduce the size of the SIFT features significantly by efficient clustering. A visterm is treated as a term in a document (image in our case) which has an Inverse Document Frequency (IDF) to indicate its discriminative power. Once visterms are extracted from an image, they can be matched efficiently against visterms extracted from the images retrieved from the venues database in a manner similar to keywords in text retrieval.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{pictures2.pdf}
\caption{Pictures taken at different stores: The top left group shows example pictures used to differentiate stores (image features). The top right and bottom left groups show pictures taken at different venues with different light intensities and different floor types respectively (color/light features). Finally, the bottom right group shows some blurred images in our collection.}\label{fig:pictures}
\end{figure}
\textbf{\textit{Color/Light Fingerprint:}}
A large number of stores have a thematic color as part of their decoration, e.g., red at McDonalds.
The wall and floor colors contribute significantly to this theme (Fig.~\ref{fig:pictures}). Floors may be covered with carpets, ceramic tiles, or wooden strips, all of which are discriminating
attributes of the ambiance. Based on this, pictures taken from different spots in a store are likely to reflect this theme.
CheckInside extracts dominant colors and light intensity from pictures of floors and walls by transforming the pixels of the floor images from the RGB space to the hue-saturation-lightness (HSL) space. This has the advantages of removing the effect of shadows of objects and people, and the reflections of light; and decoupling the floor and wall colors from the ambient light intensity \cite{azizyan2009surroundsense}.
We run the K-means clustering algorithm on the HSL image representation of all pictures taken at the same
venue. The K-means algorithm divides the pixels into $K$ clusters, such that the sum of distances from all pixels to their centroid is minimized. The centroids of these clusters, as well
as the cluster sizes, together form the color/light fingerprint of that venue.\\
\textbf{\textit{Magnetic Fingerprint:}}
The natural magnetic field has two characteristics: the uniqueness of magnetic field distortion from one location to another in a building, and its time invariance. This enables the deployment of magnetic-field distortion-based location estimation. The ambient geomagnetic fingerprint can be modeled as a vector M of three components $m_{x}, m_{y}$, and $m_{z}$ which represent the measured magnetic field in the three directions x, y and z. The collected magnetic readings are normalized using mean normalization, and the magnitude of each reading is computed \cite{galvan2014magnetic}. A comparison between the magnitude of the normalized magnetic field collected at a clothing and an electronic venues is shown in Fig.~\ref{fig:magn}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth,height=3cm,keepaspectratio]{mag1} %
\caption{Comparison between the magnitude of normalized magnetic field in an electronic and a clothing store.}
\label{fig:magn}
\end{figure}
The FFT is applied on the normalized magnetic signature to generate an energy signature independent of time and walking patterns. This will avoid the burden of
collecting magnetic information in different directions in order to construct an accurate magnetic map.\\
\textbf{\textit{SSID Fingerprint:}}
The SSID of an Access Point (AP) installed in a certain venue may be indicative of the venue name given that the vast majority of shops have a wireless Internet connection. The WiFi scan collected while the user is performing a checks-in contains a number of AP SSIDs overheard at that place. Consequently, the location's SSID fingerprint is represented as the SSID of the AP that has the strongest average RSS in that scan. However, before computing the strongest AP, APs having SSID holding common AP manufacturers or service providers names (e.g., LinkSys or Vodafone) are filtered out.\\
\textbf{\textit{ OCR Fingerprint:}}
Opportunistic images captured while users visiting a place may contain menus, store logo, or postings. CheckInside mines words from these images by incorporating the HP Tesseract OCR engine \cite{smith2007overview}. The set of words extracted from place-related images constitute its OCR fingerprint. However, the venues database manager keeps track of tips posted by users at candidate venues crawled from Foursquare.\\
\textbf{\textit{Familiarity:}}
It indicates how frequently a user visits the
same venue and/or the venue's brand and is measured by the number of check-ins
the user has performed at the venue and/or its brand. We hypothesis that
this feature will outperform popularity used in \cite{elhamshary2014checkinside} as the
place popularity is measured over all users and a certain
venue may be popular in general but the current user is not
interested in it.
\subsection{New Venue Determination}
As LBSNs are organic systems which are based on users' contributions, some venues may not be covered yet (e.g., 35\% in case of Foursquare) as illustrated in the study shown in Sec. \ref{sec:study}. When the user issues a check-in request, CheckInside should determine if the test venue is a new (i.e., visited for the first time) one, or it is already included in the venues database, to decide where to forward the test place fingerprint. If the test venue is new, the coverage extender module is used to predict its name. Otherwise, the venue ranking module is employed to match the test venue fingerprint against candidate venues retrieved from the venue database. To recognize whether the current venue is new or not, we draw on the WiFi similarity (using Eq.~\ref{eq:wifi_sim}) among the current place WiFi bind and all candidate places WiFi fingerprints. If the maximum WiFi similarity among test venue and other candidate venues is lower than a certain threshold (1.2 in our system), this indicates that the place is not covered yet by the system.
\subsection{Coverage Extender Module}
\label{sec:covext}
Given the fact that the vast majority (82\%) of stores in large shopping malls belong to well-known international or local brands, as shown in the study in Sec. ~\ref{sec:study}. CheckInside leverages this fact to extend the coverage of current LBSNs by predicting the names of venues that are not yet included in the LBSN database.
To predict the name of a venue, CheckInside relies on two subsequent approaches. The first approach uses the WiFi scan collected during the check-in at the test location. It will first determine the strongest AP at that location. The SSID of the strongest AP is compared against a predefined list of brand names (retrieved from the CheckInside database). The venue name that encounters the lowest edit distance with the strongest AP SSID is deemed as the predicted name for the new venue. To decrease the false positive (FP) rate, venue naming is confirmed only if the lowest edit distance is less than certain threshold. Nevertheless, if the first approach failed to predict the name, the system will compare the test place logical fingerprint (the complete fingerprint excluding WiFi, magnetic and location) against the logical fingerprints of all brand venues in other malls (indexed in the venues database). Our intuition is that most shops belonging to the same brand share decorations, color themes, lighting styles, may play the same type of music, and have similar mobility data. Consequently, logical fingerprints matching of test venue against brand venues will be able to predict the correct name for brand venues with high probability.
\subsection{Venues Ranking Module}
This component is responsible for ranking the candidate list generated by the Venues Database Manager.
It accomplishes this by three main components (blue in Fig. \ref{fig:arch}):
filtering, feature-based ranking, and rank aggregation.
\subsubsection{Filtering}
The function of this component is to eliminate candidate venues that are not likely to be similar to the test venue. This helps in increasing the efficiency and accuracy of the next ranking modules. Filtering is performed based on the current user location and the WiFi fingerprint. Both filters are run independently and concurrently returning a number of candidate venues. To avoid excessive filtering, each filter returns a fixed number of locations (taken the same as the Foursquare API default of 10\footnote{\url{https://developer.foursquare.com/docs/venues/search}}). The output lists of the two filters are aggregated, generating the candidate list. The number of places in the candidate list can be extended/shortened depending on the confidence in the place inference or/and the user preference.
\textbf{\textit{Filtering By Location:}}
This is performed by placing a threshold on the distance between the current user location and the candidate venue location. The metric used for distance calculation is the shortest door-to-door walking distance, rather than the euclidean distance. To speed up the retrieval of closest candidate venue to user location, we use an R-tree to index the venues database.
\textbf{\textit{Filtering By WiFi Fingerprint:}}
It is performed by computing the similarity between the test venue WiFi fingerprint and all candidate venues WiFi fingerprints using Eq.~\ref{eq:wifi_sim} and then returning the venues with the highest scores.
\subsubsection{Feature-based Ranking}
This module orders candidate venues according to their pairwise similarity with the test venue. CheckInside employs all extracted features for matching the test place against candidate places.
Each ranker orders the pruned list of nearby venues received from the filtering component based on one of the features in parallel.
\textbf{\textit{Sound ranker:}}
To compute the degree of similarity between two sound fingerprints, we use the Euclidean distance between the corresponding sound fingerprint 100-dimensional vectors.
\textbf{\textit{Image ranker:}}
We employ the technique developed in \cite{yan2010crowdsearch} for image search to our image ranking operation. Specifically, we use an inverted index constructed from the corpus of images in the venues database. The inverted index is a mapping of each visterm feature in the test images to the images in the database containing that visterm. The IDF of found visterms in a candidate venue are averaged to get the venue score.
The list of candidates are returned ranked in order of their average IDF score.
\textbf{\textit{Mobility data ranker:}}
This module computes the similarity based on visiting time ($v$), user activity ($r$), and stay duration ($d$) between each venue in the candidate list and the user test venue. The similarity is taken as the joint probability of the different mobility features at the candidate venues. In particular, the mobility similarity ($m$) between the current user mobility test data ($v, r$, and $d$) and a venue fingerprint ($F$) is given by:
\begin{equation} \label{eq:mob}
m = P((v, r, d)|F)= P(F_V=v).P(F_R=r).P(F_D=d)
\end{equation}
Where $P(F_V=v)$ can be obtained from the histogram of the user visiting time at the candidate venue, $P(F_R=r)$ from the histogram of the user activity, and $P(F_D=d)$ from the histogram of stay duration.
This metric indicates that a candidate venue is good if it has a high probability of matching the current user mobility behavior. For example, food venues would have close visiting time (e.g., at meals time), long stay durations (e.g., 30+ minutes), and similar user activity (e.g., sitting) with high probability.\\
\textbf{\textit{Color/Light ranker:}}
The color/light similarity is performed based on the Euclidean distance between their cluster centroids and
the clusters' sizes \cite{azizyan2009surroundsense}. The similarity ($S$) between fingerprints $F_{1}$ and $F_{2}$ is defined as:
\begin{equation}
S=\sum_{i,j} \dfrac{1}{\delta(i,j)} \dfrac{\textrm{sizeOf}(C_{1i})}{T_{1}}\dfrac{\textrm{sizeOf}(C_{2j})}{T_{2}}
\end{equation}
where $ C_{1i} $, $ C_{2j} $ are set of clusters for fingerprints $F_{1}$ and $F_{2}$ respectively. $T_{1}$, $T_{2}$ are the total number of pixels in clusters in $F_{1}$ and $F_{2}$ respectively, and $\delta(i,j)$ is the centroid distance between the $i^{th}$cluster of $F_{1}$ and the $j^{th}$ cluster of $F_{2}$.\\
\textbf{\textit{Magnetic field ranker:}}
It estimates the similarity between the magnetic signal vector at a test place $M_{a}$ and a candidate place $M_{b}$ by the Euclidean distance defined as:
\begin{equation}
S(M_{a},M_{b})=\sqrt{(m_{a}^{x} -m_{b}^{x}) ^{2} + (m_{a}^{y} -m_{b}^{y}) ^{2}+(m_{a}^{z} -m_{b}^{z}) ^{2}}
\end{equation}
\textbf{\textit{SSID ranker:}}
It computes the string similarity using the edit distance among the SSID of strongest AP in the WiFi bind of the test venue and candidate venues names. For a candidate venue with duplicate names, the average edit distance among all names of the candidate venue and the test venue name is deemed as their similarity measure.\\
\textbf{\textit{OCR ranker:}}
It matches venues tips against OCR text mined from images captured at the test place. As a preprocessing step, stop words and non grammatical words are filtered out. Finally, the ranker orders candidate venues based on the number of overlapping terms between the texts extracted from OCR
and from the user tips \cite{chon2012automatically}.\\
\textbf{\textit{Familiarity ranker:}}
It ranks more familiar venues with respect to the current user higher in the list.
\subsubsection{Rank Aggregation}
Once the different rankers provide their ranked lists of places, this module fuses them into a single ranked list. We experimented with CombSUM as an example of score-based methods (combine the different lists based on the assigned scores in the individual lists) and the Borda's method as an example of order-based ones (combine the lists based on the order of places in each individual list); both representing data fusion (unsupervised rank aggregation) techniques \cite{faheem2010rank}. In addition, we experimented with different learning-to-rank (supervised rank aggregation) algorithms like AdaRank \cite{xu2007adarank} and Ranking SVM \cite{joachims2002optimizing}. We compare the different techniques in Sec.~\ref{sec:perf_modules}.
\subsection{User Feedback Module}
A significant characteristic of the users' interaction with a LBSN is that the user explicitly selects a venue to check-in from the list of nearby venues, which acts as a ``ground-truth'' for the user current place. This feedback not only provides information about the performance of CheckInside venues ranking algorithm, but it also can improve the system future performance by identifying which ranker provides the best performance.
\noindent Specifically, rankers performance varies from venue to another. For example, some venues have distinct color theme (e.g., restaurants), other have unique sound signature (e.g., libraries are quiet), and no WiFi signal is overheard in some venues. After each check-in, the rankers weight is updated to reflect the degree of user satisfaction to the returned results.
We leverage this user feedback to weigh the different rankers. Initially, to ensure correctness, fake check-ins are filtered out by the incorrect check-ins detection module. Then, we start with all rankers having an equal weight. After each check-in operation, and given that the candidate list contains $l$ venues, each ranker is assigned a score of $l-i$, where $i$ is the rank of the actual venue in the ranker's list. These scores are then normalized to add to one.
\subsection{Semantic Floorplan Labelling Module}
This component is responsible for the automatic labelling of venue names on the floorplan. CheckInside starts with a floorplan with rooms and corridors highlighted which can be either manually uploaded or automatically generated from crowdsourced data~\cite{alzantot2012crowdinside}.
To enrich the floorplan with the semantic labels of the venue names, one \textbf{\emph{cannot simply }} use the user check-in information, which provides the current venue name and the current user location due to the errors inherent in the check-in process in the form of incorrect check-ins. So, as a preprocessing step, the incorrect check-in detection module will identify and remove outlier and noisy check-ins taking place outside the actual venue. Once the fake check-ins are removed, the venue location is estimated as the mean of the locations of the users who check-in at this venue. Based on the law of large numbers, this mean converges to the actual location as the number of samples increases. The venue enclosing this location on the map is tagged accordingly.
\subsection{Incorrect Check-ins Detection Module}
In LBSNs, without a method for detecting fake check-ins, the quality of location information is based solely on the honesty of users. Fake check-ins can be performed in different ways including faking GPS coordinates or checking-in a venue that is not nearby. Moreover, the indoor localization algorithm employed has inherent error in the range of few meters, which may place the user at an incorrect venue. Fake check-ins have a negative impact of the operations of LBSNs, including monetary loss and degraded quality (e.g., in recommendation services that make use of the check-ins data). In addition, fake check-ins affect the performance of several modules of CheckInside such as user feedback and semantic floorplan labelling.
To address these challenges, CheckInside uses an unsupervised outlier detection algorithm as there is no a-priori model available for identifying correct check-ins. Our approach is based on the outlier detection in the WiFi signal space. Specifically, we depend on the fact that independent \emph{correct} check-ins made at the same venue are adjacent in the WiFi signal space and tend to cluster, while fake check-ins are distributed over a larger area. Consequently, we apply an agglomerative hierarchical clustering approach to detect check-ins that are suspected to be erroneous. Later, label assignment and user feedback incorporates only check-ins tagged as correct. The system maintains all WiFi fingerprints assigned to a venue during check-ins within a time window (regardless of correctness), so that recent data can be used to periodically reclassify clusters and detect outliers for that venue. For location verification of a check-in of a user at certain venue v, we utilize the recent k RSS vectors collected by users claiming presence in v.
For the agglomerative hierarchical clustering algorithm, clusters are successively merged in a bottom-up fashion, based on the WiFi similarity metric in Eq.~\ref{eq:wifi_sim}, until the similarity falls below a pre-defined cut-off threshold $d^{*}$. The selection of appropriate value for $d^{*}$ is based on formulating the threshold identification problem as a Bayesian decision problem \cite{park2010growing}. Once check-ins are grouped into clusters, the system identifies which cluster includes the correct check-ins (the rest are assumed to contain fake check-ins).
If we assume that most users make correct binds, it is natural to take the largest cluster as the correct binding for the venue. However, when the system starts, it has not yet obtained enough check-ins and thus majority voting is not feasible. Therefore, we identify the correct cluster of check-ins $c_{v}^{*}$ given a set of check-in clusters ($\mathcal{C}_{v}$) at venue $v$ according to the following criterion:
\begin{equation}
c_{v}^{*}= \argmin_{c\in \mathcal{C}_{v}}\sum_{m\in \mathcal{N}(v)} d_{s}(c,c_{m}^{*})
\end{equation}
where $\mathcal{N}(v)$ is the set of neighboring venues to venue $v$, $c_{m}^{*}$ is the cluster of correct check-ins at neighboring venue $m$ at the time of computation, and $d_s(c,c_{m}^{*})$ is distance between the two clusters centroids. The intuition is that the correct cluster assignment for a venue is the one that is most similar to its neighboring venues.
\setlength{\belowcaptionskip}{-1pt}
\section{Evaluation}
\label{sec:eval}
CheckInside is evaluated through a multi-mall deployment that include 711 stores distributed in four malls in two different cities over six-weeks period.
\subsection{ Data Collection}
We recruited 20 participants to collect the necessary data for evaluation. While visiting places, participants capture images and record audio samples. Simultaneously the deployed data collection tool collects user traces and samples WiFi. To collect ground-truth (GPS is not available inside malls building), participants manually label the venue when they depart the place. Table~\ref{tab:desc} shows the description of collected data. The light proximity sensors are used to make sure that pictures are taken when the phone is in the user's hand. The phone orientation sensor is used to differentiate between pictures of floors (used for color/light ranking) and other positions (used in images ranking). About 22.7\% of the pictures were blurred images (Fig.~\ref{fig:pictures}).
\begin{table}[!t]
\resizebox{0.48\textwidth}{!}{
\begin{tabular}{|p{2cm}||c|c|c|c||c|c|c|}
\hline
\multirow{2}{*}{\textbf{Venue Type}} & \multicolumn{4}{|c}{\textbf{CrowdSensing}} & \multicolumn{3}{|c|}{\textbf{Foursquare}} \\ \cline{2-8}
& \textbf{\begin{tabular}[c]{@{}c@{}}\# of\\ Venue\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}\# of \\ Image\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}\# of \\ Color\\/Light \end{tabular}}& \textbf{\begin{tabular}[c]{@{}c@{}}\# of\\ Sound\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}\# of \\ Venue\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}\# of\\ Image\end{tabular}}&\textbf{\begin{tabular}[c]{@{}c@{}}\# of\\ Tip\end{tabular}}\\\hline
\textbf{Food} & 101 & 1366 & 1087 &380 &91 &1243 & 1299 \\ \hline
\textbf{Cloth.\&Fash.} & 374 & 5031 & 2914 &1237 &235 &374 & 641 \\ \hline
\textbf{Arts\&Ent.} & 23 & 287 & 183 &127 &14 &57 &138 \\ \hline
\textbf{Others} &213 &2648 & 1514 &847 &96 & 234 &277 \\ \hline\hline
\textbf{Total} & 711 &9332 &5698 &2591 & 436 & 1908 &2355 \\ \hline
\end{tabular}
}
\caption{Description of collected data.}
\label{tab:desc}
\end{table}
\setlength{\belowcaptionskip}{-6pt}
Moreover, Foursquare is crawled to extract venues attributes such as name, category, user tips, among others. The dataset crawled from Foursquare contains only 436 stores (i.e., covered by Foursquare) out of the 711 stores used in this evaluation.
\subsection{Methodology}
The participants are divided into four groups. Each group of five participants is assigned to a mall. To make the data collected as natural as possible, participants were asked to behave normally while visiting each place category. For example, participants browse items in the shelves and wait in queues at cashiers like normal customer at Clothing venues. Moreover, each place is visited five times on different days by different participants. The data is collected using different Android phones including Samsung Galaxy S plus, Nexus One, Galaxy Tab, among others. This captures the time-variant nature of the fingerprint at the same venue as well as the heterogeneity of users and devices. Finally, some participants check-in in the neighborhood of the place (e.g., corridors) while others check-in inside the place.
\subsection{Performance Results}
We start by evaluating the accuracy of different system modules and the performance of the system in different modes of operation. Finally, we quantify the advantage of CheckInside as compared to traditional LBSNs, in terms of ranking accuracy, coverage and power consumption.
\subsubsection{Performance of Different System Modules}\label{sec:perf_modules}
\noindent\textbf {1) New Venue Determination Module:} To evaluate the ability of CheckInside to recognize new places (not included in its database), we perform a two round process. At the first round, a WiFi fingerprint of each venue is matched against all WiFi fingerprints of other venues in the database (not including itself) simulating the case when the user is at a new place. At the second round, we match the WiFi fingerprint of each venue against all other venues in the database (including itself) simulating the case when the venue is covered by the system. At each case, the \textbf{maximum WiFi similarity} is calculated and compared against a predetermined similarity threshold. If the maximum similarity is less than the similarity threshold, the venue is confirmed as a new venue. Fig.~\ref{fig:newvenue} shows the effect of varying the similarity threshold (from conservative to lenient) on the CheckInside ability to recognize new venues. When the system is conservative (i.e, small threshold) in recognizing new places, it can avoid False Positive (visited places classified as new one). As the threshold increased, the module will recognize more new venues but at the cost of higher FP rates. A similarity threshold of 1.2 is selected empirically as a comprise between False Positive (FP) and True Positive (TP) rates.
\begin{figure}[!t]
\noindent\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=2.8cm]{newvenue}
\captionof{figure}{Performance of new venue recognition.}
\label{fig:newvenue}
\end{minipage}
\hfill
\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=2.8cm]{filter}
\captionof{figure}{Performance of the different venues filters.}
\label{fig:filter}
\end{minipage}
\end{figure}
\noindent\textbf{2) Venues Ranking Module:} \\
\textbf{Filtering:} Filter accuracy refers to the ability of the filter to return the actual venue within its list. Fig.~\ref{fig:filter} shows the effect of changing the candidate venues list size on the filter accuracy. It shows that both the WiFi and location-based filters have comparable performance with a slight advantage to WiFi-based filtering due to the wall-aliasing effect described before. The accuracy can be further increased by combining their output. CheckInside can achieve 100\% accuracy with a candidate list as small as 15 entries. This is compared to Foursquare that can achieve only 29.5\% for the same list size as we presented in our earlier study. This highlights that CheckInside can enhance the user experience, especially for mobile devices with small screens. This is \textbf{further} enhanced by the ranking modules.
\begin{figure}[!t]
\noindent\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=3cm]{match1}
\captionof{figure}{Performance of the different rankers.}
\label{fig:rankers}
\end{minipage}
\hfill
\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=3cm]{rankaggr}
\captionof{figure}{Comparison of rank aggregation techniques.}
\label{fig:agg}
\end{minipage}
\end{figure}
\textbf{Ranking:}
The performance of different rankers in isolation is shown in Fig.~\ref{fig:rankers}.
The WiFi and location-based rankers have the best accuracy due to relatively lower noise compared to the other rankers. The WiFi ranker, again, has better performance than the location ranker due to the wall-aliasing effect.
For the Color/Light ranker, the average accuracy for detecting the user venue is about 42\%. We observe that this ranker performs well in stores where there is a large diversity in the color and light intensities. For example, food places are usually grouped into food court area and thus for economical competition they have diverse color themes and decorations.\\
The image ranker achieves moderate accuracy (47.6\%). This is due in part to the 22.7\% collected blurred images. Another reason that made the image fingerprint less effective is that most of adjacent venues are of the same category and the majority of venues (over 50\%) are clothing and fashion shops. Therefore, pictures captured in one shop are similar to other pictures taken at adjacent shops as they contain similar shelves and items. We noticed that the image fingerprint achieves high accuracy in small size shops, where the majority of the captured pictures have a number of common interest points due to the limited area available to take photos.\\
The sound ranker offers a little discriminative power as the majority of stores have a small size at which the sound samples recorded at a venue contain noise from adjacent stores. Moreover, the majority of shops are crowded with people, which makes the sound fingerprint has similar background noise for different stores. Finally, we observe that the majority of stores are correctly identified based on their sound fingerprint are those that either have no music played in the background or are less popular venues having a small number of customers.
For the mobility ranker, the logical design of the malls, which divides the available space into blocks mostly from the same category prevents the diversity of mobility data (user activity, visiting time, and stay duration). This leads to ambiguity between adjacent stores. Another reason is that the stay duration is highly affected by the time at which users perform the check-in. For example, if a user performs a check-in just after arriving a cafe (typically identified by long stay periods), this will make the reported stay duration short, leading to incorrect identification.\\
The magnetic ranker provides the actual venue on the top-5 of the list in 49.5\% of cases. This degraded performance is due to the difference of the movement pattern of different users which affect magnetic field distortion sensed by the phone compass. The magnetic field can perform better when the magnetic field is measured at each point in all direction forming a complete circle as discussed in \cite{chung2011indoor} which is impractical. However, it can easily recognize stores having a distinct pattern of magnetic distortion by metals and/or electronics (e.g., electronic stores).\\
The OCR based ranker orders the actual venue on the top-10 of the list in 74\% of cases.
Its performance depends on the amount of words mined from images, the amount of tips as well as their overlapping terms. Consequently, it achieves best accuracy in cafe and restaurant which contain large number of tips conveying user feedback about different menu items and the images uploaded or collected opportunistically at these venues type always contain menus and/or store signage.\\
The SSID ranker performs well and it can correctly infer the user location in about 49\% of cases. The reason is that 58.7\% of venues have APs conveying SSIDs very similar to venues name. However, about 18\% of these AP are not the strongest SSID in this venue.
Finally, familiarity ranking usually used in Foursquare does not achieve the desired performance. This due to the unpredictable nature of user behavior and the fact that on average 50\% of generated of a venue check-ins are performed by first time visitors (Sec.~\ref{sec:study}). The familiarity ranker, however, may be important as a tie-breaker when other features are absent or they reported approximately the same similarity values.
Finally, the last bar in Fig.~\ref{fig:rankers} shows the CheckInside rank aggregation performance with \textbf{equal} weights for all rankers. The figure shows that the actual venue is within the first five places 99\% of the time. This is even enhanced using the feedback module as quantified next.\\
\textbf{Rank Aggregation:} Fig. \ref{fig:agg} compares the performance
of different rank aggregation algorithms. Evident from the
figure, for data fusion algorithms, the Borda's method
provides better performance than CombSUM as rankers
have high score variance which may make one ranker
dominate the others, even if it is not the best ranker.
For learning-to-rank algorithms, AdaBoost outperforms
Ranking SVM as it is designed to optimize loss functions
incorporated with any performance measure. Overall, AdaRank
outperforms all rank aggregation algorithm with slight
improvement over Borda's method. However, we opt to use
Borda's method as it does not need training, is simple
to implement, and does not require parameters tuning.
\noindent\textbf{3) User Feedback Module:} Fig.~\ref{fig:equal} shows how the weights of the different modules evolve with time starting from equal weights. The weights converge to values proportional to the discriminative abilities of the different rankers as quantified in the previous section. This is reflected on the overall accuracy as shown in Fig.~\ref{fig:feed}, where using the feedback module can detect the actual venue 87\% of the time compared to only 76\% without using user feedback.
\begin{figure}[!t]
\noindent\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=\textwidth,height=2.8cm]{weight}
\captionof{figure}{Weight evolution of different rankers starting from equal weights.}\label{fig:equal}
\end{minipage}
\hfill
\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=2.8cm]{cdffeed}
\captionof{figure}{The CDF of the rank of the actual venue in final list.}
\label{fig:feed}
\end{minipage}
\end{figure}
\noindent\textbf{4) Incorrect Check-in Detection Module:} Fig.~\ref{fig:fake} evaluates the performance of the fake check-in detection at different clustering threshold values $d^{*}$ ranging from conservative ($d^{*}$=12dB) to lenient ($d^{*}$=16dB) clustering. It shows that the trade-off between the probability of correct outlier detection and false alarm (a correct check-in classified as an outlier). Larger value of $ d^{*} $ means that a WiFi bind has high probability to join clusters which decreases the probability of detection. This is due to some fake check-ins will fall into true clusters. However, using large threshold value decrease the probability of false alarm.
\noindent\textbf{5) Semantic Floorplan Labelling Module:} To understand the accuracy of the outlier detection technique of CheckInside regarding correctly inferring the venue location on the floorplan, we select for each venue a group of erroneous check-ins (performed at corridors or at other shops), where $p_{e}$ represents the probability of erroneous check-ins. We compare to an ``Oracle'' (perfect) detector.
Fig.~\ref{fig:semantic} shows that the outlier detection algorithm of CheckInside can provide from 9 to 19\% enhancement in detecting the actual venue location over a wide range of $p_e$. Note that, even with the Oracle detector, there are still errors in labelling the venue due to the inherent errors in the location determination system used. Unsurprisingly, at high error rates, the detector has a low accuracy as all detected neighbors have errors. We believe, however, that the performance is reasonable to typical values of $p_e$.
\begin{figure}[!t]
\noindent\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=2.8cm]{fake3}
\captionof{figure}{The performance of fake check-in detection at different clustering thresholds.}\label{fig:fake}
\end{minipage}
\hfill
\begin{minipage}[t]{0.485\linewidth}
\includegraphics[width=1\textwidth,height=2.8cm]{semgen}
\captionof{figure}{Semantic labelling accuracy at different check-in errors ($p_e$) for different detectors.}
\label{fig:semantic}
\end{minipage}
\end{figure}
\noindent\textbf{6) Coverage Extender Module:}
The coverage enhancement is evaluated as the percentage of venues that are discovered by CheckInside and are not included in the Foursquare database. To evaluate the ability of CheckInside to increase the coverage of current LBSNs, we performed a cross validation experiment where all the venues of one mall are used as the test and the venues of the other three malls are used as the database.
The ability of CheckInside to predict venues \textbf{names} using APs SSIDs is shown in Fig.~\ref{fig:cov}(a). The Fig. shows the prediction accuracy at different values of edit distances among venues names and strongest AP's SSID. As the edit distance increased, more venues name can be predicted at the cost of false prediction. When the distance threshold reaches a certain value, no more venues are added as the coverage extension is bounded by the number of brand uncovered venues. Eventually, the system is able to predict the names of about 31\% of uncovered venue without any false prediction. Fig.~\ref{fig:cov}(b) shows the logical localization based name prediction accuracy by plotting venues coverage ratio for different venues categories. It shows that this approach can predict the names of 34\% of uncovered venues increasing the coverage ratio of the system from 65\% (the Foursquare ratio including the granularity mismatch) to 77\%. The logical localization based prediction accuracy depends on the category and it is better for categories that are more likely to have many brand venues and has a diversity in their ambient fingerprint, e.g., restaurants chains.
Finally, using the two approaches subsequently will allow CheckInside to extend the coverage of traditional LBSNs by 37\%.
\begin{figure}[]
\centering
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth,height=2.8cm]{coverage_label}
\caption{The prediction accuracy of uncovered venues names at different edit distance thresholds.}
\end{subfigure}
\begin{subfigure}[t]{0.24\textwidth}
\includegraphics[width=\textwidth,height=2.8cm]{coverage_comp}
\caption{Extended coverage using CheckInside for different categories compared to Foursquare.}
\end{subfigure}
\caption{Evaluation of the coverage extender module.}
\label{fig:cov}
\end{figure}
\subsubsection{Performance in Different Modes of Operation}
Fig.~\ref{fig:modes} shows the performance of CheckInside in different modes of operation that can trade-off accuracy and privacy. In particular, we compare the full system using all sensors (Full) with using only location information derived from the indoor localization technique (Loc. only), and the system when the camera is turned off (Cam. off), mic is turned off (Mic. off) and when both are tuned off (Mic.-Cam. off). The figure shows that CheckInside can provide 58\% better performance in estimating the exact venue than a simple system that is based only on location estimation. This confirms the usefulness of the semantic fingerprint. The figure also shows that CheckInside can maintain high accuracy in detecting the exact venue location (71\%) even when the privacy-sensitive sensors are turned off.
\subsubsection{Comparison with other Systems } \label{sec:perf_comp}
We compare the performance of CheckInside to Foursquare (as a typical LBSN) and the place naming approach in \cite{chon2013autonomous} that extends traditional LBSNs to do semantic fingerprint matching using the information available about the venues from LBSN (i.e., visit time, popularity, tips, and images SIFT and gist features) but without taking the venue physical location nor the phone sensors (except the camera) into account for places inference.
As illustrated in Fig.~\ref{fig:total_comp}(a), CheckInside correctly inferred the exact venue in 87\% of the cases compared to less than 36\% for the best of the other two systems. Moreover, CheckInside can infer the correct venue within the top 5 venues 99\% of the time as compared to 62\% for the closest system, leading to a better user experience.
Fig.~\ref {fig:total_comp}(b) shows the distance error between the actual venue and the top venue suggested by different systems. As evident from the figure, CheckInside median distance error is less than 7m as compared to less than 52m for the closest systems. This highlights that CheckInside can provide better value for location-based business.
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.24\textwidth}
\includegraphics[width=\textwidth,height=2.8cm]{cdfcomp}
\caption{Rank of actual venue.}
\end{subfigure}
\begin{subfigure}[b]{0.24\textwidth}
\includegraphics[width=\textwidth,height=2.8cm]{cdfcompdist.eps}
\caption{Distance error.}
\end{subfigure}
\caption{Comparing ranking of CheckInside against other systems.}
\label{fig:total_comp}
\end{figure}
Fig. \ref{fig:ener} compares the power consumption of CheckInside with the other two systems. The
power is calculated using the PowerTutor profiler \cite{gordon2013powertutor} and
the Android APIs using the HTC Nexus One cell phone.
The figure shows that the camera is the most energy hungry sensors. However, due to the {CheckInside}{} triggered sensing scheme, its energy consumption is still better than other techniques that use the camera continuously.
Moreover, when the camera is turned off, e.g. in the privacy-preserving mode, {CheckInside}{} power consumption is comparable to the most power efficient system, with superior accuracy advantage (Fig.~\ref{fig:modes}). In addition, noting that the inertial sensors are always running to detect the phone orientation changes and WiFi may be already turned on by the phone user, {CheckInside}{} typically consumes little extra energy in addition to the normal phone operation.
\begin{figure}[!t]
\noindent\begin{minipage}[t]{0.48\linewidth}
\includegraphics[width=1\textwidth,height=2.6cm]{modes}
\caption{The actual venue ranking for different modes of operations of CheckInside.}
\label{fig:modes}
\end{minipage}
\hfill
\begin{minipage}[t]{0.48\linewidth}
\includegraphics[width=1\textwidth,height=2.6cm]{ener6.eps}
\caption{The energy footprint of {CheckInside}{} and other systems.}
\label{fig:ener}
\end{minipage}
\end{figure}
\section{Related Work}
\label{sec:related}
Previous work focus on three main categories: venues discovery, venues description, and venues recommendation.
\textbf{Venues Discovery}
means to learn significant locations that are semantically meaningful to people such as home or work. Techniques
\cite{kim2011towards,zheng2010collaborative} mine the user GPS trajectory and other extracted features (e.g., duration spent at each location) to infer this information. Other techniques, e.g., PlaceSense \cite{kim2009discovering}, use RF beacons to extract the visited places, while others, e.g., SenseLoc \cite{kim2010sensloc}, rely on both WiFi and GPS; and accelerometer to detect visited places and distance travelled by users. CheckInside complements these techniques by providing finer-grained labelling and categorization of indoor locations with richer semantics.
Another category of techniques, e.g., SurroundSense\cite{azizyan2009surroundsense}, use \textbf{manually-created} semantic fingerprints to infer the user \textbf{location}. CheckInside, on the other hand, automatically creates these fingerprints in a way transparent to the user (i.e., more ubiquitous). Moreover, it addresses the incorrect association between the current user location and the check-in venue. Finally, it provides a complete system for both indoor LBSNs and semantic floorplan labelling.
\textbf{Venues Description}
refers to assigning either category (e.g., restaurant, drug store)\cite{chon2012automatically,semsense}, business naming (e.g., KFC)\cite{chon2013autonomous}, informal labels (e.g., my hometown) \cite{pontes2012we}, or activities or function associated with the location (e.g., eating, playing football, PoIs) \cite{lian2011collaborative,transit,map++,dejavu,semanticslam} to venues.
Some of venues description techniques depend on data collected from smartphones by crowd-sensing, e.g., \cite{chon2012automatically,lian2011collaborative}. The CSP system \cite{chon2012automatically} exploits opportunistically captured images
and audio clips from smartphones to link place visits
with place categories (e.g., store, restaurant). Other techniques, e.g.,
\cite{pontes2012we}, exploit existing large-scale data collections (e.g., Yelp PoI database) or location-based community-generated content (e.g., Foursquare). The closest work to ours is the place naming technique in \cite{chon2013autonomous} that integrates OCR text from images, mobility data, and image features with information about venues available from social networks to provide category (e.g., food), business (e.g., KFC), or personal naming (i.e, home and work) of venues visited by users. This system, however, focuses only on the \textbf{offline analysis} of data collected within known venues (i.e.,\textbf{ it assumes zero localization error}) and does not distinguish between indoor an outdoors venues. CheckInside, on the other hand, targets indoor LBSNs, exploits more sensing dimensions available in mobile devices to address intentional or accidental location and check-in inaccuracies and enhances the ranking performance of LBSNs, is oriented for indoor operation, and works in real-time to provide the user with an accurate list of nearby venues. Moreover, CheckInside generates a semantic floorplan of indoor buildings and leverages the user implicit feedback in the check-in operation to adapt the system dynamically. Finally, our system increases the venues coverage percentage of current LBSNs.
\textbf{Venues Recommendation}
With millions of users of LBSNs, rich knowledge has accumulated about places visited by users which enables suggesting meaningful locations to the user. In \cite{park2007location}, the system matches users
profile data (age, gender, cuisine preferences, and income) against the price and
category of a restaurant using a Bayesian network model. Other systems, e.g., \cite{ye2010location}, exploit users' rating for places available online (e.g., Yelp) to suggest places to their user; while other recommendation system \cite{cao2010mining} employ user-generated trajectories to find interesting locations.
CheckInside can provide richer and fine-grained venue information, which can be used by these system to provide better recommendations.
\section{Discussion }
\label{sec:disc}
\textbf{Other Applications:}
CheckInside is currently applied to shopping malls. However, the same concept can be extended to other venue types (e.g., educational buildings). In addition, the user visiting patterns can be analyzed to provide valuable statistics (i.e., indoor analytics).
\noindent \textbf{Energy-Performance Tradeoff:} CheckInside fuses different
phone sensors to identify the user current place accurately,
which may consume more energy. To further reduce
energy consumption, in addition to triggered sensing and leveraging energy-efficient sensors, users can select energy-friendly
modes (e.g. disable energy-hungry sensors like camera and
the sound sensor), possibly reducing the system accuracy.
Nonetheless, CheckInside still outperforms the
of other systems in this case with significantly better energy consumption (Figs. \ref{fig:modes} and \ref{fig:ener}).
\section{Conclusion}
\label{sec:conclude}
We presented the CheckInside as a fine-grained indoor location-based social network. CheckInside leverages data mined opportunistically from the users' phone sensors and data about venues extracted from traditional LBSNs to fingerprint each venue. It then applies a number of filtering and ranking steps to create a ranked list of candidate locations that are returned to the user to select a venue to check-in at. Check-ins performed by users are forwarded to the incorrect check-in detection module to handle noise in the check-in data. The user implicit feedback from the correct check-in operation is used to dynamically adjust the system parameters. We also presented novel approaches for semantic labelling of the building floorplan and extending the venues coverage of current LBSNs.
Extensive evaluation of CheckInside in four malls shows that CheckInside can infer the actual venue 99\% of the time within the top 5 venues in the candidate list. In addition, it increases the coverage of current LBSNs by 37\% by predicating names of uncovered venues.
\section{ACKNOWLEDGMENTS}
This work is supported by a Google Research Award.
\input{checkinside_CR_v2.bbl}
\begin{IEEEbiography}[{\includegraphics[width=1in,height=2in,clip,keepaspectratio]{moustafa.jpg}}]{Moustafa Elhamshary} is a PhD student at Egypt-Japan University of science and Technology (E-JUST), Egypt. Currently, he is a research fellow at Graduate School of Information Science and Technology, Osaka University, Japan. His research interests include location determination technologies, mobile sensing, human activity recognition and location-based social networks.
\end{IEEEbiography}
\begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{anas.jpg}}]{Anas Basalamah} is an Assistant Professor at the Computer Engineering Department of Umm Al Qura University, Saudi Arabia. He is a cofounder and director of the Wadi Makkah Technology Innovation Center at Umm Al-Qura University. He did his MSc and PhD Degrees at Waseda University, Tokyo in 2006, 2009 respectively. He worked as a Post Doctoral Researcher at the University of Tokyo and the University of Minnestoa in 2010, 2011 respectively. He cofounded two tech startups: Navibees and Averos. His areas of interest include embedded networked sensing, ubiquitous computing, participatory and urban sensing.
\end{IEEEbiography}
\begin{IEEEbiography}[{\includegraphics[width=1in,height=1.25in,clip,keepaspectratio]{youssef.jpg}}]{Moustafa~Youssef}
is an Associate Professor at
Alexandria University and Egypt-Japan University
of Science and Technology (E-JUST),
Egypt. His research
interests include mobile wireless networks,
mobile computing, location determination technologies,
pervasive computing, and network
security. He has fifteen issued and pending
patents. He is an associate editor for the ACM TSAS, a previous area editor of the ACM MC2R, and served on the organizing
and technical committees of numerous prestigious conferences. Prof. Youssef is
the recipient of the 2003 University of Maryland Invention of the Year
award, the 2010 TWAS-AAS-Microsoft Award for Young Scientists, the 2012 Egyptian State
Award, the 2013 and 2014 COMESA Innovation Award, multiple Google Research Awards, the 2013 ACM SIGSpatial GIS Conference Best Paper Award, among others. He is also an ACM Distinguished Scientist.
\end{IEEEbiography}
\end{document}
|
\section{Introduction}
In recent years, a new generation of experiments has measured with unprecedented precision the energy spectrum of \gls{UHECR}, demonstrating the existence of a cutoff at the GZK energies \cite{AugerSpectrum,Hires}. However, it is not yet possible to determine the mass of those particles with enough precision to constrain the astrophysical scenarios for their origin, and thus distinguish the origin of such a cutoff, which can be due to the exhaustion of the energy at the sources or to the interaction of the \gls{UHECR} with the \gls{CMB}.
\gls{UHECR} interact with atmospheric atomic nuclei producing a cascading process of particle reactions commonly designated by Extensive Air Shower (EAS). EAS can be measured by detecting the radiation emitted by their passage through the atmosphere or by sampling the secondary particles at ground.
The reconstruction of the primary \gls{UHECR} properties relies on our understanding of EAS physics and, in particular, of hadronic particle interactions. Our knowledge of hadronic interactions is limited, which is translated into
phenomenological models including parameters supported by experiments only up to LHC energies.
Moreover, the available accelerator data do not cover the full kinematic region of interest, namely the forward region, nor the full possible interaction systems such as pion-Nitrogen
interactions, which are of the utmost importance for the description of the shower development.
In summary, the different compositions of the UHECR, uncertainties in the hadronic interactions extrapolations, and the possibility of new physics scenarios at the highest energies often share the same phase-space of EAS observables, making it difficult to disentangle hadronic physics effects from UHECR mass determination.
The Pierre Auger Collaboration\cite{PAO_0} interprets the evolution of the electromagnetic shower maximum with increasing energy as an indication towards a heavier primary composition \cite{Xmax2014,XmaxInt2014}.
On the other hand, hadronic interaction models fail to accurately predict the number of muons that reach the ground. At $10^{19}$ eV the mean muon number in simulations would have to be increased by $30\%$ (EPOS-LHC\xspace) to $80\%$ (QGSJ{\textsc{et}}-II.04\xspace) \cite{MuonN2014} to match the number of muons deduced from the depth of the electromagnetic shower maximum. Moreover, measurements of the depth at which the muon production rate reaches its maximum at energies above $10^{19.2}$ \cite{MPD2014} are strongly inconsistent with EPOS-LHC\xspace predictions.
In each hadronic interaction, the produced $\pi^0$ carry around $25\%$ of the energy. They decay almost instantly into photons feeding the electromagnetic cascade. The rest of the energy is carried mainly by charged pions and also other mesons and barions, which continue the hadronic cascade. They eventually decay into muons whenever their energy drops below their critical energy, $ \mathcal{O}$(100 GeV). In this way, the muons act as tracers of the hadronic development: their study gives access to the details of the hadronic cascade. Assuming that the composition can be properly understood, a discrepancy between observed and predicted muon numbers, might indicate a change in the fraction of energy carried by the neutral pions, hadronic cross sections, or hadronic multi-particle production.
This paper explores a complementary scenario: a change in the energy distribution of the muons at production (leaving the total number of produced muons unaltered) might produce changes in the muon measurements performed at ground. In particular, it could mimic (or contribute to changes in) the number of muons arriving at ground. Note that the energy distribution of the muons is directly linked to the angular distribution of muon trajectories, the probability of decay, the interactions with the geomagnetic field and Coulomb scattering. In summary, we study the sensitivity of muon observables at ground to changes in the energy of muons at production, keeping their overall number constant.
The paper is organised as follows. In section \ref{section:Procedure}, we describe the muon distributions at production. In section \ref{subsection: Muons Energy}, we present the details regarding the modification of the muon energy spectrum and its propagation to the ground. The results obtained for the muon observables at ground are shown and discussed in sections \ref{section: Number of muons} and \ref{section: Signal of muons}. Finally, we summarise all results in section \ref{section: Sensitivity}.
\section{Distribution of muons at production}
\label{section:Procedure}
According to \cite{CazonTransportModel,Cazon2004}, the positions where muons are produced (where the parent mesons decay) are confined to a relatively narrow cylinder, of a few tens of meters, around the shower axis.
Thus, muons can be approximated as being produced along the shower axis every d$X$ (depth in g/cm$^{2}$ of crossed atmosphere), within a given energy and transverse momentum interval $\text{d}E_i$ and $\text{d}p_t$, respectively:
\begin{equation}
\frac{\text{d}^3N}{\text{d}X \text{d}E_i \text{d}cp_t} = F(X, E_i, cp_t) \:.
\label{eq: dNdXdE}
\end{equation}
In \cite{CazonTransportModel}, it was proven that $F(X, E_i, cp_t)$ contains all the relevant information needed to fully determine the distributions at the ground. Using the transport model described also in \cite{CazonTransportModel}, it was thus possible to fast simulate the effects on the observed muon distributions at the ground after introducing modifications at production through modifying $F(X, E_i, cp_t)$
The total number of muons produced in a shower is given by
\begin{equation}
N_0 = \int F(X, E_i, cp_t) \text{d}E_i \text{d}cp_t \text{d}X \:.
\label{eq: hX}
\end{equation}
The projection $h(X)= \int F(X, E_i, cp_t) \text{d}E_i \text{d}cp_t$ is the so called \textit{total/true} \gls{MPD} distribution.
It represents the production rate of muons in $\mathrm{g/cm^{2}}$. Its shape and features are discussed in \cite{MuonUSP}. This distribution gives the position where the charged pions decay, which means that it tracks the longitudinal development of the hadronic cascade. The \textit{apparent} MPD-distribution, the one that is experimentally accessible, corresponds to the distribution of the production depths of muons that arrived at a given location on the ground after propagation, and thus would be affected by changes in the energy spectrum. The depth at which the apparent MPD distribution reaches the maximum is denoted as $X^{\mu}_{max}$. Note that, only the apparent $X^{\mu}_{max}$ is experimentally accessible, so in this paper, it will correspond to the apparent maximum.
In this paper, a library of showers simulated with CORSIKA (v7.4100) \cite{CORSIKA} was created, and each individual $F(X, E_i, cp_t)$ distribution recorded. The library contains samples of 100 showers for each primary (proton and iron) at zenith angles ($\theta$) $0^\circ$, $20^\circ$, $40^\circ$, $60^\circ$ and $70^\circ$, and energy $10^{19}$ eV, with the models EPOS-LHC\xspace and QGSJ{\textsc{et}}-II.04\xspace.
The kinetic energy cuts were set to 0.05 GeV for muons and hadrons,
and the altitude was set to 1400 m a.s.l., with the magnetic field of Malarg\"{u}e, Argentina, site of the Pierre Auger Observatory.
\FloatBarrier
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_hdNdlogE_X2_Log_v2}
\caption{ Muon Spectrum $f_{X}(E_i)$ }
\label{fig: MuSpectrum a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_hdNdlogEF_X2}
\caption{ $f_{X}(E_i)/f_{X,p, EPOS}(E_i)$}
\label{fig: MuSpectrum b}
\end{subfigure
\end{adjustwidth}
\caption[]{Muon production spectrum at the depth $X\in \left[X^\mu_{max} -25,X^\mu_{max}+25\right] \mathrm{ g/cm^{2}}$, in $40^\circ$ showers (a) and fraction of the spectrum with respect to the EPOS-LHC\xspace proton model (b). Results shown for proton (red) and iron (blue), EPOS-LHC\xspace (dashed) and QGSJ{\textsc{et}}-II.04\xspace (full) models and to EPOS-LHC\xspace proton model considering $\delta=-0.1$ (dashed gray) and $\delta=+0.1$ (dashed green), from eq. \ref{eq: dist}.
\label{fig:MuSpectrum}}
\end{figure}
\section{Modification of the muon energy spectrum}
\label{subsection: Muons Energy}
The shape of the transverse momentum distribution of muons within EAS is universal, and does not depend on the primary particle type, zenith angle, or hadronic interaction model \cite{CazonTransportModel}. Therefore we have chosen
to study the effects of changes in the energy of muons at the production point by scaling $| \vec{p}|$ or $E_i$ while maintaining the transverse momentum $p_t$ constant. This necessarily implies a change in the angle of the 3-momentum with respect to the shower axis $\alpha$, given that $\sin \alpha=\frac{cp_t}{E_i}$.
For a given depth $X$, the energy distribution $E_i$ can be given by:
\begin{equation}
f_{X}(E_i) = \frac{\text{d}^2N}{\text{d}E_i\text{d}X} =\int F(X,E_i,cp_t) \text{d} cp_t \: .
\end{equation}
In figure \ref{fig: MuSpectrum a}, $f_{X}(E_i)$ is plotted for the depth $X\in \left[X^\mu_{max} -25,X^\mu_{max}+25\right] \mathrm{ g/cm^{2}}$, while in figure \ref{fig: MuSpectrum b}, the muon spectrum of proton and iron QGSJ{\textsc{et}}-II.04\xspace and iron EPOS-LHC\xspace is compared to the one from proton EPOS-LHC\xspace. The spectrum is harder for iron primaries than for proton primaries and is harder in the case of QGSJ{\textsc{et}}-II.04\xspace.
\begin{comment}
This might be due to lower $X^{\mu}_{max}$ in the QGSJ{\textsc{et}}-II.04\xspace giving an higher critical energy of the pions, given that the shower develops in less dense air in average.
\end{comment}
The high energy tails of the muon spectrum follow a power law $E^{-2.6}$, which is connected to the cascading process of the parent pion\cite{Cazon}.
Thus, a power law-like modification of the energy spectrum is introduced, with the form:
\begin{equation}
f'_{X}(E_i)= f_{X}(E_i) \cdot E_{i}^{\delta} \:.
\label{eq: dist}
\end{equation}
This transformation is able to effectively modify the spectrum making it harder or softer, while maintaining the main characteristics essentially unchanged. The parameter $\delta$ is the number modulating the modification. In figure \ref{fig: MuSpectrum a}, the EPOS-LHC\xspace proton is changed by $\delta=-0.1$ and $\delta=+0.1$, in green and black lines respectively. The corresponding ratios are displayed in figure \ref{fig: MuSpectrum b}. Note that the modification becomes a linear evolution as a function of $\log_{10} E$ as $E^{\delta} \simeq 1+\ln E$. It can be seen that EPOS-LHC\xspace with $\delta=+0.1$ is very similar to QGSJ{\textsc{et}}-II.04\xspace iron, and thus $\delta=\pm0.1$ can be taken as the typical order of magnitude for the uncertainty of the energy spectrum of muons.
Figure \ref{fig: MuSpectrumTheta a} displays the average production energy for both models and compositions. It also includes the average energy for proton EPOS-LHC\xspace changed with $\delta=\pm0.1$. In figure \ref{fig: MuSpectrumTheta b}, we can see the evolution of the average energy with the $\delta$ parameter for the modified proton EPOS-LHC\xspace model (the shadow bands corresponds to the different models and composition).
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_Av_LogE_GProd_v2}
\caption{$\langle\log_{10}(E_{i}/\text{GeV})\rangle$ all muons at production}
\label{fig: MuSpectrumTheta a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_DifferLogE_}
\caption{$\langle\log_{10}(E_{i}/\text{GeV})\rangle\DD{=\pm0.1}-\langle\log_{10}(E_{i}/\text{GeV})\rangle_{\delta=0} $}
\label{fig: MuSpectrumTheta b}
\end{subfigure
\end{adjustwidth}
\caption[]{Average energy at production ($\langle\log_{10}(E_{i}/\text{GeV})\rangle$) of all muons (a), and difference between the average logarithm of the energy (modified minus original), as a function of the $\delta$ parameter (b). The shadowed bands correspond to the different models and compositions, with the respective averages in dashed lines.
\label{fig: MuSpectrumTheta}}
\end{figure}
\subsection{Fast transport model}
\label{subsection: Transport}
Full CORSIKA simulations are time and space consuming. Nevertheless, there is no need to repeat a full simulation each time the energy of muons at production is changed.
Muons can be easily propagated to ground by means of a separate fast transport code that uses the modified distribution $F^\prime(X,E_i,cp_t)=F(X,E_i,cp_t)\cdot E_i^\delta$ as the only input (as demonstrated in \cite{CazonTransportModel}).
The transport code is implemented in two steps: muons are produced along the shower axis at a depth $X$ (which corresponds to a distance from ground $z$), with an energy $E_i$, transverse momentum $p_t$ and a random polar angle, sampled from the $F(X, E_i, cp_t)$-distributions with weight $w_i$. The weight represent a bunch of muons and can be chosen to reduce the computation time. Muons are then propagated to ground following a straight line according to their 3-momentum, and accounting for the continuous energy loss; in a second step, the multiple scattering and magnetic field effects were included, the impact point on ground was corrected, and finally, the energy loss, arrival time delay, and decay probability are re-evaluated. At the end of the fast simulations, the distributions of muons at the ground is obtained as a function of: final weight $w_f$(due to the decay probability), distance to the shower axis $r$ and polar angle $\zeta$ of the impact point at ground, energy at ground $E_f$, and the different contributions to the overall arrival time delay with respect to a plane shower front, among them, the time delay due to subliminal velocities $t_\epsilon$, which is referred as {\it kinematic delay} \cite{CazonTransportModel}.\\
In addition to energy loss processes, multiple scattering, and magnetic field deflections, muons can decay in flight with a certain probability, in which case some information is lost. The latter phenomenon plays a fundamental role in shaping the distributions of muons at the ground: the muon lateral distribution, time distribution and energy spectrum are suppressed in the regions dominated by low energy muons.
In figure \ref{fig: MuSpectrumAll a}, the dashed lines are the energy spectrum for all the produced muons for proton EPOS-LHC\xspace while the full line represents the energy distributions of muons at production including only those that reach the ground. Most muons are produced below 3 GeV (peaking at $\sim0.3$ GeV). A significant fraction of these low energy particles decay before reaching the ground while higher energy muons will mostly survive.
From figure \ref{fig: MuSpectrumAll b} it can be seen that at the ground the average $\langle\log_{10}(E_{i}/\text{GeV})\rangle$ for $\delta=+0.1$ with EPOS-LHC\xspace proton is similar to iron EPOS-LHC\xspace. This smaller effect, with respect to the effect of $\delta=+0.1$ at production, can be understood through atmospheric energy losses and muon decays.
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_hEnergyAll}
\caption{$\langle\log_{10}(E_{i}/\text{GeV})\rangle$ all muons }
\label{fig: MuSpectrumAll a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_Av_LogE_GProdInGround_v2}
\caption{$\langle\log_{10}(E_{i}/\text{GeV})\rangle$ arriving at ground}
\label{fig: MuSpectrumAll b}
\end{subfigure
\end{adjustwidth}
\caption[]{Muon spectrum distribution of all produced muons (in dashed lines) and the ones arriving at the ground (full lines), for a simulation of proton showers using EPOS-LHC\xspace model, at $20^{\circ}$ zenith angle, in (a). In (b), the average energy at production is plotted for muons arriving at ground.
\label{fig:MuSpectrumAll}}
\end{figure}
\section{Number of muons on the ground}
\label{section: Number of muons}
Near the shower core, cosmic ray detectors can be saturated, due to the high particle density. Far from the core, the limited size of the detectors limits the sensitivity to very low particle densities (particle per unit area). Therefore, only muons between 500 m and 2000 m are considered, and a truncated number of muons is defined as $N_\mu=\int \int_{500}^{2000} \rho_\mu(r) r dr d\zeta$. The number of muons at the ground is plotted in figure \ref{fig: MuNintegral a}. As expected, the iron samples and the EPOS-LHC\xspace samples have more muons than proton or QGSJ{\textsc{et}}-II.04\xspace samples, respectively.
From this figure it is also possible to see that a change in the proton EPOS-LHC\xspace spectrum by $\delta=\pm0.1$ results in a modification of the number of muons at ground that still remains within the values for proton and iron with the QGSJ{\textsc{et}}-II.04\xspace model.\\
In figure \ref{fig: MuNintegral b}, the ratio of the muon number at ground between the modified and original spectrum distributions (ratio=$N_{\mu}(\delta)/N_{\mu}$) for each primary and model can be seen. Notice that results are similar for the same value of $\delta$ in all primary and model samples. Moreover, one important prediction, of the muon energy spectrum modification, is the behaviour with zenith angle. The relative importance of this effect increases with zenith angle. At higher zenith angles, muons will travel, on average, larger distances before reaching the ground, making the decay of low energy muons more important.\\
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_NintegralR_v2}
\caption{$N_{\mu}$ }
\label{fig: MuNintegral a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_NFintegralR}
\caption{$N\DDsm{\mu}{=\pm0.1} / N\DDsm{\mu}{=0}$}
\label{fig: MuNintegral b}
\end{subfigure
\end{adjustwidth}
\caption[]{Number of muons arriving at ground between 500 m and 2000 m (a). Fraction of the number of muons arriving at the ground, between the changed muon spectrum and the initial spectrum, for the different models and compositions (b).
\label{fig: MuNintegral}}
\end{figure}
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_LDF_Th2_v2}
\caption{$\langle \rho_\mu (r)\rangle$ }
\label{fig: LDF a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_LDFFraq_Th2_v2}
\caption{$\langle \rho_\mu (r)\rangle\DDsm{ }{=\pm0.1}/\langle \rho_\mu (r)\rangle_{\delta=0} $}
\label{fig: LDF b}
\end{subfigure
\end{adjustwidth}
\caption[]{Lateral muon distribution on ground at $40^\circ$ (a) and ratio between the changed (with $\delta=\pm0.1$) and unchanged LDF, for all models and compositions (b).
\label{fig: LDF}}
\end{figure}
The average \gls{ldf} of the muons over a given sample of showers with the same primary and zenith angle can be built as the average number of muons per unit area at ground as a function of the perpendicular distance to the shower axis $r$. Figure \ref{fig: LDF} shows an example for $40^\circ$ showers. From figure \ref{fig: LDF b} it is possible to see that variations $\delta=\pm0.1$ induces a change of the order of 10\% in the absolute muonic LDF. Moreover, it can be noted that the LDF is more sensitive to changes in the muon energy spectrum for lower distances to the shower core.
A modified NKG\cite{Nishimura} function
\begin{equation}
f_{LDF}(r,\rho_{1000},\beta_{\mu}) =\rho_{1000}\left(\frac{r}{r_{opt}}\right)^{\beta_{\mu}} \left(\frac{r+r_{scale}}{r_{opt}+r_{scale}}\right)^{\beta_{\mu}} \:,
\label{eq: LDF}
\end{equation}
with $r_{opt}=1000$ m and $r_{scale}=700$ m, was chosen to fit the average LDF, where $\rho_{1000}$ (normalization) and $\beta_{\mu}$ (LDF {\it slope}) were left as free parameters. The range of distances to the core allowed in the fit ($r\in$[500,2000] m) justifies this particular choice of only one degree of freedom for the shape parameter $\beta_{\mu}$.
The results for $\beta_{\mu}$ are plotted in figure \ref{fig: MuRhoBeta}. The slope is very similar in all samples, and the differences around $\sim5\%$ come essentially from the composition and not from the model. The effect of changing the muon spectrum is similar for all models and primaries and is around $\sim2\%$.
It is interesting to see that the $\beta_{\mu}$ parameter is not very sensitive to a change in the spectrum, while the total number of muons is more sensitive. More important is that the dependence of the muon number with zenith angle changes with the energy spectrum modification, which is a distinctive behaviour, while $\beta_{\mu}$ is constant.
These features could be used to identify discrepancies between data and simulations regarding the muon energy spectrum. \\
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_RhoBeta_v2}
\caption{$\beta_{\mu}$ }
\label{fig: MuRhoBeta a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_RFRhoBeta}
\caption{$\beta\DDsm{\mu}{=\pm0.1}/\beta\DDzMU{\mu}$}
\label{fig: MuRhoBeta b}
\end{subfigure
\end{adjustwidth}
\caption[]{The $\beta_{\mu}$ parameter of eq. \ref{eq: LDF}, fitted to the lateral distribution of muons on the ground (a). Comparison between the changed spectrum models and unchanged models, $\beta\DDsm{\mu}{=\pm0.1}/\beta\DDsm{\mu}{=0}$ (b).
\label{fig: MuRhoBeta}}
\end{figure}
\FloatBarrier
\begin{figure}[!b]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_XmaxApp_v2}
\caption{$\langle X_{max}^{\mu}\rangle$}
\label{fig: MuXmaxApp a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_FXmaxApp}
\caption{$\langle X_{max}^{\mu}\rangle\DD{=\pm0.1} - \langle X_{max}^{\mu}\rangle_{\delta=0} $}
\label{fig: MuXmaxApp b}
\end{subfigure
\end{adjustwidth}
\caption[]{Apparent $X_{max}^{\mu}$ seen between 1700 m and 4000 m (a). Effect of changing the spectrum between changed and unchanged models, as $\langle X_{max}^{\mu}\rangle\DD{=\pm0.1} - \langle X_{max}^{\mu}\rangle_{\delta=0} $ (b).
\label{fig: MuXmaxApp}}
\end{figure}
\subsection{Maximum of the muon production depth, $X_{max}^{\mu}$, and arrival time delay}
\label{section: ApparentXmax}
The apparent MPD-distribution is made from muons arriving at a particular region at the ground. Thus, only those muons that subtend an angle $\alpha$ compatible with the solid angle contained by the observation region to the production point will contribute to the apparent MPD-distribution. Notice that since $\sin \alpha=cp_t / E_i$, this automatically selects a fraction of the muons that were produced. Moreover, decay and other propagation effects also modify the shape of the MPD-distribution depending on the energy at which muons were produced. Altogether, despite having the same {\it total} MPD-distribution, $h(X)$, a change in the energy spectrum effectively changes the {\it apparent} MPD-distribution which is observed at ground.
The $X_{max}^\mu$ is plotted in figure \ref{fig: MuXmaxApp a} for the different studied samples. In this plot, only the muons with distances between 1700 m and 4000 m to the core were tracked\footnote{We have also verified that the relative changes in the range 1000-4000 m are practically the same with less than $\sim1.5 \text{ g/cm}^2$ difference}. In this range, there are several equivalent energy thresholds for muons due to different distances crossed in the atmosphere, but it is enough to study the overall differences.
The maximum is reached higher for iron primaries and in the EPOS-LHC\xspace model compared to the QGSJ{\textsc{et}}-II.04\xspace.
The modification in the energy spectrum with $\delta=+0.1$ changes the apparent MPD approximately $-10 \text{ g/cm}^2$ at $20^\circ$ while there is no effect at $60^\circ$. The effect appears again for higher zenith angles but with an opposite sign with respect to vertical showers ($\sim+5\text{ g/cm}^2$ at $70^\circ$).
Since the average energy of muons at production decreases with the increase of the depth, the modification of the energy spectrum by a factor $E^\delta$ increases the importance of high energy muons, which are produced in early stages of the shower. This makes $X_{max}^\mu$ to become smaller. On the other hand, for very inclined showers, for instance those with $\theta = 70^\circ$, all low energy muons are highly suppressed due to decays, and therefore this effect is not present anymore and higher order effects begin to give non-negligible contributions. These effects are essentially related with the angle of emission of the muons.
Variations of the energy spectrum produce a net change in the angular dependence of the $X_{max}^\mu$ of the apparent MPD of the order of $X_{max}^\mu(60^\circ)-X_{max}^\mu(40^\circ)\sim 10 \text{ g/cm}^2$.
In \cite{CazonTransportModel,Cazon2004} it is shown that there is a one-to-one correspondence between the delay accumulated with respect to a shower front plane, the {\it geometrical delay}, $t_{g}\simeq \frac{1}{2}\frac{r^2}{z}$, and the production depth of each individual muon through its production distance to the ground $z$. Furthermore, this is what is used by the Auger Collaboration \cite{cazon2005} to experimentally reconstruct the apparent MPD distribution of single events\cite{MPD2014,Laura}. A direct calculation shows that
the time delay changes by $\sim 5$ ns at 40 degrees and 1000 m from the shower core.
The {\it kinematic delay}, $t_\epsilon$, acts as a second order correction to the total arrival time delay, and becomes less important as the distance to the shower core increases. Experimentally, it must be subtracted from the total time delay in order to access the geometric time delay. We have found that changes in the energy spectrum of muons produce a shift which is practically constant as a function of the distance to the shower core which is of the order of $\sim \pm$ 3 ns for $\delta=\mp0.1$ (at 1700 m, see figure \ref{fig: timeE 1700}). This would correspond to a bias on the reconstructed $X^\mu_{max}$ of the order of $\sim11 \text{ g/cm}^{2}$ at 60 degrees at $r=1000$ m or $\sim4 \text{ g/cm}^{2}$ at 1700 m.
\begin{figure}[!h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_hpTimetE_logr1700}
\caption{$\langle t_{\epsilon} \rangle$ at 1700 m}
\label{fig: timeE 1700 a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_FhpTimetE_logr1700}
\caption{$\langle t_{\epsilon} \rangle\DD{=\pm0.1} / \langle t_{\epsilon}\rangle_{\delta=0} $}
\label{fig: timeE 1700 b}
\end{subfigure
\end{adjustwidth}
\caption[]{Kinematic delay time $t_{\epsilon}$ observed at 1700 m (a). Effect of changing the spectrum seen in $\langle t_{\epsilon}\rangle\DD{=\pm0.1} - \langle t_{\epsilon}\rangle_{\delta=0} $ (b) at 1700 m.\label{fig: timeE 1700}}
\end{figure}
\begin{comment}
\begin{figure}[!b]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_hpTimetE_logr1000}
\caption{$\langle t_{E} \rangle$ at 1000 m}
\label{fig: timeE 1000 a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_FhpTimetE_logr1000}
\caption{$\langle t_{E}\rangle\DD{=\pm0.1}\langle t_{E}\rangle_{\delta=0} $ at 1000 m}
\label{fig: timeE 1000 b}
\end{subfigure
\end{adjustwidth}
\caption[]{Cinamatic arrival time $t_{E}$ seen between 1000 m (a). Effect of changing the spectrum with eq. \ref{eq: dist}, $\langle t_{E}\rangle\DD{=\pm0.1} - \langle t_{E}\rangle_{\delta=0} $ at 1000 m (b).\label{fig: timeE 1000}}
\end{figure}
\end{comment}
\section{Signal in water Cherenkov detectors}
\label{section: Signal of muons}
In the previous section, the numbers of muons at ground were directly considered. Nevertheless, water Cherenkov detectors, like the ones used at the Pierre Auger Observatory or Ice Top, record signals with a non-negligible dependence on the energy of the muon. On one hand the emitted Cherenkov light depends on the velocity of the particle, until a plateau. On the other hand, high energy particles can also produce $\delta$-rays which generate an electromagnetic cascade inside of the tank giving additional contribution to the recorded signal.
The average signal emitted by a muon is drawn in figure \ref{fig: MuEnergyGround a}. The equation that translates the muon kinetic energy into signal (in arbitrary units) was obtained considering that the muon traverses vertically a volume of water with $1.2$ m thickness (similar to what happens for the Pierre Auger detectors). Two physical phenomena were considered: Cherenkov light emission and $\delta$-ray production. The first one has an explicit dependence on the muon velocity, $\beta=v/c$ and is represented in the figure by the blue dotted line. During the passage through the water the muon will lose on average $240\,$ MeV. This amount of energy was removed from the muon kinetic energy before computing the Cherenkov yield. Note that once the muon becomes \emph{relativistic} the number of emitted Cherenkov photons becomes independent of the energy of the particle. The $\delta$-ray effect was included following a parametrisation motivated from \cite{deltaRay}. For muons, this effect becomes more noticeable above $1\,$GeV and the contribution to the recorded signal increases logarithmically its energy.
In this section we calculate what would be the sensitivity to the changes on the energy spectrum of muons at production, of the signal observed by detectors similar to those used in the Auger Observatory.
In figure \ref{fig: MuEnergyGround b}, the muon energy at ground is shown in dashed lines and in full lines the same distribution weighted by the signal that muons leave in the water is drawn (from figure \ref{fig: MuEnergyGround a}). The less energetic muons do not leave signal, while the higher energy tail is increased.
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/VEMvalues2}
\caption{Muon $\langle Signal \rangle$ on water}
\label{fig: MuEnergyGround a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_hEnergyGround}
\caption{Energy distribution on ground}
\label{fig: MuEnergyGround b}
\end{subfigure
\end{adjustwidth}
\caption[]{Average signal of muons in water as function of the muon energy (a). Distribution of muons at the ground and weighted by the muon signal in dashed and full lines, respectively (b).
\label{fig: MuEnergyGround}}
\end{figure}
The size parameter $S_{1000}$ (equivalent to $\rho_{1000}$), which is proportional to the total number of muons is displayed in figure \ref{fig: MuS1000 a}. Also here a very similar behaviour to that of the total number of muons (from figure \ref{fig: MuNintegral}) is observed.
The average \gls{ldf}, similar to figure \ref{fig: LDF}, can be built for the muonic signal at the ground and fitted again using eq. \ref{eq: LDF}. The slope $\beta_S$ and the size parameter $S_{1000}$ are plotted in figures \ref{fig: MuBeta} and \ref{fig: MuS1000}, respectively.
The effect of changing the energy spectrum by $\delta$, on the parameter $\beta_S$ is similar to the one found in the previous section.
In summary, despite the dependence of the signal recorded by the water Cherenkov detectors with energy, all the tested observables present sensitivities to the muon energy spectrum that are similar to those found by considering directly the muons at the ground.
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_SBeta_v2}
\caption{$\beta_{S}$ }
\label{fig: MuBeta a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_SFBeta}
\caption{$\beta\DDsm{S}{=\pm0.1}/\beta\DDzMU{S}$}
\label{fig: MuBeta b}
\end{subfigure
\end{adjustwidth}
\caption[]{The $\beta_{S}$ parameter of eq. \ref{eq: LDF} for the signals on ground (a).Comparison between the changed spectrum models and unchanged models, $\beta\DDsm{S}{=\pm0.1}/\beta\DDsmZ{S}$ (b).
\label{fig: MuBeta}}
\end{figure}
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_S1000_v2}
\caption{$S_{1000}$ }
\label{fig: MuS1000 a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_G_SFS1000}
\caption{$S\DDsm{1000}{=\pm0.1}/S\DDzMU{1000} $ }
\label{fig: MuS1000 b}
\end{subfigure
\end{adjustwidth}
\caption[]{The $S_{1000}$ parameter of eq. \ref{eq: LDF} for the signals at the ground (a). Comparison between the changed spectrum models and unchanged models, $S\DDsm{1000,\mu}{=\pm0.1}/S\DDsmZ{1000,\mu}$ (b).
\label{fig: MuS1000}}
\end{figure}
\FloatBarrier
\section{Sensitivity of the variables to the $\delta$ modification}
\label{section: Sensitivity}
In this section the effect of changing the muon energy spectrum with eq. \ref{eq: dNdXdE} is shown for different $\delta$ parameters.
In figure \ref{fig: DELTA1 a}, the ratio $\frac{x\DD{}}{x\DD{=0} }$ is plotted for $x=\beta_{\mu}$, $\rho_{1000}$, $\beta_{S}$, $S_{1000}$ and $N_{\mu}$, at $\theta=40^\circ$. Both $\beta_{\mu}$ slopes change very little with the parameter $\delta$. As seen before, the size parameters change around $-35\%$ at $\delta=-0.4$ to $+20\%$ around $\delta=+0.4$. We saw that the effect on $\delta$ depends on the zenith angle, so in the figure \ref{fig: DELTA1 b} the parameter $\Delta_{20^\circ}^{60^\circ}\left(\delta\right)=\frac{x_{\delta }}{x_{\delta=0}}(60^\circ) - \frac{x_{\delta }}{x_{\delta=0}}(20^\circ)$ is plotted. For the size parameters, the modification affects more the results at $60^\circ$ than at $20^\circ$.
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_Differences_Th2}
\caption{ratio $\frac{x\DD{ }}{x_{\delta=0} }$ }
\label{fig: DELTA1 a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_DifferencesDelta_}
\caption{$\Delta_{20^\circ}^{60^\circ}\left(\delta\right)$ }
\label{fig: DELTA1 b}
\end{subfigure
\end{adjustwidth}
\caption[]{The effect of changing $\delta$ in the ratio $\frac{x\DD{}}{x_{\delta=0} }$ for the parameters described at $\theta=40^\circ$ (a) and the value $\Delta_{20^\circ}^{60^\circ}\left(\delta\right)=\frac{x_{\delta }}{x_{\delta=0}}(60^\circ) - \frac{x_{\delta }}{x_{\delta=0}}(20^\circ)$ as function of $\delta$ (b).
The shadow bands contain all compositions and models used and the line is their respective average value.
\label{fig: DELTA1}}
\end{figure}
\subsection{Effect on the MPD}
The energy modification on the apparent $X_{max}^{\mu}$ changes drastically with zenith angle, as can be seen in figure \ref{fig: DELTA2 a}. At lower zenith angles ($\theta<60^\circ$), the apparent maximum on the MPD increases for negative $\delta$ and decreases for positive $\delta$. At $70^\circ$, we can see the contrary behaviour and at $\sim60^\circ$ the maximum is almost constant with $\delta$. This behaviour comes from different characteristic of the muon production and transport. The number of muons in the shower depends on its stage of development which is less sensitive to changes on the zenith angle for more inclined showers.
Additionally, the modification of the muon energy spectrum changes the relative importance between low energy and high energy muons (by altering the muon survival probability) producing distortions in the MPD profile that lead to subsequent differences in the $X_{max}^{\mu}$.
Finally, for more inclined showers, the propagation effects become more important and at $70^\circ$, more muons survive with positive than with negative $\delta$.
\begin{figure}[h]
\centering
\hspace{-0.5cm}
\begin{adjustwidth}{-0.50cm}{-0.50cm}
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_DifferXmax_}
\caption{$\langle X_{max}^{\mu}\rangle\DD{}-\langle X_{max}^{\mu}\rangle_{\delta=0} $ }
\label{fig: DELTA2 a}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}\centering
\includegraphics[width=1\textwidth]{pics/FinalPlots/Plots_h_DifferencesDeltaXmax_}
\caption{$\Delta_{40^\circ}^{60^\circ}\left(\delta\right) X_{max}^{\mu}$ }
\label{fig: DELTA2 b}
\end{subfigure
\end{adjustwidth}
\caption[]{The difference between $X_{max}^{\mu}$ from the modified and the unmodified spectrum, $\langle X_{max}^{\mu}\rangle\DD{}-\langle X_{max}^{\mu}\rangle_{\delta=0} $ (a) and the value $\Delta_{40^\circ}^{60^\circ}\left(\delta\right)= \langle X_{max}^{\mu}\rangle\DD{}(60^\circ) - \langle X_{max}^{\mu}\rangle\DD{}(40^\circ)$ as function of $\delta$ (b).
The shadow bands contain all compositions and models used and the line is their respective average value.
\label{fig: DELTA2}}
\end{figure}
\FloatBarrier
\section{Summary}
\label{section:Conclusions}
We have studied the impact of changing the muon energy spectrum at production on EAS observables related to the muon shower component.
The muon content at ground increases between 5\% at $20^\circ$ zenith angle to 17\% at $60^\circ$ zenith angle when the muon spectrum is changed by $E^{\delta}$, with $\delta=+0.1$. This introduces an additional dependence on the observed number of muons at ground with respect to the zenith angle, that might be measured by experiments.
The slope of the average LDF $\beta_{\mu}$, measured through a modified NKG function for $r \in [500;2000]$ m, is relatively insensitive to variations in the energy spectrum, changing by around 2\% for $\delta=+0.1$.
The maximum of the apparent MPD $X^\mu_{max}$ changes by $10\text{ g/cm}^{2}$, and the kinematic delay changes by 3 ns independently of the distance to the core. In the MPD reconstruction algorithms, the kinematic delay is usually parametrized from models and subtracted from the total arrival time delay in order to access the pure geometric transformation from arrival times into depth. This could introduce a bias on the reconstructed apparent $X^\mu_{max}$ of the order of $\sim 5\text{ g/cm}^{2}$ at 60 degrees and 1700
m from the core, or $10\text{ g/cm}^{2}$ at 1000 m.
The parameters studied in this work are summarised in table \ref{tab: Results}, for proton EPOS-LHC\xspace simulated at $\theta=40^\circ$. Since the effect of the modified spectrum is different with zenith angle, we define the variable $\Delta_{20^\circ}^{60^\circ}(\delta)=\frac{x_{\delta}}{x_{\delta=0}}(60^\circ) - \frac{x_{\delta}}{x_{\delta=0}}(20^\circ)$, where $x_{\delta=0}$ is the parameter without modification, and it is also given in table\ref{tab: Results}.\\
It should be kept in mind, that the large number of muons on data might be caused simply by an increase in the overall number of produced muons. In that case, the shape of the muon content should be just a normalization factor at some energy for all angles. Nonetheless, if some different effect plays a role (and also due to the propagation effects in the atmosphere) the final muon normalization factor might depend on the shower zenith angle. According to \cite{GlennysICRC13}, the excess on the total signal seen in data of the Auger Observatory changes with zenith angle. This effect contains also the information about different muonic/electromagnetic ratios per angle, but it might suggest a difference in the muon normalization. In the near future, with the upgrade of the Auger Observatory(AugerPrime)\cite{AugerPrime}, it might be possible to clarify this feature and understand where the differences in the muon component of air showers come from.
\begin{table}[h]
\centering{}
\caption[]{Summary of the variations of the parameters for proton EPOS-LHC\xspace$\left(\frac{x_{\delta}}{x_{\delta=0}}\right)$, at $\theta=40^\circ$ under $\delta=\pm0.1$ modification of the spectrum, and the angular dependence of such variations as $\Delta_{20^\circ}^{60^\circ}( \delta=\pm0.1)=\frac{x_{\delta}}{x_{\delta=0}}(60^\circ)-\frac{x_{\delta}}{x_{\delta=0}}(20^\circ)$ for all variables except MPD and time, which is calculated between 40$^\circ$ and 60$^\circ$.
}
\begin{tabular}{ll|l|ll}
\toprule \toprule
& Values & Variation ($\delta=\pm 0.1$) & Angular dependence ($\delta=\pm0.1$) & Range\\
\tabularnewline \bottomrule
$N_\mu$ & $1.57\times10^{7}$ & $\pm$8\%& $\pm$11.0\% & 20-60$^\circ$\\
$S_{1000}[\text{a.u.}]$ & 1.44 & $\pm8\%$ & $\pm$10\% & 20-60$^\circ$\\
$\beta_\mu$ & 1.99 & $\pm$1.6\% & $\mp$0.7\% & 20-60$^\circ$\\
$\beta_S$ & 2.06 & $\pm$1.5\% & $\pm$ 0.6\% & 20-60$^\circ$\\
$X_{max}^{\mu}[\text{g/cm}^{2}]$& 611 & $\mp$10 & $\pm$10 & 40-60$^\circ$\\
$t_\epsilon$ [ns] & 140 & $\mp$3 & $\pm$0 & 40-60$^\circ$\\
\tabularnewline
\bottomrule \bottomrule
\end{tabular}
\label{tab: Results}
\end{table}
\normalsize
\section*{Acknowledgements}
We would like to thank J. Alvarez-Mu\~niz, C. Dobrigkeit, M. C. Esp\'irito Santo and M. Pimenta for carefully reading this manuscript. The authors wish also to thank IF/00820/2014/CP1248/CT0001, SFRH/BPD/73270/2010, SFRH/BD/89109/2012, OE, FCT-Portugal and CERN/FIS-NUC/0038/2015 for financial support.
\section*{References}
|
\section{Introduction}
\label{sec:1}
The availability of data from collider, direct and indirect searches for dark matter has raised the importance of dark matter complementarity across these search strategies. In this context, effective field theories and simplified models have become popular tools, as they can capture most of the dark matter phenomenology.
Planck measurements of the power spectrum of the cosmic microwave background radiation infer that the cold dark matter abundance should be around 27\% ($\Omega_{DM} h^2=0.12$), where $h$ is a parameter that accounts for uncertainties in the Hubble rate \cite{Ade:2013zuv}. This alone strongly constrains the viable parameter space of dark matter models.
The observation of cosmic rays and gamma rays also offers a compelling probe for dark matter \cite{Hooper:2012sr,Bringmann:2012ez,Calore:2013yia,Kopp:2013eka,Galli:2013dna,Gomez-Vargas:2013bea,Berlin:2013dva,Madhavacheril:2013cna,
Abazajian:2014fta,Bringmann:2014lpa,Gonzalez-Morales:2014eaa,Buckley:2015doa,Huang:2016pxg}. In particular, the Fermi-LAT sensitivity to continuous gamma-ray emission from dark matter annihilations taking place in Dwarf Galaxies resulted in restrictive bounds in the annihilation cross section today, namely $\sigma v < 3 \times 10^{-26} {\rm cm^3/s}$ for masses of 80 GeV and annihilation into $b\bar{b}$ quark pairs~\cite{Ackermann:2015zua}. This rules out a multitude of light WIMP (weakly interacting massive particles) models in which velocity-independent interactions occur.
Moreover, underground detectors using liquid XENON, such as XENON \cite{Aprile:2012nq} and LUX \cite{Akerib:2013tjd} that use scintillation and ionization measurements to discriminate signal from background events, observed no excess, leading to the exclusion of spin-independent WIMP-nucleon scattering cross sections larger than $10^{-45} {\rm cm^2}$ for WIMP masses of $50$~GeV. Other experiments have placed complementary limits in particular at lower masses such as SUPERCDMS, which uses Ge targets \cite{Agnese:2015nto}. The ongoing XENON1T \cite{Aprile:2015uzo} and LZ \cite{Malling:2011va} experiments are expected to bring down the limits by roughly two orders of magnitude in the absence of signal and zero background events.
Besides the indirect and direct detection probes, the Tevatron \cite{Bai:2010hh} and the LHC \cite{Goodman:2010ku,Fox:2011pm} have proven to be great laboratories to test dark matter models. In the case where the dark and visible sectors are connected by vector mediators, dijet \cite{An:2012va,Frandsen:2012rk,Alves:2013tqa,Fairbairn:2016iuf} and dilepton \cite{Profumo:2013sca,Alves:2015pea,Alves:2015mua,Allanach:2015gkd,Kahlhoefer:2015bea} bounds are by far the most stringent constraints.
Dark matter phenomenology is then dictated by gauge interactions which are determined, once the gauge group behind the origin of the vector mediator is known. The common approach is to consider simplified lagrangians that encompass both Dirac and Majorana dark matter fermions and then to compute dark matter observables; namely, relic density, annihilation and scattering cross sections, the latter being spin-independent and spin-dependent for Dirac and Majorana fermions, respectively\footnote{Dirac fermions also induce spin-dependent interactions but the spin-independent ones lead to stronger constraints.}. The simplified dark matter model approach is interesting, intuitive and serves as a guide for future work. However, they might lead to different results once embedded in a complete theory.
In the context of the B-L model, dark matter scenarios have been previously investigated. In~\cite{Li:2010rb}, the authors discussed the radiative see-saw mechanism to account for neutrino masses and focused exclusively on dark matter abundance. Supersymmetric B-L extensions~\cite{Khalil:2008ps,Burell:2011wh,Basso:2012gz} and a conformal approach~\cite{Okada:2012sg} have also been investigated. Even though later disfavored in~\cite{Mambrini:2015sia}, a global B-L symmetry has been proposed~\cite{Baek:2013fsa}. In~\cite{El-Zant:2013nta} a warm dark matter scenario was investigated.
The possibility of having one of the right-handed neutrinos to be the dark matter candidate was entertained in \cite{Sahu:2005fe,Basak:2013cga,Kaneta:2016vkq}, whereas in \cite{Rodejohann:2015lca} an additional scalar played this role. This extra scalar dark matter was also investigated in \cite{Guo:2015lxa}, but in the context of classical scale invariance.
The authors of \cite{Sanchez-Vega:2015qva,Patra:2016ofq} considered an exotic B-L model and advocated the presence of many scalar fields.
Finally, the authors of \cite{Duerr:2015wfa} studied Dirac fermion dark matter in the context of a $U(1)_{B-L}$ symmetry, but with the inclusion of LEP bounds only they discussed gamma-ray lines emissions, which turned out to be irrelevant unless one lives very close to the resonance with a dark matter quantum number under B-L larger than three.
Thus, our work supplements previous studies for the following reasons:
(i) Both fermionic and scalar dark matter realizations are discussed as well as several quantum numbers and gauge couplings options.
(ii) We investigate two-component dark matter scenarios.
(iii) We perform a detailed collider study at next-to-leading order (NLO) plus next-to-leading logarithmic (NLL) accuracy using recent dilepton data from the LHC at 13 TeV, which are often ignored due to the handy LEP limits.
(iv) Finally, the region of parameter space allowed/excluded by limits from the LHC, LEP and indirect detection experiments in dependence of the mass of the mediator, gauge couplings and dark matter mass is presented.
\section{Model}
\label{sec:model}
In the Standard Model, both baryon and lepton numbers are accidental global symmetries. Thus, a natural extension of the SM consists of gauging both quantum numbers. However, only combinations of B-L are free of triangle anomalies. Interestingly, the gauge anomalies $Tr(U(1)_{B-L} SU(2)_L^2),\ Tr(U(1)_{B-L} U(1)_Y^2)$ and $Tr(U(1)_{B-L}^3)$ vanish with the introduction of three right-handed neutrinos having charge $(-1)$ under B-L. In addition, this also leads to vanishing gravitational anomalies. Therefore, the gauged B-L symmetry naturally addresses neutrino masses through see-saw mechanisms~\cite{Mohapatra:1979ia,Minkowski:1977sc,Schechter:1980gr,Mohapatra:1980yp,Lazarides:1980nt,Keung:1983uu}. There are several ways to accommodate dark matter without spoiling the anomaly cancellation, namely:
\begin{enumerate}[(i)]
\item{ {\it \bf Dirac Fermion Dark Matter - $Z^{\prime}$ Portal with unbroken B-L}: This model introduces a vector-like Dirac fermion charged under $U(1)_{B-L}$ leaving the B-L symmetry unbroken. Dark matter phenomenology is then governed by the $Z^{\prime}$ portal. The new gauge boson mass is generated through the Stueckelberg mechanism, which leads to the following Lagrangian~\cite{Feldman:2006wb,Feldman:2007wj,Feldman:2011ms,Heeck:2014zfa}:
\begin{eqnarray}
&&\mathcal{L} \supset \bar{\chi}\gamma^{\mu}D_{\mu}\chi -M_{\chi}\bar{\chi}\chi -\frac{1}{4} F^{\prime \mu\nu}F_{\mu\nu}^{\prime} -\frac{1}{2} M_{Z^{\prime}}Z_{\mu}^{\prime}Z^{\prime \mu}+ g_{BL}\sum_{i=1}^3 (\bar{l}\gamma_{\mu}l+\bar{\nu}_i \gamma_{\mu}\nu_i)Z^{\prime\mu}\nonumber\\
&&+ \frac{g_{BL}}{3}\sum_{i=1}^6(\bar{q_i}\gamma_{\mu}q_i)Z^{\prime\mu} \nu + y_{ij}\bar{L_i}\tilde{\phi}\nu_{jR}\, ,
\label{Lagrangian1}
\end{eqnarray}
where $D_{\mu}\chi=(\partial_{\mu} +i g_{BL} n_{\chi}Z_{\mu}^{\prime})\chi$. We denote by $\tilde{\phi}$ the isospin transformation of the Higgs doublet, $\phi=(\phi^+, \phi^0)^T$, defined as $\tilde{\phi}= i\sigma_2 \phi$. The dark matter charge, $n_{\chi}$, should be different from $\pm 1$ to prohibit an additional Yukawa term involving $\chi_R$, that would lead to dark matter decay. Note that $M_{Z^{\prime}}$ is not determined by the B-L symmetry and that the right-handed neutrinos acquire mass through the usual Yukawa term. Consequently, the neutrinos are Dirac fermions with their small masses being obtained via suppressed Yukawa couplings. We emphasize that the dark matter stability is guaranteed by B-L symmetry.
}
\item { {\it \bf Dirac Fermion Dark Matter - $Z^{\prime}$ Portal with broken B-L}:
In this scenario one adds a SM singlet scalar, $S$, carrying charge 2 under the B-L symmetry. Dark matter is realized via a vector-like Dirac fermion $\chi$ as follows:
\begin{eqnarray}
&&\mathcal{L} \supset \bar{\chi}\gamma^{\mu}D_{\mu}\chi -M_{\chi}\bar{\chi}\chi -\frac{1}{4} F^{\prime \mu\nu}F_{\mu\nu}^{\prime} -\frac{1}{2} M_{Z^{\prime}}Z_{\mu}^{\prime}Z^{\prime \mu}+ g_{BL}\sum_{i=1}^3 (\bar{l}\gamma_{\mu}l+\bar{\nu}_i \gamma_{\mu}\nu_i)Z^{\prime \mu}\nonumber\\
&&+ \frac{g_{BL}}{3}\sum_{i=1}^6(\bar{q_i}\gamma_{\mu}q_i)Z^{\prime \mu} + y_{ij}\bar{L_i}\tilde{\phi}\nu_{jR}+ \lambda_S \bar{\nu}_R \nu_R S\, ,
\label{Lagrangian2}
\end{eqnarray}
where $v_{BL}$ is the vev of the singlet scalar $S$ and $M_{Z^{\prime}} = 2 g_{BL} v_{BL}$. This mass term arises after spontaneous symmetry breaking of the B-L symmetry through the scalar $S$. The mass of the new gauge boson is generated through the kinetic term of the scalar.
Interestingly, in this procedure the neutrinos are Majorana particles. The right-handed neutrinos have masses determined by the last term in Eq.\ (\ref{Lagrangian2}), whereas the active neutrinos have their masses generated through the usual see-saw type I mechanism. The dark matter stability in this case is assured by a $Z_2$ symmetry remnant from the B-L spontaneous symmetry breaking.
Another possibility would be to give different charges to the three right-handed neutrinos such as $(5,-4,-4)$, which is still anomaly-free. However, several extra fields are then needed to successfully generate neutrino masses~\cite{Ma:2015mjd}. For other different studies based on the B-L gauge symmetry see \cite{Perez:2009mu,Majee:2010ar,Li:2010rb,Basso:2011hn,Okada:2014nea,Abdelalim:2014cxa,Ovrut:2014rba,Banerjee:2015hoa,Dong:2015yra,
Gardner:2016wov,Hammad:2016trm}.
}
\item { {\it \bf Scalar Dark Matter - $Z^{\prime}$ Portal}:
Scalar dark matter in the context of B-L symmetry is also a plausible alternative to accommodate dark matter, since it requires only two new fields: a singlet scalar $S$, with charge $+2$ under B-L, and a scalar $\phi$, as dark matter, which should be charged under B-L with a quantum number different from multiples of $\pm 2$ for stability purposes~\cite{Rodejohann:2015lca}. Taking this into account, the Lagrangian of this model reads
\begin{eqnarray}
&&\mathcal{L} \supset \mu_S S^{\dagger}S + \frac{\lambda_S}{2} (S^{\dagger}S)^2 + \mu_{\phi}^2 \phi^{\dagger}\phi + \frac{\lambda_{\phi}^2}{2} (\phi^{\dagger}\phi)^2\nonumber\\
&& +\lambda_1 (\phi^{\dagger}\phi)(H^{\dagger}H)+ \lambda_2 (S^{\dagger}S)(H^{\dagger}H)+\lambda_3 (\phi^{\dagger}\phi)(S^{\dagger}S) + g_{BL}\sum_{i=1}^3 (\bar{l}\gamma_{\mu}l+\bar{\nu}_i \gamma_{\mu}\nu_i)Z^{\prime \mu}\nonumber\\
&&+ \frac{g_{BL}}{3}\sum_{i=1}^6(\bar{q_i}\gamma_{\mu}q_i)Z^{\prime \mu} + y_{ij}\bar{L_i}\tilde{\phi}\nu_{jR}+ \lambda_S \bar{\nu}_R \nu_R S\, .
\label{Lagrangian3}
\end{eqnarray}
The dark matter phenomenology~\cite{Rodejohann:2015lca} is determined by both gauge interactions, $\phi^{\dagger}\phi \rightarrow Z^{\prime} \rightarrow \bar{f}f$, and scalar interactions, $\phi^{\dagger}\phi \rightarrow h \rightarrow \bar{f}f, SS$. In the first case the dark matter phenomenology is strongly related to the gauge coupling and the $Z^{\prime}$ mass. It is very predictive and connected to collider physics. In the second, the scalar potential couplings control dark matter observables and the strong connection to collider physics is lost, such that we will not discuss it further. For a detailed study see e.g.~\cite{Rodejohann:2015lca}.
}
\end{enumerate}
\begin{figure}[!t]
\centering{}
\subfloat[s-channel annihilation process\label{fig:diagram1a}]{
\includegraphics[scale=0.53]{fermionanni}}\hspace{1.5cm}
\subfloat[t-channel annihilation process\label{fig:diagram1b}]{
\includegraphics[scale=0.53]{fermionanni2}}\hspace{1.5cm}
\subfloat[Dark matter-nucleon scattering\label{fig:diagram1c}]{
\includegraphics[scale=0.53]{fermionScatt}}\hspace{1.5cm}
\caption{Dark matter annihilation and dark matter-nucleon scattering processes in the fermion dark matter model, where $f$ stands for all SM fermions and $q$ represents the quarks. The t-channel annihilation process is only relevant for $M_{\chi} > M_{Z^{\prime}}$.}
\label{fig:diagram1}
\end{figure}
\begin{figure}[!t]
\centering{}
\subfloat[s-channel annihilation process\label{fig:diagram2a}]{
\includegraphics[scale=0.55]{scalaranni}}\hspace{2.5cm}
\subfloat[Dark matter-nucleon scattering process\label{fig:diagram2b}]{
\includegraphics[scale=0.57]{scalarScatt}}
\caption{Dark matter annihilation and dark matter-nucleon scattering processes in the scalar dark matter model.}
\label{fig:diagram2}
\end{figure}
\section{Dark Matter Abundance}
The relic abundance of dark matter is determined by solving the Boltzmann equation. The dark matter particle pair annihilates and is pair-produced in equal rate in the early Universe, but as the Universe cools down and expands, eventually the expansion rate approaches the interaction rate, and from then on the dark matter particles are only able to self-annihilate into lighter particles. Eventually, then the expansion rate prevents the dark matter particles from self-annihilating. This episode is referred to as freeze-out. In order words, the abundance of left-over dark matter particles is linked to the annihilation cross section at the freeze-out, which can be very different from the annihilation cross section today ~\cite{Profumo:2013yn}. Thus, the stronger the annihilation cross section is, the fewer remnant dark matter particles subsist today. In what follows, we discuss the abundance of the fermion and scalar dark matter in quantitative terms.
\begin{itemize}
\item {\bf Dirac Fermion}
\end{itemize}
In Figs.~\ref{fig:diagram1a} and \ref{fig:diagram1b}, we show the processes that set the dark matter abundance for the fermion. When $M_{\chi} < M_{Z^{\prime}}$, only the first diagram is relevant. $f$ stands for all SM fermions, including the right-handed neutrinos, whose masses are in the eV range in the case where the B-L symmetry in unbroken, whereas in the broken B-L scenario their masses are kept at $100$~GeV. The precise value for their masses is not relevant, and both cases lead to very similar dark matter phenomenology. For this reason, dark matter observables will be derived without explicitly specifying whether or not the B-L symmetry is broken.
\begin{figure}[!t]
\centering{}
\subfloat[$n=1/3$ and $M_{Z^{\prime}}=4$~TeV\label{fig:abundance1a}]{
\includegraphics[scale=0.45]{Abundance2.pdf}}
\subfloat[$n=1/3,\ g_{BL}=0.1$\label{fig:abundance1b}]{
\includegraphics[scale=0.45]{Abundance1.pdf}}
\caption{{\it \bf Dirac Fermion}. Abundance as a function of mass for various gauge couplings and $Z'$ boson masses. Because the model must satisfy the relic density, it features a strong dependence in the resonance region.}
\label{fig:abundance1}
\end{figure}
In Fig.~\ref{fig:abundance1} we display, for $n=1/3$, the abundance of the fermion as a function of its mass. In the left panel, Fig.~\ref{fig:abundance1a}, the $Z^{\prime}$ mass has been fixed to $4$~TeV and the gauge coupling varied in $g_{BL}\in[0.1,0.8]$, while in the right panel, Fig.~\ref{fig:abundance1b}, we keep $g_{BL}=0.1$ and vary $M_{Z^{\prime}}=2,4,6$~TeV.
From Fig.~\ref{fig:abundance1a}, it is clear that the increase in the coupling widens the resonance and therefore leads to viable dark matter masses away from $M_{Z^{\prime}}/2$, lower or higher. In addition, the larger the coupling, the larger the annihilation rate, leading to smaller abundance. Thus, one needs sufficiently large gauge couplings to enhance the annihilation rate and reach $\Omega h^2 \sim 0.1$. Notice that the resonance condition is not as needed, if couplings close to unity are used. Such large couplings arise naturally in 3-3-1 models \cite{Mizukoshi:2010ky,Alves:2011mz,Alvares:2012qv,Profumo:2013sca,Kelso:2013nwa,Cogollo:2014jia,
Dong:2014wsa,Dong:2014esa,Kelso:2014qka} and left-right models \cite{Mohapatra:1974gc,Mohapatra:1974hk,Senjanovic:1975rk,Mohapatra:1979ia,
Mohapatra:1980yp,Dias:2010vt,Heeck:2015qra,Rodejohann:2015hka,Deppisch:2015cua,Patra:2015vmp}. Other fermion dark matter models feature similar trends \cite{Esch:2013rta,Esch:2014jpa,Dutra:2015vca,Yaguna:2015mva,Ibarra:2016dlb}.
The impact of the $Z^{\prime}$ mass is shown in Fig.~\ref{fig:abundance1b}, which exhibits a series of peaks at different dark matter masses. The larger $M_{Z^{\prime}}$ gets, the heavier the dark matter mass has to be in order to achieve the right abundance. We point out that both results for fermion dark matter are presented for $n=1/3$, but they can be easily rescaled, since the abundance scales as $n^2 g_{BL}^4$. Hence, for constant relic density, a change in $n$ straightforwardly induces a quadratically inverse change in $g_{BL}$.
\begin{itemize}
\item {\bf Scalar Field}
\end{itemize}
In Fig.~\ref{fig:diagram2a} we show the Feynman diagram relevant for determining the scalar dark matter abundance. In Fig.~\ref{fig:abundance2} the abundance for two different charges under B-L, $n=1/3$ and $n=1$, is shown. The kinks in the plots are the result of the $Z^{\prime}$ threshold, i.e. when the scalar can pair annihilate into a $Z^{\prime}$ boson\footnote{This effect is also present, but much less pronounced in the fermion case discussed above.}.
Similarly to the case of fermion dark matter, the s-channel resonance regime $m_{\phi} \sim M_{Z^{\prime}}/2$ is responsible for increasing the annihilation cross section and consequently reducing the abundance to values close to the one inferred by Planck. Fig.~\ref{fig:abundance2c} shows the abundance with $n=1$ and $g_{BL}=0.8$ and for various masses of the new gauge boson, $M_{Z^{\prime}}=2,4,6$~TeV. Again, the effect of increasing $M_{Z'}$ is to simply move the resonance region to higher dark matter masses. It is noticeable that for $ g_{BL}=0.8$ the resonance region is wide enough to accommodate two different dark matter masses yielding the right abundance.
As already mentioned, the annihilation cross section grows as $n^2 g_{BL}^4$.
For $n\ll 1$, one therefore needs gauge couplings larger than one in order to satisfy the relic density constraint. On the other hand, values of $n$ closer to one enhance the dark matter-nucleon scattering rate thus severely restricting the model, as we shall see below.
As a summary, we have seen in this part that both Dirac fermions and scalars can be viable dark matter candidates of the Universe as long as the annihilation rate occurs not very far from the resonance.
\begin{figure}[!t]
\centering
\subfloat[$M_{Z^{\prime}}=2$~TeV and $n=1/3$ for $g_{BL}=0.4,0.8,1$.\label{fig:abundance2a}]{
\includegraphics[scale=0.45]{scalarDMplot1.pdf}}\hfill
\subfloat[$M_{Z^{\prime}}=2$TeV and $n=1$ for $g_{BL}=0.4,0.8,1$.\label{fig:abundance2b}]{
\includegraphics[scale=0.45]{scalarDMplot2.pdf}}\\
\subfloat[$g_{BL}=0.8$ and $n=1$ for $M_{Z^{\prime}}=2,4,6$~TeV.\label{fig:abundance2c}]{
\includegraphics[scale=0.45]{scalarDMplot3.pdf}}
\caption{ {\it \bf Scalar Field}. Abundance as a function of mass. The kinks in the plots are the result of the $Z^{\prime}$ threshold, i.e. when the scalar can pair annihilate to produce $Z^{\prime}$ bosons.}
\label{fig:abundance2}
\end{figure}
\section{Indirect Dark Matter Detection}
In this B-L model, dark matter self-annihilations take place through vector-like gauge mediation. Therefore, they occur at a similar rate for all SM fermions. That said, one can use gamma-ray observations of dwarf galaxies from the Fermi-LAT satellite to constrain the annihilation cross section into SM fermions, which after hadronization processes produce gamma rays. Fermi-LAT has been able to exclude annihilation cross sections into $b\bar{b}$ of $3\times 10^{-26} {\rm cm^3/s}$ for masses around $1-80$~GeV \cite{Ackermann:2015zua}. There are additional complementary constraints in the literature \cite{Galli:2011rz,Hooper:2012sr,Berlin:2013dva,Gomez-Vargas:2013bea,Kopp:2013eka,Calore:2013yia,Weniger:2013hja,Galli:2013dna,Madhavacheril:2013cna,Gonzalez-Morales:2014eaa,Bringmann:2014lpa,Abazajian:2014fta,Elor:2015bho,Bertoni:2015mla,Buckley:2015doa,Caputo:2016ryl}, which lie in the same ballpark. We therefore decided to adopt the Fermi-LAT collaboration results throughout. In both the fermion and scalar dark matter models, the right relic density is achieved for annihilation cross sections smaller than $3\times 10^{-26} {\rm cm^3/s}$. Since only heavy dark matter particles are viable, much heavier than $100$ GeV, the indirect detection limits are rather subdominant to collider and direct detection ones and for this reason not shown throughout.
\begin{figure}[!t]
\centering
\subfloat[$M_{Z'}=4$ TeV\label{fig:xsectiona}]{
\includegraphics[scale=0.45]{scatteringMZp4TeV.pdf}}
\subfloat[$M_{Z'}=6$ TeV\label{fig:xsectionb}]{
\includegraphics[scale=0.45]{scatteringMZp6TeV.pdf}}
\caption{{\it \bf Dirac Fermion}. Spin-independent dark matter-nucleon scattering cross section as a function of the dark matter mass with $n=1/3$ and $M_{Z^{\prime}}=4,6$~TeV for different gauge couplings, $g_{BL}=0.1,0.4,0.8$. The current limit from LUX-2015 (solid line)~\cite{Akerib:2015rjg}, preliminary limit from LUX-2016 (dotted-dashed)\cite{LUX2016} and the one projected from XENON1T (dashed line) for two years of data taking~\cite{Aprile:2015uzo} are also shown.}
\label{fig:xsection}
\end{figure}
\begin{figure}[!t]
\centering
\subfloat[$M_{Z'}=2$ TeV and $g_{BL}\in\{0.1,0.4,0.8\}$\label{fig:xsectionscalara}]{
\includegraphics[scale=0.45]{scalarDMplot4.pdf}}
\subfloat[$g_{BL}=0.4$ and $M_{Z'}\in\{2,4,6\}$ TeV\label{fig:xsectionscalarb}]{
\includegraphics[scale=0.45]{scalarDMplot5.pdf}}
\caption{{\it \bf Scalar Field}. Scattering cross section as a function of the dark matter mass for $n=1$ and various values of $M_{Z'}$ and coupling $g_{BL}$. Predictions are compared to current and projected bounds from LUX-2015 (solid), LUX-2016 (dotted-dashed) and XENON1T (dashed).}
\label{fig:xsectionscalar}
\end{figure}
\section{Direct Dark Matter Detection}
Direct dark matter detection relies on the measurement of nuclear recoil energies down to energies below 10~keV. The method is based on the use of discriminating variables such as ionization, heat, and scintillation efficiencies to disentangle possible dark matter events from nuclear background rates and mis-identified electron recoils, see~\cite{Freese:2012xd,Cremonesi:2013bma,Klasen:2015uma,Undagoitia:2015gya,Queiroz:2016awc,Mayet:2016zxu} for recent reviews. The measurement of the recoil energy is translated into the plane dark matter-nucleon scattering cross section vs.\ mass, once the dark matter velocity distribution and the local density is set. Since no excess of events has been observed, only limits in this same plane have been derived. LUX experiment provides the world-leading limits on both the spin-independent and spin-dependent scattering cross sections, with the former being more stringent, which we refer as LUX2015 in the figures. However, LUX just presented their new limit with 332 live days, which improves by a factor of four the latest one \cite{LUX2016}. The limit seems to be preliminary, but we have incorporated in the figures with a dotted-dashed line, labelled as LUX2016.
Since in our setup, both Dirac fermion and scalar dark matter models exhibit larger spin-independent rates, we will use the spin-independent bounds. Moreover, we present the projected bounds from the ongoing XENON1T experiment, which is expected to surpass the LUX2015 sensitivity by two orders of magnitude with two years of data taking \cite{Aprile:2015uzo}. In the following, we discuss the results for dark matter-nucleon scattering cross sections for both candidates.
\subsection{Dirac Fermion}
In Fig.~\ref{fig:diagram1c} we show the Feynman diagram responsible for dark matter-nucleon scattering. Fig.~\ref{fig:xsection} shows the spin-independent dark matter-nucleon scattering cross section as a function of the dark matter mass with $n=1/3$ and $M_{Z^{\prime}}=4$~TeV (\ref{fig:xsectiona}) and $M_{Z'}=6~$TeV (\ref{fig:xsectionb}) for different gauge couplings $g_{BL}\in\{0.1,0.4,0.8\}$. In both figures, current limits from LUX2015 (solid line) \cite{Akerib:2015rjg}, preliminary LUX2016 \cite{LUX2016}, and projected limits from XENON1T (dashed line) are superimposed.
The curves read from top to bottom: blue is for $g_{BL}=0.8$, red for $g_{BL}=0.4$, and pink for $g_{BL}=0.1$. The dark blobs in the figure reproduce the right relic abundance.
From Fig.~\ref{fig:xsectiona}, it is clear that one needs to use gauge couplings smaller than $0.8$ in order to have a viable dark matter candidate with masses below $2$~TeV. If no dark matter signal is seen, the XENON1T experiment is expected to exclude gauge coupling values larger than $0.4$, if the dark matter mass is demanded to be below $8$~TeV. Ramping up the $Z^{\prime}$ mass to $6$ TeV ameliorates the situation, and couplings as low as 0.8 can be allowed in the entire mass range.
This range will, however, be entirely probed by XENON1T, whereas this experiment will only probe dark matter masses below 1.5 TeV for a coupling of 0.4.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.6]{ATLAS-Zp-13TeV-limit.pdf}
\caption{Inclusive total cross section for $pp\rightarrow Z'\rightarrow \ell\bar{\ell}$ at NLO+NLL in the $\mathrm{U(1)}_{B-L}$ models for various values of $g_{B-L}$ as function of the mass of the heavy resonance $M_{Z'}$.}
\label{fig:atlas_excl}
\end{figure}
\begin{figure}[!h]
\centering
\subfloat[Exclusion limit in the plane $M_{Z'}-g_{B-L}$\label{fig:exclusion_contoura}]{
\includegraphics[width=0.45\textwidth]{Exclusion_limit_Zp_gbl_13.pdf}}
\subfloat[Exclusion limit in the plane $M_{Z'}/g_{B-L}-g_{B-L}$\label{fig:exclusion_contourb}]{
\includegraphics[width=0.45\textwidth]{Exclusion_limit_Zp_ov_gbl_13.pdf}}
\caption{LHC exclusion limits for the $\mathrm{U(1)}_{B-L}$ model.}
\label{fig:exclusion_contour}
\end{figure}
\subsection{Scalar Field}
In Fig.~\ref{fig:xsectionscalar} we display the scattering cross section as a function of the dark matter mass with $n=1$ and various values of the new gauge boson mass, $M_{Z^{\prime}}\in\{2,4\}$ TeV, and gauge couplings, $g_{BL}\in\{0.1,0.4,0.8\}$. In both plots, Figs.~\ref{fig:xsectionscalara} and~\ref{fig:xsectionscalarb}, the predictions are compared with current bound from LUX2015 (solid), preliminary from LUX2016 (dotted-dashed), and projected from XENON1T (dashed). The blobs represent points with the right relic density. The value of $n=1$ has been selected in order to simplify the identification of points satisfying the correct dark matter abundance. As before, results can be rescaled taking into account the scaling of the scattering cross section, $n^2 g_{BL}^4/M_{Z^{\prime}}^4$. That is, the result for $n=1,\ g_{BL}=0.4$, is equivalent to the one with $n=1/3$ and $g_{BL}=0.7$.
From Fig.~\ref{fig:xsectionscalara}, one sees that LUX2015 already rules out a large region of the model parameter space, forcing the use of suppressed gauge couplings, e.g. $g_{BL} \sim 0.1$, for $M_{Z^{\prime}}=2$~TeV. Note also that the projected limits from XENON1T might fiercely exclude couplings larger then 0.1.
Similarly, Fig.~\ref{fig:xsectionscalarb} shows the spin-independent cross section as a function of the dark matter mass for various values of $M_{Z'}$ and fixed $g_{BL}=0.4$ and $n=1$. The LUX experiment excludes $Z^{\prime}$ masses above $4$~TeV, whereas XENON1T has the potential to rule out masses larger than $6$~TeV, which is in the ballpark of the LHC-14 TeV sensitivity to gauge bosons with an integrated luminosity of 300 fb$^{-1}$~\cite{Ferrari:2000sp,Lindner:2016lxq}. Analogous conclusions would be drawn for $n=1/3$ by simply shifting the gauge coupling as mentioned before.
It is important to keep in mind that collider bounds on the model have been ignored up to now. Including them would lead to the exclusion of some of the points considered above. These limits will be included later on, when we present our results in a more informative plane, that is, $M_{Z^{\prime}}$ vs.\ $g_{BL}$. In what follows, we derive updated limits on the mass of a new neutral gauge boson using $13$~TeV dilepton data from the LHC and compare with the well known LEP bounds.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.8]{MzpgBLplot1TeV.pdf}
\caption{Allowed region of parameters for a $1$~TeV Dirac fermion as dark matter. The green curve delimits the region of parameter space with the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is ruled out by LUX2015 (XENON1T), the blue region is excluded by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).}
\label{fig:5}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.8]{MzpgBLplot2TeV.pdf}
\caption{Allowed region of parameters for a $2$~TeV Dirac fermion as dark matter. The green curve delimits the region of parameter space with the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is ruled out by LUX2015 (XENON1T), the blue region is excluded by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).}
\label{fig:6}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.8]{MzpgBLplot3TeV.pdf}
\caption{Allowed region of parameters for a $3$~TeV Dirac fermion as dark matter. The green curve delimits the region of parameter space with the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is ruled out by LUX2015 (XENON1T), the blue region is excluded by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).}
\label{fig:7}
\end{figure}
\section{Collider Limits}
\label{sec:coll_limits}
The ATLAS and CMS collaborations have performed extensive analyses to search for new heavy resonances in both dilepton and dijet signals. In the absence of any excess event over the Standard Model background, the two experiments derived lower bounds on the mass of the \ensuremath{Z'}\xspace-boson, with dileptons offering stronger limits than dijets due to relatively fewer background events. These bounds are limited to a given model, and typically the experiments express their results assuming simplified models such as the Sequential SM (SSM) or the GUT-inspired $\mathrm{E}_6$ models.
In this work, however, we re-interpreted their results in terms of the B-L model in question\footnote{See also \cite{Batell:2016zod} for displaced vertices limits in the B-L model, which are weaker for the region of interest.}. In particular, the ATLAS collaboration~\cite{ATLAS-CONF-2015-070} analyzed 3.2 fb$^{-1}$ of $pp$ collisions at $\sqrt{s}=13$ TeV searching for new phenomena in the dilepton final state and extracted the limit $M^{SSM}_{Z'}\geq 3.4$ TeV\footnote{Note that the width of the heavy resonance was fixed to $3\%$ of its mass.}. To calculate the total production cross section of a heavy neutral resonance \ensuremath{Z'}\xspace and its subsequent decay into leptons, we use the public code RESUMMINO~\cite{Fuks:2013vua}, in which we implemented the appropriate couplings. RESUMMINO implements threshold resummation for total cross sections, $p_T$-resummation for the $p_T$ distribution of heavy gauge bosons, as well as a joint resummation matched to the fixed-order NLO calculation.
When it comes down to interpreting dilepton resonance searches from ATLAS to a model different from the ones aforementioned, one needs to carefully compute the propagator width. In the B-L model, the width, $\Gamma_{Z^{\prime}}$ is proportional to $g_{BL}^{2} M_{Z'}$ and was estimated using PYTHIA 8.215~\cite{Sjostrand:2006za, Sjostrand:2007gs}. It was found to follow the relation
\begin{equation}
\frac{\Gamma_{Z'}(g_{BL})}{M_{Z'}}= \frac{\Gamma_{Z'} (g_{BL}=0.7)}{M_{Z'}}\left(\frac{g_{BL}}{0.7}\right)^{2}= 3\%\left(\frac{g_{BL}}{0.7}\right)^{2}\,
\label{eq:gamma_rel}
\end{equation}
to a very good precision. Therefore, it is clear that for any perturbative values of $g_{BL}$, the $Z'$-boson can be considered as a narrow resonance. For our numerical study we use the CT14~\cite{Dulat:2015mca} NLO PDF set with $\alpha_{S}(M_Z)=0.118$. Following~\cite{ATLAS-CONF-2015-070}, we cut on the transverse mass of the lepton pair, $q_{\ell\ell}^2 \geq 500$ GeV. For each value of mass, $M_{Z'}$, the electroweak coupling constant $\alpha_{EW}$ is evolved to $\alpha_{EW}(M_{Z'}^2)$. Finally, we set the factorization and renormalization scales such that $\mu_F=\mu_R=M_{Z'}$. With these settings, we were able to reproduce to a good level ($\sim$2-3\%) the ATLAS predictions for the SSM.
In Fig.~\ref{fig:atlas_excl}, we show the inclusive total cross section for the process, $pp\rightarrow Z'\rightarrow \ell\bar{\ell}$ calculated at NLO+NLL for the B-L model for various values of the gauge coupling $g_{B-L}$ and as a function of the mass of the heavy resonance. From this, it is straightforward to estimate the lower bound on the mass of the resonance. In Fig.~\ref{fig:exclusion_contoura}, we exhibit this limit in the plane $M_{Z'}$ vs.\ $g_{BL}$, while Fig.~\ref{fig:exclusion_contourb} shows the same limit in the plane $M_{Z'}/g_{BL}$ vs.\ $g_{BL}$.
Comparing with the SSM result obtained by ATLAS, we see that the exclusion bound for the B-L model is weaker. Note that in a recent analysis~\cite{Guo:2015lxa} the LHC bounds for $\sim5$ fb$^{-1}$ of data and $8$~TeV center-of-mass energy were computed. The conclusion was that for $M_{Z^{\prime}} < 3$~TeV the LHC bounds are stronger than those from LEP, which is in very good agreement with our results obtained at $13$~TeV with $3.2\, {\rm fb^{-1}}$ of data. For the SSM, ATLAS results for 13 TeV with 3.2 fb$^{-1}$ are a bit stronger than those at $8$~TeV and $20\, {\rm fb^{-1}}$, which uses much more data than the analysis in~\cite{Guo:2015lxa}. In addition to that, our results rely on the inclusion of NLO+NLL order effects, which improves our limits. Thus, the collider limits in \cite{Guo:2015lxa} seem to be overoptimistic. Moreover, an assessment of the LHC sensitivity to the B-L model at 13 TeV, was recently performed in \cite{Okada:2016gsh} without inclusion of detector effects and NLO corrections. There, the authors have found a limit much stronger than ours, namely $M_{Z^{\prime}} > 3$~TeV for $g_{BL}=0.01$.
We are now ready to combine the relic density, direct detection and collider constraints in the model. To do so, perhaps it is more informative to gather the results in the plane $M_{Z^{\prime}}$ vs.\ $g_{BL}$, since these two parameters basically define the B-L symmetry.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.8]{MzpgBLplot1TeVScalar.pdf}
\caption{Allowed region of parameters for a $1$~TeV scalar field as dark matter. The green curve delimits the region of parameter space with the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is ruled out by LUX2015 (XENON1T), the blue region is excluded by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).}
\label{fig:8}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.8]{MzpgBLplot2TeVScalar.pdf}
\caption{Allowed region of parameters for a $2$~TeV scalar field as dark matter. The green curve delimits the region of parameter space with the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is ruled out by LUX2015 (XENON1T), the blue region is excluded by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).}
\label{fig:9}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.8]{MzpgBLplot3TeVScalar.pdf}
\caption{Allowed region of parameters for a $3$~TeV scalar field as dark matter. The green curve delimits the region of parameter space with the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is ruled out by LUX2015 (XENON1T), the blue region is excluded by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).}
\label{fig:10}
\end{figure}
\section{Combined Results}
\subsection{Dirac Fermion}
In this section we outline the viable parameter space in an arguably more informative plane, i.e. $M_{Z^{\prime}}$ vs.\ $g_{BL}$ with charge $n=1/3$ under B-L throughout. We combine our findings from relic density, direct detection and collider searches for both the Dirac fermion and scalar dark matter models.
In all figures, the green curve delimits the region of parameter space yielding the right abundance ($\Omega h^2=0.11-0.12$), the pink (gray) shaded region is excluded by LUX2015 (XENON1T), the blue region is ruled out by dilepton data from the LHC, and the solid red (dashed) lines represent the current (old) LEP-II bounds, namely $M_{Z^{\prime}}/g_{BL}> 7$~TeV ($M_{Z^{\prime}}/g_{BL}> 6$~TeV).
In Fig.~\ref{fig:5} we collect these results for a $1$~TeV Dirac fermion, which features a $Z^{\prime}$ resonance of $2$~TeV. Since the annihilation cross section grows with $n^2 g_{BL}^4/(4m_{\chi}^2-M_{Z^{\prime}}^2)^2$, we can see that for small gauge couplings one needs to live very close to the resonance to obtain the right relic density, but as we increase the coupling, the regions relatively far from the resonance become viable. The annihilation cross section is typically small, leading to overabundant dark matter. Therefore one needs to either use large gauge couplings or be near the resonance region to increase the annihilation cross section and bring down the relic abundance to the correct value. Interestingly, LUX2015 limits on the spin-independent scattering cross section exclude a large region of parameter space, especially large values of the coupling. The linear behavior of direct detection limits occurs simply because the scattering cross section scales as $n^2g_{BL}^4/M_{Z^{\prime}}^4$. Consequently larger couplings are more strongly constrained by direct detection, but since $g_{BL}$ and $M_{Z^{\prime}}$ decrease simultaneously in the plane the direct detection limits are simply lines. The inclination is determined by the magnitude of the limit. For instance, XENON1T in two years of data is expected to improve LUX2015 bound by about two orders of magnitude, thus the steeper inclination. It is quite remarkable that XENON1T by itself may rule out almost the entire parameter space of the model. LHC-13 TeV limits based on dilepton data already now exceed the revised LEP-II bound and the LUX sensitivity for this model for gauge couplings smaller than 0.4.
In Figs.~\ref{fig:6}-\ref{fig:7} similar results for $m_{\chi}=2,3$~TeV are also shown. The model is less constrained as the dark matter mass increases for two reasons: \begin{inparaenum}[(i)]\item the direct detection limits are weakened as a result of fewer dark matter events. Indeed, since the local density, $\rho_{\odot}=n_{\chi}M_{\chi}$, is fixed, we have less dark matter events as we increase the mass; \item the resonance is located at $M_{\chi} \sim M_{Z^{\prime}}/2$ and therefore moves upwards along the $M_{Z^{\prime}}$ axis, towards a weakened LUX and XENON1T limit.\end{inparaenum}
\subsection{Scalar Field}
The possibility of having a singlet scalar dark matter in the B-L model is very much constrained\footnote{As aforementioned, we keep the same color scheme for all figures.}. In Fig.~\ref{fig:8} we present the result for $M_{\phi}=1$~TeV. First, we note that as in the Dirac fermion case, for sufficiently large values of the gauge coupling, there are regions of parameter space away from the $Z^{\prime}$ resonance at $2$~TeV where the correct relic density is achieved. Then, it is clear that there exists a strong degree of complementarity among dilepton, LUX2015 and LEP limits. Combined they fiercely exclude almost the entire parameter space of the model for $M_{\phi}=1$~TeV.
Only at the resonance is the model capable of satisfying all constraints and reproduce the right dark matter abundance. Strikingly, XENON1T is expected to rule out the possibility of having a $1$~TeV scalar dark matter particle in the B-L model. Note that decreasing the dark matter mass will not be sufficient as the direct detection constraints then get stronger.
Similarly, increasing the scalar mass to around 2-3~TeV does not have much impact as shown in Figs.~\ref{fig:9}-\ref{fig:10}.
Finally for a mass of 2 TeV, there is a tiny region right at the peak of the $Z^{\prime}$ resonance that might survive the projected XENON1T bound. At this point, the result must be taken with a grain of salt, since the precise XENON1T sensitivity would be required to draw any definite conclusion. Our findings agree approximately with~\cite{Rodejohann:2015lca}, but there the authors used an outdated XENON1T reach.
\subsection{Mixed Dark Matter Scenario}
Two-component dark matter is a plausible scenario. There is no fundamental reason to have one WIMP comprising the entire dark matter of the Universe. In the situation where solid signals come from direct detection and indirect dark matter searches, two-component dark matter arises as a promising framework. Several publications in the past have focused on two- or multi-component dark matter~\cite{Dienes:2011ja,Daikoku:2011mq,Dienes:2012cf,Bhattacharya:2013hva,Kajiyama:2013rla,Biswas:2013nn,Geng:2013nda,
Anandakrishnan:2013tqa,Geng:2014dea,Dienes:2014via,Queiroz:2014ara,Allahverdi:2014ppa,Allahverdi:2014bva,Allahverdi:2014eca,
Biswas:2015sva,Bian:2014cja,Esch:2014jpa,Bae:2014efa,Bae:2015efa,Bae:2015rra,Alves:2016bib}.
In Fig.~\ref{fig:11} we investigate the possibility of having two-component dark matter (fermion plus scalar) making up the total abundance. All the points are consistent with direct detection limits. As an example, we fix $n=1/3$ for the fermion and $n=1$ for the scalar and let the dark matter mass free. A scan in the plane $M_{Z^{\prime}}$ vs.\ $g_{BL}$ is performed looking for regions where $\Omega h^2=0.11-0.12$.
We have learned in the previous sections that scalar dark matter is more constrained than the Dirac fermion case, and for this reason we chose to exhibit several regimes for the two component dark matter based on the scalar abundance.
Blue circles represent the scenario where the scalar makes up for 30\% of the total abundance; pink squares correspond to 50\% of the total abundance; green triangles correspond to 70\% of the total abundance; and gray diamonds correspond to 90\% of the total abundance. Limits from the LHC (blue curve) and LEP (red curves) are also shown.
Notice that there are large regions of parameter space, where a two-WIMP dark matter scenario is realized within a well motivated theory.
Since the interactions that govern the scalar dark matter abundance are not very efficient, the scalar-dominated regime easily overcloses the Universe. The way out is to use sufficiently large gauge couplings and live near the $Z^{\prime}$ resonance region, enhancing the annihilation cross section and consequently bringing down the abundance to the proper value. Basically, all points in Fig.~\ref{fig:11} are in the neighbourhood of the resonance, except those for $g_{BL} \sim 1$, where one can obtain the right relic density while being slightly away from the resonance. This feature was observed in Figs.~\ref{fig:xsection}-\ref{fig:xsectionscalar}.
The points representing different regimes overlap, because we are scanning over the dark matter mass, which largely changes the abundance of the Dirac fermion dark matter. Therefore, for the same $g_{BL}$ one might have different abundances for the scalar and fermion fields, which explains the overlapping. In summary, Fig.~\ref{fig:11} shows a UV complete realization of a two component dark matter scenario.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.7]{mixedDM.pdf}
\caption{Scan of the parameter space, in which a two-component dark matter scenario can be successfully realized and account for the entire dark matter of the Universe in agreement with direct detection limits. We have superimposed limits from the LHC (blue curve) and LEP (red curves). The points with different shapes represent different scalar dark matter contributions to the overall dark matter abundance. Blue circles represent the scenario where the scalar makes up for 30\% of the total abundance; pink squares correspond to 50\% of the total abundance; green triangles correspond to 70\% of the total abundance; and gray diamonds correspond to 90\% of the total abundance.}
\label{fig:11}
\end{figure}
\section{Conclusions}
Supplementing the SM with an extra $\mathrm{U}(1)_{B-L}$ gauge symmetry is an appealing possibility. In this paper, we studied the dark matter phenomenology of simplified models exhibiting such a gauge symmetry and in particular the possibilities of having Dirac fermion as well as scalar dark matter with and without broken B-L symmetry. In this context, we determined the impact of constraints coming from indirect and direct detection experiments as well as collider limits. Bounds from LUX2015, LUX2016 and projected bounds from XENON1T have been considered along with the famous LEP limit. In addition, we re-interpreted dilepton searches from the LHC at 13 TeV and extracted competitive limits for the model.
While XENON1T projected bounds have a very good potential to exclude most of if not all the parameter space for scalar dark matter, we found that Dirac fermion dark matter would still be viable in a larger region of the parameter space. Interestingly, it was shown that the LHC limits that were extracted from dilepton production are already better than the LEP bounds for small gauge couplings. Finally, we also considered a mixed dark matter scenario, in which the relic abundance is realized as a combination of both fermion and scalar dark matter. In this case, numerous points satisfying the required relic density, collider, direct and indirect dark matter constraints were found, showing that a minimal and successful two component dark matter model is realized.
\section*{Acknowledments}
The work of M.K.\ was supported by the BMBF under contract 05H15PMCCA
and by the Helmholtz Alliance for Astroparticle Physics (HAP).
The work of F.L.\ was partially supported by the U.S.\ Department of Energy
under Grant No.\ DE-SC0010129. The authors thank Carlos Yaguna, Alexandre Alves, and Juri Smirnov for discussions.
\bibliographystyle{JHEP}
|
\section{Introduction}
\label{sec:introduction}
In many real world problems, the distribution of data between classes is imbalanced. Learning from imbalanced datasets is an important research problem with many applications.
The fundamental issue in imbalanced learning is the ability of imbalanced data to significantly compromise the performance of standard learning algorithms \cite{HH:09}. Generally, there are three primary reasons that can cause this problem \cite{weiss2004mining}.
The first reason is that the lack of data in the minority class makes it difficult to detect regularities within the minority class. Thus, the learned decision boundaries are less likely to approximate the true decision boundaries.
Second, many classification algorithms utilize a general bias for better generalization and to avoid overfitting during learning. However, such bias can adversely affect the ability to learn the minority class. Inductive bias also plays a key role with respect to the minority class. Most classification algorithms prefer more common classes in the presence of uncertainty (i.e., they are biased in favor of the class priors).
Last but not least, noise exerts a greater impact on the minority class, because in this case it is more difficult for a classifier to distinguish noise from minority data. This is especially so in extreme cases where the number of noisy samples is greater than actual minority samples. The problem of overfitting rises again, when modifying the classifier to learn the minority data correctly.
To address these problems, numerous research efforts have been devoted to imbalanced learning in recent years. The majority of techniques that solve the imbalanced learning problem fall into two categories: cost-sensitive methods and sampling-based methods. In the next section, we review related work on sampling-based methods.\footnote{Go to (https://github.com/xzhang311/CGMOS.git) for codes of this project.}
\subsection{Related work}
A number of solutions to the class-imbalance problem were previously proposed both at the data and algorithmic levels \cite{chawla2004editorial}. There are mainly three groups of methods that can solve imbalanced learning problem \cite{HH:09} including sampling methods, cost sensitive methods, and kernel methods. Sampling-based methods are very effective and easy to use when solving imbalanced learning problems. In addition, sampling-based methods can be used together with methods in the other two groups to further improve performance. In such approaches a sampling technique is used to modify an imbalanced dataset to produce a balanced distribution. It has been shown that for most imbalanced datasets, sampling techniques do improve classification accuracy.
The basic sampling methods include undersampling and oversampling. Undersampling reduces majority class samples while oversampling increases minority class samples. While several works achieving data balance through undersampling have been proposed in the past \cite{liu2009exploratory}\cite{ZJMI03}, more research efforts have been devoted to oversampling due to the fact that oversampling does not discard information.
The simplest form of oversampling is duplication of minority class samples. This approach decreases the overall level of class imbalance, but may lead to overfitting \cite{Drummond03c4.5}. SMOTE \cite{CNV:02} is a fundamental approach for oversampling using data synthesis. To balance the dataset, SMOTE randomly selects a seed sample and synthesizes a new sample by applying a linear interpolation between the seed sample and one of its neighbors. Large research efforts have been devoted to feature space data synthesis based on SMOTE. Several methods integrate data synthesis as a part of the learning procedure. For example, by introducing SMOTE in each iteration of boosting, SMOTEBoost \cite{chawla2003smoteboost} increases the number of minority class samples and focus on these cases in each boosting iteration. Using the same idea of boosting, DataBoost-IM \cite{guo2004learning} and RAMOBoost \cite{chen2010ramoboost} discover samples difficult to classify during each iteration of boosting, which are used to guide the oversampling in both the majority and minority classes.
In another group of minority oversampling approaches, the data synthesis procedure is independent of the learning processes. Such methods give preferences to different regions of a dataset by assigning weights to samples in the dataset. These weights can then generate a probability distribution which is used for randomly drawing samples. In such approaches the data synthesis can be completed in one step. Methods in this group include Borderline-SMOTE \cite{HH:05}, Adasyn \cite{HH:08}, \cite{barua2011novel} and MWMOTE \cite{barua2014mwmote}. All of these methods synthesize more samples along decision boundaries. However, these methods do not have objective functions to systematically guide the process of oversampling and so do not have a systematic way to decide on where new data should be synthesized. Thus, such approaches cannot measure the impact of each synthetic sample. As a result, there are several potential problems. One is that the oversampling procedure may sacrifice the performance of the majority class in order to improve the performance of the minority class in the classification. Another is that synthetic minority samples themselves can be misclassified and affect the performance in the minority class.
\subsection{Novel Contribution}
The proposed approach, CGMOS, is a member of the SMOTE family that can achieve data oversampling in a single step. To address some of the shortcomings in existing approaches, we propose a novel oversampling strategy by systematically considering the performance of both minority and majority classes. Based on a Bayesian classification framework, our proposed approach computes the influence of minority data addition on the certainty of the entire dataset. CGMOS thus can synthesize new samples that will improve the overall certainty of the entire dataset in classification. We prove that during training CGMOS is guaranteed to perform better than SMOTE when using Bayesian classification. To validate the proof, We further show experimentally that CGMOS outperforms known approaches when tested on real-world data set collections using different classifiers.
\section{Problem Formulation}
\label{sec: problem formulation}
In this paper, we address the binary classification problem for imbalanced datasets. Let $D=\{(x_j, y_j)\}_{j=1}^n$ be a training dataset, where $x_j \in \mathfrak{R}^m$ are features and $y_j \in \{l_{\mbox{mjr}}, l_{\mbox{mnr}}\}$ are ground truth class labels. We begin by formally defining the certainty of imbalanced binary classification using a Bayesian framework, where a kernel density estimation (KDE) is used to estimate the samples' probability density function (PDF). We then show how CGMOS can synthesize more samples according to the certainty estimation.
\subsection{Definition of Certainty}
Suppose $(x_j, y_j)$ is any tuple in the training dataset $D$, where $x_j$ is a feature vector and $y_j$ is the ground truth label of $x_j$.
A Bayesian classifier maps $x_j \rightarrow l$, $l\in\{l_{\mbox{mjr}}, l_{\mbox{mnr}}\}$ using following rule.
\begin{align*}
l = \left\{
\begin{array}{cl}
l_{\mbox{mnr}} & \text{if} \;\; \frac{P(l_{\mbox{mnr}} | x_j)}{P(l_{\mbox{mjr}} | x_j)} > 1\\
l_{\mbox{mjr}} & \text{otherwise}
\end{array}\right.
\end{align*}
where the posterior probability $P(l|x_j)$ is computed using Bayes' rule:
\begin{align*}
P(l | x_j) = \frac{P(x_j | l)P(l)}{P(x_j)}; \;\;\;\; l\in\{l_{\mbox{mjr}}, l_{\mbox{mnr}}\}
\end{align*}
Uncertainty is commonly used in machine learning algorithms. In this work, we use the posterior probability $P(y_j | x_j)$ to define certainty. This is because in classification, the posterior probabilities $P(y_j | x_j)$ reflect the certainty of assigning a sample to a correct label, where higher numbers indicate classification results with a stronger certainty.
\newline
\newline
\noindent \textbf{Definition 1. (Certainty)} Let $(x_j, y_j)$ be any tuple in $D$, where $x_j$ is a feature vector and $y_j$ is the ground truth label of $x_j$. The certainties for samples in the majority and minority class are respectively defined as:
\begin{equation}
C(y_j=l_{\mbox{mjr}} | x_j) = P(y_j=l_{\mbox{mjr}}|x_j)
\end{equation}
\begin{equation}
C(y_j=l_{\mbox{mnr}} | x_j) = P(y_j=l_{\mbox{mnr}}|x_j)
\end{equation}
It should be noted that in the case of binary classification the definition of certainty above is related up to some constants to the uncertainty defined in \cite{2013Sharma} based on margin confidence.
\subsection{PDF Estimation}
There are two general ways to estimate a density function: parametric or non-parametric. In this work we use a non-parametric model so as to not depend on a specific distribution model. We use kernel density estimation (KDE)\cite{elgammal2002background}\cite{zhang2006bayesian} to estimate the likelihood $P(x_j|l)$, $l\in \{l_{\mbox{mjr}}, l_{\mbox{mnr}}\}$.
Assuming that the data is independent and identically distributed (i.i.d) and drawn from some distribution with an unknown density $P(x_j | l)$, we have using KDE:
\begin{equation}
\begin{split}
P(x_j | l) =& \frac{\sum_{k=1}^n K(\frac{x_j-x_k}{h_k}) \mathrm{I}(y_k=l)}{\sum_{k=1}^n \mathrm{I}(y_k=l)}
\end{split}
\end{equation}
where $l \in \{l_{\mbox{mjr}}, l_{\mbox{mnr}}\}$, $I(\cdot)$ is an indicator function, and $K(\cdot)$ is a kernel function which has zero mean and integrates to one. Given any sample $x_k$, the bandwidth $h_k$ of the sample $x_k$ controls the effective range of the kernel and smoothness of the density function. Intuitively one wants to choose $h_k$ as small as the data allows to exhibit as many underlying structures of the data as possible. Small bandwidth, however, will result in a noisy estimate. In this work, for any sample $x_k$, we calculate a bandwidth $h_k$ as a scaled average distance between $x_k$ and its $q$ nearest neighbors:
\begin{equation}
h_k = \sigma \cdot \frac{\sum_{x \in N(x_k)} \| x-x_k \|}{q}
\end{equation}
where $N(x)$ is the set of the $q$ nearest neighbors of $x_k$ and $\sigma > 0$ is a scale factor applied to the distance. We will discuss selection of parameters $\sigma$ and $q$ in Section \ref{sec: results}.
\subsection{Oversampling Seed Selection}
In most classification algorithms, samples close to decision boundaries have less certain classification results. In order to achieve better predictions for such samples, many existing approaches synthesize data directly along the boundaries. However, this is risky and the expected performance improvement is not guaranteed. There are two primary reasons. First, samples from both classes are mixed in regions near the boundaries. Synthetic samples if added to these regions are less predictable and hard to learn. Second, adding synthetic minority samples to these regions may adversely impact the majority class, which may in turn decrease the performance of the majority class in classification. Instead of unguided oversampling near the boundaries, our proposed approach targets adding samples by considering the certainties of both the minority and majority classes before and after adding the samples. The synthetic samples thus are added to locations that can improve the overall certainty of the original data and boost the performance of the classification.
CGMOS uses a similar procedure as SMOTE when synthesizing a new sample. The sample is produced by interpolating between one seed sample and some of its neighbors. However, instead of randomly drawing a seed sample for interpolation, CGMOS assigns each sample $(x_i, y_i) \in D$ a weight $W(x_i)$ which is used to determine the probabilities of $x_i$ being chosen for interpolation. A higher weight results in a higher probability of a point being selected.
To compute $W(x_i)$, we suppose that a new sample will be added to the same location as $x_i$. The weight $W(x_i)$ is computed as a relative certainty change\footnote{Measuring absolute certainty increments will not work, because measuring magnitude will give higher preference to parts which already have high certainty.} comparing the certainty before and after the sample is added. With a new sample added at location $x_i$, we update the certainty for all $(x_j, y_j) \in D$ and denote it as $C_{+i}(y_j | x_j)$.
\newline
\newline
\noindent \textbf{Definition 2. (Relative Certainty Change)} The relative certainty change of label $y_j$ assigned to feature $x_j$ due to adding a minority example at location $x_i$ is defined by:
\begin{equation}
R_{+i}(y_j | x_j)= \frac{C_{+i}(y_j | x_j) -C(y_j | x_j)}{C(y_j | x_j)}
\label{eqn: relative diff}
\end{equation}
where $C(y_j | x_j)$ is the certainty before addition.
When computing $W(x_i)$, CGMOS considers the relative certainty changes of examples from both the majority and the minority classes. $W(x_i)$ is computed as the average value of relative certainty changes of all samples in the dataset.
\begin{equation}
\begin{split}
W(x_i)= 1+\frac{1}{n}\sum_{j=1}^n R_{+i}(y_j | x_j)
\end{split}
\label{eqn: weight}
\end{equation}
Given $W(x_i)$ for all $x_i \in D$, it is easy to see $W(x_i)>0$. We compute a normalization factor $z$ so that $\frac{1}{z}\sum_{i=1}^n W(x_i)=1$. Therefore, the oversampling procedure can randomly choose sample for interpolation according $W(x_i)/z$. The interpolation phase of CGMOS is the same as SMOTE\cite{CNV:02}.
A demonstration of CGMOS is shown in Fig. \ref{fig: demoofdiffinsertion}. In this figure, samples in both the majority and minority classes are randomly drawn based on Gaussian distribution, where the means of the two datasets are on the same horizontal line, and the mean of the majority is to the right of the minority. The majority class contains 2000 samples and the minority class contains 400 samples. Color in part 1 of the figure indicates the certainty of each example with respect to its class, where red indicates high certainty. We highlight 3 regions (A, B, C) in the minority class. Samples in region A have relative high certainties, sample in region B has low certainties and region C is a boundary region in which samples have the lowest certainties. Part 2 of the figure shows the weight of each example as computed by our approach where red indicates high values. Region B has higher values and is where CGMOS will synthesize most of the samples.
To show the certainty changes induced by adding samples at different locations of the dataset, in part 3 of the figure we add one minority sample and move its location with a fixed step size from left to right on a horizontal line passing through the two classes. We then compute the relative certainty changes for all samples in both classes. As can be observed, by measuring relative certainty changes, CGMOS will assign a higher weight to samples in region B. The figure also shows that by oversampling more in region B, the certainty of the entire dataset gets improved, because the relative certainty changes are positive.
\begin{figure}[ht]
\centerline{ %
\begin{tabular}{c}
\resizebox{0.48\textwidth}{!}{\rotatebox{0}{
\includegraphics{./combined-crop.pdf}}}
\end{tabular}}
\caption{Demonstration of CGMOS. In first two figures, diamonds represent minority samples and circles represent majority samples. The positions of synthesized data points are labeled using a star symbol on a horizontal line passing through the center. The x and y axes represent features. In the bottom figure the x axis indicates a location where a sample was added (in correspondence with the first two figures) whereas the y-axis indicates the relative certainty change.}
\label{fig: demoofdiffinsertion}
\end{figure}
\section{Theoretical Guarantee Over SMOTE}
Several existing approaches claim handling imbalanced learning better than SMOTE. Such claims are normally validated using empirical tests without a theoretical guarantee and in some instances may not extend to new datasets. In this section we provide a theoretical guarantee showing that CGMOS is expected to work better than SMOTE in training process.
Let $D=\{(x_j, y_j)\}_{j=1}^n$ be a training dataset. Let $W(D)=\{W(x_i)\}_{i=1}^n$ be the sample weights computed using Eqn. \ref{eqn: weight}.
\newline
\newline
\textbf{Lemma 1.} \textit{Given a set of weights $\{W(x_i)\}_{i=1}^n$ as defined above and a normalization factor $z$ given by $z=\sum_{i=1}^n W(x_i)$, it must be that $\sum_{i=1}^n W(x_i)^2 \geq \frac{z^2}{n}$.}
\newline
\newline
\textbf{Proof}
Let $W$ be an n-dimensional vector whose elements are $W(x_i)$. Let $I$ be an n-dimensional vector whose elements are all 1. Using the Cauchy-Schwarz inequality we have: $|W\cdot I|\leq \|W\|\cdot\|I\|$. Thus, $|\sum_{i=1}^n W(x_i)|\leq \sqrt{\sum_{i=1}^n W(x_i)^2 }\sqrt{n}$ using the fact that $\sum_{i=1}^n W(x_i)=z$, we thus have $\sum_{i=1}^n W(x_i)^2 \geq \frac{z^2}{n}$. $\;\;\;\blacksquare$
\newline
\noindent\textbf{Definition 3. (Addition Likelihood Ratio)} Let $\theta$ denote the non-parametric likelihood estimate $P(x_j|l)$, $l\in \{l_{\mbox{mjr}}, l_{\mbox{mnr}}\}$ before a new sample $x_i$ is added, and $\theta'$ denote the non-parametric likelihood estimate after the new sample is added. The addition likelihood ratio $r_{+i}(y_j|x_j)$ of example $x_j$ by adding data to $x_i$ location is defined as the ratio between the likelihood estimate after the new addition and the likelihood estimate before the new addition:
\begin{equation}
r_{+i}(y_j|x_j) \equiv P(y_j|x_j ; \theta') / P(y_j|x_j; \theta).
\label{eqn. likelihoodratio}
\end{equation}
\noindent \textbf{Lemma 2.} \textit{The addition likelihood ratio $r_{+i}(y_j|x_j)$ is related to the relative certainty change ratio $R_{+i}(y_j|x_j)$ by: }
\begin{equation}
r_{+i}(y_j|x_j)=1+R_{+i}(y_j|x_j).
\end{equation}
\noindent\textbf{Proof} According to the definition of the certainty, we have $C_{+i}(y_j|x_j)=P(y_j|x_j; \theta')$ and $C(y_j|x_j; \theta)=P(y_j|x_j; \theta)$. Then $P(y_j|x_j; \theta')=r_{+i}(y_j|x_j) P(y_j|x_j; \theta)$ according to the definition of likelihood ratio. Given Eqn. \ref{eqn: relative diff}, we have that $R_{+i}(y_j | x_j)=\frac{r_{+i}(y_j|x_j) P(y_j | x_j; \theta) - P(y_j |x_j; \theta)}{P(y_j |x_j; \theta)}$. By simplifying this equation, we thus have
\begin{equation}
r_{+i}(y_j|x_j)=1+R_{+i}(y_j|x_j). \;\;\;\blacksquare
\end{equation}
The addition likelihood ratio defined in Eqn. \ref{eqn. likelihoodratio} measures the gain in adding a new point, where higher gains are desired. Note that while the gain is normally close to 1 it may be bigger or smaller than 1.
\newline
\noindent\textbf{Definition 4. (Average gain)} The average gain when adding sample $x_i$ is defined by:
\begin{equation}
\bar{r}_{+i}=\frac{1}{n} \sum_{j=1}^n r_{+i}(y_j|x_j)
\end{equation}
\noindent\textbf{Lemma 3.} \textit{Given the average gain, it must be that:}
\begin{equation}
\bar{r}_{+i}=W(x_i).
\end{equation}
\noindent\textbf{Proof} Using the definition of $W(x_i)$ we have $W(x_i)=\frac{1}{n}\sum_{j=1}^n R_{+i}(y_j|x_j)$. Using Lemma 2 we can replace $r_{+i}(y_j | x_j)-1$ with $R_{+i}(y_j | x_j)$. Hence:
\begin{equation}
\bar{r}_{+i}=\frac{1}{n}\sum_{j=1}^n R_{+i}(y_j | x_j) + 1\equiv W(x_i) \;\;\;\blacksquare
\end{equation}
The average gain is an indicator of the benefit of CGMOS. We show that the expected average gain is higher in proposed approach compared with SMOTE.
\newline
\noindent \textbf{Theorem 1.} \textit{The expected average gain in CGMOS is higher or equal to that of SMOTE.}
\newline
\noindent\textbf{Proof} For CGMOS the expected average gain is given by:
\begin{equation}
E_p\equiv E[\bar{r}_{+i}] = \sum_{i=1}^n \bar{r}_{+i} \frac{W(x_i)}{z}
\end{equation}
where $z$ is the normalization factor as defined earlier.
Using Lemma 3:
\begin{equation}
E_p=\sum_{i=1}^n W(x_i)\frac{W(x_i)}{z}=\frac{1}{z}\sum_{i=1}^n W^2(x_i).
\end{equation}
\newline
\noindent For SMOTE the expected average gain is given by:
\begin{equation}
E_s\equiv E[\bar{r}_{+i}] = \sum_{i=1}^n \bar{r}_{+i} \frac{1}{n}
\end{equation}
Using Lemma 3:
\begin{equation}
E_s = \frac{1}{n}\sum_{i=1}^n W(x_i) =\frac{z}{n}
\end{equation}
Using Lemma 1:
\begin{equation}
E_p= \frac{1}{z}\sum_{i=1}^n W^2(x_i)\geq \frac{1}{z}\frac{z^2}{n}= E_s \;\;\;\blacksquare
\end{equation}
\section{Results and Discussion}
\label{sec: results}
\subsection{Datasets}
30 real-world datasets were randomly chosen from the UCI machine learning repository \cite{Lichman:2013} for empirical testing of CGMOS. Most of the datasets were released within the past 10 years. As some of the datasets contain samples of more than two classes, we convert such datasets to a binary classification problem by keeping the class with the least data and merging all other classes. A summary of the test collections is provided in Table \ref{tab: realdata}.
\begin{table*}[ht]
\begin{center}
\scalebox{1}
{
\begin{tabular}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}l|l|l|l!{\VRule[1.5pt]}l!{\VRule[1.5pt]}l|l|l|l!{\VRule[1.5pt]}}
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}c!{\VRule[1.5pt]}}{{\bf Name}} & \multicolumn{1}{c|}{{\bf S \#}} & \multicolumn{1}{c|}{{\bf F \#}} & \multicolumn{1}{c|}{{\bf R}} & \multicolumn{1}{c!{\VRule[1.5pt]}}{{\bf Year}} & \multicolumn{1}{c!{\VRule[1.5pt]}}{{\bf Name}} & \multicolumn{1}{c|}{{\bf S \#}} & \multicolumn{1}{c|}{{\bf F \#}} & \multicolumn{1}{c|}{{\bf R}} & \multicolumn{1}{c!{\VRule[1.5pt]}}{{\bf Year}}\\
\specialrule{1.5pt}{0pt}{0pt}
\textbf{BankMarket} & 45211 & 17 & 0.13 & 2012 & \textbf{Libras} & 360 & 91 & 0.07 & 2009 \\ \hline
\textbf{BloodService} & 748 & 5 & 0.31 & 2008 & \textbf{MultipleFs} & 2000 & 649 & 0.11 & 1998 \\ \hline
\textbf{BreastCancer} & 400 & 9 & 0.53 & 1988 & \textbf{Parkinson} & 1040 & 26 & 0.02 & 2014 \\ \hline
\textbf{BreastTissue} & 106 & 10 & 0.15 & 2010 & \textbf{PlanRelax} & 182 & 13 & 0.4 & 2012 \\ \hline
\textbf{CarEvaluation} & 1730 & 6 & 0.04 & 1997 & \textbf{QSAR} & 1055 & 41 & 0.51 & 2013 \\ \hline
\textbf{Card'graphy} & 2126 & 23 & 0.09 & 2010 & \textbf{SPECT} & 268 & 22 & 0.26 & 2001 \\ \hline
\textbf{CharacterTraj} & 2860 & 3 & 0.04 & 2008 & \textbf{SPECTF} & 134 & 44 & 0.26 & 2001 \\ \hline
\textbf{Chess} & 3198 & 22 & 0.91 & 1989 & \textbf{SeismicBumps} & 2584 & 19 & 0.07 & 2013 \\ \hline
\textbf{ClimateSim} & 540 & 18 & 0.09 & 2013 & \textbf{Statlog} & 2310 & 19 & 0.17 & 1990 \\ \hline
\textbf{Contraceptive} & 1474 & 9 & 0.29 & 1997 & \textbf{PlatesFaults} & 1941 & 27 & 0.03 & 2010 \\ \hline
\textbf{Fertility} & 100 & 10 & 0.14 & 2013 & \textbf{TAEvaluation} & 151 & 5 & 0.49 & 1997 \\ \hline
\textbf{Haberman} & 306 & 3 & 0.36 & 1999 & \textbf{UKnowledge} & 403 & 5 & 0.1 & 2013 \\ \hline
\textbf{ILPD} & 580 & 10 & 0.4 & 2012 & \textbf{Vertebral} & 310 & 6 & 0.48 & 2011 \\ \hline
\textbf{ImgSeg} & 2310 & 19 & 0.17 & 1990 & \textbf{Customers} & 440 & 8 & 0.48 & 2014 \\ \hline
\textbf{Leaf} & 342 & 16 & 0.24 & 2014 & \textbf{Yeast} & 1484 & 8 & 0.04 & 1996 \\ \specialrule{1.5pt}{0pt}{0pt}
\end{tabular}
}
\end{center}
\caption{Summary of the datasets used in our experiments, where S\#, F\#, and R stand for the number of samples, the number of features, and imbalance ratio (defined as \#minority/\#majority).}
\label{tab: realdata}
\end{table*}
\begin{figure}[ht]
\centerline{ %
\begin{tabular}{cc}
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./roc_nb-crop.pdf}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./roc_knn-crop.pdf}}}
\\
b-kde & knn
\\
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./roc_svm-crop.pdf}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./roc_nn-crop.pdf}}}
\\
svm & nn
\\
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./roc_rf-crop.pdf}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./roc_boost-crop.pdf}}}
\\
rf & Adaboost.M1
\\
\end{tabular}}
\caption{ROC curves of classification results. From left to right, up to down, we show the results of 6 different classifiers: b-kde, knn, svm, nn, rf and Adaboost.M1. Curves in blue are the results of the proposed CGMOS.}
\label{fig: roc}
\end{figure}
\begin{figure}[t]
\centerline{ %
\begin{tabular}{ccc}
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./icr_nb-crop.pdf}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./icr_knn-crop.pdf}}}
\\
b-kde & knn
\\
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./icr_svm-crop.pdf}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./icr_nn-crop.pdf}}}
\\
svm & nn
\\
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./icr_rf-crop.pdf}}}
&
\resizebox{0.23\textwidth}{!}{\rotatebox{0}{
\includegraphics{./icr_boost-crop.pdf}}}
\\
rf & Adaboost.M1
\\
\end{tabular}}
\caption{Comparison of results when increasing the number of data synthesized for the minority class. The curves measure the average AUC of the ROC curves. Curves in blue are the results of the proposed CGMOS.}
\label{fig: increasingnum}
\end{figure}
\subsection{Compared Approaches}
According to a survey of imbalanced learning \cite{HH:09}, there are mainly three groups of methods addressing imbalanced learning: sampling methods, cost sensitive methods, and kernel methods. The proposed CGMOS belongs to the sampling group. Thus, we compare CGMOS to five other oversampling methods in this group: SMOTE\cite{CNV:02}, Borderline-SMOTE\cite{HH:05}, ADASYN\cite{HH:08}, MWMOTE\cite{barua2014mwmote} and RAMOBoost\cite{chen2010ramoboost}. Since oversampling by duplication is broadly used in many applications as a baseline, we add it to our evaluation as well. To demonstrate the improvement of these oversampling strategies, we include in the comparison raw data with no oversampling. It should be noted that sampling methods are often combined with cost sensitive methods and kernel methods to further boost learning. \cite{chawla2004editorial}\cite{chawla2003smoteboost}\cite{guo2004learning}.
\subsection{Base classifiers}
We match the compared classifiers to classifiers used in other SMOTE extension evaluations. Six well-known classifiers are tested in experiments. The first is the Bayesian classifier based on kernel density estimation described in Section \ref{sec: problem formulation} (b-kde). The second is a K nearest neighbors classifier (knn). The third is a support vector machine classifier using RBF kernel (svm). The fourth one is a neural network (nn) with one hidden layer. We use in addition two ensemble methods: a random forest implementing the C4.5 decision tree \cite{Quinlan:1993} (rf) and Adaboost.M1 \cite{fy:1996}. All hyper-parameters of the classifiers tested were determined by cross validation to ensure the best performance of each method.
\subsection{Evaluation metric}
Finding an appropriate evaluation metric for different tasks is challenging, since different evaluation metrics are designed for different purposes. The datasets used in this paper cover from financial application to medical treatment. To achieve an general evaluation and avoid bias, we follow the method in \cite{CNV:02}\cite{HH:05}\cite{HH:08}\cite{barua2014mwmote}\cite{chen2010ramoboost} and use different metrics to evaluate the performance of the proposed CGMOS oversampling algorithm.
Among these evaluation metrics, the most frequently adopted ones are $Precision$ and $Recall$ when the focus of evaluation is focus on one specific class such as problems in text classification, information extraction, natural language processing and bioinformatics. In these areas of application the number of examples belonging to one class is often substantially lower than the overall number of examples, which basically are imbalance learning problems. $Precision$ and $Recall$ are defined as:
\begin{align*}
Precision &= \frac{TP}{(TP + FP)}\\
Recall &= \frac{TP}{(TP+FN)}
\end{align*}
However, these two metrics share an inverse relationship between each other. A quick inspection on the $Precision$ and $Recall$ formulas readily yields that solely use each of these two metrics only provide a limit view of an algorithm under test. As $Recall$ provides no insight to how many examples are incorrectly labeled as positive and $Precision$ cannot assert how many positive examples are labeled incorrectly. Specifically, the $\fscore$ combines $Precision$ and $Recall$ as measure of the effectiveness of classification in terms of a ration of the weighted importance on either $Recall$ or $Precision$, which is defined as:
\begin{align*}
\fscore = \frac{(1+\beta^2)\cdot Precision \cdot Recall}{(\beta^2\cdot Precision)+ Recall}.
\end{align*}
We use $\beta=1$ to treat $Precision$ and $Recall$ equally in all evaluations. As a result, $\fscore$ provides more insight into the functionality of a classifier.
As $\fscore$ measures the harmonic mean of $Precision$ and $Recall$, we also compute $Gscore$ which is the geometric mean of $Precision$ and $Recall$ and is able to evaluate the degree of inductive bias in terms of a ratio of positive accuracy and negative accuracy \cite{HH:09}.
\begin{align*}
\gscore=\sqrt{Precision \cdot Recall}
\end{align*}
As both $\fscore$ and $\gscore$ concentrate their measures on one class (positive examples) \cite{sokolova2006beyond}, to have a general way of comparing our test results, we altered the positive examples between the majority and minority classes when computing $\fscore$ and $\gscore$. Thus we show $\fscore$ and $\gscore$ for the majority and the minority classes separately.
\begin{table*}[]
\begin{center}
\scalebox{1.0}{
\begin{tabular}{lc!{\VRule[1.5pt]}c|c|c|c!{\VRule[1.5pt]}c|c|c|c!{\VRule[1.5pt]}}
\Cline{1.5pt}{3-10}
& \textbf{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{\textbf{Minority}} & \multicolumn{4}{c!{\VRule[1.5pt]}}{\textbf{Majority}} \\ \Cline{1.5pt}{2-10}
\multicolumn{1}{l!{\VRule[1.5pt]}}{} & \textbf{AUC} & \textbf{Precision} & \textbf{Recall} & \textbf{Fscore} & \textbf{Gscore} & \textbf{Precision} & \textbf{Recall} & \textbf{Fscore} & \textbf{Gscore} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{b-kde}} & & \multicolumn{4}{c!{\VRule[1.5pt]}}{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Original}} & 0.797 & 0.139 & 0.033 & 0.054 & 0.068 & 0.830 & \textbf{0.995} & \textbf{0.905} & \textbf{0.909} \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Dup}} & 0.733 & 0.385 & 0.454 & 0.417 & 0.418 & 0.869 & 0.742 & 0.801 & 0.803 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{SMOTE}} & 0.807 & 0.488 & 0.705 & 0.577 & \textbf{0.587} & 0.833 & 0.644 & 0.726 & 0.733 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{B-SMOTE}} & 0.774 & 0.258 & 0.456 & 0.330 & 0.343 & 0.846 & 0.671 & 0.748 & 0.754 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{MWMOTE}} & 0.794 & 0.396 & \textbf{0.754} & 0.520 & 0.547 & 0.836 & 0.557 & 0.669 & 0.682 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{ADASYN}} & 0.802 & 0.395 & 0.632 & 0.487 & 0.500 & 0.817 & 0.598 & 0.691 & 0.699 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{RAMOboost}} & 0.748 & 0.358 & 0.343 & 0.350 & 0.350 & 0.860 & 0.822 & 0.841 & 0.841 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{CGMOS}} & \textbf{0.842} & \textbf{0.536} & 0.517 & \textbf{0.526} & 0.526 & \textbf{0.908} & 0.815 & 0.859 & 0.860 \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{knn}} & & \multicolumn{4}{c|}{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Original}} & 0.821 & \textbf{0.701} & 0.521 & 0.598 & 0.604 & 0.902 & \textbf{0.942} & \textbf{0.922} & \textbf{0.9217} \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Dup}} & 0.810 & 0.519 & 0.732 & 0.607 & 0.616 & 0.921 & 0.818 & 0.867 & 0.868 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{SMOTE}} & 0.827 & 0.506 & \textbf{0.804} & 0.621 & 0.638 & 0.925 & 0.805 & 0.861 & 0.863 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{B-SMOTE}} & 0.811 & 0.494 & 0.736 & 0.591 & 0.603 & 0.927 & 0.790 & 0.853 & 0.856 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{MWMOTE}} & 0.832 & 0.504 & 0.792 & 0.616 & 0.632 & \textbf{0.928} & 0.794 & 0.856 & 0.858 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{ADASYN}} & 0.825 & 0.495 & 0.786 & 0.607 & 0.623 & \textbf{0.929} & 0.786 & 0.851 & 0.854 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{RAMOboost}} & 0.827 & 0.540 & 0.684 & 0.604 & 0.608 & 0.918 & 0.847 & 0.881 & 0.881 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{CGMOS}} & \textbf{0.840} & 0.544 & 0.766 & \textbf{0.636} & \textbf{0.646} & 0.925 & 0.842 & 0.882 & 0.883 \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{svm}} & & \multicolumn{4}{c|}{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Original}} & 0.792 & \textbf{0.632} & 0.587 & 0.609 & 0.609 & 0.882 & 0.935 & 0.908 & 0.908 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Dup}} & 0.815 & 0.543 & 0.436 & 0.484 & 0.487 & \textbf{0.981} & 0.861 & 0.917 & 0.919 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{SMOTE}} & 0.844 & 0.579 & 0.726 & 0.644 & 0.648 & 0.879 & 0.844 & 0.861 & 0.861 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{B-SMOTE}} & 0.832 & 0.475 & 0.729 & 0.575 & 0.588 & 0.893 & \textbf{0.959} & \textbf{0.924} & \textbf{0.925} \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{MWMOTE}} & 0.830 & 0.547 & 0.647 & 0.593 & 0.595 & 0.880 & 0.884 & 0.882 & 0.882 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{ADASYN}} & 0.827 & 0.536 & 0.654 & 0.589 & 0.592 & 0.880 & 0.755 & 0.813 & 0.815 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{RAMOboost}} & 0.842 & 0.556 & 0.673 & 0.609 & 0.611 & 0.968 & 0.852 & 0.906 & 0.908 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{CGMOS}} & \textbf{0.864} & 0.555 & \textbf{0.788} & \textbf{0.651} & \textbf{0.661} & 0.943 & 0.830 & 0.883 & 0.885 \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{nn}} & & \multicolumn{4}{c|}{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Original}} & 0.801 & \textbf{0.632} & 0.412 & 0.499 & 0.510 & 0.892 & \textbf{0.962} & \textbf{0.925} & \textbf{0.9258} \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Dup}} & 0.843 & 0.543 & 0.777 & 0.639 & 0.650 & 0.926 & 0.819 & 0.869 & 0.871 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{SMOTE}} & 0.840 & 0.555 & 0.750 & 0.638 & 0.645 & 0.921 & 0.820 & 0.868 & 0.869 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{B-SMOTE}} & 0.841 & 0.475 & 0.779 & 0.590 & 0.608 & 0.924 & 0.802 & 0.859 & 0.861 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{MWMOTE}} & 0.841 & 0.547 & 0.778 & 0.642 & 0.652 & 0.927 & 0.812 & 0.866 & 0.867 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{ADASYN}} & 0.842 & 0.536 & \textbf{0.786} & 0.637 & 0.649 & 0.929 & 0.803 & 0.861 & 0.863 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{RAMOboost}} & 0.841 & 0.556 & 0.743 & 0.636 & 0.643 & 0.919 & 0.830 & 0.872 & 0.873 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{CGMOS}} & \textbf{0.865} & 0.579 & 0.750 & \textbf{0.653} & \textbf{0.659} & \textbf{0.933} & 0.845 & 0.887 & 0.888 \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{rf}} & & \multicolumn{4}{c|}{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Original}} & 0.872 & \textbf{0.699} & 0.534 & 0.606 & 0.611 & 0.909 & \textbf{0.956} & \textbf{0.932} & \textbf{0.932} \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Dup}} & 0.873 & 0.682 & 0.641 & 0.661 & 0.661 & 0.917 & 0.924 & 0.921 & 0.921 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{SMOTE}} & 0.875 & 0.667 & 0.655 & 0.661 & 0.661 & 0.920 & 0.917 & 0.918 & 0.918 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{B-SMOTE}} & 0.867 & 0.653 & 0.637 & 0.645 & 0.645 & 0.920 & 0.906 & 0.913 & 0.913 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{MWMOTE}} & 0.878 & 0.658 & 0.651 & 0.655 & 0.655 & 0.920 & 0.922 & 0.921 & 0.921 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{ADASYN}} & 0.876 & 0.663 & 0.669 & 0.666 & 0.666 & 0.919 & 0.915 & 0.917 & 0.917 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{RAMOboost}} & 0.874 & 0.686 & 0.618 & 0.650 & 0.651 & 0.915 & 0.933 & 0.924 & 0.924 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{CGMOS}} & \textbf{0.884} & 0.685 & \textbf{0.678} & \textbf{0.681} & \textbf{0.681} & \textbf{0.923} & 0.926 & 0.924 & 0.924 \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Adaboost.M1}} & & \multicolumn{4}{c|}{} & \multicolumn{4}{c!{\VRule[1.5pt]}}{} \\
\specialrule{1.5pt}{0pt}{0pt}
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Original}} & 0.868 & \textbf{0.699} & 0.572 & 0.629 & 0.632 & 0.906 & \textbf{0.944} & \textbf{0.925} & \textbf{0.9247} \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{Dup}} & 0.865 & 0.622 & 0.708 & 0.662 & 0.664 & 0.922 & 0.873 & 0.897 & 0.897 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{SMOTE}} & 0.867 & 0.608 & 0.714 & 0.657 & 0.659 & 0.923 & 0.880 & 0.901 & 0.901 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{B-SMOTE}} & 0.864 & 0.581 & 0.724 & 0.644 & 0.648 & \textbf{0.927} & 0.861 & 0.893 & 0.893 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{MWMOTE}} & 0.868 & 0.600 & 0.708 & 0.650 & 0.652 & 0.922 & 0.880 & 0.901 & 0.901 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{ADASYN}} & 0.867 & 0.599 & 0.726 & 0.657 & 0.660 & 0.925 & 0.873 & 0.898 & 0.899 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{RAMOboost}} & 0.865 & 0.631 & 0.699 & 0.663 & 0.664 & 0.922 & 0.882 & 0.901 & 0.902 \\ \hline
\multicolumn{1}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}}{\textbf{CGMOS}} & \textbf{0.871} & 0.619 & \textbf{0.728} & \textbf{0.670} & \textbf{0.672} & 0.925 & 0.882 & 0.903 & 0.903 \\
\specialrule{1.5pt}{0pt}{0pt}
\end{tabular}
}
\end{center}
\caption{A summary of AUC, $Precision$, $Recall$, $\fscore$ and $\gscore$ of all competitors for the majority and minority classes produced by 6 classifiers on the artificial datasets.}
\label{tab: results_real_data}
\end{table*}
\begin{table*}[t]
\begin{center}
\scalebox{1.0}{
\centering
\begin{tabular}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}c|c|c|c|c|c|c|c!{\VRule[1.5pt]}}
\specialrule{1.5pt}{0pt}{0pt}
& \textbf{CGMOS} & \textbf{Original} & \textbf{Dup} & \textbf{SMOTE} & \textbf{B-SMOTE} & \textbf{MWMOTE} & \textbf{ADASYN} & \textbf{RAMOboost}\\
\specialrule{1.5pt}{0pt}{0pt}
\textbf{BankMarket} & \textbf{0.728} & 0.661 & 0.708 & 0.718 & 0.710 & 0.721 & 0.710 & 0.723 \\ \hline
\textbf{BloodService} & \textbf{0.733} & 0.653 & 0.648 & 0.649 & 0.651 & 0.720 & 0.714 & 0.728 \\ \hline
\textbf{BreastCancer} & 0.992 & 0.992 & \textbf{0.993} & 0.992 & 0.989 & 0.991 & 0.991 & 0.992 \\ \hline
\textbf{BreastTissue} & \textbf{0.984} & 0.899 & 0.946 & 0.932 & 0.917 & 0.937 & 0.908 & 0.943 \\ \hline
\textbf{CarEvaluation} & \textbf{0.997} & 0.995 & 0.845 & \textbf{0.997} & 0.994 & 0.996 & \textbf{0.997} & 0.995 \\ \hline
\textbf{Card'graphy} & \textbf{0.977} & 0.976 & 0.939 & 0.962 & 0.956 & 0.925 & 0.957 & 0.960 \\ \hline
\textbf{CharacterTraj} & 0.985 & 0.962 & 0.717 & 0.985 & 0.978 & 0.981 & \textbf{0.988} & 0.909 \\ \hline
\textbf{Chess} & \textbf{0.977} & 0.974 & 0.959 & 0.973 & \textbf{0.977} & 0.974 & 0.975 & 0.959 \\ \hline
\textbf{ClimateSim} & \textbf{0.908} & \textbf{0.908} & 0.861 & 0.902 & 0.863 & 0.901 & 0.901 & 0.882 \\ \hline
\textbf{Contraceptive} & \textbf{0.724} & 0.705 & 0.699 & 0.712 & 0.702 & 0.705 & 0.702 & 0.705 \\ \hline
\textbf{Fertility} & \textbf{0.673} & 0.615 & 0.594 & 0.634 & 0.592 & 0.604 & 0.639 & 0.638 \\ \hline
\textbf{Haberman} & \textbf{0.651} & 0.623 & 0.577 & 0.600 & 0.593 & 0.594 & 0.587 & 0.586 \\ \hline
\textbf{ILPD} & 0.707 & 0.687 & 0.693 & \textbf{0.715} & 0.703 & 0.702 & 0.693 & 0.703 \\ \hline
\textbf{ImgSeg} & \textbf{0.999} & 0.998 & \textbf{0.999} & 0.997 & 0.998 & 0.998 & 0.997 & 0.998 \\ \hline
\textbf{Leaf} & \textbf{0.908} & 0.880 & 0.782 & 0.852 & 0.775 & 0.836 & 0.839 & 0.821 \\ \hline
\textbf{Libras} & \textbf{0.945} & 0.922 & 0.859 & 0.929 & 0.886 & 0.936 & 0.923 & 0.883 \\ \hline
\textbf{MultipleFs} & \textbf{0.998} & \textbf{0.998} & 0.997 & \textbf{0.998} & 0.997 & 0.997 & 0.996 & 0.997 \\ \hline
\textbf{Parkinson} & 0.841 & 0.676 & 0.692 & 0.834 & 0.791 & 0.837 & \textbf{0.842} & 0.760 \\ \hline
\textbf{PlanRelax} & 0.472 & 0.457 & 0.494 & 0.469 & 0.445 & 0.467 & \textbf{0.488} & 0.464 \\ \hline
\textbf{QSAR} & \textbf{0.901} & 0.886 & 0.879 & 0.895 & 0.863 & 0.886 & 0.886 & 0.882 \\ \hline
\textbf{SPECT} & \textbf{0.820} & 0.772 & 0.803 & 0.808 & 0.811 & 0.752 & 0.801 & 0.799 \\ \hline
\textbf{SPECTF} & 0.819 & 0.819 & 0.800 & 0.805 & 0.816 & 0.812 & \textbf{0.825} & 0.795 \\ \hline
\textbf{SeismicBumps} & \textbf{0.743} & 0.735 & 0.712 & 0.727 & 0.740 & 0.732 & 0.715 & 0.691 \\ \hline
\textbf{Statlog} & \textbf{0.998} & 0.992 & 0.996 & \textbf{0.998} & 0.990 & 0.996 & 0.976 & 0.996 \\ \hline
\textbf{PlatesFaults} & \textbf{0.956} & 0.928 & 0.844 & 0.954 & 0.920 & 0.943 & \textbf{0.956} & 0.881 \\ \hline
\textbf{TAEvaluation} & \textbf{0.748} & 0.682 & 0.644 & 0.703 & 0.671 & 0.707 & 0.665 & 0.657 \\ \hline
\textbf{UserKnowledge} & \textbf{0.958} & 0.837 & 0.919 & 0.953 & 0.947 & 0.951 & 0.950 & 0.888 \\ \hline
\textbf{Vertebral} & \textbf{0.890} & 0.839 & 0.869 & 0.855 & 0.829 & 0.860 & 0.794 & 0.872 \\ \hline
\textbf{Customers} & \textbf{0.952} & 0.930 & 0.943 & 0.946 & 0.884 & 0.902 & 0.946 & \textbf{0.952} \\ \hline
\textbf{Yeast} & \textbf{0.925} & 0.792 & 0.844 & 0.907 & 0.898 & 0.900 & 0.906 & 0.851 \\
\specialrule{1.5pt}{0pt}{0pt}
\textbf{Average} & \textbf{0.864} & 0.827 & 0.808 & 0.844 & 0.830 & 0.842 & 0.842 & 0.830 \\
\specialrule{1.5pt}{0pt}{0pt}
\end{tabular}
}
\end{center}
\caption{A summary of AUC of 8 oversampling algorithms over all 30 datasets used in our evaluation. The AUC is averaged over all 6 base classifiers used in the evaluation. It could be seen from above table that CGMOS achieves best AUC measures for 24 datasets out of 30. By average, the AUC of CGMOS is at least 2 percent higher than all other competitors.}
\label{tab: results_real_data_per_file}
\end{table*}
Although, both $\fscore$ and $\gscore$ are great evaluation metrics, they are still less effective in some situations. So we also employ the ROC graphs \cite{fawcett2004roc}\cite{fawcett2006introduction}\cite{mohri2005confidence} in the evaluation. ROC graph is a two-dimensional graph, while $FP$ $rate$ and $TP$ $rate$ are its X axis and Y axis respectively.
An ROC graph basically manifest its usefulness by showing relative trade-off between benefit (true positive) and cost (false positive). One attractive property make ROC graph a good metric in imbalanced learning lies in the facts that ROC curve is insensitive to changes in class distribution. Because of this property, it is easier to see the performances of models trained by dataset oversampled by different algorithms. The goal in ROC space is to let curves be as close to upper-left-hand corner as possible, in which case the ratio between benefit and cost is maximized. To compare all test results in a more straightforward way, we also compute area under an ROC curve (AUC) which reduce the ROC performance to a single scalar value representing expected performance of the ROC curve.
\subsection{Results}
This section presents the performance of CGMOS and all the other methods on 30 real-world datasets. The same experiment procedure as the one in the experiments of the artificial dataset was conducted. All results are averged from 10 rounds of 10-folds cross-validations. A summary of the experiment results is shown in Table \ref{tab: results_real_data} and ROC graphs are shown in Figure \ref{fig: roc}.
\begin{table}[ht]
\begin{center}
\scalebox{0.75}{
\begin{tabular}{!{\VRule[1.5pt]}l!{\VRule[1.5pt]}c|c|c|c|c|c!{\VRule[1.5pt]}}
\specialrule{1.5pt}{0pt}{0pt}
& \makebox[3em]{\textbf{Knn}} & \makebox[3em]{\textbf{Rf}}& \makebox[3em]{\textbf{B-kde}} & \makebox[3em]{\textbf{Nn}} & \makebox[3em]{\textbf{Svm}} & \makebox[3em]{\textbf{Boost}}\\
\specialrule{1.5pt}{0pt}{0pt}
\textbf{Original} & 5e-5 & 1e-4 & 0.004 & 1e-4 & 0.026 & 0.04\\\hline
\textbf{Dup} & 2e-6 & 5e-5 & 3e-6 & 0.03 & 0.049 & 0.004\\\hline
\textbf{SMOTE} & 0.003 & 2e-4 & 6e-6 & 0.018 & 0.006 & 0.046\\\hline
\textbf{B-SMOTE} & 4e-6 & 7e-6 & 2e-5 & 5e-4 & 0.047 & 5e-4\\\hline
\textbf{MWMOTE} & 0.046 & 4e-5 & 1e-5 & 0.003 & 0.005 & 0.007\\\hline
\textbf{ADASYN} & 8e-6 & 7e-5 & 9e-5 & 0.005 & 1e-4 & 0.003\\\hline
\textbf{RAMOboost} & 2e-6 & 5e-5 & 3e-6 & 0.001 & 0.045 & 0.035\\
\specialrule{1.5pt}{0pt}{0pt}
\end{tabular}
}
\end{center}
\caption{A summary of $p$-values of statistical significant tests of classification results using CGMOS against each of all the other competitors. }
\label{tab: signrank}
\end{table}
Considering the classification results of the minority class, it can be observed that the proposed approach outperforms most of the compared methods under all classification algorithms in terms of $\fscore$ and $\gscore$. For $\fscore$ and $\gscore$ of the majority class, the proposed approach in most cases is only second to the original data without oversampling. This is because the original dataset is imbalanced and it favors the majority class more than the minority class during classification. Overall, CGMOS achieves the best AUC over all tests. This is because the proposed approach takes into account both of the majority and minority classes and increases the certainties of the two classes while oversampling.
The same conclusion can be made from the ROC curves shown in Fig. \ref{fig: roc}. It could be seen from the ROC curves that the proposed approach has the highest values almost everywhere. The proposed approach achieves the best result when random forest is used as the classifier. For b-kde as the classifier, the proposed approach gets the largest improvement
since the design of the proposed approach uses b-kde for certainty computations.
To get a closer view of the performances of all compared methods on each dataset, we show the AUC results of CGMOS and all other compared methods for each dataset in Table \ref{tab: results_real_data_per_file}. The table shows that by average the AUC of CGMOS is 2 percent higher than SMOTE whose AUC is 2nd highest.
Previous studies show that it is not necessary for a learning procedure to obtain best classification results when a dataset is perfectly balanced\cite{batista2004study}\cite{weiss2003learning}. How much to oversample is usually empirically determined \cite{chawla2004editorial}. To evaluate this aspect we performed another experiment in which we synthesized increasing number of minority samples and investigated how different amounts of new samples impact classification results.
Let $\delta$ denote the difference of data samples between the majority and the minority class. We performed multiple experiments where in each round we synthesized $k\delta$ new samples of the minority class where $k$ gradually increased from $0.5$ to $5$. The classification results are shown in Figure \ref{fig: increasingnum}. As can be observed in the results, CGMOS achieves the best results in all cases. Also, observe that when increasing the number of data samples added, the results of CGMOS are much more robust compared with other approaches. Note that the results of some methods such as Dup(b-kde), B-SMOTE(knn) and B-SMOTE(Adaboost.M1) are even lower than the results at the starting point where datasets are not oversampled. This highlights the advantage of CGMOS when handling oversampling on boundary samples.
\subsection{Statistical Significance Analysis}
We evaluate the statistical significance of the classification results of all competitors. Statistical significance plays a critical role in determining whether a null hypothesis should be rejected or retained, where the term null hypothesis refers to a general statement that sample observations result purely from chance. For a null hypothesis to be rejected as false, the result has to be identified as being statistically significant.
To determine whether to reject a null hypothesis, a $p$-value has to be calculated, which is the probability of observing an effect given that the null hypothesis is true \cite{JLD:11}. The null hypothesis is rejected if $p$-value is less than the significance level. The significance level is the probability of rejecting the null hypothesis given that it is true. The lower the significance level the more confident we can be in replicating the results and usually the significance level is set at $5\%$. Then a sample observation is determined to be statistically significant if $p$-value is less than $5\%$, which is formally written as $p<0.05$ \cite{SM:06}.
We follow the same protocols used in \cite{demvsar2006statistical}\cite{chen2010ramoboost}\cite{barua2014mwmote} and choose to use Wilcoxon signed-ranks test in this paper. Wilcoxon signed-ranks test is a nonparametric statistical procedure for comparing two samples that are paired, or related \cite{GC09}. Different from $t$-test \cite{BF:08}\cite{zimmerman1997teacher}\cite{demvsar2006statistical} whose null hypothesis is that the mean difference between pairs is zero, the null hypothesis of Wilcoxon signed-ranks test is that the median difference between pairs of observations is zero.
The test results are shown in Table \ref{tab: signrank}. It could be seen from the table that the $p$-value of all tests are smaller than 0.05 and pass the test.
\section{Conclusion}
In this paper, we address the imbalanced binary classification problem by proposing a novel minority oversampling strategy. Different from existing approaches, CGMOS does not randomly synthesize new data along decision boundaries. Instead, CGMOS computes the Bayes classification certainties for both the majority and minority classes and then synthesize new samples based on improvement of the certainties for samples in both classes. We prove that CGMOS can achieve better classification results compared with SMOTE. In addition, experimental results show that CGMOS outperforms known oversampling techniques using various metrics.
{\small
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Converse proofs in network information theory consist of sequences of inequalities derived by applying fundamental laws, such as the positivity of conditional entropy, mutual information, and conditional mutual information, as well as problem constraints, including any specified independences and Markov chains created from encoding and decoding constraints \cite{YeungBook,YoungHanKimBook}. The aim of these sequences of inequalities is to obtain regions that interrelate exclusively those quantities of interest, typically including various encoding rates and fidelity criteria for lossy or lossless reproduction. While intermediate rate region expressions presented in theorems often involve auxiliary random variables, not explicitly included in the problem, whose distributions are left free other than to obey a certain series of constraints, the region of interest is only ultimately obtained after optimization over the distribution of these auxiliaries \cite{Blahut_TIT_4_72}, yielding an ultimate region relating exclusively the quantities of interest.
It remains relatively uncommon to recognize proving a converse or outer bound in network information theory as a polyhedral projection, yet this process is indeed mathematically equivalent to polyhedral projection. This fact has been made clearest in the context of network coding (broadly defined) problems \cite{YeungBook}, including determining the capacity regions for multi-source network coding \cite{yanmultisourceTIT}, as well as storage repair tradeoffs for distributed storage systems with exact repair \cite{TianJSAC433}, and cache size vs. rate during the delivery phase in coded caching \cite{ChaoCache}. From a more fundamental standpoint, proving information inequalities \cite{Zhang_TIT_07_98,BookInequalities}, as well as those inequalities that linear codes must obey \cite{Kinser2011NewIneqSubspaceArra,DFZ2009Ineqfor5var}, have also been posed in a manner reminiscent of polyhedral projection. \S \ref{sec:convProofExamp} of this paper provides a series of examples demonstrating of each of these problems can be explicitly formulated as polyhedral projection problems.
A key benefit of formulating rate region calculation problems as polyhedral projection problems is that it enables them to be solved via computer algorithms rather than undergoing the arduous process of deriving them by hand. In \S \ref{sec:convProofExamp} we draw a significant distinction between \emph{verification}, which, given a putative inequality among the quantities of interest, provides a sequence of inequalities deriving it, and that of exhaustive \emph{generation} which determines a minimal set of all inequalities relating exclusively the quantities of interest. We review that verification can be posed as a linear program, while generation is posed as a polyhedral projection. Prior computational toolkits for information theory such as the information theoretic inequality prover (ITIP) \cite{itip}, xITIP \cite{xitip}, and miniITIP \cite{minitip} have aimed at \emph{verification}, while the toolkit accompanying this manuscript, the \emph{information theoretic converse prover} ITCP performs exhaustive generation by implementing the highly efficient form of polyhedral projection described within.
From a computational perspective, proving converses in network information theory via polyhedral projection is only feasible for relatively small problems, as the common formulation involves a projecting a polyhedron with a number of indeterminates that is exponential in the number of sources and messages. This implies that in order to push computer derived proofs in network information theory as far and to as large problems as possible, the polyhedral projection algorithm should be designed to reduce the complexity of the projection process as much as possible. An important first step in this direction, noted by other authors \cite{Xu_ISIT_08,CsirmazBenson} in the particular context of deriving non-Shannon information inequalities, is to utilize methods such as the Convex Hull Method \cite{lassezchm,jayantchm} or Benson's algorithm \cite{CsirmazBenson,benson1998}, which work directly in the projection space by solving carefully designed linear programs over the polyhedron to project, rather than Fourier Motzkin \cite{PermuterFME,HuynhLL92} which suffers from substantially higher complexity owing to its incremental removal of dimensions.
This manuscript focuses on a second key aspect of reducing the complexity of proving network information theory converses via polyhedral projection: exploiting symmetry. First, \S \ref{sec:typesOfSymmetry} reviews several different notions of symmetry for polyhedra, then describes their application both the polyhedra to project, and the resulting answer, for proving each of the types of information theoretic outer bounds given as examples in \S \ref{sec:convProofExamp}. Next, \S \ref{sec:symCHM} describes three distinct ways known symmetry groups can be exploited in polyhedral projection, forming a new algorithm for symmetry exploiting polyhedral projection symCHM. The complexity reductions of polyhedral projection afforded by each of these three symmetry techniques is illustrated and quantified in the context of specific example network information theory problems from \S \ref{sec:convProofExamp}. Finally, \S \ref{sec:ITCP} describes a software package for the GAP system \cite{GAP4}, ITCP \cite{ITCPsoftware}, which accompanies the manuscript and implements the methods described in \S \ref{sec:symCHM}, providing examples showing how to utilize it to calculate outer bounds for network information theory problems. A parallel series of results and algorithms in a companion manuscript \cite{Apte_NCRR}, and a second software package, the \emph{information theoretic achievability prover} (ITAP) \cite{ITAPsoftware} take up the issue of providing achievability proofs to match the converse proofs derived via techniques in this manuscript. Together, ITCP and ITAP have been used to determine the rate regions of millions of new network information theory problems \cite{Congduan_MDCS,Li_Operators} and a new theory relating network coding problems of different sizes has been developed to best harness them to solve network information theory problems of arbitrary scale \cite{Li_Operators}.
\section{Converse Proofs in Network Information Theory are Polyhedral Projections}\label{sec:convProofExamp}
This section identifies proving converses for a series of network information theory problems as polyhedral projection.
\subsection{Network Coding Rate Regions}\label{sec:netCod}
\begin{figure}
\centerline{\includegraphics[scale=0.4]{GeneralNetworkCoding1}}
\caption{\label{fig:generalnetwork}A general network model $\textsf{A}$.}
\vspace{-0.4cm}
\end{figure}
Following \cite{Li_Operators} which builds upon \cite{yanmultisourceTIT} and \cite{YeungBook}, a network coding problem consists of a labeled directed acyclic hypergraph $(\mathcal{V},\mathcal{E})$ as in Fig.\,\ref{fig:generalnetwork}, consisting of a set of nodes $\mathcal{V}$ and a set $\mathcal{E}$ of directed hyperedges in the form of ordered pairs $e=(v,\mathcal{A})$ with $v\in \mathcal{V}$ and $\mathcal{A} \subseteq \mathcal{V} \setminus v$. The nodes can be partition into three sets $\mathcal{V}=\mathcal{S} \cup \mathcal{G} \cup \mathcal{T}$: the sources $\mathcal{S}$ which have no incoming edges and exactly one outgoing edge, the intermediate nodes $\mathcal{G}$ which have incoming and outgoing edges, and the sinks $\mathcal{T}$ which no outgoing edges. The outgoing edge from each source node $s\in \mathcal{S}$ carries a source random variable $Y_s$ with entropy $H(Y_s)$. Each outgoing edge $e\in \mathcal{E},\ e=(v,\mathcal{A})$ from an intermediate node $v \in \mathcal{G}$ carries a message random variable $U_e$ that must be encoded exclusively from the messages on those edges $\textrm{In}(e)=\textrm{In}(v) = \left\{ e' \in \mathcal{E} | e'=(v',\mathcal{A}'),\ v \in \mathcal{A}'\right\}$ coming into $v$. The sinks are labeled with a demand function $\beta:\mathcal{T} \rightarrow 2^{\mathcal{S}}$ with $\beta(t), t\in \mathcal{T}$ indicating the subset of sources which must be decodable from the messages on edges $\textrm{In}(v)$ incoming at $t$. While this is the most common and straightforward representation of a network coding problem, it enables a substantial redundancy and un-necessary components to be included, and hence \cite{Li_Operators} presents both a more concise problem representation and a series of minimality conditions removing un-necessary components that are extraneous to the model.
Defining $\mathcal{L}_{i},i=1,3,4',5$ as network constraints representing source
independence, coding by intermediate nodes, edge capacity constraints, and sink nodes decoding constraints respectively,
\begin{IEEEeqnarray*}{rCl}
\mathcal{L}_{1} & = & \{[\mathbf{h}^T,\boldsymbol{r}^T]^T :h_{\boldsymbol{Y}_{\mathcal{S}}}=\Sigma_{s\in\mathcal{S}}h_{Y_{s}}\} \label{eq:rrcondef1} \\
\mathcal{L}_{3} & = & \{[\mathbf{h}^T,\boldsymbol{r}^T]^T:h_{\boldsymbol{U}_{{\rm Out}(g)}|(\boldsymbol{Y}_{\mathcal{S}\cap\mathrm{In}(g)}\cup \boldsymbol{U}_{\mathcal{E}_U\cap\mathrm{In}(g)})} =0,g\in \mathcal{G}\} \label{eq:rrcondef2} \\
\mathcal{L}_{4'}&=&\{[\mathbf{h}^T,\boldsymbol{r}^T]^T :R_e\geq h_{U_{e}}, \forall e\in\mathcal{E}_U\} \label{eq:Lratefree}\\
\mathcal{L}_{5} & = & \{[\mathbf{h}^T,\boldsymbol{r}^T]^T :h_{\boldsymbol{Y}_{\beta(t)}|\boldsymbol{U}_{\text{In}(t)}}=0,\forall t\in\mathcal{T}\} \label{eq:rrcondef4},
\end{IEEEeqnarray*}
and denoting $\mathcal{L}_{13} = \mathcal{L}_1 \cap \mathcal{L}_3$, $\mathcal{L}_{4'5} = \mathcal{L}_{4'}\cap\mathcal{L}_5$ and $\mathcal{L}_{\mathsf{A}} = \mathcal{L}_1 \cap \mathcal{L}_3 \cap \mathcal{L}_{4'} \cap \mathcal{L}_5$, a seminal result of Yan, Yeung, and Zhang \cite{YeungBook,yanmultisourceTIT} can be translated \cite{Li_PhDdissertation,CongduanTranIT2015Arxiv} to the following theorem.
\begin{theorem}[\cite{YeungBook,yanmultisourceTIT}]
\label{thm:rateregion}
The rate region of a network $\mathsf{A}$ is expressible as
\begin{equation}
\mathcal{R}_{c}(\mathsf{A})=\mathrm{Proj}_{\boldsymbol{r},\boldsymbol{\omega}}(\overline{\rm{con}(\Gamma_{N}^{*}\cap\mathcal{L}_{13})}\cap\mathcal{L}_{4'5}),\label{eq:generalrateregionfree}
\end{equation}
where ${\rm con}(\mathcal{B})$ is the conic
hull of $\mathcal{B}$, and $\mathrm{Proj}_{\boldsymbol{r},\boldsymbol{\omega}}(\mathcal{B})$
is the projection of the set $\mathcal{B}$ on the coordinates $\left[\boldsymbol{r}^T,\boldsymbol{\omega}^T \right]^T$ where $\boldsymbol{r} = \left[ R_e | e\in\mathcal{E}_U\right]$ and $\boldsymbol{\omega} = \left[H(Y_s) | s\in\mathcal{S}\right]$.
\end{theorem}
This is only an implicit characterization as it involves the region of entropic vectors $\Gamma^*_N$ \cite{YeungBook}, which is unknown and even non-polyhedral \cite{Matus_ISIT_2007} for $N\geq 4$. However, a key motivation behind the result is that $\Gamma^*_N$ can be replaced with a polyhedral outer bound $\Gamma^{\textrm{out}}_N$, e.g. the Shannon outer bound \cite{YeungBook}
\begin{equation*}
\Gamma_N := \left\{ \boldsymbol{h} \in \mathbb{R}^{2^{\mathcal{N}}\setminus \emptyset} \left| \begin{array}{c} h_{\mathcal{N}} - h_{\mathcal{N}\setminus \{i\}} \geq 0 \\
h_{i\mathcal{K}} + h_{j\mathcal{K}} - h_{\mathcal{K}} - h_{ij\mathcal{K}}
\geq 0 \\
\forall i,j \in \mathcal{N}\ \forall \mathcal{K} \subseteq \mathcal{N} \setminus \{i,j\} \end{array} \right. \right\},
\end{equation*}
obtaining the outer-bound $\mathcal{R}_o \supseteq \mathcal{R}_c$ via the polyhedral projection
\begin{equation}\label{eq:netCodRateRegion}
\mathcal{R}_o = \mathrm{Proj}_{\boldsymbol{r},\boldsymbol{\omega}} \left(\Gamma^{\textrm{out}}_N \cap\mathcal{L}_{\mathsf{A}} \right)
\end{equation}
It is precisely in this manner that a generating a converse proof in network coding can be viewed as a process of polyhedral projection. An \emph{explicit polyhedral outer bound} (EPOB) to the network coding rate region is a description of $\mathcal{R}_o$ as a series of inequalities or rays involving exclusively the rate dimensions $\mathbf{r}$ and $\mathbf{\omega}$. The resulting converse proofs track which inequalities and equalities in the parent polyhedron $\Gamma^{\textrm{out}}_N \cap\mathcal{L}_{\mathsf{A}}$ which must be summed to obtain each of the inequalities describing the projection $\mathcal{R}_o$.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{idsc_ex1.pdf
\end{center}
\caption{The IDSC instance $\textsf{A}$, with 3 source messages labeled $1,2,3$ and 3 encoders whose output messages are labeled labeled $4,5,6$. On the extreme right hand side are $7$ decoders, each of which demands a subset of the source messages.}\label{fig:idsc_ex}
\end{figure}
The next example solidifies these concepts.
\exmpl{[Network Coding Rate Region] Consider the network coding problem depicted in Fig. \ref{fig:idsc_ex}, in which a series of three sources $1,2,3$ are all made to available to each of three encoders $4,5,6$, each of which creates single message which is then overhead by a subset of decoders, with each decoder demanding only one of the three sources. This type of network coding problem with no intermediate nodes is called an independent distributed source coding (IDSC) problem. The set of constraints $\mathcal{L}_{\mathsf{A}}$ for this problem are
\begin{equation} \label{eq:ncCons}
\begin{aligned}
h_1+ h_2+ h_3 &= h_{1,2,3} ,& h_{1,2,3,4,5,6} &= h_{1,2,3}\\
h_{1,4,5,6} & = h_{4,5,6} ,& h_{2,4} &= h_4\\
h_{2,5} &= h_5 ,& h_{2,6} &= h_6\\
h_{3,4,5} &= h_{4,5} ,& h_{3,4,6} &= h_{4,6}\\
& h_{3,5,6} = h_{5,6} &
\end{aligned}
\end{equation}
With $\Gamma_{\textsf{out}}$ defined as the intersection of all Shannon-type information inequalities for $6$ discrete random variables, the associated EPOB $\mathcal{R}_o$ is given by the following inequalities
\begin{equation}\label{eq:epobex}
\begin{aligned}
\omega_1 &\geq 0 \\
\omega_2 &\geq 0 \\
\omega_3 &\geq 0 \\
\omega_2 &\leq R_4\\
\omega_2 &\leq R_5\\
\omega_2 &\leq R_6\\
2\omega_2+\omega_3 &\leq R_4+R_5\\
2\omega_2+\omega_3 &\leq R_4+R_6\\
2\omega_2+\omega_3 &\leq R_5+R_6\\
2\omega_1+6\omega_2+3\omega_3 &\leq R_4+R_5+R_6\\
\omega_1+4\omega_2+2\omega_3 &\leq 2R_4 + R_5 + R_6\\
\omega_1+4\omega_2+2\omega_3 &\leq R_4 + 2R_5 + R_6\\
\omega_1+4\omega_2+2\omega_3 &\leq R_4 + R_5 + 2R_6\\
\end{aligned}
\end{equation}
}{ex:run}
We will use this problem as a running example throughout the manuscript to illustrate the concepts, and we will demonstrate the algorithms we will describe can generate the rate region description EPOB (\ref{eq:epobex}) from the problem description shown in Fig. \ref{fig:idsc_ex}.
\subsection{Storage Repair Tradeoff in Exact Repair Distributed Storage Systems}
Of substantial recent research interest has been the tradeoff between the amount that can be stored and the amount of information that must be transferred to repair a failed disk in certain coded designs for very large distributed information storage systems \cite{dimakis10}. By drawing graphs with receivers representing different decoding conditions and different repair conditions the storage repair tradeoff can be viewed as being derived from the capacity region of a network coding problem, with the twist in exact repair that some decoders must reproduce not original sources, but instead edges in the computations\cite{TianJSAC433}. Nevertheless, utilizing the constraints $\mathcal{L}_{\mathsf{A}}$ described in the previous section with the adaptation of reproduction of edges, and a common rate limitation for edges of the same type (storage or repair), one can obtain the storage repair tradeoff as an instance of a polyhedral projection problem \cite{TianJSAC433}. While must of the research on these problems is typically focussed on a single source representing all of the information to be stored and an associated two dimensional tradeoff between storage and repair, multisource formulations \cite{Apte_CISS_2014}, for instance reflecting data with different latency requirements, or heterogenous storage sizes and repair bandwidths, are possible and can be derived as polyhedral projections of largely the same form as (\ref{eq:netCodRateRegion}).
\subsection{Cache Delivery Tradeoff in Coded Caching Systems}
A similar variant of network coding problems draws motivation of the possibility of storing encodings across high demand content at wireless basestations or receivers during off peak hours with the goal of reducing traffic on broadcast links during peak hours. A natural tradeoff studies the size of the cache and what must be broadcasted under worst case content demands. By drawing appropriate hypergraphs with different sources representing different content to request, and different hyperedges representing cache contents and the broadcasted common message under different content requests, and with a common rate limitation for the broadcast messages, and a separate common rate limitation for the cache messages \cite{ChaoCache}, it can be seen that these problems fall into the class of network coding problems described in \S \ref{sec:netCod}. Determining their tradeoffs can thus be posed as polyhedral projection problems of the form in (\ref{eq:netCodRateRegion}).
\subsection{Non-Shannon Information Inequalities}
The most common way to generate tighter outer bounds on the entropy region $\bar{\Gamma}^*_N$ for $N\geq 4$ is through polyhedral projection of a constrained Shannon outer bound in a larger series of variables \cite{Xu_ISIT_08,BookInequalities,Kaced13}. Indeed, given a collection of $N$ random variables in which one wants to derive the new non-Shannon inequality, one augments them with $k$ new ``constrained'' random variables generated according to a ``copy'' construction of being independent of some random variables while having a matching the conditional distribution with some others \cite{Xu_ISIT_08,BookInequalities}. These copy constraints imply linear equalities between certain subset entropies in all of the $N+k$ random variables, which are intersected with $\Gamma_{N+k}$ or a polyhedral entropy outer bound created in a previous step in the process. Then the entropies containing any of the $k$ constrained random variables are dropped via polyhedral projection. The resulting cone is still an outer bound to entropy, but can be smaller than the Shannon outer bound. As such, the process of proving non-Shannon inequalities itself can be posed as a polyhedral projection in which a Shannon outer bound is projected after intersecting it with linear constraints.
\subsection{Linear Rank Inequalities}
Inequalities for the dimensions of Minkowski sums of subsets of a set of subspaces can be derived through a similar process of polyhedral projection \cite{Hammer2000ShannEntr,KinserSubspace,DFZ2009Ineqfor5var}. These inequalities reflect extra constraints that the rates and entropies of messages created through linear network codes must obey \cite{DFZ2009Ineqfor5var,DFZ_Insuff}. The most common method utilized to generate new such inequalities observes that random variables that are created through linear transformations (over a finite field) of uniformly distributed random variables (over the same finite field), must have \emph{common information}. In particular, if $X$ and $Y$ are two random variables associated with such a linear construction, then there exists a random variable $Z$, the \emph{common information} such that $H(Z|X)=H(Z|Y) = 0$ and $H(Z) = I(X;Y)$ \cite{Hammer2000ShannEntr,DFZ2009Ineqfor5var}. Forcing the existence of common information restricts the type of distributions that are possible for $X$ and $Y$, but these include all $X$ and $Y$ created through such a linear construction. Thus, one can obtain outer bounds on the region of entropic vectors (and hence subspace dimensions) that be be achieved via such linear transformations from uniforms via the following process. One again considers $N$ such subspaces, then creates $k$ random variables representing such common informations. One then intersects series of inequalities that must be obeyed by such rank vectors, which for instance include the Shannon type inequalities $\Gamma_{N+k}$, with the equalities (cases of $H(Z|X)=H(Z|Y) = 0$ and $H(Z) = I(X;Y)$) implied by the definition of the $k$ common informations. Then, one projects out any dimensions associated with the common information variables. The resulting polyhedral cones inequalities hold for any $N$ subspace ranks. This process, together with subspace constructions achieving each extreme ray, has been utilized to determine all of the inequalities among subset-sum ranks of $N\leq 5$ subspaces \cite{Hammer2000ShannEntr,DFZ2009Ineqfor5var}, and is at present being used by several scholars to study the cases $N\geq 6$. By putting them as outer bounds (not for entropy in general, but those entropies achievable with linear codes) in (\ref{eq:netCodRateRegion}), one can obtain outer bounds on the region of rates achievable with linear network codes.
\section{Polyhedral Symmetries in Network Information Theory} \label{sec:typesOfSymmetry}
Each of the polyhedral bounds to multiterminal information theory problems obtained through polyhedral projection as reviewed in \S \ref{sec:convProofExamp} can exhibit high amounts of symmetry, which in turn may be exploited to reduce the complexity of obtaining an explicit polyhedral outer bound. In this section, we first present in \S \ref{sec:polyNot} and \S \ref{sec:polySym} formal definitions for various notions of symmetry for a polyhedron, and then demonstrate in \S \ref{sec:entSym} and \S \ref{sec:netCodSym} how these definitions apply to the types of polyhedral cones encountered in the network information theory polyhedral projection problems detailed in \S \ref{sec:convProofExamp}.
\subsection{Related Concepts from Polyhedra}\label{sec:polyNot}
\newcommand{\textrm{homog}}{\textrm{homog}}
\newcommand{\mathcal{P}}{\mathcal{P}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathbf{s}}{\mathbf{s}}
\newcommand{\mathbf{t}}{\mathbf{t}}
\newcommand{\mathcal{C}}{\mathcal{C}}
\newcommand{\mathsf{CSG}}{\mathsf{CSG}}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\mathbf{v}}{\mathbf{v}}
\newcommand{\mathcal{V}}{\mathcal{V}}
\newcommand{\psi}{\psi}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathsf{ASG}}{\mathsf{ASG}}
\newcommand{\mathbb{GL}}{\mathbb{GL}}
\newcommand{\mathsf{RSG}}{\mathsf{RSG}}
\newcommand{\mathsf{PSG}}{\mathsf{PSG}}
\newcommand{\mathbb{A}}{\mathbb{A}}
\newcommand{\mathcal{F}}{\mathcal{F}}
Polyhedra have several important cascading notions of symmetry formalized through groups. To properly define these groups, notation for several important concepts in polyhedra must be fixed. A polyhedron is typically expressed as the set of solutions to a system of linear inequalities
\begin{equation}
\mathcal{P} = \left\{ \mathbf{x} \in \mathbb{R}^d | \mathbb{A}' \mathbf{x} \geq \mathbf{b}' \right\}
\end{equation}
for some $\mathbb{A}' \in \mathbb{R}^{M\times d}$ and $\mathbf{b}' \in \mathbb{R}^{M}$. When working computationally with polyhedra, it is common to consider exclusively polyhedral cones, for which $\mathbf{b}' = \mathbf{0}$, as when $\mathbf{b}'$ is non-zero we can pass to the \emph{homogenization}
\begin{equation}
\textrm{homog}(\mathcal{P}) = \left\{ \left[\begin{array}{c} x_0 \\ \mathbf{x} \end{array} \right] \in \mathbb{R}^{d+1} \left| -x_0 \mathbf{b}' + \mathbb{A} \mathbf{x}' \geq \mathbf{0} \right. \right\}
\end{equation}
which is a polyhedral cone of one higher dimension.
A face $\mathcal{F}$ of a polyhedral cone $\mathcal{C}$ is a convex subset $\mathcal{F} \subset \mathcal{C}$ such that any line segment in $\mathcal{C}$ with a relative interior point in $\mathcal{F}$ must also have both of its end points in $\mathcal{F}$ \cite{Rockafellar_Book}. A polyhedral cone $\mathcal{C}$ is said to be pointed if it contains no lines, and the discussion for the remainder of the manuscript refers to pointed polyhedral cones.
A face of a pointed polyhedral cone that is one dimensional is called an \emph{extreme ray} $\mathcal{R}$ of the cone, and can be represented with a vector $\mathbf{v}$ as $\mathcal{R} = \left\{\gamma\mathbf{v} | \gamma \geq 0\right\}$. Pointed polyhedral cones have a finite number of such extreme rays, and thus in addition to being represented as the set of solutions to a finite homogeneous system of linear inequalities
\begin{equation}
\mathcal{C} = \left\{ \mathbf{x} \in \mathbb{R}^d \left| \mathbb{A} \mathbf{x} \geq \mathbf{0} \right. \right\}
\end{equation}
with $\mathbb{A}$ having linearly independent columns, pointed polyhedral cones can also be represented as the conic hull of a finite set $\mathcal{V}$ containing representatives of the set of their extreme rays $\mathcal{T}$
\begin{equation}
\mathcal{C} =\textrm{conic}(\mathcal{T}) = \left\{ \sum_{\mathbf{v} \in \mathcal{V}} \gamma_{\mathbf{v}} \mathbf{v} \left| \gamma_{\mathbf{v}} \geq 0\ \forall \mathbf{v} \in \mathcal{V} \right. \right\}.
\end{equation}
The matrix representatives $\mathbb{A}$ for the inequalities and $\mathbb{V}=[\mathbf{v} | \mathbf{v} \in \mathcal{V}]^T$ for the extreme rays are thus known as a \emph{double descriptions pair}.
Likewise, each higher dimensional face $\mathcal{F}$ of a pointed polyhedral cone $\mathcal{C}$ can be represented as the conic hull $\mathcal{F} = \textrm{conic}(\mathcal{A})$ of a certain subset $\mathcal{A} \subseteq \mathcal{V}$ of the extreme ray representatives. The set of faces of the cone form a lattice via the containment relation called the face lattice\cite{FukudaPolyCompNotes}.
\subsection{Symmetry Groups of Polyhedra}\label{sec:polySym}
The most expansive and general notion of symmetry for a polyhedral cone $\mathcal{C}$ is the \emph{combinatorial symmetry group}.
\defn{[Combinatorial Symmetry Group] The combinatorial symmetry group $\mathsf{CSG}(\mathcal{C})$ of a polyhedral cone is the subgroup of the symmetric group of order $| \mathcal{T} |$ consisting of those permutations (bijections $\psi:\mathcal{T} \rightarrow \mathcal{T}$) of the set extreme rays that leave the face lattice invariant.}
In other words, $\psi$ is a combinatorial symmetry if, for any subset $\mathcal{A} \subseteq \mathcal{T}$ whose conic hull $\textrm{conic}(\mathcal{A})$ is a face of $\mathcal{C}$, $\psi(\mathcal{A}) = \left\{ \psi(\mathcal{R}) | \mathcal{R} \in \mathcal{A} \right\}$ is also a face of $\mathcal{C}$.
While the combinatorial symmetry group is the most general notion of symmetry of a polyhedral cone, since it is represented as a subgroup of the symmetric group of order $|\mathcal{T}|$, for cones even of moderate size, it rapidly becomes infeasible to represent, determine, and work with. Hence, one often narrows scope to the \emph{affine symmetry group}.
\defn{[Affine Symmetry Group] The affine symmetry group $\mathsf{ASG}$ consists of only those combinatorial symmetries $\psi$ which can be represented as resulting from an invertible linear transformation from the general linear group $\mathbb{A}_{\psi} \in \mathbb{GL}$, in the sense that for each extreme ray $\mathcal{R}\in\mathcal{T}$, $\psi(\mathcal{R}) = \mathbb{A}_{\psi} \mathcal{R}$. }
A further subgroup of the affine symmetry group is the restricted affine symmetry group $\mathsf{RSG}$, can be defined after selecting a series of extreme ray representatives $\mathcal{V}$, and replacing each extreme ray $\mathcal{R} = \{\gamma \mathbf{v} | \gamma \geq 0 \}$ with its particular selected representative $\mathbf{v}$, effectively selecting a scale for this representative.
\defn{[Restricted Symmetry Group] The restricted symmetries of a polyhedral cone $\mathcal{C}$ with selected extreme ray representatives $\mathcal{V}$ are then those bijections $\psi:\mathcal{V} \rightarrow \mathcal{V}$ for whom there exists a $\mathbb{A}_{\psi} \in \mathbb{GL}$ such that $\psi(\mathbf{v}) = \mathbb{A}_{\psi} \mathbf{v}$.}
It is important to note that the same polyhedral cone can thus have multiple restricted symmetry groups via the selection of difficult scalings for the representatives of the extreme rays.
A final, smallest symmetry group for a polyhedron, are the \emph{coordinate permutation symmetries}.
\defn{[Coordinate Permutation Symmetries] The coordinate permutation symmetry group
$\mathsf{PSG}(\mathcal{C})$ of a polyhedral cone, are those affine symmetries $\psi(\mathcal{R}) = \mathbb{A}_{\psi} \mathcal{R}$ which can be represented by permutation matrices, so that the columns of $\mathbb{A}_{\psi}$ are the columns of the identity matrix.}
\subsection{Polyhedral Symmetry Groups and the Entropy Region}\label{sec:entSym}
To illustrate these ideas and definitions in an information theoretic context, in this section
we will consider a sequence of examples concerning the entropy cone $\bar{\Gamma}^*_N$ and its outer bound, the Shannon outer bound $\Gamma_N$.
Let's begin by considering the case of the entropy region for two variables $\Gamma^*_2 = \Gamma_2$. This three dimensional polyhedral cone can be described by three inequalities
\begin{equation}
\Gamma^*_2 = \left\{ \left[\begin{array}{c} h_1 \\ h_2 \\ h_{12} \end{array} \right] \left| \begin{array}{c} h_{12} - h_1 \geq 0 \\ h_{12}- h_2 \geq 0 \\ h_1+h_2 -h_{12} \geq 0 \end{array} \right. \right\}
\end{equation}
and also can be represented as the conic hull of three extreme rays
\begin{equation}
\Gamma_2^* = \left\{ \left[ \begin{array}{c} 1 \\ 0 \\ 1 \end{array} \right], \left[ \begin{array}{c} 0 \\ 1 \\ 1 \end{array} \right],\left[ \begin{array}{c} 1 \\ 1 \\ 1 \end{array} \right] \right\}
\end{equation}
Each subset of two of these three extreme rays forms a two dimensional face, so any permutation of the three rays leaves the face lattice invariant, implying that $\mathsf{CSG}(\Gamma^*_2) = \mathbb{S}_3$, the symmetric group of order three. As there are an equal number of rays as there are dimensions, each of the six ray permutations in $\mathbb{S}_3$ is representable as an invertible matrix transformation, and so the affine symmetry group and restricted affine symmetry group match the combinatorial symmetry group $\mathsf{RSG}(\Gamma^*_2)=\mathsf{ASG}(\Gamma^*_2) = \mathsf{CSG}(\Gamma^*_2)$. The coordinate permutation group, however is a strict subgroup of order 2, which can be viewed as $\mathsf{PSG}(\Gamma_2)=\mathbb{S}_2$, as it contains only the identity and the permutation swapping the coordinates $h_1$ and $h_2$ and leaving $h_{12}$ fixed.
Moving on to $\Gamma_3 = \bar{\Gamma}^*_3$, we have the inequalities
\begin{equation*}
\bar{\Gamma}^*_3=\left\{ \begin{bmatrix} h_1 \\ h_2 \\ h_{12} \\ h_{3} \\ h_{13} \\ h_{23} \\ h_{123} \end{bmatrix} \in \mathbb{R}^7 \left| \begin{array}{c} -h_{12} +h_{123} \geq 0 \\ -h_{13}+h_{123} \geq 0 \\ h_{23}-h_{123} \geq 0 \\ h_1 + h_2 -h_{12} \geq 0 \\ h_1 + h_3 - h_{13} \geq 0 \\ h_2 + h_3 - h_{23} \geq 0 \\ -h_1 + h_{12} + h_{13} - h_{123} \geq 0 \\
-h_2 + h_{12} + h_{23} - h_{123} \geq 0 \\
-h_3 + h_{13} + h_{23} - h_{123} \geq 0
\end{array} \right. \right\}
\end{equation*}
The extreme ray representation $\bar{\Gamma}^*_3=\textrm{conic}(\mathcal{V})$ is
\begin{equation}\label{eq:extrRepShan3}
\mathcal{V}:=\left\{
\begin{bmatrix}
0 \\ 1 \\ 1 \\ 1 \\ 1 \\ 1 \\ 1
\end{bmatrix},
\begin{bmatrix}
1 \\ 0 \\ 1 \\ 1 \\ 1 \\ 1 \\ 1
\end{bmatrix},
\begin{bmatrix}
1 \\ 1 \\ 1 \\ 1 \\ 1 \\ 1 \\ 1
\end{bmatrix},
\begin{bmatrix}
0 \\ 1 \\ 1 \\ 0 \\ 0 \\ 1 \\ 1
\end{bmatrix},
\begin{bmatrix}
1 \\ 0 \\ 1 \\ 0 \\ 1 \\ 0 \\ 1
\end{bmatrix},
\begin{bmatrix}
0 \\ 0 \\ 0 \\ 1 \\ 1 \\ 1 \\ 1
\end{bmatrix},
\begin{bmatrix}
1 \\ 1 \\ 1 \\ 0 \\ 1 \\ 1 \\ 1
\end{bmatrix},
\begin{bmatrix}
1 \\ 1 \\ 2 \\ 1 \\ 2 \\ 2 \\ 2
\end{bmatrix}
\right\}
\end{equation}
where we will denote in the same order $\mathcal{V}:=\{\mathbf{v}_1,\mathbf{v}_2,\ldots,\mathbf{v}_8\}$. In addition to the 8 one dimensional faces formed by these rays, and the 1 seven dimensional face formed by all of $\bar{\Gamma}^*_3$ itself, some computation finds 166 other faces, comprised of the numbers of faces of dimensions as indicated in Table \ref{tbl:numFaces}. Finding the subgroup of $\mathbb{S}_8$ that fixes these faces using \textsf{GAP}, we find the combinatorial symmetry group $\mathsf{CSG}(\bar{\Gamma}^*_3)$ includes 72 permutations generated by the following 5 permutations: swapping $\mathbf{v}_1$ with $\mathbf{v}_2$, swapping $\mathbf{v}_1$ with $\mathcal{R}_7$, swapping $\mathbf{v}_5$ with $\mathbf{v}_6$, swapping $\mathbf{v}_4$ with $\mathbf{v}_5$, and swapping $\mathbf{v}_3$ with $\mathbf{v}_8$.
In fact, each of the generators for the combinatorial symmetry group is expressible with an invertible linear transformation among the extreme ray representatives depicted in (\ref{eq:extrRepShan3}), which proves that the group of restricted symmetries is in this instance again equivalent to the group of combinatorial symmetries. Indeed, swapping $\mathbf{v}_1$ with $\mathbf{v}_2$ and leaving the other rays fixed can be achieved by transforming $h_1 \mapsto h_2+h_{13}-h_{23}=H(1|3)+I(2;3)$ and $h_2 \mapsto h_1 + h_{23} - h_{13} = H(2|3)+I(1;3)$ and leaving all other coordinates unchanged. Similarly, swapping $\mathbf{v}_1$ with $\mathbf{v}_7$ and leaving the other rays fixed can be achieved by transforming $h_1 \rightarrow h_3+h_{12}-h_{23} = H(1|2)+I(2;3)$.
Thus, for $\Gamma_3=\bar{\Gamma}^*_3$, like $\Gamma_2=\Gamma_2^*$, the $\mathsf{CSG}(\bar{\Gamma}^*_3)=\mathsf{ASG}(\Gamma^*_3) =\mathsf{RSG}(\Gamma^*_3)$
The coordinate permutation group $\mathsf{PSG}(\Gamma_3)=\mathbb{S}_3$, as it contains only the identity and the permutation swapping the coordinates $h_1$ and $h_2$ and leaving $h_{12}$ fixed.
\begin{table*}
\centering
\begin{tabular}{cccccccccccccccc}
& 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 \\
$\Gamma_2$ & 3 & 3 & 1 & & & & & & & & & & & & \\
$\Gamma_3$ & 8 & 27& 49& 51& 30& 9& 1 & & & & & & & & \\
$\Gamma_4$ & 41 & 510 & 3246 & 12654 & 32957 & 60130 & 78868 & 75241 & 52232 & 26112 & 9189 & 2188 & 330 & 28 & 1 \\
\end{tabular}
\caption{Numbers of faces of each dimension for the first few Shannon outer bounds.}\label{tbl:numFaces}
\end{table*}
Moving on to $\Gamma_4 \supsetneq \bar{\Gamma}_4^*$. In general, the Shannon outer bound $\Gamma_N$ is defined via the $\binom{N}{2}2^{N-2} + N$ elemental inequalities $H(X_i | X_{\mathcal{N}\setminus \{i\}}) \geq 0, i \in\mathcal{N}$ and $I(X_i;X_j | X_{\mathcal{K}}) \geq 0,\ i,j \in \mathcal{N},\ \mathcal{K} \subseteq \mathcal{N}\setminus \{i,j\}, i\neq j$ with $\mathcal{N} =\{1,\ldots,N\}$. At $N=4$, there are thus 28 such inequalities, and the cone has 41 extreme rays \cite{Hammer2000ShannEntr}, while tedious computation finds the number of faces of each dimension summarized in Table \ref{tbl:numFaces}. The combinatorial symmetry group $\mathsf{CSG}(\Gamma_4)$ is a subgroup of $\mathbb{S}_{28}$ which fixes this set of faces. Calculating such a stabilizer subgroup through ordinary routines is daunting for such a large group, so finding the generators and cardinality of $\mathsf{CSG}(\Gamma_N)$ for $N\geq 4$ via a generic such approach appears a daunting task. However, by leveraging the software \textsf{sympol} \cite{sympol} or the GAP package \textsf{polyhedral} \cite{polyhedralSikiric}, determining the affine symmetry group $\mathsf{ASG}(\Gamma_4)$ and restricted affine symmetry group $\mathsf{RSG}(\Gamma_4)$ is possible. By performing this calculation, and a similar one for $\mathsf{ASG}(\Gamma_5),\mathsf{ASG}(\Gamma_6),\mathsf{ASG}(\Gamma_7)$, and studying the structure of these groups with GAP, the following observation can be deduced.
\thml{[Affine Symmetries of the Shannon Outer Bound] For $N\in\{3,4,5,6,7\}$ the affine symmetry group of the Shannon outer bound, $\mathsf{ASG}(\Gamma_N)=\mathsf{RSG}(\Gamma_N)$ is isomorphic to $\mathbb{C}_2 \times \mathbb{S}_N \times S_N$, the direct product between the cyclic group of order $2$, the symmetric group of order $N$, and the symmetric group of order $N$. The normal subgroup isomorphic to $\mathbb{C}_2$ fixes each of the elemental inequalities of the form $H(X_i|X_{\mathcal{N}\setminus i})\geq0$, while swapping $I(X_i;X_j|X_{\mathcal{K}})\geq 0$ with $I(X_i;X_j|X_{(\mathcal{K}\cup\{i,j\})^c})\geq 0$ for each $i,j\in\mathcal{N}$, $i\neq j$ and $\mathcal{K}\subseteq \mathcal{N}\setminus \{i,j\}$. One normal subgroup isomorphic to $S_N$ permutes the elemental inequalities $H(X_i|X_{\mathcal{N}\setminus i})\geq0$ to $H(X_{\pi(i)}|X_{\mathcal{N}\setminus \{\pi(i)\}})$ for each $\pi \in \mathbb{S}_{\mathcal{N}}$ while leaving all of the other elemental inequalities of the form $I(X_i;X_j|X_{\mathcal{K}})\geq 0$ fixed. A second normal subgroup leaves each of the elemental inequalities of the form $H(X_i|X_{\mathcal{N}\setminus i})\geq0$ fixed while permuting $I(X_i;X_j|X_{\mathcal{K}})\geq 0$ to $I(X_{\pi(i)};X_{\pi(j)}|X_{\pi(\mathcal{K})}) \geq 0$ for each $\pi \in \mathbb{S}_{\mathcal{N}}$. Accordingly $|\mathsf{ASG}(\Gamma_N)|=2(N!)^2$.}{thm:shanSym}
As the coordinate permutation symmetry group $\mathsf{PSG}(\Gamma_N)=\mathbb{S}_N$ is created by random variable permutations, it is clear that the number of affine symmetries of the entropy region is double the square of the number of coordinate permutation symmetries, and is hence far larger.
\subsection{Polyhedral Symmetries in Multiterminal Converse Proofs}\label{sec:netCodSym}
When proving a converse via polyhedral projection for the types of network information theory problems summarized in \S \ref{sec:convProofExamp}, in addition to the symmetries of the entropy cone and its outer bound, the symmetries of the network information theory problem play a role. In this section, we describe some of these problem symmetries in the context of the type of general network coding problems as described in \S \ref{sec:netCod} and encapsulated in (\ref{eq:netCodRateRegion}), as each of the remaining types of problems described in the remaining subsections of \S \ref{sec:convProofExamp} each have the same form as (\ref{eq:netCodRateRegion}).
In a polyhedral projection problem, there are two polyhedra: the parent polyhedron and the polyhedron resulting from the projection, and both of these can exhibit polyhedral symmetries. We will observe that both of these types of symmetries can aid in reducing the complexity of computing an explicit rate region. For projections of the form (\ref{eq:netCodRateRegion}), one can deduce some of the symmetries of the parent polyhedron by restricting the symmetries of the selected outer bound to entropy $\Gamma^o_N$ to those that fix the set of linear constraints $\mathcal{L}_{\mathsf{A}}$, as summarized in the following lemma.
\leml{[Computing Some Network Coding Symmetries] Subgroups of the symmetries of the parent polyhedron $\Gamma^o_N\cap\mathcal{L}_{\mathsf{A}}$ in (\ref{eq:netCodRateRegion}) can be determined as follows
\begin{eqnarray}
\mathsf{ASG}(\Gamma^o_N\cap\mathcal{L}_{\mathsf{A}}) &\geq& \textrm{Stabilizer}(\mathsf{ASG}(\Gamma^o_N),\mathcal{L}_{\mathsf{A}}) \\
\mathsf{RSG}(\Gamma^o_N\cap\mathcal{L}_{\mathsf{A}}) &\geq& \textrm{Stabilizer}(\mathsf{RSG}(\Gamma^o_N),\mathcal{L}_{\mathsf{A}}) \\
\mathsf{PSG}(\Gamma^o_N\cap\mathcal{L}_{\mathsf{A}}) &\geq& \textrm{Stabilizer}(\mathsf{PSG}(\Gamma^o_N),\mathcal{L}_{\mathsf{A}})
\end{eqnarray}
where $ \textrm{Stabilizer}$ determines the stabilizer subgroup of the first argument that fixes the second argument, and the groups act on the set of constraints $\mathcal{L}_{\mathsf{A}}$ in the natural way (multiplication of each inequality by an invertible matrix for $\mathsf{ASG}$ and $\mathsf{RSG}$ and permutation of random variable labels for $\mathsf{PSG}$).
}{lem:someSym}
However, as the following example shows, because they remove both dimensions and inequalities rendered redundant, the equalities in the problem constraints $\mathcal{L}_{\mathsf{A}}$ can in fact yield symmetries beyond those among those of $\Gamma^o_N$, as they remove some of these inequalities, enabling a larger group.
\exmpl{[Network Coding Constraint Symmetry] Substituting the equalities (\ref{eq:ncCons}) associated with the constraints for the network coding problem of Example \ref{ex:run} depicted in Fig. \ref{fig:idsc_ex} into the inequalities describing the entropy cone $\Gamma_6$, one obtains a cone with 264 inequalities, of which 167 can be removed as being redundant after substitution of the equalities (\ref{eq:ncCons}), leaving a cone with 79 inequalities and 54 dimensions associated with the subset entropies remaining after the substitutions from (\ref{eq:ncCons}). Calculating this cone's restricted symmetry group with the GAP package \textsf{polyhedral} \cite{polyhedralSikiric}, one finds that $\mathsf{RSG}(\Gamma_N \cap \mathcal{L}_{\mathsf{A}})$ is a group with 20 generators and $|\mathsf{RSG}(\Gamma_N \cap \mathcal{L}_{\mathsf{A}})|$ is 358,318,080. Recalling that $|\mathsf{RSG}(\Gamma_6)|= 2(6!)^2=1,036,800$ it is easily discerned that the substitution of the equalities have removed so many inequalities that $|\mathsf{RSG}(\Gamma_N \cap \mathcal{L}_{\mathsf{A}})|$ is far larger than $\textrm{Stabilizer}(\mathsf{RSG}(\Gamma^o_N),\mathcal{L}_{\mathsf{A}})$.}{ex:bigGroup}
Nonetheless, when the polyhedral cone $\Gamma_N \cap \mathcal{L}_{\mathsf{A}}$ becomes too high dimensional to be amenable to direct computation of symmetry groups, Lem. \ref{lem:someSym} can be used to infer some polyhedral symmetries implied by the problem. With respect to sets of linear constraints $\mathcal{L}_{\mathsf{A}}$ arising from network coding problems in particular $ \textrm{Stabilizer}(\mathsf{PSG}(\Gamma^o_N),\mathcal{L}_{\mathsf{A}})$ is known as the \emph{network symmetry group} \cite{Li_Operators}.
\defn{[Network Symmetry Group] The network symmetry group of a network coding rate region problem of the form described in \S \ref{sec:netCod}, is the subgroup of random variable permutations $\mathbb{S}_N$ that fixes the constraint set $\mathcal{L}_{\mathsf{A}}$ setwise.}
A key benefit of the network symmetry group for network coding rate region problems is that it also forms at least a subgroup of the coordinate symmetry group of the projected rate region $\mathsf{PSG}(\mathcal{R}_o)$. Indeed, permuting the network coding source and edge rates with the same permutation as their associated random variables leaves the rate region fixed.
\exmpl{[Network Coding Rate Region Symmetry] The network symmetry group for the problem from Example \ref{ex:run} depicted in Fig. \ref{fig:idsc_ex} is the subgroup of random variable label permutations $\mathbb{S}_{6}$ that fixes the constraints depicted in (\ref{eq:ncCons}). As will be shown with another example in \S \ref{sec:ITCP}, the GAP package ITCP provided along with the manuscript can calculate this network symmetry group as $\langle (5,6),(4,5) \rangle$, which is a group of order $6$ whose generators swap encoders $5$ and $6$, and encoders $4,5$ independently. Observe that $\mathcal{R}_o$ in (\ref{eq:epobex}) is fixed under this group, and hence it is a symmetry of both $\mathcal{R}_o$ and the parent polyhedron. While this group is tiny relative to the restricted symmetry group for this problem uncovered in Example \ref{ex:bigGroup}, even it is already sufficient to demonstrate a substantial reduction in the computational complexity of polyhedral projection in the coming sections.
}{ex:netSymGroup}
\section{Polyhedral Projection with CHM} \label{sec:chm}\label{sec:symCHM}
With the notions of polyhedral symmetries described in \S \ref{sec:polySym}, and the demonstration of their substantial size and non-triviality in network information theory problems provided by the examples in \S \ref{sec:entSym} and \S \ref{sec:netCodSym}, we are ready to turn to the subject of how to exploit symmetry when performing polyhedral projection to determine network information theory rate region as described in \S \ref{sec:convProofExamp}.
The symmetry exploiting polyhedral projection algorithm we will describe builds upon the convex hull method \cite{lassezchm} for projecting polytopes (bounded polyhedra). As such, we will first review the operation of the convex hull method in this section. Thereafter, in \S \ref{sec:symchm} we will detail three different ways that known symmetry groups can be exploited to reduce the complexity of the convex hull method, yielding together a new algorithm \textsf{symChm}.
\subsection{Review: The Convex Hull Method}\label{sec:chm}
Denote by $\textsf{proj}_k(\mathcal P)$ the projection of a polyhedron $\mathcal{P}\subseteq \mathbb{R}^d$ onto its first $k$ dimensions
\begin{equation}
\textsf{proj}_k(\mathcal{P})\triangleq \left\{\mathbf{x} \in \mathbb{R}^k \left| \exists \mathbf{y} \in \mathbb{R}^{d-k}\ \textrm{ s.t. }\ (\mathbf{x},\mathbf{y}) \in \mathcal{P}\right. \right\}
\end{equation}
CHM \cite{lassezchm} is an algorithm to project polytopes by building successively better inner bounds to the projected polytope $\textsf{proj}_k (\mathcal P)$ via the solution of carefully selected linear programs over $\mathcal P$. The pseudocode for our implementation \cite{jayantchm} of CHM is listed as algorithm \ref{algchm}. Note that we assume that the input polyhedron $\mathcal P$ is bounded and full-dimensional, so that the dimension of its affine hull is $d$. Tests for full-dimensionality and methods for the elimination of any redundant variables and inequalities of the input can be be implemented based on the guidelines in \cite{fukuda2000frequently}.
\begin{algorithm}
\caption{Convex Hull Method}\label{algchm}
\DontPrintSemicolon
\KwIn{Polyhedron $\mathcal{P}$ and $k$, the dimension of projection}
\KwOut{Vertex-facet set pair $(V,H)$ of $\textsf{proj}_k(\mathcal{P})$}
$(V,H)\leftarrow$ \textsf{initialhull}($\mathcal{P},k$)\;
\While{$\exists \{\mathbf h\mathbf x\geq \mathbf b\}\in H$ s.t. \textsf{isterminal}($\mathcal{P},\mathbf{h},\mathbf{b})= 0$ } {
$\mathbf{v}\leftarrow$ \textsf{extremepoint}($\mathcal{P},\mathbf{h}$)\;
$H\leftarrow\textsf{updatehull}(\mathbf{v},V,H)$\;
$V\leftarrow V\cup{\mathbf{v}}$
}
\Return{$(V,H)$}\;
\end{algorithm}
\begin{procedure}
\DontPrintSemicolon
\KwIn{Polyhedron $\mathcal{P}$ and $k$, the dimension of projection}
\KwOut{$(V,H)$ corresponding to initial hull of projection}
$V\leftarrow \phi$\;
$\mathbf{p}_1\leftarrow$ \textsf{extremepoint}($\mathcal{P},[1,0,\hdots 0]_{1\times k}$)\;
$\mathbf{p}_2\leftarrow$ \textsf{extremepoint}($\mathcal{P},[-1,0,\hdots 0]_{1\times k}$)\;
$V\leftarrow V\cup\{\mathbf{p}_1,\mathbf{p}_2\}$\;
$\textbf{h}\leftarrow \textsf{hyperplane}(V)$\;\label{proc:inithull_hyperplane}
$i\leftarrow 3$\;
\While{$i\leq k$\label{w1}}{
$\mathbf{p}_i\leftarrow$ \textsf{extremepoint}($\mathcal{P},\mathbf{h}$)\;
$V\leftarrow V\cup\{\mathbf{p}_i\}$\;
$\mathbf{h}\leftarrow$\textsf{hyperplane}($V$)\;
$i\leftarrow i+1$\;\label{w2}
}
$H\leftarrow$ \textsf{facets}($V$)\;\label{proc:inithull_facets}
\Return{$(V,H)$}\;
\label{inithull}
\caption{initialhull($\mathcal{P},k$)}\label{alg:initialHull}
\end{procedure}
\begin{procedure}
\DontPrintSemicolon
\KwIn{Set of vertices $V$ and corresponding set of inequalities $H$, new vertex $\mathbf{v}$}
\KwOut{Set of inequalities $H'$ corresponding to $V'=V\cup\{\mathbf{v}\}$}
$\mathcal C\leftarrow \textsf{homog}(H)^\circ$\;\label{uph:homog}
$\{\mathbf{r}_1,...,\mathbf{r}_t\}\leftarrow$ \textsf{DDiteration}($\mathcal{C},\{[1 \hspace{2mm} v_1,\hdots, v_d]\mathbf{y}\leq 0\}$)\;
$H'\leftarrow\{\textsf{ineq}(\mathbf r_1),\hdots,\textsf{ineq}(\mathbf r_t)\}$\;
\Return{$H'$}\;
\label{updatehull}
\caption{updatehull($\mathbf{v},V,H$)}
\end{procedure}
\noindent The algorithm relies on the fact that, if $\mathbf{c} \in \mathbb{R}^k$, then a point on the boundary of $\textrm{proj}_k(\mathcal P)$ that attains the solution to the linear program with cost vector $\mathbf{c}$ over $\textrm{proj}_k(\mathcal P)$, can be found by projecting the extreme point in $\mathcal P$ attaining the solution of the linear program with cost vector $[\mathbf{c}^T,\mathbf{0}_{d-k}^T]^T$ over $\mathcal P$, so that
\begin{eqnarray}
\min_{\mathbf{x} \in\textrm{proj}_k \mathcal P} \mathbf{c}^T \mathbf{x} &=&
\min_{\mathbf{y} \in\mathcal P} [\mathbf{c}^T,\mathbf{0}_{d-k}^T] \mathbf{x}, \ \ \textrm{and}, \\
\arg \min_{\mathbf{x} \in\textrm{proj}_k \mathcal P} \mathbf{c}^T \mathbf{x} &=& \textrm{proj}_k \left( \arg \min_{\mathbf{y} \in\mathcal P} [\mathbf{c}^T,\mathbf{0}_{d-k}^T] \mathbf{x} \right).
\end{eqnarray}
Furthermore, any extreme point of $\textsf{proj}_k(\mathcal P)$ can be described as a projection of some extreme point of $\mathcal P$, meaning that finding all extreme points of $\textsf{proj}_k(\mathcal P)$ is simply a matter of solving a series of linear programs over $\mathcal P$ and projecting the point attaining the optimum solution.
\subsection{Boundedness Transformation}\label{chm:boundedness}
Note that CHM is made to project polytopes, but when calculating the bounds on network coding rate regions and performing the other types of network information theory calculations described in \S \ref{sec:convProofExamp}, we will be interested in projecting pointed polyhedral cones in $\mathbb R^{d}_{\geq \mathbf 0}$, as all linear information inequalities are homogeneous inequalities while constraints arising from a network coding instance are homogeneous equations. Fortunately, an unbounded polyhedron $\mathcal C$ can be transformed to create a polytope $\mathcal{B}(\mathcal C)$ such that the projection of the unbounded polyhedron $\mathcal C$ can be obtained from projection of $\mathcal B(C)$. While there are several such transformations, including the one in \cite{lassezchm}, we describe a transformation that is more efficient in terms of dimension of $\mathcal B(\mathcal C)$ (=$d$). For an arbitrary polyhedron in $\mathcal P\subseteq \mathbb R^{d}_{\geq \mathbf 0}$, one can apply the transformations described in this section to $\textsf{homog}(\mathcal P)\subseteq \mathbb R^{d+1}_{\geq \mathbf 0}$. Let $\mathbf{Hx}\geq \mathbf 0$ be the inequality description associated with a polyhedral cone $\mathcal C$. This cone can be transformed into a polytope
$
\mathcal C'=\{\mathbf{x} \in \mathbb R^{d}_{\geq \mathbf 0}| \mathbf{Hx}\geq \mathbf 0\ \wedge\ (\mathbf{1}_k^T,\mathbf{0}_{d-k}^t )\mathbf{x}\leq 1\}
$.
\leml{
The inequality representation of $\textsf{proj}_k(\mathcal C')$ is equal to the inequality description of $\textsf{proj}_k(\mathcal C)$ plus the additional inequality $\mathbf{1}_k^T\mathbf x_{1:k}\leq 1$.
}{lem:bounded1}
\noindent \textbf{Proof}: The inequalities in the inequality representation of $\textsf{proj}_k(\mathcal C)$ must be the only homogeneous inequalities in the representation of $\mathcal C'$, as adding a non-homogeneous inequality $[\mathbf{1}_k^T,\mathbf 0_{d-k}^t ]\mathbf x\leq 1$ to $\mathcal C$ cannot affect their minimality. Next, we must show that $\mathbf{1}_k^T\mathbf x_{1:k}\leq 1$ is the only non-homogeneous inequality that is non-redundant for $\textsf{proj}_k(\mathcal C')$. Assuming the contrary, there must be an inequality $\mathbf a^T\mathbf x_{1:k}\leq b$ that is also non-redundant for $\textsf{proj}_k(\mathcal C')$. Every inequality bounding $\textsf{proj}_k(\mathcal C')$ can be obtained as a conic combination of inequalities in the inequality representation of $\mathcal C'$. Hence, $(b,\mathbf a^T)=\boldsymbol\lambda \mathbf H+\lambda_0(1,\mathbf 1_{k}^T)$, where $\boldsymbol\lambda$ is a non-negative real vector and $\lambda_0$ is a non-negative real number. Furthermore, since $(\mathbf{1}_k^T,\mathbf 0_{d-k}^t )\mathbf x\leq 1$ is the only non-homogeneous inequality in the inequality representation of $\mathcal C'$, $\lambda_0\neq 0$. Now, $\boldsymbol\lambda \mathbf H\geq 0$ is a homogeneous inequality satisfied by projection, and hence can be written as a conic combination of non-redundant homogeneous inequalities in the inequality representation of $\textsf{proj}_k(\mathcal C')$. This means, $\mathbf a^T\mathbf x\leq b$ can be written as a conic combination of inequalities bounding $\textsf{proj}_k(\mathcal C)$ and $\mathbf{1}_k^T\mathbf x_{1:k}\leq 1$, contradicting our assumption that it is non-redundant w.r.t. $\textsf{proj}_k(\mathcal C')$. $\blacksquare$
Using lemma \ref{lem:bounded1}, given the inequality description of $\textsf{proj}_k(\mathcal C')$, we can obtain the inequality representation of $\textsf{proj}_k(\mathcal C)$ by simply deleting the only non-homogeneous inequality present in the representation.
\exmp{[Boundedness Transformation for Ex. \ref{ex:run}] In context of the IDSC instance in Fig. \ref{fig:idsc_ex}, the bounding inequality is of the form
\begin{equation}
\omega_1+\omega_2+\omega_3+R_4+R_5+R_6\leq 1 .
\end{equation}
}
\subsection{Convex Hull Method: low-level details}\label{polyhedra:chminithull}
The first step in CHM, encapsulated in procedure \textsf{initialhull}, is the construction of an inner bound to $\textsf{proj}_k(\mathcal P)$, which is the convex hull of $d+1$ points on the boundary of $\textsf{proj}_k(\mathcal P)$, that is itself full-dimensional. In order to obtain the initial inner bound for the process, we first obtain two boundary points of $\textrm{proj}_k \mathcal P$ by solving two linear programs,
$\min_{\mathbf{y}\in\mathcal P} [-1,\mathbf{0}_{d-1}^T] \mathbf{y}$ and $\min_{\mathbf{y}\in\mathcal P} [1,\mathbf{0}_{d-1}^T] \mathbf{y}$. We then select the normal vector $\mathbf c$ of any hyperplane containing these points (proc. $\textsf{hyperplane}$), and new boundary points of the projection are obtained as $\textrm{proj}_{k} \arg\min_{\mathbf{y}\in\mathcal P} [\mathbf{c},\mathbf{0}_{d-k}] \mathbf{y}$ and $\textrm{proj}_k \arg\min_{\mathbf{y}\in\mathcal P} [-\mathbf{c},\mathbf{0}_{d-k}] \mathbf{y}$. This process of finding a hyperplane containing all previously known boundary points and finding new boundary points is repeated until we obtain $k+1$ boundary points of the projection that give a full dimensional initial inner bound of the projection (in other words, until we obtain $k+1$ convex-independent boundary points in $\mathbb R^k$). This process needs to be repeated at most $k+1$ times, owing to the full-dimensionality of $\mathcal P$. Given the set $V$ containing $k+1$ boundary points of the projection forming a full-dimensional inner bound, computing the associated inequality description $\textsf{conv}(V)$ corresponds to a $k+1\times k+1$ matrix inversion (proc. $\textsf{facets}$). This gives us a double description pair of the first full dimensional inner bound of the projection.
\exmp{[Initial Hull for Ex. \ref{ex:run}] We now describe the construction of initial hull in the context of IDSC instance in Fig. \ref{fig:idsc_ex}, which in this case is the polytope $\mathcal B_1\triangleq\textsf{conv}(\{\mathbf v_i| i\in \{1,\ldots,7\}\})$. The vertices of $\mathcal B_1$ are obtained by solving a series of linear programs with carefully chosen objective functions. The summary of the $\textsf{initialhull}(\cdot)$ procedure can be found in table \ref{tab:inithull_v}. The first and second Linear programs have objective functions $\omega_1$ and $-\omega_1$ respectively. At any later step $i>2$, the normal vector (denoted as $\mathbf h$ in $\textsf{initialhull}(\cdot)$ line \ref{proc:inithull_hyperplane}) of any of the several hyperplanes containing vertices $\mathbf v_1,\hdots,\mathbf v_{i-1}$ is computed using procedure $\textsf{hyperplane}(\cdot)$. The objective function is then $\mathbf h$ itself or its negation $-\mathbf h$. The assumption, that the projection polytope is full-dimensional, implies that at least one of these two linear programs must yield a vertex that is convex independent w.r.t. the vertices found so far. Once, we have found $k+1=7$ convex independent vertices, the inequality description of the convex hull of $\mathbf v_1,\hdots,\mathbf v_7$ is obtained using the $\textsf{facets}(\cdot)$ procedure (line \ref{proc:inithull_facets} of $\textsf{initialhull}(\cdot)$). Internally, the procedure $\textsf{facets}(\cdot)$ constructs the inequality description by inverting the matrix $\mathbf A_1$ associated with the inequality description of the cone $\mathcal P_1\triangleq\mathcal C(\mathcal B_1)^\circ$. The inequality description of $\mathcal P_1$ is $\mathcal P_1(\mathbf A_1,\mathbf 0)$ where,
\begin{equation}
\mathbf A_1 \triangleq \begin{bmatrix}
-1 & -\frac{1}{2} & 0 & 0 & 0 & -\frac{1}{2} & 0\\
-1 & 0 & 0 & 0 & 0 & 0 & 0\\
-1 & -\frac{1}{2} & 0 & 0 & 0 & 0 & -\frac{1}{2}\\
-1 & 0 & 0 & 0 & 0 & -1 & 0\\
-1 & 0 & -\frac{1}{4} & 0 & -\frac{1}{4} & -\frac{1}{4} & -\frac{1}{4}\\
-1& 0 & 0 & 0 & -1 & 0 & 0 \\
-1 & 0 & 0 & -\frac{2}{5} & \frac{1}{5} & \frac{1}{5} & \frac{1}{5}\\
\end{bmatrix}
\end{equation}
The extreme rays of $\mathcal P_1$, obtained by computing $\mathbf A_1^{-1}$, are given by the columns of matrix $\mathbf Y$ where,
\begin{equation}
\mathbf Y\triangleq
\begin{bmatrix}
0 & 1 & 0 & 0 & 0 & 0 & 0\\
1 & -1 & 0 & -1 & 0 & 0 & 0\\
1 & -1 & -1 & -2 & 1 & -1 & 0\\
\frac{1}{2} & -1 & -\frac{1}{2} & -1 & 0 & -\frac{1}{2} & 2\\
0 & -1 & 0 & 0 & 0 & 1 & 0\\
0 & -1 & 0 & 1 & 0 & 0 & 0\\
-1 & -1 & 1 & 1 & 0 & 0 & 0\\
\end{bmatrix}
\end{equation}
The inequality description of $\mathcal B_1$ can be then obtained by reading off the columns of matrix $\mathbf Y$ as,
\begin{equation}\label{eq:inithull_h}
\begin{aligned}
2 R_6 &\leq 2\omega_1+2\omega_2+\omega_3 & (\mathbf h_1) \\
\omega_1 +\omega_2 + \omega_3 + R_4 + R_5 + R_6 &\leq 1& (\mathbf h_2)\\
2\omega_2+\omega_3 &\leq 2 R_6 & (\mathbf h_3)\\
\omega_1+2\omega_2+\omega_3 &\leq R_5+R_6 & (\mathbf h_4)\\
-\omega_2 &\leq 0 & (\mathbf h_5)\\
2\omega_2 +\omega_3 &\leq 2R_4 & (\mathbf h_6)\\
-\omega_3 &\leq 0 & (\mathbf h_7)
\end{aligned}
\end{equation}
}
\begin{table*}
\vspace{3mm}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
LP & objective function & new vertex & homogenized vertex\\
\hline\hline
1 & $\omega_1$ & $\mathbf{v}_1=(\frac{1}{2},0,0,0,\frac{1}{2},0)$ & $\mathbf{r}_1=(\frac{1}{2},0,0,0,\frac{1}{2},0)$ \\
\hline
2 & $-\omega_1$ & $\mathbf{v}_2=(0,0,0,0,0,0)$ & $\mathbf{r}_2=(0,0,0,0,0,0)$ \\
\hline
3 & $\omega_1-R_2$ & $\mathbf{v}_3=(\frac{1}{2},0,0,0,0,\frac{1}{2})$ & $\mathbf{r}_3=(\frac{1}{2},0,0,0,0,\frac{1}{2})$ \\
\hline
4 & $-\omega_1+R_2$ & $\mathbf{v}_4=(0,0,0,0,1,0)$ & $\mathbf{r}_4=(0,0,0,0,1,0)$ \\
\hline
5 & $\omega_2$ & $\mathbf{v}_5=(0,\frac{1}{4},0,\frac{1}{4},\frac{1}{4},\frac{1}{4})$ & $\mathbf{r}_5=(0,\frac{1}{4},0,\frac{1}{4},\frac{1}{4},\frac{1}{4})$ \\
\hline
6 & $-\omega_2$ & $-$ & $-$\\
\hline
7 & $\omega_2-R_1$ & $-$ & $-$\\
\hline
8 & $-\omega_2+R_1$ & $\mathbf{v}_6=(0,0,0,1,0,0)$ & $\mathbf{r}_6=(0,0,0,1,0,0)$ \\
\hline
9 & $\omega_3$ & $\mathbf{v}_7=(0,0,\frac{2}{5},\frac{1}{5},\frac{1}{5},\frac{1}{5})$ & $\mathbf{r}_7=(0,0,\frac{2}{5},\frac{1}{5},\frac{1}{5},\frac{1}{5})$ \\
\hline
\end{tabular}
\vspace{3mm}
\caption{Summary of the linear programs solved during the $\textsf{initialhull}(\cdot)$ procedure in CHM.}\label{tab:inithull_v}
\end{center}
\end{table*}
Once the initial hull is computed, at each stage of the CHM algorithm, a DD pair is maintained for the current inner bound to $\textrm{proj}_k \mathcal P$. Each inequality in the inequality description of the inner bound carries with it a label, indicating whether or not it is terminal or non-terminal. The initial inner bound has all of its inequalities labelled as non-terminal. A non-terminal inequality $\mathbf{c}^T \mathbf{x} \geq b$ is then selected (proc. \textsf{isterminal}), and the linear program $\min_{\mathbf{x} \in \mathcal P} [\mathbf{c}^T\mathbf{0}_{d-k}^T] \mathbf{x} $ is solved over the high dimensional polyhedron. If the solution obtained is $b$, the inequality is marked as terminal. Otherwise, the extreme point of $\mathcal P$, obtained in the process, i.e. $\mathbf{v}=\arg\min_{\mathbf{x} \in \mathcal P} [\mathbf{c}^T\mathbf{0}_{d-k}^T] \mathbf{x} $ attaining the solution is projected to get a new boundary point $\textrm{proj}_k \mathbf{v}$ of $\textrm{proj}_k \mathcal P$. The procedure \textsf{extremepoint}($\mathcal{P},\mathbf{c}$) thus returns $\textrm{proj}_k \left( \arg \min_{\mathbf{y} \in\mathcal P} [\mathbf{c}^T,\mathbf{0}_{d-k}^T] \mathbf{x} \right)$. The DD pair of the inner bound is then updated (proc. \textsf{updatehull}) by adding the new boundary point, viewing it as a new inequality in the polar cone, via a single DD algorithm update step \cite{fukuda1996double} (proc. \textsf{DDiteration}), and any new inequalities thus introduced are marked as non-terminal. In this double description iteration, the new extreme point (ray of homogenization) is viewed as a new inequality for the polar of this homogenization, and the resulting new extreme rays are, by passing again back through the polar, are thus directly interpreted (by procedure \textsf{ineq}) as new inequalities of the inner bound created by adding this ray. Finally, a new non-terminal facet is selected and the process is repeated until all of the facets are labelled as terminal, at which point the inner bound has been proven equal to $\textrm{proj}_k \mathcal P$.
To illustrate these remaining iterations of CHM after the initial hull is computed, the next example provides a summary of the remaining iterations for the case of the running example.
\exmp{[CHM for Ex. \ref{ex:run}] Table \ref{tab:chmsummary} gives a summary of all linear programs solved, new vertices revealed, and non-facets revealed after initial hull is computed in the context of IDSC instance in Fig. \ref{fig:idsc_ex}.}
\begin{table*}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
LP & objective function (facet) & new vertex & non-facets & new facets\\
\hline\hline
10 & $-2\omega_1-2\omega_2-\omega_3+2R_6$ ($\mathbf h_1$) & $\mathbf{v}_8=(0,0,0,0,0,1)$ & $\mathbf h_1$ & $\begin{aligned} -\omega_1 & \leq 0 & (\mathbf h_9)\\ 2\omega_2+\omega_3 &\leq 2R_5 & (\mathbf h_{10})\end{aligned}$ \\
\hline
11 & $2\omega_2+\omega_3-2R_6$ ($\mathbf h_{3}$) & $\mathbf v_9=(0,0,\frac{1}{3},\frac{1}{3},\frac{1}{3},0)$ & $\mathbf h_{3}$ & $\begin{aligned} \omega_2 &\leq R_6 & (\mathbf h_{11})\\ 2\omega_2+\omega_3 & \leq R_4+R_6 & (\mathbf h_{12}) \end{aligned}$\\
\hline
12 & $2\omega_2+\omega_3-2R_4$ ($\mathbf h_{6}$) & $\mathbf v_{10}=(0,0,\frac{1}{3},0,\frac{1}{3},\frac{1}{3})$ & $\mathbf h_{6}$ & $\begin{aligned} \omega_2 &\leq R_4 & (\mathbf h_{13})\\ 2\omega_2+\omega_3 &\leq R_4 + R_5 & (\mathbf h_{14})\\ \omega_1+4\omega_2+2\omega_2 &\leq 2R_4+R_5+R_6 & (\mathbf h_{15}) \end{aligned}$\\
\hline
13 & $-\omega_3$ ($\mathbf h_7$) & - & - & - \\
\hline
14 & $2\omega_2+\omega_3-R_5$ $(\mathbf h_{10})$ & $\mathbf v_{11}=(0,0,\frac{1}{3},\frac{1}{3},0,\frac{1}{3})$ & $\mathbf h_{10}$ & $\begin{aligned} \omega_2 &\leq R_5 & (\mathbf h_{16})\end{aligned}$\\\hline
15 & $\omega_2-R_5$ ($\mathbf h_{16}$) & - & - & - \\ \hline
16 & $2\omega_2+\omega_3-R_4-R_6$ ($\mathbf h_{12}$) & - & - & - \\ \hline
17 & $\omega_3-R_4$ ($\mathbf h_{13}$) & - & - & - \\ \hline
18 & $\omega_1+4\omega_2+2\omega_3-2R_4-R_5-R_6$ ($\mathbf h_7$) & - & - & - \\ \hline
19 & $\omega_1+2\omega_2+\omega_3-R_5-R_6$ ($\mathbf h_{4}$) & $\mathbf v_{12}=(\frac{1}{2},0,0,\frac{1}{2},0,0)$ & $\mathbf h_{4}$ & $\begin{aligned} 2\omega_2+\omega_3 &\leq R_5+R_6 & (\mathbf h_{17})\\
2\omega_1+6\omega_2+3\omega_3 &\leq 2R4+2R_5+2R_6 & (\mathbf h _{18}) \\
\omega_1+4\omega_2+2\omega_3 &\leq R4+2R_5+R_6 & (\mathbf h _{19}) \\
\omega_1+4\omega_2+2\omega_3 &\leq R4+R_5+2R_6 & (\mathbf h _{20})
\end{aligned}$\\\hline
20 & $-\omega_1$ ($\mathbf h_{9}$) & - & - & - \\ \hline
21 & $2\omega_2+\omega_3-R_4-R_5$ ($\mathbf h_{14}$) & - & - & - \\ \hline
22 & $\omega_1+\omega_2+\omega_3-R_4-R_5-R_6$ ($\mathbf h_2$) & - & - & - \\ \hline
23 & $-\omega_2$ ($\mathbf h_{16}$) & - & - & - \\ \hline
24 & $\omega_2-R_5$ ($\mathbf h_{16}$) & - & - & - \\ \hline
25 & $\omega_1+4\omega_2+2\omega_3 - R4-R_5-2R_6$ ($\mathbf h_{20}$) & - & - & - \\ \hline
26 & $2\omega_1+6\omega_2+3\omega_3-2R_4-2R_5-2R_6$ ($\mathbf h_{18}$) & - & - & - \\ \hline
27 & $2\omega_2+\omega_3-R_5-R_6$ ($\mathbf h_{17}$) & - & - & - \\ \hline
28 & $\omega_1+4\omega_2+2\omega_3 -\leq R4-2R_5-R_6$ ($\mathbf h_{19}$) & - & - & - \\ \hline
\end{tabular}
\vspace{3mm}
\caption{Summary of the linear programs solved after the $\textsf{initialhull}(\cdot)$ procedure in CHM.}\label{tab:chmsummary}
\end{center}
\end{table*}
These remaining iterations also make use of the $\textsf{updatehull}(\cdot)$ procedure, involving a double description set, and the next example demonstrate how this procedure works for the first update after the initial hull for Example \ref{ex:run}.
\exmp{[$\textsf{updatehull}$ immediately after \textsf{initialhull}, Ex. \ref{ex:run}]
The input to the $\textsf{updatehull}(\cdot)$ procedure is the inequalities (see \eqref{eq:inithull_h}) and vertices (see table \ref{tab:inithull_v}) of $\mathcal B_1$, along with the newly revealed vertex $\mathbf v_8$. Since $\mathbf v_8$ was obtained as the optimum vertex of the linear program with cost function $-2\omega_1-2\omega_2-\omega_3+2R_6$, it violates inequality $\mathbf h_1$. In this case, it is the only inequality in the inequality description of $\mathcal B_1$ violated by $\mathbf v_8$. Thus, the inequality description of $\mathcal B_2\triangleq \textsf{conv}(\{\mathbf v_i| i\in \{1,\ldots,8\}\})$ must contain all inequalities in the inequality description of $\mathcal B_1$ except $\mathbf h_1$. As for the new inequalities that hold for $\mathcal B_2$, the original CHM algorithm proposed by Lassez et. al. would resort to a combinatorial search, where the search space would be all $6$-subsets of $\{\mathbf v_i| i\in \{1,\ldots,8\}\}$, any of which define a unique supporting halfspace for $\mathcal B_2$, provided it contains $\mathbf v_8$ . The authors' implementation of CHM, instead reduces the problem of updating $\mathcal B_1$ with $\mathbf v_8$ to that of intersecting $\mathcal C(\mathcal B_1)^\circ$ with a halfspace obtained by putting $\mathbf v_8$ through homogenization and polar transformation. To avoid confusion, we present a description of this procedure, in the context of $\mathcal B_1$ itself, as opposed to $\mathcal C(\mathcal B_1)^\circ$ (which can be thought of as primal vs dual interpretations of the same procedure, with duality in question being the polar duality).
In general, let $J^+$ and $J^-$ be the sets of inequalities that strictly satisfy and violate the newly revealed vertex respectively. The main lemma of the double description method (see \cite{fukuda1996double}, lemma 3) dictates that it suffices to consider conic combinations of pairs of inequalities in $J^+\times J^-$ to construct the new inequalities of the inequality description of the augmented inner bound. Fig. \ref{fig:ddex_b} describes the incidence relationships between newly revealed vertex $\mathbf v_8$ and the facets of $\mathcal B_1$. In this case, $J^+=\{\mathbf h_1\}$ and $J^-=\{\mathbf h_3,\mathbf h_4\}$. Furthermore, the strengthened main lemma of the double description method (see \cite{fukuda1996double}, lemma 8), shows that it suffices to restrict attention to the pairs $(\mathbf h_i,\mathbf h_j)\in J^+\times J^-$ that are adjacent i.e. the intersection of the associated facets have dimension $k-2$ where $k$ is the dimension of the polytope in question. Fukuda et al. \cite{fukuda1996double} also provide simple tests of adjacency, and a technique for computing conic combination coefficients to be used to combine the inequalities, which we omit for the sake of conciseness. In our example, every pair of inequalities in $J^+\times J^-$ is adjacent, and they can be combined as follows to obtain new inequalities $\mathbf h_9$ and $\mathbf h_{10}$.
\begin{equation}
\begin{aligned}
-2\omega_1 -& 2\omega_2 &-\omega_3 &+R_6&\leq 0 & (\mathbf h_1)\\
+& 2\omega_2 &+\omega_3 &-R_6 &\leq 0 & (\mathbf h_3)\\
\cline{1-6}
-2\omega_1 & & & & \leq 0 & (\mathbf h_9)
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
-2\omega_1 &- 2\omega_2 &-\omega_3 & &+R_6&\leq 0 & (\mathbf h_1)\\
+\omega_1 &+ 2\omega_2 &+\omega_3 &-R_5 &-R_6 &\leq 0 & (\mathbf h_4)\\
\cline{1-7}
&\omega_2 &+\frac{1}{2}\omega_3 &-R_5 & & \leq 0 & (\mathbf h_{10})
\end{aligned}
\end{equation}
In addition to adding these new inequalities, the routine \textsf{updatehull}, updates the vertex facet incidence relationships for the new inner bound $\mathcal{B}_2$, to those shown in Fig. \ref{fig:ddex_c}.
}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=1.2]{ddstep_example1_a.pdf
\end{center}
\caption{Complement of the vertex-facet incidence graph of inner bound $\mathcal B_1$, produced at the end of procedure $\textsf{initialhull}(\cdot)$. The nodes on top represent the facets (see \eqref{eq:inithull_h}), while the nodes at the bottom represent the vertices of $\mathcal B_1$ (see table \ref{tab:inithull_v}). The grayed-out vertices lie outside $\mathcal B_1$, and are yet undiscovered. The numbers on the edges give the strictness with which an inequality is satisfied. }\label{fig:ddex_a}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=1.2]{ddstep_example1_b.pdf
\end{center}
\caption{The graph in Fig. \ref{fig:ddex_a}, superimposed with incidence relationship between facets $\mathbf h_1,\hdots,\mathbf h_7$ and the newly revealed vertex $\mathbf v_8$. The colors red, yellow and green represent violation, equality, or strict satisfaction of a particular inequality by $\mathbf v_8$. The inequalities violated by $\mathbf v_8$ ($\mathbf h_1$) are combined with the inequalities strictly satisfied by $\mathbf v_8$ ($\mathbf h_3,\mathbf h_4$) to produce the new inequalities that hold for the augmented inner bound ($\mathbf h_9,\mathbf h_{10}$). }\label{fig:ddex_b}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=1.2]{ddstep_example1_c.pdf
\end{center}
\caption{Complement of the vertex-facet incidence graph of inner bound $\mathcal B_2$, produced at the end of procedure $\textsf{updatehull}(\cdot)$ after augmenting $\mathcal B_1$ with newly revealed vertex $\mathbf v_8$.}\label{fig:ddex_c}
\end{figure}
\subsection{CHM Complexity: Worst Case and a Series of Examples}
In this subsection, we consider a simple family of HMSNC instances and gauge the practical performance of CHM for computation of the associated EPOBs, along with a discussion of worst case complexity of CHM and its various components. When computing an EPOB for a HMSNC instance with $N$ random variables, with $\Gamma_{\textsf{out}}=\Gamma_N$, one starts with the polytope defined by the inequality representation containing elemental Shannon-type inequalities \cite{yeungframework}, network constraints and the bounding inequality $\sum_{i\in[k]}\boldsymbol \omega_i+\sum_{j\in[N]\setminus [k]} R_j\leq 1$. This polytope is then projected to $\boldsymbol \omega,\mathbf r$ coordinates using CHM and from the resultant inequality representation we simply remove the bounding inequality, giving us the inequality representation of the EPOB.
\begin{figure}[h]
\begin{center}
\includegraphics[width=3in]{u2k.pdf
\end{center}
\caption{The runtime, number linear programs solved, and no. of facets of rate region outer bound, for computing EPOB for a $U^2_k$ network vs $k$, the size of network, with $\Gamma_{\textsf{out}}=\Gamma_k$ i.e. the Shannon outer bound.}\label{fig:u2k}
\end{figure}
\exmpl{$U^2_k$ networks \cite{zongpenguniform} are a family of matroidal networks, that can be constructed from uniform matroids of rank $2$ on ground set of size $k$, using the construction of Dougherty, Freiling and Zeger\cite{DFZMatroidNetworks}. Fig. \ref{fig:u2k} shows the plot of no. of linear programs solved by CHM and time taken by CHM as a function of $k$, which is the size of a $U^2_k$ network. The number of linear programs solved by CHM, which in this case, is the sum of number of facets and number of vertices of the EPOB, grows linearly w.r.t. $k$. On the other hand, the runtime grows exponentially w.r.t. $k$. The growth in runtime reflects the exponential growth in the dimension of the polyhedra being projected, as the outer bound $\Gamma_k$, which is used to compute these EPOBs, lives in a $2^k-1$ dimensional space.
}{u2kexmp}
Now we point out worst case complexities of some of the building blocks of CHM, which are daunting, in keeping with the worst case complexities of most polyhedral computation algorithms. Note that the rate region outer bounds computed in example \ref{u2kexmp} are very \textit{well-behaved} when compared to McMullen's upper bound of $\mathcal O(m^{\left\lfloor\frac{k}{2}\right\rfloor})$\cite{ziegler1995lectures} on the no. of vertices of a $k$ dimensional polytope specified by $m$ inequalities, as are many polyhedra. It is possible, and in fact highly likely based on experience with examples, that the EPOBs for HMSNC instances and other network information theory problems satisfy a tighter worst case bound on the number of extreme rays when $\Gamma_{\textsf{out}}=\Gamma_N$. This is related to the enumeration of extremal polymatroids, i.e. extreme rays of the cone $\Gamma_N$, which remains an open problem, although it is lower bounded by the number of connected matroids \cite{Nguyen1978,Li_Allerton_13,Mayhew2008415}, which has been more carefully studied.
Owing the capability to generate severely degenerate polyhedra having many bases per extreme point, the simplex method, which is used in CHM to solve linear programs exactly by employing rational arithmetic, in worst case complexity analysis can require exponentially many pivots in $d$ for all known pivoting methods. Hence, in the worst case, assuming that the only points of projection obtained during the course are the extreme points of the projection, CHM could solve $\mathcal O(m^{\left\lfloor\frac{k}{2}\right\rfloor})+m$ linear programs, one per vertex and facet of projection, with each having exponential runtime, given that the projection has $m$ facets. Furthermore, the double description method, used in CHM to incrementally build the projection polytopes, does not have an output sensitive time complexity \cite{Bremner1996convhull}, as it is known to be susceptible to the insertion order of inequalities. In the case of CHM, this means that the order in which we find the extreme points of the projection, greatly affects the number of rays in intermediate double description pairs. Every iteration of DD method, the number of extreme rays can get squared, in the worst case.
On the other hand, compared with Fourier-Motzkin, the most commonly taught method for small polyhedral projections, CHM can have vastly improved computational complexity. In Fourier-Motzkin, one calculates not only the projected final polyhedron, but also every intermediate projected polyhedron of intermediate dimension between that of the parent polyhedron. These polyhedra of intermediate dimension typically have far larger numbers of inequalities and extreme rays than those of the parent or projected polyhedron, hence for problems such as the network coding rate regions, non-Shannon inequalities, linear rank inequalities, etc. in which the number of dimensions being removed is exponential in the number of dimensions being kept, algorithms such as CHM which work directly and exclusively in the projected polyhedron by solving linear programs have substantially lower complexity than Fourier-Motzkin.
While the worst case analysis of the building blocks of CHM paints an ominous picture, both simplex method and double description method have been practically successful in solving moderate to large problem instances, which is essentially what the authors expect of CHM, and what Ex. \ref{u2kexmp} and nearly one million other worked examples \cite{Li_Operators} demonstrates.
As solving linear programs and computing double descriptions updates are inherent to algorithms aiming to reduce complexity relative to Fourier Motzkin by working directly in the projected space, it is typically in the polyhedral computation literature to assess their complexity in terms of numbers of linear programs that must be solved, the dimensions and numbers of constraining inequalities in these linear programs, and the number of double descriptions steps that must be performed \cite{FukudaPolyCompNotes}. As each step of CHM solves a linear program which reveals a new extreme point of the projected region, it is clear that the number of linear programs which must be solved is equal to the number of extreme points of the projected region, with each such linear program having a dimension and number of inequalities consistent with the parent polyhedron. Similarly, CHM must compute an equal number of double descriptions steps. Bearing in mind that solving linear programs with simplex and computing double descriptions steps can in worst case have high complexity, but are routinely performed in many contexts, and in fact are necessary for any projection algorithm working exclusively in the projection dimensions, in the next section we shift our attention further reducing the complexity of CHM by decreasing the number of linear programs, their dimensions, and double descriptions steps that must be solved.
\section{A Polyhedral Projection Algorithm that Exploits Symmetry}\label{sec:symchm}
Having assembled the necessary ingredients in previous sections, in this section we are prepared to present the main contribution of the manuscript: a polyhedral projection algorithm that can exploit known symmetry groups of the projected polyhedron and the parent polyhedron, and the demonstration of its complexity reduction benefits when computing network information theory rate regions.
The projection algorithm we have created, named symCHM, whose pseudocode is presented in \ref{symchm}, exploits symmetry to reduce the complexity of CHM in three different ways. Additionally, substantial storage/memory reduction is achieved. These reductions build from symmetry exploiting techniques in the polyhedral computation literature.
Bremner et. al. \cite{Bremner09polyhedralrepresentation} and Rehn \cite{RehnSymmetries} consider the problem of exploiting the knowledge of restricted affine symmetries in polyhedral representation conversion, which is the problem of computing the inequality representation of a polyhedron, given the extreme ray representation (or vice versa). On the other hand, B\"odi et. al \cite{herr09arxiv,bodyhighsymmlpalgo} study how subgroups of RSG consisting of permutation matrices can be utilized to reduce the dimension of linear and integer programs. The results of B\"{o}di, and other papers explaining how to exploit constraint symmetries to reduce the dimensions of linear programs however, are easily extended the more general notions of symmetries, such as affine and restricted affine symmetries, considered in this manuscript.
In a nutshell, the symmetry exploiting complexity reductions applied to CHM to create symCHM, and the associated parts of the pseudocode in Alg. \ref{symchm} that we be described in the remaining parts of this section in detail are as follows. As does CHM, symCHM builds a series of increasingly larger inner bounds to the projected polytope by solving carefully selected linear programs over the parent polytope. However, each step of the inner bounds created by symCHM are forced to include the known symmetries of the projected polyhedron. This enables memory savings (\S \ref{sec:memred}), by only storing one representative for each equivalence class of symmetric inequalities and extreme rays created by this known symmetry group. It also enables substantial computational savings (\S \ref{sec:symUp}), as it enables the number of linear programs that must be solved to be reduced from the number of extreme points of the projected polyhedron to the number of equivalence classes (under the known symmetry group) of extreme points. A second type of complexity reduction (\S \ref{sec:symDD}) is achieved by adapting techniques from Bremner et. al. \cite{Bremner09polyhedralrepresentation} to reduce the complexity of the double descriptions step that must be performed when each new equivalence class (under symmetry) of extreme points are revealed. Finally, a third type of complexity reduction (\S \ref{sec:symLp}) is achieved, under a method similar to that exploited in B\"{o}di et. al. \cite{herr09arxiv,bodyhighsymmlpalgo}, by reducing the dimension of the linear program necessary to check whether a given equivalence class of facets (under the action of the symmetry group), is terminal. Exploring further implications of this final idea, in \S \ref{sec:symLpOth}, we point out that in certain network information theory problems focussed not on entire rate regions, but scalar quantities, this final complexity reduction via dimension reduction is especially helpful.
\begin{algorithm}[h]
\caption{Symmetry exploiting CHM}\label{symchm}
\DontPrintSemicolon
\KwIn{Polyhedron $\mathcal{P}\subseteq \mathbb R^n$, projection dimension $d<n$ and symmetry group $G\leq S_k$ of $\textsf{proj}_k(\mathcal P)$}
\KwOut{Transversal pair $(\mathcal T_{ V},\mathcal T_{ H})$ of $\textsf{proj}_d(\mathcal{P})$}
$(\mathcal T_V,\mathcal T_H)\leftarrow$ \textsf{syminitialhull}($\mathcal{P},k$)\;
\While{$\exists \{\mathbf h\mathbf x\geq \mathbf b\}\in \mathcal T_H$ s.t. \textsf{isterminal}($\mathcal{P},\mathbf{h},\mathbf{b})= 0$ \label{w2_1}} {
$\mathbf{v}\leftarrow$ \textsf{extremepoint}($\mathcal{P},\mathbf{h}$)\;\label{line3}
$\mathcal T_H\leftarrow$\textsf{symupdatehull}($\mathbf{v},V,H$)\;
$\mathcal T_V\leftarrow \mathcal T_V\cup\{\mathbf{v}\}$\label{w2_2}
}
\Return{$(\mathcal T_V,\mathcal T_H)$}\;
\end{algorithm}
\begin{procedure}[h]
\DontPrintSemicolon
\KwIn{Transversal pair $(\mathcal T_V,\mathcal T_H)$, vertex $\mathbf v$ and group $G$}
\KwOut{Transversal $\mathcal T_H'$ associated with $\mathcal T_V'=\mathcal T_V \cup\{\mathbf{v}\}$}
$(\mathcal T_{V_{\mathcal C}},\mathcal T_{H_{\mathcal C}})\leftarrow \textsf{homog}(\mathcal T_H)^\circ$\;
$\{\mathbf{r}_1,...,\mathbf{r}_t\}\leftarrow$ symDD($\mathcal T_{V_{\mathcal C}},\mathcal T_{H_{\mathcal C}},\{(1 \hspace{2mm} \mathbf v^T)\mathbf{y}\leq 0\},G$)\;
$\mathcal T_H'\leftarrow\{\textsf{ineq}(\mathbf r_1),\hdots,\textsf{ineq}(\mathbf r_t)\}$\;
\Return{$\mathcal T_H'$}\;
\label{proc:symupdatehull}
\caption{symupdatehull($\mathbf{v},V,H$)}
\end{procedure}
\begin{procedure}[h]
\label{symdd}
\DontPrintSemicolon
\KwIn{Transversal pair $\mathcal T_{V_{\mathcal C}}, \mathcal T_{H_{\mathcal C}}$ of the set of extreme rays and facets of a cone $\mathcal C\subseteq \mathbb R^{d+1}$, inequality $\mathbf a^T\mathbf x\leq 0$, group $G$}
\KwOut{Transversal of set of extreme rays of $\mathcal C\bigcap\cap_{g\in G}\{\mathbf a^T\mathbf x\leq 0\}^g$}
$V_{\mathcal C}\leftarrow \mathcal T_{V_{\mathcal C}}^G$, $H_{\mathcal C}\leftarrow \mathcal T_{H_{\mathcal C}}^G$\;\label{ln:ddofc1}
$(P,N,Z)\leftarrow \textsf{DD}({V_{\mathcal C}}, {H_{\mathcal C}},\mathbf a)$\;\label{ln:dd}
$V_{\mathcal C_{\mathbf v^{\leq}}}\leftarrow P\cup Z$\;
${H_{\mathcal C_{\mathbf v^{\leq}}}}\leftarrow H_{\mathcal C}\cup \{\mathbf a^T\mathbf x\geq 0\}$\;\label{ln:cineq}
$(V_{\mathcal C_{\mathbf v^=}},H_{\mathcal C_{\mathbf v^=}})\leftarrow \textsf{tightenfacet}(P\cup Z,{H_{\mathcal C_{\mathbf v^{\leq}}}})$\;\label{ln:tighten}
$\mathcal A\leftarrow\textsf{repDD}((V_{\mathcal C_{\mathbf v^=}},\mathbf a,G)$\;\label{ln:repdd}
\For{$\mathbf a'\in \mathcal A$}
{$(P,N,Z)\leftarrow \textsf{DD}(V_{\mathcal C_{\mathbf v^=}},H_{\mathcal C_{\mathbf v^=}},\mathbf a')$\;\label{ln:smalldd}
$V_{\mathcal C_{\mathbf v^=}}\leftarrow P\cup Z$\;
$H_{\mathcal C_{\mathbf v^=}}\leftarrow H_{\mathcal C_{\mathbf v^=}}\cup\{\mathbf a'^T\mathbf x\leq 0\}$
}
$\mathcal T_{V_{\mathcal C'}}\leftarrow \textsf{nonisomorphic}((P\setminus{N^G}) \cup \textsf{lift}(V_{\mathcal C_{\mathbf v^=}},\mathbf a),G)$\;\label{ln:newtrans}
\Return{$\mathcal T_{V_{\mathcal C'}}$}
\caption{symDD($\mathcal T_{V_{\mathcal C}},\mathcal T_{H_{\mathcal C} },\mathbf a^T\mathbf x\leq \mathbf 0,G$)}
\end{procedure}
\subsection{Space Reduction via Symmetry}\label{sec:memred}
Let $ V$ and $ H$ be the set of vertices and facets respectively of $\textsf{proj}_k(\mathcal P)$. For an affine symmetry $g\in G$ and a vertex $\mathbf v$ of $\textsf{proj}_k(\mathcal P)$ denote by $\mathbf v^g$ to be the vertex to which $\mathbf v$ maps to under action of $g$ and let $\mathbf v^G$, the \emph{orbit} of $\mathbf v$ under $G$, be the set of all vertices to which $\mathbf v$ can map to under action of $G$. $\mathbf v^G$ contains all vertices that are $G$-\textit{equivalent} to $\mathbf v$. The set of all orbits in $V$ under action of $G$ forms a partition of $ V$ and is denoted as $\mathcal O_V$. Since each facet is simply the convex hull of a set of vertices, the action of the ASG can be extended to $H$ i.e. we define the orbit of a facet $h\in H$ denoted as $h^G$ and $\mathcal O_H$ to be set of all orbits of facets. The transversal $\mathcal T$ of a set of orbits $\mathcal O$ is a set containing one representative per orbit in $\mathcal O$, and transversals of $\mathcal O_V$ of $\mathcal O_H$ are denoted as $\mathcal T_V$ and $\mathcal T_H$ respectively. Based on the aforementioned McMullen's Upper Bound, it is possible that the size of the double description of $\textsf{proj}_k(\mathcal P)$ is prohibitively large. In this case, we can trade space requirement for orbit computation, which is a basic procedure in computational group theory \cite{seressbook}, by storing only the transversals of the inequality and vertex orbit sets.
\subsection{Exploiting Symmetry to Solve Fewer Linear Programs}\label{sec:symUp}
The transversal $\mathcal T_H$ of the facets of the inner bound carries with it an indicator variable indicating if it is terminal or non-terminal. At an intermediate step in the algorithm, a non-terminal facet $\mathbf{c}^T \mathbf{x} \geq b$ is selected from the current inner bound's facet transversal $\mathcal T_H$. Just as in CHM, the linear program $\min_{\mathbf{y} \in \mathcal P} [\mathbf{c}^T, \mathbf{0}_{d-k}^T] \mathbf{y}$ is solved, and if the result is $b$, the facet is marked as terminal. If the result is not $b$, the projection of the extreme point attaining the minimum, $\textrm{proj}_k \arg \min_{\mathbf{y} \in \mathcal P} [\mathbf{c}^T, \mathbf{0}_{d-k}^T] \mathbf{y}$, is added to the transversal $\mathcal T_V$. This act of adding this single extreme point $\mathbf{v}$ to the inner bound's vertex transversal has the same effect as having added the entire orbit $\mathbf{v}^G$ to the full list of extreme points in CHM. In ordinary CHM, to find these extreme points, $|\mathbf{v}^G|$ linear programs would have had to be solved, whereas in symCHM, only one LP is required to obtain all of them, followed by an orbit computation.
Similarly, if a facet of an inner bound is found to be terminal, all the facets is its orbit under $G$ can be labeled as terminal, amounting to further reduction in the number of LPs solved.
\exmp{[Fewer LPs for Ex. \ref{ex:run} ]
In case of the example in Fig. \ref{fig:idsc_ex}, as stated in Ex. \ref{ex:netSymGroup}, the network symmetry group $G$ is of order $6$, generated by permutations $(4,5)$ and $(5,6)$. To begin with, the knowledge of symmetry can be used to make the initial inner bound symmetric. Thus all permutations of vertices of initial inner bound $\mathcal B_1$ under the network symmetry group an be added to $\mathcal B_1$ and $\mathcal B_1$ is augmented each time using an iteration of the double description method much on the lines of the inner bound updates in CHM. Tables \ref{tab:syminithull_v} and \ref{tab:syminithull_h} show the vertex and facet orbits of the symmetric initial inner bound $\mathcal B_1^G$ so constructed. The key complexity reduction here is that additional vertices are obtained via simple orbit computation as opposed to substantially more expensive linear programs. Fig. \ref{fig:syminithull} shows vertex facet incidences between vertex orbits and facet orbits of $\mathcal B_1^G$.
}
\begin{table}[h]
\vspace{3mm}
\begin{center}
\begin{tabular}{|c|l|}
\hline
Orbit Label & Member vertices\\
\hline\hline
$\mathcal O_V^1 $& $\begin{aligned}\mathbf v_2&=\left(0,0,0,0,0,0\right)\end{aligned}$ \\
\hline
$\mathcal O_V^2$ & $\begin{aligned}\mathbf v_4&=(0,0,0,0,1,0)\\\mathbf v_6&=(0,0,0,1,0,0)\\\mathbf v_{8} &= (0,0,0,0,0,1)\end{aligned} $ \\
\hline
$\mathcal O_V^3 $& $\begin{aligned}\mathbf v_7&=\left(0,0,\frac{2}{5},\frac{1}{5},\frac{1}{5},\frac{1}{5}\right)\end{aligned}$ \\
\hline
$\mathcal O_V^4 $& $\begin{aligned}\mathbf v_5&=\left(0,\frac{1}{4},0,\frac{1}{4},\frac{1}{4},\frac{1}{4}\right)\end{aligned}$ \\
\hline
$\mathcal O_V^5$ & $\begin{aligned}\mathbf v_1&=\left(\frac{1}{2},0,0,0,\frac{1}{2},0\right)\\\mathbf v_3&=\left(\frac{1}{2},0,0,0,0,\frac{1}{2}\right)\\\mathbf v_{12} &= \left(\frac{1}{2},0,0,\frac{1}{2},0,0\right)\end{aligned}$ \\
\hline
\end{tabular}
\vspace{3mm}
\caption{Vertex orbits of the symmetric initial hull $\mathcal B_1^G$ obtained by procedure \textsf{syminitialhull}.}\label{tab:syminithull_v}
\end{center}
\end{table}
\begin{table}[h]
\vspace{3mm}
\begin{center}
\begin{tabular}{|c|r|}
\hline
Orbit Label & Member facets\\
\hline\hline
$\mathcal O_H^1$ & $\begin{aligned}2\omega_2+\omega_3 &\leq 2R_6 & (\mathbf h_3)\\ 2\omega_2+\omega_3 &\leq 2R_4 & (\mathbf h_6)\\2\omega_2+\omega_3 &\leq 2R_5 & (\mathbf h_{10})\end{aligned} $ \\
\hline
$\mathcal O_H^2 $& $\begin{aligned}-\omega_1 &\leq 0 & (\mathbf h_9)\end{aligned}$ \\
\hline
$\mathcal O_H^3 $& $\begin{aligned}-\omega_2 &\leq 0 & (\mathbf h_5)\end{aligned}$ \\
\hline
$\mathcal O_H^4 $& $\begin{aligned}-\omega_3 &\leq 0 & (\mathbf h_7)\end{aligned}$ \\
\hline
$\mathcal O_H^5$ & $\begin{aligned}\omega_1+\omega_2+\omega_3+R_4+R_5+R_6 &\leq 1 & (\mathbf h_2)\end{aligned}$ \\
\hline
$\mathcal O_H^6 $& $\begin{aligned}2\omega_1+6\omega_2+3\omega_3 &\leq 2R_4+ 2R_5+2R_6 & (\mathbf h_{18})\\ \end{aligned}$ \\
\hline
\end{tabular}
\vspace{3mm}
\caption{Facet orbits of the symmetric initial hull $\mathcal B_1^G$ obtained by procedure \textsf{syminitialhull}.}\label{tab:syminithull_h}
\end{center}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[scale=1.4]{syminithull1.pdf
\end{center}
\caption{Complement of the incidence graph between vertex orbits and facet orbits of the symmetric initial inner bound $\mathcal B_1^G$ obtained by $\textsf{syminitialhull}(\cdot)$ procedure.}\label{fig:syminithull}
\end{figure}
While the previous example provided some detail regarding the mechanics of this complexity reduction with the running example, of additional interest is the types of complexity reductions of this type that can be achieved in other problems of interest. The next example addresses this question.
\begin{figure}
\centering
\includegraphics[width=3in]{idscAllShrink-jw.pdf}
\caption{Six hypergraph network coding problems that are independent distributed source coding problems that have network symmetry groups of order 6.}\label{fig:sixIDSC}
\end{figure}
\exmp{[Reduction in $\#$ of LPs] Consider the 6 problems displayed in Fig. \ref{fig:sixIDSC}, which include the running example and 5 other small network coding/IDSC problems, each with network symmetry group of order 6 as displayed under their diagram. Fig. \ref{fig:symCHMcr} displays the number of linear programs that CHM must solve versus the number of linear programs that symCHM must solve to compute the rate region for these problems. In each instance, a large reduction is achieved in the number of linear programs that must be solved.}
\begin{figure}
\centering
\includegraphics[width=3.1in]{symchm_barplot_dim-eps-converted-to.pdf}
\caption{The symmetry exploiting complexity reductions described in \S \ref{sec:symUp} and \S \ref{sec:symLp} enable the number of linear programs solved by CHM (left) to be reduced to the number of linear programs solved by symCHM (right). Furthermore, for checking whether a facet is terminal, the linear program can be reduced to the dimensions indicated by the colors in the legend.}\label{fig:symCHMcr}
\end{figure}
\subsection{Exploiting Symmetry when Updating the Inner Bound's Double Description Pair}\label{sec:symDD}
When a new extreme point $\mathbf{v}$ is added to the transversal $\mathcal T_V$, the transversal of the inequalities $\mathcal T_H$ must be updated (proc. \textsf{symupdatehull}) to reflect the new inequalities that the addition of the extreme points $\mathbf{v}^G$ to the symmetric inner bound creates (we call this new polytope the \textit{symmetric improvement}). In ordinary CHM, this would have been done through of $|\mathbf{v}^G|$ steps of the DD method applied to the complete inequality description of the symmetric inner bound. However, based on Lemma \ref{chm_incid}, which is the same insight from which the incidence decomposition method \cite{Bremner09polyhedralrepresentation} for representation conversion of symmetric polyhedra is derived, we can perform DD steps of smaller size (proc. \textsf{symDD}) to obtain the new facets that must be added to the transversal. The size of a DD step, in this context is the number of extreme rays in the input double description step.
\leml{Let $\mathcal P_k^{(\ell)}$ be an inner bound on $\textsf{proj}_k(\mathcal P)$ whose ASG has $G_p$ as a subgroup, and let $\mathbf v$ be a new vertex of a symmetric improvement $\mathcal P_k^{(\ell + 1)}$. If, in $\mathcal P_k^{(\ell + 1)}$, $\{f_1,\hdots,f_t\}$ is the set of facets incident to $\mathbf v$ then, $\{f_1^g,\hdots,f_t^g\}$ is the set of facets incident to $\mathbf v^g$.}{chm_incid}
\noindent Lemma \ref{chm_incid} ensures that as long as we calculate the facets of $\mathcal P_k^{(\ell+1)}$ incident to $\mathbf v$ correctly, and include any of these facets that are $G$-inequivalent into the new transversal $\mathcal T_H$ after removing those non-terminal inequalities that the new extreme points violate, the new facet transversal will reflect all of the $G$-inequivalent facets of $\mathcal P_k^{(\ell +1)}$. The key issue in calculating the facets incident to $\mathbf v$ in $\mathcal P_k^{(\ell +1)}$ correctly is that there may be some vertices in $\mathbf v^G \setminus \{\mathbf v\}$ that are adjacent to $\mathbf v$.
To address this issue, we first determine the double description pair of cone $\mathcal C_{\mathbf v^=}$, given as,
\begin{equation}
\mathcal{C}_{\mathbf{v}^=} \triangleq \mathcal{C} \cap \{ [1\ \mathbf{v}^T] \mathbf{x} = 0\} ,
\end{equation}
where $\mathcal C$ is the homogenized polar of the current inner bound $\mathcal{C} = \textrm{homog}(\mathcal P_k^{\ell})^{\circ}$. The double description pair associated with $\mathcal C$ is computed from the associated transversals in line \ref{ln:ddofc1} of proc. \textsf{symDD}. The double description pair of $\mathcal{C}_{\mathbf{v}^=}$ is determined by first determining the double description pair of $\mathcal C_{\mathbf v^\leq}=\mathcal C\cap \{ (1, \mathbf{v}^T) \mathbf{x} \leq 0\}$ through an iteration of double description step, in lines \ref{ln:dd} and \ref{ln:cineq} of proc. \textsf{symDD}. The procedure $\textsf{DD}$, used for this purpose, returns sets $P,Z$ and $N$ of extreme rays, where $P$ and $N$ are the extreme rays of $\mathcal C$ that strictly satisfy and violate the inequality $\{ (1, \mathbf{v}^T) \mathbf{x} \leq 0\}$ respectively, while the set $Z$ contains the extreme rays of $\mathcal C$ that evaluate to zero w.r.t. $\{ (1\ \mathbf{v}^T) \mathbf{x} \leq 0\}$ in addition to the new extreme rays that are computed as conic combinations of rays in $P$ and $N$. The set of extreme rays of $\mathcal C_{\mathbf v^=}$ is the set $Z$, i.e. the rays of $\mathcal C_{\mathbf v\geq}$ that have inequality $\{ (1, \mathbf{v}^T) \mathbf{x} = 0\}$ incident to them, while its inequality representation is formed by the subset of inequalities in $\hat{H_{\mathcal C}}$ that are adjacent to $\{ (1, \mathbf{v}^T) \mathbf{x} \leq 0\}$. The computation of double description pair of $\mathcal C_{\mathbf v^=}$ is condensed into proc. \textsf{tightenfacet} on line \ref{ln:tighten} of proc. \textsf{symDD}. Note that cone $\mathcal C_{\mathbf v^=}$ has dimension one lower that of $\mathcal C_{\mathbf v^\leq}$ along with having only a subset of its extrem rays and facets. The symmetry exploitation is achieved by using only $\mathcal C_{\mathbf v^=}$ from this poin onwards, to compute the symmetric update. To check to see if any vertices in $\mathbf v^G\setminus \{\mathbf v\}$ are adjacent to $\mathbf v$, and if so, which ones, we can determine the set $\mathcal A$, defined as,
\begin{equation}\label{eq:Aset}
\mathcal{A} = \left\{ (1,\mathbf z^T) \left| \mathbf{z} \in \mathbf{v}^G\setminus \{\mathbf{v}\} \min_{\mathbf{x} \in \mathcal{C}_{\mathbf{v}^=}} (1\ \mathbf{z}^T) \mathbf{x} < 0 \right. \right\}
\end{equation}
Procedure \textsf{repDD} (line \ref{ln:repdd} of \textsf{symDD}) is used to compute the set $\mathcal A$.
The rays of $V_{\mathcal C_{\mathbf v^=}}$ are further refined by adding the inequalities $\{ \mathbf{w}^T \mathbf{x} \leq 0\}$ for each $\mathbf{w} \in \mathcal{A}$ if any, through $|\mathcal{A}|$ further DD steps. This refinement is carried out in line \ref{ln:smalldd} of proc. \textsf{symDD}. The new inequality transversal of $\mathcal{P}_{k}^{(\ell+1)}$ is can be created by removing any $G$-equivalent inequalities from $\mathcal{P}_k^{\ell}$'s inequality description (i.e. rays in the homogenized polar) in these $| \mathcal{A} | + 1$ DD steps, and by adding the representatives of the new rays introduced at the end of these $|\mathcal{A}|+1$ DD steps, in line \ref{ln:newtrans} of proc. \textsf{symDD}, where proc. \textsf{lift} is used to embed the refined set of rays of $\mathcal C_{\mathbf v^=}$ into the higher dimensional space in which $\mathcal C$ existed.
The consideration of symmetry in this step of updating the inequality description of the inner bound, reduced both the number of DD steps required for CHM and their size. Indeed, only $|\mathcal{A}|+1$ DD steps must be performed to find the result of adding $|\mathbf{v}^G|$ new extreme points. Also, the size of each these DD steps is substantially smaller, since the cone $\mathcal{C}_{\mathbf{v}^{=}}$ is being dealt with in the $|\mathcal{A}|$ latter DD steps, rather than $\mathcal{C}$.
\begin{figure*}
\begin{center}
\includegraphics[scale = 0.5]{symddcube}
\vspace{2mm}
\caption{Symmetric update of an inner bound of a 3D cube. Part (a) shows the 3D cube in question with symmetry group $S_3$ which is the projection polytope to be computed, part (b) shows a 3D simplex that forms a symmetric inner bound to the projection, by the virtue of a symmetry group $S_3$. The inequalities and extreme points of (c) are shown in \eqref{eq:simplexv} and \eqref{eq:simplexh} respectively. Part (c) shows the updated inner bound obtained by adding vertex $\mathbf v=(1,0,1)$ to the description. Polar of homogenization of the polytope in (c) lies in $(\mathbb R^4)^\circ$, which is referred to as $\mathcal C_{\mathbf v^{\leq}}$, whose rays are shown in \eqref{eq:cvgeqrays}. The cone in part (d) is $\mathcal C_{\mathbf v^=}$, that lives in $(\mathbb R^3)^\circ$, whose rays are shown in \eqref{eq:cveqrays}. Parts (e) and (f) show the lower dimentional double description steps performed to obtain the symmetric update. The extreme rays of the cone in part (f) are shown in \eqref{eq:lastconev}}\label{cubeupdate}
\end{center}
\end{figure*}
\exmp{[Symmetric DD Update for a Cube]
This is a simple example elaborating how procedure \textsf{symDD} works. Consider a cube $\mathcal P_3\subseteq \mathbb R^3$ (Fig. \ref{cubeupdate}-a) with inequality representation,
\begin{equation}
H = \{0\leq\mathbf x_i\leq 1,i\in \{1,2,3\}\},
\end{equation}
and extreme points,
\begin{equation}
V=\left\{\begin{aligned} & (0,0,0),(0,0,1),(0,1,0),(0,1,1),\\ & (1,0,0),(1,0,1),(1,1,0),(1,1,1) \end{aligned}\right\}
\end{equation}
$\mathcal P_3$ is the projection of any hypercube $\mathcal P\subseteq \mathbb R^d,d\geq 4$. The symmetric group $S_3$ is a group of symmetries of $\mathcal P_3$, as $\mathcal P_3$ is stabilized setwise under any permutation of coordinate dimensions. A simplex $P_3^{(l)}\subseteq \mathcal P_3\subseteq \mathbb R^3$ (Fig. \ref{cubeupdate}-b) is the convex hull of $V^{(l)}\subseteq V$,
\begin{equation}\label{eq:simplexv}
V^{(l)}=\{(0,0,0),(0,0,1),(0,1,0),(1,0,0)\},
\end{equation}
and has inequality representation,
\begin{equation}\label{eq:simplexh}
H^{(l)} = \{0\leq\mathbf x_i,i\in \{1,2,3\}\}\cup\{\mathbf x_1+\mathbf x_2+\mathbf x_3\leq 1\}.
\end{equation}
Let $\mathcal P_3^{(l)}$ be the inner bound at the $l$th iteration of symCHM. Now consider the problem of computing the inequality representation $H^{(l+1)}$ of the symmetric update $P_3^{(l+1)}$, i.e. the convex hull of $V^{(l)}\cup \mathbf v^{S_3}$ where $\mathbf v$ is the vertex $(1,0,1)$, which is presumably found by solving a linear program over the aforementioned hypercube $\mathcal P$ (line \ref{line3} of symCHM). The first input to procedure $\textsf{symDD}$ is the transversal of the orbits of the rays of $\mathcal C=\textsf{homog}(P_3^{(l)})^\circ\subseteq (\mathbb R^4)^\circ$, given as,
\begin{equation}
\mathcal T_{\mathcal C}=\{ (0,-1,0,0),(-1,1,1,1)\},
\end{equation}
which bears one to one correspondance with transversal of orbits of $S_3$ in $H^{(l)}$. The second input is the inequality $ \mathbf y_0+\mathbf y_1+\mathbf y_3\leq 0$, which corresponds to vertex $\mathbf v$, after homogenization and polar. The initial double description step inserts this inequality in $\mathcal C$, giving the double description pair of cone $\mathcal C\cap\{\mathbf y|\mathbf y_0+\mathbf y_1+\mathbf y_3\leq 0\}\subseteq (\mathbb R^4)^\circ$ which corresponds to the polytope shown in Fig. \ref{cubeupdate}-c, in the original space. $\mathcal C\cap\{\mathbf y|\mathbf y_0+\mathbf y_1+\mathbf y_3\leq 0\}$ has inequality representation,
\begin{equation}
\begin{aligned}
H^{\mathcal C_{\mathbf v\geq}} &= T_1\cup T_2\cup T_3 \text{ where, }\\
& \begin{aligned} T_1 &=\left\{\mathbf y_0+\mathbf y_i\leq 0,\forall i\in \{1,2,3\}\right\}\\ T_2 &= \left\{\mathbf y_0\leq 0\right\}\\ T_3 &= \left\{\sum_{i\in \{0,1,3\}}\mathbf y_i\leq 0\right\}\end{aligned}
\end{aligned},
\end{equation}
where the last inequality is the newly added inequality, and extreme rays,
\begin{equation}\label{eq:cvgeqrays}
V^{\mathcal C_{\mathbf v^{\leq}}}=\left\{\begin{aligned}& (0,-1,0,0),(0,0,0,-1),(0,0,-1,0),\\
& (-1,0,1,1),(-1,1,1,0)\end{aligned}\right\}
\end{equation}
The first two rays in $V^{\mathcal C_{\mathbf v^{\leq}}}$ belong to the set $P$ in line \ref{ln:dd} of \textsf{symDD} and last three rays belong to set $Z$, while the set $N$ contains the ray $(-1,1,1,1)$, which in original space corresponds to the inequality $\mathbf x_1+\mathbf x_2+\mathbf x_3\leq 1$, is not part of the inequality representation of the symmetric updates, and is to be removed from consideration along with all its permutations in line \ref{ln:newtrans} at the end of proc. \textsf{symDD}($N^{S_3}=N$ in this case).
The cone $\mathcal C_{\mathbf v^{=}}$ is obtained by making last inequality in $H^{\mathcal C}$ tight, and has inequality representation,
\begin{equation}
H^{\mathcal C_{\mathbf v^{=}}}=\left\{\begin{aligned} & -\mathbf y_1\leq 0,-\mathbf y_1+\mathbf y_2-\mathbf y_3\leq 0, \\
& -\mathbf y_3\leq 0,-\mathbf y_1-\mathbf y_2\leq 0\end{aligned}\right\},
\end{equation}
which is obtained by substituting $\mathbf y_0=-\mathbf y_1-\mathbf y_3$ in the first 4 inequalities of $H^{\mathcal C}$. Note that the last inequality in $H^{\mathcal C_{\mathbf v^{=}}}$ is redundant, i.e. the cone $\mathcal C_{\mathbf v^{=}}$ remains unchanged even after removing this inequality. Such redundancy can be detected, using e.g. the algebraic adjacency oracle described by Fukuda et al. in \cite{fukuda1996double}. In the framework of Bremner et. al.\cite{Bremner09polyhedralrepresentation}, this would require solving of a linear program, a situation which is alleviated in our case by the knowledge of both descriptions of $\mathcal C$.
Cone $\mathcal C_{\mathbf v^{=}}$ is shown in Fig. \ref{cubeupdate}-d, and has extreme rays obtained from set $Z$ in line \ref{ln:dd} of \textsf{symDD},
\begin{equation}\label{eq:cveqrays}
V^{\mathcal C_{\mathbf v^=}}=\{(1,1,0),(0,-1,0),(0,1,1)\}.
\end{equation}
Next, we form inequalities $\mathbf y_0+\mathbf y_2+\mathbf y_3\leq 0$ and $\mathbf y_0+\mathbf y_1+\mathbf y_2\leq 0$, corresponding to other vertices in the orbit of $\mathbf v$ i.e. $\mathbf v^G\setminus\{\mathbf v\}=\{(0,1,1),(1,1,0)\}$.
Substituting $\mathbf y_0=-\mathbf y_1-\mathbf y_3$ in these inequalities, we get inequalities $-\mathbf y_1+\mathbf y_2\leq 0$ and $\mathbf y_2-\mathbf y_3\leq 0$. The set $\mathcal A$ is formed by checking if any rays in $V^{\mathcal C_{\mathbf v^=}}$ fail to satisfy these inequalities. In this case, ray $(0,1,1)$ fails inequality $-\mathbf y_1+\mathbf y_2\leq 0$ and ray $(1,1,0)$ fails inequality $\mathbf y_2-\mathbf y_3\leq 0$, implying that $\mathcal A=\{(1,0,1,1),(1,1,1,0)\}$. Standard DD steps are now performed (line \ref{ln:smalldd} of \textsf{symDD}), for inserting inequalities $-\mathbf y_1+\mathbf y_2\leq 0$ (obtained by substituting $\mathbf y_0=-\mathbf y_1-\mathbf y_3$ in $\mathbf y_0+\mathbf y_2+\mathbf y_3\leq 0$) and $\mathbf y_2-\mathbf y_3\leq 0$ in $\mathcal C_{\mathbf v^{=}}$ (obtained by substituting $\mathbf y_0=-\mathbf y_1-\mathbf y_3$ in $\mathbf y_0+\mathbf y_1+\mathbf y_2\leq 0$), as shown in Fig. \ref{cubeupdate}-e and Fig. \ref{cubeupdate}-f respectively. The cone in Fig. \ref{cubeupdate}-f has extreme rays,
\begin{equation}\label{eq:lastconev}
\{(0,-1,0),(0,0,1),(1,1,1),(1,0,0)\},
\end{equation}
which after prepending coordinate $\mathbf y_0=-\mathbf y_1-\mathbf y_3$ become,
\begin{equation}
\left\{\begin{aligned} & (0,0,-1,0),(-1,0,0,1),\\
& (-2,1,1,1),(-1,1,0,0)\end{aligned}\right\} \subseteq (\mathbb R^4)^\circ
\end{equation}
In the original space, these correspond to the inequalities,
\begin{equation}\label{eq:incidentineq}
\left\{-\mathbf x_2\leq 0,\mathbf x_3\leq 1, \mathbf x_1+\mathbf x_2+\mathbf x_3\leq 2,\mathbf x_1\leq 1\right\}
\end{equation}
which are the inequalities incident to vertex $\mathbf v=(1,0,1)$ in the symmetric update. The inequalities incident to rays in $\mathbf v^G\setminus \{\mathbf v\}$ are all obtained as permutations of those incident to $\mathbf v$, according to lemma \ref{chm_incid}. Finally, the inequality description of the symmetric update is obtained by permuting inequalities in \eqref{eq:incidentineq} under $S_3$ and taking union with the inequalities associated with set $P\setminus N^{S_3}$.
\begin{equation}
H^{(l+1)}=\left\{0\leq \mathbf x_i\leq 1,\forall i\in\{1,2,3\}\right\}\cup\left\{\sum_{i\in\{1,2,3\}}\mathbf x_i\leq 2\right\}
\end{equation}
}
The previous example was included for the purposes of demonstrating the need for consideration of adjacency of different elements of the orbit of the newly discovered extreme point. The next example continues the demonstration of the running example \ref{ex:run} for the symmetric bound update double descriptions step.
\exmp{[Symmetric Update for Ex. \ref{ex:run}]
the running example in Fig. \ref{fig:idsc_ex}, after construction the initial symmetric inner bound $\mathcal B_1^G$, consider the situation where a new vertex $\mathbf v_{10}$ is obtained by solving a linear program. The orbit of $\mathbf v_{10}$, i.e. $\mathbf v_1^G$ where $G$ is the network symmetry group, contains two additional vertices $\mathbf v_{9}$ and $\mathbf v_{11}$. The problem we now face is that of constructing the inequality description of convex hull of vertices of $\mathcal B_1^G$ and $\mathbf V_{10}^G$. In the procedure $\textsf{symDD}(\cdot)$, an initial asymmetric double description step is carried out to compute the inequality description of $\mathcal B_1^G$ augmented by $\mathbf v_{10}$. Following equations illustrate the construction of new inequalities that hold for the inner bound obtained by augmenting $\mathcal B_1^G$ by vertex $\mathbf v_{10}$ via an iteration of the double description method,
\begin{equation}
\begin{aligned}
\frac{1}{2}*&(2\omega_2 +& \omega_3 & &-2R_6&\leq 0 )& (\mathbf h_3)\\
+\frac{1}{2}*&(2\omega_2 +& \omega_3 &-2R_4& &\leq 0 )& (\mathbf h_6)\\
\cline{1-6}
& 2\omega_2 +& \omega_3 &-R_4 &-R_6&\leq 0 & (\mathbf h_{12})
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\frac{1}{2}*&( & -\omega_3 & &\leq 0 )& (\mathbf h_7)\\
+\frac{1}{2}*&(2\omega_2 +& \omega_3 &-2R_4&\leq 0 )& (\mathbf h_6)\\
\cline{1-6}
& \omega_2 +& &-R_4 &\leq 0 & (\mathbf h_{13})
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\frac{1}{2}*&(2\omega_2 +& \omega_3 & &-2R_5&\leq 0 )& (\mathbf h_{10})\\
+\frac{1}{2}*&(2\omega_2 +& \omega_3 &-2R_4& &\leq 0 )& (\mathbf h_6)\\
\cline{1-6}
& 2\omega_2 +& \omega_3 &-R_4 &-R_5&\leq 0 & (\mathbf h_{14})
\end{aligned}
\end{equation}
\begin{equation}
\begin{aligned}
\frac{1}{2}*&(2\omega_1+&6\omega_2 +& 3\omega_3 & -2R_4 &-2R_5 & -2R_6 &\leq 0 )& (\mathbf h_{18})\\
+\frac{1}{2}*&(&2\omega_2 +& \omega_3 &-2R_4& & &\leq 0 )& (\mathbf h_6)\\
\cline{1-8}
&\omega_1&+ 4\omega_2 +& 2\omega_3 &-2R_4 &-R_5 &-R_6 &\leq 0 & (\mathbf h_{15})
\end{aligned}
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[scale=1.4]{symddstep_example1_a.pdf
\end{center}
\caption{Illustration of the asymmetric Double Description step in line \ref{ln:dd} of procedure $\textsf{symDD}(\cdot)$. The colors red, yellow and green represent violation, equality, or strict satisfaction of a particular inequality by $\mathbf v_{10}$. The inequality violated by $\mathbf v_{10}$ i.e. $\mathbf h_6$ is combined with each inequality strictly satisfied by $\mathbf v_{10}$ i.e. $\mathbf h_2,\mathbf h_3,\mathbf h_{10}$, and $\mathbf h _{2}$, to produce the new inequalities that hold for the augmented inner bound i.e. $\mathbf h_{12},\mathbf h_{14},\mathbf h_{15},\mathbf h_{13}$.}\label{symdd_asym}
\end{figure}
Fig. \ref{symdd_asym} graphically describes the asymmetric double description step just carried out. We call the resultant asymmetric inner bound $\mathcal B_2$. After the asymmetric double description step, the cone $\mathcal C_{\mathbf v_{10}^=}$ is constructed, as described in table \ref{tab:embed_cveq}. Additional double description steps, if needed, are now carried out on the cone $\mathcal C_{\mathbf v_{10}^=}$. In this case, no additional steps are needed as the set $\mathcal A$ (given in \eqref{eq:Aset}) is empty. Thus the new symmetric inner bound $\mathcal B_2^G$ can be constructed by simply considering permutations of facets incident to $\mathbf v_{10}$ in $\mathcal B_2$. The vertex and facet orbits of $\mathcal B_2^G$ so constructed are described in tables \ref{tab:symsecond_v} and \ref{tab:symsecond_h} respectively, while Fig. \ref{fig:symsecond} shows the vertex-facet orbit incidences of $\mathcal B_2^G$.
}
\begin{table*}
\vspace{3mm}
\begin{center}
\begin{tabular}{|l|c|c|}
\hline
Vertex of $\mathcal B_2$ adjacent to $\mathbf v_{10}$ & Homogenized Polar Inequality & $\mathcal C_{\mathbf v_{10}^=}$ Inequality\\
\hline\hline
$\mathbf v_1=\left(\frac{1}{2},0,0,0,\frac{1}{2},0\right)$ & $\omega^\circ+\frac{1}{2}\omega_1^\circ+\frac{1}{2}R_5^\circ\leq 0$ & $\frac{1}{2}\omega_1^\circ-\frac{1}{3}\omega_3^\circ+\frac{1}{6}R_5^\circ-\frac{1}{3}R_6^\circ\leq 0$ \\
\hline
$\mathbf v_3=\left(\frac{1}{2},0,0,0,0,\frac{1}{2}\right)$ & $\omega^\circ+\frac{1}{2}\omega_1^\circ + \frac{1}{2}R_6^\circ\leq 0$ & $\frac{1}{2}\omega_1^\circ -\frac{1}{3}\omega_3^\circ - \frac{1}{3}R_5^\circ + \frac{1}{6}R_6^\circ\leq 0$ \\
\hline
$\mathbf v_5=\left(0,\frac{1}{4},0,\frac{1}{4},\frac{1}{4},\frac{1}{4}\right)$ & $\omega^\circ+\frac{1}{4}\omega_2^\circ + \frac{1}{4}R_4^\circ + \frac{1}{4}R_5^\circ+ \frac{1}{4}R_6^\circ\leq 0$ & $\frac{1}{4}\omega_2^\circ - \frac{1}{3}\omega_3^\circ + \frac{1}{4}R_4^\circ - \frac{1}{12}R_5^\circ - \frac{1}{12}R_6^\circ\leq 0$ \\
\hline
$\mathbf v_4=(0,0,0,0,1,0)$ & $\omega^\circ+R_5^\circ\leq 0$ & $-\frac{1}{3}\omega_3^\circ+\frac{2}{3}R_5^\circ -\frac{1}{3}R_6^\circ\leq 0$ \\
\hline
$\mathbf v_{8} = (0,0,0,0,0,1)$ & $\omega^\circ+R_6^\circ\leq 0$ & $-\frac{1}{3}\omega_3^\circ-\frac{1}{3}R_5^\circ +\frac{2}{3}R_6^\circ\leq 0$ \\
\hline
$\mathbf v_2=\left(0,0,0,0,0,0\right)$ & $\omega^\circ\leq 0$ & $-\frac{1}{3}\omega_3^\circ-\frac{1}{3}R_5^\circ -\frac{1}{3}R_6^\circ\leq 0$ \\
\hline
$\mathbf v_7=\left(0,0,\frac{2}{5},\frac{1}{5},\frac{1}{5},\frac{1}{5}\right)$ & $\omega^\circ+\frac{2}{5}\omega_3^\circ +\frac{1}{5}R_4^\circ + \frac{1}{5}R_5^\circ+ \frac{1}{5}R_6^\circ\leq 0$ & $\frac{1}{15}\omega_3^\circ +\frac{1}{5}R_4^\circ - \frac{2}{15}R_5^\circ - \frac{2}{15}R_6^\circ\leq 0$ \\
\hline
\end{tabular}
\vspace{3mm}
\caption{Construction of cone $\mathcal C_{\mathbf v_{10}^=}$. A new dummy variable $\omega$ is used to correspond to the extra dimension added by homogenization. $\circ$ on top of a variable corresponds to the polar dual of that variable. Thus, a vertex $\mathbf v_1=\left(\frac{1}{2},0,0,0,\frac{1}{2},0\right)$ is homogenized to get a ray $\left(1,\frac{1}{2},0,0,0,\frac{1}{2},0\right)$ which is interpreted as an inequality in the second column under polar duality. Vertex $\mathbf v_{10}=\left(0,0,\frac{1}{3},0,\frac{1}{3},\frac{1}{3}\right)$ under the same mapping yiends inequality $\omega^\circ + \frac{1}{3}\omega_3^\circ+\frac{1}{3}R_5^\circ +\frac{1}{3}R_6^\circ\leq 0$, which is made tight to give $\omega^\circ = -\frac{1}{3}\omega_3^\circ-\frac{1}{3}R_5^\circ -\frac{1}{3}R_6^\circ$. It is then substituted in column 2 inequalities to yield column 3 inequalities describing $\mathcal C_{\mathbf v_{10^=}}$}\label{tab:embed_cveq}
\end{center}
\end{table*}
\begin{table}[h]
\vspace{3mm}
\begin{center}
\begin{tabular}{|c|l|}
\hline
Orbit Label & Member vertices\\
\hline\hline
$\mathcal O_V^1 $& $\begin{aligned}\mathbf v_2&=\left(0,0,0,0,0,0\right)\end{aligned}$ \\
\hline
$\mathcal O_V^2$ & $\begin{aligned}\mathbf v_4&=(0,0,0,0,1,0)\\\mathbf v_6&=(0,0,0,1,0,0)\\\mathbf v_{8} &= (0,0,0,0,0,1)\end{aligned} $ \\
\hline
$\mathcal O_V^3 $& $\begin{aligned}\mathbf v_7&=\left(0,0,\frac{2}{5},\frac{1}{5},\frac{1}{5},\frac{1}{5}\right)\end{aligned}$ \\
\hline
$\mathcal O_V^4 $& $\begin{aligned}\mathbf v_5&=\left(0,\frac{1}{4},0,\frac{1}{4},\frac{1}{4},\frac{1}{4}\right)\end{aligned}$ \\
\hline
$\mathcal O_V^5$ & $\begin{aligned}\mathbf v_1&=\left(\frac{1}{2},0,0,0,\frac{1}{2},0\right)\\\mathbf v_3&=\left(\frac{1}{2},0,0,0,0,\frac{1}{2}\right)\\\mathbf v_{12} &= \left(\frac{1}{2},0,0,\frac{1}{2},0,0\right)\end{aligned}$ \\
\hline
$\mathcal O_V^6 $& $\begin{aligned}\mathbf v_9&=\left(0,0,\frac{1}{3},\frac{1}{3},\frac{1}{3},0\right)\\ \mathbf v_{10}&=\left(0,0,\frac{1}{3},0,\frac{1}{3},\frac{1}{3}\right)\\\mathbf v_{11}&=\left(0,0,\frac{1}{3},\frac{1}{3},0,\frac{1}{3}\right) \end{aligned}$ \\
\hline
\end{tabular}
\vspace{3mm}
\caption{Vertex orbits of the second symmetric inner bound $\mathcal B_2^G$ obtained by procedure $\textsf{symDD}(\cdot)$.}\label{tab:symsecond_v}
\end{center}
\end{table}
\begin{table}[h]
\vspace{3mm}
\begin{center}
\begin{tabular}{|c|r|}
\hline
Orbit Label & Member facets\\
\hline\hline
$\mathcal O_H^1$ & Deemed non-terminal \\
\hline
$\mathcal O_H^2 $& $\begin{aligned}-\omega_1 &\leq 0 & (\mathbf h_9)\end{aligned}$ \\
\hline
$\mathcal O_H^3 $& $\begin{aligned}-\omega_2 &\leq 0 & (\mathbf h_5)\end{aligned}$ \\
\hline
$\mathcal O_H^4 $& $\begin{aligned}-\omega_3 &\leq 0 & (\mathbf h_7)\end{aligned}$ \\
\hline
$\mathcal O_H^5$ & $\begin{aligned}\omega_1+\omega_2+\omega_3+R_4+R_5+R_6 &\leq 1 & (\mathbf h_2)\end{aligned}$ \\
\hline
$\mathcal O_H^6 $& $\begin{aligned}2\omega_1+6\omega_2+3\omega_3 &\leq 2R_4+ 2R_5+2R_6 & (\mathbf h_{18})\\ \end{aligned}$ \\
\hline
$\mathcal O_H^7 $& $\begin{aligned}\omega_2 &\leq R_4& (\mathbf h_{13})\\ \omega_2 &\leq R_5& (\mathbf h_{16})\\\omega_2 &\leq R_6& (\mathbf h_{11})\\ \end{aligned}$ \\
\hline
$\mathcal O_H^8 $& $\begin{aligned}\omega_1+4\omega_2+2\omega_3 &\leq 2R_4+ R_5+R_6 & (\mathbf h_{15})\\\omega_1+4\omega_2+2\omega_3 &\leq R_4+ 2R_5+R_6 & (\mathbf h_{19})\\\omega_1+4\omega_2+2\omega_3 &\leq R_4+ R_5+2R_6 & (\mathbf h_{20})\\ \end{aligned}$ \\
\hline
$\mathcal O_H^9 $& $\begin{aligned}2\omega_2+\omega_3 &\leq R_4+R_6 & (\mathbf h_{12})\\2\omega_2+\omega_3 &\leq R_4+ R_5& (\mathbf h_{14})\\2\omega_2+\omega_3 &\leq R_5+R_6 & (\mathbf h_{17})\\ \end{aligned}$ \\
\hline
\end{tabular}
\vspace{3mm}
\caption{Facet orbits of the second symmetric inner bound $\mathcal B_2^G$ obtained by procedure $\textsf{symDD}(\cdot)$.}\label{tab:symsecond_h}
\end{center}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=3.15in]{symddstep_example1_d.pdf
\end{center}
\caption{Complement of the incidence graph between vertex orbits and facet orbits of the second symmetric inner bound $\mathcal B_2^G$ obtained by $\textsf{symDD}(\cdot)$ procedure. This bound is in fact the final bound, as all its faces are terminal.}\label{fig:symsecond}
\end{figure}
In many contexts, the known symmetry group of the projected polyhedron may not include all of its symmetries. The next example shows the effect of increasing knowledge of symmetry on a symmetric double descriptions pair updates, the reduction of the number of LPs to solve, and the resulting reduction in run-time of an implementation, when projecting a hypercube.
\exmp{[Varying Symmetry Knowledge when Projecting a Hypercube]
When polytope $\mathcal P$ is a hypercube in $\mathbb R^{12}$, which we want to project down to $\mathbb R^9$, the sizes of DD steps under varying knowledge of symmetry is shown in Fig. \ref{fig:stepsizeplot}. In this case, the number of double description steps does not reduce
i.e. $\vert \mathcal A\vert=\vert\mathbf v^G\vert -1$, in line \ref{ln:repdd} of proc. \textsf{symdd}, for every symmetric update. The main tool to gauge the efficiency of the symmetric updates deescribed in this section, we consider the sizes of extreme ray descriptions input to the double description steps. The average stepsizes under knowledge of no symmetries, cyclic group $C_9$ and symmetric group $S_9$ are $978.97$, $245.38$ and $120.54$ respectively. Fig. \ref{fig:numlpplot} shows the number of LPs solved under varying knowledge of symmetry for projecting $\mathcal P$ to different dimensions, while Fig. \ref{fig:runtimesplot} shows the time required for projection vs the dimension under varying knowledge of symmetry.}
\begin{figure}[h]
\begin{center}
\includegraphics[width=3in]{cube12projddszes9.pdf}%
\end{center}
\caption{Sizes of double description steps for finding $i$th vertex of the projection of a 12-dimensional hypercube to 9 dimensions versus $i$, under varying knowledge of symmetry.}\label{fig:stepsizeplot}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=3in]{cube12projnumlp.pdf}%
\end{center}
\caption{Number of LPs solved for projection of a 12-dimensional hypercube versus the dimension of projection, under varying knowledge of symmetry.}\label{fig:runtimesplot}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=3in]{cube12projtimes.pdf}%
\end{center}
\caption{Time is seconds required for projection of a 12-dimensional hypercube versus the dimension of projection, under varying knowledge of symmetry.}\label{fig:numlpplot}
\end{figure}
\subsection{Exploiting Symmetry to Reduce the Dimension of the Linear Program}\label{sec:symLp}
The complexity reductions in the previous two sections are based on exploiting knowledge of the symmetry group of the projected polyhedron that can be inferred from those of the parent polyhedron to project. However, building on standard symmetry exploitation techniques for symmetric linear programs \cite{herr09arxiv,bodyhighsymmlpalgo}, knowledge of symmetries of the parent polyhedron can also be heavily exploited to reduce the dimension of the linear programs that must be solved.
Taking the subgroup of any known symmetries of the parent polyhedron that fixes the cost in the linear program to solve, one may reduce the dimension of the constraint space (parent polyhedron) by adding the constraints restricting it to the subspace that is fixed under this subgroup without affecting the minimal cost obtained by solving the LP. The can be executed for both known coordinate permutation symmetries, as well as known affine and restricted affine symmetries.
In the case of symCHM, this capability to reduce the dimension of the LP without affecting the cost can be exploited when checking if a given facet is terminal by solving the associated lower dimensional linear program. Only in the event that a facet is found to be non-terminal is it necessary to solve a higher dimensional linear program to ensure a new extreme point of the projection's inner bound is revealed.
\exmp{[LP Dimension Reduction for Fig. \ref{fig:sixIDSC}] By applying these dimensionality reduction using the coordinate permutation symmetries (network symmetry group), one can reduce the dimension of the LP checking for terminal facets when symCHM is applied to the networks in Fig. \ref{fig:sixIDSC} according to the legend in Fig. \ref{fig:symCHMcr}.}
\subsection{Symmetry Exploitation in Weighted Sum-Rate Computation and Related Problems}\label{sec:symLpOth}
An especially attractive application of the dimension reduction trick mentioned in the previous section applies not only to symCHM for full polyhedral projection, but also to simpler problems in network information theory such as weighted sum-rate problems.
Here, we are given an outer bound $\Gamma_{\textsf{out}}$ on the entropy region, a HMSNC instance $\textsf{A}$, weights $\lambda_1,\hdots,\lambda_k$ for each of the $k$ sources and capacities
$r_{k+1},\hdots,r_N$ of the edges. The weighted sum-rate bound associated with $\Gamma_{\textsf{out}}$, is the solution to the following linear program:
\begin{equation}\label{eq:sumratelp}
\max_{\mathcal P} \sum_{i\in [k]}\lambda_i\omega_i
\end{equation}
where,
\begin{equation}
\mathcal P\triangleq \{(\widetilde{\boldsymbol\omega},\widetilde{\mathbf r},\widetilde{\mathbf h})\in \Gamma_{\textsf{out}}\cap\mathcal L'\cap\mathcal L''| \widetilde{\mathbf r_i}=\mathbf r_i,\forall i\in[N]\setminus [k]\}
\end{equation}
A key consideration that enables the dimension reduction is the symmetries of the following vector,
\begin{equation}\label{eq:lambvec}
\boldsymbol\delta=(\lambda_1,\hdots,\lambda_k,r_{k+1},\hdots,r_N)^T
\end{equation}
under the action of the symmetry group $G$ of $\Gamma_{\textsf{out}}\cap\mathcal L'\cap\mathcal L''\subseteq \mathbb R^M$. If $G$ is an ASG of $\mathcal P$, let $G'$be the subgroup of $G$ s.t.
\begin{equation}
G'=\left\{ [\mathbf{b},\mathbf{A}] \in G \left|\begin{aligned} & [\boldsymbol\delta^T \mathbf{0}_{M-N}^T] (\mathbf{A} - \mathbb{I}_{M}) = \mathbf{0}_M,\\
& [\boldsymbol{\delta}^T \mathbf{0}_{M-N}^T] \mathbf{b} = 0 \end{aligned}\right. \right\}
\end{equation}
If $G'$ is the subgroup of $G$ that fixes vector \eqref{eq:lambvec}, the solution of the linear program in equation \eqref{eq:sumratelp} can be obtained by solving the following linear program instead,
\begin{equation}
\max_{\textsf{Fix}_{G'}(\mathcal P)} \sum_{i\in [k]}\lambda_i\omega_i
\end{equation}
where $\textsf{Fix}_{G'}(\mathcal P)$ is defined as,
\begin{equation}
\textsf{Fix}_{G'}(\mathcal P) = \left\{ \mathbf{y} \in \mathcal P \left| \mathbf{A}\mathbf{y} + \mathbf{b} = \mathbf{y},\ \forall [\mathbf{b},\mathbf{A}] \in G' \right. \right\}.
\end{equation}
The dimension of $\textsf{Fix}_{G'}(\mathcal P)$ is expected to be much smaller than that of $\mathcal P$. For example, if $G'$ is made up of permutation matrices, i.e. it is a subgroup of $S_M$, the dimension of $\textsf{Fix}_{G'}(\mathcal P)$ is equal to number of orbits of $G'$ in set $[M]$.
On the same lines as NSG, one can compute symmetry groups of access structures in secret sharing\cite{padronotes}, and directed graphs in the context of guessing games \cite{riis06}. Computation of lower bounds on worst case information ratio in secret sharing and upper bounds on guessing numbers of graphs, using information inequalities, has the same semantics as the computation of weighted sum-rate bounds in network coding, allowing symmetry to be exploited in identical manner.
\section{The Information Theoretic Converse Prover} \label{sec:ITCP}
Provided along with this manuscript is a GAP \cite{GAP4} package, the Information Theoretic Converse Prover (ITCP)\cite{jayantitcp}, that implements the described techniques and algorithms. To ensure that the proofs are exact and not riddled with floating point precision issues, the inbuilt rational arithmetic of GAP from GMP\cite{Granlund12} is exploited along with exact rational linear programming are achieved via an interface to QSopt\_ex linear programming solver \cite{qsoptex}. The default outer bound on entropy region, with respect to which EPOBs and other bounds can be computed, is the Shannon outer bound $\Gamma_N$. ITCP also contains an inbuilt database of $215$ non-Shannon inequalities that are inequivalent under random variable permutations, which includes the inequality of Zhang and Yeung \cite{zyineq} and those of Dougherty,Freiling and Zeger \cite{DBLP:journals/corr/abs-1104-3602}. The user can optionally specify a subset of these inequalities to be considered in the computation.
While the series of examples throughout the manuscript have served the purpose of illustrating the ideas, in this section we aim to demonstrate how to use the GAP package to perform the types of calculations described in the manuscript. To this end, a series of examples are accompanied with the associated transcripts from sessions with the GAP package.
The first example demonstrates how to compute a network symmetry group with ITCP.
\exmpl{[Computing a Network Symmetry Group with ITCP] Fig. \ref{fig:itcp1} contains the transcript of a GAP session in which ITCP is utilized to calculate a network symmetry group. The network coding instance used here is an IDSC \cite{li2015computer} system, that is constructed so that it has a desired symmetry group. As mentioned in \S\ref{sec:netCodSym}, the network symmetry group is the direct product of two groups the form $G_1\times G_2$ where group $G_1$ is a group of permutations of source labels while group $G_2$ is a group of permutations of edge labels. The instance considered here has size $8$ while its NSG $G$ has order $20$, and is isomorphic to $S_2\times D_5$, where $D5$ is the dihedral group corresponding to the symmetries of a regular pentagon. Starting from group $G$ and $5$ encoders, this instance can be constructed by, 1) adding decoders demanding sources $\{1,2\}$, having access to a set in the orbit of $\{4,5\}$ under $G$, and 2) adding decoders demanding sources $\{3\}$, having access to a set in the orbit of $\{4,6\}$ under $G$.}{itcpex1}
\begin{figure*}
\begin{Verbatim}[commandchars=!@|,fontsize=\small,frame=single,label=Sample ITCP session for example \ref{itcpex1}]
!gapprompt@gap>| !gapinput@# Define a size 8 IDSC instance|
!gapprompt@>| !gapinput@idsc:=[ [ [ [ 1, 2, 3 ], [ 1, 2, 3, 4, 5, 6, 7, 8 ] ],\|
!gapprompt@>| !gapinput@ [ [ 4, 5 ], [ 1, 2, 4, 5 ] ], [ [ 5, 6 ], [ 1, 2, 5, 6 ] ],\|
!gapprompt@>| !gapinput@ [ [ 6, 7 ], [ 1, 2, 6, 7 ] ], [ [ 7, 8 ], [ 1, 2, 7, 8 ] ],\|
!gapprompt@>| !gapinput@ [ [ 4, 8 ], [ 1, 2, 4, 8 ] ], [ [ 4, 6 ], [ 3, 4, 6 ] ],\|
!gapprompt@>| !gapinput@ [ [ 5, 8 ], [ 3, 5, 8 ] ], [ [ 4, 7 ], [ 3, 4, 7 ] ],\|
!gapprompt@>| !gapinput@ [ [ 5, 7 ], [ 3, 5, 7 ] ], [ [ 6, 8 ], [ 3, 6, 8 ] ] ], 3, 8 ];|
!gapprompt@gap>| !gapinput@G:=NetSymGroup(idsc);|
Group([ (5,8)(6,7), (4,5)(6,8), (4,6)(7,8), (1,2) ])
!gapprompt@gap>| !gapinput@Size(G);|
20
\end{Verbatim}
\caption{Example \ref{itcpex1}: Computing a network symmetry group with ITCP.}\label{fig:itcp1}
\end{figure*}
The next example demonstrates how to calculate an information theoretic converse EPOB via ITCP.
\exmpl{[Shannon Outer Bound for the Fano Network] Fig. \ref{fig:itcpex2} contains the transcript of a GAP session in which ITCP is utilized to calculate the Shannon outer bound to a network coding rate region. The HMSNC instance considered in this example is the Fano network, which is a well-known matroidal network \cite{DFZMatroidNetworks}, whose network symmetry group is trivial. The EPC is computed using the Shannon outer bound $\Gamma_7$. If we substitute $R_i=1,\forall i\in\{1,\ldots,7\}\setminus \{1,2,3\}$ in the rate region shown in the sample session, we get the region described by Dougherty, Freiling and Zeger in \cite{dougherty2015achievable}, which is a 3 dimensional cube, while the more general list of inequalities describing the region are calculated and given here.}{itcpex2}
\begin{figure*}
\begin{Verbatim}[commandchars=!@|,fontsize=\small,frame=single,label=Sample ITCP session for example \ref{itcpex2}]
!gapprompt@gap>| !gapinput@# define a network instance (in this case, Fano network)|
!gapprompt@>| !gapinput@F:=[ [ [ [ 1, 2 ], [ 1, 2, 4 ] ], [ [ 2, 3 ], [ 2, 3, 5 ] ],\|
!gapprompt@>| !gapinput@ [ [ 4, 5 ], [ 4, 5, 6 ] ], [ [ 3, 4 ], [ 3, 4, 7 ] ],\|
!gapprompt@>| !gapinput@ [ [ 1, 6 ], [ 3, 1, 6 ] ], [ [ 6, 7 ], [ 2, 6, 7 ] ],\|
!gapprompt@>| !gapinput@ [ [ 5, 7 ], [ 1, 5, 7 ] ] ], 3, 7 ];;|
!gapprompt@gap>| !gapinput@rlist:=NCRateRegionOB(F,true,[]);;|
!gapprompt@gap>| !gapinput@Display(rlist[2]);|
0 >= -w2
0 >= -w1
0 >= -w3
+R6 >= +w3
+R5 >= +w3
+R7 >= +w1
+R4 >= +w1
+R6 +R7 >= +w2 +w3
+R4 +R6 >= +w2 +w3
+R4 +R5 >= +w2 +w3
+R6 +R7 >= +w1 +w2
+R4 +R6 >= +w1 +w2
+R4 +R5 >= +w1 +w2
\end{Verbatim}
\caption{Example \ref{itcpex2}: Computing the Shannon outer bound on the Fano network, which has the trivial network symmetry group, with polyhedral projection via ITCP.}\label{fig:itcpex2}
\end{figure*}
We next pass to demonstrating how to use ITCP to calculate the rate region for the running example in the manuscript.
\exmpl{[Shannon Outer Bound for Ex. \ref{ex:run}] Fig. \ref{fig:itcpex3} gives the transcript of a session in GAP in which ITCP is utilized to calculate the Shannon outer bound to a rate region for a symmetric network. In such cases when NSG is non-trivial, ITCP outputs only one inequality per equivalence class under the action of NSG $G$. There are $18$ such equivalence classes for the IDSC instance under consideration, as shown in the sample ITCP session, while there are a total of 94 distinct permuted forms of these inequalities that form the facets of $\mathcal R_{\textsf{out}}$. A total of $119$ LPs are solved during this computation, while if on decides to ignore symmetries, $207$ LPs must be solved instead. The time required for computation of the EPOBs is 43.91 sec and 67.72 sec, with and without the knowledge of symmetries, respectively.}{itcpex3}
\begin{figure*}
\begin{Verbatim}[commandchars=!@|,fontsize=\small,frame=single,label=Sample ITCP session for example \ref{itcpex3}]
!gapprompt@gap>| !gapinput@rlist1:=NCRateRegionOB2(idsc,true,[]);;|
!gapprompt@gap>| !gapinput@Display(rlist1[2]);|
0 >= -w2
0 >= -w3
+R4 >= 0
+R4 +R6 >= +w3
+R4 +R5 >= +w1 +w2
+R4 +1/2 R5 +1/2 R8 >= +w1 +w2 +1/2 w3
+1/2 R4 +1/2 R5 +1/2 R6 +1/2 R7 >= +w1 +w2 +1/2 w3
+2/3 R4 +2/3 R5 +1/3 R6 +1/3 R8 >= +w1 +w2 +2/3 w3
+2/3 R4 +1/3 R5 +1/3 R6 +1/3 R7 +1/3 R8 >= +w1 +w2 +2/3 w3
+1/2 R4 +1/2 R5 +1/2 R6 +1/4 R7 +1/4 R8 >= +w1 +w2 +3/4 w3
+R4 +1/2 R5 +1/2 R6 +1/2 R7 >= +w1 +w2 +w3
+R4 +1/2 R5 +1/2 R6 +1/2 R8 >= +w1 +w2 +w3
+R4 +1/3 R5 +1/3 R6 +1/3 R7 +1/3 R8 >= +w1 +w2 +w3
+2/3 R4 +2/3 R5 +1/3 R6 +2/3 R7 +1/3 R8 >= +w1 +w2 +4/3 w3
+R4 +1/2 R5 +1/2 R6 +R7 >= +w1 +w2 +3/2 w3
+R4 +1/2 R5 +1/2 R6 +1/2 R7 +1/2 R8 >= +w1 +w2 +3/2 w3
+2 R4 +R6 +R7 >= +w1 +w2 +2 w3
+R4 +R5 +R6 +R7 >= +w1 +w2 +2 w3
\end{Verbatim}
\caption{Example \ref{itcpex3}: Using ITCP to calculate the rate region in Ex. \ref{ex:run}.}\label{fig:itcpex3}
\end{figure*}
In addition to computing explicit polyhedral outer bounds for network coding rate regions, ITCP contains driver routines to calculate upper bounds on the sum rate of a network, lower bounds on the worst case information ratio for a secret sharing problem, and upper bounds on the guessing number of a graph, all of which can exploit problem symmetry in the manner described in \S \ref{sec:symLpOth}. The remaining three examples demonstrate these capabilities, with the GAP transcript indicating the substantially lower dimensional linear program that is obtained after using the techniques from \S \ref{sec:symLpOth}.
\exmpl{[Upper Bound on a Network Coding Sum Rate with ITCP] Fig. \ref{fig:itcpex4} gives the transcript of a GAP session in which ITCP is utilized to upper bound the sum rate for a network coding problem. The selected problem is a size $5$ HMSNC instance, with symmetry group of order $2$, that contains permutations $\{(1),(3,4)\}$. By exploiting these symmetries, the dimension of the linear program can be reduced from 28 to 22.}{itcpex4}
\begin{figure*}
\begin{Verbatim}[commandchars=!@|,fontsize=\small,frame=single,label=Sample ITCP session for example \ref{itcpex4}]
!gapprompt@gap>| !gapinput@# define a network instance|
!gapprompt@>| !gapinput@N:= [[ [ [ 1 ], [ 1, 3 ] ], [ [ 1 ], [ 1, 4 ] ],[ [ 1, 2, 5 ],\|
!gapprompt@>| !gapinput@ [ 1, 2 ] ],[ [ 1, 2, 3 ], [ 2, 3 ] ],[ [ 2, 4 ], [ 1, 2, 4 ] ],\|
!gapprompt@>| !gapinput@ [ [ 2, 3, 4, 5 ], [ 3, 4, 5 ] ]] , 2, 5 ];;|
!gapprompt@gap>| !gapinput@ub:=NCSumRateUB(N,[1,1,1],[]);;|
Original LP dimension...28
LP dimension after considering symmetries...22
!gapprompt@gap>| !gapinput@ub;|
2
\end{Verbatim}
\caption{Example \ref{itcpex4}: computing a network coding sum rate bound using ITCP.}\label{fig:itcpex4}
\end{figure*}
\exmpl{[Secret Sharing Information Ratio Bound via ITCP] Fig. \ref{fig:itcpex5} gives a transcript of a GAP session in which ITCP is used to lower bound the worst case information ratio lower bounds for a specified secret sharing access structure. The access structure considered here has authorized sets $\{\{2,3\},\{3,4\},\{4,5\}\}$, where there are 4 parties labeled $\{2,3,4,5\}$ while the dealer of secret is labeled $1$. The lower bound is computed with respect to $\Gamma_N$, which is already known in the literature to be $\frac{3}{2}$, which is also achievable using multi-linear scheme (see \cite{padronotes} \S 2.8). The symmetry detected in the problem decreases the dimension of the associated linear program from 20 to 12.}{itcpex5}
\begin{figure*}
\begin{Verbatim}[commandchars=!@|,fontsize=\small,frame=single,label=Sample ITCP session for example \ref{itcpex5}]
!gapprompt@gap>| !gapinput@# define an access structure|
!gapprompt@>| !gapinput@Asets:=[[2,3],[3,4],[4,5]];;|
!gapprompt@gap>| !gapinput@lb:=SSWorstInfoRatioLB(Asets,5,[]);;|
Original LP dimension...20
LP dimension after considering symmetries...12
!gapprompt@gap>| !gapinput@lb;|
3/2
\end{Verbatim}
\caption{Example \ref{itcpex5}: computing a secret sharing information ratio bound with ITCP.}\label{fig:itcpex5}
\end{figure*}
\exmpl{[Bounding the Guessing Number of a Graph with ITCP] Fig. \ref{fig:itcpex5} gives a GAP transcript of a session wherein ITCP is used to determine upper bounds on the guessing number of a directed graph. The graph under consideration is the cycle graph $C_5$. The upper bound obtained using $\Gamma_5$, in this case, is $\frac{5}{2}$ (see eg. \cite{atkinsguessoddcycle}). The problem dimension is reduced from 25 to 5 by exploiting symmetry.}{itcpex6}
\begin{figure*}
\begin{Verbatim}[commandchars=!@|,fontsize=\small,frame=single,label=Sample ITCP session for example \ref{itcpex6}]
!gapprompt@gap>| !gapinput@# define a directed graph (in this case, the cycle graph C5)|
!gapprompt@>| !gapinput@C5:=[ [ 1, 2, 3, 4, 5 ],\|
!gapprompt@>| !gapinput@ rec( 1 := [ 2, 5 ], 2 := [ 1, 3 ], 3 := [ 2, 4 ],\|
!gapprompt@>| !gapinput@ 4 := [ 3, 5 ], 5 := [ 4, 1 ] ) ];;|
!gapprompt@gap>| !gapinput@ub:=GGnumberUB(C5,[]);;|
Original LP dimension...25
LP dimension after considering symmetries...5
!gapprompt@gap>| !gapinput@ub;|
5/2
\end{Verbatim}
\caption{Example \ref{itcpex6}: computing an upper bound on the guessing number of a directed graph.}\label{fig:itcpex6}
\end{figure*}
\section{Conclusions}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\subsection{The problem}
In this paper we are concerned with local or global $W^{1,p}$ solutions to the Poisson equation for the $p-$Laplace operator, i.e.
\begin{equation}
\label{plaplace}
-\Delta_p u:=-\mathrm{div}(|\nabla u|^{p-2}\,\nabla u)=f,\qquad \mbox{ in }\Omega,
\end{equation}
with $\Omega\subset \mathbb{R}^N$ open set.
Our analysis is confined to the super-quadratic case, i.e. throughout the whole paper we consider $p>2$.
For $f\equiv 0$, we know that
\begin{equation}
\label{karen}
|\nabla u|^\frac{p-2}{2}\,\nabla u\in W^{1,2}_{\rm loc}(\Omega;\mathbb{R}^N),
\end{equation}
This is a well-known regularity result which dates back to Uhlenbeck, see \cite[Lemma 3.1]{Uh}. We refer to \cite[Proposition 3.1]{Mi03} for a generalization of this result. If $f$ is smooth enough, the same result is easily seen to be still true. For example, it is sufficient to take
\begin{equation}
\label{fsmooth}
f\in W^{1,p'}_{\rm loc}(\Omega),
\end{equation}
where $p'=p/(p-1)$.
However, it is easy to guess that assumption \eqref{fsmooth} is far from being optimal: in the limit case $p=2$, \eqref{karen} boils down to
\[
\nabla u\in W^{1,2}_{\rm loc}(\Omega;\mathbb{R}^N).
\]
Then from {\it Calder\'on-Zygmund estimates} for the Laplacian, we know that this is true if (and only if)
\[
f\in L^2_{\rm loc}(\Omega).
\]
Thus in this case $f\in L^2_{\rm loc}$ would be the sharp assumption in order to get Uhlenbeck's result.
The main concern of this work is to prove Uhlenbeck's result for solutions of \eqref{plaplace}, under sharp assumptions on $f$.
\subsection{The main result} In this paper we prove the following result. We refer to Section \ref{sec:2} for the notation.
\begin{teo}
\label{teo:local}
Let $p>2$ and let $U\in W^{1,p}_{\rm loc}(\Omega)$ be a local weak solution of equation \eqref{plaplace}. If
\[
f\in W^{s,p'}_{\rm loc}(\Omega)\qquad \mbox{ with } \frac{p-2}{p}<s\le 1,
\]
then
\[
\mathcal{V}:=|\nabla U|^\frac{p-2}{2}\,\nabla U\in W^{1,2}_{\rm loc}(\Omega;\mathbb{R}^N),
\]
and
\[
\nabla U\in W^{\sigma,p}_{\rm loc}(\Omega;\mathbb{R}^N),\qquad \mbox{ for }0<\sigma<\frac{2}{p}.
\]
Moreover, for every pair of concentric balls $B_r\Subset B_R\Subset\Omega$ and every $j=1,\dots,N$ we have
the local scaling invariant estimate
\begin{equation}
\label{filipposobolev}
\begin{split}
\int_{B_r} \left|\mathcal{V}_{x_j}\right|^2\,dx&\le \frac{C}{(R-r)^2}\int_{B_R} |\nabla U|^p\,dx+C\left(R^{\left(s-\frac{p-2}{p}\right)}\,[f]_{W^{s,p'}(B_R)}\right)^{p'},
\end{split}
\end{equation}
for a constant $C=C(N,p,s)>0$ which blows-up as $s\searrow (p-2)/p$.
\end{teo}
\begin{oss}[Sharpness of the assumption]
The assumption on $f$ in the previous result is essentially sharp, in the sense that {\it the result is false for} $s<(p-2)/p$, see Section \ref{sec:5} for an example. Also observe that
\[
\frac{p-2}{p}\searrow 0 \quad \mbox{ as }\quad p\searrow 2,
\]
thus the assumption on $f$ is consistent with the case of the Laplacian recalled above.
\end{oss}
\begin{oss}[Comparison with other results]
The interplay between regularity of the right-hand side $f$ and that of the vector field $\mathcal{V}$ has been considered in detail by Mingione in \cite{Mi07}. However, our Theorem \ref{teo:local} does not superpose with the results of \cite{Mi07}. Indeed, the point of view in \cite{Mi07} is slightly different: the main concern there is to obtain (fractional) differentiability of the vector field
\[
\mathcal{V}:=|\nabla U|^\frac{p-2}{2}\,\nabla U,
\]
when $f$ is not regular. In particular, in \cite{Mi07} the right-hand side $f$ may not belong to the relevant dual Sobolev space and the concept of solution to \eqref{plaplace} has to be carefully defined.
\par
In order to give a flavour of the results in \cite{Mi07}, we recall that \cite[Theorem 1.3]{Mi07} proves that if $2<p\le N$ and $f$ is a Radon measure, then
\[
\mathcal{V}\in W^{\frac{\tau}{2},\frac{2}{p'}}_{\rm loc}(\Omega;\mathbb{R}^N),\qquad \mbox{ for every } 0<\tau<p'.
\]
When the Radon measure $f$ is more regular, accordingly one can improve the differentibility of $\mathcal{V}$. For example, \cite[Theorem 1.6]{Mi07} proves that if $2<p\le N$ and $f\in L^{1,\lambda}(\Omega)$ with $N-p<\lambda\le N$, then\footnote{The result of \cite[Theorem 1.6]{Mi07} is not stated for $\mathcal{V}$, but rather directly for $\nabla U$. However, an inspection of the proof reveals that one has the claimed regularity of $\mathcal{V}$, see \cite[proof of Theorem 1.6, page 33]{Mi07}.}
\[
\mathcal{V}\in W^{\frac{\tau}{2},2}_{\rm loc}(\Omega;\mathbb{R}^N),\qquad \mbox{ for every } 0<\tau<p'\,\left(1-\frac{N-\lambda}{p}\right).
\]
Here $L^{1,\lambda}(\Omega)$ is the usual {\it Morrey space}. In this case, the assumption on $f$ guarantees that a solution to \eqref{plaplace} can be defined in variational sense, we refer to \cite{Mi07} for more details.
\par
Some prior results are also due to J. Simon, who proved for example the global regularity
\[
\mathcal{V}\in W^{\frac{\tau}{2},2}(\mathbb{R}^N),\qquad \mbox{ for every } 0<\tau<p',
\]
for solutions $U$ in the whole $\mathbb{R}^N$ with right-hand side $f\in L^{p'}(\mathbb{R}^N)$, see \cite[Theorem 8.1]{SimonR5}.
Finally, even if it is concerned with the solution $U$ rather than the vector field $\mathcal{V}$, we wish to mention a result by Savar\'e contained in \cite{Savare}. This paper is concerned with {\it global regularity} for solutions of \eqref{plaplace} satisfying homogeneous Dirichlet boundary conditions.
In \cite[Theorem 2]{Savare}, it is shown that
\[
f\in W^{-1+\frac{\tau}{p'},p'}(\Omega) \quad \Longrightarrow \quad U\in W^{1+\frac{\tau}{p},p}(\Omega),\qquad \mbox{ for every } 0<\tau<1,
\]
when $\partial\Omega$ is Lipschitz continuous. This gives a regularity gain on the solution $U$ of
\[
2+\tau\,\left(\frac{1}{p}-\frac{1}{p'}\right),
\]
orders of differentiability, compared to the right-hand side $f$.
It is interesting to notice that by formally taking $\tau=2$ in the previous implication, this essentially gives the regularity gain of Theorem \ref{teo:local}.
\end{oss}
\subsection{About the proof}
Let us try to explain in a nutshell the key point of estimate \eqref{filipposobolev}. For the sake of simplicity, let us assume that $U$ is smooth (i.e. $U\in C^2$) and explain how to arrive at the a priori estimate \eqref{filipposobolev}. The rigourous proof is then based on a standard approximation procedure.
\par
For ease of notation, we set
\[
G(z)=\frac{|z|^p}{p},\qquad \mbox{ for }z\in\mathbb{R}^N,
\]
then $U\in W^{1,p}_{\rm loc}(\Omega)$ verifies
\[
\int \langle \nabla G(\nabla U),\nabla \varphi\rangle\,dx=\int f\,\varphi,
\]
for every compactly supported test function $\varphi$.
Uhlenbeck's result is just based on differentiating this equation in direction $x_j$ and then testing it against\footnote{Of course, this test function is not compactly supported. Actually, to make it admissible we have to multiply it by a cut-off function, see Proposition \ref{prop:sobolevuniforme}. This introduces some lower-order terms in the estimate, which are not essential at this level and would just hide the idea of the proof.} $U_{x_j}$. This yields
\begin{equation}
\label{equazionederivata}
\int \langle D^2 G(\nabla U)\,\nabla U_{x_j},\nabla U_{x_j}\rangle\,dx=\int f_{x_j}\,U_{x_j}\,dx.
\end{equation}
By using the convexity properties of $G$, it is easy to see that
\[
\int |\mathcal{V}_{x_j}|^2\,dx=\int \left|\left(|\nabla U|^\frac{p-2}{2}\,\nabla U\right)_{x_j}\right|^2\,dx\lesssim\int \langle D^2 G(\nabla U)\,\nabla U_{x_j},\nabla U_{x_j}\rangle\,dx.
\]
The main difficulty is now to estimate the right-hand side of \eqref{equazionederivata}, without using first order derivatives of $f$. A first na\"ive idea would be to integrate by parts: of course, this can not work, since this would let appear the Hessian of $U$ on which we do not have any estimate. A more clever strategy is to {\it integrate by parts in fractional sense}, i.e. use a duality-based inequality of the form
\[
\left|\int f_{x_j}\,U_{x_j}\,dx\right|\le \|f_{x_j}\|_{W^{s-1,p'}}\,\|U_{x_j}\|_{W^{1-s,p}},
\]
where $W^{s-1,p'}$ is just the topological dual of $W^{1-s,p}$.
The main point is then to prove that
\[
``\mbox{\it taking a fractional derivative of negative order of $f_{x_j}$}
\]
\[
\mbox{\it gives a fractional derivative of positive order\,}''
\]
i.e. we use that
\begin{equation}
\label{voto}
\|f_{x_j}\|_{W^{s-1,p'}}\lesssim \|f\|_{W^{s,p'}},
\end{equation}
see Theorem \ref{teo:filipponecas} below.
\par
In order to conclude, we still have to control the term containing fractional derivatives of $U_{x_j}$. This can be absorbed in the left-hand side, once we notice that $U_{x_j}$ is the composition of $\mathcal{V}$ of with a H\"older function. More precisely, we have
\[
U_{x_j}\simeq |\mathcal{V}_j|^\frac{2}{p},
\]
thus if $1-s<2/p$ we get
\[
\begin{split}
\|U_{x_j}\|_{W^{1-s,p}}^p&\lesssim \sup_{|h|>0}\int \frac{|U_{x_j}(\cdot+h)-U_{x_j}|^p}{|h|^2}\,dx \\
&\lesssim \sup_{|h|>0}\int \frac{|\mathcal{V}_j(\cdot+h)-\mathcal{V}_{j}|^2}{|h|^2}\,dx \simeq \int \left|\nabla \mathcal{V}_j\right|^2\,dx\lesssim\int \left| \mathcal{V}_{x_j}\right|^2\,dx.
\end{split}
\]
This would permit to obtain the desired estimate on $\mathcal{V}$, under the standing assumption on $f$.
\begin{oss}
Actually, the genesis of Theorem \ref{teo:local} is somehow different from the
above sketched proof. Indeed, the fact that fractional Sobolev regularity of $f$ should be enough to obtain $\mathcal{V}\in W^{1,2}$ appeared as natural in the framework of the {\it regularity via duality} strategy presented in \cite{LNRegDual}. This is a general strategy, first used in the much harder context of variational methods for the incompressible Euler equation by Brenier in \cite{br}, and then re-applied to Mean-Field Games (see for instance \cite{ProSan} for a simple case where this strategy is easy to understand). This strategy allows to prove estimates on the incremental ratios
\[
\frac{u(\cdot+h)-u}{h},
\]
of the solutions $u$ of convex variational problems by using the non-optimal function $u(\cdot+h)$ in the corresponding primal-dual optimality conditions.
\end{oss}
In our main result above we make the restriction $p>2$. For completeness, let us comment on the sub-quadratic case.
\begin{oss}[The case $1<p<2$] The sub-quadratic case is simpler, indeed we already know that in this case
\[
f\in L^{p'}_{\rm loc}(\Omega)\quad \Longrightarrow \quad \mathcal{V}\in W^{1,2}_{\rm loc}(\Omega;\mathbb{R}^N)\quad \Longrightarrow\quad \nabla U\in W^{1,p}_{\rm loc}(\Omega;\mathbb{R}^N),
\]
see for example\footnote{In \cite{de} as well the result is stated directly for $\nabla U$. However, it is easily seen that the very same proof leads to the stronger result for $\mathcal{V}$.} \cite[Theorem]{de} by de Thelin.
We can have an idea of the proof of this result
by still following the guidelines sketched above. We treat the left-hand side of \eqref{equazionederivata} as before, while on the right-hand side one now performs an integration by parts and uses H\"older's inequality. These yield
\[
\left|\int f_{x_j}\,U_{x_j}\,dx\right|\le \|f\|_{L^{p'}}\,\|U_{x_j\,x_j}\|_{L^p}\qquad \mbox{ and }\qquad \|U_{x_j\,x_j}\|_{L^p}\lesssim \|\nabla \mathcal{V}\|_{L^2}\,\|\nabla U\|_{L^p}^\frac{2-p}{2},
\]
and the term containing the gradient of $\nabla \mathcal{V}$ can be absorbed in the left-hand side.
Observe that
\[
p'\searrow 2 \quad \mbox{ as }\quad p\nearrow 2,
\]
thus again the assumption on $f$ is consistent with the case of the Laplacian.
\end{oss}
\subsection{Plan of the paper}
We set the notation and recall the basic facts on functional spaces in Section \ref{sec:2}. Here, the important point is Theorem \ref{teo:filipponecas}, which proves inequality \eqref{voto}. In Section \ref{sec:3} we consider a regularization of equation \eqref{plaplace} and prove a Sobolev estimate, independent of the regularization parameter (Proposition \ref{prop:sobolevuniforme}). Then in Section \ref{sec:4} we show how to take the estimate to the limit and achieve the proof of Theorem \ref{teo:local}. We show with an example that our assumption is essentially sharp: this is Section \ref{sec:5}. The paper closes with an appendix containing some technical tools needed for the proof of Theorem \ref{teo:filipponecas}.
\begin{ack}
Both authors have been supported by the {\it Agence Nationale de la Recherche}, through the project ANR-12-BS01-0014-01 {\sc Geometrya}.
The first author is a member of the {\it Gruppo Nazionale per l'Analisi Matematica, la Probabilit\`a
e le loro Applicazioni} (GNAMPA) of the {\it Istituto Nazionale di Alta Matematica} (INdAM).
\end{ack}
\section{Preliminaries}
\label{sec:2}
\subsection{Notation}
For a measurable function $\psi:\mathbb{R}^N\to \mathbb{R}^k$ and a vector $h\in\mathbb{R}^N$, we define
\[
\psi_h(x)=\psi(x+h),\qquad \delta_h \psi(x)=\psi_h(x)-\psi(x),
\]
and
\[
\delta_h^2 \psi(x)=\delta_h(\delta_h \psi(x))=\psi_{2h}(x)+\psi(x)-2\,\psi_h(x).
\]
We consider the two vector-valued functions
\[
\nabla G(z)=|z|^{p-2}\,z \qquad \mbox{ and }\qquad V(z)=|z|^\frac{p-2}{2}\,z,\qquad \mbox{ for }z\in\mathbb{R}^N.
\]
The following inequalities are well-known, we omit the proof.
\begin{lemma}
Let $p>2$, for every $z,w\in\mathbb{R}^N$ we have
\begin{equation}
\label{1}
|z-w|\le C_1\,|V(z)-V(w)|^\frac{2}{p},
\end{equation}
\begin{equation}
\label{3}
\Big|V(z)-V(w)\Big|\le C_2\, \left(|z|^\frac{p-2}{2}+|w|^\frac{p-2}{2}\right) |z-w|,
\end{equation}
for some $C_1=C_1(p)>0$ and $C_2=C_2(p)>0$.
\end{lemma}
\subsection{Functional spaces}
We recall the definition of some fractional Sobolev spaces needed in the sequel.
Let $1\le q<\infty$ and let $\psi\in L^q(\mathbb{R}^N)$, for $0<\beta\le 1$ we set
\[
[\psi]_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}:=\sup_{|h|>0} \left\|\frac{\delta_h \psi}{|h|^{\beta}}\right\|_{L^q(\mathbb{R}^N)},
\]
and for $0<\beta<2$
\[
[\psi]_{\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)}:=\sup_{|h|>0} \left\|\frac{\delta_h^2 \psi}{|h|^{\beta}}\right\|_{L^q(\mathbb{R}^N)}.
\]
We then introduce the two Besov-type spaces
\[
\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)=\left\{\psi\in L^q(\mathbb{R}^N)\, :\, [\psi]_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}<+\infty\right\},\qquad 0<\beta\le 1,
\]
and
\[
\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)=\left\{\psi\in L^q(\mathbb{R}^N)\, :\, [\psi]_{\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)}<+\infty\right\},\qquad 0<\beta<2.
\]
We also need the {\it Sobolev-Slobodecki\u{\i} space}
\[
W^{\beta,q}(\mathbb{R}^N)=\left\{\psi\in L^q(\mathbb{R}^N)\, :\, [\psi]_{W^{\beta,q}(\mathbb{R}^N)}<+\infty\right\},\qquad 0<\beta<1,
\]
where the seminorm $[\,\cdot\,]_{W^{\beta,q}(\mathbb{R}^N)}$ is defined by
\[
[\psi]_{W^{\beta,q}(\mathbb{R}^N)}=\left(\int_{\mathbb{R}^N}\int_{\mathbb{R}^N} \frac{|\psi(x)-\psi(y)|^q}{|x-y|^{N+\beta\,q}}\,dx\,dy\right)^\frac{1}{q}.
\]
For $1<\beta<2$, the space $W^{\beta,q}(\mathbb{R}^N)$ consists of
\[
W^{\beta,q}(\mathbb{R}^N)=\left\{\psi\in L^q(\mathbb{R}^N)\, :\, \nabla\psi\in W^{\beta-1,q}(\mathbb{R}^N)\right\}.
\]
More generally, if $\Omega\subset\mathbb{R}^N$ is an open set, the space $W^{\beta,q}(\Omega)$ is defined by
\[
W^{\beta,q}(\Omega)=\left\{\psi\in L^q(\Omega)\, :\, [\psi]_{W^{\beta,q}(\Omega)}<+\infty\right\},\qquad 0<\beta<1,
\]
and the seminorm $[\,\cdot\,]_{W^{\beta,q}(\Omega)}$ is defined accordingly. We endow this space with the norm
\[
\|\psi\|_{W^{\beta,q}(\Omega)}=\|\psi\|_{L^q(\Omega)}+[\psi]_{W^{\beta,q}(\Omega)}.
\]
For an open bounded set $\Omega\subset\mathbb{R}^N$ and $0<\beta\le 1$, the {\it homogeneous Sobolev-Slobodecki\u{\i} space} $\mathcal{D}^{\beta,q}_0(\Omega)$ is defined as the completion of $C^\infty_0(\Omega)$ with respect to the norm
\[
\psi\mapsto \|\psi\|_{\mathcal{D}^{\beta,q}_0(\Omega)}:=[\psi]_{W^{\beta,q}(\mathbb{R}^N)}\qquad \mbox{ if } 0<\beta<1,
\]
or
\[
\psi\mapsto \|\psi\|_{\mathcal{D}^{1,q}_0(\Omega)}:=\|\nabla \psi\|_{L^q(\mathbb{R}^N)}\qquad \mbox{ if } \beta=1.
\]
Finally, for $0<\beta\le 1$ the topological dual of $\mathcal{D}^{\beta,q}_0(\Omega)$ will be denoted by $\mathcal{D}^{-\beta,q'}(\Omega)$. We endow this space with the natural dual norm, defined by
\[
\|F\|_{W^{-\beta,q'}(\Omega)}=\sup\left\{|\langle F,\varphi\rangle|\, :\, \varphi\in C^\infty_0(\Omega)\ \mbox{ with }\ \|\psi\|_{\mathcal{D}^{\beta,q}_0(\Omega)}\le 1\right\},\qquad F\in \mathcal{D}^{-\beta,q'}(\Omega).
\]
\begin{oss}
\label{oss:poincare}
It is not difficult to show that for $0<\beta\le 1$ we have
\[
\mathcal{D}^{\beta,q}_0(\Omega)\hookrightarrow L^q(\Omega),
\]
thanks to Poincar\'e inequality. The latter reads as
\[
\|u\|_{L^q(\Omega)}^q\le C\,|\Omega|^\frac{\beta\,q}{N}\,\|u\|^q_{\mathcal{D}^{\beta,q}_0(\Omega)},
\]
with $C=C(N,\beta,q)>0$.
\end{oss}
\subsection{Embedding results}
In order to make the paper self-contained, we recall some functional inequalities needed in dealing with fractional Sobolev spaces.
\begin{lemma}
Let $0<\beta<1$ and $1\le q<\infty$, then we have the continuous embedding
\[
\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)\hookrightarrow \mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N).
\]
More precisley, for every $\psi\in \mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)$ we have
\[
[\psi]_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}\le \frac{C}{1-\beta}\,[\psi]_{\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)},
\]
for some constant $C=C(N,q)>0$.
\end{lemma}
\begin{proof}
We already know that
\[
[\psi]_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}\le \frac{C}{1-\beta}\,\left[[\psi]_{\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)}+\|\psi\|_{L^q(\mathbb{R}^N)}\right],
\]
se for example \cite[Lemma 2.3]{brolin}.
In particular, by using this inequality for $\psi^\lambda(x)=\psi(\lambda\,x)$ with $\lambda>0$, after a change of variable we obtain
\[
\lambda^{\beta-\frac{N}{q}}\,[\psi]_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}\le \frac{C}{1-\beta}\,\left[\lambda^{\beta-\frac{N}{q}}\,[\psi]_{\mathcal{B}^{\beta,q}_\infty(\mathbb{R}^N)}+\lambda^{-\frac{N}{q}}\,\|\psi\|_{L^q(\mathbb{R}^N)}\right].
\]
By multiplying by $\lambda^{N/q-\beta}$ and taking the limit as $\lambda$ goes to $+\infty$, we get the desired inequality.
\end{proof}
\begin{prop}
\label{prop:lostinpassing}
Let $1\le q<\infty$ and $0<\alpha<\beta< 1$. We have the continuous embedding
\[
\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)\hookrightarrow W^{\alpha,q}(\mathbb{R}^N).
\]
More precisely, for every $\psi\in \mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)$ we have
\[
[\psi]^q_{W^{\alpha,q}(\mathbb{R}^N)}\le C\,\frac{\beta}{(\beta-\alpha)\,\alpha}\, \left([\psi]_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}^q\right)^\frac{\alpha}{\beta}\,\left(\|\psi\|^q_{L^q(\mathbb{R}^N)}\right)^\frac{\beta-\alpha}{\beta},
\]
for some constant $C=C(N,q)>0$.
\end{prop}
\begin{proof}
Let us fix $h_0>0$, by appealing for example to \cite[Proposition 2.7]{brolin}, we already know that
\[
[\psi]^q_{W^{\alpha,q}(\mathbb{R}^N)}\le C\,\left(\frac{h_0^{(\beta-\alpha)\,q}}{\beta-\alpha}\, \sup_{0<|h|<h_0} \left\|\frac{\delta_h \psi}{|h|^{\beta}}\right\|_{L^q(\mathbb{R}^N)}^q+\frac{h_0^{-\alpha\,q}}{\alpha}\,\|\psi\|^q_{L^q(\mathbb{R}^N)}\right).
\]
for a constant $C$ depending on $N$ and $q$ only. If we now optimize the right-hand side in $h_0$ we get the desired conclusion.
\end{proof}
\begin{prop}
Let $1\le q<\infty$ and $0<\alpha<\beta<1$. We have the continuous embedding
\[
\mathcal{B}^{1+\beta,q}_{\infty}(\mathbb{R}^N)\hookrightarrow W^{1+\alpha,q}(\mathbb{R}^N).
\]
In particular, for every $\psi\in \mathcal{B}^{1+\beta,q}_\infty(\mathbb{R}^N)$ we have $\nabla \psi\in W^{\alpha,q}(\mathbb{R}^N)$, with the following estimates
\begin{equation}
\label{02081980}
\|\nabla \psi\|^q_{L^q(\mathbb{R}^N)}\le \frac{C}{\beta^\frac{\beta+q}{\beta+1}}\,\left(\|\psi\|^q_{L^q(\mathbb{R}^N)}\right)^\frac{\beta}{\beta+1}\, \left([\psi]^q_{\mathcal{B}^{1+\beta,q}_\infty(\mathbb{R}^N)}\right)^\frac{1}{\beta+1},
\end{equation}
for some $C=C(N,q)>0$, and
\begin{equation}
\label{02082015}
[\nabla \psi]^q_{W^{\alpha,q}(\mathbb{R}^N)}\le C\left( [\psi]^q_{\mathcal{B}^{1+\beta,q}_\infty(\mathbb{R}^N)}\right)^\frac{\alpha+1}{\beta+1}\,\left(\|\psi\|^q_{L^q(\mathbb{R}^N)}\right)^\frac{\beta-\alpha}{\beta+1},
\end{equation}
for some $C=C(N,q,\alpha,\beta)>0$, which blows up as $\alpha\nearrow \beta$, $\beta\searrow 0$ or $\beta\nearrow 1$.
\end{prop}
\begin{proof}
We already know that
\[
\|\nabla \psi\|^q_{L^q(\mathbb{R}^N)}\le C\,\|\psi\|^q_{L^q(\mathbb{R}^N)}+\frac{C}{\beta^q}\, [\psi]^q_{\mathcal{B}^{1+\beta,q}_\infty(\mathbb{R}^N)},
\]
see for example \cite[Proposition 2.4]{brolin}.
By replacing $\psi$ with the rescaled function $\psi^\lambda(x)=\psi(\lambda\,x)$ and optimizing in $\lambda$, we get \eqref{02081980}.
\par
For the second estimate \eqref{02082015}, it is sufficient to observe that
\[
[\nabla \psi]^q_{\mathcal{N}^{\beta,q}_\infty(\mathbb{R}^N)}\le \frac{C}{\beta^q\,(1-\beta)^q} [\psi]^q_{\mathcal{B}^{1+\beta,q}_\infty(\mathbb{R}^N)},
\]
again by \cite[Proposition 2.4]{brolin}.
Then by using Proposition \ref{prop:lostinpassing} for $\nabla \psi$ and \eqref{02081980}, we get the desired conclusion.
\end{proof}
\subsection{An inequality for negative norms}
As explained in the Introduction, a crucial r\^ole in the proof of Theorem \ref{teo:local} is played by the following weak generalization of the so-called {\it Ne\v{c}as' negative norm Theorem} (which corresponds to $\beta=0$, see \cite{Ne}).
\begin{teo}
\label{teo:filipponecas}
Let $0<\beta<1$ and $1<q<\infty$. Let $B\subset\mathbb{R}^N$ be an open ball, for every $f\in W^{\beta,q}(B)$ we have
\begin{equation}
\label{filipponecasSI}
\|f_{x_j}\|_{\mathcal{D}^{\beta-1,q}(B)}\le C\,[f]_{W^{\beta,q}(B)},\qquad j=1,\dots,N,
\end{equation}
for a constant $C=C(N,\beta,q)>0$.
\end{teo}
\begin{proof}
We explain the guidelines of the proof, by referring the reader to Appendix \ref{sec:A} for the missing details. We first prove
\begin{equation}
\label{filipponecas}
\|f_{x_j}\|_{\mathcal{D}^{\beta-1,q}(B)}\le C\,\left(\|f\|_{L^q(B)}+[f]_{W^{\beta,q}(B)}\right),\qquad j=1,\dots,N.
\end{equation}
This estimate says that the linear operator
\[
T_j:W^{\beta,q}(B)\to \mathcal{D}^{\beta-1,q}(B),
\]
defined by the weak $j-$th derivative is continuous. It is easy to see that $T_j$ is continuous as an operator defined on $W^{1,q}(B)$ and $L^q(B)$. More precisely, the following operators
\[
\begin{array}{clll}
T_j:&W^{1,q}(B)&\to& L^q(B)\\
&&&\\
T_j:&L^q(B)&\to& \mathcal{D}^{-1,q}(B)
\end{array}
\]
are linear and continuous. In other words, inequality \eqref{filipponecas} is true for the extremal cases $\beta=0$ and $\beta=1$. We observe that $W^{\beta,q}(B)$ is an interpolation space between $W^{1,q}(B)$ and $L^q(B)$, i.e.
\[
W^{\beta,q}(B)=\left(L^q(B),W^{1,q}(B)\right)_{\beta,q},
\]
see Definition \ref{defi:inter} for the notation. We can then obtain quite easily that $T_j$ is continuous from $W^{\beta,q}(B)$ to the interpolation space between $\mathcal{D}^{-1,q}(B)$ and $L^q(B)$, i.e.
\[
\left(\mathcal{D}^{-1,q}(B),L^q(B)\right)_{\beta,q},
\]
see Lemma \ref{lm:3}. Such a space can be computed explicitely: as one may expect, it coincides with the dual Sobolev-Slobodecki\u{\i} space $\mathcal{D}^{\beta-1,q}(B)$ (see Lemma \ref{lm:4}). This proves inequality \eqref{filipponecas}.
\par
In order to get \eqref{filipponecasSI} and conclude, it is now sufficient to use a standard scaling argument.
Let us assume for simplicity that $B$ is centered at the origin, for $f\in W^{\beta,q}(B)$ and every $\lambda>0$, we define $f_\lambda(x)=f(x/\lambda)$. This belongs to $W^{\beta,q}(\lambda\,B)$, then from \eqref{filipponecas} and the scaling properties of the norms, we get
\[
\begin{split}
\lambda^{\frac{N}{q}-\beta}\,\|f_{x_j}\|_{\mathcal{D}^{\beta-1,q}(B)}=\|(f_\lambda)_{x_j}\|_{\mathcal{D}^{\beta-1,q}(\lambda\,B)}&\le C\,\left(\|f_\lambda\|_{L^q(\lambda\,B)}+[f_\lambda]_{W^{\beta,q}(\lambda\,B)}\right)\\
&=C\,\left(\lambda^\frac{N}{q}\,\|f\|_{L^q(B)}+\lambda^{\frac{N}{q}-\beta}\,[f]_{W^{\beta,q}(B)}\right).
\end{split}
\]
If we multiply by $\lambda^{\beta-N/q}$ and then let $\lambda$ go to $0$, we finally get the desired estimate.
\end{proof}
\section{Estimates for a regularized problem}
\label{sec:3}
Let $U$ and $f$ be as in the statement of Theorem \ref{teo:local}. Let $B\Subset \Omega$ be an open ball, for every $\varepsilon>0$ we consider the problem
\begin{equation}
\label{approssimato}
\left\{\begin{array}{rcll}
-\mathrm{div}\nabla G_\varepsilon(\nabla u)&=&f_\varepsilon,&\mbox{ in } B,\\
u&=&U,&\mbox{ on }\partial B,
\end{array}
\right.
\end{equation}
where:
\begin{itemize}
\item $G_\varepsilon(z)=\dfrac{1}{p}\,(\varepsilon+|z|^2)^\frac{p}{2}$, for every $z\in\mathbb{R}^N$;
\vskip.2cm
\item $f_\varepsilon=f\ast \varrho_\varepsilon$ and $\{\varrho_\varepsilon\}_{\varepsilon>0}$ is a family of standard compactly supported $C^\infty$ mollifiers.
\vskip.2cm
\end{itemize}
Problem \eqref{approssimato} admits a unique solution $u_\varepsilon\in W^{1,p}(B)$, which is locally smooth in $B$ by standard elliptic regularity.
\begin{prop}[Uniform energy estimate]
With the notation above, we have
\begin{equation}
\label{epsilon}
\begin{split}
\int_B |\nabla u_\varepsilon|^p\,dx&\le C\,\varepsilon^\frac{p-1}{2}\,\int_B|\nabla U|\,dx+C\,\int_B |\nabla U|^{p}\,dx+C\,|B|^\frac{p'}{N}\,\|f_\varepsilon\|_{L^{p'}(B)}^{p'},
\end{split}
\end{equation}
for some $C=C(N,p)>0$.
\end{prop}
\begin{proof}
The proof is standard, we include it for completeness. We take the weak formulation of \eqref{approssimato}
\[
\int \langle \nabla G_\varepsilon(\nabla u_\varepsilon),\nabla\varphi\rangle\,dx=\int f_\varepsilon\,\varphi\,dx,\qquad \mbox{ for every }\varphi\in W^{1,p}_0(B),
\]
and insert the test function $\varphi=u_\varepsilon-U$. This gives
\[
\begin{split}
\int_B \langle \nabla G_\varepsilon(\nabla u_\varepsilon),\nabla u_\varepsilon\rangle\,dx&=\int_B \langle \nabla G_\varepsilon(\nabla u_\varepsilon),\nabla U\rangle\,dx+\int_B f_\varepsilon\,(u_\varepsilon-U)\,dx\\
&\le \int |\nabla G_\varepsilon(\nabla u_\varepsilon)|\,|\nabla U|\,dx+\|f_\varepsilon\|_{L^{p'}(B)}\,\|u_\varepsilon-U\|_{L^p(B)}.
\end{split}
\]
We then observe that
\[
\langle\nabla G_\varepsilon(z),z\rangle\ge |z|^p,\qquad z\in\mathbb{R}^N,
\]
and
\[
|\nabla G_\varepsilon(z)|\le C\, \left(\varepsilon^\frac{p-2}{2}\,|z|+|z|^{p-1}\right)\le C\,\varepsilon^\frac{p-1}{2}+2\,C\,|z|^{p-1},\qquad z\in\mathbb{R}^N,
\]
for some $C=C(p)>0$. By using these inequalities, we get
\[
\begin{split}
\int_B |\nabla u_\varepsilon|^p\,dx&\le C\,\varepsilon^\frac{p-1}{2}\int_B|\nabla U|\,dx+C\int_B |\nabla u_\varepsilon|^{p-1}\,|\nabla U|\,dx+\|f_\varepsilon\|_{L^{p'}(B)}\,\|u_\varepsilon-U\|_{L^p(B)}\\
&\le C\,\varepsilon^\frac{p-1}{2}\,\int_B|\nabla U|\,dx+C\,\tau\,\int_B |\nabla u_\varepsilon|^{p}\,dx\\
&+C\,\tau^{-\frac{1}{p-1}}\,\int_B |\nabla U|^{p}\,dx+C\,|B|^\frac{1}{N}\,\|f_\varepsilon\|_{L^{p'}(B)}\,\|\nabla u_\varepsilon-\nabla U\|_{L^p(B)},
\end{split}
\]
where we used Young's inequality in the second term and Poincar\'e's inequality in the last one. By taking $\tau>0$ sufficiently small, we can then obtain
\[
\begin{split}
\int_B |\nabla u_\varepsilon|^p\,dx
&\le C\,\varepsilon^\frac{p-1}{2}\,\int_B|\nabla U|\,dx+C\,\int_B |\nabla U|^{p}\,dx\\
&+C\,|B|^\frac{1}{N}\,\|f_\varepsilon\|_{L^{p'}(B)}\,\|\nabla u_\varepsilon-\nabla U\|_{L^p(B)}.
\end{split}
\]
We can now use the triangle inequality on the last term and conclude with a further application of Young's inequality, in order to absorb the term containing $\nabla u_\varepsilon$. We leave the details to the reader.
\end{proof}
The proof of Theorem \ref{teo:local} is crucially based on the following
\begin{prop}[Uniform Sobolev estimate]
\label{prop:sobolevuniforme}
We set
\[
\mathcal{V}_\varepsilon:=V(\nabla u_\varepsilon)=|\nabla u_\varepsilon|^\frac{p-2}{2}\,\nabla u_\varepsilon.
\]
Let $(p-2)/p<s\le 1$, for every pair of concentric balls $B_r\Subset B_R\Subset B$ and every $j=1,\dots,N$ we have
\begin{equation}
\label{Vepsilon}
\begin{split}
\int_{B_r} \left|\left(\mathcal{V}_\varepsilon\right)_{x_j}\right|^2\,dx&\le \frac{C}{(R-r)^2}\int_{B_R} (\varepsilon+|\nabla u_\varepsilon|^2)^\frac{p}{2}\,dx+C\left(R^{\left(s-\frac{p-2}{p}\right)}\,[f_\varepsilon]_{W^{s,p'}(B_R)}\right)^{p'},
\end{split}
\end{equation}
for a constant $C=C(N,p,s)>0$ which blows up as $s\searrow (p-2)/p$.
\end{prop}
\begin{proof}
In what follows, for notational simplicity we omit to indicate the dependence on $\varepsilon>0$ and simply write
\[
G,\quad u,\quad \mathcal{V}\quad \mbox{ and }\quad f.
\]
For $j\in\{1,\dots,N\}$, in the weak formulation of \eqref{approssimato} we insert a test function of the form $\varphi_{x_j}$ for $\varphi\in C^\infty_0(B)$. After an integration by parts, we obtain
\[
\int \langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla \varphi\rangle\,dx=\int f_{x_j}\,\varphi\,dx.
\]
By density, this relation remains true for $\varphi\in W^{1,p}$ with compact support in $B$. We then insert the test function
\[
\varphi=\zeta^2\, u_{x_j},
\]
with $\zeta\in C^\infty_0(B_R)$ a standard cut-off function such that
\begin{equation}
\label{zzeta}
0\le \zeta\le 1,\qquad \zeta\equiv 1\ \mbox{ in } B_r,\qquad |\nabla \zeta|\le \frac{C}{R-r}.
\end{equation}
Thus we obtain
\begin{equation}
\label{apparecchio}
\begin{split}
\int \langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla u_{x_j}\rangle\,\zeta^2\,dx&=-2\,\int\langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla \zeta\rangle\,u_{x_j}\,\zeta\,dx\\
&+ \int f_{x_j}\,u_{x_j}\,\zeta^2\,dx.
\end{split}
\end{equation}
We first observe that the left-hand side is positive, since $G$ is convex. As for the right-hand side, we have
\[
\begin{split}
\int \langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla \zeta\rangle\,u_{x_j}\,\zeta\,dx
&\le \int \left|\langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla \zeta\rangle\right|\,|u_{x_j}|\,\zeta\,dx\\
&\le \int \sqrt{\langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla u_{x_j}\rangle}\\
&\times\sqrt{\langle D^2 G(\nabla u)\,\nabla \zeta,\nabla \zeta\rangle} \,|u_{x_j}|\,\zeta\,dx
\end{split},
\]
thanks to Cauchy-Schwartz inequality. By using Young's inequality in a standard fashion, from \eqref{apparecchio} and the previous inequality we can obtain
\[
\begin{split}
\int \langle D^2 G(\nabla u)\,\nabla u_{x_j},\nabla u_{x_j}\rangle\,\zeta^2\,dx&\le 4\,\int\langle D^2 G(\nabla u)\,\nabla\zeta,\nabla \zeta\rangle\,|u_{x_j}|^2\,dx\\
&+ 2\,\int f_{x_j}\,u_{x_j}\,\zeta^2\,dx.
\end{split}
\]
We now use that
\[
|z|^{p-2}\,|\xi|^2\le \langle D^2 G(z)\,\xi,\xi\rangle\le (p-1)\,(\varepsilon+|z|^2)^\frac{p-2}{2}\,|\xi|^2,\qquad z,\xi\in\mathbb{R}^N,
\]
thus with simple manipulations we get
\[
\begin{split}
\int |\nabla u|^{p-2}\,|\nabla u_{x_j}|^2\,\zeta^2\,dx&\le \frac{4\,(p-1)}{(R-r)^2}\int_{B_R} (\varepsilon+|\nabla u|^2)^\frac{p}{2}\,dx+ 2\,\int f_{x_j}\,u_{x_j}\,\zeta^2\,dx.
\end{split}
\]
We observe that
\begin{equation}
\label{Vona}
|\nabla u|^{p-2}\,|\nabla u_{x_j}|^2=\frac{4}{p^2}\,\left|\left(|\nabla u|^\frac{p-2}{2}\,\nabla u\right)_{x_j}\right|^2=\frac{4}{p^2}\,\left|\mathcal{V}_{x_j}\right|^2,
\end{equation}
thus we have
\begin{equation}
\label{rullo!}
\int \left|\mathcal{V}_{x_j}\right|^2\,\zeta^2\,dx\le \frac{(p-1)\,p^2}{(R-r)^2}\int_{B_R} (\varepsilon+|\nabla u|^2)^\frac{p}{2}\,dx+\frac{p^2}{2}\,\int f_{x_j}\,u_{x_j}\,\zeta^2\,dx.
\end{equation}
We are left with the estimate of the term containing $f_{x_j}$. By definition of dual norm, for\footnote{We exclude here the case $s=1$, since this is the easy case. It would be sufficient to use H\"older's inequality with exponents $p$ and $p'$ to conclude.} $(p-2)/p<s<1$ we have
\begin{equation}
\label{emmo?}
\int f_{x_j}\,u_{x_j}\,\zeta^2\,dx\le \|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,[u_{x_j}\,\zeta^2]_{W^{1-s,p}(\mathbb{R}^N)}.
\end{equation}
Observe that by assumption $0<1-s<2/p$, thus by Proposition \ref{prop:lostinpassing} with $\alpha=1-s$ and $\beta=2/p$, we have
\begin{equation}
\label{bleah}
\begin{split}
\left[u_{x_j}\,\zeta^2\right]^p_{W^{1-s,p}(\mathbb{R}^N)}&\le \frac{C}{\left(s-\dfrac{p-2}{p}\right)\,(1-s)}\, \left(\left[u_{x_j}\,\zeta^2\right]_{\mathcal{N}^{\frac{2}{p},p}_\infty(\mathbb{R}^N)}^p\right)^{(1-s)\frac{p}{2}}\,\left(\|u_{x_j}\|^p_{L^p(B_R)}\right)^{1-\frac{p}{2}(1-s)},
\end{split}
\end{equation}
In order to estimate the norm of $u_{x_j}\,\zeta^2$, we recall that
\[
\left[u_{x_j}\,\zeta^2\right]_{\mathcal{N}^{\frac{2}{p},p}_\infty(\mathbb{R}^N)}^p=\sup_{|h|>0} \int_{\mathbb{R}^N} \frac{|\delta_h(u_{x_j}\,\zeta^2)|^p}{|h|^2}\,dx.
\]
Then we observe that from \eqref{1}
\[
|a-b|^p\le C\, \left||a|^\frac{p-2}{2}\,a-|b|^\frac{p-2}{2}\,b\right|^{2},\qquad a,b\in\mathbb{R}.
\]
Thus we obtain
\[
\begin{split}
\int_{\mathbb{R}^N} &\frac{|\delta_h(u_{x_j}\,\zeta^2)|^p}{|h|^2}\,dx\le C\, \int_{\mathbb{R}^N} \frac{\left|\delta_h\left(|u_{x_j}|^\frac{p-2}{2}\,u_{x_j}\,\zeta^p\right)\right|^2}{|h|^2}\,dx\le C\,\int_{\mathbb{R}^N} \left|\nabla \left(|u_{x_j}|^\frac{p-2}{2}\,u_{x_j}\,\zeta^p\right)\right|^2\,dx,
\end{split}
\]
where the second inequality comes from the classical characterization of $W^{1,2}$ in terms of finite differences.
By recalling the properties \eqref{zzeta} of $\zeta$, with simple manipulations we thus obtain
\[
\left[u_{x_j}\,\zeta^2\right]_{\mathcal{N}^{\frac{2}{p},p}_\infty(\mathbb{R}^N)}^p\le C\, \int \left|\nabla \left(|u_{x_j}|^\frac{p-2}{2}\,u_{x_j}\right)\right|^2\,\zeta^2\,dx+\frac{C}{(R-r)^2}\,\int_{B_R} |u_{x_j}|^p\,dx,
\]
for a constant $C=C(N,p)>0$. Observe that
\[
\left|\nabla \left(|u_{x_j}|^\frac{p-2}{2}\,u_{x_j}\right)\right|^2=\frac{p^2}{4}\,|u_{x_j}|^{p-2}\,|\nabla u_{x_j}|^2\le \frac{p^2}{4}\,|\nabla u|^{p-2}\,|\nabla u_{x_j}|^2=|\mathcal{V}_{x_j}|^2,
\]
thanks to \eqref{Vona}. This yields
\[
\left[u_{x_j}\,\zeta^2\right]_{\mathcal{N}^{\frac{2}{p},p}_\infty(\mathbb{R}^N)}^p\le C\, \int |\mathcal{V}_{x_j}|^2\,\zeta^2\,dx+\frac{C}{(R-r)^2}\,\int_{B_R} |\nabla u|^p\,dx.
\]
By inserting this estimate in \eqref{bleah}, from \eqref{emmo?} we get
\[
\begin{split}
\left|\int f_{x_j}\,u_{x_j}\,\zeta^2\,dx\right|&\le C\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\left(\|\mathcal{V}_{x_j}\,\zeta\|^\frac{2}{p}_{L^2(B_R)}+\frac{C}{(R-r)^\frac{2}{p}}\,\|\nabla u\|_{L^p(B_R)}\right)^{(1-s)\frac{p}{2}}\\
&\times \left(\|\nabla u\|_{L^p(B_R)}\right)^{1-\frac{p}{2}(1-s)},
\end{split}
\]
for a constant $C=C(N,p,s)>0$, which blows-up as $s\searrow (p-2)/p$. We can still manipulate a bit the previous estimate and obtain
\[
\begin{split}
\left|\int f_{x_j}\,u_{x_j}\,\zeta^2\,dx\right|&\le C\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\left(\|\mathcal{V}_{x_j}\,\zeta\|^{1-s}_{L^2(B_R)}+\frac{C}{(R-r)^{1-s}}\,\|\nabla u\|_{L^p(B_R)}^{(1-s)\frac{p}{2}}\right)\\
&\times \left(\|\nabla u\|_{L^p(B_R)}\right)^{1-\frac{p}{2}(1-s)}\\
&\le C\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\|\mathcal{V}_{x_j}\,\zeta\|^{1-s}_{L^2(B_R)}\,\left(\|\nabla u\|_{L^p(B_R)}\right)^{1-\frac{p}{2}(1-s)}\\
&+\frac{C}{(R-r)^{1-s}}\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\|\nabla u\|_{L^p(B_R)},
\end{split}
\]
for a different constant $C>0$, still depending on $N,p$ and $s$ only. We now go back to \eqref{rullo!} and use the previous estimate. This gives
\[
\begin{split}
\int \left|\mathcal{V}_{x_j}\right|^2\,\zeta^2\,dx&\le \frac{C}{(R-r)^2}\int_{B_R} (\varepsilon+|\nabla u|^2)^\frac{p}{2}\,dx\\
&+C\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\|\mathcal{V}_{x_j}\,\zeta\|^{1-s}_{L^2(B_R)}\,\left(\|\nabla u\|_{L^p(B_R)}\right)^{1-\frac{p}{2}(1-s)}\\
&+\frac{C}{(R-r)^{1-s}}\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\|\nabla u\|_{L^p(B_R)}.
\end{split}
\]
We need to absorb the higher order term containing $\mathcal{V}$ in the right-hand side.
For this, we use Young's inequality with exponents
\[
p',\quad \frac{2}{1-s} \quad \mbox{ and }\quad \frac{2\,p}{2-p\,(1-s)},
\]
so to get
\[
\begin{split}
\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}&\|\mathcal{V}_{x_j}\,\zeta\|^{1-s}_{L^2(B_R)}\,\left(\|\nabla u\|_{L^p(B_R)}\right)^{1-\frac{p}{2}(1-s)}\\
&\le C\,\tau^{-\frac{(1-s)}{2}\,p'}\,(R-r)^{\left(s-\frac{p-2}{p}\right)\,p'}\,\|f_{x_j}\|^{p'}_{\mathcal{D}^{s-1,p'}(B_R)}+\tau\,\|\mathcal{V}_{x_j}\,\zeta\|^2_{L^2(B_R)}\\
&+\frac{C}{(R-r)^2}\,\|\nabla u\|_{L^p(B_R)}^p,
\end{split}
\]
which yields (by taking $\tau>0$ small enough)
\[
\begin{split}
\int \left|\mathcal{V}_{x_j}\right|^2\,\zeta^2\,dx&\le \frac{C}{(R-r)^2}\int_{B_R} (\varepsilon+|\nabla u|^2)^\frac{p}{2}\,dx+C\,(R-r)^{\left(s-\frac{p-2}{p}\right)\,p'}\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}^{p'}\\
&+\frac{C}{(R-r)^{1-s}}\,\|f_{x_j}\|_{\mathcal{D}^{s-1,p'}(B_R)}\,\|\nabla u\|_{L^p(B_R)}.
\end{split}
\]
We now apply Young's inequality once more and Theorem \ref{teo:filipponecas}, so to obtain \eqref{epsilon}.
\end{proof}
\section{Proof of Theorem \ref{teo:local}}
\label{sec:4}
Let $U\in W^{1,p}_{\rm loc}(\Omega)$ be a local weak solution in $\Omega$ of \eqref{plaplace}. We fix a ball $B_R\Subset\Omega$ and take a pair of concentric ball $B$ and $\widetilde B$ such that $B_R\Subset B\Subset \widetilde B\Subset\Omega$. There exists $\varepsilon_0>0$ such that
\begin{equation}
\label{fepsilon}
\|f_\varepsilon\|_{L^{p'}(B)}+[f_\varepsilon]_{W^{s,p'}(B)}\le \|f\|_{L^{p'}(\widetilde B)}+[f]_{W^{s,p'}(\widetilde B)}<+\infty,\qquad \mbox{ for every }0<\varepsilon<\varepsilon_0.
\end{equation}
We consider for every $0<\varepsilon<\varepsilon_0$ the solution $u_\varepsilon$ of problem \eqref{approssimato} in the ball $B$. By \eqref{epsilon} and \eqref{fepsilon} we know that $\{u_\varepsilon\}_{0<\varepsilon<\varepsilon_0}$ is bounded in $W^{1,p}(B)$, thus by Rellich-Kondra\v{s}ov Theorem we can extract a sequence $\{\varepsilon_k\}_{k\in\mathbb{N}}$ converging to $0$, such that
\[
\lim_{k\to\infty} \|u_{\varepsilon_k}-u\|_{L^p(B)}=0,
\]
for some $u\in W^{1,p}(B)$. Since $u_{\varepsilon_k}$ is the unique solution of
\[
\min\left\{\int_B G_{\varepsilon_k}(\nabla \varphi)\,dx-\int_B f_{\varepsilon_k}\,\varphi\,dx\, :\, \varphi-U\in W^{1,p}_0(B)\right\},
\]
by a standard $\Gamma-$convergence argument we can easily show that $u=U$, i.e. the limit function coincides with our local weak solution $U$.
\par
From Proposition \ref{prop:sobolevuniforme} and boundedness of $\{u_{\varepsilon_k}\}_{k\in\mathbb{N}}$, we know that $\{\mathcal{V}_{\varepsilon_k}\}_{k\in\mathbb{N}}$ is bounded in $W^{1,2}(B_r;\mathbb{R}^N)$, for every $B_r\Subset B$. Up to passing a subsequence, we can infer convergence to some vector field \(\mathcal{Z}\in W^{1,2}(B_r;\mathbb{R}^N)\), weakly in \(W^{1,2}(B_r;\mathbb{R}^N)\) and strongly in \(L^{2}(B_r;\mathbb{R}^N)\). In particular, this is a Cauchy sequence in $L^2(B_r)$ and by using the elementary inequality \eqref{1},
we obtain that $\{u_{\varepsilon_k}\}_{k\in\mathbb{N}}$ as well is a Cauchy sequence in $W^{1,p}(B_r)$. Thus we obtain
\[
\lim_{k\to\infty}\left\|\nabla u_{\varepsilon_k}-\nabla U\right\|_{L^p(B_r)}=0.
\]
We need to show that $\mathcal{Z}=|\nabla U|^{(p-2)/2}\,U$. We use the elementary inequality \eqref{3}, this yields
\[
\begin{split}
\int_{B_r} \left||\nabla u_{\varepsilon_k}|^{\frac{p-2}{2}}\,\nabla u_{\varepsilon_k}-|\nabla U|^{\frac{p-2}{2}}\,\nabla U\right|^2\,dx&\le C\, \int_{B_r} \left(|\nabla u_{\varepsilon_k}|^{\frac{p-2}{2}}+|\nabla U|^{\frac{p-2}{2}}\right)^2\,|\nabla u^{\varepsilon_k}-\nabla U|^2\,dx\\
&\le C\,\left( \int_{B_r} \left(|\nabla u_{\varepsilon_k}|^{\frac{p-2}{2}}+|\nabla U|^{\frac{p-2}{2}}\right)^\frac{2\,p}{p-2}\,dx\right)^\frac{p-2}{p}\\
&\times \left(\int_{B_r} |\nabla u_{\varepsilon_k}-\nabla U|^p\,dx\right)^\frac{2}{p}.
\end{split}
\]
By using the strong convergence of the gradients proved above, this gives that $\mathcal{Z}=|\nabla U|^\frac{p-2}{2}\,\nabla U$ and it belongs to $W^{1,2}(B_r;\mathbb{R}^N)$. By arbitrariness of the ball $B_r\Subset B$ in the above discussion, we can now take the ball $B_R$ fixed at the beginning and pass to the limit in \eqref{Vepsilon}, so to obtain the desired estimate \eqref{filipposobolev}.
\par
Finally, the fact that $\nabla U\in W^{\tau,p}_{\rm loc}(\Omega;\mathbb{R}^N)$ follows in standard way from the elementary inequality \eqref{1}. We leave the details to the reader.
\section{An example}
\label{sec:5}
We now show with an explicit example that the assumption on $f$ in Theorem \ref{teo:local} is essentially sharp.
\par
Let us take $U(x)=|x|^{-\alpha}$, which belongs to $W^{1,p}_{\rm loc}(\mathbb{R}^N)$ if and only if
\[
\alpha<\frac{N}{p}-1.
\]
Then we can compute
\[
|\nabla U|^{p-2}\,\nabla U\simeq |x|^{-(\alpha+1)\,(p-1)-1}\,x,\qquad
\mathrm{div}\left(|\nabla U|^{p-2}\,\nabla U\right)\simeq |x|^{-(\alpha+1)\,(p-1)-1},
\]
and
\[
|\nabla U|^\frac{p-2}{2}\,\nabla U\simeq |x|^{-(\alpha+1)\,\frac{p}{2}-1}\,x.
\]
We observe that
\[
|\nabla U|^\frac{p-2}{2}\,\nabla U\in W^{1,2}_{\rm loc}(\mathbb{R}^N)\quad \Longleftrightarrow\quad \alpha<\frac{N-2}{p}-1=:\widetilde\alpha.
\]
On the other hand the function $f(x)=|x|^{-(\alpha+1)\,(p-1)-1}$ belongs to $W^{s,p'}_{\rm loc}(\mathbb{R}^N)$ if
\[
(\alpha+1)\,(p-1)+1<\frac{N-s\,p'}{p'}\qquad \mbox{ i.\,e. if }\qquad \alpha<\frac{N}{p}-\frac{s+1}{p-1}-1=:\alpha_s.
\]
If we take $s<(p-2)/p$, we have
\[
\alpha_s>\frac{N-2}{p}-1=\widetilde\alpha,
\]
thus for every such $s$, if we take $\alpha=(N-2)/p-1$ we get
\[
-\Delta_p U=f\in W^{s,p'}_{\rm loc}(\mathbb{R}^N) \qquad \mbox{ and }\qquad |\nabla U|^\frac{p-2}{2}\, \nabla U\not\in W^{1,2}_{\rm loc}(\mathbb{R}^N).
\]
This shows that in Theorem \ref{teo:local} the differentiability exponent $s$ of $f$ can not go below $(p-2)/p$.
|
\section{Introduction \label{sec:Introduction}}
Traditional algorithms for machine learning often rely on the tacit
assumption of \emph{independent and identical distributed} (i.i.d.)
data. Under this assumption, the training samples on which an algorithm
is trained, and the test data on which the same algorithm is evaluated,
are taken to come from the same distribution. This assumption is critical for learning as it allows for a straightforward generalization of conclusions drawn from the training data to the test data.
As long as the data can be modelled using this assumption, it is possible to provide statistical guarantees on the performance of traditional machine learning algorithms \citep{Vapnik1998,Bishop2006}.
However, real-world
data rarely satisfy this assumption. Data collected from different
users, in different locations, or at different times do not fit the
requirement of coming from identical distributions. Not surprisingly, traditional machine learning algorithms proved to
perform poorly when trained and tested on data that are not identically
distributed.
Data that do not comply with the \iid assumption may be modelled according to the more realistic assumption of \emph{covariate shift}. The covariate shift assumption states that the marginal \rev{I put "feature" here among brackets, so it clarifies which distribution we are dealing with and it makes it clear that the term is optional and it will be dropped later.}{(feature)} distributions of training and test data are different, while the conditional distribution of the labels given the data is the same for both the training and test data \citep{Sugiyama2012}. \emph{Covariate shift adaptation} (CSA) aims at enabling supervised learning by compensating for the difference in the marginal distributions. \\
A common approach to CSA is through \emph{representation learning} (RL): given training and test data
exhibiting covariate shift, we try to learn a new representation of
the data that is not affected by covariate shift.
A recently-proposed algorithm for RL is sparse filtering (SF) \citep{Ngiam2011}. SF learns maximally sparse representations of the data in an unsupervised way, aiming at learning a new useful distribution while overlooking the problem of explicitly modelling the noiseless marginal distribution that generated of the data.
SF was shown to be an efficient and scalable algorithm, and it has been successfully applied to various real-world applications, leading to state-of-the-art performances even when paired with simple classifiers \citep{Ngiam2011}.\\
So far, however, SF has never been used in the context of covariate shift. Here, we suggest that the disregard of SF for the problem of modelling the original marginal distribution of the data makes it a potential candidate for dealing with data affected by covariate shift. In addressing only the problem of shaping a useful distribution for the learned representations, SF may learn new representations in a way that is insensitive to covariate shift. SF may then offer an opportunity not only to avoid the computationally hard problem of learning the marginal distribution of the data, but also a way to circumvent the covariate shift problem. \\
In this paper we explore the potentialities and the limitations of using SF-like algorithms for CSA. We start by asking: \emph{can SF perform CSA?} In answering this question we will analyse the conditions required to perform CSA, and prove through a theoretical analysis that SF can indeed be used for CSA if the conditional distribution of the labels is explained by a cosine metric. Given this severe limitation, we move on to ask: \emph{can we devise a SF-based variant that can perform CSA under looser and more realistic conditions than SF?} We then develop an alternative algorithm based on SF, which retains the main advantages of SF while being able to perform CSA when the conditional distribution of the labels is periodic. We provide a theoretical analysis of this new algorithm and conduct experiments to demonstrate the usefulness of this SF-based algorithm on synthetic and real-world data.\\
The paper is organized as follows.
Section \ref{sec:BG} formalizes the problem of learning under covariate shift and briefly reviews previous work on CSA and SF.
Section \ref{sec:CSA-via-SF} presents a theoretical analysis of the conditions under which SF can perform CSA.
Section \ref{sec:PSF} proposes a new SF-based algorithm to perform CSA.
Section \ref{sec:CSA-via-PSF} provides a theoretical analysis of the conditions under which the new algorithm is guaranteed to perform CSA.
Section \ref{sec:Experiments} validates our theoretical results via an empirical study. Finally, Section \ref{sec:Conclusions} draws conclusions about our study.
\section{Background and Related Work \label{sec:BG}}
This section presents the concepts and the notation that will be used later in our theoretical study and in our experimental validations. Section \ref{sec:BG_Notation} introduces the formalism used to refer to the data and to the learning problem; Section \ref{sec:BG_CS} defines the problem of learning under covariate shift; Section \ref{sec:BG_CSA} discusses CSA and reviews related work; Section \ref{sec:BG_FDL} presents the SF algorithm and related work.
\revII{Corrected}{For clarity,} Table \ref{tab:Notation} summarizes the conventions and the notation adopted throughout this paper.
\begin{table*}
\caption{Summary of notation. \label{tab:Notation}}
\begin{centering}
\begin{tabular}{cl}
\hline
$a, b, c...$ & Generic scalars \tabularnewline
$\APPLIEDFUNCTION{f}{\emptyarg}, \APPLIEDFUNCTION{g}{\emptyarg}, \APPLIEDFUNCTION{h}{\emptyarg}...$ & Generic functions \tabularnewline
$\DATAVECTOR{a}{}, \DATAVECTOR{b}{}, \DATAVECTOR{c}{}...$ & Generic vectors \tabularnewline
$\DATAVECTORelem{a}{i}$ & $i^{th}$ element of the the vector $a$ \tabularnewline
$\DATAMATRIX{A}{}, \DATAMATRIX{B}{}, \DATAMATRIX{C}{}...$ & Generic matrices \tabularnewline
$\DATAMATRIXrow{a}{i}$ & $i^{th}$ column of the the matrix $\DATAMATRIX{A}{}$ (sample) \tabularnewline
$\DATAMATRIXcol{a}{j}$ & $j^{th}$ row of the the matrix $\DATAMATRIX{A}{}$ (feature) \tabularnewline
$\DATAMATRIXelem{a}{i}{j}$ & $(i,j)^{th}$ element of the the matrix $\DATAMATRIX{A}{}$ \tabularnewline
\hline
$\absval{\emptyarg}$ & Absolute-value function \tabularnewline
$\APPLIEDFUNCTION{\lO}{\emptyarg}, \APPLIEDFUNCTION{\lI}{\emptyarg}, \APPLIEDFUNCTION{\lII}{\emptyarg}, ..., \APPLIEDFUNCTION{\lp}{\emptyarg}$ & $\lO$-, $\lI$-, $\lII$-, $\lp$-norms \tabularnewline
$\dist{A}{x}{y}$ & Distance $A$ between $x$ and $y$ \tabularnewline
\hline \hline
$\RV{X}{}$ & Random variable X \tabularnewline
$\PDF{p}{X}$ & Probability distribution function. \tabularnewline
& Shorthand for $\PDF{p}{X=x}$ or $\PDF{p_X}{x}$. \tabularnewline
$\PDF{p}{X,Y}$ & Joint distribution function \tabularnewline
$\CONDPDF{p}{X}{Y}$ & Conditional distribution function \tabularnewline
\hline
$\mom{i}{X}$ & $i^{th}$ moment of the random variable $X$ \tabularnewline
$\mean{X}$ & Expected value of the random variable $X$ \tabularnewline
$\var{X}$ & Variance of the random variable $X$ \tabularnewline
$\pdfdist{A}{p}{q}$ & $A$-distance between the pdfs $p$ and $q$ \tabularnewline
\hline \hline
$\nXsamples$ & Number of samples \tabularnewline
$\nXfeatures, \nZfeatures$ & Number of features of original and learned representations \tabularnewline
$\nYclasses$ & Number of labels \tabularnewline
\hline
$\X$ & Original data ($\X \in \Xspace$) \tabularnewline
$\Z$ & Learned representations ($\Z \in \Zspace$) \tabularnewline
$\Y$ & Labels ($\Y \in \Yspace$) \tabularnewline
$\W$ & Weight matrix ($\W \in \Wspace$) \tabularnewline
\hline
$\domx$ & Domain of data \tabularnewline
$\Xtr, \Xte, \Xta$ & Training, test, target data \tabularnewline
$\rvX$ & Random variable for the original data \tabularnewline
$\px$ & Pdf for the original data \tabularnewline
$\rvXtr, \rvXte, \rvXta$ & Random variable for training, test, target data \tabularnewline
$\rvXfeatj$ & Random variable for the $j^{th}$ feature \tabularnewline
\hline \hline
\end{tabular}
\end{centering}
\end{table*}
\subsection{Notation \label{sec:BG_Notation}}
Let us define the \emph{original data set} as the matrix $\X$ containing $\nXsamples$ data points. Each data point is represented as a vector $\x$ composed by $\nXfeatures$ features: $\x \in \domx \subseteq \origspace$. Elements of the vector $\x$ are referred as $\xij$, $1 \leq i \leq \nXsamples$ and $1 \leq j \leq \nXfeatures$. From a statistical point of view, a vector $\x$ is taken to be a random and independent sample from the multivariate marginal pdf $\px$.
Analogously, let us define the \emph{learned representation set} as the matrix $\Z$ containing the same number $\nXsamples$ of data points. Each data point is now represented as a vector $\z$ composed by $\nZfeatures$ features: $\z \in \domz \subseteq \learnspace$. Elements of the vector $\z$ are referred as $\zij$, $1 \leq i \leq \nXsamples$ and $1 \leq j \leq \nZfeatures$. From a statistical point of view, a vector $\z$ is taken to be a random and independent sample from the multivariate marginal pdf $\pz$.
Let us also consider the \emph{label set} as the vector $\Y$ made up of $\nXsamples$ categorical labels $\y \in \domy = \{1,2,\ldots,\nYclasses\}$. From a statistical point of view, a label $\y$ is taken to be a random and independent sample from the pdf $\py$.
In supervised learning, the original data is paired with the labels, $\set{\X,\Y}$, and learning means discovering a morphism $\MORPH{f}{\domx}{\domy}$ that maps original data onto the correct labels: $\APPLIEDMORPH{f}{\x}{\y}$. The degree of success of supervised learning is evaluated in terms of generalization, that is, how well the learned morphism $f$ approximates $\pygx$.
In unsupervised learning, learning may be expressed as the problem of modelling the original data through a morphism $\MORPH{g}{\domx}{\domz}$ that maps original data onto new representations: $\APPLIEDMORPH{g}{\x}{\z}$. The degree of success of unsupervised learning depends on the criteria chosen to evaluate the learned representations $\Z$ \citep{Bengio2012}. A simple way to assess the goodness of the learned representations is to evaluate their usefulness in supervised tasks, that is, evaluating whether approximating $\pygz$ is easier than approximating $\pygx$.\\
For the sake of evaluating the generalization of our algorithms, we will follow the convention of partitioning our dataset in (at least) two subsets: a training data set $\Xtr$ made up of $\nXtrsamples$ samples from $\pxtr$ for learning the morphism $f$; and a test dataset $\Xte$ made up of $\nXtesamples$ samples from $\pxte$ for testing the morphism $f$.\\
When learning from data $\X$ in the space $\domx$, the results of many machine learning algorithm \revII{Corrected}{are} dependent on the metric assumed to explain the distances and the relationships between the samples \citep{Xing2003}.
\revII{I added this sentence to quote the source referenced by the reviewer. However I kept a quite general and informal tone because actually arguing when different metrics lead to the same result is not straightforward. While in general it is sound to assert that different metrics do NOT lead to the same result (as we stated in the previous sentence with reference to Xing), arguing when they actually lead to the same result is more complicated and it depends on the learning algorithm too. For many algorithms a simple topological equivalence may be enough; for other algorithms, who may be sensitive to the absolute value of the distances, a strong equivalence or an isomorphism may be required. Discussing these different cases seem to me outside our scope and it would require the introduction of a topological notation too. Therefore I tried to make a high-level statement with a reference to the source (Runde)}{Different choices of metrics may} lead to different learning outcomes unless the metrics may be proved to be equivalent \citep{Runde2007}.
Data $\X$ are often tacitly assumed to be defined in an Euclidean metric space, such that, given two points $\xone$ and $\xtwo$, we can compute their distance as $\dist{E}{\xone}{\xtwo}= \sqrt{\sumjfeatX \left( \xonej - \xtwoj \right)^2}$. The Euclidean distance $\dist{E}{\xone}{\xtwo}$ is then often assumed to define the structure of the data $\X$. However, different distance functions $\MORPH{D}{\domx \setCartprod \domx}{\domRgrez}$ may be considered to evaluate the distance between $\xone$ and $\xtwo$, thus inducing different structures on the data $\X$.
When considering the conditional distribution $\pygx$, we say that the labelled data $\set{\X,\Y}$ \emph{has a structure explained by a metric function} $D$ if the following $ \epsilon$-$\delta$ condition is satisfied: given two data samples $\xone,\xtwo \in \domx$, if $\dist{}{\xone}{\xtwo} < \epsilon$, then $\left| \CONDPDF{p}{y}{\xone} - \CONDPDF{p}{y}{\xtwo} \right| < \delta(\epsilon)$, where $\epsilon \in \domRgrz$ is an arbitrarily small positive constant, and $\MORPH{\delta}{\domRgrz}{\domRgrz}$ is a function returning an arbitrarily small positive constant depending on $\epsilon$.
In other words, a metric function $D$ explains the data structure if it preserves locally the conditional distribution: two points $\xone,\xtwo$ evaluated as close to each other, $\dist{}{\xone}{\xtwo} \approx 0$, have also similar conditional probabilities, $\CONDPDF{p}{y}{\xone} \approx \CONDPDF{p}{y}{\xtwo}$. \\
Lastly, given a data samples $\x$, we say that $\x$ is sparse if $m \ll \nXfeatures$ components of the vector $\x$ are active (that is, have a value different from zero), while the remaining $\nXfeatures-m$ components are inactive (that is, they have the value zero); we say that $\x$ is $k$-sparse if exactly $k$ components are active \citep{Makhzani2013}.
By analogy, we may define sparsity for matrices (with reference to their components) and for random variables (with reference to their realizations).
With regard to matrices, given a data matrix $\X$ made up of $\nXsamples$ samples in $\nXfeatures$ dimensions, we also say that $\X$ is \emph{population spares} if the sample vectors $\x$ in $\X$ are sparse, while we say that $\X$ is \emph{lifetime sparse} or \emph{selective} if the feature vectors $\xj$ in $\X$ are sparse.
\subsection{Covariate Shift \label{sec:BG_CS}}
Under the \iid assumption all the data samples making up the matrix $\X$ are independent samples from the same pdf $\px$; thus, $\pxtr = \pxte$. As long as the conditional distribution of the labels remain the same across training and test data, $\pygxtr = \pygxte$, then generalization from the training data $\Xtr$ to the test data $\Xte$ is straightforward \citep{Vapnik1998,Bishop2006}.
Unfortunately, the \iid assumption rarely holds in the real-world, where training data and test data may significantly differ in their distribution; this is the case, for instance, when an algorithm is trained on well-behaved data samples collected in a controlled environment, and then it is deployed in the wild. To model this situation we explicitly rely on the assumptions of \emph{covariate shift} \citep{Shimodaira2000,Sugiyama2012}.
This assumption can be divided into two statements: (i) the marginal distribution of the training and test data are defined on the same sub-domains, $\domxtr = \domxte$, but their pdfs are different, $\pxtr \neq \pxte$, and (ii) the conditional distribution of the labels is the same for the training and test data, $\pygxtr = \pygxte$.
The first statement of the assumption of covariate shift is supposed to model the original problem of a difference between the distribution of training and test data \rev{I added this part to relate our problem of covariate shift to domain shift}{and to guarantee the identity of the sub-domains of the distributions; in case of domain shift, $\domxtr \neq \domxte$, it would not be possible to establish the identity of the conditionals because $\pygx$ would then be defined over different domains for training and test data.}
The second statement is supposed to guarantee the possibility of learning and generalizing from training to test data; if both the marginal and the conditional distribution of the training and test data were to be different, generalization would not be possible as $\pxtr$ and $\pxte$ would be completely unrelated.\\
It is worth noticing that, even if many supervised discriminative algorithms are apparently concerned only with estimating a conditional distribution $\pygx$ which is the same over training and test data, they are actually affected by covariate shift in the marginals because they optimize an average error estimated on the training domain \citep{Yamada2012}. Indeed, standard supervised algorithms, such as SVM, are affected by the density of the training data; only purely conditional models, such as Gaussian processes, are immune to covariate shift \citep{Quinonero-Candela2009}. It is therefore important, even when learning a conditional distribution through standard supervised discriminative algorithms, to perform CSA to compensate for the difference between the marginal pdfs of the training and test data.
\subsection{Covariate Shift Adaptation \label{sec:BG_CSA}}
A CSA algorithm is an algorithm designed to make learning possible under covariate shift. Often, a CSA algorithm compensates for the difference between the training and test distributions in order to allow further supervised learning; in this case, a CSA algorithm should not only tackle covariate shift, but also retain relevant discriminative information carried by the conditional distributions.
We can express these requirements more formally in the following two necessary, but not sufficient, conditions:
\begin{description}
\item[\emph{Marginal condition}:] a CSA algorithm must compensate for the difference between the marginal distributions of the training and the test data, that is, $\pdfdist{}{\pxtr}{\pxte} > \pdfdist{}{\pztr}{\pzte}$, where $\pdfdist{}{\emptyarg}{\emptyarg}$ is a measure of distance among pdfs, such as Kullback-Leibler divergence \citep{MacKay2003} or maximum mean discrepancy \citep{Gretton2012};
\item[\emph{Conditional condition}:] a CSA algorithm must preserve the identity of the conditional distributions\footnote{This condition may actually be improved by requiring not just the preservation of the original $\pygx$ but instead the learning of a better-behaved conditional $\pygz$ (as it happens in representation learning).}, that is, $\pygz = \pygx$.
\end{description}
These two conditions have to be simultaneously satisfied to achieve good CSA.
The first condition is necessary but not sufficient: if an algorithm were not to compensate for the difference in the distribution of training and test data, then no adaptation would take place; on the other hand, if an algorithm were to compensate for the difference in the distribution of training and test data, then we would still have no guarantee about CSA being successful since all discriminative information may be lost.
Similarly, the second condition, too, is necessary but not sufficient: if an algorithm were not to preserve the identity of the conditionals, then ensuing discrimination would be compromised; if an algorithm were to preserve the identity of the conditionals, then we would still have no guarantee that the distance between the marginals had been reduced.
Notice that these two conditions closely mirror the two statements in the assumption of covariate shift. The marginal condition requires the problem in the first statement of the covariate shift assumption to be addressed. The conditional condition requires that the guarantee for learning expressed in the second statement of the covariate shift assumption is preserved.\\
The statistical and machine learning literature contains several algorithms for performing CSA \citep[for a review of CSA algorithms see, for instance,][]{Jiang2008,Margolis2011}. Two types of approach to CSA are particularly relevant to our work: CSA via importance weighting and CSA via representation learning.
\subsubsection{CSA via Importance Weighting}
The first type of CSA is importance weighting (IW) \citep{Sugiyama2012}. IW rescales the loss function of a learning algorithm by the ratio between the distribution of the training data and the test data, thus emphasizing the contribution to learning of training points close to the test points and discounting the contribution of training points falling far from the test points. IW works under two conditions: (i) no domain shift, $\domxtr = \domxte$, which allows us to define the ratio between the pdf of the training samples and the pdf of the test samples; and (ii) smooth conditional distribution between the training and test domain \citep{Sugiyama2012}.
CSA is performed by rescaling the contribution to the loss function of each training sample according to the ratio between the marginal distribution of the training data and the test data. The loss function $\loss$ is modified based on each individual instance $\xtr$:
\[
\frac
{\PDF{p}{\rvXtr = \xtr}}
{\PDF{p}{\rvXte = \xtr}}
\APPLIEDFUNCTION{\loss}{\xtr}
\]
where the ratio $\frac{\pxtr}{\pxte}$ is estimated as $\frac
{\PDF {\estimation{p}} {\rvXtr} }
{\PDF {\estimation{p}} {\rvXte} }$ using density ratio estimation algorithms \citep{Sugiyama2012}.
The IW approach is theoretically well-grounded, but it has a high computational complexity due to the problem of estimating the pdfs $\pxtr$ and $\pxte$ or their ratio $\frac{\pxtr}{\pxte}$. The IW approach has been successfully applied to many problems, such as speaker identification, EEG processing, natural language processing \citep{Sugiyama2012} and emotional speech recognition \citep{Hassan2013}.
\subsubsection{CSA via Representation Learning}
The second type of CSA is representation learning (RL). In general, RL algorithms are designed to discover a new and better representation $\Z$ for the given data $\X$; better representations are intuitively described as abstract, denoised, robust or informative representations \citep{Bengio2012}. When applied to CSA, RL algorithms are enriched with the additional aim of tackling the shift in the pdfs of the training data and test data. This objective may be formally expressed in the requirement of discovering a new space $\domz$ such that $\pdfdist{}{\pxtr}{\pxte} > \pdfdist{}{\pztr}{\pzte}.$
Most of the existing RL algorithms work under the assumptions of \iid data and thus need to be redesigned to tackle covariate shift. However, some algorithms, most notably denoising auto-encoders (DAE) \citep{Vincent2008a}, have been shown to be able to perform effective CSA without any explicit modification to account for covariate shift.
A DAE module computes a new representation $\z$ of a data sample $\x$ as:
\[
\z = \APPLIEDFUNCTION {f} {\W \mathbf{\tilde{x}}_i+\mathbf{b}},
\]
where $\mathbf{\tilde{x}}_i$ is a corrupted noisy version of $\x$, $\MORPH{f}{\domR}{\domR}$ is an element-wise non-linear function, $\W \in \Wspace$ \revII{Corrected}{denotes} the generic weight matrix\footnote{Notice that the generic notation $\W$ is used to denote the weight matrix of different learning algorithms; the context disambiguates the meaning of the notation.} and $\mathbf{b} \in \learnspace$ is the bias vector.
From the learned representation $\z$ a DAE module computes a reconstruction as:
\[
\mathbf{\hat{x}}_i = \APPLIEDFUNCTION {g} {\V \z + \mathbf{c}},
\]
where $\MORPH{g}{\domR}{\domR}$ is an element-wise non-linear function, $\V \in \Vspace$ is the reconstruction weight matrix and $\mathbf{c} \in \origspace$ is the reconstruction bias vector.
According to the implementation of the DAE, the non-linear functions might be the same, $\APPLIEDFUNCTION{f}{\emptyarg}=\APPLIEDFUNCTION{g}{\emptyarg}$, and the weight matrices might be tied, $\W = \transp{\V}$.
Learning is performed by minimizing over all the parameters a reconstruction loss, such as the square error:
\[
\argminproblem{\W\in\Wspace, \V\in \Vspace, \mathbf{b}\in\learnspace, \mathbf{c}\in\origspace}{\sumisamples \left( \mathbf{\hat{x}}_i - \x \right)^2}.
\]
A solution to this optimization problem can be computed by gradient descent minimizing over all the available data. Even if no rigorous theoretical explanation has been provided to justify the use of DAE for CSA, DAE and stacked DAE \citep{Vincent2010} were demonstrated to be able to perform effective CSA \citep{Glorot2011a}.\\
A representation learning solution explicitly designed to tackle covariate shift is sub-space alignment (SSA) \citep{Fernando2013}. SSA works under the assumption that the space defined by the PCA components of the training data can be projected onto the space of the PCA components of the test data. SSA aligns the PCA spaces by computing the following representations:
\begin{eqnarray*}
\Ztr & = & \Xtr \mathbf{T} \transp{\mathbf{T}} \mathbf{U}\\
\Zte & = & \Xte \mathbf{U},
\end{eqnarray*}
where $\mathbf{T}$ is the matrix of the eigenvectors of the covariance matrix for the training data, and $\mathbf{U}$ is the matrix of the eigenvectors of the covariance matrix for the test data.\\
Other approaches to RL algorithms include methods based on the identification of pivot features \citep[such as \emph{structural correspondence learning},][]{Blitzer2006}, methods grounded in manifold learning \citep[such as \emph{geodesic flow kernels},][]{Gong2012}, and methods based on the minimization of a differentiable measure of distance between the original and the learned distribution \citep[such as \emph{kernel mean matching},][]{Huang2007, Quadrianto2009}.\\
CSA through RL allows to solve the problem of covariate shift, while, at the same time, learn a better-behaved conditional distribution $\pygz$. The RL approach has been successfully applied to many problems dealing with complex data, such as image recognition \citep{Kulis2011, Tzeng2014}, sentiment analysis \citep{Glorot2011a, Li2014} and emotional speech recognition \citep{Deng2014a}.
\subsection{Sparse Filtering \label{sec:BG_FDL}}
SF is the main representation learning algorithm belonging to the family of \emph{feature distribution learning} algorithms. While other unsupervised algorithms normally learn new representations by trying to recover the original noiseless distribution that generated the data, feature distribution learning algorithms aim at producing new representations
with a useful distribution for classification. SF learns
a mapping $ \MORPH{f}{\origspace}{\learnspace}$ projecting the data onto maximally sparse representation \citep{Ngiam2011}:
\[
\Z=\lIIcol { \lIIrow { \absval {\W \X} } },
\]
where $\W \in \Wspace$ is the weight matrix,
$\absval{\emptyarg}$ is the element-wise absolute-value function,
$\lIIrow{\emptyarg}$ is the $\lII$-normalization along the rows: $\lIIrow{\X} = \frac{\xij} {\sqrt {\sumisamples \left(\xij\right)^{2}}}$,
and
$\lIIcol{\emptyarg}$ is the $\lII$-normalization along the columns: $\lIIcol{\X} = \frac{\xij} {\sqrt{ \sumjfeatZ \left(\xij\right)^{2}}}$.
For computational reasons, the absolute-value non-linearity is implemented as a soft absolute-value function: $\FUNCTION{f}{x}{\sqrt{x^2+\epsilon}}$, where $\epsilon$ is an arbitrarily small positive constant, such as $\epsilon=10^{-8}$. SF is trained by minimizing the $\ell_1$-norm of the learned representations:
\[
\argminproblem{\W \in \Wspace} { \sumisamples \sumjfeatZ \zij},
\]
using gradient descent.
SF was shown to be an excellent algorithm for unsupervised learning as it scales very well with the dimensionality of the input, it is easy to implement, it has essentially a single tunable parameter and it was able to achieve state-of-the-art performance on image recognition and phone classification \citep{Ngiam2011}. Since its introduction, SF has been integrated in several machine learning systems tackling problems as diverse as iris recognition \citep{Raja2015}, electrical fault detection \citep{Lin2016} or terrain classification \citep{Liu2016a}.
Moreover, SF has also been the object of more theoretical studies \citep{Lederer2014, Zennaro2016}. In particular, in our earlier work \citep{Zennaro2016}, we showed that learning in SF can be explained in terms of defining filters in the original space $\origspace$ which map data samples $\x$ onto bases of the learned representation space $\learnspace$; consequently, we proved that SF can learn good representations for classification when the conditional distribution of the labels $\pygx$ is explained by a metric of cosine neighbourhoodness \citep{Zennaro2016}.
\section{Theoretical Analysis of Sparse Filtering for Covariate Shift Adaptation \label{sec:CSA-via-SF}}
This section provides a theoretical analysis of the standard SF algorithm aimed at understanding \emph{whether} and \emph{under which conditions} SF successfully performs CSA.
Section \ref{sec:CSA-via-SF_Marginal} discusses when the marginal condition for CSA is met by SF, while Section \ref{sec:CSA-via-SF_Conditional} discusses when the conditional condition for CSA is met.
\subsection{Marginal CSA Condition for SF \label{sec:CSA-via-SF_Marginal}}
In order to prove that SF meets the marginal condition for CSA, it is necessary to show that SF reduces the distance between the marginal distributions of the training and test data. Specifically, we need to show that the sub-domains of definition of the training and test data are kept the same, $\domztr = \domzte$, while the difference between the new pdfs $\pztr$ and $\pzte$ is reduced.\\
Let us start showing that, through its normalization steps, SF projects all the samples onto a common bounded sub-domain.
\begin{proposition} \label{prop:SF-Marginal-Condition-1}
The sub-domain $\domz$ of the representations $\z$ learned by SF is $\closedinterval{0}{1}^{\nZfeatures}$.
\end{proposition}
\textbf{Proof sketch.} By definition, all the representations $\z$ learned by SF are bounded through $\ell_2$-normalization within the hyper-cube $\closedinterval{0}{1}^{\nZfeatures}$. For the complete proof, see Appendix \ref{app:Proof-of-Proposition-1}. $\QED$
Thus, independently from the original sub-domain of definition of the training and test data, the sub-domain $\domz$ of the learned representations of the training and test data is always identical.\\
Now let us consider the problem of a difference in the pdfs of training and test data. We show in the following proposition that SF addresses covariate shift by reshaping the pdfs of each learned feature within precise bounds.
\begin{proposition} \label{prop:SF-Marginal-Condition-2}
For each learned feature $\zj$, the SF algorithm bounds $\mean{\rvZfeatj} \in \closedinterval{\epsilon}{1}$ and $\var{\rvZfeatj} \in \closedinterval{0}{1-\epsilon^2}$, where $\epsilon>0$ is an arbitrarily small value defined in the non-linearity of SF.
Moreover, if we make the assumption that learned representations are $\closedinterval{1}{k}$-sparse in population and lifetime, and that $\epsilon$ is negligible, then we have the bounds $\mean{\rvZfeatj} \in \closedinterval{\frac{1}{\nXsamples}}{\frac{k}{\nXsamples}}$ and $\var{\rvZfeatj} \in \closedinterval{\frac{\nXsamples-k^{2}}{\nXsamples^{2}}}{\frac{\nXsamples k-1}{\nXsamples^{2}}}$.
\end{proposition}
\textbf{Proof sketch.} By closely analysing each processing step of SF, we can bound the distribution of each feature. For the formal analysis, see Appendix \ref{app:Proof-of-Proposition-2}. $\QED$
Thus, SF tackles the problem of covariate shift by forcing all the features to have bounded expected values and bounded variances. These bounds on the distribution of the features $\rvZfeatj$ are theoretically independent of the data matrix being processed. The fact that the new learned distributions $\PDF{p}{\rvZtr_{\emptyarg,j}}$ and $\PDF{p}{\rvZte_{\emptyarg,j}}$ have the first statistical moments similarly bounded on the same interval suggests that SF is able, at least in part, to mitigate the problem of covariate shift.
More interestingly, Proposition \ref{prop:SF-Marginal-Condition-2} reveals that, under the assumption of $k$-sparsity, SF moves the center of mass of the pdf of each feature $\PDF{p}{\rvZfeatj}$ towards zero, and it also decreases the variance in proportion to the number of samples $\nXsamples$. In other words, SF not only shapes the overall pdf $\pz$ towards being mainly localized around zero, but also does the same for the individual pdf of each feature $\PDF{p}{\rvZfeatj}$.
Note that this is consistent with the interpretation of SF in terms of entropy minimization \citep{Zennaro2016}; the maximization of sparsity may be interpreted as a proxy for the minimization of entropy \citep{Principe2010,Pastor2015}, and SF can be understood as an algorithm projecting the original data onto representations with a pdf with minimal entropy. Therefore, independently from the original pdfs $\pxtr$ and $\pxte$, SF aims at learning a new representation with an entropy-minimized pdf $\pz$.
\subsection{Conditional CSA Condition for SF \label{sec:CSA-via-SF_Conditional}}
Satisfying the marginal condition for CSA is not enough to guarantee that SF always generates useful representations for classification. Indeed, if the samples were to be randomly mapped to maximally sparse representations in agreement with the marginal condition only, we would very likely lose all discriminative information carried by the data. For discriminative information to be retained, SF must satisfy the conditional condition requiring the preservation of the identity of the conditional distribution $\pygz = \pygx$.
To verify when this condition is met, we need to determine in which situation SF can preserve the identity of the conditional distributions of the labels given training and test data.
In our earlier work \citep{Zennaro2016}, we showed that SF can preserve the conditional structure of $\pygx$ explained by a metric of cosine neighbourhoodness (\revII{Corrected}{in order to be self-contained, we include} in Appendix \ref{app:preservation-cosine-neigh} the proof of the theorem proving the preservation of cosine neighbourhoodness in SF appearing in our previous work, \citet{Zennaro2016}). By expressing cosine neighbourhoodness in terms of cosine distance, it is immediate to derive the following corollary for the preservation of the structure of the conditional distribution $\pygx$.
\begin{corollary} \label{coroll:SF-Conditional-Condition}
SF preserves the structure of the conditional distribution $\pygx$ explained by the metric of cosine distance $\dist{C}{\xone}{\xtwo} = 1 - \frac{\xone \xtwo}{\APPLIEDFUNCTION{\lII}{\xone}\APPLIEDFUNCTION{\lII}{\xtwo}}$.
\end{corollary}
SF transforms the conditional distribution $\pygx$ explained by the cosine metric in the original space into a new conditional distribution $\pygz$ explained by the Euclidean distance in the learned space \citep{Zennaro2016}. Thus, if in the original space a small cosine distance implies a small difference in the conditional distribution:
\[
\left[\dist{C}{\xone}{\xtwo} < \epsilon \right] \implies
\left[\absval{\CONDPDF{p}{y}{\xone} - \CONDPDF{p}{y}{\xtwo}} < \APPLIEDFUNCTION{\delta}{\epsilon}\right],
\]
where $\epsilon$ is an arbitrarily small value $\epsilon>0$ and $\APPLIEDFUNCTION{\delta}{\epsilon}$ is a function dependent from $\epsilon$ and returning an arbitrarily small value $\APPLIEDFUNCTION{\delta}{\epsilon}>0$, then in the learned space a small Euclidean distance implies a small difference in the conditional distribution:
\[
\left[\dist{E}{\zone}{\ztwo} < \epsilon' \right] \implies
\left[\absval{\CONDPDF{p}{y}{\zone} - \CONDPDF{p}{y}{\ztwo}} < \APPLIEDFUNCTION{\delta'}{\epsilon'}\right].
\]
This, in turn, allows standard Euclidean-based classifiers to successfully process the new representations.\\
In conclusion, bringing together these results for CSA using SF, we have:
\begin{theorem}
SF meets the necessary and sufficient conditions for CSA if the structure of the conditional distribution $\pygx$ is explained by a cosine metric.
\end{theorem}
\textbf{Proof.} This theorem follows directly from Propositions \ref{prop:SF-Marginal-Condition-1} and \ref{prop:SF-Marginal-Condition-2}, proving the marginal condition, and Corollary \ref{coroll:SF-Conditional-Condition}, proving the conditional condition. $\QED$\\
SF is then able to perform some degree of CSA under the same conditions required for SF to perform useful unsupervised learning \citep{Zennaro2016}. SF can then be used to perform CSA, but only on a limited number of real-world data sets that can be expected to comply with the assumption of a conditional structure explained by the cosine metric.
\section{Periodic Sparse Filtering \label{sec:PSF}}
In the previous section we showed that the capacity of SF to perform CSA is limited by a strict requirement on the conditional distribution. In this section we propose a new SF-based algorithm that extends this requirement to a conditional distribution explained by a generic periodic structure. A periodic conditional structure could be used to model several real-world scenarios, as periodic functions would allow us to capture common regularities present in the marginal distributions of training and test data. For instance, in the common case of user-dependent data, we could model each user as described by a specific pdf $\PDF{p}{X^{user_i}}$ on a restricted sub-domain $\domx^{user_i}$; labels may then be expected to show some degree of regularity over each sub-domain, such that $ \CONDPDF{p}{Y}{X^{user_1}} = \CONDPDF{p}{Y}{X^{user_2}} = \dots = \CONDPDF{p}{Y}{X^{user_i}}$; learning the periodic behaviour over the set of training users would then allow us to generalize it to the set of test users.\\
We set out to define a new SF-based algorithm, which we call \emph{periodic sparse filtering} (PSF), by enriching the original SF algorithm with two crucial properties: (i) the ability to generate periodic filters in the original space, and (ii) the ability to capture the periodic conditional structure underlying the data from available labels.
The first property is required in order to preserve the periodic structure of the conditional distribution $\pygx$.
SF generates hyper-conical filters that can capture a radial structure, but are unable to model more complex periodic structures in the original space $\origspace$ \citep[see Section 3.11 in][]{Zennaro2016}. In order to capture periodic structures, we substitute the absolute-value non-linearity with a sinusoidal function which can generate periodic filters that regularly tile the whole original space.
The new transformation in PSF is defined as follows:
\[
\Z= \lIIcol{\lIIrow{\APPLIEDFUNCTION{g}{\W \X}}},
\]
where $\APPLIEDFUNCTION{g}{x}$ is a positive element-wise sinusoidal function, such as $1+\epsilon+\sin(x)$ or $1+\epsilon+\cos(x)$, a sinusoidal function shifted by $1+\epsilon$, with $\epsilon>0$ being an arbitrarily small constant, such as $\epsilon=10^{-8}$, that guarantees the strict positivity of the output of $\APPLIEDFUNCTION{g}{x}$.
The second property is required in order to capture the correct periodicity underlying the conditional distribution $p\left(Y\vert X\right)$.
Given a set of unlabelled data, it is possible to discover different periodic functions underlying the data; however, only few of these periodic functions can be usefully related to the conditional distribution $\pygx$. In order to direct the algorithm to discover the conditional periodic function of interest, we turned the original unsupervised adaptation algorithm into a supervised adaptation algorithm.
Let $\Xtr$ be the training data with its associated labels $\Ytr$. Assuming that we are learning a new representation in $\learnspace$, let us partition the $\nZfeatures$ learned features in $\nYclasses+1$ groups with arbitrary cardinality, corresponding to the $\nYclasses$ classes defined in $\Ytr$, plus one group for potentially unlabeled samples.
We can then re-define the learned representation matrix $\Z$ using the following block matrix notation:
\[
\Z=\left[\begin{array}{ccccc}
\submZ{1}{1} & \submZ{1}{2} & \dots & \submZ{1}{C} & \submZ{1}{C+1}\\
\submZ{2}{1} & \submZ{2}{2} & \dots & \submZ{2}{C} & \submZ{2}{C+1}\\
\dots & \dots & \dots & \dots & \dots\\
\submZ{C}{1} & \submZ{C}{2} & \dots & \submZ{C}{C} & \submZ{C}{C+1}\\
\submZ{C+1}{1} & \submZ{C+1}{2} & \dots & \submZ{C+1}{C} & \submZ{C+1}{C+1}\\
\end{array}\right],
\]
where $\submZ{i}{j}$ is the block matrix containing the $i^{th}$ group of learned features from the samples belonging to the $j^{th}$ class. This structure highlights the contribution of each group of learned features to the representation of samples belonging to a given class.
Exploiting the representation matrix in this new form, we can redefine the loss function of PSF as:
\[
\argminproblem{\W \in \Wspace}{\APPLIEDFUNCTION{\lI}{\Z} -
\sum_{c=1}^{\nYclasses} } \lambda_{c} \multiplication \APPLIEDFUNCTION{\lI}{\submZ{c}{c}},
\]
where $\lambda_{c} \in \domR$ is a scaling factor. The first term of this new loss function is the same as in SF, and its aim is to push for learning sparse representations. The second term of the loss function is the PSF addition. This term has an opposite effect compared to the first: while the first term tries to reduce and shrink the values of the elements of $\Z$, the second term tries to increase the values of the sub-matrices of $\Z$ around the diagonal. Overall, this loss function should push away mass from the components off the diagonal and push it onto the components on the diagonal. These dynamics are reminiscent of learning with energy-based models \citep{LeCun2006}, in which an energy surface over a learned space is pulled down over desired outcomes and pulled up over other outcomes; similarly, even though in a reverse way, our loss function tries to increase the mass over certain outcomes and decrease it over other outcomes. Practically, the learning algorithm is now biased towards generating sparse representations where the $c^{th}$ group of learned features tends to activate for the samples belonging to the $c^{th}$ class.\\
The pseudo-code for the new PSF algorithm\footnote{The Python source code is available online at: \url{https://github.com/FMZennaro/PSF}} is provided in Algorithm \ref{alg:PSF}. In conclusion, our PSF algorithm retains the simplicity of the original SF, but at the same time it exploits the labels available for classification to perform CSA.
\begin{algorithm} \caption{Periodic Sparse Filtering (PSF) \label{alg:PSF}}
\begin{algorithmic}[1]
\Statex \textbf{Input:} training data $\Xtr$; training labels $\Ytr$; target data for adaptation $\Xta$.
\Statex \textbf{Hyper-parameters:} learned dimensionality $\nZfeatures$; lambda vector $\lambdavector$; binary matrix $\Mask$ defining the block matrix structure of $\Z$ such that $\maskcj=1$ if the $j^{th}$ learned feature is activated by the $c^{th}$ class; gradient descent step $\GDstep$.
\Statex
\State $\X \leftarrow \Xtr \setunion \Xta$
\State $\W \leftarrow \textnormal{initialize each weight as } \pdfstdnorm$
\State $C \leftarrow \#$classes in $\Ytr$ \Statex
\Repeat
\State $\IntReprI \leftarrow \W \X$
\State $\F \leftarrow 1 + \epsilon + \sin\IntReprI$
\State $\Ftilde \leftarrow \frac{\fij}{\sqrt{\sumisamples\fij^2}}$
\State $\Z \leftarrow \frac{\ftildeij}{\sqrt{\sumjfeatZ\ftildeij^2}}$
\State $\loss_1 \leftarrow \sumisamples \sumjfeatZ \zij$
\State $\loss_2 \leftarrow \sum_{c=1}^{\nYclasses} \sum_{i: \y=c} \sum_{j: \maskcj=1 } \lambdavectorc \multiplication \zij$
\State $\loss \leftarrow \loss_1 - \loss_2$
\Statex
\State $\W \leftarrow \W - \GDstep \nabla \loss$
\Until termination condition for gradient descent is met
\Statex
\Return $\Z$
\end{algorithmic}
\end{algorithm}
\section{Theoretical Analysis of Periodic Sparse Filtering for Covariate Shift Adaptation \label{sec:CSA-via-PSF}}
This section offers a theoretical analysis of PSF, analogous to the one we conducted on SF in Section \ref{sec:CSA-via-SF}.
Section \ref{sec:CSA-via-PSF_Marginal} discusses when the marginal condition for CSA is met by PSF, while Section \ref{sec:CSA-via-PSF_Conditional} discusses when the conditional condition is met.
\subsection{Marginal CSA Condition for PSF \label{sec:CSA-via-PSF_Marginal}}
Analogously to the theoretical analysis of the marginal CSA condition for SF, we first prove that PSF projects all the samples onto a common bounded sub-domain. It is straightforward to show that PSF guarantees this property in the same way as SF does.
\begin{proposition} \label{prop:PSF-Marginal-Condition-1}
The sub-domain $\domz$ of the representations $\z$ learned by PSF is $\closedinterval{0}{1}^{\nZfeatures}$.
\end{proposition}
\textbf{Proof.} This proposition follows from the $\lII$-normalization along the rows, which is exactly the same in both SF and PSF.
Therefore, the proof for Proposition \ref{prop:SF-Marginal-Condition-1} holds here as well. $\QED$\\
Next, let us consider the problem of a difference in the pdfs of training and test data. Like SF, PSF also reshapes the pdfs of each learned feature within precise bounds as specified in the following proposition.
\begin{proposition} \label{prop:PSF-Marginal-Condition-2}
For each learned feature $\zj$, the PSF algorithm bounds $\mean{\rvZfeatj} \in \closedinterval{\epsilon}{1}$ and $\var{\rvZfeatj} \in \closedinterval{0}{1-\epsilon^2}$, where $\epsilon>0$ is an arbitrarily small value defined in the non-linearity of PSF.
Moreover, if we make the assumption that learned representations are $\closedinterval{1}{k}$-sparse in population and lifetime, and that $\epsilon$ is negligible, then we have the bounds $\mean{\rvZfeatj} \in \closedinterval{\frac{1}{\nXsamples}}{\frac{k}{\nXsamples}}$ and $\var{\rvZfeatj} \in \closedinterval{\frac{\nXsamples-k^{2}}{\nXsamples^{2}}}{\frac{\nXsamples k-1}{\nXsamples^{2}}}$.
\end{proposition}
\textbf{Proof.} The bounds on the distribution of the learned features depends on the $\lII$-normalization steps, which are the same in SF and PSF.
Therefore the proof for Proposition \ref{prop:SF-Marginal-Condition-2} holds here as well. $\QED$
\subsection{Conditional CSA Condition for PSF \label{sec:CSA-via-PSF_Conditional}}
Having proved that PSF meets the marginal condition to perform CSA, we now show that our PSF further meets the conditional condition for CSA when data has a periodic structure.
By construction, the PSF algorithm was designed with the idea of preserving a periodic structure underlying the data. It is then natural to expect that the identity of the conditional distributions, $\pygz=\pygx$, is preserved when the data exhibit such a structure. To confirm that our design works as we intended, we first prove a theorem about data structure preservation in PSF.
\begin{theorem}
Let $\xone \in \origspace$ be a point in the original space and let $\zone \in \learnspace$ be its corresponding representation learned by PSF. Then there is an infinite set of points $\x \in \origspace$ that map onto the same representation $\zone$. The set of the points $\x \in \origspace$ built from $\xone$ with period $\W{}^{-1} \periodvector \pi$, where $\W$ is the weight matrix of PSF and $\periodvector$ is a vector of integer constants in $\domz$, is included in this set.
\end{theorem}
\textbf{Proof sketch.} By analysing each processing step of PSF, we can reconstruct how periodic filters are defined in the original space and show that points falling within these filters are mapped onto identical representations. For the complete proof, see Appendix \ref{app:Proof-of-Theorem-2}. $\QED$\\
This theorem proves that the PSF algorithm can define a specific frequency for each dimension in the original space; all the points with coordinates that are multiples of these frequencies are then mapped onto the same representation (see, for illustration, a comparison between the filters learned by SF and PSF in Figure \ref{fig:filters}(b)).\\
\begin{figure}
\begin{centering}
\includegraphics[scale=0.4]{filters1}
\par\end{centering}
\caption{Sample filters in the original space $\domR^2$ learned by: (a) SF and (b) PSF \label{fig:filters}}
\end{figure}
From Theorem 2, we can immediately derive the following corollary on the preservation of the structure of the conditional distribution $\pygx$.
\begin{corollary} \label{coroll:PSF-Conditional-Condition}
PSF preserves the structure of the conditional distribution $\pygx$ explained by a periodic metric $\dist{P}{\xone}{\xtwo} = \APPLIEDFUNCTION{\lp}{\APPLIEDFUNCTION{g_\periodvector}{\xone} - \APPLIEDFUNCTION{g_\periodvector}{\xtwo}}$, where $\APPLIEDFUNCTION{g_\periodvector}{\x}$ is an element-wise periodic function with periods $\periodvector$ and $\APPLIEDFUNCTION{\lp}{\emptyarg}$ is an $\lp$-norm.
\end{corollary}
In this case, if in the original space a small periodic distance implies a small difference in the conditional distribution:
\[
\left[\dist{P}{\xone}{\xtwo} < \epsilon \right] \implies
\left[\absval{\CONDPDF{p}{y}{\xone} - \CONDPDF{p}{y}{\xtwo}} < \APPLIEDFUNCTION{\delta}{\epsilon}\right],
\]
then, in the learned space a small Euclidean distance implies a small difference in the conditional distribution:
\[
\left[\dist{E}{\zone}{\ztwo} < \epsilon' \right] \implies
\left[\absval{\CONDPDF{p}{y}{\zone} - \CONDPDF{p}{y}{\ztwo}} < \APPLIEDFUNCTION{\delta'}{\epsilon'}\right].
\]
\\
In conclusion, we can summarize our theoretical analysis of PSF for CSA in the following theorem:
\begin{theorem} \label{thm:preservation-periodicity}
PSF meets the necessary and sufficient conditions for CSA if the structure of the conditional distribution $\pygx$ is explained by a periodic metric.
\end{theorem}
\textbf{Proof.} This theorem follows directly from Propositions \ref{prop:PSF-Marginal-Condition-1} and \ref{prop:PSF-Marginal-Condition-2}, proving the marginal condition, and Theorem \ref{thm:preservation-periodicity} and Corollary \ref{coroll:PSF-Conditional-Condition}, proving the conditional condition. $\QED$\\
Compared to SF, PSF can then perform CSA under the looser requirement of a periodic structure. However this increased capacity requires additional information for directing the learning process; since many different periodic structures may be learned, PSF needs to rely on side information in the form of labels in order to extract a useful or meaningful structure.
\section{Experimental Validation \label{sec:Experiments}}
In this section we validate and test our theoretical results on the use of SF and PSF for CSA.
In Section \ref{sec:Synthetic} we start by running experiments on synthetic data sets in order to obtain a better understanding and to be able to easily visualize the effects of covariate shift and the contributions of SF and PSF. In Section \ref{sec:Real-World-Data-Experiments} we execute a series of experiments on real data sets in which we measure the effectiveness of SF and PSF against other CSA algorithms used in the machine learning literature.\\
In order to highlight the dependency of CSA algorithms on their underlying assumptions about the structure of the data and to measure SF and PSF against state-of-the-art methods, we have undertaken a comparative study using the three CSA algorithms reviewed in Section \ref{sec:BG_CSA}. We implemented the following classification systems, using a linear support vector machine (SVM) as a default classifier, unless otherwise stated:
\begin{enumerate}[label=(\roman*)]
\item \emph{SVM (without CSA)}: this model does not perform any CSA and it is used only to provide a baseline against which to evaluate the contribution of CSA algorithms.
\item \emph{SF+SVM}: this model allows for the evaluation of the CSA capacity of SF; as discussed in Section \ref{sec:CSA-via-SF}, SF is supposed to successfully perform CSA when the data has a radial structure.
\item \emph{PSF+SVM}: this model allows for the evaluation of the CSA capacity of PSF; as discussed in Section \ref{sec:CSA-via-PSF}, PSF is supposed to successfully perform CSA when the data has a periodic structure.
\item \emph{IW+LSPC}: this model implements a commonly used setting for CSA based on IW \citep{Hachiya2012} and relies on a least-square probabilistic classifier (LSPC) \citep{Sugiyama2012a}; as recalled in Section \ref{sec:BG_CSA}, IW can be expected to successfully perform CSA in absence of domain shift, $\domxtr = \domxte$, and when the conditional distribution $\pygx$ is smooth \citep{Sugiyama2012a}.
\item \emph{SSA+SVM}: this model implements the SSA algorithm \citep{Fernando2013}; as explained in Section \ref{sec:BG_CSA}, SSA can be expected to successfully perform CSA when the PCA components of the training and the test data can be projected on each other \citep{Fernando2013}.
\item \emph{DAE+SVM}: this model implements the DAE algorithm, using a setting analogous to the one considered, for instance, in \citep{Deng2014c}; while DAE has been successfully used for CSA as discussed in Section \ref{sec:BG_CSA}, the conditions under which DAE successfully perform CSA have not yet been derived theoretically.
\end{enumerate}
More details on the implementation and configuration of the above models are provided in Appendix \ref{app:Experiments}.
\subsection{Synthetic Data Set \label{sec:Synthetic}}
\begin{figure*}
\begin{centering}
\includegraphics[scale=0.45]{data2}
\caption{Synthetic data sets \emph{radial}, \emph{periodic}, \emph{smooth} and \emph{diagonal}, built respectively on the CSA assumptions of SF, PSF, IW and SSA. \label{fig:data}}
\end{centering}
\end{figure*}
We conducted experiments on elaborate synthetic data to validate and illustrate the following results:
(i) confirming the ability of SF and PSF to reduce the distance between training and test data; (ii) comparing SF and PSF against other well-known algorithms from the machine learning literature; (iii) highlighting that the success of these algorithms largely depends on whether the conditional condition for successful CSA is met by the data.\\
We generated four different data sets: \emph{radial} and \emph{periodic}, satisfying respectively the conditions of SF and PSF; \emph{smooth} and \emph{diagonal}, meeting respectively the assumptions of IW and SSA.
Figure \ref{fig:data} illustrates the four data sets (see Appendix \ref{app:Synthetic} for more details on data generation).
Data from the training distribution $\pxtr$ constitute the training data set $\Xtr$, while data from the test distribution $\pxte$ are evenly partitioned into a target data subset $\Xta$ and a test data subset $\Xte$. The training data $\Xtr$ and the target data $\Xta$ are used to train the CSA algorithm and the classification module, while the test data $\Xte$ is used only for performance evaluation.\\
All the data are processed using the CSA classification systems described above. The values chosen for the hyper-parameters of the different classification systems are detailed in Appendix \ref{app:Synthetic}.\\
In order to validate that SF and PSF meet the marginal CSA condition, we estimate the distance between the pdfs of the training and the test data before and after CSA using the Maximum Mean Discrepancy (MMD) distance \citep{Gretton2012}, and we report the \emph{percentage difference} in the distance; negative values denote a decrease in the distance between the marginal distributions.
To validate that the effectiveness of CSA algorithms for classification depends on the satisfaction of their specific conditional CSA conditions, we employ a classification task and we evaluate the percentage difference in accuracy with and without CSA. The use of a relative measure \citep[analogous to percentage drop,][]{Torralba2011} allows us to account for the differences between the CSA systems (such as in the use of labels for adaptation, type of classifiers and implementation details). For SSA+SVM, we report a single result because of its deterministic nature; for the others algorithms involving a random initialization, we report mean and standard error estimated over 10 independent trials.\\
Table \ref{tab:mmd_changes} lists the percentage difference in the MMD distance between the marginal pdfs of the training and test data following the adoption of CSA. These results confirm that both SF and PSF are able to significantly reduce the distance between the marginal pdfs of the training and the test data, thus satisfying the marginal CSA condition.
\begin{table}
\caption{Percentage difference in MMD distance between the training and the test distribution.
$-100\%$ indicates a reduction of the MMD distance of two magnitude orders after applying SF or PSF. \label{tab:mmd_changes}}
\begin{centering}
\begin{tabular}{ccc}
\hline
{} & \textbf{\emph{\footnotesize{SF}}} & \textbf{\emph{\footnotesize{PSF}}}\tabularnewline
\hline
\textbf{\emph{\footnotesize{Radial}}} & $-100.1\% \pm 0.04\%$ & $-50.5 \pm 10.2\%$ \tabularnewline
\textbf{\emph{\footnotesize{Periodic}}} & $-99.7\% \pm 0.01\%$ & $-87.3\%\pm 5.1\%$ \tabularnewline
\textbf{\emph{\footnotesize{Smooth}}} & $-89.8\% \pm 2.3\%$ & $-88.2\% \pm 3.7\%$ \tabularnewline
\textbf{\emph{\footnotesize{Diagonal}}} & $-80.3\% \pm 4.5\%$ & $-84.6\% \pm 4.6\%$ \tabularnewline
\hline
\end{tabular}
\par\end{centering}
\end{table}
\begin{table*}
\caption{Accuracy change when using different CSA systems on the four synthetic data sets. Accuracy without the CSA algorithm to left of the arrow and accuracy with the CSA algorithm to the right of the arrow. \label{tab:accuracy}}
\begin{centering}
\begin{tabular}{cccc}
\hline
{} &
\textbf{\emph{\footnotesize{SF+SVM}}} &
\textbf{\emph{\footnotesize{PSF+SVM}}} &
\textbf{\emph{\footnotesize{IW+LSPC}}}
\tabularnewline
\hline
\textbf{\emph{\footnotesize{Radial}}} & $0.342\rightarrow0.779 \pm 0.03$ & $0.342\rightarrow0.479 \pm 0.05$ & $0.336\rightarrow0.598 \pm 0.06$ \tabularnewline
\textbf{\emph{\footnotesize{Periodic}}} & $0.488\rightarrow0.5 \pm 0.01$ & $0.488\rightarrow0.568 \pm 0.02$ & $0.488\rightarrow0.512 \pm 0.0$ \tabularnewline
\textbf{\emph{\footnotesize{Smooth}}} & $0.894\rightarrow0.476 \pm 0.06$ & $0.894\rightarrow0.522 \pm 0.08$ & $0.865\rightarrow0.936 \pm 0.02$ \tabularnewline
\textbf{\emph{\footnotesize{Diagonal}}} & $0.866\rightarrow0.584 \pm 0.02$ & $0.866\rightarrow0.609 \pm 0.09$ & $0.768\rightarrow0.786 \pm 0.01$ \tabularnewline
\hline
{} &
\textbf{\emph{\footnotesize{SSA+SVM}}} &
\textbf{\emph{\footnotesize{DAE+SVM}}} &
\tabularnewline
\hline
\textbf{\emph{\footnotesize{Radial}}} & $0.342\rightarrow0.342$ & $0.342\rightarrow0.387 \pm 0.05$ \tabularnewline
\textbf{\emph{\footnotesize{Periodic}}} & $0.488 \rightarrow 0.488$ & $0.488\rightarrow0.512 \pm 0.0$ \tabularnewline
\textbf{\emph{\footnotesize{Smooth}}} & $0.89\rightarrow0.89$ & $0.894\rightarrow0.613 \pm 0.02$ \tabularnewline
\textbf{\emph{\footnotesize{Diagonal}}} & $0.86\rightarrow0.932$ & $0.866\rightarrow0.514 \pm 0.0$ \tabularnewline
\hline
\end{tabular}
\par\end{centering}
\end{table*}
\begin{table*}
\caption{Percentage change in accuracy when using different CSA systems on the four synthetic data sets. \label{tab:percentage_increase}}
\begin{centering}
\begin{tabular}{cccc}
\hline
{} &
\textbf{\emph{\footnotesize{SF+SVM}}} &
\textbf{\emph{\footnotesize{PSF+SVM}}} &
\textbf{\emph{\footnotesize{IW+LSPC}}}
\tabularnewline
\hline
\textbf{\emph{\footnotesize{Radial}}} & $\mathbf{+127.8\% \pm 9.4\%}$ & $+40.1\% \pm 14.7\%$ & $+7.8 \pm 2.1\%$ \tabularnewline
\textbf{\emph{\footnotesize{Periodic}}} & $+0.33\% \pm 0.7\%$ & $\mathbf{+16.35\% \pm 5.0\%}$ & $+4.91\%\pm 0\%$ \tabularnewline
\textbf{\emph{\footnotesize{Smooth}}} & $-46.76\% \pm 7.0\%$ & $-41.61\% \pm 9.0\%$ & $\mathbf{+8.36\% \pm 1.3\%}$ \tabularnewline
\textbf{\emph{\footnotesize{Diagonal}}} & $-32.54\% \pm 2.4\%$ & $-29.63\% \pm 10.0\%$ & $+2.74\% \pm 0.7\%$ \tabularnewline
\hline
{} &
\textbf{\emph{\footnotesize{SSA+SVM}}} &
\textbf{\emph{\footnotesize{DAE+SVM}}} &
\tabularnewline
\hline
\textbf{\emph{\footnotesize{Radial}}} & $0\%$ & $+13.16 \pm 14.1\%$ \tabularnewline
\textbf{\emph{\footnotesize{Periodic}}} & $0\%$ & $+4.92 \pm 0.0\%$ \tabularnewline
\textbf{\emph{\footnotesize{Smooth}}} & $0\%$ & $-31.45 \pm 2.6\%$ \tabularnewline
\textbf{\emph{\footnotesize{Diagonal}}} & $\mathbf{+8\%}$ & $-40.65 \pm 0.0\%$ \tabularnewline
\hline
\end{tabular}
\par\end{centering}
\end{table*}
\rev{I added the table with raw performances. This table reports mean and standard error and allows for a more precise comparison of the results.}{Table} \ref{tab:accuracy} reports the raw accuracy of all the CSA classification systems before and after introducing CSA. The results shows that, even if the baselines of the five CSA classification systems are not exactly the same, they are still comparable; therefore, it makes sense to compare the percentage change in accuracy when processing a given data set.
Table \ref{tab:percentage_increase} reports the percentage difference in accuracy for all the CSA systems on the four data sets. These results highlight clearly a direct correlation between the success in classification and the satisfaction of the conditional CSA conditions related to the assumptions on which each algorithm works.
Indeed, the most relevant positive results in Table \ref{tab:percentage_increase} are obtained when a CSA classification system is used on data matching its assumptions. SF provides by far the best improvement on the \emph{radial} data set; as shown in Table \ref{tab:accuracy}, \rev{Referencing the table we added we can now argue that the performance of PSF is relevant because it raises performance above chance level.}{PSF is the only algorithm consistently increasing the performance above chance level} on the \emph{periodic} data set; IW and SSA offers the best improvements on the \emph{smooth} and the \emph{diagonal} data set respectively.
Consistently, violation of the assumptions results in very limited improvement or even decrease in classification accuracy.
For instance, SF yields a negligible increase in accuracy when applied to the \emph{periodic} data set, due to the fact that classification with a linear SVM remains basically at a random guess level before and after CSA, but it causes a severe drop in accuracy when applied to other data sets, due to the assumption mismatch.
In contrast, PSF yields a significant improvement even when applied to the \emph{radial} data set, due to the fact that it is able to extrapolate a periodic structure under the data; however, it no longer works on the \emph{smooth} and the \emph{diagonal} data sets because probably it can not reconstruct a periodic structure from samples coming from a single period.
In general, PSF exhibits also a high variance, suggesting that, despite the supervised guidance, the algorithm is very sensitive to its initialization; different initial weight settings may lead PSF to try to extract very different periodic structures.
Finally, IW and SSA behave in a more conservative way, providing small or no improvements when their assumptions are violated, while DAE works only on the \emph{radial} and \emph{periodic} data sets but severely reduces the accuracy on the other data sets.
This comparative study also suggests that there may be a trade-off between the percentage change in performance and the strictness of the assumptions: the stricter the assumptions, the higher the percentage change, positive if the conditions are met or negative if the conditions are not met, as exemplified by SF in Table \ref{tab:percentage_increase}. On the other hand, if the assumptions are looser, the variation in performance is limited between the case when the conditions are met and the case when the conditions are not met, as illustrated by IW in Table \ref{tab:percentage_increase}. This phenomenon may be well explained in the terms of the no-free lunch theorem \citep{Wolpert1997}.
\subsection{Real-World Data Set \label{sec:Real-World-Data-Experiments}}
We further carried out experiments on real-world data to validate the following statements: (i) a classification system using PSF can provide a statistically significant improvement over the baseline system without CSA, and (ii) PSF can provide competitive performances against other CSA algorithms reviewed in Section \ref{sec:BG_CSA}. \\
Real-world data are often very complex and do not perfectly fit the simple assumptions of the CSA algorithms. In this simulation, we chose emotional speech recognition (ESR) data sets to evaluate PSF for two reasons: (i) it is well known that ESR data sets are affected by covariate shift \citep{Schuller2010}, and (ii) ESR data is user-dependent and hence it may be modeled according to the assumption of PSF, so that each user could be specified by a different pdf $\PDF{p}{X^{user_i}}$ on a specific sub-domain $\domx^{user_i}$, while the conditional distribution of the emotional labels may be approximately the same for all the users $\pygx$.
In the following experiments four well-known ESR data sets are employed: the Berlin Emotional Database (EMODB) \citep{Burkhardt2005}, the Danish Emotional Speech Database (DES) \citep{Engberg2007}, Vera am Mittag (VAM) \citep{Grimm2008} and eNTERFACE (eNT) \citep{Martin2006}. This collection of data sets is very heterogeneous, containing recordings from different speakers, in different languages, with different labels and collected with different protocols (see Table \ref{tab:ESR-datasets} for details). \\
\begin{table*}
\caption{Comparison of ESR data sets.
\emph{\#Pos Val} and \emph{\#Neg Val} refer respectively to the number of samples with positive valence and negative valence. \label{tab:ESR-datasets}}
\begin{centering}
\begin{tabular}{ccccc}
\hline
{} &
\textbf{\emph{\footnotesize{\#Speakers}}} &
\textbf{\emph{\footnotesize{Language}}} &
\textbf{\emph{\footnotesize{Recording}}} &
\textbf{\emph{\footnotesize{Labelling}}}
\tabularnewline
\hline
\textbf{\emph{\footnotesize{EMODB}}} & 10 & German & acted & discrete (7 classes) \tabularnewline
\textbf{\emph{\footnotesize{DES}}} & 4 & Danish & acted & discrete (5 classes) \tabularnewline
\textbf{\emph{\footnotesize{VAM}}} & 47 & German & natural & continuous (3 dimensions) \tabularnewline
\textbf{\emph{\footnotesize{eNT}}} & 43 & English & induced & discrete (6 classes) \tabularnewline
\hline
{} &
\textbf{\emph{\footnotesize{\#Samples (1-sec)}}} &
\textbf{\emph{\footnotesize{\#Pos Val}}} &
\textbf{\emph{\footnotesize{\#Neg Val}}}
\tabularnewline
\hline
\textbf{\emph{\footnotesize{EMODB}}} & 1211 & 289 & 922 \tabularnewline
\textbf{\emph{\footnotesize{DES}}} & 974 & 579 & 395 \tabularnewline
\textbf{\emph{\footnotesize{VAM}}} & 2495 & 167 & 2328 \tabularnewline
\textbf{\emph{\footnotesize{eNT}}} & 2988 & 877 & 2111 \tabularnewline
\hline
\end{tabular}
\par\end{centering}
\end{table*}
All the recordings are pre-processed into one-second segments and converted into 72-dimensional feature vectors based on Mel-frequency cepstrum coefficients (MFCC) \citep{Eyben2010a}.
All the labels are aligned using a standard procedure that reduces the heterogeneous labels to binary labels denoting a state of positive or negative valence (see Appendix \ref{app:Real-World} for details on feature extraction and label alignment).
We decided to use only speakers from EMODB, DES and eNTERFACE when performing CSA, excluding VAM because of the high unbalance between the valence classes (see Table \ref{tab:ESR-datasets}). We partition the data using the following protocol: in each trial, one speaker from either EMODB, DES or eNTERFACE is randomly selected; half of the samples from the selected speaker constitute the target set $\Xta$, while the remaining half constitutes the test set $\Xte$; all the samples from the three remaining data sets constitute the training set $\Xtr$. Adaptation is performed on the training and target data; classification uses training data for learning, target data for model selection and test data for evaluation. This protocol has two advantages: (i) in line with the most challenging scenarios in literature, this is a dataset-out and speaker-out scenario, in which the training set does not contain samples from the same data set or the same speaker in the test set, and (ii) preserving part of the samples only for test allows us to properly evaluate the degree of inductive generalization.\\
\rev{Here I provided more details about CV in general, and about SF and PSF in particular. Now all that is left in the appendix are just the numbers and the tables; no explanation or detail is left in the appendix. For readability, I left the tables in the appendix}{All the data are processed using the same classification systems used in the previous experiment. Hyper-parameters for the five classification systems are selected by cross-validation using a standard grid search method.} Details about the grid space for the hyper-parameters and the actual values selected through cross-validation are provided in Appendix \ref{app:Real-World}.
Moreover, to decrease the number of hyper-parameters, to reduce the computational time and to improve the results, we implemented an early stopping criterion for SF and PSF relying on the Kolmogorov-Smirnov (KS) test. The KS test is a statistical test assessing whether two univariate distributions are the same. Following \citet{Hassan2013}, we apply this test to pair of features to derive a gross estimation of the distance\footnote{This estimation is efficient, but clearly inaccurate as the distance between all the univariate distributions of $\PDF{p}{X_{1,i}}$ and $\PDF{p}{X_{2,i}}$ is not the same as the distance between the multivariate distributions $\PDF{p}{X_{1}}$ and $\PDF{p}{X_{2}}$.} between the distribution of the training $\mathbf{X^{tr}}$ and target data $\mathbf{X^{tar}}$. Given two samples $\mathbf{x}_{1}^\mathbf{tr}$ and $\mathbf{x}_{2}^\mathbf{tar}$ we apply the KS test feature-by-feature $KS\left(x_{1,j}^\mathbf{tr},x_{2,j}^\mathbf{tar}\right)$, and we average over all the feature $\mathbb{E}_j\left[KS\left(x_{1,j}^\mathbf{tr},x_{2,j}^\mathbf{tar}\right)\right]$. We stop training when the learned distribution of the training $\mathbf{Z^{tr}}$ and target data $\mathbf{Z^{tar}}$ achieve a minimum in the KS distance over the first 50 iterations.\\
In order to evaluate the performance in this unbalanced classification problem we employ the unweighed average recall (UAR): $\frac{1}{C}\sum_{c=1}^{C}recall(c)$, where $recall(c)$ denotes the recall for class $c$, computed as the number of correctly identified instances of class $c$ divided by all the instances of class $c$ \citep{Batliner2010}.
For each configuration, we report the mean and the standard error achieved over 10 independent simulations.
In order to validate the contribution of PSF to classification, we perform a paired Wilcoxon test with the null hypothesis that the classification performance with and without CSA has the same mean (under the assumption that the results are symmetrically distributed around the true mean performance). Statistics for the hypothesis test are collected from 100 independent trials.\\
\begin{figure}
\begin{centering}
\includegraphics[scale=0.5]{UAR}
\par\end{centering}
\caption{UAR Accuracies of different CSA methods along with the baselines. \label{fig:UAR}}
\end{figure}
Figure \ref{fig:UAR} shows the UAR of different classification systems on the three data sets using the protocol described above. After running the Wilcoxon test, we could reject the null hypothesis that classification with and without PSF is equivalent in the case of EMODB and eNTERFACE ($p$-value, respectively, $9.5\cdot10^{-5}$ and $6.9\cdot10^{-4}$), but not in the case of DES (at $p$-value $0.05$). Thus, the statistical test implies that PSF was indeed able to capture some relevant periodic structure in the conditional distribution when applied to the EMODB and eNTERFACE data sets.
In general, the experimental results suggest that PSF is able to provide performances better or comparable to those of the other CSA algorithms, even if it is outperformed by DAE on a specific data set (EMODB).
On the other hand, the SF algorithm failed to improve over the baseline. This was not surprising, as it can be easily explained by the fact that real-world data sets are too complex to comply with the tight assumption of a radial data structure.
Of the other CSA algorithms, DAE provided the best results with the highest UAR on EMODB and comparable UAR to other classification systems on the remaining data sets. Interestingly, IW yields good results on EMODB and eNTERFACE, but failed to provide an improvement on DES, potentially because DES speaker data may lie in sub-domains removed from the other data sets. SSA, instead, performed well on DES and eNTERFACE, but performed the worst on EMODB; this could hint at the fact that the PCA components of the EMODB test speaker data are not easily projected on the PCA components of all the other speakers.
\section{Conclusions \label{sec:Conclusions}}
In this paper, we have studied the use of SF-like algorithms for CSA. We have rigorously showed that SF can perform CSA under the strict condition of a conditional distribution $\pygx$ explained by cosine neighbourhoodness. To overcome this limitation, we have then proposed the new PSF algorithm, which is able to perform CSA under looser and more realistic conditions on $\pygx$; indeed, while SF works only when the conditional distribution has a fixed radial structure, PSF can perform CSA adapting to data sets having different forms of periodic structure. We have demonstrated the strengths and limitations of SF and PSF on both synthetic and real-world data and we have compared their performances with other standard CSA algorithms. Experimental results have confirmed our theoretical conclusions and showed that PSF could be a competitive and efficient candidate for CSA, particularly when the structure of the conditional distribution is periodic. Moreover, our simulations have showed that by making the conditions for CSA explicit, we can not only predict when a CSA algorithm is likely to work, but we can also gain useful insights into the structure of the data even if a CSA algorithm were to fail.\\
Nevertheless, some non-trivial yet interesting points are left out in our current work and will be investigated in our ongoing research.
From a theoretical viewpoint, we have not looked into the \emph{model misspecification assumption} discussed in \citet{Shimodaira2000}. The model misspecification assumption refers to a situation when the model employed to explain the data does not perfectly fit the data; this is normally due to the learning of a model from a family of simple parametric models that does not include the true model that generated the data \citep{Quinonero-Candela2009}. This situation may be explicitly stated for the algorithms in our simulations, too. Thus, future work would be directed at the model misspecification assumption in order to study theoretical bounds on the empirical risk when using SF and PSF.
From a practical viewpoint, it is also interesting to study the extension of our algorithms to the \emph{semi-supervised setting} \citep{Chapelle2006}. While we used SF and PSF in a purely unsupervised or supervised setting, our PSF has a potential to work in a semi-supervised way by processing unlabelled target samples alongside labelled source samples. In future work, we would study the dynamics of PSF in a semi-supervised scenario and compare it against other semi-supervised algorithms designed to perform CSA \citep{Margolis2011,Jiang2008}.
Our results provide a deep understanding of the potential of SF and PSF in performing CSA. This understanding may be generalized to feature distribution learning algorithms and it could be further exploited to develop novel feature distribution learning algorithms with high computational efficiency. Reasoning in terms of the marginal and conditional CSA conditions, it may be possible to design new SF-like algorithms making explicit assumptions about the conditional distribution of the data. Given prior knowledge about the structure of a data set, this knowledge may be directly integrated to develop a SF variant tailored to the data; without explicit prior knowledge, metric learning algorithms could instead be employed or integrated in SF-based algorithms to indirectly learn the conditional structure of the data.
\begin{acknowledgements}
F. M. Zennaro's work was supported by the Kilburn PhD studentship.
\end{acknowledgements}
|
\section{Introduction}
\noindent
In the last decade, there has been an increasing interest in macroscopic anomaly-driven effects. To date, several macroscopic phenomena have been shown to be related to the existence of quantum anomalies in the underlying microscopic theory. This includes, among others, the chiral magnetic (vortical) effect \cite{Fukushima:2008xe, Son:2009tf} which consists of the generation of an electromagnetic current in the direction of an external magnetic field (vortex), a massless propagating mode known as the chiral magnetic wave \cite{Kharzeev:2010gd}, and the enhancement of the electric conductivity in presence of parallel electromagnetic fields known as negative magnetoresistance~\cite{Nielsen:1983rb}.
From the phenomenological point of view, the need for massless fermions strongly restricts the variety of experiments where these effects might be verified. To date the two main test grounds for anomalous transport are the state of matter known as quark gluon plasma (QGP) and Weyl/Dirac semimetals. In heavy ion collisions (HIC) the QGP is generated for $\sim 10^{-24}s$. In HIC very strong and transient magnetic fields are generated. Due to the high temperatures reached in the QGP, it is reasonable to neglect the quark masses. Moreover, experimental evidence \cite{Heinz:2008tv} shows that the QGP is, despite the high temperatures, strongly coupled.
The second kind of systems in consideration is of a very different nature. Weyl (Dirac) semimetals are solid state crystals. These present effective Weyl (Dirac) fermionic degrees of freedom in concrete points of the Brioullin zone known as Weyl nodes. The main difference between Dirac and Weyl semimetals is that in the former the nodes for opposite chiralities coincide in momentum space while in the latter these are separated. Remarkably, one of the aforementioned effects, the ``negative magnetoresistance'', has already been measured in several Weyl and Dirac semimetals \cite{Li:2014bha, Li2015, 2015PhRvX...5c1023H, 2016NatCo...710301L}. Although it is usually said that Weyl semimetals are the 3D graphene, it is not yet clear whether the former are strongly coupled too. Therefore it is of special interest to study these systems from a strong coupling perspective. In this sense the gauge/gravity duality \cite{Maldacena:1997re} (for textbooks see \cite{Ammon:2015wua,Nastase,Schalm}) is a specially adequate theoretical tool to study these processes.
A complete study of anomalous transport effects in both HIC and condensed matter systems requires to address the question of how the system behaves out of equilibrium. Several studies in this direction include computations for fixed axial charge and dynamic magnetic fields at weak coupling to leading order in $\alpha_s$ \cite{Kharzeev:2009pj, Jimenez-Alba:2015bia} and at strong coupling via holography \cite{Yee:2009vw, Landsteiner:2013aba}. The fate of the chiral magnetic effect (CME) in presence of time-dependent external electric and magnetic fields has been studied at weak coupling for concrete configurations of the electromagnetic fields \cite{Fukushima:2015tza, Iwazaki:2009bg}. Moreover, the authors of \cite{Lin:2013sga} studied the evolution of the CME during thermalization in holography. It is desirable to have a complete description of anomalous transport for realistic systems.
Most of the past effort in this direction had its focus on the phenomenology of the QGP in HIC. However, the appearance of Weyl/Dirac semimetals has affected the order of relevance of the different features to be addressed. An example of this is the axial charge time-dependence. In the field of QGP and HIC it has been argued \cite{Lin:2013sga, Kharzeev:2001ev, Hegde:2008rm} that the estimated decay rate of axial charge is of the order of the lifetime of the plasma and therefore as a first approximation the charge can be taken to be constant. In ``chiral'' condensed matter system, however, the story is quite different.
Indeed, it is known that in these systems, the chiral magnetic current vanishes in equilibrium even at very large temperatures \cite{Basar:2013iaa}. This is due to the fact that the Weyl nodes are filled up to the same energy (the Fermi energy) and that the Nielsen-Ninomiya theorem \cite{nino} applies. For materials with the Weyl nodes placed at the same energy, one can generate a net chiral magnetic current by inducing a finite axial chemical potential. This is achieved by the non-conservation of the axial charge density in the presence of parallel electric and magnetic fields. Therefore, the mechanism studied here can be used to generate the CME in Weyl semimetals.\footnote{One can also consider shifting the relative energy of the Weyl nodes \cite{Cortijo:2016wnf}}
Motivated by these ideas, we focus in this paper on the time evolution of the CME in presence of time dependent axial charge. In order to introduce axial charge in the system in a time dependent fashion, we consider the anomaly expression in presence of external electromagnetic fields and in absence of axial currents
\begin{equation}\label{eq:anomH}
\dot{\rho}_5 \sim \mathbf{E\cdot B}\,.
\end{equation
%
This equation tells us that the evolution of the axial charge is determined by the time dependence of $\mathbf{E\cdot B}$. Concretely, we will consider a static magnetic field and a dynamic electric field.
The possible strongly coupled nature of Dirac/Weyl semimetals motivates us to apply holography. This is not, however, the only reason to consider holography for this topic. As extensively discussed in the literature, holography is nicely suited to perform real time computations which, by means of the holographic dictionary, reduce to solving the bulk system of PDEs. As shown later, our case can be reduced to one hyperbolic PDE describing a quenched current operator in presence of a magnetic field. This connects the initial motivation of this work to the topic of (quantum) quenches in holography.
In recent years, there has been a lot of interest in studying (quantum) quenches due to new experimental results. From the theoretical point of view, we have to
compute the (quantum) evolution of an isolated system in the presence of a time-dependent parameter in the Hamiltonian. As explained above, for strongly coupled quantum field theories, gauge/gravity dualities are an ideal playground to tackle such questions. One remarkable outcome of previous studies of quenches in holography is the appearance of universal scaling behaviours in the ``early time'' response \cite{Buchel:2013lla} for fast quenches. We explore these ideas within our system and determine how the anomaly affects them.
In order to study all these features, it is in principle necessary to solve the initial-boundary value problem i.e. solving PDEs. Indeed we explicitly solve the equations by means of pseudospectral methods. However, for linear systems there is a way out of this. Based on several works from the numerical GR community in flat space \cite{PhysRevD.45.2617, Nollert:1998ys, Ansorg:2016ztf}, we are able to numerically compute the residue of the QNMs of the system by just solving the ODEs that result from Laplace transforming the PDEs. In other words, one does not need to solve the explicit time dependent problem to obtain the residues. As shown in \cite{Ansorg:2016ztf}, generically, this information allows to determine the full response only from an initial time scale $\tau$ on. $\tau$ is obtained from the amplitudes and therefore depends not only on the system under consideration but on the initial and boundary conditions as well. When this applies, it means that the knowledge of \textit{all} QNM and their residues is not enough to determine the response of the system at initial times. We explore the applicability of these ideas in AdS and give a physical interpretation to $\tau$ from the dual theory point of view.
The paper is structured as follows. In section \ref{sec:model}, we present and discuss our holographic model and the ansatz that we use for the different fields. This section includes some discussion regarding the 1-point function renormalization, gauge fixing and the definition of the current. In section \ref{sec:rodrigo}, we adapt the spectral decomposition analysis developed for asymptotically flat Schwarzschild in \cite{Ansorg:2016ztf} to AdS-Schwarzschild. Concretely, we explicit compute the amplitudes of the QNMs in our system for different boundary data. In section \ref{sec:initial}, we focus on the initial time behaviour of the current. First we look at the dependence of $\tau$ on the different parameters of the system. After this, following \cite{Buchel:2013lla}, we look at the response of the current for fast quenches and its dependence on the anomaly coefficient and the magnetic field. Finally, in section \ref{sec:late} we study the late time response where we identify QNMs approaching the real axis with increasing magnetic field. This modes are responsible for a long lived oscillatory late time behaviour of the current and for a resonant behaviour under oscillatory sources. By computing the QNMs with backreaction, we argue that this is indeed a consequence of the axial anomaly in our class of holographic models. Moreover, we relate the QNMs to resonances due to Landau levels. We finish this work in section \ref{sec:discussion} with a discussion of the results and some hints on possible future directions.
\section{The model}\label{sec:model}
\noindent
In this work we consider a $U(1)\times U(1)$ model, which has been previously applied to study anomalous transport in holography \cite{Gynther:2010ed, Jimenez-Alba:2014iia, Jimenez-Alba:2014pea}. None of these studies, however, has considered the effect that dynamical generation of axial charge may have in the response. Our aim here is to compute time dependent setups for the CME in holography. Concretely we focus on the evolution of the effect in presence of electric field driven axial charge generation. This is in contrast to the usual ``by hand'' implementation of net charge via (axial) chemical potential \cite{ Yee:2009vw, Gynther:2010ed}.
The 5-dimensional matter Lagrangian consists of two photon fields $(A_\mu, V_\mu)$ coupled by a Chern-Simons (CS) term in the bulk
\begin{equation}
\mathcal{L}= -\frac{1}{4}F^{\mu\nu}F_{\mu\nu}-\frac{1}{4}H^{\mu\nu}H_{\mu\nu}+ \frac{\kappa}{2}\varepsilon^{\mu\alpha\beta\rho\lambda}A_\mu\left(F_{\alpha\beta}F_{\rho\lambda}+
3H_{\alpha\beta}H_{\rho\lambda}\right)\,,
\end{equation}%
with $F=\mathrm{d} A$, $H=\mathrm{d} V$. Here, bulk indices are denoted by Greek letters running from $0$ to $4$. Computing the divergence of the dual current operators one can show that, on-shell
\begin{equation}\label{eq:ward}
\langle\partial_i J^i_V \rangle=0\,,\hspace{2cm}\langle\partial_i J^i_A \rangle=\frac{\kappa}{2}\tilde\varepsilon^{\,ijkl}\left( F_{ij}F_{kl}+3H_{ij}H_{kl} \right),
\end{equation
with $\tilde\varepsilon^{\,ijkl}$ being the epsilon symbol in the boundary theory and Latin indices running from $0$ to $3$. These Ward identities correspond to an abelian anomaly. The Bardeen counter term has been chosen such that only one of the currents is not conserved. It is therefore customary to identify $J_V$ and $J_A$ with the dual vector and axial currents, respectively. This allows us to meaningfully consider external ``electromagnetic'' fields and study the evolution of the system in such backgrounds. Moreover, the relative factor between both CS terms is chosen so that the relative factor in the r.h.s.~of equation \eqref{eq:ward} coincides with that of the abelian part of the anomaly in QCD \cite{Bertlmann:1996xk}.
As a first step towards solving the full non-linear system we consider the probe limit of the system. This limits the validity of the results. We comment on this later, when the results are discussed. As background metric we consider Schwarzschild-AdS$_5$. In infalling Eddington-Finkelstein coordinates the metric reads
\begin{equation}
\mathrm{d} s^2=\frac{1}{\rho^2}\,\left(-f(\rho)\mathrm{d} v^2-2\,\mathrm{d} v \mathrm{d}\rho +\mathrm{d} x^2+\mathrm{d} y^2+\mathrm{d} z^2\right),\hspace{2cm}f(\rho)=1-\rho^4\,.
\end{equation}%
Note that the conformal boundary is located at $\rho=0$ and without loss of generality we have rescaled the coordinates such that the AdS radius is $L=1$ and the black-hole horizon is at $\rho=1$.
The equations of motion for the gauge fields read
\begin{align}
\nabla_\mu F^{\mu \nu}+\frac{3\kappa}{2}\varepsilon^{\nu\alpha\beta\rho\lambda}
\left(F_{\alpha\beta}F_{\rho\lambda}+H_{\alpha\beta}H_{\rho\lambda}\right)=0\,,
\nabla_\mu H^{\mu \nu}+3\kappa\,\varepsilon^{\nu\alpha\beta\rho\lambda}
F_{\alpha\beta}H_{\rho\lambda}=0\,.
\end{align}
\subsection{Setup}\noindent
As explained in the introduction, in order to control the dynamic generation of axial charge we need a time dependent $\mathbf{E\cdot B}$. We keep the magnetic field static and just set a time dependent electric field. Therefore, we consider the following ansatz
\begin{align}\label{eq:ansatz}
A_v(v, \rho), \hspace{1cm} V_y(x)=Bx, \hspace{1cm} V_z(v, \rho),
\end{align}
with boundary condition
\begin{align}\label{eq:conditions}
\dot V_z(v, \rho\rightarrow 0)=E(v)\,,
\end{align}
and the usual regularity conditions for $A_v(v,\rho)$ at the horizon. Here, the time derivate $\partial_v$ is represented by the dot. This corresponds to a time independent magnetic field and time dependent electric field both homogeneous and pointing in the $z$-direction in the boundary theory. Due to the anomaly, a non-trivial time component of the axial gauge field is needed to get equations for $V_z$ and $V_y$ consistent with the choices in equation \eqref{eq:ansatz}. The equations for this ansatz read
\begin{align}
A_v''-\frac{1}{\rho}A_v'- 12\kappa B V_z' \rho=0\,,\label{eq:eom1}\\
V_z''+ \left(\frac{f'}{f}-\frac{1}{\rho} \right)V_z' -\frac{2}{f}\dot V_z '+\frac{1}{\rho f}\dot V_z-12 \kappa
B \frac{\rho}{f} A_v'=0\,,\label{eq:eom2}\\\
\dot A_v'-12 \kappa B \rho \dot V_z=0\,,\label{eq:eom3}
\end{align}
with the notation $'$ for the radial derivative $\partial_\rho$. Integrating equation~(\ref{eq:eom3}) in time one obtains
\begin{equation}\label{eq:av}
A_v'=12 \kappa B \rho V_z+C_1(\rho)\,.
\end{equation
Substituting this result back into equation \eqref{eq:eom1}, one finds $C_1(\rho)=\rho C$. We will later see in equation \eqref{eq:regularsolultions} that $C$ is just a gauge shift for $V_z$ and will fix it to $C=0$. We can then reduce the system of PDEs to a single hyperbolic PDE for the $z-$component of the vector field
\begin{align}\label{eq:vzeq}
V_z''+ \left(\frac{f'}{f}-\frac{1}{\rho} \right)V_z' -\frac{2}{f}\dot V_z '+\frac{1}{\rho f}\dot V_z-\left(12 \kappa
B \rho\right)^2\frac{1}{f} V_z-12 \kappa B \rho^2\frac{C}{f} =0\,.
\end{align}
%
The asymptotic boundary expansions read
\begin{align}\label{eq:expansion}
V_z&\sim V_{0}+\rho\, \dot V_{0} + \rho^2\,\tilde V +\frac{1}{2} \rho^2\,\log(\rho) \ddot V_{0}+ \mathcal{O}(\rho^3),\\
A_v&\sim A_{0}+\rho^2 \tilde A + \mathcal{O}(\rho^3).
\end{align}
In order to compute the dual correlators we need to specify boundary terms. We have
\begin{equation}
-\int_{\partial \mathcal{M}} \mathrm{d}^4x\, \sqrt{-\gamma}\, \frac{1}{4} F_{ij}F^{ij} \log(\rho)-\int_{\partial \mathcal{M}}\mathrm{d}^4x\,\sqrt{-\gamma}\, \frac{1}{4} H_{ij}H^{ij} \left( \log(\rho)-\frac{1}{2} \right),
\end{equation
with $\gamma$ the induced metric on the boundary. The first term is an infinite counter-term fixed by the theory. In addition, for convenience in our numerical calculations, we fix the renormalization scheme by introducing the second (finite) gauge invariant counter term. We make this choice to avoid explicit contributions of the logarithmic coefficient in \eqref{eq:expansion} to the one point function which in our scheme is given by
\begin{equation}\label{eq:1pcons}
\langle J_z^V\rangle_\text{cons}=2 \tilde V - 12\kappa B A_{0}\,,
\end{equation
where the subscript stands for ``consistent'' current. From previous studies of the CME in holography it is known that this version of the current can be problematic \cite{Yee:2009vw, Gynther:2010ed} due to the difficulty to distinguish between the chemical potential and the dual coupling associated to $A_v$ in certain gauge. It is now well understood \cite{Gynther:2010ed} that imposing a vanishing $A_v$ at the horizon gives rise to a vanishing \textit{consistent} current in the time independent limit. One way to circumvent these issues is to compute the \textit{covariant} current instead
\begin{equation}
\langle J_z^V\rangle_\text{cov}=\langle J_z^V\rangle_\text{cons} + 12\kappa B A_{0}\,.
\end{equation
%
In Eddington-Finkelstein coordinates, however, regularity does not imply a vanishing gauge field at the horizon. Therefore we are free to choose a gauge in which the field vanishes at the boundary $A_v(\rho=0)=A_0=0$. With this choice the covariant and consistent currents are indistinguishable
\begin{equation}
\label{eq:Current}
\langle J_z \rangle\equiv\langle J_z^V \rangle= 2 \tilde V\,.
\end{equation}
From now on we stick to this choice and omit the ``vector'' superscript. Then, our problem reduces to computing the coefficient of the normalisable mode of the $V_z$ field. To this end, we must solve the dynamical equation~\eqref{eq:vzeq} with the boundary condition \eqref{eq:conditions} and additional initial data.
In order to construct the initial data, we assume that our system is initially in equilibrium, i.e. described by a stationary configuration. Mathematically, this assumption implies that all time derivatives of the boundary data $V_0(v)$ must vanish at $v=0$. In practice, this condition is hardly realised. Nonetheless, as we are going to mention afterwards, we choose specific quenches profiles for the function $V_0(v)$ in such a way that
$\dfrac{\mathrm{d}^j}{\mathrm{d} v^j} V_0(0) \sim 0$.
The time-independent equation \eqref{eq:vzeq} can be rewritten in terms of a new variable $u\equiv \rho^2$ as an inhomogeneous Legendre equation
\begin{align}\label{eq:leg}
(1-u^2) V_z''- 2 u V_z' -36 \left( \kappa B \right)^2 V_z= 3 \kappa
B \,C\,,
\end{align}
%
with regular solutions
\begin{equation}\label{eq:regularsolultions}
V_z=C_2\mathcal{P}_l(\rho^2)-\frac{C}{12 \kappa B}\hspace{2cm} l=\frac{1}{2}\left(-1+\sqrt{1-144\,(\kappa B)^2} \right)\,.
\end{equation}%
Here, $\mathcal{P}_l$ are the l-th Legendre functions of the first kind. The constant $C$ is just a shift in the field. It becomes clear that $A_v'$ in equation~\eqref{eq:av} is $C$ independent. We will from now on gauge fix it to $C=0$. The constant $C_2$ determines the initial value of the current and it can be related to the initial value of the axial chemical potential
\begin{equation}
\mu_A\equiv A_v(\rho=0)-A_v(\rho=1)=12\kappa B C_2 \int^0_1 \mathrm{d}\rho \,\mathcal{P}_l(\rho^2)\,\rho\,.
\end{equation}%
Since the adiabatic regime implies that the equilibrium formula is valid at any time, the adiabatic response is given by the time independent solution. The asymptotic expansion of the stationary solutions in equation \eqref{eq:regularsolultions} read
\begin{equation}\label{eq:legenderexpansion}
V_z\sim V_0- 18(\kappa B)^2\frac{\Gamma\left(\frac{1}{2}(l
+2) \right) \Gamma\left(\frac{1}{2}(1-l) \right)}{\Gamma\left(\frac{1}{2}(l
+3) \right)\Gamma\left(\frac{1}{2}(2-l) \right)}V_0 \rho^2+\mathcal{O}(\rho^3)\,,
\end{equation}%
with $\Gamma$ the Euler gamma functions. Therefore, in the adiabatic limit
\begin{equation}\label{eq:adiabatic}
\langle J_z \rangle= -36(\kappa B)^2\frac{\Gamma\left(\frac{1}{2}(l
+2) \right) \Gamma\left(\frac{1}{2}(1-l) \right)}{\Gamma\left(\frac{1}{2}(l
+3) \right)\Gamma\left(\frac{1}{2}(2-l) \right)} V_0(v)\, .
\end{equation}%
Moreover, for $\kappa B\gg1$ it asymptotes as
\begin{equation}\label{eq:adiabaticJ}
\langle J_z \rangle\sim -12 \kappa B V_0(v)\,.
\end{equation}%
%
It is possible to understand the above equation in terms of the effect of the Chiral Magnetic Wave. If a system displays a massless mode with a finite residue $R_0$, the response to an externally applied electric field $E(v) =\dot V_0(v)$ has a contribution of the form $\langle J(\omega) \rangle= R_0 E(\omega)/\omega$ in Fourier space. Hence the D.C. conductivity is infinite if $R_0(\omega=0) \ne 0$. For a given $V_0(v)$, it is then straightforward to see that $\langle J(v) \rangle= R_0 V_0(v)$. In our situation, $R_0 = -12 \kappa B$ for large $\kappa B$ and thus at low frequencies the D.C. conductivity scales as $\sigma(\omega) \sim - 12 \kappa B\,\delta (\omega)$.
In order to compute the current outside the adiabatic limit, we make use of the asymptotic expansion \eqref{eq:expansion} and introduce an auxiliary field $U(v,\rho)$ via
\begin{equation}
\label{eq:AuxField_U}
V_z(v,\rho) \equiv V_{0}(v)+\rho\, \dot V_{0}(v) + \rho^2\,U(v,\rho) +\frac{1}{2}\rho^2\,\log(\rho) \left(\ddot V_{0}(v)+\rho \dddot V_{0}(v) \right).
\end{equation}%
Substituting equation~\eqref{eq:AuxField_U} into the original equation \eqref{eq:vzeq} we obtain
\begin{equation}
\label{eq:DynEq_TimeDom}
\left[ - \rho\,(1-\rho^4)\, \frac{\partial^2 }{\partial \rho^2 } - \left( 3 - 7\rho^4\right)\, \frac{\partial}{\partial \rho} + (8 + \lambda^2)\,\rho^3 \right]U(v,\rho) + \left[ 2\,\rho\, \frac{\partial }{\partial \rho } + 3\ \right] \dot{U}(v,\rho) + S(v,\rho) = 0.
\end{equation}
with $\lambda = 12\,\kappa B$. By doing this, we automatically incorporate the information about the boundary condition \eqref{eq:conditions} into the inhomogeneity $S(v,\rho)$ given by
\begin{equation}
S(v,\rho) = a_4(\rho) \frac{\mathrm{d}^4}{\mathrm{d} v^4} V_0(v) + a_3(\rho) \frac{\mathrm{d}^3}{\mathrm{d} v^3} V_0(v) + a_2(\rho) \frac{\mathrm{d}^2}{\mathrm{d} v^2} V_0(v) + a_1(\rho) \frac{\mathrm{d}}{\mathrm{d} v} V_0(v) + a_0(\rho) V_0(v),
\end{equation}
and
\begin{eqnarray}
\label{eq:Source_ak}
a_4(\rho) &=& \frac{1}{2}\left(2\rho + 5\rho\log(\rho) \right), \quad a_3(\rho) = \frac{15}{2}\rho^4\log(\rho)+ \frac{\lambda^2}{2}\rho^4\log(\rho)+4\rho^4-1 \notag \\
a_2(\rho) &=& \rho^3\left(4\log(\rho)+3+\frac{\lambda^2}{2}\log(\rho) \right), \quad
a_1(\rho) = \left(3+\lambda^2\right)\rho^2, \quad a_0(\rho) = \lambda^2\,\rho\,.
\end{eqnarray}
Therefore, equation~\eqref{eq:DynEq_TimeDom} is to be solved as an initial value problem with
\begin{equation}
\label{eq:ID_U}
U(0,\rho) \equiv U_{\rm in}(\rho) = \frac{V_0}{\rho^2}\left[ \frac{P_l(\rho^2)}{P_l(0)} - 1\right].
\end{equation}
The most natural way to solve~\eqref{eq:DynEq_TimeDom} is to numerically integrate it in time. In this work, the time evolution is performed with the fully spectral code\footnote{In order to double check our results, we also evolve the equations with a Crank-Nicolson time integrator while keeping the spectral method in the spatial direction.} introduced in \cite{Macedo:2014bfa} (the numerical details are described in appendix \ref{sec:methods}). We use this highly accurate numerical method to obtain the full response of the current operator. In particular, it provides us with a reference solution, against which we can compare the results from the strategy based on the spectral decomposition of the solution $U(v,\rho)$ in terms of the quasi-normal modes.
\section{QNM amplitudes from Laplace analysis in AdS}\label{sec:rodrigo}
As stated in the introduction, solving the PDE is not the only way to determine the response of the current. Given the spectral decomposition for the dual operator \begin{equation}\label{eq:spectraldec}
\langle J_z(v) \rangle=\sum_{n=0}^\infty A_n \operatorname{e}^{-\mathrm{i}\omega_n v}\, ,
\end{equation}%
solving the problem reduces to obtaining the eigenfrequencies $\omega_n$ and the amplitudes $A_n$ for given initial and boundary data. It is well known how to obtain the quasi-normal frequencies in holography \cite{Kovtun:2005ev}. However this is not the case for the amplitudes $A_n$. Of course a possibility is to fit the data obtained from the PDE to the previous formula (see for example \cite{Heller:2013oxa}). It is, nevertheless, possible to compute $A_n$ directly without need of explicitly solving the PDEs. Although this topic has a long history in the community of numerical general relativity \cite{PhysRevD.45.2617, Nollert:1998ys} only recently \cite{Ansorg:2016ztf} has it been shown how to compute these amplitudes in the case of asymptotically flat Schwarzschild black hole for rather generic initial data.
More specifically, the authors formulate the wave-equation describing the propagation of fields on the Schwarzschild background in terms of a particular coordinate system, where the surfaces of constant time are spacelike hypersurfaces that penetrate the black-hole horizon and extend to future null infinity. By using the standard framework provided by the Laplace-transformation, they develop a semi-analytical algorithm to obtain ``eigenvalues" and ``eigenvectors" (related only to the wave-equation in question) and amplitudes (related to the particular initial data being used). Most importantly, they introduce a well defined time scale $\tau$ for which the spectral decomposition in the form \eqref{eq:spectraldec} is valid, based on the growth rate of such amplitudes.
In this section, we provide the first steps towards the application of these techniques in asymptotically AdS and we focus on the role that the $\tau$ may have in the dual theory.
\subsection{Laplace transformation}
We consider the compact form of the dynamical equation \eqref{eq:DynEq_TimeDom} for the function $U(v,\rho)$
\begin{equation}
{\boldsymbol \alpha} [U] + {\boldsymbol \beta} \left[ \frac{\partial} {\partial v} U\right] + S = 0,
\end{equation}
with given initial data $U_{\rm in}(\rho)$ specified in \eqref{eq:ID_U}. Here, ${\boldsymbol \alpha(\rho)}$ and ${\boldsymbol \beta(\rho)}$ are differential operators acting on the radial coordinate $\rho$, which can be easily read from \eqref{eq:DynEq_TimeDom}. The former is second order, whereas the latter is first order. As mentioned, the source term $S(v,\rho)$ contains information about the boundary data $V_0(v)$ and is generically written in the form
\begin{equation}
S(v,\rho) = \sum_{i=0}^{N_S} a_i(\rho) \frac{\mathrm{d}^i}{\mathrm{d} v^i} V_0(v).
\end{equation}
Here, $N_S=4$ and the functions $a_i(\rho)$ are summarised in equations~\eqref{eq:Source_ak}. Applying the Laplace transformation\footnote{Note that the parameter $s$ of the Laplace transformation is related to the Fourier parameter by the relation $s=-\mathrm{i} \omega$.}
\begin{equation}
\bar{U}(\rho;s) = {\cal L}[U(v,\rho)](s) = \int\limits_0^{\infty} \\\mathrm{d} v \, U(v,\rho) \operatorname{e}^{-sv}
\end{equation}
and taking into account that for a generic function $f(v,\rho)$
\begin{equation}
{\cal L}\left[\frac{\partial^{n}}{\partial v^n}f(v,\rho)\right](s) = s^n\,\bar{f}(\rho;s) - \sum_{k=0}^{n-1} s^{n-k-1} \frac{\partial^{k}}{\partial v^k}f(0,\rho),
\end{equation}
we obtain
\begin{eqnarray}
{\boldsymbol \alpha} [ \bar{U}] &+& s\, {\boldsymbol \beta} [ \bar{U}] = {\boldsymbol \beta} [ U_{\rm in}] - \bar{S} \quad {\rm with} \label{eq:LapTransEq}\\
\bar{S}(\rho;s) &=& \bar{V}_0(s) \sum_{i=0}^{N_S} a_i(\rho) s^i\, - \underbrace{ \sum_{i=0}^{N_S} \sum_{k=0}^{i-1} s^{i-k-1} \frac{\mathrm{d}^{k}}{\mathrm{d} v^k}V_0(0)}_{\sim \, 0}. \label{eq:LapTransSource}
\end{eqnarray}
As observed in the previous section, ideally the boundary data is such that all the time derivatives vanish initially. A condition that is only approximatively realised in practice.
\subsection{Quasi-normal amplitudes}\noindent
\noindent
Quasi-normal modes are the complex $s_n-$values for which $\phi_n(\rho)$ is a {\em regular} solution to the homogeneous equation
\begin{equation}
{\boldsymbol \alpha} [\phi_n] + s_n {\boldsymbol \beta} [\phi_n] = 0 \label{eq:QNMEq}.
\end{equation}
This equation is solved in {\texttt Mathematica} within the framework of the ``Generalised Eigenvalue problem" (see appendix \ref{sec:SpecMeth_QNM} for further details).
As discussed in \cite{Ansorg:2016ztf}, the function $\bar{U}(\rho;s)$ has poles on the quasi-normal modes. We introduce the decomposition
\begin{equation}
\bar{U}(\rho;s) = \frac{\bar{V}(\rho;s)}{s-s_n} + \bar{W}(\rho;s).
\end{equation}
Besides, we identify the differential operator on the l.h.s of equation~\eqref{eq:LapTransEq} as ${\mathbf A} (s)= {\boldsymbol \alpha} + s {\boldsymbol \beta}$ and rewrite it in the form
\begin{equation}
{\mathbf A} (s) = {\mathbf A} (s_n) + (s-s_n){\boldsymbol \beta}.
\end{equation}
Then, equation~\eqref{eq:LapTransEq} in the limit $s \rightarrow s_n$ gives as regularity condition
\begin{equation}
\label{eq:AmpliEq_ZeroOrder}
{\mathbf A} (s_n) [ \bar{V} ] = 0 \Rightarrow \bar{V} = \eta_n \, \phi_n(\rho).
\end{equation}
At second order, we obtain
\begin{equation}
\label{eq:Amplitude}
{\mathbf A} (s_n) [ \bar{W} ] + \eta_n \, {\boldsymbol \beta}[\phi_n] = {\boldsymbol \beta} [U_{\rm in}]- \bar{S}.
\end{equation}
equation~\eqref{eq:Amplitude} is to be solved simultaneously for $\bar{W}(\rho)$ and $\eta_n$ with the normalization condition $\bar{W}(\rho_0)=\bar{W}_0$ (for further details see appendix \ref{sec:SpecMeth_QNM_Amplitudes}). \\
Then, the standard recipe to obtain the spectral representation of the solution $U(v, \rho)$ is as follows:
\begin{enumerate}
\item one expresses the time-dependent solution $U(v,\rho)$ as the inverse Laplace-transformation;
\item the integration path of the inverse Laplace-transformation in the complex $s-$plane is deformed in order to include the QNMs (see \cite{Ansorg:2016ztf} for further details). By doing so, we obtain
\begin{equation}
\label{eq:SpecDecom_GenSol}
U(v,\rho) = \sum_{n=0}^{\infty} \eta_n \phi_n(\rho) \operatorname{e}^{s_n v} + \int_{\cal C} \mathrm{d} s\, \hat{U}(\rho;s) \operatorname{e}^{s v} ,
\end{equation}
with the last term corresponding to the integral along the half-circle ${\cal C}$ as $|s|\rightarrow \infty$;
\item for asymptotically flat spacetimes, it is argued in~\cite{Ansorg:2016ztf} that the validity of \eqref{eq:SpecDecom_GenSol} depends on the behaviour of $|\eta_n \phi_n(\rho)|$ for large $n$. It is shown that generically $|\eta_n \phi_n(\rho)| \sim \operatorname{e}^{-\tau \text{Re}(s_n)}$. For $v< \tau$, the function $\hat{U}(\rho;s)$ diverges exponentially as $|s|\rightarrow \infty$. On the other hand, for $v>\tau$, the contribution from the integral $\int_{\cal C}$ is zero and we obtain the desired spectral representation.
\end{enumerate}
\subsection{Numerical evidences for the recipe in AdS-Schwarzschild}\noindent
\noindent
Here, we present numerical evidence showing that the recipe discussed above may also work for asymptotically AdS spacetimes. Using \eqref{eq:Current} we identify\footnote{As detailed in appendix \ref{sec:SpecMeth_QNM_Amplitudes}, we can always normalise $\phi_n(0) = 1$.} \begin{equation}A_n =-2\, \eta_n \phi_n(0).
\end{equation}
First, for the quenches to be presented in~\eqref{eq:Quenches}, we calculate the quasi-normal amplitudes $A_n$ using equation~\eqref{eq:Amplitude} and we observe that they indeed behave as
\begin{equation}\label{eq:tautau}
|A_n|\sim \operatorname{e}^{-\tau\text{Im}[\omega_n]}\,.
\end{equation}
Here $\omega_n$ are the Fourier frequencies. We expect\footnote{It is possible to fine tune the initial data to excite only one (or several) QNMs. In this cases the argument is not valid.} this behaviour to be valid for any quench which approaches a stationary configuration as $v \rightarrow \infty$. In particular, we show in the left panel of figure~\ref{fig:123} an example for the Gaussian quench with $\kappa B=2$ and $\Lambda =0.1$, for which we read the time $\tau = 3.0$.
Second, we check that the sum in \eqref{eq:spectraldec} converges for $v>\tau$, based on
\begin{equation}
|A_n|\operatorname{e}^{-\mathrm{i}\omega_n v} \Big|_{n\rightarrow\infty} \sim \operatorname{e}^{-\text{Im}[\omega_n](\tau-v)}\operatorname{e}^{-\mathrm{i}\text{Re}[\omega_n] v}\,.
\end{equation}%
In figure \ref{fig:123} we also show the comparison of the spectral analysis to the numerical data obtained by solving the PDE for this specific quench. As mentioned before, here we obtain $\tau=3.0$ and the time evolution based on the spectral decomposition matches the numerical evolution accordingly.
In practice, since we can only take a finite number of quasi-normal modes into account, we can distinguish three regimes in the plot. The first is $v<\tau$ where the spectral decomposition diverges and does not resemble the current response at all. There is a second, intermediate, regime $\tau< v\lesssim 4$ where the fit is not completely accurate. In order to obtain a perfect fit in this region, we should also take into account higher QNMs. Finally for $v \gtrsim 4$ the spectral decomposition fits perfectly the numerical data.
The regime $v < \tau$ requires further investigations. It is still not clear to us whether the inclusion of all higher QNMs and possibly an analytic continuation can cure the deviation between the numerical time evolution and the spectral decomposition observed in figure \ref{fig:123}. We leave further investigation in this topic for future work.
\begin{figure}[h!]
\centering
\includegraphics[width=6.9cm]{./Gaussian_etaXs.pdf}
\hspace{1cm}
\includegraphics[width=6.9cm]{./Gaussian_SpecDecXTimeEvol_n12_Inset.pdf}
\caption{\label{fig:123} Left: Logarithmic plot for the amplitudes against $\text{Im}[\omega]$ for the first twelve QNMs. Right: Comparison of the spectral analysis (green) for twelve QNMs to the full current (purple) obtained by solving equation \eqref{eq:vzeq} for a generic Gaussian quench. The chosen parameters yield $\tau=3$. The zoomed image shows times very close to $\tau$ where the convergence is not yet achieved. }
\end{figure}
To summarise, $\tau$ defines the time from which an explicit solving of the time evolution can be substituted by an analysis based on the spectral decomposition i.e. on the quasi-normal modes of the system. It depends not only on the system under consideration but on the concrete initial and boundary conditions. Moreover, there is in principle no bound on the value of this quantity. This means that for specific systems or initial/boundary conditions it might be possible to get $\tau<0$.
In \cite{Szpak:2004sf}, it is argued that $\tau$ can be understood as the time it takes for (a compact support) initial data to propagate to a point whose lightcone contains all the initial data information. So far, we have not found a similar intuition in AdS. It is worth mentioning that \cite{Szpak:2004sf} assumes that the initial data is a function of compact support and does not consider the possibility of a boundary with ``ingoing'' data.
In the following, we investigate the dependence of $\tau$ on the width of the quenches and on the effective anomaly-magnetic field parameter $\kappa B$. With this we would like to explore the possibility of considering $\tau$ as a well defined notion of ``initial time'' for linear (or linearised) systems.
\section{Initial time}\label{sec:initial}\noindent
In the previous section we reviewed how to compute the residues of the QNMs of the system and discussed the mutual growth rate $\tau$. As emphasised before, this time scale is fixed not only by the system under consideration but by the initial and boundary data as well. In this section we study the dependence of $\tau$ on the different quenches and on $\kappa B$ and we explore the possibility of identifying it with the out of equilibrium or ``initial response'' time of the system.
Concretely we have computed $\tau$ as a function of the anomaly parameter $\kappa B$ for quenches of the form
\begin{equation}
\label{eq:Quenches}
V_0(v) = \left\{
\begin{array}{ccc}
\text{exp} \left \lbrace-(v-v_i)^2\Lambda\right\rbrace & & {\rm (Gaussian \, quench)} \\
\left[ 1+ \tanh\left \lbrace (v-v_i)\Lambda\right\rbrace\right]/2\, & & {\rm (tanh \, quench)},
\end{array}\right.
\end{equation}%
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./Gaussian_TauXDelta.pdf}
\hspace{1cm}
\includegraphics[width=6.9cm]{./Gaussian_slopeXkB.pdf}
\caption{\label{fig:tau1}Left: Distance of $\tau$ to the centre of the Gaussian quench against its width for several values of $\kappa \tilde{B}$. From now on we use the rescaled magnetic field $\tilde B \equiv B/(\pi T)^2$ in the plots. In the zoom the ``very fast'' quench region is shown. Right: Slopes of the lines in the l.h.s. against $\kappa \tilde{B}$. The data fits $\sim -4\kappa \tilde{B}$.}
\end{figure}
where $v_i$ fixes the ``centre of the source", and of $\Lambda$, which is inverse to the width of the signal. This quantity should be though of as the abruptness of the quench, so that $\Lambda\gg1$ implies a very fast quench. Let us remark that the main qualitative difference between the two sources is that Gaussian sources introduce no final net axial charge in the system while hyperbolic tangents do. This can be easily understood by considering that the total axial charge introduced in the system goes like\footnote{ Equation \eqref{eq:totalcharge} makes it clear that the total charge generated is of topological nature.}
\begin{equation}\label{eq:totalcharge}
\rho_5\sim \int \mathrm{d} v \, E\cdot B = B\, V_0(v)\Big|^{v_f}_{v_i}\,.
\end{equation}%
\begin{figure}[t!]
\centering
\includegraphics[width=8.cm]{./Tanh_TauXDelta1.pdf}
\caption{\label{fig:tautanh} Distance of $\tau$ to the centre of the tanh quench against $\Lambda^{-1}$ for several values of $\kappa \tilde{B}$. Note that $\tau$ deviates from the linear behaviour for fast quenches i.e. $\Lambda^{-1}$ small. In the linear regime the slope is independent of $\kappa \tilde{B}$.}
\end{figure}
As in previous studies of holographic quenches \cite{Buchel:2013lla}, we expect the quench scale $\Lambda$ to play an important role in the response of the system. For non-compact support functions like the ones we are considering, it is necessary to set some reference time to compare $\tau$ with. The most natural quantity to consider in our case is the temporal distance to the centre of the source: $\tau-v_i$. In figures \ref{fig:tau1} and \ref{fig:tautanh} we show the dependence of $\tau-v_i$ on $\Lambda$ for several values of $\kappa B$ for Gaussian and tanh quenches respectively.
For Gaussian quenches (see figure~\ref{fig:tau1}) we find a linear dependence on $\Lambda^{-1}$. In particular, for $\kappa B=0$ we observe no dependence on this quench parameter. In the r.h.s. of figure \ref{fig:tau1} we show the dependence of the slope $(\tau-v_i)/\Lambda^{-1}$ with $\kappa B$. As shown in the figure the data fits to a line with slope $-4\kappa B$. The fact that this slope is negative means that for the same quench, the initial time is smaller the higher $\kappa B$. Hence, the system is closer to the adiabatic response the stronger the anomaly term. In other words, the relaxation time of the system is smaller for larger $\kappa B$.
For tanh quenches (see figure~\ref{fig:tautanh}) we find again a linear dependence on $\Lambda^{-1}$ for small $\Lambda$ ($\Lambda\lesssim 10$). In this case, contrary to what we observed in the Gaussian quench, $\kappa B$ only shifts the lines and does not affect the slope. For larger $\Lambda$, the time scale $\tau - v_i$ shows a non-linear dependence on $\Lambda^{-1}$. Although qualitatively different, the behaviour of $\tau$ for tanh quenches is still compatible with the intuition of a smaller relaxation times the higher $\kappa B$.
Summarising, in all cases the data fits to a linear dependence on $\Lambda^{-1}$ in the regime of small enough $\Lambda$. That $\tau-v_i$ is generically negative and increases for faster quenches seems natural, since for very slow quenches one gets closer to the adiabatic regime and one expects to find $\tau-v_i\rightarrow-\infty$.
By looking at a different quantity, obtained within the initial time $v<\tau$, we will confirm this intuition in next section.
\subsection{Universality in fast quenches}\label{sec:adiabatic}\noindent
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./1b.jpg}
\hspace{1cm}
\includegraphics[width=6.9cm]{./2.pdf}
\caption{\label{fig:adiabvsnoadiab}Left: Current against time for Gaussian quenches with same height and different widths $\Lambda/(\pi T)^2=0.1$ ``slow'' (orange) and $\Lambda/(\pi T)^2=2$ ``fast'' (blue) at fixed magnetic field $\kappa \tilde{B}=1$. Dashed back lines show the adiabatic response given by equation \eqref{eq:adiabatic}. Right: Current against time for fixed Gaussian source with $\Lambda/(\pi T)^2 =0.5$ and several values of the magnetic field $\kappa \tilde{B}=0.1,0.5,4$ (blue, green, orange). For an easier comparison, the currents have been rescaled such that the corresponding adiabatic response (black, dashed) is the same.}
\end{figure}
\noindent
In the previous subsection we have seen how $\tau$ depends on $\Lambda$ and $\kappa B$. As already argued this provides us with some notion of ``initial'' response of the system. In this subsection we investigate the initial response regime defined as $v<\tau$. Let us recall that for this part of the response one has to, by definition, explicitly integrate equation~\eqref{eq:DynEq_TimeDom} in time.
Our basic motivation comes from the studies of quantum quenches in \cite{Buchel:2013lla, Das:2014hqa}. In these papers a universal response was found for certain kind of smooth, fast enough quenches in generic CFT's. Concretely the authors of \cite{Buchel:2013lla} looked at a free scalar holographic model. Among others, it was shown that the time $\tau_\text{ex}$, at which the operator deviates 5\% from the adiabatic response, follows a simple universal law for fast quenches. We perform an analogous analysis in our system.
Before we proceed to explain the details, some comments are in order.
A first observation is that our system is linear, in contrast with that of \cite{Buchel:2013lla}. As we will see this does not prevent the appearance of an universal behaviour for fast quenches. In addition to this, we emphasise that our main interest is to determine how the anomaly affects the universal behaviour. Therefore we focus on the dependence of the universality on the value of our anomaly/magnetic parameter $\kappa B$. Moreover, we focus on Gaussian sources henceforth.
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./adiabatictransition.pdf}
\hspace{1cm}
\includegraphics[width=6.9cm]{./criticlambda.pdf}
\caption{\label{fig:max-min}Left: Logarithmic plot of $\delta$ (see equation \eqref{eq:delta}) against the abruptness of the quench $\Lambda$ for several values of $\kappa \tilde{B} =0.2-1.2$ (red-blue). For high enough $\Lambda$ all cases converge to a straight line with slope 1.120. Black points highlight the critical $\Lambda_C$ as defined in
the text. Right: Critical inverse width $\Lambda_C$ against $\kappa \tilde{B}$ (black) which can be fitted by a straight line (blue).}
\end{figure}
To give a qualitative idea of the system response, we show the generic behaviour of the current in figure \ref{fig:adiabvsnoadiab}. In the l.h.s.~we fix $\kappa B$ and plot the current for two qualitatively different values of the width $\Lambda$ of Gaussian.
The qualitative difference between a ``fast'' quench (blue) and a ``slow'' quench (orange) is apparent: in the slow case the current approaches the adiabatic behaviour, given by equation~\eqref{eq:adiabatic}, depicted in black. The fast quench shows a more complicated structure and deviates significantly from its corresponding adiabatic behaviour. In the r.h.s.~we fix the width and show the response for different values of the magnetic field. Again, one can observe a qualitative difference; the lower $\kappa B$ the bigger the deviation from the adiabatic response. This simple qualitative observations seem consistent with those made in the previous section. We would like now to see whether the system shows an universal behaviour for fast quenches. This will shed some light concerning the relaxation time. In order to characterise this we have found it useful to look at the following quantity
\begin{equation}\label{eq:delta}
\delta\equiv |J_z(v_1)|-|J_z(v_2)|,
\end{equation}%
with $J_z(v_1)$ and $J_z(v_2)$ the first minima/maxima of the current, respectively.
We have computed $\delta$ as a function of $\Lambda$ for several values of the $\kappa B$. Our results are summarised in figure \ref{fig:max-min}. The plot shows that fast quenches indeed show an universal, $B$ independent, behaviour. Concretely we find
\begin{equation}
\delta\sim \Lambda^{\alpha} \quad {\rm with} \quad \alpha = 1.118 \,,
\end{equation}%
%
and a fitting error for $\alpha$ of the order $10^{-3}$. As one can observe in figure~\ref{fig:max-min}, as we increase $\kappa B$, we need faster quenches, i.e.~higher $\Lambda$, to get to the universal regime. Despite of the smoothness of the transition, it is possible to define a critical $\Lambda$ using an interpolating function for the (logarithmic) data. It is natural to identify $\Lambda_C$ with the absolute maximum of the second derivative of the interpolating function. We get same results using either splines or Hermite polynomials as interpolants and for several interpolation orders. In the left panel of figure \ref{fig:max-min} we have exaggerated the corresponding points. In the right panel we show the values of $\Lambda_C$ against $\kappa B$. By fitting this data to a curve of the form $a+bx^c$ we find $a=-0.08$, $b=4.33$ and $c=0.96$ with fitting errors 0.19, 0.19, 0.07 respectively. We conclude that the transition to the ``universal'', fast quench, regime is delayed linearly with $\kappa B$. This suggests that the effective relaxation time of the system is inversely proportional to $\kappa B$. This is in agreement with the $\tau$ behaviour for tanh quenches displayed in figure \ref{fig:tautanh} where we see that the nonlinear (``fast'') behaviour appears for higher $\Lambda$ the higher $\kappa B$.
\section{Landau resonances}\label{sec:late}\noindent
\begin{figure}[t!]
\centering
\includegraphics[width=230pt]{./nodecay.jpg}
\caption{\label{fig:latetime}Current against time for fixed Gaussian width $\Lambda/(\pi T)^2=6$ and several values of $\kappa \tilde{B}=0.5,1,2$ (orange, blue, red). Late time oscillations decay faster for smaller $\kappa \tilde{B}$.}
\end{figure}
\noindent
Let us now look at the response of the current for late times. Concretely, we consider times bigger than the transition time $\tau$ as defined in equation~\eqref{eq:tautau}. An interesting result in this regime is depicted in figure \ref{fig:latetime}. Here we show the behaviour of the current for a fixed Gaussian source and several values of the magnetic field. The decay rate of the late time oscillations decreases as we increase $\kappa B$. In particular, for $\kappa B\gtrsim 1$ this relaxation time is very small compared with all other scales in the system. This signals the existence of a QNM approaching the real axis for increasing $\kappa B$. We have computed the late time current for a variety of quenches (Gaussian, tanh, oscillatory) with analogous behaviour; the only qualitative difference we have found is the value around which the current oscillates and that is given by the total amount of axial charge induced in the system (see equation~\eqref{eq:totalcharge}).
There is, in fact, an analytic argument to see that indeed a QNM must approach the real axis with increasing $\kappa B$ by writing the perturbation equation for the relevant field component as a Schr\"odinger-type equation (see appendix \ref{sec:QNMapp} for further discussion).
Next we want to check whether this behaviour is indeed an anomaly driven phenomenon, at least for a class of holographic models. The main obstacle we face is that in the probe limit the magnetic field always appears in combination with the Chern-Simons coupling $\kappa$ in the equations. Therefore, in this limit it is not possible to disentangle anomaly from magnetic field effects. Moreover, the probe approximation only makes sense for small values of the fields and we are now interested in the fate of the QNMs as we increase $B$. This motivates us to compute the QNM spectrum of the system including backreaction. Details on the computation can be found in appendix \ref{app:QNM}. Our main goal is to address the following questions: How do the lowest QNMs behave for increasing $B$? Is any mode crossing to the upper half plane $\text{Im}[\omega] >0$? What is the role of the axial anomaly in the dynamics of the QNMs?
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./lhsfig7.pdf}
\hspace{1cm}
\includegraphics[width=6.9cm]{./rhsfig7.pdf}
\caption{\label{fig:imqnm}Left: Imaginary versus real part of the three lowest QNMs at zero frequency for $\kappa=0$, $\tilde{B}=0-4$ (green-yellow) and $\kappa=1$, $\tilde{B}=0-4$ (red-yellow). Right: QNMs' imaginary part versus magnetic field for $\kappa\in\{ 0.1, 0.2, 0.34, 0.5, 1\}$ (yellow-blue).}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./lhsfig8.pdf}
\hspace{1cm}
\includegraphics[width=6.9cm]{./rhsfig8.pdf}
\caption{\label{fig:reqnm}Left: Real part of the lowest QNM against magnetic field for $\kappa=1$ (red). For $\tilde{B}> 2 $ data fits $\sqrt{\tilde{B}}$ (dashed, black). Right: Double logarithmic plot of $\text{Re}[\omega]$ of the lowest QNM against $\tilde{B}$ for several values of $\kappa\in\{0.5, 0.67, 0.83, 1, 1.5, 2, 2.5 \}$ in the large magnetic field regime. Data (red) fits to $\sqrt{\tilde{B}}$ (dashed, black) in all cases.}
\end{figure}
We have looked at the dependence of the lowest QNMs on $\kappa$ and $B$ at zero momentum. The main message is summarized in the l.h.s. of figure \ref{fig:imqnm}. Here we show the behaviour of the three lowest QNMs as one increases the magnetic field for $\kappa=0$ and $\kappa=1$. The difference is apparent. For $\kappa=1$ the modes approach the real axis and the real part increases with B. Conversely, if we set $\kappa=0$ the modes approach the imaginary axis. There is an intermediate region, for approximately $0<\kappa\lesssim 1/2$. To illustrate this we show the fate of the lowest QNM for $\kappa\in\{ 0.1, 0.2, 0.34, 0.5, 1\}$ in the r.h.s. of figure \ref{fig:imqnm}. Here we see that in the low $\kappa$ regime the imaginary part of the frequency begins decreasing again once the magnetic field is large enough. We found similar results for higher QNMs. This implies that for low values of $\kappa$ no resonances are to be found, independently of the value of $B$. For $\kappa \gtrsim 1/2$ we find the mode approaching monotonically the real axis in the regime of B allowed by our numerics. For these values the QNMs approach the axis up to very small absolute values ($\text{Im}[\omega]\sim -10^{-6}$ for $\kappa=1/2$) of the imaginary part and resonances are found. In all cases the approach is faster the higher $\kappa$. We didn't find any mode crossing to the upper half plane for any of the values considered.
In addition to this, it is clarifying to look at the behaviour of the real part of the frequency. By fitting it to a function of the form
\begin{equation}\label{eq:fitx}
\text{Re}[\omega]=a+bB^c,
\end{equation} we find that far all non-zero values of $\kappa$ considered here, ranging from $\kappa=0.1$ to $\kappa=2.5$, the value of $c$ is compatible with $\sqrt{B}$ behaviour when $B$ is large enough. Concretely $c=0.4929$ (with a fitting error of $10^{-4}$) is the largest deviation from $c=1/2$ that we find. In the l.h.s. of figure \ref{fig:reqnm} we show the real part of the lowest QNM against magnetic field for $\kappa=1$. The zoom in the low $B$ region shows that the data deviate from the $\sqrt{B}$ behaviour at low values of $B$. In the r.h.s. of this figure we show the fit of this mode to $\sqrt{B}$ in the high $B$ regime for several, non-trivial, values of $\kappa$. This indicates that these resonances are a consequence of the presence of Landau levels in the system. At $\kappa=0$, however, the behaviour is qualitatively different, as one can already deduce from the l.h.s. of figure \ref{fig:imqnm}. In this case the real part of the frequency decreases with increasing $B$. This absence of Landau levels suggests that at $\kappa=0$ there are no fermions charged under the global symmetries induced by the bulk photons in the dual theory. Moreover the dependence of the imaginary part on $\kappa$ indicate that we can view the resonances as a footprint of the anomaly.
Let us make a remark. In principle, for large values of $\kappa$ we cannot exclude the possibility of the modes going away from the real axis again for very large values of $B$. If this is the case, then the resonances would be restricted to a certain regime of the magnetic field.
Next, we inquire the dependence of $b$ (see equation \eqref{eq:fitx}) on $\kappa$. For the data used in figure \ref{fig:reqnm} we find $b\sim\kappa^{0.510}$ with a fitting error of $10^{-3}$ for the exponent. Therefore, for high enough $\kappa$ and large enough magnetic field we conjecture that $\text{Re}[\omega] \sim \sqrt{\kappa B}$.\\
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./res1.jpg}
\hspace{1cm}
\includegraphics[width=6.9cm]{./res2.jpg}
\caption{\label{fig:resonance}Current vs time for an oscillatory source \eqref{eq:tanhsin} for $\omega=0.95\omega_{\rm c}, 0.90\omega_{\rm c}$ (left: red, blue) and $\omega=\omega_c$ (right: orange).}
\end{figure}
Finally, the slow decay rate of the current is not the only effect that ``almost normal'' modes have in the current. One can consider the possibility of directly exciting these modes by switching on an oscillatory source with appropriate frequency. In figure \ref{fig:resonance} we show the behaviour of the current within the probe limit for a source of the form
\begin{equation}\label{eq:tanhsin}
V_0(v)=(1+\tanh(v-v_i))\sin(\omega\,v)\, .
\end{equation
It corresponds to an oscillatory electric field with an amplitude envelope\footnote{The envelope plays no important role away from $v_i$ which is conveniently chosen to be small. One can check this by comparing the response of a damped oscillator to sources of the form $(\sin(\omega v))$ and $((\tanh(v-v_i)+1)\sin(\omega \,v))$.} chosen for numerical convenience. As expected, the current response is qualitatively similar to that of a damped driven oscillator. For frequencies close to the resonant frequency $\omega_{\rm c}$ (fixed by $\kappa B$) the amplitude is dramatically enhanced and several characteristic frequencies appear. At the critical value of $\omega$ the amplitude is maximally amplified and it asymptotes a certain value. The higher $\kappa B$ the stronger the resonance. \\
\section{Discussion}\label{sec:discussion}\noindent
\noindent
In this paper we have considered quenches in an holographic system with a $U(1)\times U(1)$ anomaly. Concretely, we looked at the electric current generated parallel to a magnetic field known as chiral magnetic current and its response to a parallel time-dependent external electric field. Our motivations were twofold. On the one hand, as argued in the introduction, it is necessary to address the question of how anomalous transport behaves out of equilibrium and how it reacts to the dynamical evolution of the axial charge. On the other hand, we were interested in how the axial anomaly affects previous results \cite{Buchel:2013lla} regarding the universal response to fast quenches.
As a first step, we have only considered the probe limit of the theory. The probe approximation limits the validity of the results to field configurations with a small stress-energy tensor compared to the temperature of the black hole. As a consequence, this approximation is only justified for sufficiently ``small" magnetic fields and the sufficiently ``slow" quenches.
We have already made progress in the task of including backreaction into the system and will present our results in a follow up paper. However, we remark that one of our main results, the Landau resonances for large magnetic fields, were obtained with backreaction.
In the probe limit, our setup yields in a linear hyperbolic PDE, which has to be solved as a initial/boundary value problem. We explicitly evolve this equation in time with a fully spectral method~\cite{Macedo:2014bfa}. Moreover, we also consider a different strategy to obtain the solution to the problem. Based on the discussion from~\cite{PhysRevD.45.2617, Nollert:1998ys,Ansorg:2016ztf} we read the so called quasi-normal amplitudes out of the initial/boundary data and express the solution as a spectral decomposition based on the quasi-normal modes. As discussed in~\cite{Ansorg:2016ztf,Szpak:2004sf} for asymptotically flat spacetimes, this analysis is expected to be valid for a time scale $v>\tau$, where $\tau$ is related to the growth rate of the amplitudes. We observed that the same analysis works in our asymptotically AdS system. Moreover, based on~\cite{Warnick:2013hba}, we expect this approach to be valid in generic asymptotically AdS spacetimes.
We have investigated the behaviour of $\tau$ for Gaussian and tanh quenches with the idea of associating this quantity to a well defined notion of initial time. Our results are positive, although some unsolved questions remain. We find that the qualitative dependence of $\tau$ on the abruptness of the quench and $\kappa B$ matches the expectations: $\tau$ becomes smaller close to the adiabatic regime or for higher $\kappa B$.
However, it is still to be investigated how $\tau$ depends on the precise quench protocol and whether $\tau$ can be used to identify a generic classification for the quenches. In particular, the Gaussian quench showed a rather unexpected behaviour at $\kappa B = 0$, namely the time scale $\tau$ appeared to be independent of the width. This issues deserve deeper analysis and we leave it to future work.
In addition, we have studied the behaviour of the dual current operator for fast quenches (motivated by \cite{Buchel:2013lla}) where a universal response regime was found in a holographic model with a dual scalar operator. Our results show that for fast quenches the quantity $\delta$ defined in equation \eqref{eq:delta} shows a universal behaviour: it is independent of the magnetic field and it scales with a fixed exponent with the width of the quench. Moreover, we have shown that this universal regime is suppressed the higher $\kappa B$, i.e. one needs to perform faster quenches to reach the universal regime the higher $\kappa B$.
Our results for both $\tau$ and $\delta$ indicate that the relaxation time of the system is smaller with increasing $\kappa B$. This is coherent with the notion of $\kappa B$ being a coupling of the current to the electric field\footnote{This becomes apparent if one considers the negative magnetoresistance in the case of parallel external fields $J\sim (\kappa B)^2 E$.}.
The last main point of this work is the emphasis on the dynamics of the current for late times. By explicitly computing the fate of the current after quenches in the electric field, we have shown that oscillatory, long-lived currents are produced in presence of the anomaly and strong enough magnetic fields in our holographic model. By computing the QNMs of the system with backreaction we have been able to show that this is a consequence of the presence of Landau levels\footnote{It is worth mentioning that Landau level resonances have been experimentally found in Dirac semimetals \cite{2016arXiv160102316Y, PhysRevLett.115.176404}.}. This has been done by looking at the dependence of the lowest QNMs on $\kappa$ and $B$. We have observed that for non-zero $\kappa$ the real part of the frequency goes as $\sim\sqrt B$ for large $B$, pointing towards the appearance of Landau levels for $\kappa>0$. As already discussed in the main text, the fact that the characteristic $\sqrt{B}$ behaviour is absent for $\kappa =0$ indicates that no dual fermionic d.o.f. are charged under the symmetries in consideration when the Chern-Simons term is not present in the bulk, i.e. when there is no dual anomaly.
Whether resonances are to be found in the conductivity depends not only on the magnetic field, but on $\kappa$ as well. As we saw, the imaginary part of the quasi normal frequencies asymptotes the real axis only if $\kappa \gtrsim 1/2$. It is tempting to speculate whether this is related to the existence of a quantum critical point for $\kappa>1/2$ studied in \cite{D'Hoker:2010rz, D'Hoker:2010ij, Ammon:2016szz}.
In addition, we explored the idea of directly exciting the Landau resonances by means of an oscillatory external electric field finding a resonant pattern as expected\footnote{This response was already implicit in \cite{Jimenez-Alba:2014iia, Sun:2016gpy}.}. Moreover it would be interesting to check whether long lived oscillating currents can be generated by means of an electric quench in Weyl semimetals. A more realistic treatment of this would require to introduce some (``intrinsic'') mechanism for axial charge relaxation \cite{Jimenez-Alba:2014iia}. We propose to check this statements in the weak coupling regime. Regarding this resonances, it is worth mentioning that the system shows several similarities to holographic p-wave superconductors. Both present dissipationless transport driven by a broken symmetry and both cases \cite{Ammon:2010pg} present resonances at finite frequencies in the anisotropic direction. It can be interesting to explore this analogy in more depth.
Let us finally mention that initially we had Weyl/Dirac semimetals in the back of our heads, which is described at strong coupling by \cite{Landsteiner:2015lsa}. However, we have considered a simpler model which we think captures the core properties we wanted to focus on. It would be interesting to make an analogous study with the holographic model~\cite{Landsteiner:2015lsa}.
\section*{Acknowledgements}
\noindent We thank Marcus Ansorg, Birger Boening, Markus Gardemann, Holger Gies, Markus Heinrich, Carlos Hoyos, Matthias Kaminski, Karl Landsteiner, Julian Leiber, Chris Rosen, Andrej Rostworowski, Jackson Wu and Ho-Ung Yee for fruitful discussions. A.J.~and R.M.~acknowledge financial support by \textit{Deutsche Forschungsgemeinschaft (DFG)} GRK 1523/2. R.M.~was supported by CNPq under the programme "Ci\^encia sem Fronteiras". The work of L.M. is supported by the ERC Advanced grant No.339140 Gravity, Black Holes and Strongly Coupled Quantum Matter.
\begin{appendix}
\section{Pseudospectral methods}\label{sec:methods}\noindent
In this section, we give more details on the numerical techniques used in this work. As already mentioned along the text, all the relevant equations are solved with spectral methods. These methods are widely used for boundary-value problems~\cite{Boyd00,canuto_2006_smf} (typically elliptic equations) and they find several applications for time-independent problems within numerical relativity~\cite{grandclement_2009_smn} and numerical holography~\cite{Dias:2015nua}. Recently, there has been also some development in the application of spectral methods along the time direction in dynamical scenarios~\cite{Macedo:2014bfa}.
Here, we discuss the main features of the method with emphasis on its application for eigenvalue problems and for systems of differential equations with extra parameters. Then, we review the algorithm for the fully spectral code introduced in~\cite{Macedo:2014bfa}.
\subsection{The eigenvalue problem: quasi-normal modes}\label{sec:SpecMeth_QNM}\noindent
For convenience, let us reproduce here the general representation of the ordinary differential equation characterising the QNM as we introduced in \eqref{eq:QNMEq} (for a fixed background) or \eqref{eqS::gevp} (for the backreacted system)
\[
{\boldsymbol \alpha} [\phi_n] + s_n {\boldsymbol \beta} [\phi_n] = 0.
\]
This equation is to be solved in the domain $\rho\in[0,1]$. Due to the choice of the ingoing Eddington-Finkelstein coordinates to the background metric, we obtain only a linear term in the complex parameter $s_n$ characterising the QNM\footnote{Recall that $s$ is the Laplace parameter, which is related to the Fourier frequency $\omega$ by $s=-i\omega$.}. Furthermore, thanks to this coordinate choice, the surfaces of constant time penetrate the black-hole horizon and therefore the ingoing boundary condition on the horizon is automatically realised here by the geometry, i.e., they correspond to the regularity condition of \eqref{eq:QNMEq} at $\rho=1$. In the same way, equation~\eqref{eq:QNMEq} provides us with the proper regularity conditions at $\rho=0$, so no further information is needed for the solution of equation~\eqref{eq:QNMEq}.
In order to discretise equation~\eqref{eq:QNMEq}, we fix a numerical resolution $N_\rho$ and we introduce a Chebyshev-Lobatto grid in the $\rho$ direction via
\begin{equation}
\label{eq:LobattoGrid}
\rho_j = \frac{1}{2}\left[ 1+\cos\left(\pi \frac{j}{N_\rho} \right) \right] \quad j = 0, \ldots, N_{\rho}.
\end{equation}
Then, we recall that ${\boldsymbol \alpha(\rho)}$ and ${\boldsymbol \beta(\rho)}$ are (second/first order, respectively) differential operators acting on the radial coordinate $\rho$. Hence, a discrete representation $\hat{\alpha}$ and $\hat{\beta}$ is obtained by substituting the differential operators $\partial_\rho$ and $\partial^2_{\rho \rho}$ by the discrete Chebyshev-Lobatto spectral differentiation matrices $\hat{D}_{\rho}$ and $\hat{D}^2_{\rho \rho} = \hat{D}_{\rho}\cdot \hat{D}_{\rho}$ , whose expression can be found in~\cite{canuto_2006_smf}. In its discrete form, equation~\eqref{eq:QNMEq} has the structure of a generalised eigenvalue problem and both eigenvalues $s_n$ and eigenvectors $\vec{\phi}_n$ are easily obtained with \texttt{Mathematica}.
First thing to observe is that we obtain a total of $N_\rho+1$ eigenvalues and eigenvectors. Yet, not all of them are trustful numerical solution. It is crucial that one performs a convergence test to determine which of the obtained solutions are stable and converge to fixed value as one increases the resolution $N_\rho$. Our empirical observation shows that for a given $N_{\rho}$, the first $n_\text{QNM} \sim \sqrt{N_\rho}$ QNMs correspond to a reliable numerical solution. In order to keep high-accurate solutions, we fix $N_{\rho} = 300$ and considered only the first $n_\text{QNM} = 12$ quasi-normal modes in this work. Secondly, we would like to stress that we are also interested in the eigenfunctions $\vec{\phi}_n$, since they are needed in the calculation of the QNM amplitudes $\eta_n$. Each component of $\vec{\phi}_n$ corresponds to the value of the function $\phi(\rho)$ at the grid point, i.e.,~$\phi_n^j \equiv \phi(\rho_j)$. Notice that the eigenfunctions are uniquely defined up to a normalisation constant. Thus, we work with the conveniently rescale quantity\footnote{From now on, we omit the notation $\bullet^{\rm norm}$ and $\phi(\rho)$ is always the normalised quantity.} $\vec{\phi}_n^{\rm norm} = \vec{\phi}_n/\phi_{n}^{N_{\rho}}$ which give us $\phi^{\rm norm}_n(0) = 1$.
\subsection{Equation with a free parameter: QNM amplitudes}\label{sec:SpecMeth_QNM_Amplitudes}\noindent
We now proceed with the discussion of the solution of equation~\eqref{eq:Amplitude} (which we reproduce here once more for convenience)
\[
{\mathbf A} (s_n) [ \bar{W} ] + \eta_n \, {\boldsymbol \beta}[\phi_n] = \bar{Q}.
\]
We remind that ${\mathbf A} (s_n) = {\boldsymbol \alpha} [\phi_n] + s_n {\boldsymbol \beta} [\phi_n]$. Moreover, we have simplified the source term on the r.h.s.~with the introduction of $\bar{Q}={\boldsymbol \beta} [U_{\rm in}] - \bar{S} $.
Our objective is to numerically solve equation~\eqref{eq:Amplitude} for {\em both} $W(\rho)$ and $\eta_n$. We also make use of spectral methods for this task so, in principle, the discretisation procedure in terms of the Chebyshev-Lobatto grid points~\eqref{eq:LobattoGrid} and the discrete Chebyshev-Lobatto spectral differentiation matrices is the same as described in \ref{sec:SpecMeth_QNM}.
Note, however, that we have now a total of $N_\rho +2$ unknown variables, which we collect into a single vector (with $W^i \equiv W(\rho_i)$)
\begin{equation}
\vec{X}^{\rm T} = \left( W^0, \,\ldots\,, W^{N_\rho}, \, \eta_n\right).
\end{equation}
The discrete version of equation~\eqref{eq:Amplitude} provides us only with $N_\rho +1$ equations, though. The extra condition needed to complete the system is obtained after observing that the solution $\bar{W}(\rho)$ is not unique. Assuming $\bar{W}_{\rm a}(\rho)$ is solution to equation~\eqref{eq:Amplitude}, then, $\bar{W}_{\rm b}(\rho) = \bar{W}_{\rm a}(\rho) + C \phi_n(\rho)$ also satisfies \eqref{eq:Amplitude}. Therefore, we must impose an extra normalisation condition to $W(\rho)$, which here we fix as
\begin{equation}
\label{eq:NormW}
W(0) = 1.
\end{equation}
Taking into account the discretised spectral representation of equation~\eqref{eq:Amplitude} together with the condition \eqref{eq:NormW}, and introducing the notation $\vec{\beta}_\phi = \hat\beta \cdot \vec{\phi}$, we end up with the linear system $\hat{M}\cdot \vec{X} = \vec{q}$, whose components can be explicitly expressed as
\begin{equation}
\label{eq:QNMAmplit_LinSyst}
\left(
\begin{array}{cc}
A^{i}{}_{j} & \beta_{\phi}^i \\
\delta^{j}{}_{N_\rho} & 0
\end{array}
\right)
\left(
\begin{array}{c}
W^{j} \\
\eta_n
\end{array}
\right)=
\left(
\begin{array}{c}
Q^{i} \\
1
\end{array}
\right), \quad i = 0,\ldots, N_\rho.
\end{equation}
We observe that $\eta_n$ does not depend on the normalisation $W(\rho_0) = W_0$.
While the algorithm used to solve equation~\eqref{eq:Amplitude} is essentially the one described above, we must face a caveat introduced by the logarithmic terms presented in the source term $\bar{Q}(\rho)$ [see equations~\eqref{eq:Source_ak} and \eqref{eq:LapTransSource}], which leads to an algebraic convergence rate of the spectral scheme. In fact, for generically non-vanishing boundary data $V_0(v)$, the source $\bar{Q}(\rho)$ is merely ${\cal C}^0$ due to the term $\sim \rho\log(\rho)$ and in practice, one faces difficulty in finding the amplitudes $\eta_n$ already for $n\gtrsim 2$.
To overcome this problem, we introduce a new coordinate~\cite{Kalisch:2016fkm} $z\in[0,1]$ via
\begin{equation}
\rho = \operatorname{e}^{1-{1}/{z}},
\end{equation}
which allows one to map the problematic ${\cal C}^{k-1}$ terms $\sim\rho^{k} \log(\rho)$ into the ${\cal C}^\infty$ expressions $\sim(1-1/z) e^{(k-k/z)}$. By rewriting\footnote{The point $z=0$ must be treated with care. In this limit, one obtains ${\cal W}_{,z}(0) = 0$ and this property should be explicitly implemented when constructing the matrix $\hat{M}$.} equation~\eqref{eq:Amplitude} in terms of $z$ and discretising the system with Chebyshev-Lobatto grid points $z_i$ together with differentiation matrices $\hat{D}_z$, we obtain a spectral convergence rate\footnote{Here, spectral convergence means that the convergence rate is faster than algebraic. However, the rate does not decay exponentially since ${\cal W}(z)$ is not analytic at $z=0$.} for the solution ${\cal W}(z) = W(\rho(z))$ and therefore a much more stable scheme for finding the amplitudes $\eta_n$.
A second remark regarding the efficiency of the code is related to the presence of rather huge numbers ($\sim 10^{30}-10^{50}$). Such values are a consequence of the Laplace-transformation of the boundary function $\bar{V}_0(s_n)$, which contributes significantly to the source function $\bar{Q}(\rho)$ (see \eqref{eq:LapTransSource}). Therefore, it is also convenient to rescale the vectors in \eqref{eq:QNMAmplit_LinSyst} by $\vec{q} = \bar{V}_0(s_n) \vec{p}$ and $\vec{X} = \bar{V}_0(s_n) \vec{Y}$ and solve the system $\hat{M}\cdot \vec{Y} = \vec{p}$.
In order to obtain the desired high accuracy for the first $n_{\rm QNM} = 12$ quasi-normal modes amplitudes $\eta_n$, we set the resolution to $N_z = 600$. It is true that both resolutions $N_\rho=300$ (for the solution of the eigenvalue problem - sec.~\ref{sec:SpecMeth_QNM}) and $N_z=600$ are quite extreme for codes based on spectral methods. We mention however, that for a given physical parameter $\kappa B$ the generalised eigenvalue problem and the inversion of the matrix $\hat{M}$ must be performed only once. The results can be saved and applied afterwards several times for any r.h.s $\vec{q}$. Still, a more sophisticated and efficient approach based on the work~\cite{Ansorg:2016ztf} together with a systematic study of the QNM amplitudes in a more generic context is planned to be presented in a forthcoming article.
\subsection{Hyperbolic equation: fully spectral code}\label{sec:TimeSpec}\noindent
We end this section with a brief discussion of the fully spectral code used to solve the equation \eqref{eq:DynEq_TimeDom} in terms of the coordinates $\{v,\rho\}$. A detailed description of the method can be found in~\cite{Macedo:2014bfa}. Let us start by assuming we are looking for a solution in a generic time interval $v\in [v_{\rm a}, v_{\rm b}]$. Given the initial data $U_{\rm a}(\rho)$ at $v=v_{\rm a}$, we introduce the auxiliary fields $P(v,\rho)$ via
\begin{equation}
\label{eq:AuxFields}
U(v,s) = U_{\rm a}(\rho) + (v - v_{\rm a})P(v, \rho).
\end{equation}
For prescribed numerical resolution $N_\rho$ and $N_v$, we work with the Chebyshev-Lobatto grid in the $\rho$-direction given in equation~\eqref{eq:LobattoGrid} and with the Chebyshev-Radau collocation points in the $v$-direction
\begin{equation}
\label{eq:grid}
v_k = \frac{ v_{\rm b} + v_{\rm a}}{2} + \frac{v_{\rm b} - v_{\rm a}}{2} \, \cos\left(\frac{2k\pi}{2N_v+1}\right),\,\,
k=0...N_v,
\end{equation}
The function values $V_{ki}=V$ are stored in a vector $
\vec{X}^{\rm T} = \left( V_{ki} \right)_{k=0\dots N_{v}, \, i = 0 \dots N_{\rho}},$ from which Chebyshev coefficients $c_{kj}$ of the field $P(v, \rho)$ are computed by inverting the equations
\begin{eqnarray}
\label{eq:SpecAppr}
P(v_k,\rho_i) &=& \sum_{n=0}^{N_u} \sum_{l=0}^{N_\rho}
c_{nl} T_n\left(\frac{2v_k-v_{\rm a}-v_{\rm b}}{v_{\rm b}-v_{\rm a}}\right)
T_l\left(1-2\rho_i\right),
\end{eqnarray}
with $T_j(\mu) = \cos[j\arccos(\mu)]$ the Chebyshev polynomials of the first kind. After calculating spectral approximations of the fields' derivatives, equation~\eqref{eq:DynEq_TimeDom} yields a linear algebraic system, which is solved with the iterative BiCGStab method. Furthermore, we also provide the BiCGstab method with a pre-conditioner based on a Singly Implicitly Diagonally Runge-Kutta (SDIRK) method~\cite{Macedo:2014bfa}.
After obtaining the solution $U(v, \rho)$ for $v \in [v_{\rm a}, v_{\rm b}]$, the values $U(v_{\rm b}, \rho)$ at the upper time boundary $v_{\rm b}$ serve as initial data for a subsequent time domain $v\in [v_{\rm b}, v_{\rm c}]$. This procedure allows to divide the whole time interval $v\in[0,v_{\rm final}]$ into smaller sub-intervals, with the very first one being $[0, v_{\rm a}]$. The size of each time interval $\Delta v$ can be chosen according to the quench profile. If $\Delta$ is a characteristic time length for a given quench, then we fix $\Delta v = \Delta/4$. Furthermore, we set the resolution in the time direction to $N_v \sim 25$. The radial direction requires a higher resolution $(N_\rho \sim 100)$ due to the presence of logarithmic terms in the source function $S(v,\rho)$.
\section{Schr\"odinger type analysis for the QNMs}\label{sec:QNMapp}\noindent
To understand the behaviour of the QNMs approaching the real axis in the complex $\omega$ plane in the probe limit, we rewrite the perturbation equation in the form
\begin{align}
\label{eqp} (\partial^2 +\omega^2 -V_{\rm eff}) \phi =0\,.
\end{align}
In order to do so we first change coordinates
\begin{align}
\mathrm{d} s^2= -U(r) \mathrm{d} t^2 +\frac{\mathrm{d} r^2}{U(r)} + r^2\,(\mathrm{d} x^2+\mathrm{d} y^2+\mathrm{d} z^2)\,.
\end{align}
where $U(r)= r^2-1/r^2$ is the blackening factor and we restrict ourselves to the probe limit. In this coordinates the boundary is located at $r\rightarrow\infty$. We now consider the e.o.m. to first order in perturbations $\delta V_j(r,t,x)=v_j(r) \operatorname{e}^{-\mathrm{i} \omega t+ \mathrm{i} k x}$, $\delta A_j(r,t,x)=a_j(r) \operatorname{e}^{-\mathrm{i} \omega t+ \mathrm{i} k x}$ on top of the magnetic field background in the $k=0$ limit. Since the system is linear, this just gives rise to the background equations (\ref{eq:eom1},\ref{eq:eom2},\ref{eq:eom3}). Using the constraint equation one finds
\begin{align}\label{eq:luis1}
v_z''+v_z'\left(\frac{U'(r)}{U(r)}+\frac{1}{r}\right)+\frac{r^4 \omega ^2-144\, (\kappa B )^2\, U(r)}{r^4\, U(r)^2}\, v_z=0\,,
\end{align}
where the prime denotes radial derivative and $v_z =v_z(r)$.
To attain the form \eqref{eqp}, we define
\begin{align}
\frac{\mathrm{d}\tilde{r}}{\mathrm{d} r} \equiv H(r)\,, \hspace{4cm}
v_z(r) = S(r) \tilde{v}_z(r)\,,
\end{align}
using the dot $\dot{\ }$ to denote $\partial_{\tilde{r}}$ \eqref{eq:luis1} can be rewritten as
\begin{align}
&\ddot{\tilde{v}}_z +\left(\frac{2 S'}{S H} + \frac{1}{H}\left(\frac{U'}{U}+\frac{1}{r}\right) + \frac{ H'}{H^2}\right)\, \dot{\tilde{v}}_z \nonumber \\ &\quad + \left(\frac{S'}{SH^2}\left(\frac{U'}{U}+\frac{1}{r}\right) +\frac{S''}{SH^2} -\frac{144 \, (\kappa B )^2\,}{r^4U H^2} +\frac{\omega^2}{H^2U^2} \right)\, \tilde{v}_z=0\,.\label{finaleq}
\end{align}
Imposing this expression to match with \eqref{eqp} fixes $S$ and $H$ up to an unimportant constant and leads to
\begin{align}
\left(\partial_{\tilde r}^2+\omega ^2-\frac{U(r) \left(288 \, (\kappa B )^2\,+r^3 U'(r)\right)}{2 r^4}+\frac{U(r)^2}{4 r^2}\right) \tilde{v}_z=0\,.
\end{align}
The explicit expression effective potential is
\begin{align}\label{eq:effpot}
V_{\rm eff} (\kappa^2 B^2;r) = \frac{\left(r^4-1\right) \left(576 \,(\kappa B)^2+3 r^4+5\right)}{4 r^6}\,.
\end{align}
Note that $\lim_{r\rightarrow \infty} V_{\rm eff} = \infty$, whereas $V_{\rm eff} (r=1)=0$.
\begin{figure}[t!]
\centering
\includegraphics[width=6.9cm]{./potential0.pdf}
\hspace{1cm}
\includegraphics[width=6.9cm]{./potential.pdf}
\caption{\label{fig:potential} Left: Effective potential (see equation \eqref{eq:effpot}) for $\kappa B=0$ against $r$. The curve grows monotonically with increasing $r$. Right: Effective potential for $\kappa B=0.1,1,2$ (red-blue) against $r$. The $\kappa B$ term becomes relevant close to the horizon and forms a higher barrier for increasing $\kappa B$. }
\end{figure}
We plot this function in Figure \ref{fig:potential}. As explained in \cite{Kiritsis:2015oxa} the $\kappa B=0$ case is compatible with a conductivity displaying a continuous spectrum due to the fact that the potential is unbounded. However, as $\kappa B$ increases the effective potential develops a barrier close to the horizon, creating more bound states as $\kappa B$ grows bigger. For $\kappa B \rightarrow \infty$ the barrier is infinite and we have a binding potential: there is only a discrete set of allowed frequencies and the QNMs do not display an imaginary part because they cannot reach the black hole due to the infinite barrier.
\section{Quasi Normal Modes}\label{app:QNM}\noindent
\noindent
As explained in the main text, in order to be able to vary $\kappa$ and $B$ independently we have to take backreaction into account (see e.g. \cite{D'Hoker:2009mm}). Since the magnetic field breaks rotational invariance we choose the following ansatz for the metric
\begin{equation}
\mathrm{d} s^2=\frac{1}{\rho^2}\left(-U(\rho)\, \mathrm{d} v^2 - 2\,\mathrm{d} v\,\mathrm{d} \rho+W(\rho)^2\,(\mathrm{d} x^2+\mathrm{d} y^2)+H(\rho)^2\,\mathrm{d} z^2\right).
\end{equation}
The equations of motion for the background in the so called trace reduced form read
\begin{equation}
R_{\mu\nu}=-4\,g_{\mu\nu}+\tau^2\,\left(-\frac 16\, g_{\mu\nu}\,\left( F_{\alpha\beta}F^{\alpha\beta}+H_{\alpha\beta}H^{\alpha\beta}\right)+g^{\alpha\beta}\,\left(F_{\mu\alpha}F_{\nu\beta}+H_{\mu\alpha}H_{\nu\beta}\right)\right).\label{eqs:efg}
\end{equation}
Due to diffeomorphism invariance we can set the black hole horizon to $\rho=1$. The metric function $U(\rho)$ is the blackening factor and hence has to vanish at the horizon $U(1)=0$.
Before solving the equations of motion we first consider the asymptotic expansions. Imposing asymptotically AdS
\begin{equation}
U'(0)=0, \ W(0)=1, \ H(0)=1
\end{equation}
leads to
\begin{eqnarray}
U(\rho) & =1+\rho^4 \,\left[\mathbf{u}_4+\mathcal O(\rho^2)\right]+\rho^4\,\log(\rho)\,\left[\frac{B^2 \tau^2}{3}+\mathcal O(\rho^2)\right],\label{uexp}\\
W(\rho) & =1+\rho^4 \,\left[-\frac{\mathbf{h}_4}{2}+\mathcal O(\rho^2)\right]+\rho^4\,\log(\rho)\,\left[-\frac{B^2 \tau^2}{12}+\mathcal O(\rho^2)\right],\label{wexp}\\
H(\rho) & =1+\rho^4 \,\left[\mathbf{h}_4+\mathcal O(\rho^2)\right]+\rho^4\,\log(\rho)\,\left[-\frac{B^2 \tau^2}{6}+\mathcal O(\rho^2)\right],\label{hexp}
\end{eqnarray}
where we set the term linear in $\rho$ in equation \eqref{uexp} (and therefore in equations \eqref{wexp} and \eqref{hexp}) to zero in order to fix the remaining diffeomorphisms. Furthermore we can fix $2\tau^2=1$ since it appears in the equations always as a product with the magnetic field $B$.
Using the ansatz $V_m=V_y(x)=Bx$ and $A_m=0$ we can solve the equations of motion by a spectral method. To do so we first rescale our functions properly using the expansions equation \eqref{uexp}-\eqref{hexp} and solve for given $\kappa$ and $B$ equation \eqref{eqs:efg} for the rescaled functions.
With the background solutions at hand we can look at the fluctuations (c.f. \cite{Janiszewski:2015ura}). Since the factor proportional to the Chern-Simons coupling is metric independent the metric fluctuations will decouple from the fluctuations of the gauge fields. Furthermore the (1) and (2) sector decouple as well and we have just a coupling of the (0)-(3) sector. Setting the momentum
$k$ to zero, we notice, that only the $a_0$-$v_3$ and $a_3$-$v_0$ components are coupled.
For the former the constraint equation reads
\begin{equation}
\dot a_0'(v,\rho)+\frac{12\, \kappa B \,\rho}{H(\rho)\, W(\rho)^2}\, \dot v_3(v,\rho)=0.
\end{equation}
Similar to the non-backreacted case we can integrate the equation in time and obtain
\begin{equation}
a_0'(v,\rho)=-\frac{12\, \kappa B\, \rho}{H(\rho)\, W(\rho)^2}\, v_3(v,\rho)+C_1.
\end{equation}
Like before we set the constant $C_1$ to zero. With this relation between $a_0'$ and $v_3$ we can eliminate $a_0$ in the $v_3$ equation and obtain after a Laplace transformation
\begin{equation}
{\boldsymbol \alpha}[v_{3,n}]+s_n\,{\boldsymbol \beta}[v_{3,n}]=0, \label{eqS::gevp}
\end{equation}
with
\begin{align}
\boldsymbol{\alpha}&=-(12\,\kappa B )^2\, \rho^3\,H(\rho)+\rho\, H(\rho) U(\rho) W(\rho)^4 \,\frac{\mathrm{d}^2}{\mathrm{d} \rho^2}\nonumber\\&+\left[H(\rho) W(\rho)^3 \left(2 \rho \,U(\rho) W'(\rho)\!-\!W(\rho) (U(\rho)-\rho\,U'(\rho)\right)-\rho\, U(\rho) W(\rho)^4 H'(\rho)\right]\,\frac{\mathrm{d}}{\mathrm{d} \rho}
\intertext{and}
\boldsymbol{\beta}&= \left[H(\rho) W(\rho)^3 \left(W(\rho)-2 \rho\, W'(\rho)\right)\right] +2 H(\rho) W(\rho)^4 \,\rho\,\frac{\mathrm{d}}{\mathrm{d} \rho}\,.\end{align}
We solve this equation by means of pseudospectral methods imposing the non-normalizable mode to zero. The QNMs we are interested in converge slowly for high values of the magnetic field. To ensure the convergence we improve the spectral solution by a coordinate mapping of the radial variable $\rho\mapsto\tilde\rho^2$ which moves the gridpoints in the direction of the boundary localised at $\tilde\rho=0$. Furthermore we use a grid with $N_\rho=200$.
\end{appendix}
\newpage
\bibliographystyle{JHEP}
|
\section{Introduction}\label{setting}
Consider a system with a large state space $\Omega$, hidden from view inside a box. On the outside of the box there are lightbulbs and buttons. Each lightbulb corresponds to a set $f_i \subseteq \Omega$ and is lit whenever the current state of the system is in $f_i$. We think of each set $f_i$ as containing all states sharing some negative feature $i \in [m]$ and refer to each such set as a \emph{flaw}, letting $F = \{f_1, f_2, \ldots, f_m\}$.
For example, if the system corresponds to a
graph $G$ with $n$ vertices each of which can take one of $q$ colors, then $\Omega = [q]^n$, and we can define for each edge $e_i$ of $G$ the flaw $f_i$ to contain all assignments of colors to the vertices of $G$ that assign the same color to the endpoints of $e_i$.
Following linguistic convention, instead of mathematical, we will say that flaw $f$ is present in state $\sigma$ whenever $f \ni \sigma$ and that state $\sigma$ is \emph{flawless} if no flaw is present in $\sigma$. The buttons correspond to actions, i.e., to mechanisms for state evolution. Specifically, taking action $a$ while in state $\sigma$ moves the system to a new state $\tau$, selected from a probability distribution that depends on both $\sigma$ and $a$.
Outside the box, an agent called the \emph{controller} observes the lightbulbs and pushes buttons, in an effort to bring the system to a flawless state. Specifically, if $O(\sigma) \in \{0,1\}^m$ denotes the lightbulb bitvector, with 1 corresponding to lit, the controller repeatedly applies a function $P$, called a \emph{policy}, that maps $O(\sigma)$ to a distribution over actions. Thus, overall, state evolution proceeds as follows: if the current (hidden) state is $\sigma \in\Omega$, the controller observes $O(\sigma)$ and samples an action from $P(O(\sigma))$; after she takes the chosen action, the system, internally and probabilistically, moves to a new (hidden) state $\tau$, selected from a distribution that depends on both $\sigma$ and the action taken.
Our work begins with the observation that several recent results~\cite{moser,MT,szege_meet, SrinivasanPerm, AI, JACM, HV, Harmonic, Commu} on LLL algorithmization can be seen as giving sufficient conditions for a controller as above to be able to bring the system to a flawless state quickly, with high probability. Motivated by this viewpoint we ask if conditions for LLL algorithmizations can be seen as \emph{stability criteria} and give results for more general settings, e.g., Partially Observable Markov Decision Processes (POMDPs). Given the capacity of LLL algorithmization arguments to establish convergence in highly non-convex domains, a major pain point in control theory, we believe that bringing such arguments to stochastic control is a first step in a fruitful direction. In order to move in that direction we generalize the setting described so far in two ways:
\begin{itemize}
\item
The mapping $O$ from states to observations is \emph{stochastic}: the lightbulbs are unreliable, exhibiting both false-positives and false-negatives.
\item
Both the environment surrounding the system and the implementation of actions are \emph{noisy}: the controller is not the only agent affecting state evolution and flaws may be introduced into the state for reasons unrelated to her actions, even spontaneously.
\end{itemize}
The question, naturally, is whether sufficient conditions for quick convergence to flawless states can still be established in this new setting. We answer the question affirmatively and show, in a precise mathematical sense, that the less internal conflict there is in the system, the more noise the controller can tolerate. In order to prove this we require the controller to be \emph{focused} and to \emph{prioritize}. That is, we will assume that the flaws are ordered by priority according to an arbitrary but fixed permutation $\pi$ of $F$, and we will ascribe the action taken by the controller in each step to the present flaw (focus) of highest priority (prioritization). The analysis will then take into account both how good the actions are at ridding the state of that flaw and how damaging they are in terms of introducing new flaws. In particular, with this attribution mechanism in place, and similarly to LLL algorithmization arguments, we will say that flaw $f_i$ can cause flaw $f_j$ if there exists a state transition with non-zero probability under the policy, from a state in which $f_i$ is the highest priority flaw and $f_j$ is absent to a state in which $f_j$ is present.
The main challenge we face is that in the presence of noise the causality relationship becomes dense. To overcome this we develop a new analysis in which causality is not a binary relationship, but one weighted by the \emph{frequency} of interactions. In particular, our condition guaranteeing that the controller will succeed within a reasonable amount of time allows the causality graph to become arbitrarily dense, if the frequency of interactions is sufficiently small. Turning the sparsity of the causality relationship into a \emph{soft} requirement is a major departure from the LLL setting and our main technical contribution. We do this by developing an entropy compression argument, in which we carefully amortize the entropy injected into the system to encode the effect of noise on the state trajectory. It is worth pointing out that even though our technique applies to the far more general noisy setting, in the absence of noise it recovers the main result of~\cite{JACM}, thus providing a smooth relationship between lack of internal conflict and robustness to noise.
\section{Formal Setting and Statement of Results}
In the absence of observational and environmental noise we can think of the state evolution under a policy $P$ as a random walk on a certain digraph on $\Omega$. Specifically, at each flawed state $\sigma \in \Omega$, for each action in the support of $P(O(\sigma))$, there is a bundle of outgoing arcs of total probability 1, corresponding to the state-transitions from $\sigma$ under this action. The convolution of $P(O(\sigma))$ with the distribution inside each bundle yields the state-transition probability distribution from each flawed state $\sigma$.
The presence of observational and environmental noise both distorts the transition probabilities and introduces new transitions. For example, whenever observational noise causes $O(\sigma)$ to differ from the set of flaws truly present in $\sigma$, the controller may chose an action (from the support of $P(O(\sigma))$) under which there are transitions from $\sigma$ that were not present in the noise-free digraph. We model the overall distortion induced by noise by taking the noise-free digraph, which we think of as the \emph{principal} mechanism for state evolution, reducing the probabilities on all its edges uniformly by a factor of $1-p$, and allowing the leftover probability mass to be distributed arbitrarily, in order to form the noise. More precisely:
\begin{itemize}
\item
Let $D_{\mathrm{pr}}$ be the digraph on $\Omega$ of possible state-transitions under policy $P$, with a self-loop added at every flawless state. Let $\rho_{\mathrm{pr}}$ be the $P$-induced state-transition probability distribution, augmented so that all self-loops at flawless states have probability 1.
\item
Let $D_{\mathrm{ns}}$ be an {\bf arbitrary} digraph on $\Omega$. For each vertex $\sigma$ in $D_{\mathrm{ns}}$, let $\rho_{\mathrm{ns}}(\sigma,\cdot)$ be an {\bf arbitrary} probability distribution on the arcs leaving $\sigma$.
\item
We will analyze the Markov chain on $\Omega$ which at every $\sigma \in \Omega$, with probability $p$ follows an arc in $D_{\mathrm{ns}}$ and with probability $1-p$ follows an arc in $D_{\mathrm{pr}}$. Formally, for every $\sigma \in \Omega$,
\begin{align*}
\rho(\sigma,\cdot) = (1-p) \cdot \rho_{\mathrm{pr}}(\sigma, \cdot) + p \cdot \rho_{\mathrm{ns}}( \sigma, \cdot ) \enspace.
\end{align*}
We assume that the system starts at a state $\sigma_1$, according to some unknown probability distribution $\theta$.
\end{itemize}
Requiring that the effect of noise is captured by a mixture of the original (principal) chain and an arbitrary chain is the only assumption that we make. In particular, by allowing $D_{\mathrm{ns}}$ and $\rho_{\mathrm{ns}}$ to be arbitrary we forego the need to posit specific models of observational and environmental noise, lending greater generality to our results. To see this, let $U(\sigma)$ denote the set of flaws actually present in $\sigma$ (and, slightly abusing notation, also the characteristic vector of $U(\sigma) \subseteq F$). In any step where the state transition distribution is not the principal one, we can think of this as occurring because $O(\sigma) \neq U(\sigma)$ and the distribution corresponds to $P(O(\sigma))$, or because $O(\sigma) \neq U(\sigma)$ and the distribution does not \emph{even} correspond to $P(O(\sigma))$, or because $O(\sigma) = U(\sigma)$ but, silently, the distribution followed is different from $P(O(\sigma))$. In particular, notice that whenever $O(\sigma) = \mathbf{0}$, the controller thinks she has arrived at a flawless state and, thus, authorizes a self-loop with probability 1. In such a case, the fact that the system will follow $\rho_{\mathrm{ns}}$ with probability $p$ means that we are allowing the noise not only to trick the controller to inactivity but also to silently move the system to a new state. Similarly, after the system arrives at a flawless state, i.e., $U(\sigma) = \mathbf{0}$, with probability $p$ it will then follow an arc in $D_{\mathrm{ns}}$, potentially to a flawed state. We allow this to occur to be consistent with (i) the idea that observational noise can occur at any state, even a flawless one, thus causing unneeded, potentially detrimental action, and (ii) with the idea that flaws can be introduced spontaneously from the environment at any state. Our goal is, thus, to prove that from \emph{any} initial state, after a small number of steps, the system will reach a flawless state, despite the noise. As we will see, what will matter about the noise is the extent to which noise-induced transitions introduce flaws in the state.
Let $D = D_{\mathrm{pr}} \cup D_{\mathrm{ns}}$. To avoid certain trivialities we will assume that there exists a constant $B < \infty$ such that $2^{-B} < \rho(\sigma,\tau) < 1-2^{-B}$ for every arc $(\sigma,\tau) \in D$. For each state $\sigma$, we denote the highest priority flaw present in $\sigma$ by $\pi(\sigma)$; if $\pi(\sigma) = f_i$, we label all arcs leaving $\sigma$ as $\sigma \xrightarrow{i} \cdot$, i.e., with the index of the flaw to which we attribute the transition (we use $i$ instead of $f_i$ as the label to lighten notation). We will refer to $\pi(\sigma)$ as the flaw \emph{addressed} at $\sigma$.
\begin{causality}
For an arc $\sigma \xrightarrow{i} \tau$ in $D$ and a flaw $f_j$ present in $\tau$ we say that $f_i$ causes $f_j$ if $f_j \not\ni \sigma$.
The digraph on $[m]$ where $i \rightarrow j$ iff $D$ contains an arc such that $f_i$ causes $f_j$ is the causality digraph $C(D)$.
\end{causality}
\begin{neigh}
The \emph{neighborhood} of a flaw $f_i$ in $C=C(D)$ is $\Gamma (f_i) = \{ f_i \} \cup \{f_j : i \to j \text{ exists in $C$}\}$.
\end{neigh}
For our condition we will need to bound from \emph{below} the entropy injected into the system in each step. To that end we define the potential of each flaw $f_i$ to be
\begin{equation}\label{eq:pot_def}
\mathrm{Potential}(f_i) = \min_{\sigma: \pi(\sigma) = f_i} H[ \rho(\sigma, \cdot) ] \enspace .
\end{equation}
We extend the definition to sets of flaws i.e., $\mathrm{Potential}(S) = \sum_{f \in S} \mathrm{Potential}(f)$, where $\mathrm{Potential}(\emptyset) = 0$.
In the absence of noise, $\mathrm{Potential}(f_i)$ expresses a lower bound on the diversity of ways to address flaw $f_i$, by bounding from below the ``average number of random bits consumed" whenever $f_i$ is addressed. Thus, it bounds from below the rate at which the controller explores the state space \emph{locally}.
The presence of noise may decrease or may increase the potential. For example, if all arcs in $D_{\mathrm{ns}}$ are self-loops, then the noise is equivalent to the action-buttons ``sometimes not working" and its only (and very benign) effect is to slow down the exploration by a constant factor. At the other extreme, if $D_{\mathrm{ns}}$ is the complete digraph on $\Omega$ and $\rho_{\mathrm{ns}}$ is uniform, then (unless $p$ is extremely small) the situation is, clearly, hopeless. Correspondingly, even though the potential has been greatly increased, the causality relationship is complete. We note that, trivially, the potential of each flaw is bounded from below by the minimum entropy injected by the principal alone whenever the flaw is addressed, i.e., $\mathrm{Potential}(f_i) \ge (1-p) \min_{\sigma : \pi(\sigma) = f_i} H[\rho_{\mathrm{pr}}(\sigma)]$.\smallskip
The other important characteristic of each flaw $f_i$ is its congestion, i.e., the maximum number of arcs with label $i$ that lead to the same state. For the same reason we would like the potential of a flaw to be big, we would like its congestion to be small: if arcs from different states in $f_i$ lead to the same state, then exploration slows down and the entropy injected into the system must be appropriately discounted in order to yield a good measure of the rate of state space exploration. To see this observe that $\mathrm{Potential}(f_i)$ is independent of the destinations of the arcs leaving $f_i$ and compare the case where these destinations are all distinct with the case where they all lie in a small (bottleneck) set. As the congestion due to the principal and the congestion due to noise will have different effects, we need to account for them separately. Let $A_{\mathrm{pr}}(\sigma)$ denote the support of $\rho_{\mathrm{pr}}(\sigma, \cdot)$ and $A_{\mathrm{ns}}(\sigma)$ denote the support of $\rho_{\mathrm{ns}}(\sigma, \cdot)$.
\begin{backw}
For any flaw $f_i \in F$, let
\begin{eqnarray*}
\mathrm{Congestion}_{\mathrm{pr}}(f_i) & = & \max_{\tau \in \Omega} |\{\sigma \in f_i: \tau \in A_{\mathrm{pr}}(\sigma)\}| \\
\mathrm{Congestion}_{\mathrm{ns}}(f_i) & = & \max_{\tau \in \Omega} |\{\sigma \in f_i: \tau \in A_{\mathrm{ns}}(\sigma)\}|
\enspace .
\end{eqnarray*}
Let $\mathrm{b}_{\mathrm{pr}}^{f_i} = \log_2 \mathrm{Congestion}_{\mathrm{pr}}(f_i)$. Let $\mathrm{b}_{\mathrm{ns}}^{f_i} = \log_2 \mathrm{Congestion}_{\mathrm{ns}}(f_i)$. Let $\mathrm{b}_{\mathrm{ns}} = \max_{f_i \in F} \mathrm{b}_{\mathrm{ns}}^{f_i}$.
\end{backw}
Let $C_{\mathrm{pr}}$ and $C_{\mathrm{ns}}$ be the causality graphs of $D_{\mathrm{pr}}$ and $D_{\mathrm{ns}}$, respectively, and let $\Gamma_{\mathrm{pr}}(f_i)$ and $\Gamma_{\mathrm{ns}}(f_i)$ be the corresponding neighborhoords. Let $\Delta_i = \left|\Gamma_{\mathrm{ns}} (f_i)\right|$. Recall that $h( p ) = - p \log_2 p - (1-p) \log_2 (1 - p )$ is the \emph{binary entropy} of $p \in [0,1]$. To express the lost efficiency due to noise in addressing flaw $f_i$, we let
\begin{eqnarray*}
q_i(p) & = & p\left( \Delta_i\left( \mathrm{b}_{\mathrm{ns}}+\frac{5}{2}+h(p)\right) - 2 - h(p)\right) \\
& \leq & p \, \Delta_i (\mathrm{b}_{\mathrm{ns}}+4) \enspace .
\end{eqnarray*}
Observe that $q_i(p)$ is independent of the policy and that its leading term is $p \Delta_i$. This means that, unlike the LLL, the number, $\Delta_i$, of different flaws that may be introduced when addressing a flaw can be arbitrarily large if the \emph{frequency of interactions} between flaws, captured by $p$, is sufficiently small. Our main result establishes a condition under which the probability of not reaching a flawless state within $O(\log_2 |\Omega | +m)$ steps is exponentially small. To state it wefine for each flaw $f_i$,
\[
\mathrm{Amenability}(f_i) = \mathrm{Potential}(f_i) - \mathrm{b}_{\mathrm{pr}}^{f_i} \enspace .
\]
\begin{theorem}\label{Molloy}
If for every flaw $f_i \in F$,
\begin{align}\label{eq:amenable}
\sum_{f_j \in \Gamma_{\mathrm{pr}}(f_i)} 2^{-
\mathrm{Amenability}(f_j)+ q_j(p) } < 2^{-(2+h(p))} \enspace ,
\end{align}
then there exists a constant $R>0$ depending on the slack in~\eqref{eq:amenable}, such that for every $s>1/2$, the probability of not reaching a flawless state after $Rs(\log_2 |\Omega | +m)$ steps is less than $\exp(-s)$.
\end{theorem}
\begin{remark}
In the noiseless case, i.e., when $p=0$, equation \eqref{eq:amenable} becomes an asymmetric LLL criterion. In particular, the main result of~\cite{JACM} is that if $\mathrm{b}_{\mathrm{pr}}^{f_i} = 0$ and all distributions $\rho_{\mathrm{pr}}(\sigma,\cdot)$ are uniform over their support $A_{\mathrm{pr}}(\sigma)$, then, a sufficient condition for reaching a flawless state quickly is that for every $f_i \in F$, $\sum_{f_j \in \Gamma_{\mathrm{pr}}(f_i)} 1/a_j < 1/\mathrm{e}$, where $a_j = \min_{\sigma \in f_j: \pi(\sigma) = f_j} |A_{\mathrm{pr}}(\sigma)|$. We see that in this setting our condition~\eqref{eq:amenable} recovers this, up to the constant on the right hand side, i.e., $1/4$ vs. $1/\mathrm{e}$.
\end{remark}
\section{Related Work}\label{sec:related}
\subsection{POMDPs and the Reachability Problem}
Markov Decision Processes (MDPs) are widely used models for describing problems in stochastic dynamical systems~\cite{CompetitiveMDP,Puterman,bertsekas2012}, where an agent repeatedly takes actions to achieve a specific goal while the environment reacts to these actions in a stochastic way. In an MDP the agent is assumed to be able to \emph{perfecty} observe the current state of the system and take action based on her observations. In a \emph{partially} observable Markov Decision Process (POMDP) the agent only receives limited, and possibly inaccurate, information about the current state of the system. POMDPs have been used to model and analyze problems in artificial intelligence and machine learning such as reinforcement learning~\cite{chrisman1992reinforcement,RL}, planning under uncertainty~\cite{Planning}, etc.
Formally, a discrete POMDP is defined by the following primitives (all sets are assumed finite): (i) a \emph{state space} $\Omega$, (ii) a finite alphabet of \emph{actions} $\cal A$, (iii) an observation space $\cal O$, (iv) an {\em action}-conditioned state transition model $\Pr(\tau | \sigma, a)$, where $\sigma,\tau \in \Omega$ and $a \in \cal A$, (v) an observation model $\Pr(o | \sigma)$, where $\sigma \in \Omega $ and $o \in \cal O$, (vi) a cost function $\mathrm{c}: \Omega \mapsto \mathbb{R}$ (or more generally a map from state-action pairs to the reals), and (vii) a desired criterion to minimize, e.g., expected total cumulative cost $\sum_{t=0}^\infty \mathbb E \left[ \mathrm{c}(\sigma_t) \right]$, where $\sigma_t$ is the random variable that equals the $t$-th state of the trajectory of the agent. Finally, various choices of controllers are possible. For instance, a \emph{stochastic memoryless} controller is a map from the \emph{current} observation to a probability distribution over actions, whereas a \emph{belief-based} controller conditions its actions on probability distributions over the state space (i.e., beliefs) that are sequentially updated (using Bayes rule or some approximation of it) while the agent is interacting with the environment.
Unfortunately, the problem of computing an optimal policy for a POMDP, i.e., designing a controller that minimizes the expected cost, is highly intractable~\cite{PapaTsitsi,mundhenk2000complexity} and, in general, undecidable~\cite{madani1999undecidability}. Notably, the problem remains hard even if we severely restrict the class of controllers over which we optimize~\cite{PapaTsitsi, littman1998computational, etessami2010complexity, Nikos}. As far as we know, the only tractable case~\cite{Nikos} requires both the cost function and the class of controllers over which we optimize to be extremely restricted. In particular, the controller can not observe or remember anything and must apply the same distribution over actions in every step.
An important special case that has motivated our work is the \emph{reachability} problem for POMPDs. Here, one has a set of \emph{target} states $T \subseteq \Omega$, and the goal is to design a controller that starting from a distribution $\theta$ over $\Omega$, guides the agent to a state in $T$ (almost surely) with the optimal expected total cumulative cost. As shown in~\cite{Chater}, the problem is undecidable in the general case. In the same work, for the case where the costs are positive integers and the observation model is deterministic, i.e., the observations induce a partition of the state space, the authors give an algorithm which runs in time doubly-exponential in $|\Omega|$ and returns doubly-exponential lower and upper bounds for the optimal expected total cumulative cost, using a belief-based controller. On the other hand, our work establishes a sufficient condition for a stochastic memoryless controller to reach the target set $T$ rapidly (in time logarithmic in $|\Omega|$ and linear in $|F|$), in the case where each individual observation is binary valued (set membership) and the observation model is arbitrarily stochastic. To our knowledge, this is the first tractability result for a nontrivial class of POMDPs under stochastic memoryless controllers.
\subsection{Focusing and Prioritization}\label{sec:foc}
To achieve our results the controller must be focused and prioritize. The idea of focusing was introduced by Papadimitriou~\cite{papafocus} in the context of satisfiability algorithms, and amounts to ``if it ain't broken don't fix it", i.e., state evolution should only happen by changing the values of variables that participate in at least one violated constraint. One way to implement this idea is to always first select a violated constraint (flaw) and then take actions that tend to get rid of it. This has been an extremely successful idea in practice~\cite{walksat, circumspect} and it is often materialized by selecting a \emph{random} flaw to address in each step. We remark that our methods allow, in fact, also the analysis of controllers that address a random flaw in each step, but for simplicity of exposition we chose to only present the case of a fixed permutation (prioritization).
Focusing is not only a good algorithmic idea, but also enables proofs of termination. Specifically, at the foundation of the argument of Moser and Tardos~\cite{MT} is the following observation: whenever an algorithm (focused or not) takes $t$ or more steps to reach a flawless state, say through flawed states $\sigma_1, \sigma_2, \ldots, \sigma_t$, there exists, by definition, a sequence of flaws $w_1, w_2, \ldots, w_t$ such that $\sigma_i \in w_i$. Therefore, by establishing a (potentially randomized) rule for selecting a flaw present in the state at each step, we can construct a random variable $W_t = w_1, w_2, \ldots, w_t$ to act as a \emph{witness} of the fact that the algorithm took at least $t$ steps. While, though, prima facie all constructions are equivalent, our capacity to bound the set of all possible such sequences is not. In particular, if the algorithm is focused and in each step we report the flaw on which the algorithm focused, then we can take advantage of the following observation: each appearance of a flaw $f_i$ in the witness sequence, with the potential exception of the very first, must be preceded by a distinct appearance of a flaw $f_j$ that causes $f_i$. This allows us to bound the rate at which the entropy of the set of $t$-witness sequences grows with $t$. Of course, in a general setting, there is good reason to believe that prioritization, i.e., focusing on the flaw determined by a fixed permutation, will be not be the best one can do. In particular, observe that for the same $D_{\mathrm{pr}}$, different permutations $\pi$ give rise to different causality graphs. On the other hand, at the level of generality of this work, i.e., without any assumptions about the system at hand, we can not really hope for a more intelligent choice.
\subsection{LLL algorithmization}
The Lov\'{a}sz Local Lemma (LLL)~\cite{LLL} is a non-constructive method for proving the \emph{existence} of flawless states that has served as a cornerstone of the probabilistic method. To use the LLL one provides a probability measure $\mu$ on $\Omega$, often the uniform measure, transforming flaws to (``bad") events, so that the existence of flawless states is equivalent to $\mu(\bigcup_{i=1}^m f_i) < 1$. The key quantity to control in order to prove this is negative dependence, i.e., the extent to which the probability of a bad event may be increased (boosted) by conditioning on the non-occurrence of other bad events. Roughly speaking, the LLL requires that for each bad event $f$, only a small number of other bad events should be able to boost $\mu(f)$ in this manner, whereas conditioning on the non-occurrence of all other bad events should not increase $\mu(f)$. Representing the boosting relationship in a graphical manner, with vertices corresponding to bad events pointing to their potential boosters, at a high level, the LLL requirement is that this digraph is sparse.
As one can imagine, whenever one proves that $\Omega$ contains flawless objects via the LLL it is natural to then ask if some such object can be found efficiently. Making the LLL constructive has been a long quest, starting with the work of Beck~\cite{beck_lll}, with subsequent works of Alon~\cite{alon_lll}, Molloy and Reed~\cite{mike_stoc}, Czumaj and Scheideler~\cite{Czumaj_lll}, Srinivasan~\cite{aravind_08} and others. Each of these works established a method for finding flawless objects efficiently, but with additional conditions relative to the LLL. A breakthrough was made by Moser~\cite{moser} who gave a very elegant algorithmization of the LLL for satisfiability via entropy compression. Very shortly afterwards, Moser and Tardos in a landmark paper~\cite{MT} made the LLL constructive for every product measure $\mu$. Specifically, they proved that if one starts by sampling an initial state according to $\mu$, and in every step selects an arbitrary occurring bad event and resamples its variables according to $\mu$, then with high probability a flawless state will be reached within $O(m)$ steps.
Following~\cite{MT}, several works~\cite{szege_meet, SrinivasanPerm, AI, JACM, HV, Harmonic, Commu} have extended the scope of LLL algorithmization beyond product measures. In these works, unlike~\cite{MT}, one has to also provide either an explicit algorithm~\cite{szege_meet, SrinivasanPerm}, or an algorithmic framework~\cite{JACM,Harmonic,HV, Commu}, along with a way to capture the \emph{compatibility} between the algorithm's actions for addressing each flaw $f_i$ and the measure $\mu$.
As was shown in~\cite{HV, Harmonic, Commu}, one can capture compatibility by letting
\begin{align}\label{charge}
d_i = \max_{\tau \in \Omega } \frac{\nu_i(\tau)}{ \mu(\tau) } \ge 1 \enspace ,
\end{align}
where $\nu_i(\tau)$ is the probability of ending up at state $\tau$ at the end of the following experiment: sample $\sigma \in f_i$ according to $\mu$ and address flaw $f_i$ at $\sigma$.
An algorithm achieving $d_i = 1$ is a \emph{resampling oracle} for flaw $f_i$. If $d_i =1$ for every $i \in [m]$, then it was proven in~\cite{HV} that the causality digraph equals the boosting digraph mentioned above and the condition for success is identical to that of the LLL (observe that the resampling algorithm of Moser and Tardos~\cite{MT} is trivially a resampling oracle for every flaw). More generally, ascribing to each flaw $f_i$ the \emph{charge} $\gamma(f_i) = d_i \cdot \mu(f_i)$, yields the following user-friendly algorithmization condition~\cite{Harmonic}, akin to the asymmetric Local Lemma: if for every flaw $f_i \in F$,
\begin{align} \label{LLL}
\sum_{ f_ j \in \Gamma(f_i) } \gamma(f_j) < \frac{1}{4} \enspace ,
\end{align}
then with high probability the algorithm will reach a sink after $O( \log | \Omega | + m )$ steps.
Even though the noiseless case is only tangential to the main point of this work, as an indication of the sharpness of our analysis, we point out that in the noiseless case, our condition~\eqref{eq:amenable} is identical to~\eqref{LLL} with $\gamma(f_i)$ replaced by $\chi(f_i) := 2^{-\mathrm{Potential}( f_i ) + b_{\mathrm{pr}}^{f_i}}$. In general, $\gamma(f_i)$ and $\chi(f_i)$ are incomparable. Roughly speaking, settings where $b_{\mathrm{pr}}^{f_i}$ is small and $d_i$ is large favor $\chi(f_i)$ over $\gamma(f_i)$ and vice versa, while the two meet when $b_{\mathrm{pr}}^{f_i}=0$, $\mu$ is uniform, and the transition probabilities are uniform, as in~\cite{JACM}.
In terms of techniques, as hinted in Section~\ref{sec:foc}, proofs of LLL algorithmizations consist of two independent parts. In one part, one bounds from above the probability of any witness sequence occurring, or in the case of Moser's entropic argument, bounds from below the entropy injected to the system while addressing the sequence. In the other part, one has to estimate the [entropy of the] set of possible witness sequences, using syntactic properties considerations mandated by causality: roughly speaking every occurrence of a flaw in a witness sequence, with the potential exception of the very first, must be preceded by an occurrence of some flaw that causes it. Finally, one compares the rate at which the probability of a $t$-step witness sequence decreases (or the rate at which entropy is increased) with the rate at which the [entropy of the] set of possible witness sequences increases, to establish that their product tends to 0 with $t$.
In this paper, exactly because we aim to capture the intensity of interactions between flaws under adversarial noise, we need to take a different approach. In particular, our proof can be thought of as entangling the two parts described above in order to establish that, while adversarial noise can make the imposed syntactic requirements inherited by the causality graph very weak (by making the graph extremely dense), the fact that the intensity of the noise is low, suffices to control the growth rate of the entropy of the set of witness sequences. The result is a carefully tuned argument that amortizes the entropy injected into the system against its effect on the entropy of the set of Break Forests. Key to the capacity to perform this amortization is the use of so-called Break Forests, introduced in~\cite{AI}, which localize in time the introduction of new flaws in the state. This property of Break Forests was not used in earlier works~\cite{AI,Harmonic} and allows us to use a different amortization for the flaws introduced by the principal vs.\ those introduced by noise.
\section{Termination via Compression}
Our analysis will not depend in any way on the distribution $\theta$ of the initial state. As a result, without loss of generality, we can assume that the process starts at an arbitrary but fixed state $\sigma_{\mathrm{init}}$. We let $A(\sigma)$ denote the support of $\rho(\sigma, \cdot)$, i.e., $A(\sigma)$ is the set of all states reachable by the process in a single step from $\sigma$.
\begin{definition}
We refer to the (random) sequence $\sigma_{\mathrm{init}} = \sigma_1,\ldots,\sigma_{t+1}$, entailing the first $t$ steps of the process, as the \emph{$t$-trajectory}. A $t$-trajectory is \emph{bad} iff $\sigma_1,\ldots,\sigma_{t+1}$ are all flawed.
\end{definition}
We model the set of all possible trajectories as an infinite tree whose root is labelled by $\sigma_1 = \sigma_{ \mathrm{init} }$. The root has $|A(\sigma_{\mathrm{init}})|$ children corresponding to (and labelled by) each possible value of $\sigma_2$. More generally, a vertex labeled by $\sigma$ has $|A(\sigma)|$ children, each child labeled by a distinct element of $A(\sigma)$, i.e., a distinct possible value of $\sigma_{i+1}$. Every edge of this infinite vertex-labelled tree is oriented away from the root and labelled by the probability of the corresponding transition, i.e., $\rho(\sigma,\tau)$, where $\sigma$ is the parent and $\tau$ is the child vertex. By our assumption, every such edge label is at least $2^{-B}$.
We call the above labelled infinite tree the \emph{process tree} and note that it is nothing but the unfolding of the Markov chain corresponding to the state-evolution of the process. In particular, for every vertex $v$ of the tree, the probability, $p_v$, that an infinite trajectory will go through $v$ equals the product of the edge-labels on the root-to-$v$ path. In visualizing the process tree it will be helpful to draw each vertex $v$ at Euclidean distance $-\log_2 p_v$ from the root. This way all trajectories whose last vertex is at the same distance from the root are equiprobable, even though they may entail wildly different numbers of steps (this also means that sibling vertices are not necessarily equidistant from the root). Finally, we color the vertices of the process tree as follows. For every infinite path that starts at the root determine its maximal prefix forming a bad trajectory. Color the vertices of the prefix red and the remaining vertices of the path blue.
In terms of the above picture, our goal will be to prove that there exist a critical radius $x_0$ and $\delta >0$, such that the proportion of red states at distance $x_0$ from the root is at most $1-\delta$. Crucially, $x_0$ will be polynomial, in fact linear, in $m=|F|$ and $ \log_2 | \Omega | $. Since we will prove this for every possible initial state and the process is Markovian, it follows that the probability that the process reaches distance $x$ from the root while going only through red states is at most $(1-\delta)^{\lfloor x/x_0 \rfloor}$.
To prove that red vertices thin out as we move away from the root we stratify the process tree as follows. Fix any real number $x>0$ and on each infinite path from the root mark the first vertex of probability $2^{-x}$ or less, i.e., the first vertex that has distance at least $x$ from the root. Truncate the process tree so that the marked vertices become leaves of a finite tree. Let $L(x)$ be the set of all root-to-leaf paths (trajectories) in this finite tree and let $B(x) \subseteq L(x)$ consist of the bad trajectories. Now, let $I$ be the random variable equal to an infinite trajectory of the process and let $\Sigma = \Sigma(x)$ be the random variable equal to the prefix of $I$ that lies in $L(x)$. By definition, $\sum_{\ell \in L(x)} \Pr[\Sigma = \ell] = 1$, while $\Pr[\ell] \in (2^{-x-B},2^{-x}]$ for every $\ell \in L(x)$, since $-\log_2 \rho \ge B$. Let $P = P(\Sigma)$ be the maximal red prefix of $\Sigma$ and observe that if $\Sigma \in B(x)$ then $P = \Sigma$. Therefore,
\begin{equation}\label{beauty}
H[P] \; \ge \sum_{\ell \in B(x)} \Pr[\Sigma = \ell] (-\log_2 \Pr [ \Sigma =\ell]) \; \ge \; x \sum_{\ell \in B(x)} \Pr[\Sigma = \ell] \; = \; x \Pr[\Sigma \in B(x)] \enspace .
\end{equation}
Assume now that there exist $M_0 > 0$ and $\lambda < 1$, such that $H[P] \le \lambda x + M_0$, for every $x>0$. Then~\eqref{beauty} implies that for $x_0 = 2 M_0/ (1 - \lambda )$,
\begin{equation}\label{eq:posi}
\Pr[\Sigma \in B(x_0)] \le \frac{H[P]}{x_0} \le \frac{ \lambda x_0 + M_0}{x_0} = \lambda+ \frac{1- \lambda}{2} = \frac{1 + \lambda}{2} < 1 \enspace .
\end{equation}
If $\Sigma \in B(x_0)$, we treat the reached state as the root of a new finite tree and repeat the same analysis, as it is independent of the starting state. It follows in this manner that for every integer $T \ge 1$, the probability that the process reaches a state at distance $T(x_0 + B)$ or more from the root by going only through red states is at most $\left((1+\lambda)/2\right)^T$. Thus, for any $s > 1/2$, the probability that the process reaches a state at distance
\[
E = \left\lceil\frac{2s}{1+\lambda} \right\rceil (x_0 + B) = O\left(\frac{s M_0}{1-\lambda^2} \right)
\]
or more from the root by going only through red states is at most $\left((1+\lambda)/2\right)^{\left\lceil \frac{2s} {1 +\lambda}\right\rceil} < \exp(-s)$.
Since $\rho(\sigma,\tau) < 1 - 2^{- B}$, it follows that $ - \log_2 \rho(\sigma,\tau)
> 2^{-B}$, for every arc in $D$. Thus, after $ 2^{B} E $ steps the process is always at distance $E$ or more from the root. Thus, the probability of not reaching a flawless state after $ 2^{B} E= O\left(\frac{s M_0}{1-\lambda^2} \right)$ steps is $\exp(-s)$. Therefore Theorem~\ref{Molloy} follows from the following.
\begin{theorem}\label{Amenability}
Let $\Xi = \max \{ \mathrm{b}_{\mathrm{ns} }, \mathrm{b}_{\mathrm{pr}} \}$ and $\Delta = \max_{j \in F} \Delta_j$. If there exists $\lambda < 1$ such that for all $j \in [m]$,
\[
\sum_{f_i \in \Gamma_{\mathrm{pr}}(f_j)} 2^{- (\lambda \mathrm{Potential}(f_i) - \mathrm{b}_{\mathrm{pr}}^{f_i} - q_i(p) )} < 2^{-(2+h(p))} \enspace ,
\]
then $H[P] \le \lambda x + M_0$ for every $x>0$, where $M_0 = \log_2 |\Omega | + m (\Delta+1)(\Xi+4) + \lambda B$.
\end{theorem}
\section{Break Sequences}
Recall that $\pi$ is an arbitrary but fixed ordering of the set of flaws $F$ and that the highest flaw present in each state $\sigma$ is denoted by $\pi(\sigma)$. We will refer to $\pi(\sigma)$ as the flaw \emph{addressed} at state $\sigma$, i.e., as in the noiseless case, even though the action distribution $P(O(\sigma))$ may be ``misguided" whenever $O(\sigma) \neq U(\sigma)$.
\begin{definition}
Given a bad $t$-trajectory $\Sigma$, its \emph{witness} sequence is $W(\Sigma) = w_1, \ldots, w_t = \{\pi(\sigma_i)\}_{i=1}^t$.
\end{definition}
To prove Theorem~\ref{Amenability}, i.e., to gain control of bad trajectories and thus of $H[P]$, we introduce the notion of \emph{break sequences} (see also~\cite{AI,Harmonic}). Recall that $U(\sigma)$ denotes the set of flaws present in $\sigma$.
\begin{definition}
Let $B_0 = U(\sigma_1)$. For $1 \le i\le t-1$, let $B_i = U(\sigma_{i+1}) \setminus ( U (\sigma_i) \setminus w_i )$.
\end{definition}
Thus, $B_i$ is the set of flaws ``introduced" during the $i$-th step, where if a flaw is addressed in a step but remains present in the resulting state we say that it ``introduced itself". Each flaw $f \in B_i$ may or may not be addressed during the rest of the trajectory. For example, $f$ may get fixed ``collaterally" during some step taken to address some other flaw, before the controller had a chance to address it. Alternatively, it may be that $f$ remains present throughout the rest of the trajectory, but in each step $i < j \le t-1$ some other flaw has greater priority than $f$. It will be crucial to identify and focus on the subset of flaws $B_i^* \subseteq B_i$ that \emph{do} get addressed during the $t$-trajectory, causing entropy to enter the system. Per the formal Definition~\ref{def:bs} below, the set of such flaws is $B_i^* = B_i \setminus \{O_i \cup N_i \}$, where $O_i$ comprises any flaws in $B_i$ that get eradicated collaterally, while $N_i$ comprises any flaws in $B_i$ that remain present in every subsequent state after their introduction without being addressed.
\begin{definition}\label{def:bs}
The \emph{Break Sequence} of a $t$-trajectory is $B_0^*, B_1^*, \ldots, B_{t}^*$, where for $0 \le i \le t$,
\begin{align*}
B_i^* & = B_i \setminus \{O_i \cup N_i \} \enspace , \text{where}\\
O_i & = \{f \in B_i \mid \exists j \in [i+1,t] :
f \notin U(\sigma_{j+1}) \wedge \forall \ell \in [i+1,j]: f \ne w_{\ell} \} \enspace , \\
N_i & = \{f \in B_i \mid \forall j \in [i+1, t] :
f \in U(\sigma_{j+1}) \wedge \forall \ell \in [i+1,t]: f \ne w_{\ell} \} \enspace .\
\end{align*}
\end{definition}
Given $B_0^*,B_1^*,\ldots,B_{i-1}^*$ we can determine the sequence $w_1, w_2, \ldots, w_i$ of flaws addressed inductively, as follows. Define $E_1 = B_0^*$, while for $i \ge 1$, let
\begin{equation}\label{eq:ri}
E_{i+1} = (E_{i} - w_i) \cup B_i^* \enspace .
\end{equation}
Observe that, by construction, $E_i \subseteq U(\sigma_i)$ and $w_i \in E_i$. Therefore, for every $i$, the highest flaw in $E_i$ is $w_i$.
\section{Proof of Theorem~\ref{Amenability}}
For the analysis, we will assume that the state-transition distribution $\rho$ is realized in each step by flipping a coin with bias $p$ to determine if the state transition will occur according to $\rho_{\mathrm{pr}}(\sigma, \cdot)$ or $\rho_{\mathrm{ns}}(\sigma, \cdot)$. Let $I$ be the random variable equal to an infinite trajectory of the algorithm. For any fixed real number $x>0$, we define the following random variables.
\begin{itemize}
\item
Let $\Sigma$ be the prefix of $I$ in $L(x)$.
\item
Let $P = \sigma_1, \sigma_2, \ldots, \sigma_{Z+1}$ be the maximal bad prefix of $\Sigma$. Thus, $P$ consists of $Z$ steps.
\item
For $ i \ge 1$:
\begin{itemize}
\item
Let $\sigma'_i = \sigma_i$ for $i \le Z$, while $\sigma'_i = \emptyset$ for $i > Z$.
\item
Let $r_i = w_i(\Sigma)$ for $i \le Z$, while $r_i = \emptyset$ for $i > Z$.
\item
Let $n_i$ be indicator r.v.\ that $\rho_{\mathrm{ns}}$ was employed in the $i$-th step of $P$, while $n_i = 0$ for $i > Z$.
\end{itemize}
\item Let $N
= n_1, n_2, \ldots$
\item
Let $Y = Y(P) = B_0^*, B_1^*,\ldots,$ be the break sequence of $P$, where $B^*_i = \emptyset$ for $i > Z$.
\item
Let $Y_1$ be the suffix of $Y$ starting at $B_1^*$.
\item
Let $L = |B_0^*|, |B_1^*|, \ldots$
\end{itemize}
Observe that $L$ determines $Z$ since $\sum_{i=0}^t |B_i^*| \ge t$ for all $t \le Z$, with equality holding only for $t = Z$. To pass from~\eqref{full_seq} to~\eqref{pointwise} we use that $L$ determines $Z$ and that $Y$ determines the sequence $r_1,r_2,\ldots, r_{Z}$. To pass from~\eqref{verylong} to~\eqref{poules} we use that there is a 1-to-1 correspondence between the elements of the witness sequence $r_1, \ldots, r_{Z}$ and the $Z$ elements in the sets $B_0^*, B_1^*, \ldots$ Thus,
\begin{align}
H[P] & = H[Z, \sigma_1, \sigma_2, \ldots, \sigma_{Z+1} ] \nonumber \\
& \le H[L, \sigma_1, \sigma_{2}, \ldots, \sigma_{Z+1}, Y , N] \nonumber \\
& \le H[B_0^*, L] + H[N \mid L] + H[Y_1 \mid B_0^*, L, N] + H[\sigma_{Z+1}]
+ \sum_{i \ge 2 } H[\sigma'_{i-1} \mid \sigma'_i, Y, N, L] \label{full_seq} \\
& \le H[B_0^*, L] + H[N \mid Z] + H[Y_1 \mid B_0^*, L, N] + \log_2 |\Omega|
+ \sum_{i \ge 2 } H[\sigma'_{i-1} \mid \sigma'_i, r_{i-1},n_{i-1} , Z] \label{pointwise} \\
& \le H[B_0^*, L] + h(p) \cdot \mathbb E Z + H[Y_1 \mid B_0^*, L, N] + \log_2 |\Omega|
+ \sum_{i \ge 2 } \mathbb E \left[ (1-p) \mathrm{b}_{\mathrm{pr}}^{r_{i-1}} + p \mathrm{b}_{\mathrm{ns}}^{r_{i-1}} \right]\label{verylong} \\
& \le H[B_0^*, L] + h(p) \cdot \mathbb E Z + H[Y_1 \mid B_0^*, L, N] + \log_2 |\Omega|
+ \sum_{i \ge 0} \mathbb E \mathrm{In}(B_i^*) \label{poules} \enspace ,
\end{align}
where, recalling that $\mathrm{b}_{\mathrm{ns}} = \max_{j \in [m]} \mathrm{b}_{\mathrm{ns}}^{f_j}$, we define for an arbitrary set of flaws $S$,
\begin{eqnarray}
\mathrm{In}(S) & = & (1-p) \sum_{f \in S } \mathrm{b}_{\mathrm{pr}}^{f} + p |S| \mathrm{b}_{\mathrm{ns}} \nonumber \\
& := & (1-p) \mathrm{In}_{\mathrm{pr}}(S) + p \mathrm{In}_{\mathrm{ns}}(S) \label{eq:indef} \enspace .
\end{eqnarray}
To bound the right hand side of~\eqref{poules} we prove at the end of this section Lemmata~\ref{struct_code} and~\ref{meatless} presented below. In the rest of this section, all sums over index $i$ are sums over $i \ge 1$.
\begin{lemma}\label{struct_code}
$H[B_0^*,L] \le m + 2\mathbb E Z - \mathbb E |B_0^*|$, and $\sum_i \mathbb E |B_i^*| = \mathbb E Z - \mathbb E |B_0^*|$.
\end{lemma}
Using the chain rule for entropy to write $H[Y_1 \mid B_0^*, L, N] = \sum_{i} H[B_i^* \mid B_0^*, \ldots, B_{i-1}^*, L, N]$ and combining the inequality in Lemma~\ref{struct_code} with~\eqref{poules}, we see that $H(P)$ is bounded from above by
\[
m + 2 \mathbb E Z - \mathbb E |B_0^*| + h(p) \cdot \mathbb E Z + \sum_{i} H[B_i^* \mid B_0^*, \ldots, B_{i-1}^*, L, N] + \log_2 |\Omega| + \mathbb E \mathrm{In}(B_0^*) + \sum_i \mathbb E \mathrm{In}(B_i^*)
\]
Using the equality in Lemma~\ref{struct_code} to express $\mathbb E Z$ as a sum, we see that the line above is equal to
\begin{align*}
m + & (1+h(p))\mathbb E |B_0^*| + \log_2 |\Omega| + \mathbb E \mathrm{In}(B_0^*) \\
+ &\sum_{i}
\left\{
H[B_i^* \mid B_0^*,\ldots, B_{i-1}^*, L, N] +
\mathbb E
\left[
\mathrm{In}(B_i^*) + (2+h(p)) |B_i^*|
\right]
\right\}
\enspace
.
\end{align*}
For an arbitrary set of flaws $S$, we define
\begin{eqnarray*}
q(S) & = & \sum_{ f_j \in S} q_j(p) \enspace , \\
g(S) & = & \lambda^{-1} (p( 2 + h(p)) |S| + q(S)) \enspace .
\end{eqnarray*}
\begin{definition}
Say that $C_{\mathrm{pr}}$ is $\lambda$-amenable if the conditions of Theorem~\ref{Amenability} are satisfied.
\end{definition}
\begin{lemma}\label{meatless}
If $C_{\mathrm{pr}}$ is $\lambda$-amenable, then
\begin{align*}
\sum_i \{ H[B_i^* \mid B_0^*,\ldots,B_{i-1}^*, L, N] +\mathbb E \left[\mathrm{In}(B_i^*) + (2+h(p)) |B_i^*| \right] \} \le \lambda (x+B + \mathbb E g(B_0^*)) \enspace.
\end{align*}
\end{lemma}
Lemma~\ref{meatless} thus implies that under the conditions of Theorem~\ref{Amenability},
\[
H(P) \le m +(1 + h(p) ) \mathbb E |B_0^*| + \log_2 |\Omega| + \mathbb E \mathrm{In}(B_0^*) + \lambda(x+B + \mathbb E g(B_0^*)) \enspace .
\]
Recall that $\Xi = \max \{ \mathrm{b}_{\mathrm{ns} }, \mathrm{b}_{\mathrm{pr}}\}$ and that $\Delta = \max_{j \in F} \Delta_j$. Since $\mathbb E |B_0^*| \le m$ and $\mathbb E \mathrm{In}(B_0^*) \le m \Xi$, we conclude, as claimed in Theorem~\ref{Amenability}, that $H(P) \le M_0 +\lambda x$, where
\begin{eqnarray*}
M_0
& = & \log_2 |\Omega | + m \left( 2 +h(p) + \Xi \right) + \lambda B + \lambda \mathbb E g(B_0^*) \\
& \le & \log_2 |\Omega | + m \left( 2 +h(p) + \Xi \right) + \lambda B + [p (2+h(p))+ \max_{j \in F} q_j(p)] m \\
& \le & \log_2 |\Omega | + m \left( 6 + \Xi + \max_{j \in F} q_j(p)\right) + \lambda B \\
& \le & \log_2 |\Omega | + m (\Delta+1)(\Xi+4) + \lambda B \enspace . \\
\end{eqnarray*}
\begin{proof}[Proof of Lemma~\ref{struct_code}]
We will represent $B_0^*,L$ as a binary string $s$ of length $m+2Z-|B_0^*|$. Since $B_0^* \subseteq F$ the first $m$ bits of $s$ are the characteristic vector of $B_0^*$. We encode $L$ immediately afterwards, representing the $i$-th element of $L$, for each $i \in [Z]$, as $1^{|B_i^*|}0$. Decoding, other than termination, is trivial: after reading the first $m$ bits of $s$, the rest of the string is interpreted in blocks of the form $1^*0$. To determine termination we note that, by construction, $|B_0^*| + \sum_{i=1}^j |B_i^*| - j \ge 0$ for every $j \in [Z]$ with equality holding only for $j=Z$. Therefore, decoding stops as soon as equality holds for the first time. The representation of $L$ in this manner consists of $\sum_i |B_i^*|$ ones and $Z$ zeroes, i.e., of $2Z - B_0^*$ bits, since $\sum_i |B_i^*| = Z - |B_0^*|$.
For $\sum_{i} \mathbb E |B_i^*|$ the claim follows readily from the fact $\sum_i |B_i^*| = Z - |B_0^*|$.
\end{proof}
For an arbitrary set of flaws $S$, we define
\[
\mathrm{Potential}^{-}(S) \; = \; \mathrm{Potential}(S) - g(S) \enspace .
\]
Lemma~\ref{meatless} follows trivially by combining Lemmata~\ref{meat} and \ref{neato} below.
\begin{lemma}\label{meat}
If $C_{\mathrm{pr}}$ is $\lambda$-amenable, then for every $i \ge 1$,
\begin{align*}
H \left[B_i^* \mid B_0^*,\ldots,B_{i-1}^*, L, N \right] +\mathbb E[\mathrm{In}(B_i^*) + (2+h(p)) |B_i^*|]
\le \lambda \mathbb E[\mathrm{Potential}^{-}(B_i^*) + g(r_i)] \enspace.
\end{align*}
\end{lemma}
The proof of Lemma~\ref{meat} is presented in Appendix~\ref{sec:meat}.
\begin{lemma}\label{neato}
$\sum_i \mathbb E [ \mathrm{Potential}^{-}(B_i^*) + g(r_i) ]\le \left(x+B \right) +\mathbb E g(B_0^*) $.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{neato}]
Since there is a 1-to-1 correspondence between the elements of the witness sequence $r_1, \ldots, r_{Z}$ and the $Z$ elements in the sets $B_0^*, B_1^*, \ldots$
\begin{eqnarray}\label{xesemas}
\sum_{i} \mathrm{Potential}(r_i) & = & \mathrm{Potential}(B_0^*) + \sum_{i} \mathrm{Potential}(B_i^*) \nonumber \\
& \ge & \sum_{i} \left[ \mathrm{Potential}(B_i^*) - g(B_i^*) + g(r_i) \right] - g(B_0^*) \nonumber \\
& = & \sum_i [ \mathrm{Potential}^{-}(B_i^*) + g(r_i) ] - g(B_0^*) \label{terrible} \enspace.
\end{eqnarray}
The chain rule for entropy gives~\eqref{eq:chain_chain_chain}. Since the evolution of $\Sigma$ is Markovian, inequality~\eqref{eq:almost} would have been an equality if it were not for the possibility that $r_i = \emptyset$. Finally, inequality~\eqref{eq:hahaha} follows from the definition of potential~\eqref{eq:pot_def}. Thus,
\begin{eqnarray}
H[\Sigma] & = & \sum_{i}H[ \sigma_{i+1} \mid \sigma_{i} ] \label{eq:chain_chain_chain} \\
& = & \sum_{i} \sum_{\sigma \in \Omega } \Pr[ \sigma_{i} = \sigma ] \cdot H[ \sigma_{i+1} \mid \sigma_i = \sigma ] \nonumber \\
& \ge & \sum_{i} \sum_{j \in [m]} \Pr[r_i = f_j ] \sum_{\sigma \in f_j} \Pr[ \sigma_i = \sigma \mid r_i = f_j ] \cdot H[ \rho( \sigma, \cdot ) ] \label{eq:almost} \\
& \ge & \sum_{i} \sum_{j \in [m]} \Pr[ r_i = f_j] \cdot \mathrm{Potential}(f_j) \label{eq:hahaha} \\
& = & \sum_{i} \mathbb E \mathrm{Potential}(r_i) \enspace .\label{trajectory}
\end{eqnarray}
Combining~\eqref{xesemas} and~\eqref{trajectory} with the fact $H[ \Sigma ] \le x + B$ yields the lemma.
\end{proof}
\section{Proof of Lemma~\ref{meat}}\label{sec:meat}
We need to prove that if $C_{\mathrm{pr}}$ is $\lambda$-amenable, then for every $i \ge 1$,
\begin{equation}
H \left[B_i^* \mid B_0^*,\ldots,B_{i-1}^*, L, N \right] +\mathbb E[\mathrm{In}(B_i^*) + (2+h(p)) |B_i^*|]
\le \lambda \mathbb E[\mathrm{Potential}^{-}(B_i^*) + g(r_i)] \enspace. \label{july4}
\end{equation}
Recall that $B_i^* = r_i = \emptyset$ for $i > Z$. Since $L$ determines $Z$, it follows that the conditional entropy in the left hand side of~\eqref{july4} is 0 for $i>Z$. Thus, \eqref{july4} holds trivially for $i>Z$ since $\mathrm{In}(\emptyset) = \mathrm{Potential}^-(\emptyset) = g(\emptyset) = 0$.
In the rest of this section we consider an arbitrary but fixed $1 \le i \le Z$. For any such $i$, recall that $B_0^*, B_1^{*}, \ldots,B_{i-1}^{*}$ determine $r_i$ and $L$ determines $|B_i^*|$. Therefore,
\begin{eqnarray*}
H[B_i^* \mid B_0^*,\ldots,B_{i-1}^*, L, N] &\le& H[B_i^* \mid r_i , |B_i^*|, n_i ] \nonumber \\
& =& (1-p) \, H[ B_i^* \mid r_i, |B_i^*| , n_i = 0] + p \, H[ B_i^* \mid r_i, |B_i^*|, n_i = 1] \enspace .
\end{eqnarray*}
For $j \in [m]$ let us denote $ \Pr_j [\cdot] = \Pr \left[ \cdot \mid r_i = f_j \right]$, $ \mathbb E_j [ \cdot ] = \mathbb E \left[ \cdot \mid r_i = f_ j \right]$ and $ H_j[ \cdot ] = H \left[ \cdot \mid r_i = f_ j \right]$. With this notation, and recalling~\eqref{eq:indef}, we see that~\eqref{july4} follows from the inequalities in the following lemma, i.e., by multiplying each inequality with the probability that $r_i = f_j$ and summing up over $j \in [m]$.
\begin{lemma}\label{ns_lemma}
For every $j \in [m]$,
\begin{align*}
H_j [B_i^* \mid n_i = 0 , |B_i^*|]
+ \mathbb E_j[\mathrm{In}_{\mathrm{pr}}(B_i^*) ]
+ (2 + h(p)) \, \mathbb E_j \left[ |B_i^*| \mid n_i = 0 \right]
& \le (1-p)^{-1} \lambda \mathbb E_j \mathrm{Potential}^{-}(B_i^*) \\
\\
H_j[ B_i^* \mid n_i = 1, |B_i^*|]
+ \mathbb E_j[\mathrm{In}_{\mathrm{ns}}(B_i^*) ]
+ (2 + h(p)) \, \mathbb E_j \left[ |B_i^*| \mid n_i = 1 \right]
& \le p^{-1} \lambda g(f_j) \enspace .
\end{align*}
\end{lemma}
\begin{proof}
To lighten notation, let $p_k^c = \Pr_j[ n_i = c, |B_i^*| =k ]$ and $p_k^c(S) = \Pr_j[B_i^* = S \mid n_i = c , |B_i^*| = k]$.
We start with the simpler case $c=1$. For any $j \in [m]$, recalling that $\log_2 \binom{\Delta_j}{k} \le \Delta_j \cdot h(k/\Delta_j)$, we get
\begin{eqnarray}
H_j[ B_i^* \mid n_i = 1, |B_i^*|]
+ \mathbb E_j[\mathrm{In}_{\mathrm{ns}}(B_i^*) ]
+ (2 + h(p)) \, \mathbb E_j \left[ |B_i^*| \mid n_i = 1 \right]
& \le & \nonumber \\
\sum_{k } p_k^1 \bigg( H_j[B_i^* \mid n_i = 1 , |B_i^*| = k] + k (\mathrm{b}_{\mathrm{ns}} + 2+h(p)) \bigg) &\le & \nonumber \\
\sum_{k} p_k^1 \cdot \left( \Delta_j\,h (k/\Delta_j) + k ( \mathrm{b}_{\mathrm{ns}} + 2+ h(p) ) \right) \nonumber & \le & \nonumber \\
\max_{k \in [0,\Delta_j]} \left\{ \Delta_j\,h (k/\Delta_j) + k ( \mathrm{b}_{\mathrm{ns}} + 2+ h(p) ) \right\} & < & \nonumber \\
\Delta_j\left( \mathrm{b}_{\mathrm{ns}}+\frac{5}{2}+h(p)\right) \nonumber & = & \\
p^{-1} \left( q_j(p) + p (2+h(p))\right) \label{july7} & , &
\end{eqnarray}
since $q_j(p) = p\left( \Delta_j\left( \mathrm{b}_{\mathrm{ns}}+\frac{5}{2}+h(p)\right) - 2 - h(p)\right)$.
For the case $c=0$ we need some preparation. Observe that $C_{\mathrm{pr}}$ being $\lambda$-amenable implies that $\lambda \mathrm{Potential}(f_i) - q_i(p) \ge 2 + h(p)$ for every flaw $f_i$, as otherwise~\eqref{eq:amenable} would be violated. Therefore, we see that for every set $S \subseteq F$,
\begin{eqnarray}
\lambda \mathrm{Potential}(S) & \ge & q(S) + |S| \left( 2+ h(p)\right) \label{eq:july6} \\
\mathrm{Potential}^-(S) & \ge & 0 \enspace , \label{eq:potika}
\end{eqnarray}
where~\eqref{eq:potika} follows from~\eqref{eq:july6} since $\lambda \mathrm{Potential}^{-}(S) = \lambda \mathrm{Potential}(S) - q(S) - p|S| (2+h(p))$. The positivity of $\mathrm{Potential}^-$ also implies that for every set $B_i^*$,
\begin{eqnarray}
\mathrm{Potential}^{-} (B_i^*) & = & \mathrm{Potential}^{-}
\left(
B_i^* \cap \Gamma_{\mathrm{pr}}(r_i)
\right) +
\mathrm{Potential}^{-} \left(B_i^* \cap \overline{\Gamma_{\mathrm{pr}}(r_i)}\right) \\
& \ge &
\mathrm{Potential}^{-} \left(B_i^* \cap \Gamma_{\mathrm{pr}}(r_i)\right) \enspace .\label{eq:panagos}
\end{eqnarray}
Finally, for $j \in [m]$, we let $\mathcal{S}_k^{j}$ denote the set of all $k$-subsets of $\Gamma_{\mathrm{pr}} ( f_j )$.
We can now start working towards our goal, which is to prove that for every $j \in [m]$ the first line below is non-negative. We start by invoking~\eqref{eq:panagos} to prove the first inequality below and~\eqref{eq:july6} to prove the second.
\begin{align*}
\lambda \mathbb E_j \mathrm{Potential}^{-}(B_i^*) + (1-p)\bigg(- H_j [B_i^* \mid n_i = 0 , |B_i^*|] - \mathbb E_j[\mathrm{In}_{\mathrm{pr}}(B_i^*)] - (2 + h(p)) \, \mathbb E_j \left[ |B_i^*| \mid n_i = 0 \right]\bigg) & \ge \\
\lambda \mathbb E_j \mathrm{Potential}^{-}(B_i^* \cap \Gamma_{\mathrm{pr}}(r_i)) \\
+ (1-p) \bigg(- H_j [B_i^* \mid n_i = 0 , |B_i^*|] - \mathbb E_j[\mathrm{In}_{\mathrm{pr}}(B_i^*)] - (2 + h(p)) \, \mathbb E_j \left[ |B_i^*| \mid n_i = 0 \right]\bigg) & = \\
\sum_{k} p_k^0 \sum_{S \in \mathcal{S}_k^{j}} p_k^0(S)
\left\{
\lambda \mathrm{Potential}^{-}(S) + (1-p) \bigg(\log_2 p_k^0(S) - \mathrm{In}_{\mathrm{pr}}(S) - (2 + h(p)) k \bigg)
\right\}
& = \\
\sum_{k} p_k^0 \sum_{S \in \mathcal{S}_k^{j}} p_k^0(S)
\left\{
p \lambda \mathrm{Potential}^{-}(S)
+
(1-p) \bigg(
\lambda \mathrm{Potential}^{-}(S) + \log_2 p_k^0(S) - \mathrm{In}_{\mathrm{pr}}(S) - (2 + h(p) )k
\bigg)
\right\}
& = \\
(1-p) \sum_{k} p_k^0 \sum_{S \in \mathcal{S}_k^{j}} p_k^0(S)
\left\{
\frac{p\lambda \mathrm{Potential}^{-}(S)}{1-p} + \lambda \mathrm{Potential}^{-}(S) + \log_2 p_k^0(S) - \mathrm{In}_{\mathrm{pr}}(S) - (2+h(p)) k
\right\}
& \ge \\
(1-p) \sum_{k} p_k^0 \sum_{S \in \mathcal{S}_k^{j}} p_k^0(S) \left\{ \lambda \mathrm{Potential}(S) -q(S) +\log_2 p_k^0(S) - \mathrm{In} _{\mathrm{pr}}(S) - (2 + h(p) )k \right\} & \enspace .
\end{align*}
Letting $\zeta(S) = 2^{- \lambda \mathrm{Potential}(S) + q(S) + \mathrm{In}_{\mathrm{pr}}(S) }$ it thus suffices to prove that the left hand side of~\eqref{eq:lhso} is non-negative
\begin{align}
\sum_{k} p_k^0 \left\{-(2 + h(p) )k +\sum_{S \in \mathcal{S}_k^{j}} p_k^0(S) \log_2 \frac{p_k^0(S)}{\zeta(S)}\right\} \ge & \label{eq:lhso} \\
\sum_{k} p_k^0 \left\{- (2 + h(p) )k- \log_2 \left( \sum_{S \in \mathcal{S}_k^{j}} \zeta(S) \right)\right\} \label{entropia} \enspace ,
\end{align}
where the inequality follows from the log-sum inequality. To prove that the right hand side of~\eqref{entropia} is non-negative we note that for all $k \geq 1$,
\begin{eqnarray*}
\sum_{S \in \mathcal{S}_k^{j}} \zeta(S)
\le
\left(\sum_{A \in \mathcal{S}_{k-1}} \zeta(A)\right) \left(\sum_{B \in \mathcal{S}_{1}} \zeta(B) \right)
\le \ldots \le \left( \sum_{B \in \mathcal{S}_{1}} \zeta(B) \right)^k \enspace .
\end{eqnarray*}
Therefore, since $C_{\mathrm{pr}}$ is $\lambda$-amenable,
\begin{align*}
\sum_{S \in \mathcal{S}_k^{j}} \zeta(S) \le \left(\sum_{f_i \in \Gamma_{\mathrm{pr}}(f_j)} 2^{ - \lambda \mathrm{Potential}(f_i) + {\mathrm{b}_{ \mathrm{pr} }^{f_i}} + q_i(p) } \right)^k\le \left(2^{-(2+ h(p))}\right)^k \enspace .
\end{align*}
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
The scaling methods were introduced by Pinchuk \cite{p} and Frankel \cite{f} independently in 1980's as a technique to study bounded domains in $\mathbb{C}^n$ with noncompact automorphism group. These techniques have been developed further by many authors and have become an important tool to prove the results of \cite{p}, \cite{bp1}, \cite{k} and others.
Pinchuk's scaling sequence was constructed by a sequence of compositions of stretching maps, say $\Lambda_j$, and automorphisms $\phi_j$ of the given domain $\Omega$. If the automorphism group of $\Omega$ is noncompact then, in many cases, there is a sequence $\{\phi_j\} \subset Aut(\Omega)$ which contracts compact subsets successively to some boundary point. On the other hand, a sequence of stretching maps is a divergent sequence of shear maps, the composition of $\mathbb{C}$-affine maps and triangular maps. The general expectation is that there is a subsequence of the Pinchuk scaling sequence $\{ \sigma_j := \Lambda_j \circ \phi_j \}$ convergent to the limit map, say $\widehat{\sigma}$, uniformly on compact subsets of $\Omega$. If this limit were 1-1, then it would be a re-embedding of $\Omega$ into $\mathbb{C}^n$. If $\Omega$ is a domain in $\mathbb{C}$ with smooth boundary, then the image of the limit map turns out to be a half plane. This is the special case of the Riemann mapping theorem and therefore it seems natural to hope for the convergence of $\{ \sigma_j \}$ in all dimensions.
As the first result in the higher dimensions, Pinchuk proved that his scaling sequence has a subsequence that converges to a biholomorphism uniformly on compact subsets for the class of bounded strongly pseudoconvex domains in all dimensions \cite{p}; this proves Wong-Rosay theorem \cite{w,r}. And later, Bedford and Pinchuk \cite{bp1} showed the convergence of the sequence if the domain is bounded with a finite type boundary in the sense of D'Angelo \cite{d}.
One of the difficulties in proving the convergence is that the expected limit domain $\widehat{\Omega}$ is not bounded; its Kobayashi hyperbolicity is not {\it a priori} clear. Pinchuk considers, alternatively, the convergence of the backward scaling sequence $\{ \sigma_j^{-1} \}$. If the limit domain $\widehat{\Omega}$ is well-defined in some sense, then this sequence $\{ \sigma_j^{-1} \}$ always has a convergent subsequence by Montel's theorem. Now, a question arises naturally: is the limit map of the backward sequence 1-1? For the class of bounded domains in $\mathbb{C}^2$ whose boundaries are of finite type, Bedford and Pinchuk have given a general affirmative resolution \cite{bp1} (cf. \cite{bc} also). If the limit map is surjective, additionally, then it follows that the inverse of the limit map is actually the same as a subsequential limit of the initial Pinchuk scaling sequence. This establishes the Pinchuk scaling method.
The Frankel scaling sequence follows the same principle but its construction is different. Given a domain $\Omega$, a point $p \in \Omega$ and a sequence of automorphisms $\{ \phi_j \}$, it is defined directly by $\omega_j(\mathbf{z}) := [d\phi_j|_p]^{-1} (\phi_j (\mathbf{z}) - \phi_j (p))$. If $\{\phi_j (p)\}$ converges to some boundary point of $\Omega$, then $\{\phi_j\}$ cannot converge to another automorphism. In fact, $\lim_{j \rightarrow \infty} \det{(d\phi_j|_p)} = 0$. Then $[d\phi_j|_p]^{-1}$ diverges. So, Frankel's scaling method appears to be similar to Pinchuk's. The sequence $\{[d\phi_j|_p]^{-1}\}$ stretches in some sense, whereas the sequence $\{\phi_j\}$ contracts. Now one can naturally pose the question: when does Frankel's scaling sequence form a normal family? Frankel proved that it suffices for $\Omega$ to be convex and Kobayashi hyperbolic \cite{f}.
\medskip
The purpose of this article is summarized into the following:
\medskip
In Section 2, we introduce a special continuously-varying coordinate system, pertaining to the target boundary point. Using this coordinate system, we give another proof to the convergence of the Pinchuk scaling sequence on a bounded domain in $\mathbb{C}^2$ with smooth finite-type boundary. We feel that our proof is simpler and more straightforward than that of Berteloot/C\oe ur\'{e} \cite{bc}.
Section 3 concerns the Frankel scaling sequence. The convexity was essential for its convergence to a holomorphic embedding into $\mathbb{C}^n$. There has been a question whether it converges without convexity. Here, we give a modification of the Frankel scaling sequence so that they may converge also on some nonconvex domains, using a sequence $\{\psi_j\}$ of automorphisms of $\mathbb{C}^n$ that converges to another. Two examples are given to show several aspects of the (modified) Frankel scaling sequence.
\smallskip
Finally in Section 4, we observe that the limit maps, if they exist, of Pinchuk and modified Frankel's scaling sequences are equivalent. Notice that this generalizes a theorem of Kim/Krantz in \cite{kk} for the convex case.
In Appendix, we give a proof of an existence of the coordinate system introduced in Section 2.
\medskip
\section{The Pinchuk scaling sequence}
Recall the definition of the finite type in the sense of D'Angelo \cite{d}.
\begin{defn} \rm
Let $\Omega$ be a domain in $\mathbb{C}^n$ with smooth boundary. Let $q$ be a point in $\partial \Omega$ and $\rho$ be a local defining function of $\Omega$ at $q$. The {\it type} $\Delta (q) = \Delta (\Omega, q)$ at $q$ is the positive value defined by
\begin{equation} \nonumber
\Delta (q) := \sup_{h} \frac{\nu(\rho \circ h)}{\nu(h - q)},
\end{equation}
where $\nu(f)$ is the order of vanishing of $f$ at $0$ and the supremum is taken over all nontrivial analytic discs $h$ in $\mathbb{C}^n$ with $h(0) = q$.
The point $q$ is called a {\it finite type boundary point} of $\Omega$ if $\Delta (q)$ is finite. If all the boundary points of $\Omega$ is of finite type, then $\Omega$ is called a {\it domain with finite type boundary}.
\end{defn}
From now on, the main subject is the bounded domain $\Omega$ with noncompact automorphism group. Under this condition, there is a point $p \in \Omega$ and a sequence $\{\phi_j\} \subset Aut(\Omega)$ with $\lim_{j \to \infty} \phi_j(p) = \widehat{p}$ for some boundary point $\widehat{p}$. We call such sequence $\{\phi_j(p)\}$ a {\it boundary accumulation automorphism orbit} converging to $\widehat{p}$, and present the following improvement upon the scaling theorem of Pinchuk.
\begin{thm}
Let $\Omega$ be a bounded domain in $\mathbb{C}^2$ with smooth pseudoconvex boundary. Assume that $Aut(\Omega)$ admits a boundary accumulating automorphism orbit $\{\phi_j(p)\}$ converging to $\widehat{p} \in \partial \Omega$. If $\widehat{p}$ is of finite type in the sense of D'Angelo, then there is a sequence $\{ \Lambda_j \}$ in $Aut(\mathbb{C}^2)$ such that the sequence $\{ \Lambda_j \circ \phi_j \}$ has a subsequence that converges to a biholomorphism-into $\mathbb{C}^2$ uniformly on compact subsets of $\Omega$. Moreover, the image of $\Omega$ by the limit map is of the form $\{ (w, z) \in \mathbb{C}^2 \mid \textrm{Re}\,w + P(z, \bar{z}) < 0\}$ for some subharmonic polynomial $P$ with no harmonic terms.
\end{thm}
This stretching sequence $\Lambda_j$ is an automorphism of $\mathbb{C}^2$. Indeed, the map $\Lambda_j$ is a composition of shear maps (cf. Section 2.2). This property plays an important role in proving Theorem 4.1. We call this $\Lambda_j \circ \phi_j$ the ($j$-th) {\it Pinchuk scaling map} of the {\it Pinchuk scaling sequence} $\{ \Lambda_j \circ \phi_j \}$.
\bigskip
\subsection{An admissible coordinate system for finite type boundary.} To prove Theorem 2.2, we introduce a continuously-varying coordinate system near the target boundary point.
\begin{restatable}{prop}{keylemma}
Let $\Omega$ be a domain in $\mathbb{C}^2$ with a smooth boundary. Fix a boundary point $p$ and an integer $r > 0$. Assume that the outward unit normal vector of $\partial\Omega$ at $p$ is $(1, 0)$. Then there is a neighborhood $U$ of $p$ and a continuous map $\Psi \colon (\partial\Omega \cap U) \times \mathbb{C}^2 \rightarrow \mathbb{C}^2$ which satisfies, for each $q \in \partial\Omega \cap U$, the following properties:
\begin{enumerate}
\item The map $\Psi_q := \Psi (q, \cdot)$ is the composition of a translation, a dilation and a triangular map.
\item $\Psi_q (q) = (0, 0)$.
\item The local defining function of $\Psi_q (\Omega \cap U)$ at $(0, 0)$ is represented by
\begin{equation} \nonumber
\left\{ (w, z) \mid \textrm{Re}\,w + P(z, \bar{z}) + R(z, \bar{z}) + \textrm{Im}\left( \frac{w}{c} \right) Q \left( \textrm{Im}\left( \frac{w}{c} \right), z, \bar{z} \right) < 0 \right\}
\end{equation}
where $P$ is a real-valued polynomial of degree $r$ with no harmonic terms, $c = c(q)$ is a constant satisfying $\textrm{Re}\,c \neq 0$ and $\lim_{q \to p} c(q) = 1$, $R$ and $Q$ are real-valued smooth functions with the conditions on the vanishing order $\nu(R(z, \bar{z})) > r$ and $\nu \left(Q \left( \textrm{Im}\left( \frac{w}{c} \right), z, \bar{z} \right) \right) \ge 1$.
\end{enumerate}
Moreover, if the point $q \in \partial\Omega$ is pseudoconvex of finite type $2k$ and $r = 2k$, then $P$ is a homogeneous polynomial of degree $2k$.
\end{restatable}
This coordinate system is a variation of that of \cite{c}. We shall present the proof of this in Appendix.
\bigskip
\subsection{Construction of the scaling sequence} Let $\Omega$ be a domain in the hypothesis of Theorem 2.2. Denote by $\widehat{p} := \lim_{j\rightarrow \infty} \phi_j(p)$, the orbit accumulation boundary point. Taking a unitary transformation if necessary, one may assume that the outward unit normal vector of $\partial\Omega$ at $\widehat{p}$ is $(1, 0)$. Set $r = 2k$, which is the same as the type at $\widehat{p}$, and apply Proposition 2.3 to $\widehat{p}$. Then there is a global coordinate map $\Psi_{\widehat{p}}^{\Omega} \in Aut(\mathbb{C}^2)$ with $\Psi_{\widehat{p}}^{\Omega}(\widehat{p}) = (0, 0)$ and a neighborhood $U$ of $\widehat{p}$ such that the local defining function of $\Psi_{\widehat{p}} (\Omega)$ at $(0, 0)$ is represented by:
$$
\Psi_{\widehat{p}}^{\Omega}(\Omega \cap U) = \{ (w, z) \in \Psi_{\widehat{p}}^{\Omega}(U) \mid \rho_{\widehat{p}} (w, z) < 0 \}
$$
where:
\begin{enumerate}
\item
$\rho_{\widehat{p}} (w, z) = \textrm{Re}\,w + P(z, \bar{z}) + R(z, \bar{z}) + \textrm{Im}\,w \cdot Q(\textrm{Im}\,w, z, \bar{z})$,
\item
$P$ is a real-valued homogeneous polynomial of degree $2k$ without harmonic terms,
\item $R$ and $Q$ are real-valued smooth functions with the conditions $\nu(R(z, \bar{z})) > 2k$ and $\nu(Q(\textrm{Im}\,w, z, \bar{z})) \ge 1$.
\end{enumerate}
Taking a subsequence if necessary, one may assume that $\{ \phi_j(p) \} \subset U$. Write $p_j := \Psi_{\widehat{p}}^{\Omega}(\phi_j(p))$. Then the sequence $\{p_j\}$ is in $\Psi_{\widehat{p}}^{\Omega}(U)$ and its limit is the origin $\mathbf{0} = (0, 0)$. For each $j$, let $q_j$ be an intersection point of the half-line $\{ p_j + (t, 0) \in \mathbb{C}^2 \mid t > 0 \}$ and $\partial(\Psi_{\widehat{p}}^{\Omega}(\Omega)) \cap \Psi_{\widehat{p}}^{\Omega}(U)$. Since $\partial\Omega$ is smooth, the point $q_j$ is uniquely determined for every sufficiently large $j$. Notice that the sequence $\{q_j\}$ also converges to the origin $\mathbf{0}$. Since the upper semicontinuity of D'Angelo type holds in this case, one can choose a subsequence $\{ q_{j_l} \}$ of $\{ q_j \}$ so that the type $\Delta(q_{j_l})$ is less than or equal to $2k$ for all $l$. For simplicity, denote this subsequence by $\{ q_j \}$ again.
Write $\Omega' := \Psi_{\widehat{p}}^{\Omega}(\Omega)$ and apply Proposition 2.3 again to the domain $\Omega'$, with ${\bf 0} \in \partial \Omega'$, at each boundary point $q_j$. Denote by $\Psi_j^{\Omega'} := \Psi_{q_j}^{\Omega'}$. Now one arrives at
\begin{equation} \nonumber
\Psi_j^{\Omega'} (\Psi_{\widehat{p}}^{\Omega}(\Omega \cap U)) = \{ \rho_j(w, z) < 0 \} \textrm{ in } \Psi_j^{\Omega'}(\Psi_{\widehat{p}}^{\Omega}(U))
\end{equation}
where:
\begin{enumerate}
\item
$\rho_j(w, z) = \textrm{Re}\,w + P_j(z, \bar{z}) + R_j(z, \bar{z}) + \textrm{Im}\left( \frac{w}{c_j} \right) Q_j \left( \textrm{Im}\left( \frac{w}{c_j} \right), z, \bar{z} \right)$,
\item
$P_j$ is a real-valued polynomial of degree $2k$ with no harmonic terms (not in general homogeneous),
\item
$R_j$ and $Q_j$ are real-valued smooth functions with the conditions $\nu(R_j) > 2k$ and $\nu(Q_j) \ge 1$,
\item
$c_j$ is a constant satisfying $\textrm{Re}\,c_j \neq 0$ and $\lim_{j \to \infty} c_j = 1$.
\end{enumerate}
Note that this local defining function is valid in $\Psi_j^{\Omega'}(\Psi_{\widehat{p}}^{\Omega}(U))$ since each $\Psi_j^{\Omega'}$ is a composition of a translation and a triangular map. Moreover, the continuity of the coordinate system $\Psi$ of Proposition 2.3 guarantees that $\Psi_j^{\Omega'}$ converges to the identity map uniformly on compact subsets of $\mathbb{C}^2$. Hence there is $J > 0$ and an open neighborhood $U'$ of the origin $\mathbf{0}$ so that
\begin{equation} \label{U'}
U' \subset \Psi_{\widehat{p}}^{\Omega}(U) \cap \Psi_j^{\Omega'}(\Psi_{\widehat{p}}^{\Omega}(U)) \textrm{ for all } j > J.
\end{equation}
Since $\rho_j$ converges to $\rho_{\widehat{p}}$ uniformly on compact subsets of $\mathbb{C}^2$, the convergences $P_j \rightarrow P$, $R_j \rightarrow R$ and $Q_j \rightarrow Q$ are also uniform on compact subsets of $\mathbb{C}^2$. We call this $\Psi_j^{\Omega'} \circ \Psi_{\widehat{p}}^{\Omega}$ as the ($j$-th) {\it centering map}.
\bigskip
\begin{figure}[!h]
\centering
\includegraphics[width=0.9\columnwidth]{./fig1.eps}
\end{figure}
\begin{center}
{\bf Fig. 1.} The j-th centering map $\Psi_j^{\Omega'} \circ \Psi_{\widehat{p}}^{\Omega}$.
\end{center}
\bigskip
Next, we construct the sequence $\{D_j\}$ of dilations. Write $P_j (z, \bar{z}) = \sum_{n = 2}^{2k} P_{j,n} (z, \bar{z})$ where $P_{j,n}$ is a homogeneous polynomial of degree $n$. Define $\epsilon_j := \| p_j - q_j \|$ and choose $\delta_j > 0$ satisfying:
\begin{equation} \nonumber
\max \left\{ \left\| \frac{1}{\epsilon_j} P_{jn} ( \delta_j z, \overline{\delta_j z}) \right\|_{\infty}, n = 2, 3, \cdots, 2k \right\} = 1
\end{equation}
Here, the norm $\| \cdot \|_{\infty}$ is the $l^{\infty}$ norm on the space of polynomials as a finite sequence of coefficients. Let $D_j : (w, z) \mapsto \left( \frac{w}{\epsilon_j}, \frac{z}{\delta_j} \right)$ and denote by $\Lambda_j := D_j \circ \Psi_j^{\Omega'} \circ \Psi_{\widehat{p}}^{\Omega}$. Then
$$
(w, z) \in \Lambda_j (\Omega) \cap D_j (U') \Leftrightarrow (w, z) \in D_j (U') \textrm{ and } \widetilde{\rho_j}(w, z) < 0,
$$
where
$$
\widetilde{\rho_j}(w, z) = \textrm{Re}\,w + \frac{1}{\epsilon_j}P_j(\delta_j z, \delta_j \bar{z}) + \frac{1}{\epsilon_j}R_j(\delta_j z, \delta_j \bar{z}) + \textrm{Im} \left( \frac{w}{c_j} \right) Q_j \left( \textrm{Im} \left( \frac{\epsilon_j w}{c_j} \right), \delta_j z, \delta_j \bar{z} \right).
$$
\begin{lemma}
The sequence $\{ \widetilde{\rho_j} \}$ has a subsequence that converges to $\textrm{Re}\,w + \widehat{P} (z, \bar{z})$ uniformly on compact subsets of $\mathbb{C}^2$ where $\widehat{P}$ is a nonzero subharmonic polynomial of degree less than or equal to $2k$ with no harmonic terms.
\end{lemma}
\begin{proof}
The construction of $\delta_j$ and the local uniform convergence $P_j \rightarrow P$ guarantees that the sequence $\{\frac{1}{\epsilon_j}P_j(\delta_j z, \delta_j \bar{z})\}$ has a subsequence that converges to such a polynomial $\widehat{P}$. Moreover, the inequality $\epsilon_j \gtrsim \delta_j^{2k}$ holds since $P$ is homogeneous of degree $2k$. Then the term $\frac{1}{\epsilon_j} R_j(\delta_j z, \delta_j \bar{z})$ converges to the zero function owing to the conditions on $R_j$. The term $\textrm{Im} \left( \frac{w}{c_j} \right) Q_j \left( \textrm{Im} \left( \frac{\epsilon_j w}{c_j} \right), \delta_j z, \delta_j \bar{z} \right)$ converges to the zero function by a similar reason.
\end{proof}
In order to avoid using excessive indices, we shall keep $\{ \widetilde{\rho_j} \}$ for the subsequence and denote by $\widehat{\rho} := \lim \widetilde{\rho}_j$. So,
$$
\widehat{\rho} (w, z) := \textrm{Re}\,w + \widehat{P} (z, \bar{z}).
$$
Define $\widehat{\Omega} := \{ (w, z) \in \mathbb{C}^2 \mid \widehat{\rho} (w, z) < 0 \}$. Now we introduce the normal set-convergence to control the convergence of a sequence of holomorphic functions with domains varying. The following are a modification from Section 9.2.2 of \cite{gkk}.
\begin{defn} \rm
Let $\Omega_j$ be domains in $\mathbb{C}^n$ for each $j = 1, 2, \cdots$. The sequence $\Omega_j$ is said to {\it converge normally} to a domain $\widehat{\Omega}$ if the following two conditions hold.
\begin{enumerate}
\item For any compact set $K$ contained in the interior of $\bigcap_{j>m} \Omega_j$ for some positive integer $m$, $K \subset \widehat{\Omega}$.
\item For any compact subset $K'$ of $\widehat{\Omega}$, there exists a constant $m > 0$ such that $K' \subset \bigcap_{j>m} \Omega_j$.
\end{enumerate}
\end{defn}
\begin{prop}
If $\Omega_j$ is a sequence of domains in $\mathbb{C}^n$ that converges normally to the domain $\widehat{\Omega}$, then
\begin{enumerate}
\item If a sequence of holomorphic mappings $f_j : \Omega_j \rightarrow \Omega '$ from $\Omega_j$ to another domain $\Omega '$ converges uniformly on compact subsets of $\widehat{\Omega}$, then its limit is a holomorphic mapping from $\widehat{\Omega}$ into the closure of the domain $\Omega '$.
\item If a sequence of holomorphic mappings $g_j : \Omega ' \rightarrow \Omega_j$ converges uniformly on compact subsets of $\Omega '$, if $\widehat{\Omega}$ is pseudoconvex, and if there are a point $p \in \Omega '$ and a constant $c > 0$ so that the inequality $|\det{(dg_j|_p)}| > c$ holds for each $j$, then $\lim_{j \to \infty}g_j$ is a holomorphic mapping from the domain $\Omega '$ into $\widehat{\Omega}$.
\end{enumerate}
\end{prop}
Recall that $U'$ satisfying the condition (\ref{U'}), and $\Lambda_j := D_j \circ \Psi_j^{\Omega'} \circ \Psi_{\widehat{p}}^{\Omega}$. In this sense, $D_j (U')$ converges normally to $\mathbb{C}^2$. Hence the convergence $\widetilde{\rho_j} \rightarrow \widehat{\rho}$ guarantees that a sequence of domains $\{ \Lambda_j(\Omega) \}$ converges normally to $\widehat{\Omega}$. We now have constructed the sequence $\{ \Lambda_j \circ \phi_j : \Omega \rightarrow \Lambda_j(\Omega) \}$ for the proof of Theorem 2.2. Denote by $\sigma_j := \Lambda_j \circ \phi_j$.
\bigskip
\subsection{Convergence of the Pinchuk scaling sequence $\{ \sigma_j \}$} Since the limit domain $\widehat{\Omega}$ is unbounded, the convergence of $\{ \sigma_j \}$ does not follow immediately. So Pinchuk takes first the inverse sequence $\{ \sigma_j^{-1} : \sigma_j(\Omega) \rightarrow \Omega \}$. Recall that $\sigma_j(\Omega) = \Lambda_j(\Omega)$ and that the sequence $\{\sigma_j(\Omega)\}$ converges normally to $\widehat{\Omega}$. So Proposition 2.6 and Montel's theorem guarantee that there is a subsequence converging to a holomorphic map $\widehat{\tau} : \widehat{\Omega} \rightarrow \overline{\Omega}$ uniformly on compact subsets of $\widehat{\Omega}$. We shall keep the notation $\{ \sigma_j^{-1} \}$ for this subsequence. Actually, the image of $\widehat{\tau}$ is contained in $\Omega$. Indeed, suppose that there is a point in $\widehat{\Omega}$ whose image by $\widehat{\tau}$ is in $\partial \Omega$. Then $\widehat{\tau}(\widehat{\Omega}) \subset \partial\Omega$ by the pseudoconvexity of $\Omega$. This is impossible because $\widehat{\tau}(-1, 0) = p \in \Omega$. Consequently $\widehat{\tau}(\widehat{\Omega}) \subset \Omega$.
\begin{lemma}
There is a point $z_0 \in \widehat{\Omega}$ such that $d \widehat{\tau} |_{z_0}$ is nonsingular.
\end{lemma}
The proof follows the arguments of Lemma 2 in \cite{bp2}. We shall treat this in the last part of this section.
\medskip
Assume Lemma 2.7. Then the convergence of the inverse sequence $\{ \sigma_j^{-1} \}$ guarantees the uniform convergence of the sequence $\{ \det({d\sigma_j^{-1} |_{\mathbf{z}}}) \}$ on compact subsets of $\widehat{\Omega}$ by Cauchy estimates. Actually, it converges to $\det(d\widehat{\tau}|_{\mathbf{z}})$ uniformly on compact subsets of $\widehat{\Omega}$. Notice that $\det({d\sigma_j^{-1}) |_{\mathbf{z}}}$ is nowhere vanishing for any $j$ since each $\sigma_j^{-1}$ is a biholomorphic map. Hurwitz's theorem implies that $\det(d\widehat{\tau}|_{\mathbf{z}})$ is nowhere vanishing and hence $d\widehat{\tau}|_{\mathbf{z}}$ is nonsingular for all $\mathbf{z} \in \widehat{\Omega}$. In particular, $\widehat{\tau}$ is an immersion.
Suppose that $\widehat{\tau}$ is not 1-1. Then there are distinct points $s,s' \in \widehat{\Omega}$ satisfying $\widehat{\tau} (s) = \widehat{\tau} (s')$. Choose a neighborhood $U \Subset \widehat{\tau}(\widehat{\Omega})$ of $\widehat{\tau} (s)$ so that $\widehat{\tau}^{-1} (U)$ is disconnected. Let $V_{s}$ and $V_{s'}$ be mutually disjoint, connected components of $\widehat{\tau}^{-1} (U)$ such that $s \in V_{s}$ and $s' \in V_{s'}$. Note that $U$ can be adjusted so that $V_{s}$ and $V_{s'}$ are relatively compact in $\widehat{\Omega}$. Consequently, $\{ \sigma_j^{-1} |_{V_{s}} \}$ and $\{ \sigma_j^{-1} |_{V_{s'}} \}$ converge uniformly to $\widehat{\tau} |_{V_{s}}$ and $\widehat{\tau} |_{V_{s'}}$ respectively. Note that $\widehat{\tau}(V_{s}) = \widehat{\tau}(V_{s'}) = U$, which implies that $\sigma_j^{-1} (V_{s}) \cap \sigma_j^{-1} (V_{s'}) \neq \emptyset$ for sufficiently large $j$. This contradicts the injectivity of $\sigma_j^{-1}$. This implies
\begin{prop}
$\widehat{\tau} : \widehat{\Omega} \rightarrow \Omega$ is 1-1.
\end{prop}
\medskip
Now we are concerned with the surjectivity of $\widehat{\tau}$.
\begin{lemma}
The sequence $\{ \sigma_j \}$ converges to $\widehat{\tau}^{-1}$ uniformly on compact subsets of $\widehat{\tau}(\widehat{\Omega})$.
\end{lemma}
\begin{proof}
Fix a compact set $K$ in $\widehat{\tau}(\widehat{\Omega})$ and take a compact set $K'$ such that $K \Subset K' \Subset \widehat{\tau}(\widehat{\Omega})$. Since $\widehat{\tau}^{-1}(K')$ is also compact in $\widehat{\Omega}$, $\sigma_j(\Omega)$ contains $\widehat{\tau}^{-1}(K')$ for all sufficiently large $j$. Hence the uniform convergence of the inverse scaling sequence $\{ \sigma_j^{-1} \}$ on $\widehat{\tau}^{-1}(K')$ is well-defined. So the sequence of differentials $\{ d\sigma_j^{-1} \}$ converges to $d\widehat{\tau}$ uniformly on $\widehat{\tau}^{-1}(K')$. Moreover there is $c > 1$ and $J > 0$ such that $\frac{1}{c} < |\det{(d\sigma_j^{-1})}| < c$ on $\widehat{\tau}^{-1}(K')$ for all $j > J$, since the limit map $\widehat{\tau}$ is 1-1.
Crammer's rule in Linear algebra says that
$$
(d\sigma_j^{-1})^{-1} = \frac{1}{\det({d\sigma_j^{-1}})}C_j^T
$$
where $C_j^T$ is the transpose of the cofactor matrix of $d\sigma_j^{-1}$.
Since $|\det{(d\sigma_j^{-1})}|$ has a uniformly positive lower bound on $\widehat{\tau}^{-1}(K')$, each entry of the sequence of matrices $\{ (d\sigma_j^{-1})^{-1} \}$ is uniformly bounded on $\widehat{\tau}^{-1}(K')$. Notice that $\sigma_j^{-1} \circ \widehat{\tau}^{-1}$ converges to identity uniformly on $K'$. Hence one can choose $J' > 0$ so that $K \Subset \sigma_j^{-1}(\widehat{\tau}^{-1}(K'))$ if $j > J'$. Consequently the sequence $\{ d\sigma_j \mid j > J' \}$ is uniformly bounded on $K$ by the inverse function theorem. Recall that $\sigma_j (p) = (-1, 0)$ for all $j$. Hence Montel's theorem implies that the sequence $\{ \sigma_j \}$ has a subsequence that converges to some holomorphic function $g$ uniformly on compact subsets of the interior of $K$. we shall keep the notation $\{ \sigma_j \}$ also for this convergent subsequence. Then the sequence $\{ \sigma_j^{-1} \circ \sigma_j \}$ of the identity maps converges to $\widehat{\tau} \circ g$ on $K$. Hence $g = \widehat{\tau}^{-1}$ on $K$. Since $K$ is arbitrarily chosen, $g = \widehat{\tau}^{-1}$ on all of $\widehat{\tau}(\widehat{\Omega})$ and this proves the lemma.
\end{proof}
\medskip
\begin{prop}
$\widehat{\tau} : \widehat{\Omega} \rightarrow \Omega$ is onto.
\end{prop}
\begin{proof}
Suppose that $\widehat{\tau}(\widehat{\Omega}) \subsetneq \Omega$. Choose a point $p' \in \partial (\widehat{\tau}(\widehat{\Omega})) \cap \Omega$. Since $\widehat{p} = \lim_{j \to \infty} \phi_j(p)$ is a peak point of $\Omega$, the sequence $\{ \phi_j(p') \}$ also converges to $\widehat{p}$. Now we construct the Pinchuk scaling sequence $\{ \sigma'_j := \Lambda'_j \circ \phi_j \}$ with respect to $\{ \phi_j(p') \}$ as in Section 2.2. We already observed, in Proposition 2.8, that the inverse scaling sequence $\{ {\sigma'}_j^{-1} \}$ has a subsequence converging to a 1-1 holomorphic map $\widehat{\tau'} : \widehat{\Omega}' \rightarrow \Omega$ uniformly on compact subsets of $\widehat{\Omega}'$, where $\widehat{\Omega}'$ is the limit domain of the sequence $\{ \sigma'_j (\Omega) \}$. Taking a subsequence if necessary, we may assume that the uniform convergence holds for $\sigma_j^{-1} \rightarrow \widehat{\tau}$ and ${\sigma'}_j^{-1} \rightarrow \widehat{\tau'}$ on compact subsets of $\widehat{\Omega}$ and $\widehat{\Omega}'$ respectively.
Denote by $W := \widehat{\tau}(\widehat{\Omega}) \cap \widehat{\tau'}(\widehat{\Omega}')$ and set $B_j := \sigma_j \circ {\sigma'}_j^{-1} : {\sigma'}_j(\Omega) \rightarrow \sigma_j(\Omega)$. Notice that the maps $B_j$ and $B_j^{-1}$ are polynomial automorphisms (in fact, essentially triangular) of $\mathbb{C}^2$ with degree less than or equal to $2k$. Fix a nonempty open set $W' \Subset W$. By Lemma 2.9, $\sigma_j|_{W'}$ converges uniformly to $\widehat{\tau}^{-1}|_{W'}$ and hence $\lim_{j \to \infty} B_j|_{\widehat{\tau'}^{-1}(W')} \equiv \widehat{\tau}^{-1} \circ \widehat{\tau'}|_{\widehat{\tau'}^{-1}(W')}$, a biholomorphism of ${\widehat{\tau'}^{-1}(W')}$ and $\widehat{\tau}(W')$. Since each $B_j$ is a polynomial automorphism of $\mathbb{C}^2$ and of degree less than or equal to $2k$, it converges uniformly on compact subsets of $\mathbb{C}^2$ to a polynomial map $\widehat{B}$ of degree less than or equal to $2k$. Actually the limit map $\widehat{B}$ is in $Aut(\mathbb{C}^2)$ by a similar argument as in the proof of an injectivity of $\widehat{\tau}$. Similarly, $B_j^{-1}$ converges to $\widehat{B}^{-1}$. Now Proposition 2.6 guarantees that $\widehat{B}|_{\widehat{\Omega}'} : \widehat{\Omega}' \rightarrow \widehat{\Omega}$ and $\widehat{B}^{-1}|_{\widehat{\Omega}} : \widehat{\Omega} \rightarrow \widehat{\Omega}'$ are inverse to each other and hence $\widehat{\Omega}$ and $\widehat{\Omega}'$ are biholomorphic. Notice that $\widehat{B}|_{\widehat{\Omega}'}(-1, 0) \in \widehat{\Omega}$ and $\widehat{\tau} \circ \widehat{B}|_{\widehat{\Omega}'}(-1, 0) = p'$. Therefore $p' \in \widehat{\tau}(\widehat{\Omega})$. This contradicts that $p' \in \partial (\widehat{\tau}(\widehat{\Omega})) \cap \Omega$.
\end{proof}
\medskip
\begin{rmk} \rm
The Pinchuk scaling sequence also depends on the initial point. But the argument above shows that their limit domains have to be equivalent, via a polynomial automorphism of $\mathbb{C}^2$ with its degree not more than $2k$, the type of the orbit accumulating boundary point.
\end{rmk}
\medskip
Propositions 2.8 and 2.10 imply that the limit map $\widehat{\tau} : \widehat{\Omega} \rightarrow \Omega$ of the inverse scaling sequence is a biholomorphic map. Consequently, the limit map $\widehat{\sigma}$ of the Pinchuk scaling sequence is defined on all of $\Omega$ and it satisfies $\widehat{\sigma} \equiv \widehat{\tau}^{-1}$. So the only remaining part for the proof of Theorem 2.2 is justifying Lemma 2.7.
\bigskip
{\bf Proof of Lemma 2.7.} The main idea is an estimate of the invariant metric introduced by Sibony \cite{s}. It is defined by
\begin{equation} \nonumber
F_S(p, \xi ; M) := \sup_u \left\{ \left( \sum_{i, j = 1}^{n} \frac{\partial^2 u}{\partial z_i \partial \bar{z}_j} \bigg|_p \xi_i \bar{\xi}_j \right)^{\frac{1}{2}} \colon u \in A_{M} (p) \right\},
\end{equation}
where $A_{M} (p)$ is the set of all plurisubharmonic functions of $M$ defined as follows: $u \in A_{M} (p)$ if $0 \le u \le 1$, $u(p) = 0$, $u \in C^2$ near $p$, and $\log u$ is plurisubharmonic on $M$.
\medskip
For this metric, Sibony proved;
\begin{thm}[\cite{s}]
Let $M$ be a complex manifold. If there is a bounded plurisubharmonic function $u$ of $M$ and there is a constant $\delta > 0$ such that $dd^c u$ is positive definite on a $\delta$-neighborhood of $p$, then there is an $\epsilon = \epsilon(\delta) > 0$ such that $F_S(p, \xi ; M) \ge \epsilon |\xi|$ for all $\xi$ in the holomorphic tangent space $T_p^{\mathbb{C}} M$.
\end{thm}
Recall that $\Psi_{\widehat{p}}^{\Omega}(\Omega)$ is a bounded domain in $\mathbb{C}^2$ and its local defining function at $\mathbf{0}$ is $\rho_{\widehat{p}}$. Hence the main theorem of \cite{df} by Diederich and Forn{\ae}ss says that there is a $C^{\infty}$ defining function $\beta$ and a sufficiently small $\eta > 0$ such that $-(-\beta)^{\eta}$ is a strictly plurisubharmonic bounded exhaustion function of $\Psi_{\widehat{p}}^{\Omega}(\Omega)$. Since both $\beta$ and $\rho_{\widehat{p}}$ are defining functions of $\Psi_{\widehat{p}}^{\Omega}(\Omega)$, $\lim_{{\bf z} \rightarrow {\bf 0}} \left(\frac{\beta({\bf z})}{\rho_{\widehat{p}} ({\bf z})}\right)^{\eta} = c$ for some positive constant $c$. Taking a constant multiple of $\beta$, one may assume that $c = 1$. Define $\widetilde{\beta}_j := -\epsilon_j^{-\eta}(-\beta \circ (\Psi_j^{\Omega'})^{-1} \circ D_j^{-1})^{\eta}$. Since the equality $-\widetilde{\rho_j} = \epsilon_j^{-1} (-\rho_{\widehat{p}} \circ (\Psi_j^{\Omega'})^{-1} \circ D_j^{-1})$ holds due to the construction of $\widetilde{\rho_j}$ and $\rho_{\widehat{p}}$, we have the following formula:
\begin{equation} \nonumber
\widetilde{\beta}_j = -\left(\frac{\beta}{\rho_{\widehat{p}}} \circ (\Psi_j^{\Omega'})^{-1} \circ D_j^{-1} \right)^{\eta} \cdot (-\widetilde{\rho_j})^{\eta}.
\end{equation}
Recall that $\widetilde{\rho_j} \rightarrow \widehat{\rho}$ uniformly on compact sets of $\mathbb{C}^2$. Since $(\Psi_j^{\Omega'})^{-1} \circ D_j^{-1}$ converges to ${\bf 0}$ uniformly on compact sets of $\mathbb{C}^2$, the convergence $\widetilde{\beta}_j \rightarrow -(-\widehat{\rho})^{\eta}$ is also uniform on compact sets of $\mathbb{C}^2$. Note that the plurisubharmonicity of each $\widetilde{\beta}_j$ guarantees the plurisubharmonicity of the limit map $-(-\widehat{\rho})^{\eta}$. Recall that $\widehat{\rho} (w, z) = \textrm{Re}\,w + \widehat{P} (z, \bar{z})$ for a certain nonzero subharmonic polynomial $\widehat{P}$ without harmonic terms. Hence $e^{-(-\widehat{\rho})^{\eta}}$ is strictly plurisubharmonic almost everywhere on $\widehat{\Omega}$. In particular, there is a point $q_0 \in \widehat{\Omega}$ such that $e^{-(-\widehat{\rho})^{\eta}}$ is strictly plurisubharmonic at $q_0$. So $e^{\widetilde{\beta}_j}$ is uniformly strictly plurisubharmonic at $q_0$ for all sufficiently large $j$. Now Theorem 2.11 guarantees that there is an $\epsilon > 0$ such that $F_S(q_0, \xi; \Lambda_j(\Omega)) \ge \epsilon |\xi|$ holds whenever $\xi \in \mathbb{C}^2$ and $j$ is sufficiently large. Note that this metric is invariant under biholomorphic transformations, and hence
\begin{equation} \nonumber
F_S(\tau_j(q_0), d \tau_j |_{q_0} (\xi); \Omega) \ge \epsilon |\xi|
\end{equation}
for all $\xi \in \mathbb{C}^2$ and $j$ sufficiently large.
\medskip
Let $q' \in \Omega$ be the limit point of $\{ \tau_j(q_0) \}$. Then the uniform convergence $d \tau_j |_{q_0} \rightarrow d \widehat{\tau} |_{q_0}$ on compact subsets of $\widehat{\Omega}$ implies that $F_S(q', d \widehat{\tau} |_{q_0} (\xi); \Omega) \ge \epsilon |\xi|$ for all $\xi \in \mathbb{C}^2$. Consequently, $d \widehat{\tau}$ is nonsingular at $q_0$ and this proves Lemma 2.7.
\medskip
\section{The Frankel scaling sequence}
\begin{defn}[\cite{f}] \rm
Let $\Omega$ be a domain in $\mathbb{C}^n$ with a point $p \in \Omega$, and $\{ \phi_j \}$ be a sequence in $Aut(\Omega)$. Then the {\it Frankel scaling sequence} $\{ \omega_j : \Omega \rightarrow \mathbb{C}^n \}$ with respect to $(\Omega, p, \{\phi_j\})$ is defined by
\begin{equation}
\omega_j (\mathbf{z}) := [d\phi_j |_p]^{-1} (\phi_j(\mathbf{z}) - \phi_j (p))
\end{equation} \nonumber
where $\mathbf{z} = (z_1, \cdots, z_n) \in \mathbb{C}^n$.
\end{defn}
Notice that $\omega_j(p) = 0$ and $d\omega_j|_p = I$ for all $j$, where $I$ is the identity map. So its construction appears to be more intrinsic than Pinchuk's. However, the convergence is known only for the convex Kobayashi hyperbolic domains \cite{f}. In fact, we show here that there is a non-convex domain for which the Frankel scaling sequence diverges, even though the domain is biholomorphic to a bounded convex domain.
\begin{ex} \rm
Let $\Omega_1 = \{ (w, z) \in \mathbb{C}^2 \mid \textrm{Re}\,w + |z|^4 < 0 \}$. Note that $\Omega_1$ is biholomorphic to the Thullen domain $\{ |w|^2 + |z|^4 < 1 \}$, which is bounded and convex. Fix a point $p = (-1, 0) \in \Omega_1$. Consider a sequence $\left\{ \phi_j (w, z) := \left( \frac{1}{j^4}w, \frac{1}{j}z \right) \right\}$ of automorphisms of $\Omega_1$. Then the Frankel scaling map $\omega_j^{\Omega_1}$ with respect to $(p, \phi_j)$ is
\begin{equation} \nonumber
\omega_j^{\Omega_1} (w, z) = [d\phi_j |_p]^{-1} (\phi_j(w, z) - \phi_j (p)) = (w+1, z).
\end{equation}
Consequently, the sequence $\{ \omega_j^{\Omega_1} (w, z) \}$ converges to $(w+1, z)$.
\medskip
On the other hand, let $\psi$ be defined by $\psi(w, z) = (w - 2z^2, z)$ and $\Omega_2 := \psi(\Omega_1)$. Notice that $\Omega_2 = \{ (w, z) \in \mathbb{C}^2 \mid \textrm{Re}\,w + z^2 + \overline{z}^2 + |z|^4 < 0 \}$ and $\psi(p) = p$. Denote by $\widetilde{\phi_j} := \psi \circ \phi_j \circ \psi^{-1}$ which is in $Aut(\Omega_2)$. Then the Frankel scaling map $\omega_j^{\Omega_2}$ with respect to $(p, \widetilde{\phi_j})$ is
\begin{equation} \nonumber
\omega_j^{\Omega_2} (w, z) = [d\widetilde{\phi_j} |_{\psi(p)}]^{-1} (\widetilde{\phi_j}(w, z) - \widetilde{\phi_j}(p))
= \left(\begin{array}{cc} w + 2(1 - j^2)z^2 + 1 \\ z \end{array} \right).
\end{equation}
Observe that every subsequence of $\{ \omega_j^{\Omega_2} (w, z) \}$ diverges.
\end{ex}
\begin{rmk} \rm
On the other hand, Frankel's scaling sequence can converge on a certain nonconvex domain. In such a case, the convergence is preserved through $\mathbb{C}$-affine biholomorphic transformations.
Let $\Omega$, $p \in \Omega$ and $\{\phi_j\} \subset Aut(\Omega)$ be given as in Definition 3.1. Let $\psi$ be a nonsingular $\mathbb{C}$-affine map and denote by $\widetilde{\phi_j} := \psi \circ \phi_j \circ \psi^{-1}$. Then the Frankel scaling map with respect to $(\psi(\Omega), \psi(p), \widetilde{\phi_j})$ is
\begin{equation} \nonumber
[d \widetilde{\phi_j} |_{\psi (p)}]^{-1} (\widetilde{\phi_j}(z) - \widetilde{\phi_j} (\psi (p))) = d \psi_p [d \phi_j |_p]^{-1} (\phi_j (\psi^{-1}(z)) - \phi_j (p)).
\end{equation}
Notice that the right hand side is the composition of a nonsingular matrix $d \psi_p$ and the Frankel scaling map with respect to $(\Omega, p, \phi_j)$. So the convergence is invariant under the nonsingular affine transformations.
\end{rmk}
In Example 3.2, observe that $\psi^{-1}$ removes the harmonic term $z^2 + \bar{z}^2$, whose vanishing order is smaller than the principle term $|z|^4$, in the expression of the defining function of $\Omega_2$. So one would hope that taking a coordinate change map $\Psi \in Aut(\mathbb{C}^2)$, that removes harmonic terms from the defining function of given domain at the limit point of the sequence $\{\phi_j(p)\}$, may be enough to make the (adjusted) Frankel scaling sequence converge. However, the following example shows that this is not true in general.
\begin{ex} \rm
Let $\Omega_3 = \{ (w, z) \in \mathbb{C}^2 \mid \textrm{Re}\,w + 4z\overline{z}^3 + 6|z|^4 + 4z^3 \overline{z} < 0 \}$ and fix the point $p = (1, i) \in \Omega_3$. We consider the map
\begin{equation} \nonumber
\begin{split}
&\phi_{\mu}(w, z)
\\
=& \left(\frac{1}{\mu^8}w + \frac{8i(\mu-1)}{\mu^8}z^3 - \frac{12(\mu-1)^2}{\mu^8}z^2 - \frac{8i(\mu-1)^3}{\mu^8}z + \frac{2(\mu-1)^4}{\mu^8}, \frac{1}{\mu^2}z + \frac{\mu i - i}{\mu^2}\right).
\end{split}
\end{equation}
Then $\{\phi_{\mu}\}_{\mu} \subset Aut(\Omega_3)$. Notice that $\phi_{\mu}(p)$ converges to the origin $\mathbf{0}$ as $\mu$ goes to infinity. So we investigate whether no adjustments are needed since the defining function of $\Omega_3$ has no harmonic terms in $z$. Observe that the Frankel scaling map $\omega_{\mu}^{\Omega_3}$ with respect to $(p, \phi_{\mu})$ is:
$$
\omega_{\mu}^{\Omega_3}(w, z) = (w + 8i(\mu-1)z^3 - 12(\mu-1)^2z^2 + 24i\mu(\mu-1)z + 12\mu^2 - 8\mu - 5, z - i).
$$
It is now clear that every subsequence of $\{ \omega_{\mu}^{\Omega_3} \}$ diverges.
\end{ex}
In the light of these examples, the following question arises naturally: {\it for a given $(\Omega, p, \phi_j)$ as in Definition 3.1, would there exist a map $\Psi \in Aut(\mathbb{C}^2)$ such that the Frankel scaling sequence with respect to $(\Psi(\Omega), \Psi(p), \Psi \circ \phi_j \circ \Psi^{-1})$ converges?} In Example 3.2, the map $\psi^{-1}$ performs the role for $(\Omega_2, p, \widetilde{\phi_j})$, while finding such a map is not easy in Example 3.3. We leave this for a later study. However, the following adjustment can be a reasonable alternative.
\medskip
\begin{defn} \rm
Let $\Omega$, $p$ and $\{\phi_j\}$ be as in Definition 3.1. Consider the sequence $\{ \psi_j \} \subset Aut(\mathbb{C}^n)$ converging to another automorphism $\widehat{\psi}$ uniformly on compact sets of $\mathbb{C}^n$. Now we define the {\it modification} (i.e., {\it modified Frankel scaling sequence})
\begin{equation} \nonumber
\omega_j (\mathbf{z}) := [d(\psi_j \circ \phi_j \circ \psi_j^{-1}) |_{\psi_j(p)}]^{-1} (\psi_j \circ \phi_j \circ \psi_j^{-1}(\mathbf{z}) - \psi_j \circ \phi_j (p))
\end{equation}
by $\{\psi_j\}$ of the original Frankel scaling sequence.
\end{defn}
\begin{thm}
Let $\Omega$ be a bounded domain in $\mathbb{C}^n$ which admits a boundary accumulation automorphism orbit $\{\phi_j(p)\}$. Assume that there is a sequence $\{ \psi_j \} \subset Aut(\mathbb{C}^n)$ satisfying:
\begin{enumerate}
\item $\{\psi_j\}$ converges to $\widehat{\psi} \in Aut(\mathbb{C}^n)$ uniformly on compact sets of $\mathbb{C}^n$.
\item There is a sequence $\{ D_j\}$ of $\mathbb{C}$-affine maps so that the sequence $\{ D_j \circ \psi_j \circ \phi_j \}$ converges to a certain biholomorphism-into $\mathbb{C}^n$ uniformly on compact subsets of $\Omega$.
\end{enumerate}
Then the modified Frankel scaling sequence by $\{\psi_j\}$ has a subsequence that converges to a biholomorphism-into $\mathbb{C}^n$ uniformly on compact subsets of $\widehat{\psi}(\Omega)$.
\end{thm}
\begin{proof}
Let $\sigma_j := D_j \circ \psi_j \circ \phi_j$ and $\widehat{\sigma} := \lim_{j \to \infty} \sigma_j$. Let $\omega_j$ represent the modified Frankel scaling map by $\psi_j$. Write $\widetilde{\phi_j} := \psi_j \circ \phi_j \circ \psi_j^{-1}$, then
\begin{equation} \nonumber
\omega_j (\mathbf{z}) := [d\widetilde{\phi_j} |_{\psi_j(p)}]^{-1} (\widetilde{\phi_j}(\mathbf{z}) - \psi_j \circ \phi_j(p)).
\end{equation}
Define the $\mathbb{C}$-affine map $A_j$ of $\mathbb{C}^n$ by
\begin{equation} \nonumber
A_j (\mathbf{z}) := [d\widetilde{\phi_j} |_{\psi_j(p)}]^{-1} (D_j^{-1} (\mathbf{z}) - \psi_j \circ \phi_j(p)).
\end{equation}
Now, $A_j$ makes the following diagram commute.
\bigskip
\begin{figure}[!h]
\centering
\includegraphics[width=0.6\columnwidth]{./fig2.eps}
\end{figure}
\begin{center}
{\bf Fig. 2.} A relationship of j-th scaling maps.
\end{center}
\bigskip
Notice that the map $A_j$ enjoys the following properties:
\begin{itemize}
\item $A_j |_{\sigma_j(\Omega)} = \omega_j \circ \psi_j \circ \sigma_j^{-1}$.
\item $A_j(\sigma_j(p)) = \mathbf{0}$, for all $j$.
\item $dA_j$ converges to a nonsingular matrix.
\end{itemize}
The first two properties follow directly from the construction of $A_j$, and the third is nothing but
\begin{equation} \nonumber
\lim_{j \rightarrow \infty} dA_j|_{\sigma_j(p)} = \lim_{j \rightarrow \infty} d\omega_j |_{\psi_j(p)} \circ d\psi_j |_p \circ d\sigma_j^{-1} |_{\sigma_j(p)} = d\widehat{\psi}|_p \circ d\widehat{\sigma}^{-1} |_{\widehat{\sigma}(p)}.
\end{equation}
Hence the sequence $\{A_j\}$ converges to some nonsingular $\mathbb{C}$-affine map, say $\widehat{A}$, satisfying $\widehat{A}(\widehat{\sigma}(p)) = \mathbf{0}$. Notice that $\omega_j = A_j \circ \sigma_j \circ \psi_j^{-1}$ and $\psi_j(\Omega)$ converges to $\widehat{\psi}(\Omega)$ in the sense of normal set convergence. Hence the uniform convergence of sequences $\{A_j\}$, $\{\sigma_j\}$ and $\{\psi_j^{-1}\}$ implies that the sequence $\{\omega_j\}$ converges to $\widehat{A} \circ \widehat{\sigma} \circ \widehat{\psi}^{-1}$ uniformly on compact subsets of $\widehat{\psi}(\Omega)$. This proves the theorem.
\end{proof}
\begin{rmk} \rm
If $\Omega$ is convex, then $\psi_j$ satisfying the hypotheses (1) and (2) automatically exist, affine maps, as demonstrated in \cite{kk}. It is clear now that Theorem 3.6, with Remark 3.3, generalizes the convergence theorem of Frankel for convex domains.
\end{rmk}
\medskip
Recall that the sequence of centering maps $\{\Psi_j^{\Omega'} \circ \Psi_{\widehat{p}}^{\Omega}\} \subset Aut(\mathbb{C}^2)$ constructed in Section 2.2. Combined with Theorem 2.2, Theorem 3.6 implies the following:
\begin{thm}
Let $\Omega$ be a bounded domain in $\mathbb{C}^2$ with smooth pseudoconvex boundary. Assume that $Aut(\Omega)$ admits a boundary accumulating automorphism orbit $\{\phi_j(p)\}$ converging to $\widehat{p} \in \partial\Omega$. If $\widehat{p}$ is of finite type in the sense of D'Angelo, then there exists a modified Frankel scaling sequence, by $\{\Psi_j^{\Omega'} \circ \Psi_{\widehat{p}}^{\Omega}\}$, that converges to a 1-1 holomorphic map into $\mathbb{C}^n$ uniformly on compact subsets of $\Psi_{\widehat{p}}^{\Omega}(\Omega)$.
\end{thm}
\bigskip
\section{Equivalence of two scalings}
Theorem 3.6 above can be restated as follows: If the Pinchuk scaling sequence built upon the sequence of the global centering maps $\{ \psi_j \}$ converges to $\widehat{\sigma}$, then the modified Frankel scaling sequence by $\{ \psi_j \}$ also converges to $\widehat{\omega}$, say.
Observe that $\widehat{\omega} = \widehat{A} \circ \widehat{\sigma} \circ \widehat{\psi}^{-1}$. Therefore we have
\begin{thm}
If the Pinchuk scaling sequence, built upon the sequence of the global centering maps $\{ \psi_j \}$, converges to $\widehat{\sigma}$. Then the modified Frankel scaling sequence by $\{ \psi_j \}$ converges to $\widehat{\omega}$, and
\begin{equation} \nonumber
\widehat{\omega} = \widehat{A} \circ \widehat{\sigma} \circ \widehat{\psi}^{-1}
\end{equation}
for some nonsingular $\mathbb{C}$-affine map $\widehat{A}$ and the limit map $\widehat{\psi}$ of $\{ \psi_j \}$.
\end{thm}
\begin{rmk} \rm
If the given domain is bounded and convex, Kim and Krantz \cite{kk} proved that Pinchuk's scaling sequence converges, that $\widehat{\psi}$ turns out to be a $\mathbb{C}$-affine biholomorphism, and consequently that the limits of two scalings are $\mathbb{C}$-affinely equivalent. In this regard, Theorem 4.1 generalizes their result to a nonconvex case.
\end{rmk}
\bigskip
\section{Appendix}
We now prove Proposition 2.3. For convenience, we restate the proposition.
\keylemma*
\begin{proof}
Define the translation map by $T_q(z) := z-q$. Since the outward unit normal vector of $\partial\Omega$ at $p$ is $(1, 0)$, the implicit function theorem guarantees that there is a neighborhood $U$ of $p$ such that
$$
T_p(\Omega \cap U) = \{ (w, z) \mid \textrm{Re}\,w + P_p (z, \bar{z}) + \textrm{Im}\,w \cdot Q_p (\textrm{Im}\,w, z, \bar{z}) < 0 \}
$$
for some real valued smooth function $P_p$ and $Q_p$ with the conditions $\nu(P_p) \ge 2$ and $\nu(Q_p) \ge 1$. Now, one can obtain
$$
T_q(\Omega \cap U) = \{ (w, z) \mid \textrm{Re}\,w + P_q (z, \bar{z}) + b_q \cdot \textrm{Im}\,w + \textrm{Im}\,w \cdot Q_q (\textrm{Im}\,w, z, \bar{z}) < 0 \}
$$
for some real constant $b_q$ and real valued smooth functions $P_q$ and $Q_q$ with the conditions $\nu(P_q) \ge 1$ and $\nu(Q_p) \ge 1$. Notice that all of $P_q$, $Q_q$ and $b_q$ vary continuously with respect to $q \in \partial \Omega \cap U$, and hence $\lim_{q \to p} P_q = P_p$, $\lim_{q \to p} Q_q = Q_p$ and $\lim_{q \to p} b_q = 0$. Now we take the coordinate change $S_q : (w, z) \mapsto ((1-ib_q)w, z)$ on $T_q(\Omega)$. Denote by $c_q := (1-ib_q)$. Then
$$
S_q \circ T_q(\Omega \cap U) = \left\{ (w, z) \,\Big|\, \textrm{Re}\,w + P_q (z, \bar{z}) + \textrm{Im} \left( \frac{w}{c_q} \right) Q_q \left(\textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) < 0 \right\}.
$$
Define $H_{q, 1} (z, \bar{z})$ by the harmonic part of $P_q (z, \bar{z})$ of degree $r$ as follows:
$$
H_{q, 1} (z, \bar{z}) := 2 \textrm{Re}\,h_{q, 1} (z) \textrm{ where } h_{q, 1}(z) := \sum_{j = 1}^{r} \frac{\partial^j P_q (z, \bar{z})}{\partial z^j} \bigg|_0 \cdot \frac{z^j}{j!}.
$$
Let $P_{q, 1} := P_q - H_{q, 1}$, then $P_{q, 1} (z, \bar{z})$ has no harmonic terms of degree less than or equal to $r$. Now consider the coordinate change $\psi_{q, 1} : (w, z) \mapsto (w + 2h_{q, 1} (z), z)$. Then,
\begin{equation} \nonumber
\begin{split}
\psi_{q, 1} &(S_q(T_q(\Omega \cap U)))
\\
=& \bigg\{ \textrm{Re}\,w + P_{q, 1} (z, \bar{z}) + \textrm{Im} \left( \frac{-2h_{q, 1} (z)}{c_q} \right) Q_q \left( \textrm{Im} \left( \frac{w - 2h_{q, 1} (z)}{c_q} \right), z, \bar{z} \right)
\\
&\quad + \textrm{Im} \left( \frac{w}{c_q} \right) Q_q \left( \textrm{Im}\left( \frac{w - 2h_{q, 1} (z)}{c_q} \right), z, \bar{z} \right) < 0 \bigg\}.
\end{split}
\end{equation}
Denote by
$$
R_{q, 1} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) := \textrm{Im} \left( \frac{ - 2h_{q, 1} (z)}{c_q} \right) Q_q \left( \textrm{Im} \left( \frac{ w - 2h_{q, 1} (z)}{c_q} \right), z, \bar{z} \right),
$$
and
\begin{equation} \nonumber
\begin{split}
Q_{q, 1} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) :=& \, Q_q \left( \textrm{Im}\left( \frac{w - 2h_{q, 1} (z)}{c_q} \right), z, \bar{z} \right)
\\
+& \left( \textrm{Im} \left( \frac{w}{c_q} \right) \right)^{-1} \left( R_{q, 1} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) - R_{q, 1} (0, z, \bar{z}) \right).
\end{split}
\end{equation}
Then $R_{q, 1}$ and $Q_{q, 1}$ vary continuously with respect to $q$ and the local defining function of $\psi_{q, 1} (S_q(T_q(\Omega \cap U)))$ above can be rewritten as follows:
\begin{equation} \nonumber
\begin{split}
&\psi_{q, 1} \circ S_q \circ T_q(\Omega \cap U)
\\
=& \bigg\{ (w, z) \,\Big|\, \textrm{Re}\,w + P_{q, 1} (z, \bar{z}) + R_{q, 1} (0, z, \bar{z}) + \textrm{Im} \left( \frac{w}{c_q} \right) Q_{q, 1} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) < 0 \bigg\}.
\end{split}
\end{equation}
Notice that $\nu(R_{q, 1} (0, z, \bar{z})) \ge \nu(h_{q, 1}(z)) + \nu(Q_q) > 1$. Now we construct the following functions, varying continuously with respect to $q$, inductively for $j = 1, 2, \cdots, r$.
\begin{itemize}
\item $R_{q, 0} := P_q$, $Q_{q, 0} := Q_q$.
\item $H_{q, j} (z, \bar{z}) := 2 \textrm{Re}\,h_{q, j} (z) \textrm{ where } h_{q, j}(z) := \sum_{k = j}^{r} \frac{\partial^k R_{q, j-1} (z, \bar{z})}{\partial z^k} \big|_0 \cdot \frac{z^k}{k!}$.
\item $P_{q, j} := R_{q, j-1} - H_{q, j}$.
\item $\psi_{q, j} : (w, z) \mapsto (w + 2h_{q, j} (z), z)$
\item $R_{q, j} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) := \textrm{Im} \left( \frac{- 2h_{q, j} (z)}{c_q} \right) Q_{q, j-1} \left( \textrm{Im} \left( \frac{ w-2h_{q, j} (z)}{c_q} \right), z, \bar{z} \right)$.
\item $Q_{q, j} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) := Q_{q, j-1} \left( \textrm{Im}\left( \frac{w - 2h_{q, j} (z)}{c_q} \right), z, \bar{z} \right)$ \newline \phantom{$Q_{q, j} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right)$} $+ \left( \textrm{Im} \left( \frac{w}{c_q} \right) \right)^{-1} \left( R_{q, j} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) - R_{q, j} (0, z, \bar{z}) \right)$.
\end{itemize}
Define $\psi_q := \psi_{1, r} \circ \cdots \circ \psi_{q, 1}$. Then $\psi_q$ varies continuously with respect to $q$ and satisfies:
\begin{equation} \nonumber
\begin{split}
\psi_q \circ S_q \circ T_q(\Omega \cap U) = \bigg\{ (w, z) \,\Big|\, &\textrm{Re}\,w + \Sigma_{j=1}^{r} P_{q, j} (z, \bar{z}) + R_{q, r}(0, z, \bar{z})
\\
&+ \textrm{Im} \left( \frac{w}{c_q} \right) Q_{q, r} \left( \textrm{Im} \left( \frac{w}{c_q} \right), z, \bar{z} \right) < 0 \bigg\}.
\end{split}
\end{equation}
Notice that $\Sigma_{j=1}^{r} P_{q, j} (z, \bar{z}) + R_{q, r}(0, z, \bar{z})$ has no harmonic terms of degree less than or equal to $r$, $Q_{q, r}$ has no constant term, and $\nu(R_{q, r}(0, z, \bar{z})) > r$. Set $\Psi_q := \psi_q \circ S_q \circ T_q$, $P := \Sigma_{j=1}^{r} P_{q, j}$, $R(z, \bar{z}) := R_{q, r}(0, z, \bar{z})$ and $Q := Q_{q, r}$. Then the first assertion of the proposition follows.
\smallskip
To prove the second, fix $q \in \partial\Omega \cap U$ and $r = 2k$, the type of $q$. Denote $\rho_q(w, z)$ by the local defining function of $\Psi_q (\Omega)$ near $q$ as in Proposition 2.3. Recall that $P$ has no harmonic terms. So it is sufficient to prove that $\nu(P) = 2k$.
Suppose $\nu(P) > 2k$. Then for the analytic disc $\delta_1 : z \mapsto (0, z)$, the vanishing order of a composition $\rho_q \circ \delta_1$ is larger than $2k$ and this contradicts that $\Delta(p) = 2k$. So $\nu(P) \le 2k$.
Now assume that $\nu(P) < 2k$. Then there is a type-realizing holomorphic disc $\delta_2 : z \mapsto (f(z), g(z))$, so that $\frac{\nu(\rho_q \circ \delta_2)}{\nu(\delta_2)} = 2k$. If $\nu(f) \le \nu(g)$ then $f$ is not identically zero, and hence
\begin{equation} \nonumber
\frac{\nu(\rho_q \circ \delta_2)}{\nu(\delta_2)} = \phantom{.} \frac{\min \{ \nu(\textrm{Re}\,f), \nu(P) \nu(g) \} }{\nu(f)} \le 1.
\end{equation}
The first equality holds since $P$ has no harmonic terms while $\textrm{Re}\,f$ is harmonic. But this is impossible because $\frac{\nu(\rho_q \circ \delta_2)}{\nu(\delta_2)} = 2k$. So $\nu(f) > \nu(g)$ and hence
\begin{equation} \nonumber
\frac{\nu(\rho_q \circ \delta_2)}{\nu(\delta_2)} = \frac{\min \{ \nu(\textrm{Re}\,f), \nu(P) \nu(g) \} }{\nu(g)} \le \nu(P) < 2k.
\end{equation}
The last inequality holds due to the assumption. However, this contradicts that $\frac{\nu(\rho_q \circ \delta_2)}{\nu(\delta_2)} = 2k$ and hence $\nu(P) = 2k$. Therefore the second assertion of the proposition follows.
\end{proof}
|
\subsection{DDP-vMF-means}
In the following, we analyze the small-variance asymptotics of the DDP-vMF
mixture model. We first derive the label assignment rules, followed by
the parameter updates.
\paragraph{Label Update:} First, let $\alpha = \exp(\lambda \tau)$, $q =
\exp(Q\tau)$, $\xi = \exp(\beta\tau)$, and $\tau_k = \tau w_k$, with
$\lambda\in\left[-2, 0\right]$ as before, $Q \leq 0$, and $\beta, w_k \geq 0$.
Note that $\lim_{\tau\to\infty}\frac{1-q^t}{1-q}= 1$, and thus the asymptotics
of the label assignment probability for current and new clusters is the same as
in Section \ref{sec:dpvmf}.
Hence, we focus on the assignment of a datapoint to a previously
observed, but currently not instantiated, cluster $k$. During the
$\dt{k}$ timesteps since cluster $k$ was last observed, the mean
direction $\mu_k$ underwent a random $\vMF$ walk $\mu_{k0}\to
\mu_{k1}\to \dots\to \mu_{k\dt{k}}= \mu_k$ with initial distribution
$\mu_{k0} \sim \vMF(\mu_{k0}; m_k, \tau_k)$. Therefore, the
intermediate mean directions $\{\mu_{kn}\}_{n=1}^{\dt{k}}$ must be
marginalized out when computing $p(x_i; m_k, \tau_k)$:
\begin{align}
\begin{aligned}
&p(x_i ; m_k,\tau_k) \\
&\begin{array}{l}
= \idotsint\limits_{\mu_{k0}, \dots, \mu_{k\dt{k}}}
\begin{matrix}
p(x_i |\mu_{k\dt{k}};\tau)\cdot p(\mu_0;m_k,\tau_k) \\
\cdot \prod_{n=1}^\dt{k} p(\mu_{kn} |\mu_{k(n-1)};\xi)
\end{matrix} \\
= Z(\tau) Z(\beta\tau)^\dt{k} Z(w_k\tau) \idotsint\limits_{\mu_{k0} ,\dots,
\mu_{k\dt{k}}}\exp\left(\tau f\right)
\end{array}\\
&f\! =\! x_i^T \mu_{k\dt{k}}\! + \beta \sum_{n=1}^\dt{k} \mu_{kn}^T
\mu_{k(n-1)}\! + w_k
\mu_{k0}^Tm_k \,.
\end{aligned}\label{eq:transint}
\end{align}
The integration in (\ref{eq:transint}) cannot be computed in closed form;
however, the value of the integral is only of interest in the limit
as $\tau \to \infty$. Therefore, Theorem \ref{thm:laplace}, an extension
of Laplace's approximation to general differentiable manifolds,
may be used to obtain an exact formula.
\begin{theorem}[Manifold Laplace Approximation]
Suppose $M\subset \mathbb{R}^n$ is a bounded $m$-dimensional differentiable manifold
and $f\!:\!\mathbb{R}^n\!\!\to\!\mathbb{R}$ is a smooth function on $M$. Further,
suppose $f$ has a unique global maximum on $M$, $x^\star = \argmax_{x\in M}
f(x)$. Then
\begin{align}
\lim_{\tau\to\infty} \frac{\int_M e^{\tau f(x)}}
{\left(\frac{2\pi}{\tau}\right)^m\left|\det
U^T\grad{}^2f(x^\star)U\right|^{\frac{1}{2}}
e^{\tau f(x^\star)}} = 1 \,,
\end{align}
where $U\in\mathbb{R}^{n\times m}$ is a matrix whose columns
are an orthonormal basis for the tangent space of $M$ at $x^\star$.\label{thm:laplace}
\end{theorem}
\begin{proof}
See the supplementary material. The general technique of this proof is to
transform coordinates between the manifold and its tangent plane
using the exponential map~\cite{Docarmo1992},
and then apply the multidimensional Laplace approximation in the
transformed Euclidean space.
\end{proof}
\begin{corollary}\label{cor:laplace}
Given a smooth function $f\!:\!\left(\mathbb{R}^D\right)^N \!\!\to\! \mathbb{R}$, with
a unique global maximum over the $N$-product of $\nobreak{(D-1)}$-spheres
$x^\star \in \left(\mathbb{S}^{D-1}\right)^N$,
\begin{align}
\int_{\left(\mathbb{S}^{D-1}\right)^N} e^{\tau
f(x)} \proptau{\to} e^{\tau f(x^\star)}\,.
\end{align}
\end{corollary}
\begin{proof}
This is Theorem \ref{thm:laplace} applied to
$\left(\mathbb{S}^{D-1}\right)^N$.
\end{proof}
Using Corollary \ref{cor:laplace} and the limiting approximation of the
modified Bessel function (\ref{eq:mbessapprox}) in equation (\ref{eq:transint}) yields
the following asymptotic behavior as $\tau \to \infty$:
\begin{align}
\begin{aligned}
p(x_i ; m_k, \tau_k) &\proptau{\to}
\exp\left(\tau(f^\star -1 - \beta\dt{k} -w_k) \right) \,.
\end{aligned}\label{eq:LaplacevMFTransition}
\end{align}
The only remaining unknown in the asymptotic expression, $f^\star$, can be found via constrained optimization
\begin{align}
\begin{aligned}
\hspace{-.3cm}\max_{\{\mu_{kn}\}_{n=1}^{\dt{k}}} \; & x_i^T \mu_{k\dt{k}} + \beta
\sum_{n=1}^\dt{k} \mu_{kn}^T \mu_{k(n-1)} + w_k \mu_{k0}^T m_k\; \\
& \text{ s.t. } \mu_{kn}^T\mu_{kn} = 1 \; \forall n \in \{ 0,\dots, \dt{k} \} \,.
\end{aligned}\label{eq:expopt}
\end{align}
The optimization (\ref{eq:expopt}) has a closed-form solution:
\begin{align}
\begin{aligned}
\mu_{k0} &= \frac{w_k m_k + \beta \mu_{k1}}{|| w_k m_k + \beta \mu_{k1}||_2}\\
\mu_{kn} &= \frac{\mu_{k(n+1)} + \mu_{k(n-1)}}{||\mu_{k(n+1)} +
\mu_{k(n-1)}||_2} \; { \scriptstyle \forall n \in \{1,\dots, \dt{k}-1 \} }\\
\mu_{k\dt{k}} &= \frac{x_i + \beta \mu_{k(\dt{k}-1)}}{||x_i + \beta
\mu_{k(\dt{k}-1)}||_2}\,.
\end{aligned}
\end{align}
These ternary relationships enforce that the optimal vMF
mean directions along the random walk lie on the geodesic between $m_k$ and $x_i$.
Therefore, this walk can be described geometrically by three angles, as shown in
Fig.~\ref{fig:transitionGeometry} (with $\bar{x}_k = x_i$): the angle $\phi$ between consecutive $\mu_{kn}$, the angle $\eta$ between $x_i$
and $\mu_{k \dt{k}}$, and the angle $\theta$ between $m_k$ and $\mu_{k0}$.
\begin{figure}
\centering
\begin{subfigure}{0.27\columnwidth}
\includegraphics[width=\textwidth]{{./figures/vmf3-crop}.pdf}
\caption{$\tau=100$}
\end{subfigure}
\begin{subfigure}{0.27\columnwidth}
\includegraphics[width=\textwidth]{{./figures/vmf1-crop}.pdf}
\caption{$\tau=1$}
\end{subfigure}
\hspace{0.1cm}
\begin{subfigure}{0.40\columnwidth}
\scalebox{0.8}{
\input{./angleDrawing.tex}
}
\end{subfigure}
\caption{Left: 2D vMF distributions. Right: geometry of the maximum likelihood
setting of $\mu_{k0}, \mu_{k1}, \dots,
\mu_{k\dt{k}}$ for the transition distribution.
\label{fig:transitionGeometry}}
\end{figure}
Given these definitions, standard trigonometry yields a set of three equations
in $\phi$, $\eta,$ and $\theta$:
\begin{align}
\begin{aligned}
w_k \sin(\theta^\star) = \beta \sin(\phi^\star) = \sin(\eta^\star) \\
\zeta = \theta^\star + \dt{k} \phi^\star + \eta^\star =
\arccos(m_k^T x_i ) \,,
\end{aligned}\label{eq:etathetaphi}
\end{align}
where $\zeta$ is the full angle between $x_i$ and $m_k$.
Since (\ref{eq:etathetaphi}) cannot be solved in closed-form, Newton's method
is used to compute $\phi^\star$, $\theta^\star$, and $\eta^\star$, which in turn
determines $f^\star$:
\begin{align}
f^\star = w_k \cos(\theta^\star) + \beta \dt{k} \cos(\phi^\star) +
\cos(\eta^\star) \,.
\end{align}
Returning to (\ref{eq:LaplacevMFTransition}), the transition asymptotics are
\begin{align}
p(x_i ; m_k;\tau_k) \proptau{\to}\exp\left(\hspace{-.2cm}\begin{array}{c}
\tau w_k(\cos(\theta^\star)-1)\\
+\tau \beta\dt{k}(\cos(\phi^\star)-1)\\
+\tau (\cos(\eta^\star)-1)\end{array}\hspace{-.2cm}\right)\,.
\end{align}
Substituting this into Eq.~\eqref{eq:ddpAssignmentDistri} with the earlier
definition $q=\exp(\tau Q)$, and taking the limit $\tau\to\infty$
yields the assignment rule $z_i = \argmax_{k} J_k$, where
\begin{align}
\hspace{-.2cm}J_k\! =\! \left\{\hspace{-.2cm}\begin{array}{l l}
\lambda +1 & k = K+1\\
\mu_k^T x_i & k\leq K, |\mathcal{I}_k| > 0\\
\hspace{-.1cm}\left(\begin{aligned} \dt{k} \beta &( \cos (\phi^\star ) -1) \\
+ w_k&( \cos(\theta^\star)-1) \\
+ \cos&(\eta^\star) + \Delta t_{k} Q \end{aligned} \right)
&k\leq K, |\mathcal{I}_k| = 0 \,.
\end{array}\right. \label{eq:labelupdate}
\end{align}
Note that if $Q\dt{k} < \lambda$, cluster $k$ can be removed
permanently as it will never be revived again.
Furthermore, for any $Q \leq \lambda$ all clusters are removed after each
timestep, and the algorithm reduces to DP-vMF-means.
\paragraph{Parameter Update:}
The parameter update rule for DDP-vMF-means comes from the asymptotic
behavior of (\ref{eq:ddpParameterDistri}) as $\tau \to \infty$. The
analysis for any new cluster is the same as that in Section
\ref{sec:dpvmf}, so our focus is again on the transitioned mean
direction posterior $p(\mu_{k\dt{k}}| \textbf{x}, \textbf{z}; m_k, \tau_k)$ (recall
that $\mu_k=\mu_{k\dt{k}}$ in the definition of the random $\vMF$
walk). This distribution can be expanded, similarly to
(\ref{eq:transint}), as:
\begin{align}
\hspace{-.2cm}\begin{aligned}
&p(\mu_{k\dt{k}}| \textbf{x}, \textbf{z}; m_k, \tau_k) \\
&\begin{array}{l}
= \idotsint\limits_{\mu_{k0}, \dots, \mu_{k(\dt{k}-1)}}
\begin{matrix}
p(\textbf{x} |\mu_{k\dt{k}};\tau)\cdot p(\mu_0;m_k,\tau_k) \\
\cdot \prod_{n=1}^{\dt{k}} p(\mu_{kn} |\mu_{k(n-1)};\xi)
\end{matrix} \\
= Z(\tau)^{|\mathcal{I}_k|} Z(\beta\tau)^\dt{k} Z(w_k\tau) \hspace{-.3cm}\idotsint\limits_{\mu_{k0} ,\dots,
\mu_{k(\dt{k}-1)}}\hspace{-.2cm}\exp\left(\tau f\right)
\end{array}\\
&f\! =\! \sum_{i\in\mathcal{I}_k}x_i^T \mu_{k\dt{k}}\! + \beta \sum_{n=1}^\dt{k} \mu_{kn}^T
\mu_{k(n-1)}\! + w_k
\mu_{k0}^Tm_k \,.
\end{aligned}\label{eq:ddpPosterior}
\end{align}
Define $\bar{x}_k = \sum_{i\in\mathcal{I}_k}x_i$.
Once again, applying Corollary \ref{cor:laplace}, the limit $\tau\to\infty$
removes the integrals over the marginalized mean directions. However, in
contrast to the label assignment update, $\mu_{k\dt{k}}$ is not marginalized
out. Therefore, an additional maximization with respect to $\mu_{k\dt{k}}$ to find
the concentration point of the posterior yields
\begin{align}
\hspace{-.15cm}\begin{aligned}
&p(\mu_{k\dt{k}}| \textbf{x}, \textbf{z}; m_k, \tau_k) \proptau{\to}\\
&\quad\exp(\tau(f^\star - |\mathcal{I}_k|
-\dt{k} \beta - w_k)) \\
& f^\star = w_k \cos(\theta^\star) + \beta \dt{k} \cos(\phi^\star) + ||\bar{x}_k||_2 \cos(\eta^\star)\,.
\end{aligned}\label{eq:paramupdateasym}
\end{align}
Analyzing the geometry of the geodesic between $\bar{x}_k/\|\bar{x}_k\|_2$ and $m_k$ (Fig.~\ref{fig:transitionGeometry}) there exist $\phi^\star$, $\theta^\star$ and
$\eta^\star$ such that
\begin{align}
\begin{aligned}
w_k \sin(\theta^\star) = \beta \sin(\phi^\star) = ||\bar{x}_k||_2
\sin(\eta^\star) \\
\zeta = \theta^\star + \dt{k} \phi^\star + \eta^\star = \arccos(m_k^T
\tfrac{\bar{x}_k}{\|\bar{x}_k\|_2}) \,,
\end{aligned}\label{eq:paramupdate}
\end{align}
which can be solved via Newton's method. Given the solution,
$\mu_{k}$ can be obtained by rotating
$\tfrac{\bar{x}_k}{\|\bar{x}_k\|_2}$ by angle $\eta^\star$ on the geodesic
shown in Fig.~\ref{fig:transitionGeometry} towards
$m_k$,
\begin{align}
\hspace{-.5cm}\begin{aligned}
\mu_{k} &= R(\eta^\star) \tfrac{\bar{x}_k}{ ||\bar{x}_k||_2}\,.
\end{aligned}
\end{align}
\paragraph{Weight Update:}
After the iteration of label and parameter updates has converged, the
weight $w_k$ must be updated for all clusters to reflect
the new uncertainty in the mean direction of cluster $k$.
This can be done by examining (\ref{eq:paramupdateasym}): Since
at the maximum of a $\vMF(\mu; m_k, w_k\tau)$ density, $\exp(\tau w_k m_k^T\mu)
= \exp(\tau w_k)$, $w_k$ is updated to $f^\star$.
\subsection{DP-vMF-means\label{sec:dpvmf}}
In this section, we provide a small-variance asymptotic analysis of the
label and parameter update steps of the Gibbs sampling algorithm for
the DP vMF mixture, yielding deterministic updates. As
with \emph{k}-means, the label assignments are computed sequentially
for all datapoints before the means are updated, and the process is
iterated until convergence. Pseudocode can be found in the supplement.
\paragraph{Label Update:} To derive a hyperspherical analog to
DP-means~\cite{jordan2012dpmeans}, consider the limit of the label
sampling step (\ref{eq:dplabelsample}) as $\tau \rightarrow \infty$.
The normalizer $Z(||\tau x_i + \tau_0 \mu_0||_2)$ approaches
\begin{align}
\begin{aligned}
Z(||\tau x_i + \tau_0 \mu_0||_2)
&\proptau{\to}
\exp(-\tau) \,,
\end{aligned}\label{eq:vMFnormalizerLimit}
\end{align}
where overscript $\ensuremath{\tau^{(*)}}$ denotes proportionality up to a finite
power of $\tau$, and where we have used the fact that as
$\tau\to\infty$, the modified Bessel function of the first kind
satisfies~\cite{Abramowitz65}
\begin{align}
\Bessel{D/2-1}(\tau) &= \frac{\exp(\tau)}{\sqrt{2\pi \tau}} \left(1 -
O\left(\frac{1}{\tau}\right) \right)
\proptau{\to} \exp\left(\tau\right)\,.\label{eq:mbessapprox}
\end{align}
To achieve a nontrivial result, the asymptotic behavior of $Z(\tau)$
must be matched by $\alpha$, so let $\alpha = \exp(\lambda \tau)$ to
obtain
\begin{align} \label{eq:marginalAssymptotics}
\frac{\alpha Z(\tau) Z(\tau_0)}{Z(||\tau x_i + \tau_0 \mu_0||_2 )}
&\proptau{=} Z(\tau) \exp(\tau(\lambda+1))\, .
\end{align}
Therefore, as $\tau\to\infty$, the label sampling step becomes
\begin{align}
\!\!\!\!\begin{aligned}
&\lim_{\tau \rightarrow \infty} p(z_i=k | \textbf{z}_{-i}, {\boldsymbol\mu}, \textbf{x}; \tau, \mu_0,
\tau_0) \\
&= \!\!\lim_{\tau \rightarrow \infty}
\!\left\{\begin{array}{ll}
\!\!\begin{matrix}
\frac{ |\mathcal{I}_k| e^{\tau (x_i^T \mu_k - \lambda-1)}}{ \sum_{j=1}^K
|\mathcal{I}_j| e^{\tau (x_i^T \mu_j - \lambda -1)} + c(\tau)} \\
\end{matrix} & k\leq K \\
\!\!\begin{matrix}
\frac{ c(\tau)}{ \sum_{j=1}^K |\mathcal{I}_j| e^{\tau (x_i^T \mu_j -
\lambda -1)} + c(\tau)} \\
\end{matrix} & k=K+1 \,,
\end{array}\right.
\end{aligned}
\end{align}
where we have used that the normalizers $Z(\tau)$ of $\vMF(x_i|
\mu_k;\tau)$ and Eq.~\eqref{eq:marginalAssymptotics} cancel and
$c(\tau) \proptau{=} 1$. Thus, as $\tau\to\infty$, sampling from
$p(z_i | \textbf{z}_{-i}, {\boldsymbol\mu}, \textbf{x}; \tau, \mu_0, \tau_0)$ is equivalent to the following
assignment rule:
\begin{align}
\label{eq:dpvMFassignment}
z_i = \argmax_{k\in\{1, \dots, K+1\}} \left\{\begin{array}{ll}
x_i^T\mu_k & k\leq K\\
\lambda+1 & k=K+1 \,.
\end{array}\right.
\end{align}
Since $-1 \leq x_i^T\mu_k \leq 1$, the parameter $\lambda$
can be restricted to the set $\lambda \in \left[-2, 0\right]$ without
loss of generality.
Intuitively $\lambda$ defines the maximum angular spread
$\phi_\lambda$ of clusters about their mean direction, via $\lambda =
\cos(\phi_{\lambda})-1$.
Note, that upon assigning a datapoint to a new cluster,
i.e.~$z_i = K+1$, the mean of that cluster is initialized to
$\mu_{K+1} = x_i$.
Finally, if an observation $x_i$ is the last one in its
cluster, the cluster is removed prior to finding the new label for
$x_i$ using Eq.~(\ref{eq:dpvMFassignment}).
\paragraph{Parameter Update:} Taking $\tau\to\infty$ in
the parameter posterior for cluster $k$ from
Eq.~\eqref{eq:vMFposterior} causes $\tau_0$ and $\mu_0$
to become negligible. Hence the parameter update becomes:
\begin{align}
\mu_k = \frac{\sum_{i\in\mathcal{I}_k}x_i}{\|\sum_{i\in\mathcal{I}_k}x_i\|_2} \quad\forall k \in \{1, \dots, K\} \,.
\end{align}
\paragraph{Objective Function:}
From Eq.~\eqref{eq:dpvMFassignment} we can see that assigning a
datapoint $x_i$ to cluster $k$ provides a score of $x_i^T\mu_k$,
whereas adding a new cluster provides a score of $\lambda+1 -
x_i^T\mu_{K+1} = \lambda$, since new mean directions are initialized
directly to $\mu_{K+1} = x_i$. Hence, the objective function that DP-vMF-means maximizes is
\begin{align}
J_{\text{DP-vMF}} = \sum_{k=1}^K \sum_{i\in\mathcal{I}_k} x_i^T\mu_k +
\lambda K \,.
\end{align}
\section{Optimistic Iterated Restarts (OIR)}
In our implementation of the algorithm we pay special attention to
speed and parallel execution to enable real-time performance for
streaming RGB-D data.
Observe that the main bottleneck of
DP-based hard clustering algorithms,
such as the proposed (D)DP-vMF-means, DP-means~\cite{jordan2012dpmeans} or
Dynamic means~\cite{campbell2013dynamic},
is the inherently sequential assignment of labels:
due to the creation of new clusters, the label assignments depend on
all previous assignments. While this is a key feature of the streaming
clustering algorithms, it poses a computational hindrance.
We address this issue with an optimistic parallel label assignment
procedure inspired by techniques for database concurrency
control~\cite{pan2013optimistic}.
First, we compute assignments in parallel (e.g.~on a GPU). If all
datapoints were assigned only to instantiated clusters, we output the
labeling. Otherwise, we find the lowest observation id $i$ that
modified the number of clusters, apply the modification, and
recompute the assignments for all observations $i' > i$ in parallel.
Thus, per data-batch, DP-vMF-means restarts once per new cluster, while
DDP-vMF-means restarts once for each new or revived cluster.
\section{Introduction}\label{sec:intro}
\input{intro.tex}
\section{Related Work}\label{sec:related}
\input{relatedWork.tex}
\section{Von-Mises-Fisher Mixture Models}
\input{vMF.tex}
\section{Dirichlet Process vMF-MM}\label{sec:dpvmf}
\input{dpvMF.tex}
\section{Dependent Dirichlet Process vMF-MM}\label{sec:ddpvmf}
\input{ddpvMF}
\input{implementation.tex}
\input{results.tex}
\section{Conclusion}\label{sec:conclusion}
\input{conclusion.tex}
\subsubsection*{Acknowledgements}
This work was partially supported by ONR MURI N00014-11-1-0688 and
ARO MURI W911NF-11-1-0391.
{\small
\input{bibl.bbl}
}
\end{document}
\section{Results}\label{sec:results}
\subsection{Evaluation of the DP-vMF-means Algorithm}
\input{./figureNYU.tex}
\begin{wrapfigure}{r}{0.20\columnwidth}
\vspace{-.4cm}
\hspace{-.55cm}\includegraphics[width=0.25\columnwidth]{./figures/rndSphereDataIwUncertain_K_30_spehre_0_cropped.png}
\caption*{\hspace{-.6cm}Synthetic Data}
\vspace{-.4cm}
\end{wrapfigure}
\textbf{Synthetic Data:}
First, we evaluate the behavior of the DP-vMF-means{} algorithm in
comparison to its parametric cousin, the spkm{} algorithm, on
synthetic 3D spherical data sampled from $K_\text{T} = 30$ true vMF
distributions. All evaluation results are shown as the mean and
standard deviation over 50 runs. The left plot of
Fig.~\ref{fig:synthEval} depicts the inferred number of clusters $K$ on
the horizontal axis as a function of the respective parameters of the
two algorithms: the number of clusters $K$ for spkm{} and the
parameter $\phi_\lambda$ for DP-vMF-means{} (recall that $\phi_\lambda =
\cos^{-1}(\lambda+1)$ as defined in Sec.~\ref{sec:dpvmf}). This figure
demonstrates the ability of the DP-vMF-means algorithm to discover the
correct number of clusters $K_\text{T}$, and the relative insensitivity
of the discovered number of clusters with respect to its parameter
$\phi_\lambda$.
The middle and right hand plots show two measures for clustering
quality. The normalized Mutual Information
(NMI)~\cite{strehl2003cluster}, depicted in the middle, is computed
using the true labels. DP-vMF-means{} achieves an almost perfect NMI of
$0.99$, while spkm{} only reaches $0.94$ NMI even with $K=K_\text{T}$.
The slightly superior performance of DP-vMF-means stems from its
enhanced ability to avoid local optima due to the way labels are
initialized: while spkm{} is forced to initialize $K$ cluster
parameters, DP-vMF-means{} starts with an empty set and adds clusters on
the fly as more data are labelled.
The NMI results are corroborated by the silhouette
score~\cite{rousseeuw1987silhouettes}, shown to the right in
Fig.~\ref{fig:synthEval}.
The silhouette
score is an internal measure for clustering quality that can be
computed without knowledge of the true clustering, and is used
to tune parametric clustering algorithms.
With a
maximum of $0.92$ DP-vMF-means{} reaches a close to perfect silhouette
score, indicating well-separated, concentrated clusters. Again, spkm{}
does not reach the same clustering performance even for $K=K_\text{T}$
for the same aforementioned reasons.
\input{./figureDDP.tex}
For additional reference we ran two sampling-based inference
algorithms for the DP-vMF-MM model.
The CRP-based inference of~\cite{bangert2010using}
was aborted after running for two days
without convergence.
This inefficiency has been noted
previously~\cite{Jain00asplit-merge}.
A more efficient alternative is the finite Dirichlet process (FSD)
approximation~\cite{ishwaran2002exact} to the DP-vMF-MM. The inference
can be parallelized and yielded results within minutes.
After $5000$ iterations the sampler converged to an incorrect number of
$23.0 \pm 1.67$ clusters on average over 10 sampler runs. These
convergence issues have also been noted before~\cite{chang13dpmm}. The
last sample of the different runs achieves an average NMI of $0.94 \pm
0.01$ and a silhouette score of $0.74\pm 0.04$.
\textbf{NYU v2 depth dataset:}
In this experiment, the DP-vMF-means{} and spkm{} algorithms were
compared on the NYU v2 RGB-D dataset~\cite{Silberman:ECCV12}. Surface
normals were extracted from the depth
images~\cite{Holz11RoboCup} and preprocessed with total variation
smoothing~\cite{RosmanWTKB12}.
We quantify the clustering quality in terms of the average silhouette
score over the clusterings of the 1449 scenes of the NYU v2 depth
dataset. Since we do not possess the true scene labeling, we use the
silhouette quality metric as a proxy for the NMI metric; this was
motivated by results of the synthetic experiment.
Across the whole NYU v2 dataset, the DP-vMF-means{} algorithm achieves
the highest average silhouette score of $0.75$ for $\phi_\lambda^\star
= 100^\circ$ as depicted in Fig.~\ref{fig:nyuStats}. The histogram
over the number of inferred clusters by DP-vMF-means{} for
$\phi_\lambda^\star$ indicates the varying complexity of the scenes
ranging from three to eleven. The clear peak at $K=4$ coincides with
the highest silhouette score for spkm{} ($0.73$) and explains the only
slightly lower silhouette score of spkm{}: most scenes in the dataset
exhibit four primary directions.
Figure~\ref{fig:nyuQualitative} shows a qualitative comparison of the
scene segmentation implied by the clustering of surface normals. In
comparison to spkm{}, the DP-vMF-means{} clustering results show the
ability of the algorithm to adapt the number of clusters to the scene
at hand. If the right number of clusters is selected for the spkm{}
clustering, the results have similar quality; however, the number of
clusters is generally not known a priori
and varies across scenes. This demonstrates two major advantages of
DP-vMF-means{} over spkm{}: (1) DP-vMF-means{} is less sensitive to the
parameter setting (see Fig,~\ref{fig:synthEval}, left) and (2) it is
easier to choose $\phi_\lambda$ than $K$ since it intuitively
corresponds to the maximum angular radius of a cluster, which can be
gauged from the type of data and its noise characteristics. For this
experiment $\phi_\lambda=100^\circ$ is justified by the typical
Manhattan structure~\cite{coughlan1999manhattan} of the indoor environment plus $10^\circ$ to
account for sensor noise.
\subsection{Evaluation of the DDP-vMF-means Algorithm}
\paragraph{Real-time Directional Segmentation:}
In fields such as mobile robotics or augmented reality, it is uncommon
to observe just a single RGB-D frame of a scene; more typically, the
sensor will observe a temporal sequence of frames. The following
experiment demonstrates the temporally consistent clustering capability
of the DDP-vMF-means algorithm on surface normals extracted from a sequence
of depth images recorded in an indoor environment. Each frame is
preprocessed in $11$ms using edge-preserving smoothing
with a hybrid CPU-GPU guided filter~\cite{he2010guided}.
We compare against the ad-hoc approaches of clustering on a
frame-by-frame basis using DP-vMF-means, both with and without
initializing the algorithm from the previous frame's clusters. The
former is referred to as sequential DP-vMF-means (sDP-vMF-means).
sDP-vMF-means achieves a greedy frame-to-frame label consistency,
but, unlike DDP-vMF-means, it cannot reinstantiate previous clusters after
multiframe lapses.
Motivated by the DP-vMF-means evaluation, all algorithms were run
with $\phi_\lambda=100^\circ$. For DDP-vMF-means $\beta = 10^5$ and $Q
=\frac{\lambda}{400}$.
The differences in labeling consistency can be observed in rows
two and three of Fig.~\ref{fig:ddpvmf}, which shows the percentage of
normals associated with a specific cluster. While DDP-vMF-means is
temporally consistent and reinstantiates the lime-green and red
clusters, observed in the first half of the run, DP-vMF-means
erroneously creates new clusters. We do not depict the percentages of
surface normals associated with the clusters for the batch DP-vMF-means
algorithm, since the there is no label consistency between time-steps
as can be observed in the last row of Fig.~\ref{fig:ddpvmf}.
The average run-time per frame was $28.4$~ms for batch DP-vMF-means,
$12.8$~ms for sDP-vMF-means, $20.4$~ms for DDP-vMF-means, and $13.6$~ms for spkm
with $K=5$.
The increased running time of batch DP-vMF-means is a result of
clustering each batch of surface normals in isolation; OIR label
assignment needs several restarts to
assign labels to all surface normals. By initializing the clusters from a
previous frame, sDP-vMF-means only incurs labeling restarts if a
new cluster is observed, and hence has significantly lower run time.
DDP-vMF-means is slightly slower than sDP-vMF-means since it is
keeping track of both observed and unobserved clusters.
|
\section{Introduction and Overview}
It is a widely held belief that the universal long-distance behavior of a quantum phase of matter at zero temperature can be encoded into an effective field theory.\footnote{It is hard to make this rigorous since neither the notion of a phase of matter nor that of an effective field theory has been formalized.} In the case of gapped phases of matter, the extreme infrared should be described by a Topological Quantum Field Theory.
It has been rigorously shown that the ground state of any gapped 1+1d Hamiltonian with a short-range interaction can be approximated by a Matrix Product State \cite{Hastings}. This representation is very efficient, especially in the translationally-invariant case, and is well-suited to the Renormalization Group analysis. In particular, it leads to a classification of Short-Range Entangled Phases of 1+1d matter in terms of group cohomology
\cite{ChenGuWenone,ChenGuWentwo,FidkowskiKitaev}.
It is natural to ask about the connection between these two approaches to gapped phases of matter. In this note we answer this question in the case of bosonic systems with a finite symmetry $G$. For simplicity, we assume that all elements of $G$ act unitarily (i.e. we do not allow for time-reversing symmetries). The case of time-reversing symmetries and fermionic phases will be addressed in separate publications \cite{KTY2}.
In brief, the results are as follows. We show that a standard-form MPS is naturally associated with a module $M$ over a finite-dimensional semisimple algebra $A$. The universality class of the MPS depends only on the center $Z(A)$. On the other hand, every unitary 2d TQFT has a state-sum construction which uses a semisimple algebra as an input. Further, given a module $M$ over this algebra, one naturally gets a particular state in the TQFT space of states. We show that this state is precisely the MPS associated to the pair $(A,M)$. Since the TQFT depends only on $Z(A)$, we reproduce the fact that the universality class of the MPS depends only on $Z(A)$.
In the case of an MPS with a symmetry $G$, a similar story holds. A $G$-equivariant MPS is encoded in a $G$-equivariant module $M$ over a $G$-equivariant semisimple algebra $A$. Such an algebra can be used to give a state-sum construction of a $G$-equivariant TQFT, while every $G$-equivariant module $M$ gives rise to a particular state. This state is an equivariant MPS state. Again, different $A$ can give rise to the same TQFT. This leads to an equivalence relation on $G$-equivariant algebras which is a special case of Morita equivalence.
An indecomposable phase with symmetry $G$ is therefore associated with a Morita-equivalence class of indecomposable $G$-equivariant algebras. The classification of such algebras is well known \cite{Ostrik} and leads to an (also well-known \cite{ChenGuWenone,ChenGuWentwo,FidkowskiKitaev}) classification of bosonic 1+1d gapped phases of matter with symmetry $G$. In the special case of Short-Range Entangled gapped phases, we recover the group cohomology classification of SPT phases.
A. K. would like to thank P. Etingof and V. Ostrik for discussions. A.T. is grateful to I.~Saberi and D.~Williamson for helpful conversations. While this paper was nearing completion, we learned that closely related results have been obtained by K. Shiozaki and S. Ryu. This paper was supported in part by the U.S. Department of Energy, Office of Science, Office of High Energy Physics, under Award Number DE-SC0011632.
\section{Matrix Product States at RG Fixed Points}
\subsection{Matrix Product States}\label{sec:mps}
In this section, we review Matrix Product States (MPS) and extract the algebraic data that characterizes them at fixed points of the Renormalization Group (RG). We find that a fixed point MPS is described by a module over a finite-dimensional semisimple algebra. We discuss the notion of a gapped phase and argue that they are classified by finite-dimensional semisimple commutative algebras. Given a fixed point MPS and the corresponding semisimple algebra $A$, the commutative algebra characterizing the gapped phase is the center of $A$, denoted $\mathcal A = Z(A)$.
The models we consider are defined on Hilbert spaces that are tensor products of finite-dimensional state spaces $A$ on the sites of a 1D chain. We are interested in Hamiltonians with an energy gap that persists in the thermodynamic limit of an infinite chain. A large class of examples of gapped systems come from local commuting projector (LCP) Hamiltonians; that is, $H=\sum h_{s,s+1}$, where the $h_{s,s+1}$ are projectors that act on sites $s$, $s+1$ and commute with each other. Since the local projectors commute, an eigenstate of $H$ is an eigenstate of each projector. It follows that the gap of $H$ is at least $1$. Thus LCP Hamiltonians are gapped in the thermodynamic limit. In one spatial dimension, ground states of gapped Hamiltonians are efficiently approximated by an ansatz called a matrix product state (MPS) \cite{Hastings}, which we recall below.\footnote{We only consider translationally-invariant MPS.} From each MPS, one can construct a LCP parent Hamiltonian, which has the MPS as a ground state. To discuss and classify 1D gapped Hamiltonians, it suffices to consider the parent Hamiltonians of the MPS that approximate their ground states.
Consider a closed chain of $N$ sites, each with a copy of a \emph{physical} Hilbert space $A \simeq{\mathbb C}^d$ and two copies $V^L$, $V^R$ of a \emph{virtual} space ${\mathbb C}^D$. We identify $V^L=V$ and $V^R=V^*$ and choose a Hilbert space structure on $V$. Between each adjacent pair $(s,s+1)$ of sites, place the maximally entangled state\begin{equation}\ket{\omega}_{s,s+1}=\sum_{i=1}^D\ket{i}\otimes\ket{i}\in V^R_s\otimes V^L_{s+1}\end{equation}
An MPS tensor\footnote{More generally, the tensors ${\mathcal P}_s$ may depend on the site index $s$. But any translationally-invariant state has an MPS representation with a site-independent tensor \cite{MPS2}.} is a linear map ${\mathcal P}:V^L\otimes V^R\rightarrow A$. The MPS associated to ${\mathcal P}$ is the state
\begin{equation}\ket{\psi_{\mathcal P}}=\left({\mathcal P}_1\otimes{\mathcal P}_2\otimes \cdots\otimes{\mathcal P}_N\right)\left(\ket{\omega}_{12}\otimes\ket{\omega}_{23}\otimes\cdots\otimes\ket{\omega}_{N1}\right)\in A\otimes A\otimes \cdots \otimes A.\end{equation}
Since $\ket{\psi_{\mathcal P}}$ lies in the image of ${\mathcal P}^{\otimes N}$, we do not lose generality by truncating $A$ to ${\rm im}\, P$. We will assume we have done so in the following. Equivalently, we assume that the adjoint MPS tensor $T={\mathcal P}^\dagger$ is injective\footnote{To avoid confusion, we stress that injectivity of $T$ is unrelated to the notion of an injective MPS in the sense of \cite{MPS}. In particular, while we will always assume that $T$ is injective, we will not assume that the the ground state of the parent Hamiltonian is unique.}. The MPS wavefunction can be expressed as a trace of a product of matrices, hence its name. In the basis $\{e_i\}_{i=1,\ldots,d}$ of $A$, the conjugate state takes the form
\begin{equation}\langle {\psi_T}\vert=\sum_{i_1\cdots i_N=1}^d\Tr[T(e_{i_1})\cdots T(e_{i_N})]\langle{i_1\cdots i_N}\vert\end{equation}
There may be many different ways to represent a given state in $A^{\otimes N}$ in an MPS form. Even the dimension of the virtual space $V$ is not uniquely defined. In general, it is not immediate to read off the properties of the state $\psi_T$ from the tensor $T$.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\node at (0,0) {\ldots};
\draw[line width=2pt] (0.5,0)--node[above] {$\mu$} (2,0) circle(2pt)--node[above] {$\nu$} (4,0) circle (2pt)--node[above] {$\rho$} (6,0) circle(2pt)--node[above] {$\sigma$} (7.5,0);
\node at (8,0) {\ldots};
\draw (2,0)--(2,2) node[pos =0.5, left]{$T^{i}_{\ \mu \nu}$} node[right] {$i$}; \draw (4,0)--(4,2) node[pos =0.5, left] {$T^j_{\ \nu \rho}$} node[right] {$j$}; \draw (6,0)--(6,2) node[pos =0.5, left] {$T^k_{\ \rho \sigma}$} node[right] {$k$};
\end{tikzpicture}
\caption{An MPS represented as a tensor network} \label{fig:f1}
\end{figure}
For the tensor $T$, one can construct a LCP Hamiltonian $H_T$, called the \emph{parent Hamiltonian}\footnote{There is a more general notion of a parent Hamiltonian where $h$ is any operator with this kernel; however, we will always take $h$ to be the projector.} of $\ket{\psi_T}$, which has $\ket{\psi_T}$ as a ground state. It is given as a sum of $2$-site terms $h_{s,s+1}$ that project onto the orthogonal complement of
$\ker h=({\mathcal P}\otimes{\mathcal P})(V\otimes\ket{\omega}\otimes V^*)$. Explicitly,\begin{equation}\label{parent}H_T=\sum_s h_{s,s+1}\hspace{5mm}\text{ where }\hspace{5mm}h_{s,s+1}=\mathds{1}-({\mathcal P}_s\otimes{\mathcal P}_{s+1})\delta({\mathcal P}_s^+\oplus{\mathcal P}_{s+1}^+)\end{equation}where $\delta$ is the projector onto $(V_s\otimes\ket{\omega}\otimes V_{s+1}^*)$ and ${\mathcal P}_s^+:=(T_s
{\mathcal P}_s)^{-1}T_s$ is a left inverse of ${\mathcal P}_s$. The local projectors $h_{s,s+1}$ commute, so $H_T$ is gapped. $\ket{\psi_T}$ is annihilated by $h_{s,s+1}$, $\forall s$ and therefore also by $H_T$.
In general, $H_T$ has other ground states. Consider a state of the form\begin{equation}\label{X-state}\ket{\psi_T^X}=\left({\mathcal P}_1\otimes{\mathcal P}_2\otimes\cdots\otimes{\mathcal P}_N\right)\left(\ket{\omega}_{12}\otimes\ket{\omega}_{23}\otimes\cdots\otimes\ket{\omega^X}_{N1}\right)\end{equation}
for some virtual state \begin{equation}\ket{\omega^X}=\sum_{i=1}^DX_{ij}\ket{i}\otimes\ket{j}\in V^*\otimes V\end{equation}where $X$ is a matrix that commutes with $T(a)$ for all $a\in A$. Note that $\ket{\omega^\mathds{1}}=\ket{\omega}$ and so $\ket{\psi_T^\mathds{1}}=\ket{\psi_T}$. The states \eqref{X-state} are clearly annihilated by $h_{s,s+1}$ for $s\ne N$. To see that they are annihilated by $h_{N1}$, note that tensor $T(e_i)XT(e_j)$ is expressible as a linear combination of tensors $T(e_i)T(e_j)$ if and only if $X$ commutes with every $T(e_i)$. The conjugate states have wavefunctions\begin{equation}\label{GS}\langle{\psi_T^X}\vert =\sum \Tr[X^\dagger T(e_{i_1})\cdots T(e_{i_n})]\langle{i_1\cdots i_N}\vert\end{equation}
We will refer to these states as \emph{generalized MPS}.
It turns out that all ground states of $H_T$ can be written as generalized MPS. One can always take $T$ to be an isometry with respect to some inner product on $A$ and the standard inner product
\begin{equation}\langle M|N\rangle =\Tr[M^\dagger N]\quad M,N\in{\rm End}(V)
\end{equation}on ${\rm End}(V)$. For an orthogonal basis $\{e_i\}$ of $A$, $\Tr[T(e_i)^\dagger T(e_j)]=\delta_{ij}$. Consider the case $N=1$. An arbitrary state
\begin{equation}\bra{\psi}=\sum_i a_i\bra{i}
\end{equation}can be written in generalized MPS form \eqref{GS} if one takes
\begin{equation}X=\sum_ja_jT(e_j)^\dagger\end{equation}Thus generalized MPS with commuting $X$ are the only ground states. Neither the number of generalized MPS nor the number of ground states depends on $N$; thus, the argument extends to all $N$.
Suppose the data $(A_1,V_1,T_1)$ and $(A_2,V_2,T_2)$ define two MPS systems with parent Hamiltonians $H_1$ and $H_2$. Consider the composite system $(A_1\otimes A_2,V_1\otimes V_2,T_1\otimes T_2).$ It has ${\mathcal P}={\mathcal P}_1\otimes{\mathcal P}_2$ and $\delta=\delta_1\otimes\delta_2$. Then\begin{eqnarray}h_{A\otimes B}&=&\mathds{1}_{A_1\otimes A_2}-{\mathcal P}^2\delta{\mathcal P}^{+2}_{A_1\otimes A_2}\nonumber\\&=&\mathds{1}_{A_1}\otimes\mathds{1}_{A_2}-{\mathcal P}^2\delta{\mathcal P}^{+2}_{A_1}\otimes {\mathcal P}^2\delta{\mathcal P}^{+2}_{A_2}\nonumber\\&=&(\mathds{1}_{A_1}-{\mathcal P}^2\delta{\mathcal P}^{+2}_{A_1})\otimes\mathds{1}_{A_2}+\mathds{1}_{A_1}\otimes(\mathds{1}_{A_2}-{\mathcal P}^2\delta{\mathcal P}^{+2}_{A_2})\nonumber\\&=&h_{A_1}\otimes\mathds{1}_{A_2}+\mathds{1}_{A_1}\otimes h_{A_2}\end{eqnarray}where the penultimate line follows from the fact that ${\mathcal P}^2\delta{\mathcal P}^{+2}$ is a projector. Therefore, the composite parent Hamiltonian is\begin{equation}\label{stackedham}H_{A\otimes B}=H_{A_1}\otimes\mathds{1}_{A_2}+\mathds{1}_{A_1}\otimes H_{A_2}.\end{equation}
\subsection{RG-fixed MPS and gapped phases}
Under real-space renormalization group (RG) flow \cite{RG}, adjacent pairs of sites are combined into blocks with physical space $A\otimes A$. The MPS form of the state is preserved, with the new MPS tensor being
\begin{equation}
T'(a\otimes b)=T(a)T(b),
\end{equation}
where on the r.h.s. the multiplication is matrix multiplication. We also define ${\mathcal P}^\prime=T'^\dagger$. Though an RG step squares the dimension of the codomain of the MPS tensor, the rank is bounded above by $D^2$, and so the truncated physical space ${\rm im}({\mathcal P}^\prime)$ never grows beyond dimension $D^2$.
An \emph{RG fixed MPS tensor} is an MPS tensor such that ${\mathcal P}$ and ${\mathcal P}^\prime$ have isomorphic images and are identical (up to this isomorphism) as maps. That is, there exists an injective map $\mu: A\rightarrow A\otimes A$ such that
\begin{equation}
\mu\circ {\mathcal P}={\mathcal P}^\prime.
\end{equation}
If we denote $m=\mu^\dagger$, this is equivalent to
\begin{equation}
T(m(a\otimes b))=T(a)T(b).
\end{equation}
Since $T$ was assumed to be injective, this equation completely determines $m$. Similarly, the fact that matrix multiplication is associative implies that $m:A\otimes A\rightarrow A$ is an associative multiplication on $A$. The map $T:A\rightarrow {\rm End}(V)$ then gives $V$ the structure of a module over $A$. Since $T$ is injective, this module is faithful (all nonzero elements of $A$ act nontrivially). The statement that $X$ commutes with $T$ in the ground state of the parent Hamiltonian is the statement that $X$ is a module endomorphism of $V$.
As previously stated, a state in $A^{\otimes N}$ may have multiple distinct MPS descriptions. Given an RG fixed MPS, one can always choose $T$ to have a certain \emph{standard form} \cite{MPS}. When this is done, the matrices $T(a)$ are simultaneously block-diagonalized, for all $a\in A$. Moreover, if we denote by $T^{(\alpha)}$ the $\alpha^{\rm{th}}$ block, say of size $L_\alpha\times L_\alpha$, then the matrices $T^{(\alpha)}(e_i)$ span the space of $L_\alpha\times L_\alpha$ matrices. That is, $T^{(\alpha)}$ defines a surjective map from $A$ to the space of $L_\alpha\times L_\alpha$ matrices.
For an RG-fixed MPS tensor in its standard form, one can easily see that $A$ is a direct sum of matrix algebras. Indeed, each block $A^\alpha$ defines a surjective homomorphism $T^\alpha$ from $A$ to the algebra of $L_\alpha\times L_\alpha$ matrices, and if an element of $A$ is annihilated by all these homomorphisms, then it must vanish. Thus we get a decomposition
\begin{equation}\label{sumofmatrixalgebras}
A=\oplus_{\alpha} A^\alpha,
\end{equation}
where each $A^\alpha=(\ker T^\alpha)^\perp$ is isomorphic to a matrix algebra. We stress that some of these homomorphisms might be linearly dependent, so the number of summands may be smaller than the number of blocks in the standard form of $T$. An algebra of such a form is semisimple, that is, any module is a direct sum of irreducible modules. More specifically, any module over a matrix algebra of $L\times L$ matrices is a direct sum of several copies of the obvious $L$-dimensional module. This basic module is irreducible. If, for a particular $A^\alpha$, $T$ contains more than one copy of the irreducible module, the corresponding blocks in the standard form of $T$ are not independent.
The ground-state degeneracy is simply related to the properties of the algebra $A$. Namely, the number of ground states is equal to the number of independent blocks in a standard-form MPS, or equivalently the number of summands in the decomposition (\ref{sumofmatrixalgebras}). Since the center of a matrix algebra consists of scalar matrices and thus is isomorphic to ${\mathbb C}$, one can also say that the number of ground states is equal to the dimension of $\mathcal A=Z(A)$.
Two gapped systems are said to be in the same phase if their Hamiltonians can be connected by a Local Unitary (LU) evolution, i.e. if they are related by conjugation with a finite-time evolution operator for a local time-dependent Hamiltonian \cite{ChenGuWenzero}. Clearly, the ground-state degeneracy is the same for all systems in a particular phase. In fact, for 1+1d gapped bosonic systems, it completely determines the phase \cite{MPS,ChenGuWenone}.
It is convenient to introduce an addition operation $\oplus$ on systems and phases. Given two 1+1d systems with local Hilbert spaces $A_1$ and $A_2$, we can form a new 1+1d system with the local Hilbert space $A_1\oplus A_2$. The Hamiltonian is taken to be the sum of the Hamiltonians of the two systems plus projectors which enforce the condition that neighboring ``spins'' are either both in the $A_1$ subspace or in the $A_2$ subspace. The ground state degeneracy is additive under this operation. A phase is called decomposable if it is a sum of two phases, otherwise it is called indecomposable. Clearly, it is sufficient to classify indecomposable phases.
It is easy to see that if $A$ decomposes as a sum of subalgebras, the corresponding phase is decomposable. Further, an indecomposable semisimple algebra $A$ is isomorphic to a matrix algebra. The corresponding ground state is unique. Moreover, while the parent Hamiltonians for different matrix algebras are different, they all correspond to the same phase \cite{ChenGuWenone}, i.e. are related by a Local Unitary evolution. Hence the phase is determined by the number of components in the decomposition \eqref{sumofmatrixalgebras}, or in other words, by $Z(A)$.
\section{Topological Quantum Field Theory}
We have seen above that an RG-fixed MPS state is associated with a finite-dimensional semisimple algebra $A$, and that the universality class of the corresponding phase depends only on the center of $A$. On the other hand, it is known since the work of Fukuma, Hosono, and Kawai \cite{FHK} that for any finite-dimensional semisimple algebra $A$ with an invariant scalar product one can construct a unitary 2D TQFT, and that the isomorphism class of the resulting TQFT depends only on the center of $A$. In this section we show that this is not a mere coincidence, and that the ground states of this TQFT can be naturally written in an MPS form, with an RG-fixed MPS tensor.
\subsection{State-sum construction of 2d TQFTs}
A (closed) 2D TQFT associates a space of states $\mathcal A$ to an oriented circle, and a vector space $\mathcal A^{\otimes n}$ to $n$ disjoint oriented circles. Further, suppose we are given an oriented bordism from $n$ circles to $l$ circles, i.e. a compact oriented 2d manifold $\Sigma$ whose boundary consists of $l$ circles oriented in the same way as $\Sigma$ and $n$ circles oriented in the opposite way. A 2d TQFT associates to $\Sigma$ a linear map from $\mathcal A^{\otimes n}$ to $\mathcal A^{\otimes l}$. This map is invariant under diffeomorphisms. Also, gluing bordisms taking care that orientations agree corresponds to composing linear maps.
Let us mention some special cases. If $\Sigma$ is closed (i.e. has an empty boundary), then the 2D TQFT associates to it a linear map ${\mathbb C}\rightarrow{\mathbb C}$, i.e. a complex number $Z_\Sigma$, called the partition function. If $\Sigma$ is a pair-of-pants bordism from two circles to one circle, the corresponding map $m: \mathcal A\otimes \mathcal A\rightarrow \mathcal A$ defines an associative, commutative product on $\mathcal A$. The cap bordism defines a symmetric trace function ${\rm Tr}:\mathcal A\rightarrow {\mathbb C}$ such that the scalar product
$\eta(a,b)={\rm Tr}(ab)$ is symmetric and non-degenerate. These data make $\mathcal A$ into a commutative Frobenius algebra. It is known that a two-dimensional TQFT is completely determined by the commutative Frobenius algebra structure on $\mathcal A$ \cite{Atiyah,MS,Abrams}. The state-operator correspondence identifies $\mathcal A$ with the algebra of local operators. This Frobenius algebra encodes the $2$- and $3$-point functions on the sphere, from which all other correlators, including the partition function, can be reconstructed.
In 2d there is an essentially trivial family of unitary oriented TQFTs parameterized by a positive real number $\lambda$. The partition function of such a TQFT on a closed oriented 2d manifold $\Sigma$ is $\lambda^\chi(\Sigma)$, while the Hilbert space attached to a circle is one-dimensional. Such 2d TQFTS are called invertible, since the partition function is a nonzero number for any $\Sigma$. Since, by the Gauss-Bonnet theorem, $\chi(\Sigma)$ can be expressed as an integral of scalar curvature, tensoring a 2d TQFT by an invertible 2d TQFT is equivalent to redefining the TQFT action by a local counterterm which depends only on the background curvature. One usually disregards such counterterms. In what follows we will follow this practice and regard TQFTs related by tensoring with an invertible TQFT as equivalent.
Every unitary oriented 2d TQFT\footnote{More precisely, every equivalence class of unitary oriented 2d TQFTs, in the sense explained in the previous paragraph.} has an alternative construction called the state-sum construction \cite{FHK}, which is combinatorial and manifestly local. The input for this construction is a finite-dimensional semisimple algebra $A$, which is not necessarily commutative. To compute the linear maps associated to a particular bordism $\Sigma$, one needs to choose a triangulation of $\Sigma$. Nevertheless, the result is independent of the choice of the triangulation. The connection between the not-necessarily commutative algebra $A$ and the commutative algebra $\mathcal A$ is that $\mathcal A$ is $Z(A)$, the center of $A$. From the perspective of open-closed TQFTs, $A$ is the algebra of states on the interval for a particular boundary condition. The scalar product on $\mathcal A$ is also fixed by the structure of $A$.
Let us describe the state-sum construction for the partition function $Z_\Sigma$ of a closed oriented 2D manifold $\Sigma$, following \cite{FHK}. Fix a basis $e_i,$ $i\in S$, of $A$. We define the following tensors:\begin{equation}\label{gandC}
\eta_{ij}=\eta(e_i,e_j)=\Tr_A P_i P_j\quad\quad C_{ijk}=\Tr_A P_i P_j P_k
\end{equation}
Here $P_i:A\rightarrow A$ is the operator of multiplication by $e_i$.
The tensor $\eta_{ij}$ is symmetric and non-degenerate (if the algebra $A$ is semi-simple); the tensor $C_{ijk}$ is cyclically symmetric. We also denote by $\eta^{ij}$ the inverse to the tensor $\eta_{ij}$. Note also that $C_{ijk}$ is related to the structure constants $C^i{}_{jk}$ in this basis by
\begin{equation}
C^i{}_{jk}=\sum_l \eta^{il}C_{ljk}.
\end{equation}
Let $T(\Sigma)$ be a triangulation of $\Sigma$. A coloring of a 2-simplex $F$ of $T(\Sigma)$ is a choice of a basis vector $e_i$ for each 1-simplex $E\in \partial F$. A coloring of $T(\Sigma)$ is a coloring of all 2-simplices of $T(\Sigma)$. Note that each 1-simplex of $T(\Sigma)$ has two basis vectors attached to it, one from each 2-simplex that it bounds. The weight of a coloring is the product of $C_{ijk}$ over 2-simplices and $\eta^{ij}$ over 1-simplices, where the cyclic ordering of indices for each 2-simplex is determined by the orientation of $\Sigma$. The partition function is the sum of these weights over all colorings.
Topological invariance of $Z_\Sigma$ can be shown as follows. It is known that any two triangulations of a smooth manifold are related by a finite sequence of local moves \cite{Pachner}. In two dimensions, there are two moves - the 2-2 move and the 3-1 move, depicted in Figure \ref{fig:f2} - which swap two or three faces of a tetrahedron with their complement. Invariance of the state-sum under the 2-2 ``fusion'' move reads\begin{equation}\label{2-2}C_{ij}{}^pC_{pk}{}^l=C_{jk}{}^pC_{ip}{}^l\end{equation}Similarly the 3-1 move reads\begin{equation}\label{3-1}C_{i}{}^{mn}C_{nl}{}^kC^{l}{}_{mj}=C_{ij}{}^k\end{equation}These axioms are satisfied by any finite-dimensional semisimple algebra $A$ \cite{FHK}; therefore, the partition sum is a topological invariant\footnote{In two dimensions, there is no difference between topological and smooth manifolds.}.
\begin{figure}
\centering
\begin{tikzpicture} [scale =0.5]
\draw (0,0)--(3,2)--(6,0)--(3,-2)--cycle; \draw (3,2)--(3,-2);
\draw [<->] (7,0)--(8.5,0);
\draw (9.5,0)--(12.5,2)--(15.5,0)--(12.5,-2)--cycle; \draw (9.5,0)--(15.5,0);
\end{tikzpicture} \hspace{1cm}
\begin{tikzpicture}[scale = 0.6]
\draw (0,0)--(2,2*1.732)--(4,0)--cycle;
\draw (0,0)--(2,2*0.58)--(4,0);
\draw (2,2*0.58)--(2,2*1.738);
\draw[<->] (5,0.8)--(6.5,0.8);
\begin{scope}[shift={(7.5,0)}]
\draw (0,0)--(2,2*1.732)--(4,0)--cycle;
\end{scope}
\end{tikzpicture}
\caption{The 2-2 (left) and the 3-1 (right) Pachner moves} \label{fig:f2} \end{figure}
\subsection{Open-closed 2d TQFT}
So far we have discussed what is known as closed 2D TQFTs. That is, the boundary circles were interpreted as spacelike hypersurfaces, and thus each spatial slice had an empty boundary. The notion of a TQFT can be extended to incorporate spatial boundaries; such theories are called open-closed TQFTs.
In such a theory a spatial slice is a compact oriented manifold, possibly with an nonempty boundary. That is, it is a finite collection of oriented intervals and circles. A bordism between such spatial slices is a smooth oriented \emph{surface with corners}: paracompact Hausdorff spaces for which each point has a neighborhood homeomorphic to an open subset of a half-plane. Surfaces with corners are homeomorphic, but typically not diffeomorphic, to smooth surfaces with a boundary.
The corner points subdivide the boundary of the bordism into two parts: the initial and final spatial slices, and the rest. We will refer to the initial and final spatial slices as the cut boundary, while the rest will be referred to as the brane boundary. The cut boundary can be thought of as spacelike, while the brane boundary is timelike. Bordisms are composed along their cut boundary (hence the name), while on the brane boundary one needs to impose boundary conditions (known as D-branes in the string theory context, hence the name). More precisely, if $\mathcal C$ is the set of boundary conditions, one needs to label each connected component of the brane boundary with an element of $\mathcal C$.
An open-closed 2d TQFT associates a vector space $V_{MM'}$ to every oriented interval with the endpoints labeled by $M,M'\in\mathcal C$, and a vector space $\mathcal A$ to every oriented circle. To a collection of thus labeled compact oriented 1D manifolds it attaches the tensor product of spaces $V_{MM'}$ and $\mathcal A$. To every bordism with corners labeled in the way explained above, it attaches a linear map from a vector space of the `incoming'' cut boundary to the vector space of the ``outgoing'' cut boundary. Gluing bordisms along their cut boundaries corresponds to composing the linear maps.
Just like in the case of a closed 2d TQFT, one can describe algebraically the data which are needed to construct a 2d open-closed TQFT. We refer to \cite{MS} for details. Suffice it to say that each space $V_{MM}$ is a (possibly noncommutative) Frobenius algebra, and each space $V_{MM'}$ is a left module over $V_{MM}$ and a right module over $V_{M'M'}$. That is, to every element $x\in V_{MM}$ one associates a linear operator $T^M(x): V_{MM'}\rightarrow V_{MM'}$ so that composition of elements of $V_{MM}$ corresponds to the composition of linear operators: $T^M(x) T^M(x')=T^M(xx')$ (and similarly for $V_{M'M'}$). Also, for every $M\in\mathcal C$ there is a map $\iota^M: \mathcal A\rightarrow V_{MM}$ which is a homomorphism of Frobenius algebras. The dual map $\iota_M: V_{MM}\rightarrow \mathcal A$ is known as the generalized boundary-bulk map. In particular, if we act with $\iota_M$ on the identity element of the algebra $V_{MM}$, we get a distinguished element $\psi_M\in \mathcal A$ called the boundary state corresponding to the boundary condition $M$. Geometrically, $\psi_M$ is the element of $\mathcal A$ which the open-closed TQFT associates to an annulus whose interior circle is a brane boundary labeled by $M$, while the exterior circle is an outgoing cut boundary.
One may wonder if it is possible to reconstruct the open-closed TQFT from the closed TQFT. The answer turns out to be yes if $\mathcal A$ is a semisimple, i.e. if every module over $\mathcal A$ is a sum of irreducible modules \cite{MS}.\footnote{This might seem like a rather uninteresting case, since by the Wedderburn theorem every commutative semisimple algebra is isomorphic to a sum of several copies of ${\mathbb C}$. But as explained below unitarity forces $\mathcal A$ to be semisimple. Also, in the case of TQFTs with symmetries and fermionic TQFTs the classification of semisimple algebras is more interesting.} Then $\mathcal C$ is the set of finite-dimensional modules over $\mathcal A$, and $V_{MM'}$ is the space of linear maps from the module $M$ to the module $M'$ commuting with the action of $\mathcal A$ (i.e. $V_{MM'}$ is the space of module homomorphisms). Conversely, one can reconstruct the algebra $\mathcal A$ from any ``sufficiently large'' brane $M\in\mathcal C$: if we assume that the module $M$ is faithful (i.e. all nonzero elements of $\mathcal A$ act nontrivially), then $\mathcal A=Z(V_{MM})$.
The state-sum construction generalizes to the open-closed case
\cite{LaudaSS}. Let us describe it for a semisimple $\mathcal A$, and assuming
that the bordism $\Sigma$ only has a brane boundary. Each connected
component of $\partial\Sigma$ is then labeled by a brane $M\in\mathcal C$. We
pick a sufficiently large brane $M_0$ such that $\mathcal A=Z(V_{M_0 M_0})$. Let
$A=V_{M_0M_0}$. We also choose a basis $f^M_\mu$, $\mu\in S_M$ in each
module $M$. Denote the matrix elements of the action of $A$ on $M$ by
$T_M^{\mu}{}_{\nu i}$. We choose a triangulation of $\Sigma$, which also
gives us a triangulation of each connected component of the boundary.
2-simplices of $\Sigma$ are labeled as before. Label boundary 0-simplices
on any $M$-labeled boundary component by the basis vectors $f^M_\mu$. Thus
each boundary 1-simplex is labeled by a basis vector of $A$ and a pair of
basis vectors of a module. We assign a weight to each 2-simplex and each
interior 1-simplex before. We also assign a weight to each boundary
1-simplex as follows. Suppose the boundary 1-simplex is labeled by
$e_i\in A$ and $f^M_\mu,f^M_\nu\in M$. Then the weight of the boundary
1-simplex is $T_M^{\mu}{}_{\nu i}$. The total weight is the product of
weights of all 2-simplices and all 1-simplices (both interior and exterior).
Due to the introduction of brane boundaries, there are two more moves, called the 2-2 and 3-1 \emph{elementary shellings} and depicted in Figures \ref{fig:f3} and \ref{fig:f4}, that must be considered when demonstrating topological invariance \cite{LaudaSS}. They yield conditions
\begin{equation}\label{b2-2}T^\mu_{M\rho i} T^\rho_{M\nu j} =C^k_{ij} T^\mu_{M\nu k}\end{equation}
and
\begin{equation}\label{b3-1}T^\mu_{M\rho j}T^\rho_{M\nu k} C^{jk}_i =T^\mu_{M\nu i}\end{equation}
respectively. The first one is the definition of a module, and the second one follows from the semisimplicity of $A$. Therefore the state-sum is a well-defined open-closed TQFT. Moreover, such structures are precisely those required to define a topologically invariant state-sum.
\begin{figure}
\centering
\begin{tikzpicture}[scale =0.8]
\node at (0.5,-1.732/2) {\ldots};
\draw [line width =2pt] (0,0)--(3,0) circle (3pt) node[above] {$\mu$}-- node[below] {$i$} (4,-1.732) circle (3pt) node[below] {$\rho$}--node[below] {$j$} (5,0) circle (3pt) node[above] {$\nu$}--(8,0);
\draw (1,0)--(2,-1.732)--(3,0);
\draw (5,0)--(6,-1.732)--(7,0);
\draw(1,-1.732)--(7,-1.732);
\node at (7.5,-1.732/2) {\ldots};
\begin{scope}[shift={(9,-0.8)}]
\draw[<->] (0,0)--(1.5,0);
\end{scope}
\begin{scope} [shift={(12,0)}]
\node at (-0.5,-1.732/2) {\ldots};
\draw[line width=2pt] (0,0)--(1,0) circle (3pt) node[above] {$\mu$}--node[above] {$k$} (3,0) circle (3pt) node[above] {$\nu$}--(4,0);
\draw (0,-1.732)--(1,0)--node [below] {$i$} (2,-1.732)--node [below] {$j$} (3,0)--(4,-1.732);
\draw (-0.5,-1.732)--(4.5,-1.732);
\node at (5,-1.732/2) {\ldots};
\end{scope}
\end{tikzpicture}
\begin{tikzpicture}
\end{tikzpicture}
\caption{An elementary shelling representing $T^\mu_{\rho i} T^\rho_{\nu j} =C^k_{ij} T^\mu_{\nu k}$ \eqref{b2-2}. The thick line is a physical boundary.}
\label{fig:f3}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}
\draw[line width=2pt] (1,0)--(2,0) circle (3pt) node[above]{$\mu$}--node[above] {$j$} (3, 1.732) circle (3pt) node[above]{$\rho$}--node[above] {$k$} (4,0) circle(3pt) node[above]{$\nu$}-- (5,0);
\draw (2,0)--node[above] {$i$} (4,0);
\draw[<->] (6,0.5)--(7.5,0.5);
\begin{scope}[shift = {(9,0)}]
\draw [line width=2pt] (0,0)--(1,0) circle (3pt) node[above] {$\mu$}-- node[above] {$i$} (2,0) circle (3pt) node[above] {$\nu$}--(3,0);
\end{scope}
\end{tikzpicture}
\caption{An elementary shelling representing $T^\mu_{\rho j}T^\rho_{\nu k} C^{jk}_i = T^\mu_{\nu i}$ \eqref{b3-1}.}
\label{fig:f4}
\end{figure}
\subsection{Unitary TQFTs and semisimplicity}
The state-sum construction defines a perfectly good topological invariant for any finite-dimensional semisimple algebra $A$; however, if it is to model an actual physical system, its space of states must carry a Hilbert space structure, and linear maps corresponding to bordisms must be compatible in some sense with this structure. To be precise, for any oriented bordism $\Sigma$ whose source is a disjoint union of $n$ circles and whose target is a disjoint union of $l$ circles, let $-\Sigma$ denote its orientation-reversal. $-\Sigma$ has $l$ circles in its source and $n$ circles in its target. A 2d TQFT attaches to $\Sigma$ a linear map $\mathcal A^{\otimes n}\rightarrow \mathcal A^{\otimes l}$, and to $-\Sigma$ a linear map $\mathcal A^{\otimes l}\rightarrow \mathcal A^{\otimes n}$. A unitary structure on a 2d TQFT is a Hilbert space structure on $\mathcal A$ such that the maps corresponding to $\Sigma$ and $-\Sigma$ are adjoint to each other. For an open-closed 2D TQFT, we require that the state-space assigned to each boundary-colored interval has a non-degenerate Hermitian metric, and that cobordisms with nonempty brane boundary also satisfy the Hermiticity condition.
Let $\langle a,b\rangle$ denote the Hilbert space inner product of $a,b\in\mathcal A$. Since $\mathcal A$ also has a bilinear scalar product $\eta$, we can define an antilinear map
\begin{equation}\label{CPT}
*: \mathcal A\rightarrow\mathcal A, \quad a\mapsto a^*,
\end{equation}
such that $\langle a,b\rangle=\eta(a^*,b).$ It can be shown that this map is an involution (i.e. $a^{**}=a$) and an anti-automorphism (i.e.
$(ab)^*=b^*a^*$) \cite{HQFT}. This can also be expressed by saying that $\mathcal A$ is a $*$-algebra. Conversely, one can show that any commutative Frobenius $*$-algebra such that the sesquilinear product $\eta(a^*,b)$ is positive-definite gives rise to a unitary 2d TQFT \cite{HQFT}.
A corollary of this result is that for a unitary 2d TQFT the algebra $\mathcal A$ is semisimple. To see this, note first that any nonzero self-adjoint element $a$, $a=a^*$, cannot be nilpotent. Indeed, if $n$ is the smallest $n$ such that $a^n=0$, then $a^{2m}=0$, where $m=\lfloor(n+1)/2\rfloor$. Then $\left<a^m|a^m\right>=\left<1|a^{2m}|1\right>=0$, and therefore $a^m=0$. Since $n\leq m$, repeat with $n^\prime=m$ until $n=1$, i.e. $a=0$. Now we can use the result of \cite{lifeQM} which says that a $*$-algebra with no nilpotent self-adjoint elements (apart from zero) is semisimple.
By the Artin-Wedderburn theorem, a finite-dimensional semisimple algebra over complex numbers is isomorphic to a sum of matrix algebras. Since $\mathcal A$ is also commutative, this means that is isomorphic to a sum of several copies of ${\mathbb C}$. Frobenius and $*$-algebra structures exist and are unique up to isomorphism. This means that the only invariant of the 2d TQFT is the dimension of $\mathcal A$, i.e. the ground-state degeneracy of the corresponding phase.
As discussed above, for a semisimple algebra $\mathcal A$ boundary conditions correspond to finite-dimensional modules over $\mathcal A$. It is easy to see that for the open-closed TQFT to be unitary, the algebra $V_{MM}$ must also have a Hilbert space structure such that\begin{equation}\label{unitarity}T(a)^\dagger=T(a^*)\end{equation}Such a structure always exists and is unique. Thus a boundary condition for a unitary 2d TQFT can be simply identified with a module over $\mathcal A$. One can use any faithful module over $\mathcal A$ as an input for the state-sum construction.
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.7]
\draw[dashed] (0,0)--(1,1.732)--(2,0)--cycle;
\draw[line width=1pt] (-1,2)--(1,1.732/3)--(3,2); \draw[line width=1pt] (1,1.732/3)--(1,-2);
\end{tikzpicture}
\caption{The Poincare dual of a triangle}\label{fig:f5}
\end{figure}
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale = 0.4]
\draw (0,2)--(2,0)--(4,2);
\draw (2,0)--(2,-4);
\draw (0, -6)--(2,-4)--(4,-6);
\begin{scope}[shift={(5,0)}]
\draw [<->] (0, -2)--(1.5,-2);
\end{scope}
\begin{scope}[shift={(7,-2)}]
\draw (0,2)--(2,0)--(0,-2);
\draw (2,0)--(6,0);
\draw (8,2)--(6,0)--(8,-2);
\end{scope}
\end{tikzpicture}
\hspace{1.5cm} \begin{tikzpicture}[scale=0.6]
\draw (0,2)--(1.5,0)--(2.5,0)--(4,2);
\draw (1.5,0)--(2,-1.732/2)--(2.5,0);
\draw (2,-1.732/2)--(2,-2-1.732/2);
\begin{scope}[shift={(4, -0.5)}]
\draw[<->] (0,0)--(1.5,0);
\end{scope}
\begin{scope}[shift={(6,-0.5)}]
\draw (0,2)--(2,0)--(4,2);
\draw (2,0)--(2,-2);
\end{scope}
\end{tikzpicture}
\caption{The dual 2-2 (left) and 3-1 (right) moves} \label{fig:f6}
\end{figure}
\subsection{State-sum construction of the space of states}
We have discussed above the state-sum construction of the partition function $Z(\Sigma)$ for an oriented 2d manifold $\Sigma$ without boundary (or more generally, with only brane boundary). More generally, one also needs to describe in similar terms the state space $\mathcal A$ and a linear map $\mathcal A^{\otimes n}\rightarrow \mathcal A^{\otimes l}$ for every bordism $\Sigma$ whose source is a disjoint union of $n$ circles and target is a disjoint union of $l$ circles. That is, one needs to describe $Z(\Sigma)$ for the case when $\Sigma$ has nonempty cut boundary.
Consider a bordism $\Sigma$ with a nonempty cut boundary. For simplicity let us assume that there is no brane boundary; the general case is a trivial generalization, but requires a more cumbersome notation. We choose a triangulation ${\mathcal T}$ of $\Sigma$. It induces a triangulation of each boundary circle. We label the edges of 2-simplices with basis elements of $A$, as before. The only difference is that boundary 1-simplices have only one label rather than two. If we assign the weights to every 2-simplex and every internal 1-simplex as before and sum over the labelings of internal 1-simplices, we get a number $Z_{\mathcal T}(\Sigma)$ which depends on the labelings of the boundary 1-simplices. Suppose some boundary circle is divided into $N$ intervals. Then a labeling by $e_{i_1},\ldots, e_{i_N}$ corresponds to a vector
\begin{equation}
e_{i_1}\otimes \ldots\otimes e_{i_N}\in A^{\otimes N}.
\end{equation}
We can think of the number $Z_{{\mathcal T}}(\Sigma)$ computed by the state-sum as a matrix element of linear map from
\begin{equation}
A^{\otimes N_1}\otimes \ldots \otimes A^{\otimes N_n}
\end{equation}
to
\begin{equation}
A^{\otimes M_1}\otimes \ldots\otimes A^{\otimes M_l},
\end{equation}
where $N_1,\ldots,N_n$ denote the number of 1-simplices in the source circles, and $M_1,\ldots,M_l$ denote the number of 1-simplices in the target circles of $\Sigma$. It can be shown \cite{FHK} that the map $Z_{{\mathcal T}}(\Sigma)$ does not depend on the triangulation of $\Sigma$, provided we fix the triangulation of the boundary circles.
$Z_{\mathcal T}(\Sigma)$ is not yet the desired $Z(\Sigma)$ because it depends on the way the boundary circles are triangulated. To get rid of this dependence, we need to restrict this map to a certain subspace in each source factor $A^{\otimes N_i}$ and project to a certain subspace in each target factor $A^{\otimes M_j}$. Both tasks are accomplished by means of projectors $C_N: A^{\otimes N}\rightarrow A^{\otimes N}$. The projector $C_N$ is simply $Z_{{\mathcal T}_N}(C)$, where $C$ is a cylinder and ${\mathcal T}_N$ is any triangulation of $C$ such that both boundary circle are subdivided into $N$ intervals. The image of each $C_N$ is a certain subspace of $A^{\otimes N}$ isomorphic to $Z(A)$ \cite{FHK}. Restricting $Z_{\mathcal T}(\Sigma)$ to these subspaces and then projecting to the image of each $C_{M_j}$ gives us the desired map
\begin{equation}
Z(\Sigma): \mathcal A^{\otimes n}\rightarrow \mathcal A^{\otimes l},
\end{equation}
where $\mathcal A=Z(A)$.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw[line width=2pt] (0,0)--node [above] {$\mu$} (2,0) circle (2pt)--node[above] {$\nu$} (4,0);
\draw (2,0)--node[left] {$k$} (2,-1) circle (0.1)--node[left] {$i$} (1,-1-1.732/2);
\draw (2,-1)--node[left] {$j$} (3,-1-1.732/2);
\begin{scope}[shift={(5,-1)}]
\draw[<->] (0,0)--(0.75,0);
\end{scope}
\begin{scope}[shift={(7,0)}]
\draw[line width =2pt](0,0)--node[above]{$\mu$} (1,0) circle (2pt)--node[above] {$\rho$} (2,0) circle(2pt) --node[above]{$\nu$} (3,0) circle(2pt);
\draw (1,0)--node[left] {$i$} (1,-1); \draw(2,0)--node[left] {$j$} (2,-1);
\end{scope}
\end{tikzpicture}
\caption{The dual shelling of \eqref{b2-2}. A filled dot represents $T$, while an empty dot represents $C$.}
\label{fig;f7}
\end{figure}
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw [line width=2pt,fill] (0,0)--node[above] {$\mu$} (1,0) circle (3pt) -- node[above] {$\rho$} (2,0) circle (3pt) -- node[above] {$\nu$} (3,0);
\draw (1,0)--node[left] {$j$} (1.5,-1.732/2) circle (0.1)--node[right] {$k$} (2,0);
\draw(1.5, -1.732/2)--node [right] {$i$} (1.5,-1.732/2 -1);
\draw[<->] (4,-0.5)--(4.75,-0.5);
\begin{scope}[shift={(5.5,0)}]
\draw [line width=2pt,fill] (0,0)--node[above] {$\mu$} (1.5,0) circle (3pt) -- node[above] {$\nu$} (3,0);
\draw (1.5,0)--node[left] {$i$}(1.5,-1.5);
\end{scope}
\end{tikzpicture}
\caption{The dual shelling of \eqref{b3-1}, representing $T^{j \ }_{\mu \rho} T^{k \ }_{\rho \nu} C_{ijk} = T^{i \ }_{\mu \nu}$ } \label{fig:f8}
\end{figure}
\subsection{MPS from TQFT}
Let us consider the special case when $\Sigma$ is an annulus such that one of the circles is a cut boundary, while the other one is a brane boundary corresponding to an $A$-module $M$. Let $T(a)\in{\rm Hom}(M,M)$ represent an action of $a\in A$ in this module. For definiteness, we choose the cut boundary to be the source of $\Sigma$, while the target is empty. Thus $Z(\Sigma)$ is a linear map $\mathcal A\rightarrow {\mathbb C}$. It is the dual of the boundary state corresponding to the module $M$.
Let us now pick a triangulation of the annulus such that the cut boundary is divided into $N$ intervals. Then $Z_{\mathcal T}(\Sigma)$ is a linear map $A^{\otimes N}\rightarrow{\mathbb C}$ which depends only on ${\mathcal T}$ and $N$. We claim that this map is the dual of the MPS state with the dual MPS tensor given by $T: A\rightarrow {\rm Hom}(M,M)$.
To see this, it is convenient to reformulate the state-sum on the Poincare dual complex. This complex is built from the triangulation ${\mathcal T}(\Sigma)$ by replacing $k$-cells with $(2-k)$-cells, as in Figure \ref{fig:f5}. The dual of a triangulation is not a simplicial complex but a more general cell complex; since we will only be interested in the edges and vertices of this dual complex, we will refer to it as a \emph{skeleton} for $\Sigma$. The Pachner moves are the same for skeleton as for triangulations, see Figure \ref{fig:f6}. Recall that for a unitary TQFT, one can choose $\eta_{ij}=\delta_{ij}$, so that indices may be freely raised and lowered; nonetheless, keeping track of index positions now will pay off later when we generalize to equivariant theories. Choose a direction for each edge; the state-sum does not depend on this choice. Choose these directions so that all edges on incoming boundaries are incoming and all edges on outgoing boundaries are outgoing. To define a state-sum on a skeleton, label its non-boundary edges with elements $e_i$ and assign structure coefficients $C$ to each non-boundary vertex according to orientation and using lower indices for incoming arrows and upper for outgoing. With these conventions, the Pachner moves algebrize to \eqref{2-2} and \eqref{3-1} as before. To incorporate brane boundaries, color brane boundary edges by elements $v_\mu$ and attach the module tensor $T$ to each boundary vertex. The boundary moves recover \eqref{b2-2} and \eqref{b3-1}. The dual state-sum is naturally a tensor network: it defines a circuit between the incoming and outgoing legs. Note that the ``virtual'' module indices are all contracted, so these legs are physical.
Consider the triangulation, shown in Figure \ref{fig:f9a}, of the annulus with boundary condition $T$ on one of its boundary components. Its state-sum defines a state in the physical space $\mathcal A^N$. We claim that this state is the fixed point MPS $\ket{\psi_T}$. The proof of this fact is straightforward: by Pachner invariance, the annulus and MPS tensor networks are equivalent, see Figure \ref{fig:f9}.
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.5\linewidth}
\centering
\begin{tikzpicture}[scale=0.8
\draw[line width=4pt] (0,0)--(10,0);
\draw (1,0) -- (3,0) node[pos=0.5, above] {$T$};
\draw (1,0) -- (3,0) node[pos=1.5, above] {$T$};
\draw (1,0) -- (3,0) node[pos=2.5, above] {$T$};
\draw (1,0) -- (3,0) node[pos=3.5, above] {$T$};
\coordinate [label={above right:$C$}] (C) at (1.732,-1); \coordinate [label={above right:$C$}] (C) at (2.732,-1.27); \coordinate [label={above right:$C$}] (C) at (3.732,-1); \coordinate [label={above right:$C$}] (C) at (4.732,-1.27); \coordinate [label={above right:$C$}] (C) at (5.732,-1); \coordinate [label={above right:$C$}] (C) at (6.732,-1.27);
\coordinate [label={above right:$C$}] (C) at (7.732,-1);
\draw[line width=1pt] (1,0) -- (2,-1.732) {};
\draw[line width=1pt] (2,-1.732) -- (3,0) -- (4,-1.732) -- cycle {};
\begin{scope}[shift={(2,0)}]
\draw[line width=1pt] (2,-1.732) -- (3,0) -- (4,-1.732) -- cycle {};
\end{scope}
\begin{scope}[shift={(4,0)}]
\draw[line width=1pt] (2,-1.732) -- (3,0) -- (4,-1.732) -- cycle {};
\end{scope}
\draw[line width=1pt] (9,0) -- (8,-1.732) {};
\draw[line width=1pt] (0,-1.732) -- (10, -1.732) {};
\node at (-0, -1.732/2) {\ldots};
\node at (10,-1.732/2) {\ldots};
\end{tikzpicture}
\caption{Annulus with upper boundary colored by module $T$}\label{fig:f9a}
\end{subfigure}%
\begin{subfigure}[t]{0.5\linewidth}
\centering
\begin{tikzpicture}[scale=0.8]
\draw[line width=4pt] (0,0)--(2,0) circle (3pt) --(4,0) circle (3pt) --(6,0) circle(3pt) --(8,0);
\draw (2,0)--(2,-1) circle (3pt)--(1,-1-1.732/2) ;
\draw (4,0)--(4,-1) circle(3pt);
\draw(2,-1)--(3,-1-1.732/2) circle(3pt)--(4,-1);
\draw (4,-1)--(5,-1-1.732/2) circle (3pt)--(6,-1)--(6,0);
\draw(6,-1) circle(3pt)--(7,-1-1.732/2);
\draw(3,-1-1.732/2)--(3,-1.6-1.732/2);
\draw (5,-1-1.732/2)--(5,-1.6-1.732/2);
\node at (8,-1) {\ldots};
\node at (0,-1) {\ldots};
\end{tikzpicture}
\caption{The Poincare dual of \textbf{(a)}}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}
\centering
\begin{tikzpicture}[scale=0.7]
\draw [line width =4pt] (0,0)--(2,0) circle (3pt) -- (4,0) circle (3pt) -- (6,0) circle (3pt) --(8,0) circle (3pt)--(10,0);
\draw (2,0)--(1,-1.732);
\draw (4,0)--(5, -1.732) circle (0.15)-- (6,0); \draw (5,-1.732)--(5,-1-1.732);
\draw(8,0)--(9,-1.732);
\node at (10,-2) {\ldots};
\node at (0,-2) {\ldots};
\end{tikzpicture}
\caption{The move \eqref{b2-2} applied to \textbf{(b)}}
\end{subfigure}%
\begin{subfigure}[b]{0.5\linewidth}
\centering
\begin{tikzpicture}[scale=0.7]
\draw [line width =4pt] (0,0)--(1,0) circle (3pt) -- (5,0) circle (3pt) -- (9,0) circle (3pt) --(10,0);
\draw(1,0)--(1,-2);
\draw (5,0)--(5,-2);
\draw(9,0)--(9,-2);
\node at (10,-1) {\ldots};
\node at (0,-1) {\ldots};
\end{tikzpicture}
\caption{The move \eqref{b3-1} applied to \textbf{(c)}}
\end{subfigure}
\vfill
\caption{The equivalence of the annulus to the tensor network representation of an MPS} \label{fig:f9}
\end{figure}
More generally, one can insert a local observable on the brane boundary of the annulus. Such a local observable is parameterized by $X\in{\rm Hom}(M,M)$ which commutes with $T(a)$ for all $a\in A$. The corresponding dual state is $\tr[X^\dagger TT\cdots T]$, i.e. it is a generalized MPS state, with $A$ being the physical space.
Since the linear operators $T(a)$ satisfy $T(a)T(b)=T(ab)$, all these MPS states are RG-fixed MPS states. The RG-step is described by the algebra structure on $A$, $m: A\otimes A\rightarrow A$. Moreover, the MPS is automatically in a standard form. The module $T:A\rightarrow{\rm End}(V)$ is semisimple, so it has a decomposition into simple modules $T^{(\alpha)}:A\rightarrow{\rm End}(V^{(\alpha)})$. The collection of spaces ${\rm End}(V^{(\alpha)})$ form a block-diagonal subspace of ${\rm End}(V)$. Since $V^{(\alpha)}$ is simple, $T^{(\alpha)}$ surjects onto the block ${\rm End}(V^{(\alpha)})$.
The parent Hamiltonian of the MPS on an $N$-site closed chain has a TQFT interpretation as well: it is the linear map $C_N=Z_{{\mathcal T}_N}(C):A^{\otimes N}\rightarrow A^{\otimes N}$ assigned to a triangulated cylinder $C$ whose boundary consists of two circles triangulated into $N$ intervals. As previously stated, $C_N$ projects onto a subspace $\mathcal A=Z(A)\subset A^{\otimes N}$, precisely the space of ground states of the parent Hamiltonian. In the continuum TQFT, topological invariance requires that the cylinder is the identity; this is consistent with our already having projected to $\mathcal A$ in defining the continuum state spaces.
We have seen that a unitary TQFT is completely determined by its space of states $\mathcal A$ on a circle and that each finite-dimensional commutative algebra $\mathcal A$ defines a unitary TQFT. Therefore, the classification of unitary TQFTs is quite simple: there is one for every positive integer $n$, in agreement with the MPS-based classification of gapped phases \cite{ChenGuWenone,ChenGuWentwo,FidkowskiKitaev}.
\section{Equivariant TQFT and Equivariant MPS}
In this section, we generalize the relation between 2D TQFT and MPS states to systems with a global symmetry $G$. We show that both $G$-equivariant TQFTs and $G$-equivariant RG-fixed MPS states are described by semisimple $G$-equivariant algebras. In particular, we show that invertible $G$-equivariant TQFTs correspond to short-range entangled phases with symmetry $G$, and that both are classified by $H^2(G,U(1))$.
\subsection{$G$-equivariant Matrix Product States}
Let $G$ be a finite symmetry group acting on the physical space $A$ via a unitary representation $R$, $g\mapsto R(g)\in{\rm End}(A)$. A $G$-invariant MPS tensor is a map ${\mathcal P}: U\otimes U^*\rightarrow A$ equivariant in the following sense:
\begin{equation}
R(g){\mathcal P}(X)={\mathcal P}\left(Q(g) X Q(g^{-1})\right),
\end{equation}
where the linear maps $Q(g)\in{\rm End}(U)$ form a projective representation of $G$. Let $T={\mathcal P}^\dagger$. In terms of $T$, the equivariance condition looks as follows:
\begin{equation}
T(R(g)a)=Q(g) T(a) Q(g)^{-1},
\end{equation}
for any $a\in A$ and any $g\in G$.
The dual MPS state corresponding to $T$ is
\begin{equation}\label{equivMPS}
\langle \psi_T\vert=\sum_{i_1,\ldots,i_N}\Tr_U[T(e_{i_1})\ldots T(e_{i_N})] \langle i_1\ldots i_N\vert
\end{equation}
It is easy to see that the state $\psi_T$ is $G$-invariant, thanks to the equivariance condition on $P$. More generally, let $X\in{\rm End}(U)$. Note that ${\rm End}(U)$ is a genuine (not projective) representation of $G$. Then the generalized MPS state $\Tr[X TT\ldots T]$ transforms in the same way as $X$.
\subsection{$G$-equivariant TQFT}
Roughly speaking, a definition of a $G$-equivariant TQFT is obtained from the definition of an ordinary TQFT by replacing oriented manifolds with oriented manifolds with principal $G$-bundles. This reflects the intuition that a model with a global non-anomalous symmetry $G$ can be coupled to a background $G$ gauge field. (For a finite group $G$, there is no difference between a $G$ gauge field and a principal $G$-bundle.)
Some care is required regarding marked points and trivializations. Namely, each source and each target circle must be equipped with a marked point and a trivialization of the $G$-bundle at this point. This means that the holonomy of the gauge field around the circle is a well-defined element $g\in G$, rather than a conjugacy class. A $G$-equivariant TQFT associates a vector space $\mathcal A_g$ to a circle with holonomy $g$. A generic $G$-equivariant bordism has more than one marked point, and the holonomies between marked points along chosen paths are well-defined elements of $G$ as well. Of course, these holonomies depend only on the homotopy classes of paths. For example, a $G$-equivariant cylinder bordism has two marked points (one for each boundary circle) and depends on two arbitrary elements of $G$.
On the other hand, a $G$-equivariant torus, regarded as bordism with an empty source and empty target, has no marked points and depends on two commuting elements of $G$ defined up to an overall conjugation.
One can describe a $G$-equivariant TQFT purely algebraically in terms of a $G$-crossed Frobenius algebra \cite{HQFT,MS}. This notion generalizes the commutative Frobenius algebra $\mathcal A$ and encodes the linear maps $Z(\Sigma,{\mathcal P})$ in a fairly complicated way.
We will use instead a state-sum construction of 2D equivariant TQFTs which is manifestly local. Its starting point is a finite-dimensional semi-simple $G$-equivariant algebra $A$. This is an algebra with an action of $G$ which preserves both the scalar product $\eta$ and the multiplication map $m: A\otimes A\rightarrow A$. That is, $G$ acts on $A$ via a linear representation $R(g)$, $g\in G$, such that\begin{equation}\label{gauge1}m(R(g)a\otimes R(g) b)=R(g) m(a\otimes b)\end{equation}and
\begin{equation}\label{etaortho}
\eta(R(g)a,R(g)b)=\eta(a,b).
\end{equation}The condition \eqref{etaortho} says that $R(g)$ is orthogonal with respect to $\eta$. As a consequence, if $R(g)$ commutes with the anti-linear map \eqref{CPT}, it is unitary with respect to the Hilbert space inner product.
A large class of examples of $G$-equivariant algebras is obtained by taking $A={\rm End}(U)$, where $U$ is a vector space, and $G$ acts on $U$ via a projective representation $Q(g)$. It is clear that this gives rise to a genuine action of $G$ on ${\rm End}(U)$ which preserves the usual matrix multiplication on ${\rm End}(U)$. Moreover, the standard Frobenius structure
\begin{equation}
\eta(a,b)={\rm Tr}(ab)
\end{equation}
is clearly $G$-invariant.
A $G$-equivariant module over a $G$-equivariant algebra $A$ is a vector space $V$ with compatible actions of both $A$ and $G$. That is, for every $a\in A$ we have a linear map $T(a):V\rightarrow V$ such that $T(a)T(a')=T(aa')$, and for every $g\in G$ we have an invertible linear map ${\mathcal Q}(g):V\rightarrow V$ such that ${\mathcal Q}(g){\mathcal Q}(g')= {\mathcal Q}(gg')$. The compatibility condition that they satisfy reads\begin{equation}\label{gauge2}
T(R(g)a)={\mathcal Q}(g)T(a){\mathcal Q}(g)^{-1}\end{equation}
If we take $A={\rm End}(U)$, where $U$ is a projective representation of $G$ with a 2-cocycle $\omega\in H^2(G,U(1))$, then $U$ is not a $G$-equivariant module over $A$ unless $\omega$ vanishes. However, if $W$ is a projective representation of $G$ with a 2-cocycle $-\omega$, then $U\otimes W$ is a $G$-equivariant module.\footnote{In fact, the category of projective representations of $G$ with a 2-cocycle $-\omega$ is equivalent to the category of $G$-equivariant modules over ${\rm End}(U)$, and the equivalence sends a projective representation $W$ to $U\otimes W$.}
Equivariant TQFTs admit a lattice description as well. It is simplest to describe a Poincare dual formulation in the sense of Section 3.5; spaces in this formulation also have direct interpretations as tensor networks. A trivialized background gauge field is represented on a skeleton as a decoration of each oriented edge with an element $g\in G$. Flipping the orientation of the edge replaces $g$ with $g^{-1}$. We require that the field is flat: that the product of the group elements around the boundary of each face is the identity element.\footnote{In the triangulation picture, we require the product of all group elements corresponding to edges entering a particular vertex to be the identity element.} In a basis $e_i$, $i\in S$ of $A$, the weight of a coloring of the skeleton is the product of the structure constants $C^{ijk}$ over vertices (with the cyclic order given by the orientation) and a factor $\eta(R(g)e_i,e_j)=R(g)^k{}_i\eta_{jk}$ for each edge directed from $i$ to $j$ labeled by $g$. The partition sum is the sum of these weights over all colorings; we emphasize that the group labels represent a background gauge field and are not summed. To incorporate brane boundaries, choose a $G$-equivariant module $V$ over $A$. Fix a basis $f_\mu$ of $V$. For each brane boundary vertex, label its adjacent boundary edges each with a basis element, so that each boundary edge has a total of two labels. The weight of a skeleton with a brane boundary is a product of $C$'s and $R$'s as well as a module tensor $T$ for each brane boundary vertex and a matrix element ${\mathcal Q}(g)^{\mu}_{\ \nu}$ for each brane boundary edge.
As before, topological invariance of the state-sum amounts to checking the conditions \eqref{2-2}, \eqref{3-1}, \eqref{b2-2}, and \eqref{b3-1}. These are satisfied by any finite-dimensional semisimple $A$. In order for the equivariant state-sum to constitute a well-defined equivariant TQFT, it must also be independent of the choice of trivialization of the background gauge field; in order words, it must be gauge invariant. A gauge transformation by $h\in G$ on a vertex acts by changing the decorations of the three edges whose boundary contains the vertex: incoming edges with $g$ become $hg$, outgoing $gh^{-1}$, as in Figure \ref{fig:gauge}. Invariance under a gauge transformation on a vertex in the interior is ensured by axioms \eqref{gauge1} and \eqref{etaortho} of a $G$-equivariant algebra. For vertices in the brane boundary, the analogous result follows from the $G$-equivariant module condition \eqref{gauge2}.\footnote{Here it is crucial that linear transformations ${\mathcal Q}(g)$ form an ordinary (i.e. not projective) representation of $G$.} Finally, invariance under simultaneously reversing an edge direction and inverting its group label is enforced by the axiom \eqref{etaortho}.
\begin{figure}
\centering
\begin{tikzpicture}
\draw[->] (-1,1)--(0,0.5) node[above] {$g_2$};
\draw (0,0.5)--(1,0);
\draw[->] (3,1)--(2,0.5) node[above] {$g_1$} ;
\draw (2,0.5)--(1,0);
\draw[->] (1,0)--(1,-1) node[right] {$g_3$};
\draw (1,-1)--(1,-2);
\draw[->] (3,-0.3)--(4,-0.3);
\begin{scope} [shift = {(5,0)}]
\draw[->] (-1,1)--(0,0.5) node[above] {$hg_2$};
\draw (0,0.5)--(1,0);
\draw[->] (3,1)--(2,0.5) node[above] {$hg_1$} ;
\draw (2,0.5)--(1,0);
\draw[->] (1,0)--(1,-1) node[right] {$g_3h^{-1}$};
\draw (1,-1)--(1,-2);
\end{scope}
\end{tikzpicture}
\caption{A lattice gauge transformation at the vertex by $h$}
\label{fig:gauge}
\end{figure}
\subsection{$G$-equivariant semisimple algebras}
The classic Wedderburn theorem implies that every finite-dimensional semisimple algebra is a sum of matrix algebras. Let us discuss a generalization of this result to the $G$-equivariant case following \cite{Ostrik,Etingof}.
First, we can write every $G$-equivariant semisimple algebra as a sum of indecomposable ones, so it is sufficient to classify indecomposable $G$-equivariant semisimple algebras. A large class of examples is given by algebras of the form ${\rm End}(U)$, where $U$ is a projective representation of $G$. Another set of examples is obtained as follows: let $H\subset G$ be a subgroup. Consider the space of complex-valued functions on $G$ invariant with respect to left translations by $H$, i.e. $f(h^{-1}g)=f(g)$ for all $g\in G$ and all $h\in H$. The group $G$ acts on this space by right translations:
\begin{equation}
(R(g)f)(g')=f(g'g)
\end{equation}
Pointwise multiplication makes this space of functions into an associative algebra, and it is clear that the $G$-action commutes with the multiplication. This $G$-equivariant algebra is indecomposable for any $H$.
The most general indecomposable $G$-equivariant semisimple algebra is a combination of these two constructions called the induced representation $\text{Ind}_H^G{\rm End}(U)$ \cite{Ostrik,Etingof}. One picks a subgroup $H\subset G$ and a projective representation $(U,Q)$ of $H$. Here $U$ is a vector space and $Q$ is a map $H\rightarrow{\rm End}(U)$ defining a projective action with a 2-cocycle $\omega\in H^2(H,U(1))$. Then one considers the space of functions on $G$ with values in ${\rm End}(U)$ which have the following transformation property under the left $H$ action:
\begin{equation}
f(h^{-1}g)=Q(h) f(g) Q(h)^{-1}
\end{equation}
It is easy to check that the right $G$ translations act on this space of functions. Pointwise multiplication makes this space into a $G$-equivariant algebra, and one can show that it is indecomposable. To summarize, indecomposable $G$-equivariant semisimple algebras are labeled by triples $(H,U,Q)$, where $H\subset G$ is a subgroup, and $(U,Q)$ is a projective representation of $H$. All these algebras are actually Frobenius algebras: the trace function $A\rightarrow{\mathbb C}$ is given by
\begin{equation}
\sum_{g\in G} {\rm Tr}_U f(g)
\end{equation}
A $G$-equivariant module over such an algebra $A$ is obtained as follows. Start with an $H$-equivariant module $(M,{\mathcal Q})$ over ${\rm End}(U)$. Here $M$ is a module over ${\rm End}(U)$ and ${\mathcal Q}: H\rightarrow{\rm End}(M)$ is a compatible action of $H$ on $M$. As explained above, $M$ must have the form $U\otimes W$, where
$W$ carries a projective action $S(h)$ of $H$ with a 2-cocycle $-\omega$.
Then consider functions on $G$ with values in $M$ which transform as follows under the left $H$-translations:
\begin{equation}
m(h^{-1}g)=(Q(h)\otimes S(h)) m(g),\quad m:G\rightarrow U\otimes W.
\end{equation}
The group $G$ acts on this space by right translations, and it is easy to see that the pointwise action of $A=(H,U,Q)$ makes it into a $G$-equivariant module over $A$. One can show that any $G$-equivariant module over such an $A$ is a direct sum of modules of this sort.
\subsection{$G$-equivariant MPS from $G$-equivariant TQFT}
It is sufficient to consider indecomposable TQFTs and $G$-equivariant algebras. Let us begin with the case $H=G$. Then the algebra $A=(G,U,Q)$ is isomorphic to the algebra ${\rm End}(U)$, and a $G$-equivariant module over it is simply a vector space $M$ with a $G$-equivariant action of ${\rm End}(U)$. In other words, $M=U\otimes W$, where $U$ carries a projective representation of $G$ with the 2-cocycle $\omega$, and $W$ carries a projective representation of $G$ with a 2-cocycle $-\omega$.
Consider an annulus whose outer boundary is labeled by a brane $M$ and whose inner boundary is a cut boundary. Let us triangulate both boundary circles into $N$ intervals. Let $g_{i,i+1}$ be the element of $G$ labeling the interval between the $i^{\rm th}$ and $(i+1)^{\rm th}$ points on the boundary. We also assume that the holonomy of the gauge field between the points labeled by $1$ on the two boundary circles is trivial. We get the the following dual state:
\begin{equation}
\label{Gdec}\langle\psi_T\vert=\sum \Tr_{U\otimes W}[T(e_{i_1}){\mathcal Q}(g_{12}) \cdots T(e_{i_N}) {\mathcal Q}(g_{N 1})] \langle i_1 \cdots i_N \vert
\end{equation}
Note that although $T(e_i)$ is an operator on $U\otimes W$, it has the form $T(e_i)\otimes {\bf 1}_W$. Therefore, if $g_{i,i+1}=1$ for all $i$, the trace over $W$ gives an overall factor $\dim W$, and up to this factor we get the equivariant MPS (\ref{equivMPS}). Inserting an observable $X\in{\rm End}(U)$ on the brane boundary, we get a generalized equivariant MPS. The case when $X\in{\rm End}(U\otimes W)$ does not give anything new, since the trace over $V$ factors out.
The generalized equivariant MPS (cf. eq \ref{GS})
\begin{equation}
\langle \psi^X_T | = \sum \Tr[X^{\dagger} T(e_{i_1}) \cdots T(e_{i_n})] \langle i_1 \cdots i_n |
\end{equation}
may be charged under the action of $h\in G$:
\begin{align}
R(h)^{\otimes N}\bra{\psi_T^X} = \sum \Tr[X^{\dagger} T(e_{i_1}) \cdots T(e_{i_n})] \langle (h^{-1} \cdot i_1) \cdots (h^{-1} \cdot i_n) | \nonumber \\
= \sum \Tr[X^{\dagger} T(h \cdot e_{i_1}) \cdots T(h \cdot e_{i_n})] \langle i_1 \cdots i_n | \nonumber \\
= \sum \Tr[Q(h^{-1}) X^{\dagger} Q(h) T(e_{i_1}) \cdots T(e_{i_n}) ] \langle i_1 \cdots i_n |
\end{align}
Let us now consider the case when $H$ is a proper subgroup of $G$ and $A=\text{Ind}_H^G{\rm End}(U)$, for some projective representation $U$ of $H$. If we choose right $H$-coset representatives $g_a$, $a\in H\backslash G$, and a basis $e_i$ in ${\rm End}(U)$, then a basis in $A$ is given by $e_i^a$. Similarly, if $f_\mu$ is a basis in an $H$-equivariant module $U\otimes W$, then a basis in the corresponding $G$-equivariant module $M$ is $f^a_\mu$.
The action of $A$ on $M$ is diagonal as far as the $a$ index is concerned. Therefore the dual state corresponding to a triangulated annulus with $g_{i,i+1}=1$ for all $i$ vanishes unless all $a$ indices are the same. Then
\begin{equation}
\langle\psi_T\vert =\dim (W) \sum_{a,i_1,\ldots,i_N} \Tr_U [T(e_{i_1}) \cdots T(e_{i_N})] \langle i_1 a\, i_2 a\,\cdots i_N a \vert.
\end{equation}
This state has equal components along all $|H\backslash G|$ directions. We can get a state concentrated at a particular value of $a$ by inserting a suitable observable $X\in{\rm End}(M)$ on the brane boundary. Such an observable must commute with the action of $A$, so it must have the form $X^\mu_\nu { }^a_b= f(a)\delta^\mu_\nu\delta^a_b.$ Choosing the function $f(a)$ to be supported at a particular value of $a$ gives a generalized MPS state supported at this value of $a$.
The symmetry group $G$ acts transitively on $H\backslash G$. This suggests that we are dealing with a phase where the symmetry $G$ is spontaneously broken down to $H$, so that we get $|H\backslash G|$ sectors labeled by the index $a$. To confirm this, consider the partition function of this TQFT on a closed oriented 2-manifold $\Sigma$ with a trivial $G$-bundle. After we choose a skeleton of $\Sigma$, we can represent this $G$-bundle by labeling every 1-simplex with the identity element of $G$. In addition, every 1-simplex is labeled by a pair of basis vectors of $A$. Since both the multiplication in the algebra $A$ and the scalar product are pointwise in $H\backslash G$, the partition function receives contributions only from those labelings where all $a$ labels are the same. Furthermore, turning on a gauge field which takes values in $H$ does not destroy this property. We conclude that the theory has superselection sectors labeled by elements of $H\backslash G$, and each sector has unbroken symmetry $H$.
\subsection{Twisted-sector states}
Now let us not assume that $g_{i,i+1}=1$, but instead allow the gauge field around the circle to have a nontrivial holonomy. Let us take $H=G$ first, i.e. the case of unbroken symmetry. Consider the MPS \eqref{Gdec}. Applying a gauge transformations (by $g_{k-1,k}\cdots g_{2,1}$ at vertex $k$) to the boundary vertices it can be written as
\begin{equation} \langle\psi_{T,g}\vert=\sum \Tr_{U\otimes W}[{\mathcal Q}(g)T(e_{i_1}) \cdots T(e_{i_N})] \bigotimes_{k=1}^N R(g_{(k-1),k}\cdots g_{2,1})^{i_k}{}_{j_k}\langle j_k \vert
\end{equation}
where $g=g_{N,1}g_{(N-1),N}\cdots g_{1,2}$ is the holonomy of the gauge field. This is LU equivalent to the state
\begin{equation} \langle\psi_{T,g}\vert=\sum \Tr_{U\otimes W}[{\mathcal Q}(g)T(e_{i_1}) \cdots T(e_{i_N})] \langle i_1 \cdots i_N \vert
\end{equation}
so we have effectively set $g_{i,i+1}=1$ for all $i\ne N$ and $g_{N,1}=g$. Note that ${\mathcal Q} = Q \otimes S$, so the trace factors into a product of a trace over $U$ and a trace over $W$. The latter gives us an overall factor, and we have
\begin{equation}
\langle\psi_{T,g}\vert =\Tr_W [S(g)] \sum \Tr_{U}[Q(g_{N1})T(e_{i_1}) \cdots T(e_{i_N})] \langle i_1 \cdots i_N \vert.
\end{equation}
This state transforms under $h \in G$ into
\begin{align}
R(h)^{\otimes N}\bra{\psi_{T,g}} = (\Tr_W[S(g)]) \sum \Tr[Q(g)T(e_{i_1}) \cdots T(e_{i_n})] \langle (h^{-1} \cdot i_1) \cdots (h^{-1} \cdot i_n) | \nonumber \\
= (\Tr_W[S(g)])\sum \Tr[Q(h)^{ -1} Q(g) Q(h) T(e_{i_1}) \cdots T(e_{i_n})] \langle i_1 \cdots i_n | \nonumber \\
=(\Tr_W[S(g)]) \omega(g,h) \omega(h^{-1}, gh) \sum \Tr[ Q(h^{-1}gh)T(e_{i_1}) \cdots T(e_{i_n})] \langle i_1 \cdots i_n | \end{align}
Note that the $g$-twisted sector becomes the $hgh^{-1}$-twisted sector.
Now suppose $H$ is a proper subgroup of $G$. Since $T$ acts pointwise in the $a$ label, while $G$ acts on $a\in H\backslash G$ by right translations, the annulus state vanishes unless the holonomy around the circle is in $H$. This confirms once again that $H$ is the unbroken subgroup. Indeed, when the holonomy does not belong to the unbroken subgroup, there must be a domain wall somewhere on the circle. Its energy is nonzero in the thermodynamic limit, so the TQFT space of states must be zero-dimensional for holonomies not in $H$.
If $\mathcal A_g$ denotes the space of states in the $g$-twisted sector, the space $\mathcal A=\oplus_g\mathcal A_g$ has an automorphism $\alpha_h:=R(h)^{\otimes N}$ for each $h\in G$ such that $\alpha_h(\mathcal A_g)\subset\mathcal A_{hgh^{-1}}$. $\mathcal A$ is the $G$-graded vector space underlying the $G$-crossed Frobenius algebra that defines the associated $G$-equivariant TQFT \cite{HQFT,MS}.
\subsection{Morita equivalence}
We have seen that to any semisimple $G$-equivariant algebra one can associate a $G$-equivariant 2d TQFT. But different algebras may give rise to the same TQFT. In particular, we would like to argue that the TQFT corresponding to an indecomposable algebra $A=(H,U,Q)$, where $(U,Q)$ is a projective representation of $H$, depends only on the subgroup $H$ and the 2-cocycle $\omega$, but not on the specific choice of $(U,Q)$.
To show this, note first of all that the partition function vanishes if the holonomy does not lie in $H$ (this again follows from the fact that multiplication in the algebra $A$ is pointwise with respect to the $a$ index). Thus it is sufficient to consider oriented 2-manifolds with $H$-bundles. Further, if $U$ and $U'$ are projective representations of $H$ with the same 2-cocycle, then $U'=U\otimes W$, where $W$ is an ordinary representation of $H$. Thus we only need to show that the partition functions corresponding to algebras $(H,U,Q)$ and $(H,U\otimes W,Q\otimes S)$ are the same, where $S:H\rightarrow{\rm End}(W)$ is a representation of $H$. But it is clear from the state sum construction that the two partition functions differ by a factor which is the partition function of two dimensional $H$-equivariant TQFT corresponding to the algebra $(H,W,S)$.
We reduced the problem to showing that the $H$-equivariant TQFT constructed from the algebra $(H,W,S)$ is trivial when $(W,S)$ is an ordinary (not projective) representation of $H$. This is straightforward: the equation $S(h_1)\ldots S(h_n)=S(h_1\ldots h_n)$ and the flatness condition for the $H$ gauge field imply that the partition function is independent of the $H$-bundle, and for the trivial $H$-bundle the partition function is the same as for the trivial TQFT with $A={\mathbb C}$.
From the mathematical viewpoint, $G$-equivariant algebras with the same $H$ and $\omega$ are {\it Morita-equivalent}\footnote{More accurately, algebras with the same $H$ and $\omega$, \emph{up to conjugation in} $G$, are Morita-equivalent. In physical contexts, however, it is typical to keep track of the embedding of the unbroken symmetry $H$ in the full symmetry group $G$. Therefore, the classification of physical gapped $G$-symmetric phases is slightly more refined than that of Morita classes.} \cite{Ostrik}. Thus we have shown that Morita-equivalent algebras lead to identical $G$-equivariant TQFTs.\footnote{Strictly speaking, we only showed this for closed 2d TQFTs, but the argument easily extends to the open-closed case.}
\subsection{Stacking phases}
Consider two gapped systems built from algebras $A_1$ and $A_2$. Recall from Section \ref{sec:mps} that the stacked system \eqref{stackedham} is built from the tensor product algebra $A_1\otimes A_2$. Although we have not discussed parent Hamiltonians of $G$-equivariant MPS, an analogous stacking operation can be defined for $G$-symmetric gapped phases by way of the connection to TQFT. Now suppose $A_1$ and $A_2$ are $G$-equivariant algebras. It is clear from the $G$-equivariant state sum construction that the partition functions for the algebra $A_1\otimes A_2$ are products of those for $A_1$ and $A_2$ and that the Hilbert spaces are tensor products. Thus the MPS ground states, which determine a phase and which are realized in TQFT, stack like the tensor product of $G$-equivariant algebras.
It is a tedious but straightforward exercise to check that the result of stacking the phase labeled by subgroup-cocycle pair $(H,\omega)$ with the phase $(K,\rho)$ is the phase\begin{equation}\label{HK}(H\cap K,\omega|_{H\cap K}+\rho|_{H\cap K})^{\oplus[G:HK]}\end{equation}where $\omega|_{H\cap K}$ denotes the restriction of $\omega$ to the intersection subgroup $H\cap K$ and $[G:HK]$ denotes the index of the subgroup $HK$ in $G$, assuming $H$ and $K$ are normal in $G$.
Let us consider a simple example: take $G={\mathbb Z}_2\times{\mathbb Z}_2=\langle a,b\rangle$, where $a$ and $b$ are commuting elements of order $2$. For the subgroup $H=G$, there are two cohomology classes $\omega\in H^2({\mathbb Z}_2\times{\mathbb Z}_2,U(1))$. Let $\omega_1$ denote the nontrivial class. For each of the other subgroups $H=\langle a\rangle$, $\langle b\rangle$, $\langle ab\rangle$, $1$, there is a unique cocycle. Thus the classification of ${\mathbb Z}_2\times{\mathbb Z}_2$-equivariant phases looks like Figure \ref{fig:classif}.
\begin{figure}[h]
\centering
\begin{tabular}{ c | c | c }
$(H,\omega)$ & type of phase & name \\
\hline
$(\langle a,b\rangle,1)$ & trivial & 1 \\
$(\langle a,b\rangle,\omega_1)$ & symmetry-protected & $\omega$ \\
$(\langle a\rangle,1)$ & broken symmetry & A \\
$(\langle b\rangle,1)$ & broken symmetry & B \\
$(\langle ab\rangle,1)$ & broken symmetry & C \\
$(1,1)$ & broken symmetry & 0 \\
\hline
\end{tabular}
\caption{Indecomposable phase classification for the $G={\mathbb Z}_2\times{\mathbb Z}_2$}
\label{fig:classif}
\end{figure}
According to \eqref{HK}, the stacking rules are
$$1\otimes 1=1,\quad 1\otimes\omega=\omega,\quad 1\otimes A=A,\quad 1\otimes B=B,\quad 1\otimes C=C,\quad 1\otimes 0=0$$$$\omega\otimes\omega=1,\quad\omega\otimes A=A,\quad\omega\otimes B=B,\quad\omega\otimes C=C,\quad\omega\otimes 0=0$$$$A\otimes A=A^{\oplus 2},\quad B\otimes B=B^{\oplus 2},\quad C\otimes C=C^{\oplus 2},\quad A\otimes B=0,\quad B\otimes C=0,\quad C\otimes A=0$$$$A\otimes 0=0^{\oplus 2},\quad B\otimes 0=0^{\oplus 2},\quad C\otimes 0=0^{\oplus 2},\quad 0\otimes 0=0^{\oplus 4}$$
\subsection{Symmetry Protected Topological Phases}
Finally, let us discuss the case of Short-Range Entangled (SRE) phases with symmetry $G$. According to one definition \cite{Kitaevtalk}, an SRE phase is one that is invertible under the aforementioned stacking operation. Such phases have a one-dimensional space of ground states for every $G$-bundle on a circle.
Since the space of states of a decomposable TQFT on a circle with a trivial bundle has dimension greater than one, a TQFT corresponding to an SRE phase must be indecomposable. We showed that when $H$ is a subgroup of $G$, the space of states is zero-dimensional whenever the holonomy does not lie in $H$. Hence an equivariant TQFT built from an indecomposable $G$-equivariant algebra $(H,U,Q)$ cannot correspond to an SRE unless $H=G$.
These SRE phases are all Symmetry Protected Topological (SPT) phases - phases that are trivial if we ignore symmetry. A $G$-equivariant algebra of the form ${\rm End}(U)$, where $U$ is a projective representation of $G$, is simply a matrix algebra if we ignore the $G$ action. Hence the corresponding non-equivariant TQFT is trivial; the corresponding Hamiltonian is connected to the trivial one by a Local Unitary transformation. Hence SPT phases with symmetry $G$ are labeled by 2-cocycles $\omega\in H^2(G,U(1))$. This is a well-known result \cite{ChenGuWenone,ChenGuWentwo,FidkowskiKitaev}.
|
\section{\label{sec:intro}Introduction}
\textit{Introduction--}
In the recent few years, there has been considerable interest in the search for novel degeneracies associated with nonsymmorphic symmetries in electronic structures\cite{2016arXiv160303093B,Wang2016,kanedouble,muechler2016tilted,schoop2016dirac}. Since these degeneracies usually have topological origins, their study is a natural extension of the larger program of discovering new topological phases
, both theoretically and experimentally\cite{RevModPhys.83.1057, RevModPhys.82.3045, moore2010birth, chen2009experimental, zhang2009topological,RevModPhys.83.1057, RevModPhys.82.3045, moore2010birth, chen2009experimental, zhang2009topological,yu2010quantized, chang2013experimental, PhysRevLett.95.146802, bernevig2006quantum,PhysRevLett.50.1395, PhysRevLett.48.1559, Sun2011NoLandau, Xu2016, He2016,lee2017band}.
Such phases, which are protected by symmetry and/or a nontrivial topological index, possess interesting physically manifestations like boundary states, quantized response or exotic quasi-particle excitations. While a large number of topological phases have been theoretically identified and classified for different symmetry classes and dimensions\cite{Schnyder2008, Kitaev:2009mg, MHua2016}, only a handful have been experimentally realized in electronic systems. This is fundamentally due to the limited tunability of the Fermi level and atomic configurations.
Hence the push towards the realization of topological phases in alternative, artificial systems like photonic\cite{haldane2008possible, PhysRevA.78.033834, wang2008reflection, Dong2016,soskin2016singular,goryachev2016reconfigurable}, phononic\cite{salerno2014dynamical,nash2015topological,wang2015topological,susstrunk2015observation,yang2015topological,zhu2015topologically,fleury2015floquet,paulose2015selective,ong2016transport,huber2016topological,liu2016topological,lu2016observation,lee2017dynamically} and cold atom\cite{sun2012topological,furukawa2015excitation,qin2016topological} systems both in 2D and 3D\cite{Parameswaran2013, PhysRevB.90.085304, PhysRevB.93.045429, Lu:15, lu2016symmetry, 2016arXiv160701862K}, where topological invariants can be defined in analogy to those in conventional electronic systems. Photonic systems are particularly convenient for probing novel topological physics\cite{PhysRevLett.110.076403, xiao2015,2016arXiv160702918C,He03052016} due to their exactly solvable governing equations and lack of fundamental length scale\cite{joannopoulos2011photonic}. Indeed, topological phases have been discovered in various photonic systems with different symmorphic symmetries. Nontrivial edge modes have been observed in two-dimensional (2D) photonic crystals with $C_4$\cite{WenXiao-Gang}, $C_6$\cite{PhysRevLett.114.223901} or mirror symmetry\cite{chen2014experimental}, and topological Weyl points and nodel lines have been found in three-dimensional (3D) photonic crystals with gyroid structures or screw symmetry\cite{lu2013weyl,type2}. One important
advantage of photonic crystals is that they can contain features of any desired shape, i.e. an ellipsoid, which is impossible to realize in electronic systems. This additional freedom shall play a crucial role in our implementation of 2D nonsymmorphic symmetry groups.
Motivated by the richness of nonsymmorphic symmetry, we present specially designed 2D photonic crystals symmetric under the four nonsymmorphic wallpaper groups $pg$, $pmg$, $p4g$ and $pgg$. Compared to previous proposals involving strongly spin-orbit coupled systems\cite{kane2015dirac}, our lattice structures are extremely simple and amenable to experimental realization, consisting of only two inequivalent elliptical dielectric structures per unit cell.
Various combinations of topologically robust Dirac points (DPs) and gapless line nodes exist depending on the nonsymmorphic symmetry group.
Tuning our photonic crystals while preserving nonsymmorphic symmetry gives rise to various phenomena with potential technological applications. With $pg$ symmetry preserved, we show that a Lifshitz transition to a type-II DP occurs across a large range of realistic photonic rod shapes and dielectric constants. Analogous Lifshitz transitions have attracted considerable interest in the Weyl semimetal community\cite{soluyanov2015type,muechler2016tilted, wang2016mote,deng2016experimental,koepernik2016tairte}, and in our context leads to the phenomenon of anomalous refraction where an incident ray produces not one but two refracted rays. The sensitivity of the photonic bandstructure to the lattice symmetry brings forth the possibility of optical devices with mechanically-induced properties, as detailed in our stoplight device proposal.
Beginning with a pedagogical justification of the appearance of line degeneracies in a tight-binding (TB) Hamiltonian with $pg$ symmetry, we then extend the discussion to point degeneracies (Dirac points) emerging in higher orbitals. With the help of homotopy arguments, we analyze their topological properties, as well as numerically demonstrate their robustness through a protected $Z_2$ topological number. Finally, we detail the occurrence of Lifshitz transitions to type-II DPs, which are physically manifested through anomalous refraction. We conclude by proposing a stoplight device based on the C4-symmetry protected doubled DP.
\textit{TB model construction--}
A crystal with nonsymmorphic symmetry maps into itself under a combination of a point symmetry operation (i.e. reflection) and a fractional unit cell translation.
Bandstructure degeneracies appear due to the existence of higher-dimensional projective representations of the nonsymmorphic symmetry group at certain momenta.
To understand the effect of nonsymmorphic symmetry, we first introduce the TB description of a photonic system\cite{albert2002photonic}.
We focus on the case where waves propagate in an electric field parallel to the rod axis, forming the so-called harmonic transverse magnetic (TM) modes. The $n$-th mode $\vec E_{n,\mathbf{k}}=E_{n,\mathbf{k}}(\mathbf{r})\hat{z}$ obeys Maxwell's equation
\begin{equation}
\nabla^{2}\vec E_{n,\mathbf{k}}(\mathbf{r})=-\frac{\omega^2_{n,\mathbf{k}}}{c^2}\varepsilon_p(\mathbf{r})\vec E_{n,\mathbf{k}}(\mathbf{r})
\end{equation}
where $\varepsilon_p(\mathbf{r})$ is the dielectric function of the periodic medium, and $\omega^2_{n,\mathbf{k}}$ is its frequency. If we rescale the modes via $\phi_{n,\mathbf{k}}(\mathbf{r})=\sqrt{\varepsilon_p(\mathbf{r})}E_{n,\mathbf{k}}(\mathbf{r})$, the eigenvalues of the Hermitian operator $H=-\frac{1}{\sqrt{\varepsilon_p(\mathbf{r})}}\nabla^{2}\frac{1}{\sqrt{\varepsilon_p(\mathbf{r})}}$ form
the effective band structure viz.
\begin{equation}
H\phi_{n,\mathbf{k}}(\mathbf{r})=\frac{\omega^2_{n,\mathbf{k}}}{c^2}\phi_{n,\mathbf{k}}(\mathbf{r}).
\label{TB0}
\end{equation}
This equation is the direct analog of the TB Schr\"{o}dinger's equation of an electronic system, where $\phi_{n,\mathbf{k}}$ represent its Bloch states. The hopping terms of this photonic TB Hamiltonian can be determined from the overlaps of the single dielectric rod eigenstates, which are well-localized like the orbitals of a single atom. Solving Eq.~\eqref{TB0}, we obtain the photonic band structure $\omega^2_{n,\mathbf{k}}/c^2$.
A nonsymmorphic crystal has at least two different components i.e. ``atoms'' in its unit cell, which are separated
by a non-primitive lattice vector.
The effective TB Hamiltonian can be expressed in the normalized basis
\begin{equation}
\psi_{\alpha,\nu,\mathbf{k}}(\mathbf{r}) =
\sum_{\mathbf{R}}e^{i\mathbf{k}\cdot (\mathbf{R}+\mathbf{r}_{\alpha})}
\varphi_\nu(\mathbf{r} -\mathbf{R} - \mathbf{r}_{\alpha}),
\end{equation}
in analogy to the Wannier basis for electronic systems\cite{vanderbilt,lee2013,lee2014lattice,chaoming}. Here $\varphi_\nu(\mathbf{r} -\mathbf{R} - \mathbf{r}_{\alpha})$ are the L\"{o}wdin orbitals representing the TM modes, where
$\mathbf{R}$ is the usual lattice vector, $\mathbf{r}_{\alpha}$ the position vector of atom $\alpha$, and $\nu$ its orbital degree of freedom.
Due to the non-primitive lattice vector, the basis obeys extra constraints in addition to the Bloch condition: We have $\psi_{\alpha,\nu,\mathbf{k}+\mathbf{G}}(\mathbf{r}) = e^{i\mathbf{G}\cdot\mathbf{r}_{\alpha}} \psi_{\alpha,\nu,\mathbf{k}}(\mathbf{r})$,
where $\mathbf{G}$ is any reciprocal lattice vector. Hence the off-diagonal TB Hamiltonian matrix elements defined by $H_{\alpha\beta}(\mathbf{k}) = \int d\mathbf{r}\psi_{\alpha,\nu,\mathbf{k}}^*(\mathbf{r})\hat{H}\psi_{\beta,\nu,\mathbf{k}}(\mathbf{r})$ are periodic only up to a phase (i.e. form a projective representation of lattice translation):
\begin{equation}
H_{\alpha\beta}(\mathbf{k}+\mathbf{G}) = e^{i\mathbf{G}\cdot\mathbf{r}_{\Delta}}H_{\alpha\beta}(\mathbf{k}),
\end{equation}
where $\mathbf{r}_{\Delta} = \mathbf{r}_{\beta}-\mathbf{r}_{\alpha}$.
Lattices with different nonsymmorphic symmetries can be achieved with photonic cavities of different positions and orientations. Here, we shall implement them using dielectrics shaped as elliptical cylinders.
\begin{figure}[!t]
\centering
\includegraphics[width=.96\linewidth]{orbital}%
\caption{The $s$, $p_x$, $p_y$ and $2s$ orbitals above each elliptical cylinder in the photonic lattice unit cell. Note that they are neither isotropic nor aligned with the $x$ and $y$ axes, like the ellipses themselves.}
\label{fig:orbital}
\end{figure}
\begin{figure}[!t]
\subfloat[]{\label{fig:pg}%
\includegraphics[width=0.23\textwidth]{pg.pdf}%
}
\subfloat[]{\label{fig:pmg}%
\includegraphics[width=0.23\textwidth]{pmg.pdf}%
}\\
\subfloat[]{\label{fig:p4g}%
\includegraphics[width=0.23\textwidth]{p4g.pdf}%
}
\subfloat[]{\label{fig:pgg}%
\includegraphics[width=0.23\textwidth]{pgg.pdf}%
}
\caption{
The lattice structure, nodal positions and band dispersions for the four lattices we considered, with $pg$, $pmg$, $p4g$ and $pgg$ symmetry respectively (a to d). Results from our effective TB model (colored lines, see Supp. Materials~\cite{Note1}) agree closely with COMSOL simulation results (black dashed lines).
(a) The $pg$ bands from the $p_x$ and $p_y$ orbitals. There exists a Dirac point (DP) $P_1$ along $Y$-$M$ and a nodal line along $M$-$X$ protected by $\{m_y|\tau_x\}$.
(b) By rotating the ellipses in the $pg$ lattice to respect an additional mirror symmetry $m_x$, we obtain the $pmg$ lattice. The nodal line remains unchanged but the DP $P_2$ is now along $\Gamma$-$Y$.
(c) The lattice with $p4g$ symmetry, which has mirror symmetry along two diagonals and an additional glide reflection symmetry $\{m_x|\tau_y\}$.
The nodal lines persist but there is now a doubly degenerate DP $P_3$ at the $\Gamma$ point.
(d) The $pgg$ lattice obtained by breaking the $C_4$ symmetry of the $p4g$ lattice through arbitarry rotation of the ellipses.
The previously doubly degenerate DP decomposes into two singly degenerate DPs between $\Gamma$ and $\pm X$, one of which is visible here.
}
\label{fig:band}
\end{figure}
\textit{Line nodes from nonsymmorphic symmetry--}
As a first illustration of how nonsymmorphic symmetry can lead to degeneracies, consider the lowest two bands of the photonic crystal. These two bands correspond to the two $|s\rangle$ orbitals above the two inequivalent elliptical dielectric regions (labeled as $A$ and $B$).
Note that the $\ket{s}$ orbitals are not isotropic due to the anisotropy of the elliptical cylinder. From Fig.~\ref{fig:pg}, we see that the photonic crystal (PhC) has the symmetry of one of the simplest nonsymmorphic group $pg$, which only contains glide reflections.
The glide reflection operators are conventionally denoted by $g_y = \{m_y|\tau_x\}$, where $m_y \psi(x,y,z)=\psi(x,-y,z)$ and $\tau_x \psi(x,y,z)=\psi(x+\frac{a}{2},y,z)$.
Denoting orbital overlaps by $J^{AB}_{x,y}$, an immediate consequence of this glide reflection symmetry is that $J^{BA}_{x,y}=J^{AB}_{x,y}=J^{AB}_{-x,y}$ and $J^{AA}_{x,y}=J^{BB}_{x,y}$. Hence $H_{AB}(\pi/a,k_y)=0$ and $H_{AA}(\pi/a,k_y)=H_{BB}(\pi/a,k_y)$, i.e. we have a degenerate line node along $k_x=\pi/a$ (line $MX$). Analogous arguments hold for generic line degeneracies at the BZ boundary (see Fig.~\ref{fig:band} for more examples.)
\textit{Protected Dirac points--}
Besides protecting line nodes, nonsymmorphic symmetry also protects Dirac crossings in the photonic bandstructure by protecting the $Z_2$ topological number of the 1D Berry phase\cite{kariyado2013symmetry}. As detailed in the Supplement\cite{Note1}, point degenaracies must exist at the $Z_2$ jumps. By adjusting the relative positions and orientations of the cylinders in our photonic crystal, various nonsymmorphic symmetries $pg$, $pmg$, $pgg$ and $p4g$ (Fig.~\ref{fig:band}) can be implemented, each giving rise to protected Dirac points in certain bands. One observes the splitting, fusion and motion of these DPs as the cylinders are continuously modified.
Perturbing the 4-orbital TB model around each degeneracy\cite{sun2012topological,Note1} yields an effective 2-band Hamiltonian $H_{\mathbf{k}} = \mathbf{h}_{\mathbf{k}} \cdot \bm{\sigma}$ characterized by the $ \mathbf{h}_{\mathbf{k}}$ vector, where $\bm{\sigma}$ are the Pauli matrices. When sublattice symmetry is respected, as in the $pmg$, $pgg$ and $p4g$ symmetry groups, $\mathbf{h}_{\mathbf{k}}$ is confined to a plane and a winding number $w$ can be defined for the mapping $\mathbf{h}_{\mathbf{k}}$: $\mathbb{S}^1 \mapsto \mathbb{S}^1$ along a closed loop around the gapless point:
\begin{equation}
w = \oint\frac{d\mathbf{k}}{2\pi}\left[\frac{h_1}{|\mathbf{h}|}\nabla\frac{h_2}{|\mathbf{h}|}-\frac{h_2}{|\mathbf{h}|}\nabla\frac{h_1}{|\mathbf{h}|}\right] \in \mathbb{Z}.
\label{winding}
\end{equation}
To elucidate our findings in more detail:
\paragraph{pg group:}
We consider the same lattice as before (Fig.~\ref{fig:pg})
, but now focus on the four bands spanned by orbitals $\ket{A,p_x}$, $\ket{B,p_x}$, $\ket{A,p_y}$ and $\ket{B,p_y}$.
A Dirac point $P_1$ exists along $Y$-$M$, with gap opening up if time reversal or nonsymmorphic $pg$ symmetry is broken, i.e. by using a distorted magneto-optical dielectric. Along MX, a line node exists for reasons explained earlier.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{hvectorplot.pdf}
\caption{The configurations of $\mathbf{h}$ near the DPs of lattices with $pmg$, $p4g$ and $pgg$ symmetries. (a) and (b) depict the two DPs ($ \pm P_2$ between $\Gamma$ to $\pm Y$) of the $pmg$ lattice, with windings $w = -1$ and $w = 1$.
(c) The $w=-2$ $\mathbf{h}$ around the quadratic degeneracy ($P_3$) at $\Gamma$ for the $p4g$ symmetric lattice.
(d) With $p4g$ broken to $pgg$, the above $w=-2$ degeneracy splits into to two DPs ($\pm P_4$) along $-X$ to $X$, each with winding $w=-1$.
}
\label{fig:hvector}
\end{figure}
\paragraph{pmg group:} Besides glide reflection symmetry as in $pg$, the lattice also contains an additional mirror symmetry (Fig.~\ref{fig:pmg}). Within the bands spanned by $\ket{A,s}$, $\ket{B,s}$, $\ket{A,p_y}$ and $\ket{B,p_y}$, Dirac cones $\pm P_2$ between $\Gamma$ and $\pm Y$ appear without fine-tuning. Here, it is the mirror symmetry subgroup of $pmg$ that is essential in protecting the Dirac crossing.
By contrast, the nodal line requires the symmetry under the glide operation.
\paragraph{p4g group:} $p4g$ symmetry consists of mirror symmetries along the two diagonals and glide reflection symmetries $g_x=\{m_x |\tau_y\}$ and $g_y=\{m_y |\tau_x\}$ (Fig. 2(c)).
Due to the extra $C_4$ rotational symmetry, the $\Gamma$ point hosts a quadratically degenerate point $P_3$ in the space of orbitals $\ket{A,p_y}$, $\ket{B,p_y}$, $\ket{A,2s}$ and $\ket{B,2s}$, with a $\mathbf{h}_{p4g}$ winding of $w=-2$ (Fig.~\ref{fig:hvector}(c)).
To understand exactly which symmetry subgroup is necessary for protecting this double degeneracy, we proceed to break the $C_4$ rotational symmetry next.
\paragraph{pgg group:} Upon breaking $C_4$ rotation symmetry by rotating each of the elliptical cylinders, we obtain the $pgg$ lattice (Fig.2(d)) from the previous $p4g$ lattice (Fig.2(c)).
The quadratic degeneracy at $\Gamma$ decomposes into two linear ($w=-1$) DPs $P_4$ located either along $\Gamma$-$ X$, $\Gamma$-$ Y$ or their mirror inverses, depending on how the $C_4$ symmetry was broken. For the case of $\Gamma$-$X$ shown, $P_4$ is gapped by breaking both $g_x$ and $C_2$ (but not $g_y$). Hence either $g_x$ and parity symmetry can protect $P_4$, but only $g_x$ can confine $P_4$ along $\Gamma$-$ X$.
\textit{Lifshitz transition and anomalous refraction --}
Interestingly, nonsymmorphic symmetry protects the point degeneracies so robustly that a Dirac cone can ``tilt over'' and still remain gapless upon large parameter tuning. When a type-I (upright) Dirac cone tilts over into a type-II (tilted over) Dirac cone, the isofrequency ``Fermi'' surface undergoes a topological change known as a Lifshitz transition, from an isolated point to a pair of intersecting lines (Fig.~\ref{fig:pgII}). Its 3D analog has attracted considerable attention\cite{type2}, especially in the context of Weyl semimetals\cite{soluyanov2015type,muechler2016tilted, wang2016mote,deng2016experimental,koepernik2016tairte,type2}. In our PhCs, a Lifshitz transition can be induced across a wide range of nonsymmorphic symmetry preserving deformations, particularly when the dielectric constant $\epsilon_r$ or aspect ratio of the ellipses are varied (Fig. \ref{fig:phase}).
Near a tilted Dirac point, the Hamiltonian generically assumes the form
\begin{equation}
\mathcal{H}_{II}(\delta \bold{k}) = v_x \delta k_x \sigma_x + v_y \delta k_y \sigma_y + (u_x\delta k_x+u_y \delta k_y) \mathbb{I},
\label{HII}
\end{equation}
where $\delta \bf{k}$ is the displacement from the DP. In our case, the linear $u_y$ term is forbidden by glide symmetry. The tilt $ \eta=u_x/v_x$ is controlled by the last term: $\eta=0$ for an untilted type-I DP, and $|\eta|>1$ for a type-II DP.
\begin{figure}[!t]
\subfloat[]{\label{fig:type2lat}%
\includegraphics[width=0.15\textwidth]{latticever2.pdf}%
}
\subfloat[]{\label{fig:secondtype}%
\includegraphics[width=0.15\textwidth]{secondtypeline.pdf}%
}
\subfloat[]{\label{fig:phase}%
\includegraphics[width=0.17\textwidth]{phase.pdf}%
}
\caption{ The $pg$ lattice (a) and bandstructure (b) hosting type-II (tilted over) Dirac points. (c) Phase diagram for Lifshitz transition dependence on dielectric constant $\epsilon_r$ and aspect ratio $r_2/r_1$ of the ellipses, with $r_1=0.16a$ and orientation angles $\pm 80^\circ$. Type-I/II regions are marked in red/blue, while no DP exists in the white regions.
}
\label{fig:latticeandphase}
\end{figure}
From Eq. \ref{HII}, the isofrequency contour is given by
\begin{equation}
\delta\omega=\omega-\omega_0=\eta v_x \delta k_x \pm \sqrt{v_x^2 \delta k_x^2+v_y^2\delta k_y^2},
\end{equation}
where $\omega_0 = 0.811c/a$ is the frequency of the DP for our $pg$ lattice (Figs. \ref{fig:latticeandphase} and \ref{fig:type2}). Due to the unique double multiplicity of isofrequency lines near the DP, an incident light ray on the PhC will be anomalously separated into two refracted rays within the PhC. As derived in the Supp. Materials~\cite{Note1}, the two anomalous refraction angles corresponding to an incident angle $\theta$ are given by
\begin{equation}
\phi^{\pm} = \pm \tan^{-1} \frac{ v_y^2 |\delta k_y|}{\eta v_x(\delta\omega-\eta v_x \delta k_x) + v_x^2 \delta k_x},
\end{equation}
where $|\delta k_y| = \sqrt{(\delta\omega-\eta v_x \delta k_x)^2-v_x^2 \delta k_x^2}/v_y$ and $\delta k_x = \frac{\omega}{c}\sin\theta$. For frequencies near $\omega_0$, $|\phi^\pm|\approx \tan^{-1}\left[\frac{v_y}{v_x}\frac1{\sqrt{\eta^2-1}}\right]$, which suggests that anomalous refraction \emph{requires} $|\eta|>1$. This is contrasted with ordinary optical media where only one refracted ray is observed.
\begin{figure}[t]
\subfloat[]{\label{fig:pgII}%
\includegraphics[width=0.23\textwidth]{3Dtype2}%
}
\subfloat[]{\label{fig:typeIIrefract}%
\includegraphics[width=0.23\textwidth]{typeII.pdf}%
}
\caption{(a) 3D frequency plot near a type-II Dirac point, with characteristic cross-shaped isofrequency lines due to its tilt.
(b) Anomalous refraction from a type-II DP. Left panel: The refraction angles are aligned with the refracted group velocities, which are determined by the normals of the isofrequency lines. These lines intersect with the incident wavevector at the original tangential wavevector component $k_y$.
}
\label{fig:type2}
\end{figure}
\textit{Stoplight device--} The sensitive dependence of the photonic dispersion on lattice symmetry also suggests a potential stoplight device application. When the symmetry is reduced from $p4g$ to $pgg$ by breaking $C_4$ rotation,
the quadratic DP splits into two linearly dispersive DPs (Fig.~\ref{fig:p4g}). A simple representative Hamiltonian is
\begin{equation}H_{split}(k_x,k_y,k_0)=t H_d(k_x,k_y,k_0)\circ H_d(k_x,k_y,-k_0)
\label{Hsplit}
\end{equation}
where $\circ$ is the Hadamard product, $t$ is a constant, $k_0$ controls the splitting and $H_d(k_x,k_y,k_0) = (k_x-k_0)\sigma_x + k_y\sigma_y$ gives a single DP located at $k_0$. For an incident ray along the x-axis, Eq. \ref{Hsplit} gives a resultant group velocity $v_x^g = {\partial \omega}/{\partial k_x}|_{k_0} = 2t k_0 $.
\begin{figure}[t]
\subfloat[]{\label{fig:device}%
\includegraphics[width=0.26\textwidth]{device_a}%
}
\subfloat[]{\label{fig:curve}%
\includegraphics[width=0.21\textwidth]{device_b_0330}%
}
\caption{(a) Stoplight device with group velocity controlled by external applied voltage. An electric field deforms the silicone elastomer elliptic cylinders and breaks the $p4g$ symmetry to $pgg$, thereby modifying the dispersion. (b) The good agreement between COMSOL simulation results and the effective Hamiltonian in Eq. \ref{Hsplit}.
}
\end{figure}
The splitting $k_0$ can be dynamically controlled in suitable dielectrics exhibiting electric field induced strain, such as silicone elastomers\cite{pelrine2000high,kornbluh2000ultrahigh} with electrical permittivities within our desired range ($\epsilon_r = 11.82$)\cite{zhang2004}. As shown in Fig.~\ref{fig:device}, our proposed stoplight device consists of a $p4g$/$pgg$ symmetric PhC with elliptic cylinders made with 5\% 81-R hardener dissolved in silicone fluid blended with a 40\% copper-phthalocyanine oligomer. An appreciable strain of $12\%$ can be induced by a realistic applied electric field of $E \approx 25V/\mu m $ along the y-axis\cite{zhang2004}.
With external applied electric field, the cylinders lengthen along the x-axis and shorten along the y-axis, breaking the symmetry from $p4g$ to $pgg$ and giving rise to nonzero $k_0$. Assuming negligible Poisson ratio, $k_0$ is empirically fitted to $k_0 = b E_a^D$ where $b = 0.16\mu m/aV, D = 0.605$ and $t = 0.162ac$, which agrees well with COMSOL simulation results as shown in Fig.~\ref{fig:curve}. Since it is easy to precisely control the applied voltage, our device will can function robustly as an optical switch that controls, or slows, light propagation significantly with precision. An added advantage is that since light propagation is stopped by $C_4$ symmetry, the elliptic cylinders may be replaced by cuboids or other $C_4$-symmetric shapes for implementation convenience.
\textit{Conclusion--}
In this work, we proposed very simple 2D photonic lattices whose bandstructures possess highly tunable line nodes and Dirac points protected by nonsymmorphic symmetry. Consisting of easily fabricated elliptical dielectric rods, these lattices can realize all the nonsymmorphic wallpaper symmetry groups $pg$, $pmg$, $p4g$ and $pgg$.
Impressively, there exist large parameter regimes where the symmetry protected point degeneracy undergoes a Lifshitz transition into a type-II Dirac cone. Such exotic bandstructure topology have been associated with spectacular response properties\cite{soluyanov2015type}, and in our case results in anomalous refraction. The sensitivity of the bandstructure to lattice $p4g$ symmetry also leads to our proposal for a realistic light-stopping device.
\begin{acknowledgments}
We thank Shanhui Fan, Quan Zhou, Pinaki Sengupta, Yuhan Liu, Qixian Liao and Guang-Jie Li for useful discussions.
\end{acknowledgments}
|
\section{Introduction}
Since the discovery of quantum mechanics in the beginning of the last century, our classical understanding of the concept of observables has undergone a drastic change. It is by now widely accepted that, in the microscopic world, measured values of a physical quantity, termed \lq observable\rq\ in quantum mechanics, are intrinsically random, and that certain combinations of quantum observables do not admit coexistence, as exemplified typically by the pair of observables corresponding to the position and the momentum of a particle.
Such remarkable characteristics of quantum observables impose a strong limitation to the mathematical framework to be employed for describing their probabilistic behaviour; namely, it is no longer possible, in general, to assign probability spaces for the description of the joint behaviour of their arbitrary combinations in the classical sense.
Nonetheless, various attempts have been made to construct a proper mathematical framework for the probabilistic description of the combination of quantum observables that resembles the Kolmogorovian style of formulation of classical probability theory.
Extending the notion of probability has since been one of the major trends, which yielded the
extended notion of probability which goes generally by the name of `quasi-probability' or `pseudo-probability'.
Among the most celebrated proposal is the Wigner-Ville (WV) distribution \cite{Wigner_1932,Ville_1948}, commonly known as the Wigner function in the physics community, which is primarily considered for a canonically conjugate pair of quantum observables to describe their joint behaviour.
Another, though less known, example is the Kirkwood-Dirac (KD) distribution \cite{Kirkwood_1933,Dirac_1945}, which is structured differently but is meant to serve a similar purpose for arbitrary pairs.
Historically, those proposals including the WV and KD distributions have been made more or less in a heuristic manner, and as such, the general mathematical framework for the study, including the prescription for the concrete construction of such distributions to a pair of arbitrary quantum observables, which may comprehensively be termed `quasi-joint-probability' (QJP) distributions, is still underdeveloped, not to mention a transparent overview of the relations among the QJPs. We know, for instance, that
both the WV and KD distributions retain similar properties to the standard joint-probability distributions defined for a pair of classical random variables,
but they exhibit their own outstanding queerness in that the former admits negative numbers to be assigned whereas the latter takes even complex numbers. However, we still do not know whether the peculiar properties of joint-probability including those of the WV and KD distributions, which have occasionally been considered a serious impediment to their physical interpretation, are a norm of QJP distributions, or there can be other types of examples which share classical properties of joint-probability in different aspects. The theme of this paper revolves around the concept of QJP distributions of quantum observables, with the first objective being to present a mathematically solid framework to address some of their problems in a more systematic and lucid manner.
Another motivation of this paper comes from the recent rise of interest in the novel quantum observable called the \emph{weak value},
which has been put forward by Aharonov and co-workers \cite{Aharonov_1988} based on their time-symmetric formulation of quantum mechanics \cite{Aharonov_1964} proposed more than a half century ago. In simple terms, the weak value
\begin{equation}
A_{w} := \frac{\langle \psi^{\prime}, A \psi \rangle}{\langle \psi^{\prime}, \psi \rangle}
\end{equation}
is a physical quantity that supposedly characterises the value of the observable $A$ in the process specified by an initial state $\vert\psi \rangle$ and a final state $\vert\psi^{\prime} \rangle$ both specified in advance. Unlike the standard physical value which is given by one of the eigenvalues of an observable $A$,
the weak value admits a definite value for any $A$, and is envisaged to be meaningful even for a set of non-commutable observables simultaneously.
This inspired a new insight for analysing the quantum nature of the system as well as for
understanding various counter-intuitive phenomena in quantum mechanics based on the weak value. For instance, the complex-valued nature of $A_{w}$ allows for a direct measurement of the wave function, offering a novel technique to rival the existing technology of quantum tomography. This in turn alludes us to contemplate on the possible trajectory of a particle \cite{Lundeen_2011,Mori_2015}, a notion which has conventionally been deemed untenable due to the incompatibility of measuring the position and the momentum simultaneously.
The weak value also admits novel physical interpretations on such fundamental aspects of quantum mechanics as the wave-particle duality and the local existence of the physical quantity itself, offering us a possible resolution to some of the quantum paradoxes, including the three-box paradox \cite{Aharonov_1991}, Hardy's paradox \cite{Yokota_2009} and the Cheshire cat paradox \cite{Aharonov_2005}.
Despite its growing attention, the status of the weak value in quantum mechanics is still not solid, and especially its physical interpretation is still open to debate. One of the recent strategies in addressing this question has been to investigate its relations to quasi-probabilities, specifically those to the KD distribution \cite{Ozawa_2011, Hofmann_2014}. In this paper, we shall follow this line of study and show, among others, that a novel geometric/statistical interpretation emerges from these distributions. This necessitates a sound mathematical basis of QJP distributions, which we will provide in the course of our discussions.
The main theme of this paper is thus to obtain a more coherent understanding of the formalism of QJP distributions of quantum observables, and subsequently to apply the results in some areas of the foundational problems of quantum mechanics. In view of this, the key problems regarding QJP distributions may be to\begin{enumerate}
\item
provide a reasonably solid mathematical framework for the study of QJP distributions based on measure and integration theory, and possibly on the theory of generalised functions,
\item
present a viable scheme to address the inherent indefiniteness/arbitrariness to the possible candidates for QJP distributions of non-commuting pairs of quantum observables, a methodical way for their constructions, and the relation between each of the candidates, and
\item
devise a procedure for measuring such various candidates of QJP distributions in a systematic manner.
\end{enumerate}
We shall address these problems from two complementary approaches: one from a bottom-up, strictly operational construction realised by carefully reviewing the mathematical description of the conditioned measurement scheme, and the other from a top-down viewpoint realised by applying the results of spectral theorem for normal operators and its Fourier transforms.
The results of the study shall be subsequently applied to the analysis for the physical interpretation of the weak value. To this end, we first concentrate on the $L^{2}$ structures which the QJP distributions naturally induce, and observe that they furnish a statistical interpretation of the geometric structures introduced on the space of observables in the underlying Hilbert space, analogously to those introduced in the space of random variables in classical probability theory. Geometric concepts such as orthogonal projections and inner products are accordingly endowed with statistical interpretations as `conditionings' and `correlations', respectively, and in addition the representation of linear operators by functions provides us with a convenient tool for evaluating statistical quantities involved. These observations form a basis to perform further study on the weak value in general. As a result,
the weak value $A_{w}$ is given a geometric/statistical interpretation: either as the orthogonal projection of an observable $A$ on the subspace generated by another observable $B$ which is determined by one of the predetermined states entering in the weak value, or equivalently, as the conditioning of $A$ given $B$ with respect to the QJP distribution under consideration.
Although we shall not discuss it here, we mention that
this interpretation also leads to a set of novel and remarkable inequalities of uncertainty relations for approximation/estimation which are capable of treating both the standard position-momentum inequality and the time-energy inequality \cite{Lee_2016}.
As for the practical outcomes of our argument laid out for QJP distributions,
we mentioned earlier the systematic construction of QJP distributions and the geometric/statistical interpretation of the weak value, but each of these can be made more explicit as follows.
First, for the systematic construction of QJP distributions, we furnish a general prescription which ensures that it can describe the joint behaviour of an arbitrary pair of quantum observables.
Specifically, inspired by the observations made on the Fourier transform of the product spectral measure of two simultaneously measurable observables $A$ and $B$, we introduce a mixture $\#(s,t)$ of the disintegrated components of $e^{-isA}$ and $e^{-itB}$ with real parameters $s, t$
for arbitrary pairs of (generally non-commuting) observables $A$ and $B$, and thereby define the QJP distribution of the pair by the inverse Fourier transform of the distribution $(s,t) \mapsto \langle \psi, \#(s,t) \psi \rangle / \|\psi\|^{2}$ to a given quantum state $|\psi\rangle$.
Each of the QJP distributions is then found to possess reasonable properties to be qualified as what its name suggests to be, and one can confirm that both the WV distribution and the KD distribution do belong to this class. The inherent arbitrariness observed to the candidates for QJP distributions is then understood as the possible variety of the way one could mix the disintegrated components of the unitary operators, which originates directly from the non-commutative nature of the pair of the observables $A$ and $B$.
A concrete measurement scheme for members of a specific subfamily of QJP distributions is further proposed.
For the geometric/statistical interpretation of the weak value, on the other hand,
we start by noting that, as distributions, each QJP distribution naturally induces an $L^{2}$ structure. We will then find that the QJP distributions provide convenient methods of representing geometric structures in terms of the inner products of the form
\begin{equation}\label{def:intro_inn_prod}
\llangle B, A \rrangle_{\psi,\alpha} := \frac{1 + \alpha}{2} \cdot \frac{\langle B\psi, A\psi\rangle}{\|\psi\|^{2}} + \frac{1 - \alpha}{2} \cdot \frac{\langle A\psi, B\psi\rangle}{\|\psi\|^{2}}, \quad -1 \leq \alpha \leq 1,
\end{equation}
which can be introduced on the space of operators in the underlying Hilbert space by integration of functions. With this inner product, we are allowed to consider orthogonal projections onto the subspaces $\mathfrak{E}_{\psi}(B)$ of operators generated by self-adjoint operators $B$, and find that the orthogonal projections can be interpreted as conditioning given $B$ with respect to the QJP distributions under consideration. The projection
\begin{equation}
P_{\alpha}(A|B;\psi) = \int_{\mathbb{R}} \left( \frac{1 + \alpha}{2} \cdot \frac{\langle b, A \psi \rangle}{\langle b, \psi \rangle} + \frac{1 - \alpha}{2} \cdot \frac{\langle \psi, A b \rangle}{\langle \psi, b \rangle} \right) dE_{B}(b)
\end{equation}
of the observable $A$ on the subspace $\mathfrak{E}_{\psi}(B)$ is further found to be described by the weak value\footnote{
Here, $E_{B}$ denotes the unique spectral measure associated to the self-adjoint operator $B$ (more on this in Section~\ref{sec:ups_II_pre}). Intuitively, $E_{B}(b) = |b\rangle\langle b|$ is the projection associated with each of the eigenvalues $b$ of $B$, and thus $dE_{B}(b) = |b\rangle\langle b| db$ in a laxer expression.},
providing us with its proper geometric/statistical interpretation (Proposition~\ref{prop:orth_proj_cond_qe}).
Having furnished a general introduction to the topic of QJP distributions of quantum observables and the weak value along with a brief summary of the content, we now
give the outline of the present paper.
After this introductory section, we organize the main body, Section 2 to Section 7, into the following three logical groups of mutually interrelated topics:
\begin{enumerate}
\renewcommand{\labelenumi}{(\Alph{enumi})}
\item {\bf QJP: Heuristic Construction}\hspace{5pt} Four sections starting from Section~\ref{sec:ups_I} to \ref{sec:ps_II} are devoted to a heuristic and bottom-up construction of QJP distributions of a pair of quantum observables.
This is accomplished by a thorough analysis on the mathematical formalism of two measurement schemes. One is the
standard scheme, which we call the \lq unconditioned measurement (UM) scheme\rq, in which we measure an observable $A$ under a given state as conventionally done (Section~\ref{sec:ups_I} to \ref{sec:ups_II}). The other is what we call the \lq conditioned measurement (CM) scheme\rq, in which under a given state we measure an observable $A$ along with another observable $B$ whose outcome is used for conditioning (Section~\ref{sec:ps_I} to \ref{sec:ps_II}).
Each of these analyses will be conducted on the level of (conditional) expectations and (conditional) probabilities.
\begin{enumerate}
\renewcommand{\labelenumii}{(\roman{enumii})}
\item {\bf UM I}\hspace{5pt}
We start by reviewing, in Section~\ref{sec:ups_I}, the UM scheme by a standard operator-centric approach, and investigate how one could reclaim the information of the target system by that means.
\item {\bf UM II}\hspace{5pt}
Subsequently, in Section~\ref{sec:ups_II}, we take a closer look on the UM scheme in the level of probabilities, where the quantity of interest is now not only the statistical average, but also the `raw' probability measure describing the probabilistic behaviour of the measurement outcomes of the meter observable, and discuss how one could recover the probability measure describing the outcomes of the target observable.
\item {\bf CM I}\hspace{5pt}
From Section~\ref{sec:ps_I} onward, we turn our attention to the CM scheme.
In Section~\ref{sec:ps_I}, we first conduct, in a parallel manner as we have done in the preceding Section~\ref{sec:ups_I}, an analysis in the operator level, where now the quantity of interest becomes the conditional expectation of the meter observable given another conditioning observable $B$ of the target system.
\item {\bf CM II}\hspace{5pt}
In Section~\ref{sec:ps_II}, the study of the CM scheme is given a probabilistic approach, where the quantity of interest is the Wigner-Ville distribution of a pair of canonically conjugate observables on the meter system conditioned by the outcome of the conditioning observable $B$ of the target system. We then see that this implies the existence of the concept of QJP distributions of pairs of generally non-commuting observables.
\end{enumerate}
\item {\bf QJP: Formal Definition}\hspace{5pt}
Inspired by the heuristic arguments employed in the operational analyses over the preceding four sections, we devote Section~\ref{sec:qp_qo} to the top-down construction of QJP distributions for arbitrary pairs of generally non-commutating quantum observables. We shall then summarise our findings obtained through Section~\ref{sec:ups_I} to Section~\ref{sec:ps_II} from a rather aerial viewpoint, discussing where the heuristic arguments and observations in the preceding sections find their places in this relatively general framework.
\item {\bf Application to the Interpretation of Weak Values}\hspace{5pt}
As an application of the mathematical formalism provided so far, in Section~\ref{sec:app} we conduct a study on the quantum analogue of correlations, which can be defined even for a pair of non-commuting observables. This leads us to the aforementioned geometric/statistical interpretation of the weak value as conditional quasi-expectations.
\end{enumerate}
We shall finally summarise our results and give some concluding remarks in the last Section~\ref{sec:sc}.
Prior to our main discussions, however, we wish to say a few words about the mathematical preliminaries we supposed for the readers in preparing this paper.
The formalism that we intend to provide necessarily requires, on top of the mandatory functional analysis, moderate acquaintance to measure and integration theory, preferably some familiarity with the basic terminologies in general topology, and ideally insight into the basic ideas of the theory of generalised functions. The obvious difficulty is then to find a decent balance between rigour and generality on one side, and accessibility on the other. To achieve this balance as much as possible, and assure our entire arguments to be fully accessible without any prior knowledge of advanced mathematics, we have included at the beginning of each section a subsection entitled Reference Materials containing a rather lengthy introduction of mathematical concepts that are used in the subsequent discussions. While the authors took care in introducing these mathematical concepts and their results in a self-contained manner to respect their logical sequence, these Reference Materials are primarily intended to serve as a convenient place to summarise the basic concepts and results in a crash-course, and as such, the mathematical theories presented there are not intended to be learned from scratch. For those who are interested in the mathematics itself are advised to be referred to standard textbooks on the respective topics, {\it e.g.}, for general topology \cite{Kelley_1975,Munkres,Querenburg_2011}, measure and integration theory \cite{Elstrodt_2011, Amann_1998a, Amann_1998b, Amann_2001, Rudin_1976, Rudin_1986}, functional analysis
\cite{Rudin_1991, Werner_2011},
and also those specifically targeting the audience from the physics community \cite{Reed_Simon_1975, Reed_Simon_1980, Goldhorn_2009, Goldhorn_2010}.
Naturally, those who are already familiar with the preparatory materials may safely skip them and directly go to the main arguments that follow.
Admittedly, the style of discussion found in this paper is heavily oriented toward mathematical rigorousness and logical clarity rather than brevity and physical intuition, especially compared to those found in the majority of the literature in physics.
However, in spite of the possible initial hesitation that may be expected for the general readers due to the unfamiliarity of the style, the authors decided to adopt it in the belief
that this way of presentation has its own merit, and that the costs will outweigh the rewards in the end. In fact, several important concepts and results from the branches of mathematics mentioned above (specifically, measure and integration theory and functional analysis) are quite indispensable in understanding some of the interesting results obtained in this paper. This is so, for instance, in defining the conditional quasi-expectations (to which Aharonov's weak value belongs as a special case) in terms of the Radon-Nikod{\'y}m derivative to understand their properties (Section~\ref{sec:psI_cond_quasi_exp}), in formulating the problem of the `limit of amplification' by conditioning in terms of essential suprema (Section~\ref{sec:psI_amplification}), in defining a family of QJP distributions of a combination of generally non-commuting quantum observables by the method of hashing (Section~\ref{sec:qp_qo}), and in providing geometric and `statistical' interpretation of conditional quasi-expectations (Section~\ref{sec:app}). The authors hope that the readers will not be discouraged by these mathematical materials, but rather enjoy them to go through the discussions and reach the fruit of the physical results they finally brings forth.
\paragraph{Mathematical Notations Employed}
Throughout this paper, we denote by $\mathbb{K}$ either the real field $\mathbb{R}$ or the complex field $\mathbb{C}$, and define $\mathbb{K}^{\times} := \mathbb{K} \setminus \{0\}$. In order to avoid confusion, we denote the collection of all natural numbers including $0$ by $\mathbb{N}_{0}$, and $\mathbb{N}^{\times} := \mathbb{N}_{0} \setminus \{0\}$. Since our primary interest is on quantum mechanics, Hilbert spaces are always assumed to be complex. Conforming to the convention in physical literature, we denote the complex conjugate of a complex number $c \in \mathbb{C}$ by $c^{*}$, and an inner product $\langle \,\cdot\, , \,\cdot\, \rangle$ defined on a complex linear space is anti-linear in its first argument and linear in the second.
For simplicity, we adopt the natural units where we specifically have $\hbar = 1$, unless stated otherwise.
\newpage
\section{Unconditioned Measurement I: In Terms of Expectations}\label{sec:ups_I}
We start by providing a brief review on the archetype of the indirect measurement scheme widely known as the von Neumann measurement scheme. The scheme will be referred to as the \emph{unconditioned measurement} (UM) scheme in generic terms throughout this paper, primarily in order to contrast it with the \emph{conditioned measurement} (CM) scheme (which includes the \emph{post-selected measurement} scheme as a special case) discussed later.
\subsection{Reference Materials}\label{sec:ups_I_pre}
As a preamble to this section, we here include three introductory topics that form the basis of our study. We start by collecting some of the basic terminologies and results of measure and integration theory, based on which modern probability theory was established by Kolmogorov {\it et al.}
Subsequently, we provide a brief note on both the \emph{Schr{\"o}dinger representation} and the \emph{Weyl representation} of the canonical commutation relations (CCR), which will be extensively employed in describing the meter system in our measurement scheme. We finally close this subsection by providing a short summary on the precise definition of tensor products of Hilbert spaces and that of self-adjoint operators. Since these materials are included just to make our presentation self-contained, those who are already familiar with the subject may safely skip the contents and proceed directly to Section~\ref{sec:ups_I_ups}.
\subsubsection{A Crash-Course into Measure and Integration Theory}\label{sec:usp_I_MI}
We begin by presenting some of the most basic concepts and results of measure and integration theory, starting from the definition of measure spaces up to the construction of the Lebesgue integration, followed by the definition of $L^{p}$ spaces.
\paragraph{$\sigma$-algebras and Measurable Spaces}
Let $X$ be any set, and let $\mathfrak{P}(X)$ denote the power set%
\footnote{The symbol $\mathfrak{A}$ is the capital letter of the Fraktur typeface of `A' as in `Algebra', $\mathfrak{B}$ for `B' as in `Borel', $\mathfrak{E}$ for `E' as in `Erzeuger (generator)', $\mathfrak{O}$ for `O' as in `offen (open)' and $\mathfrak{P}$ for `P' as in `Potenz (power)' (some of them introduced shortly after).}
of $X$, {\it i.e.}, the collection of all subsets of $X$. A family $\mathfrak{A} \subset \mathfrak{P}(X)$ of subsets of $X$ is called a \emph{$\sigma$-algebra}
over $X$, if it satisfies the following conditions:
\begin{enumerate}
\item $X \in \mathfrak{A}$.
\item $A \in \mathfrak{A}$ implies $A^{c} := X \setminus A \in \mathfrak{A}$.
\item For any sequence $(A_{n})_{n \geq 1}$ of subsets of $X$, $\bigcup_{n=1}^{\infty} A_{n} \in \mathfrak{A}$ holds.
\end{enumerate}
Given a $\sigma$-algebra $\mathfrak{A}$ over $X$, each element $A \in \mathfrak{A}$ is called a \emph{measurable set}, and the ordered pair $(X, \mathfrak{A})$ is called a \emph{measurable space}.
\paragraph{Generator of a $\sigma$-algebra}
A trivial, but important property of $\sigma$-algebras is that, for any collection $(\mathfrak{A}_{i})_{i \in I}$ of $\sigma$-algebras over $X$ indexed by an index set $I$, the intersection $\bigcap_{i \in I} \mathfrak{A}_{i} = \{ A \in \mathfrak{P}(X) : A \in \mathfrak{A}_{i}, \forall i \in I\}$ is itself a $\sigma$-algebra over $X$.
This leads to the following basic fact: For any collection $\mathfrak{E} \subset \mathfrak{P}(X)$ of subsets of $X$, there exists a smallest (with respect to the set inclusion) $\sigma$-algebra encompassing $\mathfrak{E}$, namely, the intersection of all $\sigma$-algebras that encompass $\mathfrak{E}$. The intersection is called the \emph{$\sigma$-algebra generated by $\mathfrak{E}$}, denoted as $\sigma(\mathfrak{E})$, and $\mathfrak{E}$ is in turn called the \emph{generator} of $\sigma(\mathfrak{E})$.
\paragraph{Borel $\sigma$-algebras}
Let $X$ be a metric (or, in general, a topological) space, and let $\mathfrak{O}$ denote the collection of all open sets of $X$. We call the $\sigma$-algebra generated by $\mathfrak{O}$, the \emph{Borel $\sigma$-algebra of $X$}, and denote it by $\mathfrak{B}(X) := \sigma(\mathfrak{O})$. We prepare a special symbol for the special case $X = \mathbb{R}^{n}$ ($n \in \mathbb{N}^{\times}$), in which we denote the Borel $\sigma$-algebra of $\mathbb{K}^{n}$ by $\mathfrak{B}^{n} := \mathfrak{B}(\mathbb{R}^{n})$, which is among the most well-known examples of $\sigma$-algebras that, incidentally, also plays an important role in quantum theory. For simplicity, we occasionally denote $\mathfrak{B} := \mathfrak{B}^{1}$ whenever there is no risk of confusion.
\paragraph{Measures and Measure Spaces}
Let $(X, \mathfrak{A})$ be a measurable space. A map $\mu: \mathfrak{A} \to \overline{\mathbb{R}}$ from the $\sigma$-algebra $\mathfrak{A}$ to the \emph{extended real line} $\overline{\mathbb{R}} := \mathbb{R} \cup \{-\infty, \infty\}$ is called a \emph{measure}, if $\mu$ satisfies the following conditions:
\begin{enumerate}
\item $\mu(\emptyset) = 0$.
\item $\mu \geq 0$.
\item For any sequence $(A_{n})_{n \geq 1}$ of pairwise disjoint subsets of $X$, the \emph{countable additivity}
\begin{equation}\label{def:count_add}
\mu\left(\bigcup_{n=1}^{\infty} A_{n}\right) = \sum_{n=1}^{\infty} \mu(A_{n})
\end{equation}
holds.
\end{enumerate}
Given a measure $\mu$ over a measurable space $(X, \mathfrak{A})$, the ordered triple $(X, \mathfrak{A}, \mu)$ is called a \emph{measure space}.
\paragraph{Lebesgue-Borel Measure}
As a concrete example, we make notes on the $n$-dimensional \emph{Lebesgue-Borel measure} $\beta^{n}$ ($n \in \mathbb{N}^{\times}$) defined on the measurable space $(\mathbb{R}^{n}, \mathfrak{B}^{n})$, which is among the most well-known and important examples of measure spaces. To this end, we first recall that a measure $\mu$ on $(\mathbb{R}^{n}, \mathfrak{B}^{n})$ is called \emph{translation invariant}, if
\begin{equation}
\mu(B + a) = \mu(B), \quad B \in \mathfrak{B}^{n},
\end{equation}
holds for any $a \in \mathbb{R}^{n}$, where $B + a := \{ x + a: x \in B\}$. The Lebesgue-Borel
measure $\beta^{n}$ is then specified as the unique translation invariant measure on $(\mathbb{R}^{n}, \mathfrak{B}^{n})$ that satisfies the normalisation condition $\beta^{n}(]0,1]^{n}) = 1$, where
\begin{equation}
]0,1]^{n} := \{ x \in \mathbb{R}^{n} : 0 < x_{i} \leq 1\,\,\, \text{for} \,\, 1 \leq i \leq n, \text{ $x_{i}$ is the $i$th coordinate of $x$}\}.
\end{equation}
This is the measure which is implicitly assumed for the most case in performing the usual integration by the symbol
\begin{equation}
\int_{-\infty}^{\infty} f(x)\, dx := \int_{\mathbb{R}} f(x)\ d\beta(x),
\end{equation}
which is a common practice in the physics community (the precise definition of the integral on the r.~h.~s. will be presented shortly after). The proof of the existence and uniqueness of the Lebesgue-Borel measure will be found in most elementary textbooks on the topic.
\paragraph{Measurable Functions}
Let $(X, \mathfrak{A})$ and $(X^{\prime}, \mathfrak{A}^{\prime})$ be measurable spaces. A map $f: X \to X^{\prime}$ is called \emph{$\mathfrak{A}$-$\mathfrak{A}^{\prime}$ measurable} (or just \emph{measurable} for short, whenever the measure spaces concerned are obvious by context), if $f^{-1}(\mathfrak{A}^{\prime}) \subset \mathfrak{A}$ holds. In particular, we call a map $f:X \to X^{\prime}$ from a metric (or a topological) space $X$ to another metric (or a topological) space $X^{\prime}$ \emph{Borel-measurable} if it is $\mathfrak{B}(X)$-$\mathfrak{B}(Y)$ measurable. An important fact to note is that a continuous map $f: X \to X^{\prime}$ is necessarily Borel-measurable.
\paragraph{Numerical Functions}
In integration theory, it proves fruitful to consider not only real functions
$f: X \to \mathbb{R}$, but also functions that take values in the extended real line $\overline{\mathbb{R}}$, which is called a \emph{numerical function}. One naturally equips $\overline{\mathbb{R}}$ with the ordering $-\infty < a < +\infty$, $a \in \mathbb{R}$, and may also define agreeable operations of addition, subtraction and multiplication,
where most of them should be self-evident, except for the following rather arbitrary definition
\begin{equation}
0 \cdot (\pm \infty) := (\pm \infty) \cdot 0 := 0, \quad \infty -\infty := -\infty + \infty := 0.
\end{equation}
We then define the $\sigma$-algebra on $\overline{\mathbb{R}}$ by
\begin{equation}
\overline{\mathfrak{B}} := \{ B \cup E : B \in \mathfrak{B}, \, E \subset \{-\infty, +\infty\}\},
\end{equation}
where, in particular, its restriction on the real line gives $\overline{\mathfrak{B}}|_{\mathbb{R}} = \mathfrak{B}$. We then say that a numerical function $f: (X, \mathfrak{A}) \to (\overline{\mathbb{R}}, \overline{\mathfrak{B}})$ is measurable, if it is $\mathfrak{A}$-$\overline{\mathfrak{B}}$ measurable. Throughout this paper, we denote by $\mathcal{M}^{+}(\mathfrak{A})$ (or occasionally by $\mathcal{M}^{+}$, whenever the $\sigma$-algebra concerned is evident by context) the collection of all measurable non-negative numerical functions.
\paragraph{Lebesgue Integration}
In introducing the concept of integration, we proceed in three steps: We first define the integration for non-negative step functions, then extend the treatment to functions belonging to $\mathcal{M}^{+}$, and finally discuss the integrability of measurable numerical or complex functions.
\begin{enumerate}
\item {\it Integration of Step Functions. \hspace{15pt}}
Let $(X, \mathfrak{A}, \mu)$ be a measure space. A measurable function $f: (X, \mathfrak{A}) \to (\mathbb{R}, \mathfrak{B})$ is called a \emph{step function} (staircase function, simple function), if it takes only finite distinct values in $\mathbb{R}$. The collection of all measurable non-negative step functions will be denoted by $\mathcal{T}^{+}$. One readily sees that a non-negative step function $f \in \mathcal{T}^{+}$ admits an expression
\begin{equation}\label{eq:step_function_expression}
f = \sum_{k=1}^{m} a_{k}\chi_{A_{k}},
\end{equation}
where $a_{1}, \dots, a_{m} \geq 0$ are non-negative real numbers, $A_{1}, \dots, A_{m} \in \mathfrak{A}$ are measurable sets, and $\chi_{A}$ denotes the characteristic function
\begin{align}\label{def:characteristic_function}
\chi_{A}(x) =
\begin{cases}
1, & x \in A, \\
0, & x \notin A,
\end{cases}
\end{align}
of the subset $A \subset X$.
We then define the \emph{($\mu$-)integral of $f$ (over $X$)} as
\begin{equation}\label{def:lebesgue_integral_01}
\int_{X} f\ d\mu := \sum_{k=1}^{m} a_{k}\mu(A_{k}),
\end{equation}
whose value lies in $[0,\infty]$. Note that, although the expression (\ref{eq:step_function_expression}) is non-unique due to the possible choice of the
measurable sets used,
the definition \eqref{def:lebesgue_integral_01} is well-defined since the outcome of the integral is independent of the choice.
\item
{\it Integration of Functions in $\mathcal{M}^{+}$. \hspace{15pt}}
Now that we have defined the Lebesgue integral of non-negative step functions, we next define the integral of non-negative measurable numerical functions. For $f \in \mathcal{M}^{+}$, the Lebesgue integral of $f$ is defined as
\begin{equation}\label{def:lebesgue_integral_02}
\int_{X} f\ d\mu := \sup \left\{ \int_{X} s\ d\mu : 0 \leq s \leq f, s \in \mathcal{T}^{+} \right\}.
\end{equation}
The above definition \eqref{def:lebesgue_integral_02} is consistent with that for step functions \eqref{def:lebesgue_integral_01} introduced earlier, for one readily checks that the integral coincides for $f \in \mathcal{T}^{+} \subset \mathcal{M}^{+}$.
\end{enumerate}
Before we move on to the final step, we introduce some useful notations. We let $\mathbb{K}$ denote either the real field $\mathbb{R}$ or the complex field $\mathbb{C}$, and we understand them to be respectively equipped with the Borel $\sigma$-algebra $\mathfrak{B}$ or $\mathfrak{B}^{2}$. Analogously, we let
\begin{equation*}
\hat{\mathbb{K}} := \overline{\mathbb{R}} \text{ or } \mathbb{C}, \text{ respectively equipped with the $\sigma$-algebra $\hat{\mathfrak{B}} := \overline{\mathfrak{B}}$ or $\mathfrak{B}^{2}$}
\end{equation*}
for later convenience. For a numerical function $f : X \to \overline{\mathbb{R}}$, we define its positive and negative parts as
\begin{equation}
f^{\pm}(x) := \max( \pm f(x), 0).
\end{equation}
One then sees that a function $f: X \to \hat{\mathbb{K}}$ is measurable if and only if all the positive and negative parts of both the real and imaginary parts $(\,\mathrm{Re} f)^{\pm}$, $(\,\mathrm{Im} f)^{\pm}$ of $f$ are measurable.
Given the necessary preparations, we finally obtain the following definition:
\begin{definition*}[Lebesgue Integral]
Under the assumptions above, a function $f: X \to \hat{\mathbb{K}}$ is called $\mu$-integrable (or simply integrable) over $X$ if $f$ is measurable, and all the four integrals
\begin{equation}
\int_{X} (\,\mathrm{Re} f)^{\pm}\ d\mu,\qquad \int_{X} (\,\mathrm{Im} f)^{\pm}\ d\mu
\end{equation}
are finite. The value
\begin{equation}
\int_{X} f\ d\mu := \int_{X} (\,\mathrm{Re} f)^{+}\ d\mu - \int_{X} (\,\mathrm{Re} f)^{-}\ d\mu + i \int_{X} (\,\mathrm{Im} f)^{+}\ d\mu - i \int_{X} (\,\mathrm{Im} f)^{-}\ d\mu
\end{equation}
is then called the ($\mu$-)integrable of $f$ (over $X$) or the Lebesgue integral of $f$ (over $X$ with respect to $\mu$).
\end{definition*}
\noindent
By definition, linearity
\begin{equation}
\int_{X} (af(x) + bg(x))\, d\mu(x) = a\int_{X} f(x)\, d\mu(x) + b\int_{X} g(x)\, d\mu(x),
\end{equation}
of the integration naturally follows as expected.
For a measurable set $A \in \mathfrak{A}$, the use of the shorthand
\begin{equation}
\int_{A}f\ d\mu := \int_{X} \chi_{A} \cdot f\ d\mu
\end{equation}
is common, where $\chi_{A}$ is the characteristic function of the measurable set.
\paragraph{Probability Spaces and Expectation Values}
A measure space $(X,\mathfrak{A},\mu)$ is called a \emph{probability space}, if the measure is normalised by unity $\mu(X) = 1$. Given a probability space $(X,\mathfrak{A},\mu)$ and a $\mu$-integrable function $f$, the total integration of $f$ is occasionally denoted by
\begin{equation}
\mathbb{E}[f;\mu] := \int_{X} f\ d\mu,
\end{equation}
and called the \emph{expectation value} of $f$ under $\mu$.
\paragraph{Dominated Convergence Theorem}
The advantage of the Lebesgue integration (over the familiar Riemann counterpart) especially manifests itself when dealing with convergence. For later use throughout this paper, we make a note of one of the most powerful and oft-used theorems regarding the interchange of limit and integration. To this end, we first furnish some terminologies. Let $(X, \mathfrak{A}, \mu)$ be a measure space, and let a statement $E$ be defined on each element $x \in X$. We say that the statement $E$ holds \emph{($\mu$-) almost everywhere} (abbreviation: ($\mu$)-a.e.), if there exists a measurable set $N \in \mathfrak{A}$ with $\mu(N) = 0$ such that the statement $E$ holds for $ x \in X \setminus N$.
\begin{theorem*}[Dominated Convergence Theorem]
Let $(X, \mathfrak{A}, \mu)$ be a measure space, and let $f, f_{n}: X \to \hat{\mathbb{K}}$ ($n \in \mathbb{N}^{\times}$) be measurable. If the sequence of the functions converge point-wise $\lim_{n \to \infty} f_{n} = f$ $\mu$-a.e., and if moreover there exists an $\mu$-integrable function $g \in \mathcal{M}^{+}$ such that $|f_{n}| \leq g$ holds $\mu$-a.e. for all $n \in \mathbb{N}^{\times}$, then
\begin{equation}
\lim_{n \to \infty} \int_{X} \left| f_{n} - f \right| \ d\mu = 0
\end{equation}
holds, which in particular implies
\begin{equation}
\lim_{n \to \infty} \int_{X} f_{n}\ d\mu = \int_{X} f\ d\mu.
\end{equation}
\end{theorem*}
\paragraph{$L^{p}$ Spaces}
Having provided the definition of the Lebesgue integration, we close this subsection by introducing an important class of function spaces: $L^{p}$. Let $\mathcal{L}^{p}(\mu)$, $1 \leq p <\infty$, denote the space of all measurable functions $f: X \to \mathbb{K}$ for which its $L^{p}$-norm
\begin{equation}\label{def:Lp_norm}
\|f\|_{p} := \left( \int_{X} |f|^{p}\ d\mu \right)^{1/p}
\end{equation}
is finite. For $p = \infty$, we let $\mathcal{L}^{\infty}(\mu)$ denote the space of all $f$ for which its \emph{essential supremum}
\begin{equation}\label{def:ess_sup}
\|f\|_{\infty} := \inf\{\lambda \in [0,\infty] : |f| \leq \lambda \text{ $\mu$-a.e.} \}
\end{equation}
is finite (such a function is called \emph{essentially bounded}). The term essential supremum is justified by the fact that the evaluation $|f| \leq \|f\|_{\infty}$ $\mu$-a.e. universally holds (to see this, observe that if $\|f\|_{\infty} < \infty$ is given, $\{ x : |f| > \|f\|_{\infty}\} = \bigcup_{n=1}^{\infty}\{|f| > \|f\|_{\infty} + 1/n\}$ is a set of measure zero). Now, by identifying two functions $f, g \in \mathcal{L}^{p}(\mu)$ by the equivalence relation $f \sim g \Leftrightarrow f = g\ \text{$\mu$-a.e.}$, we obtain a quotient space $L^{p}(\mu) := \mathcal{L}^{p}(\mu)/\sim$. For simplicity, it is customary to denote an element of $L^{p}(\mu)$ by its representative $f \in \mathcal{L}^{p}(\mu)$ whenever there is no risk of confusion. For $f \in L^{p}(\mu)$, one finds that the quantity $\|f\|_{p}$, $1 \leq p \leq \infty$ is well-defined (irrespective of the choice of the representative), and that this in fact provides a norm on $L^{p}(\mu)$, called the $L^{p}$-norm. The norm $\|\cdot\|_{p}$ is also known to be complete and hence makes $L^{p}(\mu)$ into a Banach space. The case $p=2$ is of particular interest in the context of quantum mechanics, where the integration,
\begin{equation}\label{def:L2-inner_product}
\langle g, f \rangle := \int g^{*}f\ d\mu,
\end{equation}
defines an inner product that satisfies $\langle f, f \rangle = \|f\|_{2}^{2}$, making $L^{2}(\mu)$ into a Hilbert space.
As a special case, we are mostly interested in the choice $(\mathbb{R}^{n}, \mathfrak{B}^{n}, \beta^{n})$ of the measure space. Conforming to convention in physical literature, we prepare a special symbol for the $L^{p}$ spaces of it and denote $L^{p}(\mathbb{R}^{n}) := L^{p}(\beta^{n})$.
\paragraph{H{\"o}lder's Inequality}
Among the most important inequality regarding $L^{p}$-spaces is the H{\"o}lder's inequality.
\begin{theorem}[H{\"o}lder's Inequality]
Let $1 \leq p,q \leq \infty$, $\frac{1}{p} + \frac{1}{q} = 1$, where we understand $1/\infty :=0$, and let $f, g : X \to \hat{\mathbb{K}}$ be measurable. Then,
\begin{equation}
\|fg\|_{1} \leq \|f\|_{p}\|g\|_{q}
\end{equation}
holds.
\end{theorem}
\noindent
For the specific choice $p,q = 2$, the resulting inequality has its own name as the Cauchy-Schwarz Inequality.
\subsubsection{Rudimentary Techniques in handling the CCR}
While the contents of the following topics are widely known, we include this material mainly for reader's convenience, and also for self-consistency and reference.
\paragraph{Schr{\"o}dinger Representation of the CCR}
We start by recalling the definition of the Schwartz space.
A function $f :\mathbb{R}^{n} \to \mathbb{K}$ is called \emph{rapidly decreasing} when
\begin{equation}
\lim_{|x| \to \infty} x^{\gamma}f(x) = 0
\end{equation}
holds for any $\gamma := (\gamma_{1}, \dots, \gamma_{n})$ with $\gamma \in \mathbb{N}_{0}^{n}$. Here, the multi-index symbol $\gamma\in \mathbb{N}^{n}_{0}$ is understood to be used as
\begin{equation}\label{def:use_alpha}
x^{\gamma} := x_{1}^{\gamma_{1}} \cdots x_{n}^{\gamma_{n}}, \quad D^{\gamma} := (D_{1})^{\gamma_{1}} \cdots (D_{n})^{\gamma_{n}},
\end{equation}
where $D_{i} := \partial/\partial x_{i}$ is the partial differentiation operator with respect to the variable $x_{i}$.
The space
\begin{equation}
\mathscr{S}(\mathbb{R}^{n}) := \{f \in C^{\infty}(\mathbb{R}^{n}): D^{\gamma}f\ \text{is rapidly decreasing},\ \gamma \in \mathbb{N}^{n}_{0} \},
\end{equation}
is then called the \emph{Schwartz space}, and its elements are in turn called \emph{Schwartz functions}. The Schwartz space is known to be a dense subspace $\mathscr{S}(\mathbb{R}^{n}) \subset L^{p}(\mathbb{R}^{n})$ for $1 \leq p < \infty$. A well-known example of Schwartz functions is provided by the form,
\begin{equation}
x^{\gamma}e^{-a|x|^{2}} \in \mathscr{S}(\mathbb{R}^{n}), \quad \gamma \in \mathbb{N}^{n}_{0},\ \, a>0.
\end{equation}
Specifically, the Gaussian wave-functions, which also appear later in our analysis, are among the most oft-used members of the Schwartz space belonging to this class.
Now that we have the necessary definitions, we return to the main topic of this subsection and, for simplicity, confine ourselves to the case $n=1$ without loss of generality.
We start by introducing a pair of important operators $\hat{x}$ and $\hat{p}$ on the Hilbert space $L^{2}(\mathbb{R})$.
Among these, $\hat{x}: \mathrm{dom}(\hat{x}) \to L^{2}(\mathbb{R})$ is an operator on $L^{2}(\mathbb{R})$ defined by the multiplication of $x$ on a function $f$,
\begin{equation}\label{def:position_operator}
\hat{x} : f(x) \mapsto xf(x),
\end{equation}
with its domain,
\begin{equation}
\mathrm{dom}(\hat{x}) := \{ f \in L^{2}(\mathbb{R}) : xf \in L^{2}(\mathbb{R}) \}.
\end{equation}
The operator $\hat{x}$ is known to be self-adjoint and is called the (one-dimensional) \emph{position operator}.
Next, consider the operator $-i D$ defined on $\mathscr{S}(\mathbb{R})$ with $D := d/dx$ being the usual differential operator in our case $n=1$.
The operator $-iD : \mathscr{S}(\mathbb{R}) \to L^{2}(\mathbb{R})$ is known to be essentially self-adjoint,
which allows us to define the (one-dimensional) \emph{momentum operator} by its self-adjoint extension%
\footnote{While the explicit identification of the domain of the operator $\hat{p}$ is not quite straightforward, we mention that it is given by
\begin{equation}\label{eq:domain_p}
\mathrm{dom}(\hat{p}) = \left\{ f \in L^{2}(\mathbb{R}) : f|_{J} \in \mathrm{AC}(J) \text{ for all compact sub-intervals } J \subset \mathbb{R},\ \frac{df}{dx} \in L^{2}(\mathbb{R}) \right\},
\end{equation}
where $f|_{J}$ denotes the restriction of the function $f$ on the interval $J$, and $\mathrm{AC}(J)$ denotes the space of all absolutely continuous functions on $J$. Here, a function $f: [a,b] \to \mathbb{K}$ is called \emph{absolutely continuous}, if for every $\epsilon > 0$, there exists a $\delta > 0$ such that
\begin{equation}
\sum_{k=1}^{n}(b_{k} - a_{k}) < \delta \quad \Rightarrow \quad \sum_{k=1}^{n}|f(b_{k}) - f(a_{k})| < \epsilon
\end{equation}
holds for arbitrary partitions $a \leq a_{1} < b_{1} \leq a_{2} < b_{2} \leq \dots \leq a_{n} < b_{n} \leq b$, $n \in \mathbb{N}^{\times}$ of the interval $[a,b]$.
It is known that a function $f: [a,b] \to \mathbb{K}$ is absolutely continuous if and only if $f$ is differentiable almost everywhere (hence $df/dx$ in \eqref{eq:domain_p} is well-defined), its derivative is Lebesgue integrable $df/dx \in L^{1}(J)$, and that
\begin{equation}
f(t) - f(s) = \int_{s}^{t} \frac{df}{dx}\ dx, \quad s, t \in [a,b],\ s \leq t
\end{equation}
holds ({\it cf.} fundamental theorem of calculus).
},
\begin{equation}\label{def:momentum_operator}
\hat{p} := \overline{-i D}.
\end{equation}
Here, the overline on a closable operator denotes its closure, which in the case of an essentially self-adjoint operator is equivalent to its (unique) self-adjoint extension.
One then verifies that the pair $\{\hat{x}, \hat{p}\}$ satisfies the familiar (one-dimensional) \emph{canonical commutation relations (CCR)},
\begin{gather}
[\hat{x}, \hat{p}] = iI,
\label{eq:CCR_01} \\
[\hat{x}, \hat{x}] = 0, \quad [\hat{p}, \hat{p}] = 0,
\label{eq:CCR_02}
\end{gather}
on the subspace $\mathscr{S}(\mathbb{R}) \subset L^{2}(\mathbb{R})$, where $I$ denotes the identity operator and $[X, Y] := XY - YX$ denotes the commutator for operators $X, Y$, whose domain is understood to be $\mathrm{dom}([X,Y]) := \mathrm{dom}(XY) \cap \mathrm{dom}(YX)$.
In general, let $\{\mathcal{H}, \mathcal{D}, \{Q,P\}\}$ be a combination consisting of a Hilbert space $\mathcal{H}$, its dense subspace $\mathcal{D} \subset \mathcal{H}$, and a pair of self-adjoint operators $\{Q,P\}$ on $\mathcal{H}$. We say that $\{\mathcal{H}, \mathcal{D}, \{Q,P\}\}$ is a (one-dimensional) \emph{representation of the CCR}, if the CCR
\begin{gather}
\label{eq:ccr}
[Q, P] = iI, \\
\label{eq:ccr2}
[Q, Q] = 0, \quad [P, P] = 0
\end{gather}
hold on the domain $\mathcal{D}$ fulfilling
\begin{equation}
\label{eq:weyl_subspace}
\mathcal{D} \subset \mathrm{dom}(QQ) \cap \mathrm{dom}(QP) \cap \mathrm{dom}(PQ) \cap \mathrm{dom}(PP).
\end{equation}
One then concludes from the above argument that the combination,
\begin{equation}
\left\{ L^{2}(\mathbb{R}), \mathscr{S}(\mathbb{R}), \{\hat{x}, \hat{p}\}\right\},
\end{equation}
gives a concrete example for the representation of the CCR, called the (one-dimensional) \emph{Schr{\"o}dinger representation of the CCR}.
\paragraph{Weyl Representation of the CCR}
We call a combination $\{\mathcal{H}, \{Q, P\}\}$ consisting of a Hilbert space $\mathcal{H}$ and a pair of self-adjoint operators $\{Q, P\}$, a (one-dimensional) \emph{Weyl representation of the CCR}, if $\{Q, P\}$ satisfies the \emph{Weyl relations}:
\begin{equation}\label{eq:weyl01}
e^{isQ}e^{itP}=e^{-istI}e^{itP}e^{isQ},
\end{equation}
\begin{equation}\label{eq:weyl02}
e^{isQ}e^{itQ}=e^{itQ}e^{isQ}, \quad e^{isP}e^{itP}=e^{itP}e^{isP},
\end{equation}
for $s, t \in \mathbb{R}$.
One of the advantages of the Weyl relations, as compared to the CCR, is that they deal only with unitary operators, for which no particular consideration for the domain of the involved operators is necessary because of their boundedness. Fortunately, in the present case one can actually prove that the pair $\{\hat{x}, \hat{p}\}$ of the position and momentum operators introduced earlier satisfy the Weyl relations \eqref{eq:weyl01} and \eqref{eq:weyl02} on $L^{2}(\mathbb{R})$. This implies that
the Schr{\"o}dinger representation of the CCR $\{L^{2}(\mathbb{R}), \mathscr{S}(\mathbb{R}), \hat{x}, \hat{p}\}$ furnishes an example of the Weyl representation of the CCR, at least in the case of the configuration space $\mathbb{R}$. One also finds that this is true for the Euclidean configuration space $\mathbb{R}^n$.
One may naturally be interested in how the Weyl representation of the CCR relates to the standard representation of the CCR. To this end, we first begin by collecting some of the necessary definitions and basic theorems.
Recall that a vector-valued map $F: U \to V$ from an open subset $U \subset \mathbb{R}$ to a normed space $V$ is called \emph{strongly continuous} at $t_{0} \in U$ if
\begin{equation}
\lim_{u \to 0} \|F(u + t_{0}) - F(t_{0})\| = 0
\end{equation}
with respect to the norm $\|\cdot \|$ on $V$,
and in turn, strongly continuous on $U$ if it is strongly continuous at every point of $U$.
The map $F$ is then called \emph{strongly differentiable} at $t_{0} \in U$ with strong derivative $F^{\prime}(t_{0}) \in V$ if
\begin{equation}
\lim_{u \to 0} \left\| \frac{F(u + t_{0}) - F(t_{0})}{u} - F^{\prime}(t_{0}) \right\| = 0
\end{equation}
holds, and accordingly strongly differentiable on $U$ if it is strongly differentiable at every point of $U$.
We will occasionally write its strong derivative in either of the notations,
\begin{equation}
\frac{dF(t_{0})}{dt} = \frac{dF}{dt}(t_{0}) = \left. \frac{d}{dt} F(t) \right|_{t=t_{0}} := F^{\prime}(t_{0}).
\end{equation}
Now, let $A : \mathcal{H} \supset \mathrm{dom}(A) \to \mathcal{H}$ be a self-adjoint operator on a Hilbert space $\mathcal{H}$, and consider a one-parameter unitary group $\{e^{itA}\}_{t \in \mathbb{R}}$ (defined by means of functional calculus).
Then, Stone's theorem on one-parameter unitary groups states that, on account of the boundedness of the unitary operator,
for a fixed $|\phi\rangle \in \mathcal{H}$
the unitary group yields a strongly continuous vector-valued map,
\begin{equation}
F : t \mapsto e^{itA}|\phi\rangle, \quad t \in \mathbb{R},
\end{equation}
for any self-adjoint operator $A$.
However, consideration of the domain $\mathrm{dom}(A)$ becomes necessary when
differentiation of the map is considered. In fact, the map is strongly differentiable on $\mathbb{R}$ if and only if $|\phi\rangle \in \mathrm{dom}(A)$, in which case the derivative reads
\begin{equation}
\frac{dF}{dt}(t) = ie^{itA}A|\phi\rangle = iAe^{itA}|\phi\rangle.
\end{equation}
Returning to our main topic, we rewrite the r.~h.~s. of \eqref{eq:weyl01} to obtain
\begin{equation}
e^{isQ}e^{itP} = e^{itP}e^{is(Q - tI)}, \quad s, t \in \mathbb{R}.
\end{equation}
Considering the vector-valued map,
\begin{equation}\label{eq:weyl_func}
s \mapsto e^{isQ}e^{itP}|\psi\rangle = e^{itP}e^{is(Q - tI)}|\psi\rangle,
\end{equation}
for a fixed $|\psi\rangle \in \mathrm{dom}(Q)= \mathrm{dom}(Q - tI)$, $t \in \mathbb{R}$, one concludes from the above argument that the r.~h.~s. of \eqref{eq:weyl_func} is strongly differentiable at all $s \in \mathbb{R}$ with the derivative
\begin{align}
\frac{d}{ds} \left( e^{itP}e^{is(Q - tI)}|\psi\rangle \right)
&= e^{itP} \left( \frac{d}{ds}e^{is(Q - tI)}|\psi\rangle \right) \nonumber \\
&= ie^{itP}e^{is(Q - tI)}(Q - tI)|\psi\rangle, \quad s, t \in \mathbb{R}.
\end{align}
Note here that the first equality follows from the linearity and boundedness (hence, continuity) of the unitary operator $e^{itP}$.
Turning to the l.~h.~s. of \eqref{eq:weyl_func}, differentiability implies that $e^{itP}|\psi\rangle \in \mathrm{dom}(Q)$, whereby
one has
\begin{equation}
\frac{d}{ds} \left( e^{isQ}e^{itP}|\psi\rangle \right) = ie^{isQ}Qe^{itP}|\psi\rangle.
\end{equation}
Combining the two results, one duly obtains
\begin{equation}\label{eq:weyl_diff}
e^{isQ}Qe^{itP}|\psi\rangle = e^{itP}e^{is(Q-tI)} (Q-tI) |\psi\rangle, \quad s, t \in \mathbb{R}.
\end{equation}
Taking $s=0$, one finds the validity of the operator identity $Q e^{itP} = e^{itP}(Q - tI)$, or equivalently
\begin{equation}\label{eq:weak_weyl}
e^{-itP}Q e^{itP} = Q - tI, \quad t \in \mathbb{R},
\end{equation}
on the subspace $\mathrm{dom}(Q)$. This shows how the unitary adjoint action generated by $P$ results in a parallel translation $Q \mapsto Q - tI$ on its conjugate operator $Q$%
\footnote{Note that what we are discussing here is something more than just proving the Campbell-Baker-Hausdorff formula.}.
Now, if one further considers the vector-valued map by rewriting \eqref{eq:weak_weyl},
\begin{equation}\label{eq:diff_2_start}
Q e^{itP}|\psi\rangle = e^{itP}Q|\psi\rangle - te^{itP}|\psi\rangle,\quad t \in \mathbb{R},
\end{equation}
one proves the differentiability of the r.~h.~s. for the choice of the initial state $|\psi\rangle \in \mathrm{dom}(PQ) \cap \mathrm{dom}(P)$, which yields
\begin{equation}
\frac{d}{dt} \left( e^{itP}Q|\psi\rangle - te^{itP}|\psi\rangle \right) = (ie^{itP}PQ - e^{itP} - ite^{itP}P ) |\psi\rangle, \quad t \in \mathbb{R}.
\end{equation}
Turning to the l.~h.~s., differentiability also leads to
\begin{align}
\frac{d}{dt} \left( Q e^{itP}|\psi\rangle \right)
&= Q \left( \frac{d}{dt} e^{itP}|\psi\rangle \right) \nonumber \\
&= iQ e^{itP}P|\psi\rangle, \quad t \in \mathbb{R},
\end{align}
where, in particular, $e^{itP}P|\psi\rangle \in \mathrm{dom}(Q)$ is implied, and the first equality is due to the closedness of the operator $Q$ (recall that a self-adjoint operator is necessarily closed). By combining the above two results, one has
\begin{equation}
\left(iQ e^{itP}P\right)|\psi\rangle = \left(ie^{itP}PQ - e^{itP} - ite^{itP}P\right)|\psi\rangle, \quad t \in \mathbb{R}.
\end{equation}
Taking $t=0$, we learn that this in particular leads to the operator identity,
\begin{equation}
QP = PQ + iI \quad \Leftrightarrow \quad [Q, P] = iI,
\end{equation}
on the subspace $\mathrm{dom}(PQ) \cap \mathrm{dom}(P)$. One also sees from this result that the choice $|\psi\rangle \in \mathrm{dom}(PQ) \cap \mathrm{dom}(P)$ automatically implies $|\psi\rangle \in \mathrm{dom}(QP)$.
Proceeding further from \eqref{eq:weyl02} by analogous reasoning, one eventually obtains the CCR
(\ref{eq:ccr}) and (\ref{eq:ccr2}) on the domain (\ref{eq:weyl_subspace}).
In the case where $\mathcal{D}$ is dense, one sees that a Weyl representation of the CCR $\{\mathcal{H}, \{Q,P\}\}$ together with the subspace $\mathcal{D}$ indeed gives a representation of the CCR. In fact, in the case where $\mathcal{H}$ is separable, $\mathcal{D}$ is known to be dense.
In passing, we mention that the importance of the Weyl relations becomes evident when one considers configuration spaces,
other than the Euclidean space $\mathbb{R}^n$, where no reasonable counterpart of the CCR can be defined.
For instance, when the configuration space is given by a coset space $G/H$ where $G$ is a Lie group and $H$ its subgroup (typical examples being the spheres $S^n \simeq O(n+1)/O(n)$), one can readily adopt the inherent group theoretic structure of the configuration space to define the Weyl relations extended to the space.
Unlike the Euclidean case, such extended Weyl relations are known to admit a multiple of inequivalent representations.
\subsubsection{Tensor Product of Hilbert Spaces and Self-adjoint Operators}
We finally provide a brief review on tensor products of Hilbert spaces and those of self-adjoint operators. Although the topic is elementary, we find it beneficial to give a summary of its precise definition in consideration of its extensive use due to the nature of this paper focusing on indirect measurement schemes.
\paragraph{Algebraic Tensor Products}
Let $V, W$ be $\mathbb{K}$-vector spaces. We call an ordered pair
\begin{equation}
(V \otimes W,\, \otimes)
\end{equation}
consisting of a vector space $V \otimes W$ and a bilinear map $\otimes: V \times W \to V \otimes W$, an (algebraic) \emph{tensor product} of vector spaces $V$ and $W$, if for any $\mathbb{K}$-vector space $Z$ and a bilinear map $T: V \times W \to Z$, there exists a unique linear map $\widetilde{T}: V \otimes W \to Z$ for which the diagram
\begin{equation}
\xymatrix{
V \times W
\ar[rd]_{T}
\ar[r]^{\otimes}
& V \otimes W
\ar[d]^{\widetilde{T}}\\
& Z
}
\end{equation}
commutes%
\footnote{
We say that a diagram is a \emph{commutative diagram}, or more casually, \emph{the diagram commutes}, if all directed paths in the diagram with the same start and endpoints lead to the same result by composition.
} (universal property of (algebraic) tensor products).
Each element of $V \otimes W$ is called a \emph{tensor}, and the bilinear map $\otimes$ is called the \emph{tensor map}, the image of which
is denoted by
\begin{equation}
v \otimes w := \otimes(v,w).
\end{equation}
The thus defined tensor products are in fact unique up to isomorphism. Indeed if $(V \otimes W,\, \otimes)$ and $(V^{\prime} \otimes^{\prime} W^{\prime},\, \otimes^{\prime})$ were two of such, then by first letting $Z = V^{\prime} \otimes W^{\prime}$ and $T = \otimes^{\prime}$ in the above diagram, and then subsequently by changing roles of $(V \otimes W,\, \otimes)$ and $(V^{\prime} \otimes^{\prime} W^{\prime},\, \otimes^{\prime})$, one concludes that $\widetilde{\otimes}$ and $\widetilde{\otimes^{\prime}}$ are linear bijections with $\widetilde{\otimes^{\prime}} \circ \widetilde{\otimes} = I$. In this sense, we may refer to $(V \otimes W,\, \otimes)$ as \emph{the} tensor product of $V$ and $W$, and forget about the way how it is constructed%
\footnote{One finds several concrete constructions of tensor products in various literatures. See, for example \cite{Roman_2008}.}.
One of the basic facts worth of special note is that, given two bases $\{e_{i}\}_{i \in I}$ and $\{f_{j}\}_{j \in J}$ of $V$ and $W$, respectively, the tensors $\{e_{i} \otimes f_{j}\}_{i \in I, j \in J}$ form a basis of $V \otimes W$.
\paragraph{Tensor Product of Hilbert Spaces}
We are specifically interested in tensor products of Hilbert spaces. For a pair of Hilbert spaces $(\mathcal{H}_{1}, \langle \,\cdot\, , \,\cdot\, \rangle_{\mathcal{H}_{1}})$ and $(\mathcal{H}_{2}, \langle \,\cdot\, , \,\cdot\, \rangle_{\mathcal{H}_{2}})$, we denote by
\begin{equation}
(\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}, \,\widehat{\otimes}\,)
\end{equation}
their algebraic tensor product defined from their purely algebraic structures described as above. We then introduce
\begin{equation}
\langle \phi_{1} \,\widehat{\otimes}\, \phi_{2},\, \psi_{1} \,\widehat{\otimes}\, \psi_{2} \rangle := \langle \phi_{1}, \psi_{1} \rangle_{\mathcal{H}_{1}} \langle \phi_{2}, \psi_{2} \rangle_{\mathcal{H}_{2}},\quad \phi_{i}, \psi_{i} \in \mathcal{H}_{i}
\end{equation}
defined for pairs of all tensors of the form $D := \{v \otimes w : v \in V, w \in W\}$,
and let it extend linearly on whole $\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2} = \mathrm{Span} (D)$. Here,
\begin{equation}\label{def:lin_span}
\mathrm{span} (S) := \{k_{1}v_{1} + \cdots + k_{n}v_{n} : k_{i} \in \mathbb{K}, v_{i} \in S, n \in \mathbb{N}\}
\end{equation}
denotes the subspace of a $\mathbb{K}$-vector space $V$ spanned by a nonempty set $S \subset V$, {\it i.e.} the set of all finite linear combinations of vectors belonging to $S$. It is routine to check that the thus defined extension $\langle \,\cdot\,, \,\cdot\, \rangle_{\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}}$ is well-defined, and one moreover proves that the extension in fact makes itself an inner product on $\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}$, making the pair $(\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2},\, \langle \,\cdot\,, \,\cdot\, \rangle_{\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}})$ into a pre-Hilbert space ({\it i.e.}, an inner product space). The tensor map $\,\widehat{\otimes}\,$ can be also shown to be continuous with respect to the topology that the inner product generates. We then finally define the completion of the pre-Hilbert space, and denote it by
\begin{equation}
(\mathcal{H}_{1} \otimes \mathcal{H}_{2},\, \langle \,\cdot\,, \,\cdot\, \rangle_{\mathcal{H}_{1} \otimes \mathcal{H}_{2}}).
\end{equation}
The new space $(\mathcal{H}_{1} \otimes \mathcal{H}_{2},\, \langle \,\cdot\,, \,\cdot\, \rangle_{\mathcal{H}_{1} \otimes \mathcal{H}_{2}})$ is a Hilbert space by construction, and together with the continuous extension $\otimes$ of the bilinear map $\,\widehat{\otimes}\,$, is called the (topological) \emph{tensor product of the Hilbert spaces} $\mathcal{H}_{1}$ and $\mathcal{H}_{2}$.
The map $\otimes$ is called the \emph{tensor map} and the elements of $\mathcal{H}_{1} \otimes \mathcal{H}_{2}$ are called \emph{tensors}.
\begin{comment}
Let $\mathcal{H}_{1}, \mathcal{H}_{2}$ be Hilbert spaces, and let $(\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}, \,\widehat{\otimes}\,)$ denote their algebraic tensor product, which is defined as a unique (up to isomorphism) pair of an algebraic linear space $\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}$ and a bilinear map $\,\widehat{\otimes}\,: \mathcal{H}_{1} \times \mathcal{H}_{2} \to \mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}$ with the property that, for any bilinear map $T: \mathcal{H}_{1} \times \mathcal{H}_{2} \to \mathcal{K}$ there exists a unique linear map $\widetilde{T}: \mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2} \to \mathcal{K}$ for which the diagram
\begin{equation}
\xymatrix{
\mathcal{H}_{1} \times \mathcal{H}_{2}
\ar[rd]_{T}
\ar[r]^{\,\widehat{\otimes}\,}
& \mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}
\ar[d]^{\widetilde{T}}\\
& \mathcal{K}
}
\end{equation}
commutes (universal property of tensor products).
\end{comment}
\begin{comment}
A sesquilinear form on $\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}$ induced by those on $\mathcal{H}_{1}, \mathcal{H}_{2}$ through
\begin{equation}
\langle \psi_{1} \,\widehat{\otimes}\, \psi_{2}, \phi_{1} \,\widehat{\otimes}\, \phi_{2} \rangle := \langle \psi_{1}, \phi_{1} \rangle_{\mathcal{H}_{1}} \langle \psi_{2}, \phi_{2} \rangle_{\mathcal{H}_{2}}
\end{equation}
defines a natural inner product on $\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}$, and thus makes $(\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}, \langle \cdot, \cdot \rangle)$ a inner product space. We denote by $(\mathcal{H}_{1} \otimes \mathcal{H}_{2}, \langle \cdot, \cdot \rangle)$ the completion of $(\mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}, \langle \cdot, \cdot \rangle)$ with respect to the inner product, and together with the natural extension $\otimes$ of the bilinear map $\,\widehat{\otimes}\,$, is called the (topological) \emph{tensor product of the Hilbert spaces} $\mathcal{H}_{1}$ and $\mathcal{H}_{2}$. Note that $(\mathcal{H}_{1} \otimes \mathcal{H}_{2}, \langle \cdot, \cdot \rangle)$ is a Hilbert space in itself by definition, and is sometimes referred to by itself as the tensor product. The map $\otimes$ is called the \emph{tensor map} and the elements of $\mathcal{H}_{1} \otimes \mathcal{H}_{2}$ are called \emph{tensors}.
\end{comment}
\paragraph{Tensor Product of Linear Operators}
A pair of linear operators $A_{i}: \mathcal{H}_{i} \supset \mathrm{dom}(A_{i}) \to \mathcal{H}_{i}$, $i= 1, 2$, defines
a natural bilinear map
\begin{equation}
A_{1} \times A_{2}: \mathrm{dom}(A_{1}) \times \mathrm{dom}(A_{1}) \to \mathcal{H}_{1} \times \mathcal{H}_{2}, \quad (|\phi_{1}\rangle, |\phi_{2}\rangle) \mapsto (A_{1}|\phi_{1}\rangle,\, A_{2}|\phi_{2}\rangle).
\end{equation}
From the universal property of the algebraic tensor product mentioned above, one readily sees the existence of a unique linear map
\begin{equation}\label{eq:pre_pre_tensor_product_of_operators}
A_{1} \,\widehat{\otimes}\, A_{2}: \mathrm{dom}(A_{1}) \,\widehat{\otimes}\, \mathrm{dom}(A_{1}) \to \mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}
\end{equation}
that makes the diagram
\begin{equation}
\xymatrix{
\mathrm{dom}(A_{1}) \times \mathrm{dom}(A_{2})
\ar[d]_{A_{1} \times A_{2}}
\ar[r]^{\,\widehat{\otimes}\,}
\ar@{.>}[dr]|\circlearrowleft
& \mathrm{dom}(A_{1}) \,\widehat{\otimes}\, \mathrm{dom}(A_{2})
\ar[d]^{A_{1} \,\widehat{\otimes}\, A_{2}}\\
\mathcal{H}_{1} \times \mathcal{H}_{2}
\ar[r]^{\,\widehat{\otimes}\,}
& \mathcal{H}_{1} \,\widehat{\otimes}\, \mathcal{H}_{2}
}
\end{equation}
commute. Note in particular that the diagram implies
\begin{equation}
A_{1} \,\widehat{\otimes}\, A_{2}(|\phi_{1}\rangle \,\widehat{\otimes}\, |\phi_{2}\rangle) = |A_{1}\phi_{1}\rangle \,\widehat{\otimes}\, |A_{2}\phi_{2}\rangle, \quad |\phi_{i} \rangle \in \mathrm{dom}(A_{i}),\ i=1,2.
\end{equation}
Extending both the domain and the range of \eqref{eq:pre_pre_tensor_product_of_operators}, we can think of
\begin{equation}\label{eq:pre_tensor_product_of_operators}
A_{1} \,\widehat{\otimes}\, A_{2} : \mathcal{H}_{1} \otimes \mathcal{H}_{2} \supset \mathrm{dom}(A_{1}) \,\widehat{\otimes}\, \mathrm{dom}(A_{2}) \to \mathcal{H}_{1} \otimes \mathcal{H}_{2}
\end{equation}
as an operator on the Hilbert space $\mathcal{H}_{1} \otimes \mathcal{H}_{2}$.
\begin{comment}
Recall that the \emph{graph} of a linear map $T: \mathcal{H} \supset \mathrm{dom}(T) \to \mathcal{K}$ is the subspace defined by
\begin{equation}
\mathrm{gr}(T) := \{(|\phi\rangle, T|\phi\rangle) : |\phi\rangle \in \mathrm{dom}(T) \} \subset \mathcal{H} \times \mathcal{K}.
\end{equation}
$B$ is called \emph{closed}, when its graph $\mathrm{gr}(T)$ is a closed subset of $\mathcal{H} \times \mathcal{K}$ with respect to the product topology. $B$ is in turn called \emph{closable} when there exists a closed extension $\widetilde{T}$ of $B$, in the sense that $\widetilde{T}$ is closed and $\mathrm{gr}(T) \subset \mathrm{gr}(\widetilde{T}) \Leftrightarrow T \subset \widetilde{T}$. For closable $B$, the smallest closed extension is called the \emph{closure} of $B$, and is denoted by $\overline{T}$.
\end{comment}
\paragraph{Tensor Product of Self-Adjoint Operators}
Now, for a pair of densely defined closable operators $A_{i}: \mathcal{H}_{i} \supset \mathrm{dom}(A_{i}) \to \mathcal{H}_{i}$, $i= 1, 2$, the operator \eqref{eq:pre_tensor_product_of_operators} itself is known to be closable, whereby we define the \emph{tensor product}
\begin{equation}\label{def:tensor_operator}
A_{1} \otimes A_{2} := \overline{A_{1} \,\widehat{\otimes}\, A_{2}}
\end{equation}
of the pair by its closure. Specifically, since self-adjoint operators are densely defined and closed, the tensor product \eqref{def:tensor_operator} is always well-defined.
Although self-adjointness is not preserved in general by taking \eqref{eq:pre_tensor_product_of_operators}, its essential self-adjointness is at least known to be guaranteed.
As the closure of an essentially self-adjoint operator, this makes the tensor product \eqref{def:tensor_operator} itself self-adjoint, which is precisely the definition of the \emph{tensor product of self-adjoint operators}.
\subsection{Unconditioned Measurement}\label{sec:ups_I_ups}
Now that we have reviewed the necessary materials, we begin our study on the unconditioned measurement scheme. Suppose that the experimenter wishes to extract information of the combination of a given but unknown observable $A$ and a state $|\phi\rangle \in \mathcal{H}$ of the target system, without direct access to it. To accomplish this,
one first arranges an auxiliary meter system $\mathcal{K}$ equipped with a pair of observables $\{Q, P\}$ for which $\{\mathcal{K}, \{Q, P\}\}$ gives a Weyl representation of the CCR.
As we have seen above, the choice $\mathcal{K} = L^{2}(\mathbb{R})$, $Q = \hat{x}$ and $P = \hat{p}$ gives a concrete example.
One then prepares the meter system in a certain initial state represented by the vector $|\psi\rangle \in \mathcal{K}$, and combines the two systems into the direct product state $|\phi \otimes \psi \rangle \in \mathcal{H} \otimes \mathcal{K}$. Choosing an observable $Y$ of the meter system $\mathcal{K}$
either by $Y = Q$ or $Y = P$, the composite system is subjected to a von Neumann type interaction,
\begin{equation}\label{intro:von_Neumann_interaction}
|\Psi^{g}\rangle := e^{-igA \otimes Y} |\phi \otimes \psi\rangle, \quad g \in \mathbb{R},
\end{equation}
{\it i.e.}, a unitary evolution on the composite system parametrised by a real number $g$, which is often interpreted as the intensity, its time duration, or the combination thereof, of the interaction between the two systems.
Finally, the experimenter performs local measurement of an observable $X$ of the meter system $\mathcal{K}$ by choosing either by $X = Q$ or $X = P$ (chosen independently of $Y$), or equivalently $I \otimes X$ on the generally entangled composite state $|\Psi^{g}\rangle$ after the interaction (see figure~\ref{fig:ucm}).
\begin{figure}
\floatbox[{\capbeside\thisfloatsetup{capbesideposition={right,top},capbesidewidth=0.7\textwidth}}]{figure}[\FBwidth]
{\caption{A graphical illustration of the unconditioned measurement scheme. The figure is to be read from top to bottom. The initial state preparation stage of both the target and the meter systems is depicted in the top part, and the manner in which the two quantum systems undergoes a von Neumann type interaction is illustrated in the middle part. The composite system after the interaction, which is depicted in the bottom part, generally becomes entangled. One finally performs a measurement of an observable $X$ on the meter system.}\label{fig:ucm}}
{\includegraphics[width=0.3\textwidth]{ucm.eps}}
\end{figure}
As a preparation for further analysis, we first introduce the reduced density operator
\begin{equation}\label{def:meterstate}
\psi^{g} := \mathrm{Tr}_\mathcal{H}[|\Psi^{g}\rangle\langle\Psi^{g}|]
\end{equation}
representing the state of the meter system $\mathcal{K}$ after the measurement%
\footnote{Here we are adopting, instead of the more common usage
$\rho^{g}$, a slightly unusual notation $\psi^{g}$ to denote the generically mixed state of the meter. This we do because
we wish to reserve the letter $\rho$ for the density of some absolutely continuous complex measures (see Section~\ref{sec:absolute_continuity}).
However, our notation has an advantage on its own in that, if we also write the state as $|\psi^{g}\rangle$ when it is pure as we usually do,
the correspondence between the two, $\psi^{g}$ and $|\psi^{g}\rangle$ (both represent the same state), becomes obvious.
}
obtained by taking the partial trace of the composite state $|\Psi^{g}\rangle$ with respect to the target system $\mathcal{H}$. The quantity of interest for our measurement is thus the expectation value
\begin{align}\label{eq:exp_x_outcome}
\mathbb{E}[I \otimes X;\Psi^{g}]
&:= \frac{\langle \Psi^{g}, (I \otimes X) \Psi^{g} \rangle}{\|\Psi^{g}\|^{2}} \nonumber \\
&= \frac{\mathrm{Tr}_{\mathcal{K}}\left[ X \mathrm{Tr}_{\mathcal{H}}\left[ |\Psi^{g}\rangle\langle\Psi^{g}| \right] \right]}{\mathrm{Tr}_{\mathcal{K}}\left[ \mathrm{Tr}_{\mathcal{H}} \left[ |\Psi^{g}\rangle\langle\Psi^{g}| \right] \right]} \nonumber \\
&= \frac{\mathrm{Tr}_{\mathcal{K}}\left[ X \psi^{g} \right] }{\mathrm{Tr}_{\mathcal{K}}\left[ \psi^{g} \right]} \nonumber \\
&=: \mathbb{E}[X;\psi^{g}],
\end{align}
of the observable $I \otimes X$ on the composite state $|\Psi^{g}\rangle$ after the interaction, which can interchangeably be written in terms of that of the local observable $X$ on the density matrix $|\psi^{g}\rangle$ of the meter system.
\paragraph{Main Objective of this Subsection}
The main objective of this subsection is to demonstrate the following basic proposition, which provides the sufficient condition for its well-definedness and its explicit evaluations. For definiteness, we shall from now on fix $Y=P$ without loss of generality.
\begin{proposition}[Unconditioned Measurement I]\label{prop:UCM_I}
In the context of the UM scheme, let $Y=P$ for definiteness. Given the right choices
\begin{enumerate}
\item If $X = Q$: $|\phi\rangle \in \mathrm{dom}(A)$, $|\psi\rangle \in \mathrm{dom}(X)$,
\item If $X = P$: $|\phi\rangle \in \mathcal{H}$, $|\psi\rangle \in \mathrm{dom}(X)$,
\end{enumerate}
of the initial states of both the target and the meter systems,
depending on the choice of the observable $X$ on the meter system to be measured, the composite state after the interaction lies in $|\Psi^{g}\rangle \in \mathrm{dom}(I \otimes X)$, $g \in \mathbb{R}$. The expectation value \eqref{eq:exp_x_outcome} thus remains finite for all range of the interaction parameter, which reads
\begin{equation}\label{prop:UCM_I_formula}
\mathbb{E}[X;\psi^{g}] =
\begin{cases}
\mathbb{E}[Q; \psi] + g\, \mathbb{E}[A; \phi], &\quad (X = Q) \\
\mathbb{E}[P; \psi], &\quad (X = P)
\end{cases}
\end{equation}
for each of the choice of $X$.
\end{proposition}
\paragraph{Some Operator Identities}
Before we move on to the proof, we make notes on some important operator identities that will be extensively used throughout this paper.
Our analysis is based on the following operator identities on the composite Hilbert space $\mathcal{H} \otimes \mathcal{K}$, similar to those of \eqref{eq:weyl01} and \eqref{eq:weyl02}.
\begin{lemma}
Let $H$ and $K$ be Hilbert spaces, and let $A$ be a self-adjoint operator on $\mathcal{H}$, and $\{Q, P\}$ be a pair of self-adjoint operators on $\mathcal{K}$ for which $\{\mathcal{K}, \{Q, P\}\}$ defines a Weyl representation of the CCR. Then, the operator equalities
\begin{gather}
e^{isI \otimes Q}e^{itA \otimes P} = e^{-istA \otimes I}e^{itA \otimes P}e^{isI \otimes Q}, \quad s, t \in \mathbb{R}, \label{eq:weyl_analogue01} \\
e^{isI \otimes P}e^{itA \otimes P} = e^{itA \otimes P}e^{isI \otimes P}, \quad s, t \in \mathbb{R},\label{eq:weyl_analogue02}
\end{gather}
hold.
\end{lemma}
\begin{proof}
Since \eqref{eq:weyl_analogue02} is trivial, we only need to prove \eqref{eq:weyl_analogue01}. To this end, we first consider the special case where the self-adjoint operator $A$ on the target system $\mathcal{H}$ has a spectrum $\sigma(A)$ of finite cardinality. Letting $\sigma(A) = \{a_{1}, \dots, a_{N}\}$, $N \in \mathbb{N}^{\times}$ be any enumeration of its eigenvalues,
the spectral decomposition of $A$ reads
\begin{equation}\label{eq:spect_decomp_fin}
A = \sum_{n = 1}^{N} a_{n} \Pi_{a_{n}},
\end{equation}
where $\Pi_{a_{n}}$ is the projection on the eigenspace associated with the eigenvalue $a_{n}$. In the case where the eigenspace is one-dimensional (or non-degenerate), one may write $\Pi_{a_{n}} = | a_{n} \rangle\langle a_{n} |$ with the eigenstate $| a_{n} \rangle$ for which $A |\phi_{n}\rangle = a_{n} |a_{n}\rangle$ holds (more on the topic of spectral decomposition in Section~\ref{sec:spectral_theorem}).
Now, for an arbitrary self-adjoint operator $Z$ on the meter system $\mathcal{K}$, one may expect from the defining property $\Pi_{a_{n}}^2 = \Pi_{a_{n}}$ of projections that the formal computation
\begin{align}\label{eq:int_op_fin}
e^{itA \otimes Z}
&= \sum_{k = 0}^{\infty} \frac{(it)^{k}}{k!} \left( A \otimes Z \right)^{k} \nonumber \\
&= \sum_{k = 0}^{\infty} \frac{(it)^{k}}{k!} \left( \left( \sum_{n = 1}^{N} a_{n}^{k} \Pi_{a_{n}} \right) \otimes Z^{k} \right) \nonumber \\
&= \sum_{k = 0}^{\infty} \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes \frac{(ita_{n}Z)^{k}}{k!} \right) \nonumber \\
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes e^{ita_{n}Z} \right), \quad t \in \mathbb{R},
\end{align}
is legitimate. This in fact turns out to be correct as an operator identity on $\mathcal{H} \otimes \mathcal{K}$ with full rigour, which can be proven in a fairly straightforward manner by means of rudimentary techniques of functional calculus.
One then has
\begin{align}
e^{isI \otimes Q}e^{itA \otimes P}
&= \left( I \otimes e^{isQ} \right) \left( \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes e^{ita_{n}P} \right) \right) \nonumber \\
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes \left( e^{isQ}e^{ita_{n}P} \right) \right) \nonumber \\
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes \left( e^{ista_{n}}e^{ita_{n}P}e^{isQ} \right) \right) \nonumber \\
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes e^{ita_{n}(P -sI)} \right) \left( I \otimes e^{isQ} \right) \nonumber \\
&= e^{itA \otimes (P-sI)}e^{is I \otimes Q} \nonumber \\
&= e^{-istA \otimes I}e^{itA \otimes P}e^{isI \otimes Q}, \quad s, t \in \mathbb{R},
\end{align}
which proves (\ref{eq:weyl_analogue01}) for our special case, where we have used \eqref{eq:weyl01} in the third step. Returning to the general case in which $A$ is now an arbitrary self-adjoint operator, one observes that the well-definedness of both the left-most and right-most hand sides of the above equality remains valid. From this, one may expect that the same result also holds for the general case, which indeed turns out to be true (as usual, one may prove this without much difficulty through rudimentary techniques of functional calculus).
\end{proof}
\paragraph{Measurement Outcomes}
We now return to the main problem of this subsection. We are interested in finding the condition for which \eqref{eq:exp_x_outcome} is well-defined, and subsequently in obtaining an explicit formula in terms of the components of both the target and the meter system. Since most of the techniques employed here is the same as those introduced in Section~\ref{sec:ups_I_pre}, we shall proceed by sketching the proofs.
\begin{proof}[Proof of Proposition~\ref{prop:UCM_I}]
Let us begin by choosing the operator $X=Q$ for the measurement of the meter system, and thereby rewrite the r.~h.~s. of \eqref{eq:weyl_analogue01} to obtain
\begin{equation}
e^{isI \otimes Q}e^{itA \otimes P} = e^{itA \otimes P}e^{is
(\overline{I \otimes Q - tA \otimes I})},\quad s, t \in \mathbb{R}
\end{equation}
for better usability%
\footnote{Note here that the sum of two (possibly unbounded) self-adjoint operators is not necessarily self-adjoint. Fortunately, essential self-adjointness is at least assured for the sum of $I \otimes Q$ and $A \otimes I$ for our case. We may thus take the self-adjoint extension of their sum in order to ensure its self-adjointness (more to this in Section~\ref{sec:sim_meas_obs}).}.
By differentiating both sides of the above equality and taking $s=0$, an analogous argument given earlier for obtaining \eqref{eq:weak_weyl} leads to the operator identity,
\begin{equation}\label{eq:weak_weyl_analogue01}
e^{-itA \otimes P} (I \otimes Q) e^{itA \otimes P} = \overline{I \otimes Q - tA \otimes I}, \quad t \in \mathbb{R},
\end{equation}
on
the subspace $\mathrm{dom}(\overline{I \otimes Q - tA \otimes I})$. This ensures that, if $|\Phi\rangle \in \mathrm{dom}(\overline{I \otimes Q - tA \otimes I})$, then one has
\begin{equation}\label{eq:eqeta}
e^{itA \otimes P}|\Phi\rangle \in \mathrm{dom}(I \otimes Q).
\end{equation}
Here, it may be worthwhile to note the analogy between \eqref{eq:weak_weyl} and \eqref{eq:weak_weyl_analogue01}.
To put this in our context, let $t = -g$ above. If one chooses $|\psi\rangle \in \mathrm{dom}(Q)$ as the meter state, and likewise assumes $|\phi\rangle \in \mathrm{dom}(A)$ as the system state prepared prior to the interaction, one has in particular $|\phi \otimes \psi \rangle \in \mathrm{dom}(\overline{I \otimes Q + gA \otimes I})$. Then, equating $|\Phi\rangle = |\phi \otimes \psi \rangle$ in \eqref{eq:eqeta}, we find
\begin{equation}\label{eq:int_dom_Q}
|\Psi^{g}\rangle = e^{-igA \otimes P} |\phi \otimes \psi\rangle \in \mathrm{dom}(I \otimes Q), \quad g \in \mathbb{R}.
\end{equation}
This guarantees that the expectation value \eqref{eq:exp_x_outcome} of the observable $I \otimes Q$ on the composite state $|\Psi^{g}\rangle$ remains finite and is given by
\begin{align}\label{eq:ups_exp_Q}
\mathbb{E}[I \otimes Q; \Psi^{g}]
&:= \frac{\langle \Psi^{g}, (I\otimes Q) \Psi^{g}\rangle}{\|\Psi^{g}\|^{2}} \nonumber \\
&= \frac{\langle \phi \otimes \psi, (e^{igA \otimes P} (I \otimes Q) e^{-igA \otimes P}) \phi \otimes \psi\rangle}{\|\phi\|^{2}\|\psi\|^{2}} \nonumber \\
&= \frac{\langle \phi \otimes \psi, (\overline{I \otimes Q + gA \otimes I}) \phi \otimes \psi\rangle}{\|\phi\|^{2}\|\psi\|^{2}} \nonumber \\
&= \mathbb{E}[Q; \psi] + g \mathbb{E}[A; \phi], \quad g \in \mathbb{R},
\end{align}
for any such combination of the initial states.
Evidently, for the choice $X=P$, one finds the validity of the operator identity
\begin{equation}\label{eq:weak_weyl_analogue02}
e^{itA \otimes P} (I \otimes P) e^{-itA \otimes P} = I \otimes P, \quad t \in \mathbb{R},
\end{equation}
on the subspace $\mathrm{dom}(I \otimes P)$ by analogous reasoning.
From this, one readily concludes that the expectation value of $I \otimes P$ reads
\begin{equation}\label{eq:ups_exp_P}
\mathbb{E}[I \otimes P; \Psi^{g}] = \mathbb{E}[P; \psi], \quad g \in \mathbb{R},
\end{equation}
which is well-defined for any choice of the state $|\psi\rangle \in \mathrm{dom}(P)$ of the meter system and $g \in \mathbb{R}$, irrespective of the initial choice of the state $|\phi\rangle \in \mathcal{H}$ of the target system.
\end{proof}
\subsection{Recovery of the Target Profile}
Now that we have revealed the explicit behaviour of the measurement outcomes of the meter, we are thus interested in recovering the information of the target system from it. As one may expect from the statement in Proposition~\ref{prop:UCM_I}, the information of the target system (which should essentially consist of the specification of the pair of $A$ and $|\phi\rangle$) manifests itself in the form of the expectation value $\mathbb{E}[A;\phi]$. In recovering the desired information, one subsequently recognises from \eqref{prop:UCM_I_formula} that it fully suffices to examine only the outcomes of the measurement of the observable $X$ conjugate to $Y$, and there is no use for that of the choice $X=Y$ (this is to be contrasted with the conditional measurement we discuss later). Specifically, one finds below that there are two typical techniques in obtaining the desired information: one is to investigate the behaviour of the measurement outcome \eqref{prop:UCM_I_formula} in the strong region $g \to \pm \infty$ of the interaction parameter, and the other is to examine the local behaviour of it around $g = 0$, which shall be respectively called the \emph{strong unconditioned measurement} and the \emph{weak unconditioned measurement} in this paper.
\subsubsection{Strong Unconditioned Measurement}\label{sec:ups_I_sups}
Our result \eqref{prop:UCM_I_formula} shows that the expectation value of the measurement of $X = Q$ behaves linearly with respect to $g$, and that its growth is proportional to the expectation value $\mathbb{E}[A; \phi]$ of the target observable. The experimenter would thus divide the measurement outcomes of $Q$ by $g$ and then take the limit of the strong coupling $g \to \pm \infty$ (or equivalently ${g^{-1} \to 0}$):
\begin{align}\label{eq:ups_op_recovery}
\lim_{g^{-1} \to 0} \frac{\mathbb{E}[Q; \psi^{g}]}{g}
&= \mathbb{E}[A; \phi] + \lim_{g^{-1} \to 0} \frac{\mathbb{E}[Q; \psi]}{g} \nonumber \\
&= \mathbb{E}[A; \phi].
\end{align}
allowing the recovery of the desired information of the target system $\mathbb{E}[A; \phi]$ in the form of expectation values%
\footnote{Alternatively, one may consider the shift of the expectation value,
\begin{align}
\Delta_{X}(g)
:= \mathbb{E}[X;\psi^{g}] - \mathbb{E}[X;\psi^{0}]
=
\begin{cases}
g \mathbb{E}[A; \phi], & (X=Q), \\
0, & (X=P),
\end{cases}
\end{align}
for $g \in \mathbb{R}$ as a quantity directly related to the observable $A$ of the system.
For the choice $X=Q$, one then simply has
\begin{equation}
\frac{\Delta_{Q}(g)}{g} = \mathbb{E}[A; \phi], \quad g \in \mathbb{R}^{\times},
\end{equation}
which might be a more straight-forward way to be employed practically.}.
\subsubsection{Weak Unconditioned Measurement}\label{sec:ups_I_wups}
The same information may be obtained by examining the weak region ($g \to 0$) of the interaction. Indeed, one trivially finds from \eqref{prop:UCM_I_formula} that
\begin{align}\label{eq:ups_weak_I}
\left. \frac{d^{n}}{dg^{n}}\mathbb{E}[Q; \psi^{g}] \right|_{g=0} =
\begin{cases}
\mathbb{E}[Q; \psi], & n = 0, \\
\mathbb{E}[A; \phi], & n = 1, \\
0, & n \geq 2,
\end{cases}
\end{align}
which implies that the expectation value $\mathbb{E}[A; \phi]$ of our interest may also be obtained as the first differential coefficient ($n=1$) of the measured outcome at $g=0$.
\subsubsection{Discussion}\label{sec:ups_I_discussion}
While this whole section consisted of rather trivial results, the line of arguments presented here serves as the baseline of our analysis throughout this paper. Namely, we first examine the full behaviour of the target of our measurement (for this section, is was the expectation value \eqref{eq:exp_x_outcome} of the observable $X$ of the meter) and intend to obtain an explicit description of how the profile of the initial configuration of the the target system gets mixed into that of the meter system through the interaction (which, for the current case, is the result \eqref{prop:UCM_I_formula}). We then intend to extract the information of the target system (for this section, it is the expectation value $\mathbb{E}[A;\phi]$) by separating it from the measurement outcomes. Specifically, we find that examining either the strong or the weak region of the interaction parameter $g$ reveals itself useful for this purpose, and this should be the strategy that we take in the subsequent sections.
In the next section, the UM scheme is analysed in depth in terms of probabilities, following the same line as described above. Specifically, while the distinction between the strong and the weak measurements looked rather vague at the operator level, we shall see shortly that these two strategies are recognised to be qualitatively different from the viewpoint of probabilities.
\newpage
\section{Unconditioned Measurement II: In Terms of Probabilities}\label{sec:ups_II}
We have so far conducted an analysis of the UM scheme on the operator level, where the quantity of interest is the \emph{expectation value} of an observable.
However, one may be interested in the raw information that the measurement provides, {\it i.e.}, the \emph{probability} describing the behaviour of each measurement outcomes, which is the target of our study in this section.
\subsection{Reference Materials}\label{sec:ups_II_pre}
To prepare for our discussion, we here provide a concise summary on the topic of complex measures and integration with respect to them. We next make a brief review on the spectral theorem for self-adjoint operators and recall the general framework for describing the ideal measurement of a quantum observable. Subsequently, we expound on density functions and see how this relates to the description by measures.
\subsubsection{The Space of Complex Measures}
As a preparation in dealing with the spectral theorem for self-adjoint operators, we collect below the basic definitions and results regarding complex measures and integration with respect to them.
\paragraph{Signed Measures, Jordan Decomposition and Total Variation}
Let $(X, \mathfrak{A})$ be a measurable space. A map $\nu : \mathfrak{A} \to \overline{\mathbb{R}}$ is called a \emph{signed measure}, if it satisfies the following properties:
\begin{enumerate}
\item $\nu(\emptyset) = 0$.
\item $\nu(\mathfrak{A}) \subset ]-\infty, + \infty]$ or $\nu(\mathfrak{A}) \subset [-\infty, + \infty[$.
\item Countable additivity \eqref{def:count_add} holds for any sequence $(A_{n})_{n \geq 1}$ of pairwise disjoint subsets of $X$.
\end{enumerate}
They are, in a sense, generalisations of the concept of the standard measures by allowing negative numbers to be assigned to each measurable sets. A signed measure $\nu$ is called \emph{finite} if $\nu(\mathfrak{A}) \subset \mathbb{R}$. One of the most important properties of a signed measure is described by the \emph{Jordan decomposition theorem}, which states that every singed measure $\nu$ has the \emph{Jordan decomposition}, {\it i.e.}, a unique decomposition of $\nu$ into a difference
\begin{equation}
\nu = \nu^{+} - \nu^{-}
\end{equation}
of two measures $\nu^{+}$ and $\nu^{-}$, respectively called the \emph{positive} and \emph{negative variation} of $\nu$, and at least one of which being finite. Here, the positive and negative variations are \emph{singular} to one another, denoted as $\nu^{+}\perp\nu^{-}$, in the sense there exists a decomposition of $X = P \cup N$ into two measurable sets such that $\nu^{+}(N) = 0$ and $\nu^{-}(P) = 0$ holds. The Jordan decomposition is minimal in the following sense: Given any decomposition $\nu = \rho - \sigma$ of $\nu$ into two measures $\rho$, $\sigma$, at least one of which being finite, then $\nu^{+} \leq \rho$, $\nu^{-} \leq \sigma$ holds.
Let $\mathbf{M}(\mathfrak{A})$ denote the collection of all finite signed measures. One readily sees that $\mathbf{M}(\mathfrak{A})$ becomes an $\mathbb{R}$-linear space, equipped with the natural addition $(\mu + \nu)(A) := \mu(A) + \nu(A)$ and scalar multiplication $(c\mu)(A) := c\mu(A)$ for $\mu, \nu \in \mathbf{M}(\mathfrak{A})$ and $c \in \mathbb{R}$.
Now, let $\nu = \nu^{+} - \nu^{-}$ be the Jordan decomposition of $\nu \in \mathbf{M}(\mathfrak{A})$, and define a new measure by their sum
\begin{equation}
|\nu| := \nu^{+} + \nu^{-},
\end{equation}
called the \emph{variation} of $\nu$. We then define its \emph{total variation} by $\|\nu\| := |\nu|(X)$, which is nothing but the evaluation of the whole space $X$ by the non-negative measure $|\nu|$.
One proves that the total variation defines a norm on $\mathbf{M}(\mathfrak{A})$, and in fact makes $( \mathbf{M}(\mathfrak{A}), \| \cdot \|)$ into a real Banach space.
\paragraph{Complex Measures}
Let $(X, \mathfrak{A})$ be a measurable space. A map $\nu : \mathfrak{A} \to \mathbb{C}$ is called a \emph{complex measure}, when it is countably additive \eqref{def:count_add}.
One sees that $\nu$ is a complex measure if and only if both its real and imaginary parts $\,\mathrm{Re}\,[\nu]$, $\,\mathrm{Im}\,[\nu]$ are finite signed measures. Analogous to the case of signed measures, the collection $\mathbf{M}_{\mathbb{C}}(\mathfrak{A})$ of all complex measures on $(X, \mathfrak{A})$ becomes a $\mathbb{C}$-linear space, equipped with the natural addition and scalar multiplication.
For a complex measure $\nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{A})$, we define the \emph{variation} of a measurable set $A \in \mathfrak{A}$ by
\begin{equation}
|\nu|(A) := \sup \left\{ \sum_{j=1}^{\infty} |\nu(A_{j})| : A_{j} \in \mathfrak{A} \text{ disjoint for} \,\, j \geq 1, \, A = \bigcup_{j=1}^{\infty} A_{j} \right\},
\end{equation}
and also its \emph{total variation},
\begin{equation}\label{def:totvar}
\|\nu\| := |\nu|(X).
\end{equation}
The definition coincides with the previous definition when $\nu$ happens to be a signed measure. The total variation $\|\nu\|$ of $\nu$ is known to be the smallest positive measure $\mu$ on $(X, \mathfrak{A})$ satisfying $|\nu(A)| \leq \mu(A)$, $A \in \mathfrak{A}$. In parallel to the case of signed measures, one finds that the total variation defines a norm on the linear space $\mathbf{M}_{\mathbb{C}}(\mathfrak{A})$ and makes $(\mathbf{M}_{\mathbb{C}}(\mathfrak{A}), \| \cdot \|)$ into a complex Banach space.
\paragraph{Integration over Complex Measures}
It is now tempting to define integration with respect to complex measures, as a natural extension to that defined for (standard) measures. For a complex measure $\nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{A})$, we let $\rho := \,\mathrm{Re}\,[\nu]$, $\sigma := \,\mathrm{Im}\,[\nu]$ and consider the intersection of the spaces
\begin{equation}
\mathcal{L}^{1}(\nu) := \mathcal{L}^{1}(\rho^{+}) \cap \mathcal{L}^{1}(\rho^{-}) \cap \mathcal{L}^{1}(\sigma^{+}) \cap \mathcal{L}^{1}(\sigma^{-}),
\end{equation}
where $\rho^{\pm}$ and $\sigma^{\pm}$ respectively being the positive and negative variations of $\rho$ and $\sigma$.
We then define the Lebesgue integral of $f \in \mathcal{L}^{1}(\nu)$ with respect to $\nu$ by
\begin{equation}
\int_{X} f\ d\nu := \int_{X} f\ d\rho^{+} - \int_{X} f\ d\rho^{-} + i \int_{X} f\ d\sigma^{+} - i \int_{X} f\ d\sigma^{-}.
\end{equation}
Linearity of the Lebesgue integral with respect to the complex measure follows naturally as expected.
\paragraph{New Measure from Old}
There are several ways to construct a new (complex) measure from a given measure. We mention below two of the most important manners that are frequently employed throughout this paper.
\begin{enumerate}
\renewcommand{\labelenumi}{(\Alph{enumi})}
\item {\it Measure with Density. \hspace{12pt}}
Let $(X, \mathfrak{A}, \mu)$ be a measure space. Given a $\mu$-integrable function $f : X \to \mathbb{C}$, one may define a complex measure by
\begin{equation}\label{def:Mass_mit_Dichte}
\nu(A) := \int_{A} f\ d\mu, \quad A \in \mathfrak{A}.
\end{equation}
The complex measure constructed in this manner is occasionally called the \emph{complex measure with the density $f$ with respect to $\mu$}, and we write it as $\nu = f \odot \mu$. A measurable function $g : X \to \hat{\mathbb{K}}$ is known to be $(f \odot \mu)$-integrable, if and only if the product $g \cdot f$ is $\mu$-integrable, in which case the equality
\begin{equation}\label{eq:Transformationsformel_Mass_mit_Dichte}
\int_{X} g \ d(f \odot \mu) = \int_{X} g \cdot f \ d\mu
\end{equation}
holds.
\item {\it Image Measure. \hspace{12pt}}
Let $(X, \mathfrak{A}, \mu)$ be a measure space. Given another measurable space $(Y, \mathfrak{B})$ and a measurable map $f : X \to Y$, one may construct a new measure on $(Y, \mathfrak{B})$ by
\begin{equation}\label{def:Bildmass}
f(\mu)(B) := \mu(f^{-1}(B)), \quad B \in \mathfrak{B},
\end{equation}
called the \emph{image measure (push-forward measure) of $\mu$ with respect to $f$}. A measurable function $g : Y \to \hat{\mathbb{K}}$ is known to be $f(\mu)$-integrable, if and only if the composition $g \circ f$ is $\mu$-integrable, in which case the the \emph{change of variables formula}
\begin{equation}\label{eq:Transformationsformel_Bildmass}
\int_{Y} g \ df(\mu) = \int_{X} g \circ f \ d\mu
\end{equation}
holds.
\end{enumerate}
\paragraph{Measure Algebra}
The space of complex measures has an additional well-known structure regarding convolutions.
The \emph{convolution} of the two complex measures $\mu, \nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ is defined by
\begin{align}\label{def:convolution_measure}
(\mu \ast \nu)(B) := \int_{\mathbb{R}^{n}} \mu(B - x)\ d\nu(x), \quad B \in \mathfrak{B}^{n}.
\end{align}
One can easily confirm that the convolution is a bilinear operation, and is moreover shown to be associative $\mu \ast (\nu \ast \rho) = (\mu \ast \nu) \ast \rho$ and commutative $\mu \ast \nu = \nu \ast \mu$. Together with the evaluation $\|\mu \ast \nu\| \leq \|\mu\| \|\nu\|$ based on the total variation norm \eqref{def:totvar}, one sees that the convolution makes the complex Banach space $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ into a complex commutative Banach algebra, called the \emph{measure algebra} of $\mathfrak{B}^{n}$. The measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ has a multiplicative identity $e$ given by the delta measure $e = \delta_{0}$ centred at the origin, that is,
\begin{equation}
\mu \ast \delta_{0} = \delta_{0} \ast \mu = \mu
\end{equation}
holds for all $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$.
Here, the \emph{delta measure} (or the \emph{Dirac measure}) $\delta_{a}$ is a finite measure centred at $a \in \mathbb{R}^{n}$ defined by
\begin{equation}\label{def:delta_measure}
\delta_{a}(B) =
\begin{cases}
1, & a \in B, \\
0, & a \notin B,
\end{cases}
\qquad B \in \mathfrak{B}^{n},
\end{equation}
characterised by the integral
\begin{equation}
\int_{\mathbb{R}^{n}} f(x)\ d\delta_{a}(x) = f(a),
\end{equation}
whenever the integration is well-defined. It is essentially the same object as the delta distribution that appears in the theory of generalised functions.
\subsubsection{The Space of Density Functions}\label{sec:absolute_continuity}
For later use, we are particularly interested in the special subspace of the space $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ of complex measures, namely, the space of \emph{absolutely continuous} complex measures with respect to the Lebesgue-Borel measure $\beta^{n}$. We shall provide a concise review on its definition, make comments on its relation to the space of complex \emph{density functions}, and sees that the subspace reveals itself to be a sub-algebra of the measure algebra.
\paragraph{Absolute Continuity and Density Functions}
Let $\mu$ and $\nu$ be signed (or complex) measures on a measurable space $(X, \mathfrak{A})$. We say that $\nu$ is \emph{$\mu$-continuous} or \emph{absolutely continuous with respect to $\mu$}, written as $\nu \ll \mu$, if $\mu(A) = 0$ implies $\nu(A) = 0$ for all $A \in \mathfrak{A}$. A signed measure $\mu$ is called \emph{$\sigma$-finite} if there exists a sequence $(A_{n})_{n \geq 1}$ of disjoint measurable sets $A_{n} \in \mathfrak{A}$ satisfying $X = \bigcup_{n=1}^{\infty} A_{n}$ and $|\mu(A_{n})| < \infty$ ($n \in \mathbb{N}^{\times}$). By definition, finite measures are always $\sigma$-finite. The Lebesgue-Borel measure $\beta^{n}$ is among the most important examples of $\sigma$-finite measures. The following theorem is of great importance.
\begin{theorem*}[Radon-Nikod{\'y}m Theorem for Complex Measures]
Let $\mu$ be a $\sigma$-finite measure and $\nu \ll \mu$ be a complex measure. Then, $\nu$ has a density with respect to $\mu$, that is, there exists a $\mu$-integrable function $\rho : X \to \mathbb{C}$ such that $\nu = \rho \odot \mu$, and $\rho$ is unique $\mu$-a.e. If $\nu$ happens to be positive, then one may choose $\rho \geq 0$.
\end{theorem*}
\noindent
In the above situation of the Radon-Nikod{\'y}m theorem, the function $\rho$ satisfying $\nu = \rho \odot \mu$ is called the \emph{Radon-Nikod{\'y}m derivative} (or more casually, the \emph{density}), and is denoted by
\begin{equation}
\rho =: \frac{d\nu}{d\mu}.
\end{equation}
This is nothing but to say that
\begin{equation}\label{def:abs_cont}
\nu(A) = \int_{A} \frac{d\nu}{d\mu}(x)\ d\mu(x), \quad A \in \mathfrak{A},
\end{equation}
holds, if explicitly written out. For a $\nu$-integrable function $f$, a direct application of \eqref{eq:Transformationsformel_Mass_mit_Dichte} leads to
\begin{equation}\label{prop:Mass_mit_Dichte}
\int_{\mathbb{X}} f(x)\ d\nu(x) = \int_{X} f(x) \frac{d\nu}{d\mu}(x)\ d\mu(x),
\end{equation}
in which the notation for the Radon-Nikod{\'y}m derivative (which might at first seems strange) reveals its advantage.
\paragraph{Absolute Continuity with respect to the Lebesgue-Borel Measure}
We are particularly interested in the sub-family $L^{1}(\mathfrak{B}^{n}) \subset \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ consisting of complex measures that are absolutely continuous with respect to the Lebesgue-Borel measure $\beta^{n}$ on $(\mathbb{R}^{n},\mathfrak{B}^{n})$. Whenever there is no risk of confusion, members of $L^{1}(\mathfrak{B}^{n})$ shall occasionally be referred to as \emph{absolutely continuous measures}, simply without reference to the base measure $\beta^{n}$. One readily finds that the collection $L^{1}(\mathfrak{B}^{n})$ forms a linear subspace of $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$.
Now, uniqueness $\beta^{n}$-a.e. of the Radon-Nikod{\'y}m derivative allows us to define a linear map
\begin{equation}\label{def:abs_cont_ident}
L^{1}(\mathfrak{B}^{n}) \to L^{1}(\mathbb{R}^{n}),\ \mu \mapsto \frac{d\mu}{d\beta^{n}},
\end{equation}
which maps an absolutely continuous complex measure to its density.
Conversely, one may construct a new complex measure given an integrable function $f \in L^{1}(\mathbb{R}^{n})$ by $\nu := f \odot \beta^{n}$. From this, one obtains a bijective linear map between the space of absolutely continuous complex measures $L^{1}(\mathfrak{B}^{n})$ and the space of integrable functions $L^{1}(\mathbb{R}^{n})$, associating an absolutely continuous complex measure $\nu \in L^{1}(\mathfrak{B}^{n})$ to its density $d\nu/d\beta^{n} \in L^{1}(\mathbb{R}^{n})$. In this manner, one may identify a specific subspace of the space of complex measures with that of integrable functions as
\begin{equation}\label{def:abs_measure_identification}
L^{1}(\mathfrak{B}^{n}) \cong L^{1}(\mathbb{R}^{n}),
\end{equation}
and may translate and interpret various properties of complex measures in terms of density functions.
To discuss how this works, let $d\nu/d\beta^{n} \in L^{1}(\mathbb{R}^{n})$ be the density of $\nu \in L^{1}(\mathfrak{B}^{n})$ with respect to the Lebesgue-Borel measure. One confirms from \eqref{prop:Mass_mit_Dichte} that, for any measurable function $g$, the equality
\begin{equation}\label{eq:Mass_mit_Dichte_AC}
\int_{\mathbb{R}^{n}} g(x)\ d\nu(x) = \int_{\mathbb{R}^{n}} g(x)\frac{d\nu}{d\beta^{n}}(x)\ d\beta^{n}(x)
\end{equation}
holds whenever the integration exists. In this manner, one may replace the Lebesgue integration of $g$ with respect to the complex measure $\nu$ (the l.~h.~s.) by that with respect to the Lebesgue-Borel measure with the help of the (possibly more familiar notion of) density function $d\nu/d\beta^{n}$ (the r.~h.~s.).
\begin{comment}
$\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ also has a involution $\nu \mapsto \nu^{\ast}$ defined by
\begin{equation}
\nu^{\ast}(B) := \overline{\nu(-B)},\quad B \in \mathfrak{B},
\end{equation}
where $-B := \{-x : x \in B\}$, and it is indeed an involution since $(\mu + \nu)^{\ast} = \mu^{\ast} + \nu^{\ast}$, $(\lambda \nu)^{\ast} = \overline{\lambda}\nu^{\ast}$, $\nu^{\ast\ast} = \nu$ and $(\mu \ast \nu)^{\ast} = \nu^{\ast} \ast \mu^{\ast}$ holds for all $\mu, \nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B})$, $\lambda \in \mathbb{C}$. Summing up, the convolution and involution together makes $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ into a unital Banach $\ast$-algebra.
\end{comment}
\paragraph{Convolution Algebra}
The space $L^{1}(\mathfrak{B}^{n})$ of absolutely continuous complex measures is readily shown to be a topologically closed subset (with respect to the topology induced by the total variation norm $\|\cdot \|$ in \eqref{def:totvar}) of the Banach space $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$. This implies that the subspace $L^{1}(\mathfrak{B}^{n})$ is itself a Banach space.
One then finds that the linear bijection \eqref{def:abs_cont_ident} between the two Banach spaces actually defines an \emph{isometric (linear) isomorphism}, which is to say that
\begin{equation}\label{eq:isometry}
\|\nu\| = \left\|\frac{d\nu}{d\beta^{n}}\right\|_{1}
\end{equation}
holds for all $\nu \in L^{1}(\mathfrak{B}^{n})$, where the l.~h.~s. is the total variation norm \eqref{def:totvar} of the complex measure $\nu$ and the r.~h.~s. is the $L^{1}$-norm \eqref{def:Lp_norm} of its density function.
We next see how this bijection plays with convolution. To this end, we first recall that a linear subspace $\mathfrak{I}$ of a commutative algebra $\mathfrak{A}$ is called an \emph{ideal} if it `absorbs' multiplication by elements of $\mathfrak{A}$, {\it i.e.},
\begin{equation}
i \in \mathfrak{I},\ a \in \mathfrak{A} \quad \Rightarrow \quad i \cdot a = a \cdot i \in \mathfrak{I}.
\end{equation}
In fact, it is known that the subspace $L^{1}(\mathfrak{B}^{n})$ forms an ideal of the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$, which is to say that
\begin{equation}
\mu \in L^{1}(\mathfrak{B}^{n}),\ \nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n}) \quad \Rightarrow \quad \mu \ast \nu = \nu \ast \mu \in L^{1}(\mathfrak{B}^{n}).
\end{equation}
In passing, the density of the convolution $\mu \ast \nu$ above is given by the convolution of the density of $\mu$ and the complex measure $\nu$ as
\begin{equation}\label{eq:density_ac_cmeas}
\frac{d(\mu \ast \nu)}{d\beta^{n}} = \frac{d\mu}{d\beta^{n}} \ast \nu,
\end{equation}
in which we understand the convolution of an integrable function $f \in L^{1}(\mathbb{R}^{n})$ and a complex measure $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ to be
\begin{equation}
(f \ast \mu)(x) := \int_{\mathbb{R}^{n}} f(x-y)\ d\mu(y),
\end{equation}
where the integral is well-defined $\beta^{n}$-a.e. for $x \in \mathbb{R}^{n}$.
In particular, being an ideal trivially implies that the space $L^{1}(\mathfrak{B}^{n})$ of absolutely continuous complex measures is closed under the operation of convolution, {\it i.e.}, it forms a \emph{sub-algebra} of the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$. Applying \eqref{prop:Mass_mit_Dichte} to \eqref{eq:density_ac_cmeas}, one concludes that the density of the convolution of two absolutely continuous complex measures $\mu, \nu \in L^{1}(\mathfrak{B}^{n})$ is given by the convolution of their densities as
\begin{equation}\label{eq:density_ac_ac}
\frac{d(\mu \ast \nu)}{d\beta^{n}} = \frac{d\mu}{d\beta^{n}} \ast \frac{d\nu}{d\beta^{n}},
\end{equation}
in which we understand the familiar convolution of two integrable functions $f, g \in L^{1}(\mathbb{R}^{n})$ to be
\begin{equation}\label{def:convol_func_func}
(f \ast g)(x) := \int_{\mathbb{R}^{n}} f(x-y)g(y)\ d\beta^{n}(y),
\end{equation}
where the integral is well-defined $\beta^{n}$-a.e. for $x \in \mathbb{R}^{n}$. Equality \eqref{eq:density_ac_ac} implies that, equipped with the convolution \eqref{def:convol_func_func}, the space $L^{1}(\mathbb{R}^{n})$ of integrable functions becomes a Banach algebra that is isomorphically mapped to the sub-algebra $L^{1}(\mathfrak{B}^{n}) \subset \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ by the \emph{isometric algebra isomorphism} \eqref{def:abs_cont_ident}. Incidentally, the sub-algebra $L^{1}(\mathfrak{B}^{n}) \cong L^{1}(\mathbb{R}^{n})$ of the measure algebra is given its own name, and is occasionally called the \emph{convolution algebra}.
At this point, we note that the convolution algebra $L^{1}(\mathfrak{B}^{n})$ is a proper sub-algebra of the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ in general, {\it i.e.}, not every complex measure may be represented by integrable functions. This can be readily seen by observing that the delta measure $\delta_{a}$ centred at $a \in \mathbb{R}^{n}$ \eqref{def:delta_measure} does not admit a description by density functions. Intuitively, such a density function, if existed, would be given by the `delta function' centred at $a$, but it is actually a distribution and not a member of $L^{1}(\mathbb{R}^{n})$ as required. This leads to the basic fact that the convolution algebra $L^{1}(\mathbb{R}^{n})$ is non-unital, {\it i.e.}, it
lacks a multiplicative identity in the sense that there is no element $e \in L^{1}(\mathbb{R}^{n})$ for which
\begin{equation}
e \ast f = f
\end{equation}
holds for all $f \in L^{1}(\mathbb{R}^{n})$. This should be contrasted to the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$, which always possesses a multiplicative identity.
\subsubsection{Product Measures}
Given two measure spaces $(X, \mathfrak{A}, \mu)$ and $(Y, \mathfrak{B}, \nu)$, we intend to construct a `product measure' on the product space $X \times Y$ so that $\rho(A \times B) = \mu(A) \nu(B)$ holds for all $A \in \mathfrak{A}$, $B \in \mathfrak{B}$. As its domain of definition, we let
\begin{align}\label{def:product_set_of_sigma_algebras}
\mathfrak{A} \ast \mathfrak{B} := \{A \times B : A \in \mathfrak{A}, B \in \mathfrak{B} \}
\end{align}
and define
\begin{equation}
\mathfrak{A} \otimes \mathfrak{B} := \sigma(\mathfrak{A} \ast \mathfrak{B})
\end{equation}
to be the \emph{product-$\sigma$-algebra} of $\mathfrak{A}$ and $\mathfrak{B}$. The following fact and definition is of importance.
\begin{definition*}[Product Measure]
Given two measure spaces $(X, \mathfrak{A}, \mu)$ and $(Y, \mathfrak{B}, \nu)$, let both $\mu$ and $\nu$ be $\sigma$-finite. Then there exists a unique measure $\mu \otimes \nu : \mathfrak{A} \otimes \mathfrak{B} \to \overline{\mathbb{R}}$ such that
\begin{equation}\label{def:product_measure}
\mu \otimes \nu (A \times B) = \mu(A) \nu(B), \quad A \in \mathfrak{A},\ B \in \mathfrak{B}
\end{equation}
holds. The measure $\mu \otimes \nu$ is $\sigma$-finite and is called the product measure of $\mu$ and $\nu$.
\end{definition*}
\noindent
The integration with respect to the product measure $\mu \otimes \nu$ of two $\sigma$-finite measures $\mu$ and $\nu$ can be performed by iterated integration of each of the respective variables. This is the essence of the following \emph{Fubini's Theorem}, which belongs to one of the most oft-used theorems of integration theory.
\begin{theorem*}[Fubini's Theorem]
Let $\mu$ and $\nu$ be $\sigma$-finite. Then, the following statements hold:
\begin{enumerate}
\item
If $f: X \otimes Y \to \hat{\mathbb{K}}$ is $\mu \otimes \nu$-integrable, then $f(x, \cdot)$ is $\nu$-integrable for almost all $x \in X$. Moreover
\begin{equation}
A := \{x \in X : f(x,\cdot) \text{ is not $\nu$-integrable }\} \in \mathfrak{A};
\end{equation}
and likewise
\begin{equation}
B := \{y \in Y : f(\cdot,y) \text{ is not $\mu$-integrable }\} \in \mathfrak{B}.
\end{equation}
The functions
\begin{equation}
x \mapsto \int_{Y} f(x,y)\ d\nu(y) \quad x \mapsto \int_{X} f(x,y)\ d\mu(x)
\end{equation}
are respectively $\mu$-integrable on $A^{c}$ and $\nu$-integrable on $B^{c}$, and the equalities
\begin{align}
\int_{X\times Y} f\ d\mu\otimes\nu
&= \int_{X} \left( \int_{Y} f(x,y)\ d\nu(y) \right)\ d\mu(x) \nonumber \\
&= \int_{Y} \left( \int_{X} f(x,y)\ d\mu(x) \right)\ d\nu(y)
\end{align}
hold.
\item
If $f: X \otimes Y \to \hat{\mathbb{K}}$ is $\mu \otimes \nu$-integrable, and one of the integrals
\begin{equation}
\int_{X \times Y} |f|\ d\mu\otimes\nu,\ \int_{X} \left( \int_{Y} |f(x,y)|\ d\nu(y) \right)\ d\mu(x),\ \int_{Y} \left( \int_{X} |f(x,y)|\ d\mu(x) \right)\ d\nu(y)
\end{equation}
is finite, then all three of them are finite and agree, $f$ is $\mu \otimes \nu$-integrable, and the statements under \textup{(i)} hold.
\end{enumerate}
\end{theorem*}
\subsubsection{Measure on Topological Spaces}
Let $X$ be a metric space (or a topological space). One may naturally be interested in how the topology relates to the complex measures defined on the Borel $\sigma$-algebra $\mathfrak{B} := \mathfrak{B}(X)$ generated by it. To this end, we briefly review one of the prominent results in the study of this realm, namely the famous \emph{Riesz-Markov-Kakutani Representation Theorem}.
In order to avoid complexity, we shall only deal with the case where the given measurable space is $(\mathbb{R}^{n}, \mathfrak{B}^{n})$. Observing now that a complex measure $\nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ generates an (algebraic) linear map $f \mapsto \int_{\mathbb{R}^{n}} f d\nu$ that maps a function to a complex number, the opposite question is then our interest, namely: what class of linear functionals admits representation by integration with respect to some complex measure?
\paragraph{Riesz-Markov-Kakutani Representation Theorem}
Let $C_{0}(\mathbb{R}^{n})$ be the space of all continuous functions $f: \mathbb{R}^{n} \to \mathbb{C}$ that vanish at infinity, in the sense for every $\epsilon > 0$ there exists a compact subset $K \subset \mathbb{R}^{n}$ for which $|f|K^{c}| < \epsilon$ holds. The space $C_{0}(\mathbb{R}^{n})$ equipped with the supremum norm $\|f\|_{\infty} := \sup \{|f(x)| : x \in \mathbb{R}^{n} \}$ is known to be a Banach space.
Now for each $\nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$, the map
\begin{equation}
I_{\nu} : f \mapsto \int_{\mathbb{R}^{n}} f\ d\nu, \quad f \in C_{0}(\mathbb{R}^{n})
\end{equation}
gives rise to a continuous ({\it i.e.}, bounded) $\mathbb{C}$-linear functional from $C_{0}(\mathbb{R}^{n})$ to $\mathbb{C}$, for indeed the evaluation
\begin{equation}
|I_{\nu}(f)| \leq \|\nu\| \cdot \|f\|_{\infty}
\end{equation}
holds.
The Riesz-Markov-Kakutani representation theorem is a classical theorem in measure and integration theory stating that the converse is also true, which is to say that, for any continuous $\mathbb{C}$-linear functional $I \in C_{0}^{\prime}(\mathbb{R}^{n})$, there exists a unique complex measure $\nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ for which
\begin{equation}
I(f) = \int_{\mathbb{R}^{n}} f\ d\nu, \quad f \in C_{0}(\mathbb{R}^{n})
\end{equation}
holds. The precise statement is given as follows.
\begin{theorem*}[Riesz-Markov-Kakutani Representation Theorem for Euclidian Spaces]
The correspondence
\begin{align}\label{eq:RMK-thm_cor}
\begin{split}
&\Phi : \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n}) \to C_{0}^{\prime}(\mathbb{R}^{n}), \\
&\Phi(\nu)(f) := \int_{\mathbb{R}^{n}} f\ d\nu, \quad \nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n}),\, f \in C_{0}(\mathbb{R}^{n}),
\end{split}
\end{align}
that maps a complex measure to a continuous linear functional on $C_{0}$ is a bijection, which moreover satisfies
\begin{equation}\label{eq:Riesz_rep_isometry}
\|\Phi(\nu)\| = \|\nu\|.
\end{equation}
In other words, the space of complex measures $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ is isomorphic to the topological dual of $C_{0}(\mathbb{R}^{n})$, and can be mapped to each other by an isometric isomorphism.
\end{theorem*}
\noindent
Here, the norm on $\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ on the r.~h.~s. of \eqref{eq:Riesz_rep_isometry} is naturally the total variation norm, and the norm on the topological dual $C_{0}^{\prime}(\mathbb{R}^{n})$ (the l.~h.~s.) is the operator norm defined by
\begin{equation}
\|I\| := \sup_{\|f\|_{\infty}\leq 1} |I(f)|, \quad I \in C_{0}^{\prime}(\mathbb{R}^{n}).
\end{equation}
In this sense we identify
\begin{equation}\label{eq:RMK-thm}
\mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n}) \cong C_{0}^{\prime}(\mathbb{R}^{n}),
\end{equation}
and may interchangeably interpret a continuous $\mathbb{C}$-linear functional on the space $C_{0}(\mathbb{R}^{n})$ as a complex measure on the measurable space $(\mathbb{R}^{n}, \mathfrak{B}^{n})$, and vice versa.
\subsubsection{Spectral Theorem and its Consequences}\label{sec:spectral_theorem}
We next provide a concise review on some of the basic facts regarding the spectral theorem for self-adjoint operators, which is just the generalisation of the familiar eigendecomposition theorem for Hermitian matrices on finite-dimensional vector spaces to the arbitrary dimensional case. In order to avoid confusion with operators, Borel sets on $\mathbb{R}^{n}$ shall occasionally be denoted by $\Delta \in \mathfrak{B}^{n}$ in place of $B$, especially when we are working in the context of quantum mechanics.
\paragraph{Spectral Measures}
Closely associated to the notion of complex measures is that of spectral measures on a Hilbert space $\mathcal{H}$. Let $L(\mathcal{H})$ denote the space of all bounded operators on $\mathcal{H}$, and recall that a map
\begin{equation}
E : \mathfrak{B}^{n} \to L(\mathcal{H}), \quad \Delta \mapsto E(\Delta)
\end{equation}
is called an $n$-dimensional \emph{spectral measure} (or \emph{projection-valued measure}), if each $E(\Delta)$, $\Delta \in \mathfrak{B}^{n}$ is an orthogonal projection on $\mathcal{H}$ and satisfies
\begin{enumerate}
\item $E(\emptyset) = 0$, $E(\mathbb{R}^{n}) = I$,
\item for pairwise disjoint $\Delta_{1}, \Delta_{2}, \dots \in \mathfrak{B}^{n}$,
\begin{equation}
\sum_{i=1}^{\infty}E(\Delta_{i}) |\phi\rangle = E\left(\bigcup_{i=1}^{\infty}\Delta_{i}\right) |\phi\rangle, \qquad \forall |\phi\rangle \in \mathcal{H}.
\end{equation}
\end{enumerate}
The support of a spectral measure $E$ on $\mathfrak{B}^{n}$ is defined as the smallest Borel set $\Delta \in \mathfrak{B}^{n}$ that satisfies $E(\Delta) = I$.
An important point is that
a spectral measure $E$ and a pair of vectors $|\phi\rangle, |\phi\rangle \in \mathcal{H}$ induce a complex measure on $\mathfrak{B}^{n}$ given by
\begin{equation}\label{def:complex_measure_A}
\Delta \mapsto \langle \phi^{\prime}, E(\Delta) \phi \rangle, \quad \Delta \in \mathfrak{B}^{n}.
\end{equation}
\paragraph{Spectral Theorem of Self-adjoint Operators}
Having recalled the necessary definitions, we now state the \emph{spectral theorem for self-adjoint operators}, which constitutes one of the most important mathematical ingredients in quantum mechanics.
\begin{theorem*}[Spectral decomposition of self-adjoint operators]
Let $A : \mathcal{H} \supset \mathrm{dom}(A) \to \mathcal{H}$ be self-adjoint. Then there exists a unique one-dimensional spectral measure $E_{A}$ supported on the spectrum $\sigma(A) \subset \mathbb{R}$ of $A$ satisfying
\begin{equation}\label{thm:sd}
\langle \phi^{\prime}, A \phi \rangle = \int_{\sigma(A)} a\ d\langle \phi^{\prime}, E_{A}(a) \phi \rangle, \quad \forall |\phi\rangle \in \mathrm{dom}(A),\ \forall |\phi^{\prime}\rangle \in \mathcal{H},
\end{equation}
where the r.~h.~s. of the equality is understood as the Lebesgue integral with respect to the complex measure $\Delta \mapsto \langle \phi^{\prime}, E_{A}(\Delta) \phi \rangle$
induced from $E_{A}$ and the pair of vectors $|\phi\rangle$ and $|\phi^{\prime}\rangle$.
\end{theorem*}
\noindent
Under the situation above, the self-adjoint operator $A$ is occasionally written symbolically as
\begin{equation}\label{thm:sd_op}
A = \int_{\sigma(A)} a\ dE_{A}(a),
\end{equation}
in terms of integration with respect to its spectral measure.
\paragraph{Finite-dimensional Case}
To see the meaning of the above formula, we make a brief note on how the familiar eigendecomposition theorem for Hermitian matrices appears as a special case of the general statement. Let $A$ be a Hermitian matrix on an $N$-dimensional complex Hilbert space $\mathcal{H} := \mathbb{C}^{N}$, $N \in \mathbb{N}^{\times}$. The eigendecomposition theorem states that, there exists an orthonormal basis $ \mathcal{B}_{A} := \{|a_{1}\rangle, \dots, |a_{N}\rangle\}$ of $\mathcal{H}$ with real numbers $a_{1}, \dots, a_{N} \in \mathbb{R}$ such that
\begin{equation}
A |a_{i}\rangle = a_{i} |a_{i}\rangle, \quad i = 1, \dots, N,
\end{equation}
hold.
For each eigenvalue $a \in \sigma(A) = \{a_{1}, \dots, a_{N}\}$ of $A$, we have the projection $\Pi_{a}$
onto the subspace,
\begin{equation}
\mathcal{H}_{a} := \mathrm{span} \{ |a\rangle \in \mathcal{B}_{A} : A |a\rangle = a |a\rangle \}
\end{equation}
spanned by the collection of all eigenvectors associated with $a$. As we noted before, when the eigenstate $| a \rangle$ is non-degenerate for $a$, or the subspace $\mathcal{H}_{a}$ is one-dimensional, we may write $\Pi_a = | a \rangle \langle a |$.
With the projection $ \Pi_a $ in hand, the spectral measure of $A$ is defined by
\begin{equation}\label{def:spectrum_finite}
E_{A}(\Delta) := \sum_{a \in \sigma(A) \cap \Delta} \Pi_a, \quad \Delta \in \mathfrak{B},
\end{equation}
with the convention $\sum_{a \in \emptyset}\Pi_a := 0$.
One readily verifies that $E_{A}$ is indeed a spectral measure supported on its spectrum $\sigma(A)$, and subsequently sees that the projection $\Pi_{a} = E_{A}(\{a\})$ is nothing but the image of the spectral measure $E_{A}$ on the Borel set $\{a\} \in \mathfrak{B}$ consisting of a single eigenvalue $a \in \sigma(A)$ of the observable $A$. One then finds
\begin{align}\label{fin_sd_op}
A = \sum_{a \in \sigma(A)} a\, \Pi_a = \sum_{a \in \sigma(A)} a\, E_{A}(\{a\})
\end{align}
in accordance with \eqref{eq:spect_decomp_fin}, and subsequently proves
\begin{equation}\label{fin_sd}
\langle \phi^{\prime}, A \phi \rangle = \sum_{a \in \sigma(A)} a\, \langle \phi^{\prime}, E_{A}(\{a\}) \phi \rangle, \quad \forall |\phi^{\prime}\rangle, |\phi\rangle \in \mathcal{H}.
\end{equation}
The spectral decomposition formula \eqref{thm:sd} and the formal expression \eqref{thm:sd_op} are respectively just the generalisations of the finite dimensional versions \eqref{fin_sd} and \eqref{fin_sd_op}.
\paragraph{Functional Calculus}
By means of the spectral decomposition of a self-adjoint operator, one may create a new set of operators from it. Let $A : \mathcal{H} \supset \mathrm{dom}(A) \to \mathcal{H}$ be a self-adjoint operator on a Hilbert space $\mathcal{H}$, and let $E_{A}$ the unique one-dimensional spectral measure associated with it. Given a measurable complex function $f : \mathbb{R} \to \mathbb{C}$, the integral
\begin{equation}
\langle\phi^{\prime}, f(A) \phi\rangle = \int_{\sigma(A)} f(a)\ d\langle \phi^{\prime}, E_{A}(a)\phi\rangle
\end{equation}
defines a unique linear operator $f(A)$ on $\mathcal{H}$, where
\begin{equation}
|\phi\rangle \in \mathrm{dom}(f(A)) := \left\{ |\phi\rangle \in \mathcal{H} : \int_{\sigma(A)} \left| f(a) \right|^{2}\ d\langle \phi, E_{A}(a)\phi\rangle < \infty \right\}
\end{equation}
is any vector belonging to its domain, and $|\phi^{\prime}\rangle \in \mathcal{H}$. The operator $f(A)$ is occasionally written symbolically as
\begin{equation}\label{def:func_calc}
f(A) = \int_{\sigma(A)} f(a)\ dE_{A}(a),
\end{equation}
in terms of integration with respect to its spectral measure.
\paragraph{Born Rule and Quantum Measurement}
The axiom of quantum mechanics states that a quantum observable is represented by a self-adjoint operator $A : \mathcal{H} \supset \mathrm{dom}(A) \to \mathcal{H}$ on a Hilbert space $\mathcal{H}$, and that the probabilistic behaviour of the outcomes of an ideal measurement of $A$ on the state $|\phi\rangle \in \mathcal{H}$ is described by the probability measure,
\begin{equation}\label{def:prob_measrue_A}
\Delta \mapsto \mu_{A}^{\phi}(\Delta) := \frac{\langle\phi, E_{A}(\Delta) \phi\rangle}{\|\phi\|^{2}}, \quad \Delta \in \mathfrak{B}.
\end{equation}
Here, the spectral measure $E_{A}$ is induced from $A$ by the spectral theorem,
and the Born rule proclaims that the measurement outcome be given by one of the elements in the spectrum $\sigma(A)$ and that
$\mu_{A}^{\phi}(\Delta)$ provides the probability of finding the measurement in the measurable set $\Delta \in \mathfrak{B}$.
Given $|\phi\rangle \in \mathrm{dom}(A)$, one then realises from the spectral theorem \eqref{thm:sd} that the statistical average of the measurement outcomes of $A$ gives the expectation value,
\begin{align}\label{def:expectation_A}
\int_{\mathbb{R}} a\ d\mu_{A}^{\phi}(a)
= \frac{\langle \phi, A \phi \rangle}{\|\phi\|^{2}} =: \mathbb{E}[A;\phi],
\end{align}
where the l.~h.~s of the first equality is understood to be the Lebesgue integral with respect to the probability measure \eqref{def:prob_measrue_A}.
\subsubsection{Observables admitting a Description by Density Functions}
While the analysis based on probability measures provides an adequately general framework to work with, we find it useful to prepare a terminology for a special class of observables for which probability density functions, not just probability measures, are available to fully describe the behaviour of the measurement outcomes.
\paragraph{Observable admitting a description by probability density functions}
In this paper, we simply say that
an observable $A$ \emph{admits a description by probability density functions}, if the probability measure \eqref{def:prob_measrue_A} induced by the spectral measure of $A$ is absolutely continuous with respect to the Lebesgue-Borel measure for every choice of the quantum state $|\phi\rangle \in \mathcal{H}$, which is to say that, if for every $|\phi\rangle \in \mathcal{H}$, there exists an integrable function $\rho_{A}^{\phi} \in L^{1}(\mathbb{R})$ such that
\begin{equation}\label{def:absolutely_continuous}
\mu_{A}^{\phi}(\Delta)
= \int_{\Delta} \rho_{A}^{\phi}(a)\ d\beta(a), \quad \Delta \in \mathfrak{B},
\end{equation}
holds.
A well-known example of it is provided by the one-dimensional position operator $\hat{x}$ on $L^{2}(\mathbb{R})$ defined in \eqref{def:position_operator}. Indeed, one proves that the spectral measure of $\hat{x}$ is given by the multiplication of the characteristic function \eqref{def:characteristic_function} as
\begin{equation}
E_{\hat{x}}(\Delta) : \psi(x) \mapsto \chi_{\Delta}(x)\psi(x), \quad \psi \in L^{2}(\mathbb{R}),
\end{equation}
for each $B \in \mathfrak{B}$, so that
\begin{align}
\langle \psi_{1}, E_{\hat{x}}(\Delta) \psi_{2} \rangle
&= \int_{\mathbb{R}} \psi_{1}^{*}(x) \chi_{\Delta}(x) \psi_{2}(x)\ d\beta(x) \nonumber \\
&= \int_{\Delta} \psi_{1}^{*}(x) \psi_{2}(x)\ d\beta(x), \quad \Delta \in \mathfrak{B},
\end{align}
holds. Specifically, this implies that
\begin{equation}
\mu_{\hat{x}}^{\psi}(\Delta) = \int_{\Delta} \frac{|\psi(x)|^{2}}{\| \psi \|^{2}_{2}}\ d\beta(x), \quad \Delta \in \mathfrak{B},
\end{equation}
where the denominator of the integrand of the r.~h.~s. denotes the square of the $L^{2}$-norm of $\psi \in L^{2}(\mathbb{R})$ (see \eqref{def:Lp_norm}). One thus concludes that the density of the probability measure $\mu_{\hat{x}}^{\psi}$ is provided by
\begin{equation}
\rho_{\hat{x}}^{\psi}(x) = \frac{|\psi(x)|^{2}}{\| \psi \|^{2}_{2}}.
\end{equation}
Incidentally, it is known that each member of the pair of observables $\{Q, P\}$ that satisfies the Weyl relations \eqref{eq:weyl01} and \eqref{eq:weyl02} admits descriptions in terms of density functions.
However, it should be noted that this is not always the case in general: an observable with the spectrum consisting of a finite number of discrete eigenvalues (such as spin) provides a simple counterexample. To see this, let $A$ be such an observable with $N \in \mathbb{N}^{\times}$ distinct eigenvalues, and let $\sigma(A) = \{a_{1}, \dots, a_{N} \}$ be any enumeration of its spectrum. A straightforward application of \eqref{def:spectrum_finite} leads to
\begin{equation}\label{eq:prob_meas_fin_spec_obs}
\mu_{A}^{\phi} = \sum^{N}_{n=1} \mu_{A}^{\phi}(\{a_{n}\}) \cdot \delta_{a_{n}},
\end{equation}
in which one sees that the probability measure $\mu_{A}^{\phi}$ is given by the weighted sum of delta measures centred at each eigenvalue. Obviously, since each of the delta measures is not absolutely continuous, the resultant probability measure does not admit a description by density functions.
For later use, we also note that, once the observable $A$ admits a description in terms of probability density functions,
then the complex measure \eqref{def:complex_measure_A} is also absolutely continuous for an arbitrary pair of vectors $|\phi^{\prime}\rangle, |\phi\rangle \in \mathcal{H}$. That this is the case can be seen by a straightforward application of the polarisation identity
\begin{align}\label{def:polarisation_identity}
\langle\phi^{\prime}, T\phi \rangle
&= \frac{1}{4} \left\{ \langle\phi^{\prime} + \phi, T(\phi^{\prime} + \phi) \rangle
- \langle\phi^{\prime} - \phi, T(\phi^{\prime} - \phi) \rangle \right. \nonumber \\
& \qquad \left. + i\langle\phi^{\prime} + i\phi, T(\phi^{\prime} + i\phi) \rangle
- i\langle\phi^{\prime} - i\phi, T(\phi^{\prime} - i\phi) \rangle \right\}
\end{align}
with respect to the operator $T$ valid for any pair of vectors $|\phi\rangle, |\phi^{\prime}\rangle \in \mathrm{dom}(T)$, where we simply replace $T=E_{A}(\Delta)$ for each $\Delta \in \mathfrak{B}$.
\subsubsection{Simultaneously measurable Observables}\label{sec:sim_meas_obs}
For reference, we briefly review the basic mathematical definitions and facts involved in describing measurements of simultaneously measurable observables, including the simultaneous measurement of local observables on the tensor product of Hilbert spaces.
\paragraph{Strong Commutativity of Self-adjoint Operators}
Let $A$ and $B$ be self-adjoint operators on a Hilbert space $\mathcal{H}$, and let $E_{A}$ and $E_{B}$ be their respective spectral measures. We say that the pair of operators $A$ and $B$ \emph{strongly commutes}, if
\begin{equation}
E_{A}(\Delta_{A})\, E_{B}(\Delta_{B}) = E_{B}(\Delta_{B})\,E_{A}(\Delta_{A}), \quad \Delta_{A}, \Delta_{B} \in \mathfrak{B}
\end{equation}
holds as an operator equality. Note that the strong commutativity of $A$ and $B$ implies its (familiar) commutativity $AB = BA$. On the other hand, it is known that the converse is in general not true in the case where either (or both) of the operators happens to be unbounded. The term strong commutativity is named after this fact, for it indicates a stronger condition than mere commutativity.
\paragraph{Product Spectral Measures}
It is a basic result of functional analysis that, given such a pair of $A$ and $B$ of strongly commuting self-adjoint operators, there exists a unique two-dimensional spectral measure
$E_{A, B}$ called the \emph{product spectral measure} of $A$ and $B$, for which
\begin{align}\label{eq:product_spectral_measure}
E_{A, B}(\Delta_{A} \times \Delta_{B}) = E_{A}(\Delta_{A})\, E_{B}(\Delta_{B}) = E_{B}(\Delta_{B})\, E_{A}(\Delta_{A}), \quad \Delta_{A}, \Delta_{B} \in \mathfrak{B}
\end{align}
holds. This is a straightforward operator-valued analogue of product measures in measure theory.
With a pair of vectors $|\phi\rangle, |\phi^{\prime}\rangle \in \mathcal{H}$ being specified, this gives rise to a complex measure on $(\mathbb{R}^{2}, \mathfrak{B}^{2})$, defined by
\begin{align}
\Delta \mapsto \left\langle \phi^{\prime}, E_{A, B}(\Delta) \, \phi \right\rangle, \quad \Delta \in \mathfrak{B}^{2}.
\end{align}
In the context of quantum mechanics, for a given pair of simultaneously measurable quantum observables represented by strongly commuting self-adjoint operators $A$ and $B$, the probabilistic behaviour of the outcomes of an ideal simultaneous measurement of both the observables on the state $|\phi\rangle \in \mathcal{H}$ is described by the \emph{joint-probability distribution}
\begin{equation}\label{def:prob_measrue_A_simul}
\Delta \mapsto \mu_{A,B}^{\phi}(\Delta) := \frac{\langle\phi, E_{A, B}(\Delta)\, \phi\rangle}{\|\phi\|^{2}}, \quad \Delta \in \mathfrak{B}^{2},
\end{equation}
of the pair of observables $A$ and $B$ on the state $|\phi\rangle$,
which is a two-dimensional probability measure on the measurable space $(\mathbb{R}^{2}, \mathfrak{B}^{2})$. Here, the r.~h.~s. of \eqref{def:prob_measrue_A_simul} is interpreted as the probability of finding the outcomes of a simultaneous measurement of both observables in the Borel set $\Delta \in \mathfrak{B}^{2}$.
Note that the measurement outcomes of $A$ and $B$ may not be independent, {\it i.e.}, the equality,
\begin{align}
\mu_{A,B}^{\phi}(\Delta_{A} \times \Delta_{B}) = \mu_{A}^{\phi}(\Delta_{A}) \cdot \mu_{B}^{\phi}(\Delta_{B}), \quad \Delta_{A}, \Delta_{B} \in \mathfrak{B},
\end{align}
may not necessarily hold, or in other words, the joint-probability distribution is not necessarily the product measure $\mu_{A,B}^{\phi} \neq \mu_{A}^{\phi} \otimes \mu_{B}^{\phi}$ of each of the respective measurements, in general.
\paragraph{Functional Calculus regarding simultaneously measurable Observables}
Given a pair of strongly commuting self-adjoint observables $A$ and $B$, one readily confirms
\begin{equation}
A = \int_{\mathbb{R}} a\ dE_{A,B}(a,b), \quad B = \int_{\mathbb{R}} b\ dE_{A,B}(a,b).
\end{equation}
As for the sum and product of the observables, we first note the following basic fact.
\begin{lemma}\label{lem:st_strong_commutative}
Let a pair of self-adjoint operators $A$ and $B$ strongly commute. Then,
\begin{enumerate}
\item The operators $A$ and $B$ commute with each other on $\mathrm{dom}(AB) \cap \mathrm{dom}(BA)$, and the anti-commutator%
\footnote{
Here, the domain of the anti-commutator $\{X,Y\} := XY + YX$ of the pair of operators $X$, $Y$ are understood to be $\mathrm{dom}(\{X,Y\}) := \mathrm{dom}(XY) \cap \mathrm{dom}(YX)$.
}
\begin{equation}
\{A,B\} := AB + BA
\end{equation}
is essentially self-adjoint.
\item $A+B$ is essentially self-adjoint.
\end{enumerate}
\end{lemma}
\noindent
As a direct consequence, we thus have the operator equalities
\begin{align}
\overline{A + B} &= \int_{\mathbb{R}^{2}} (a + b)\ dE_{A,B}(a,b) \\
\frac{\overline{AB + BA}}{2} &= \int_{\mathbb{R}^{2}} ab\ dE_{A,B}(a,b)
\end{align}
worth of special notice.
As above, overlines on closable operators denote their closures, and specifically for essentially self-adjoint operators, their self-adjoint extensions.
\paragraph{Composite Systems}
We comment on the special case of the above situation in which the Hilbert space of our interest is the tensor product $\mathcal{H} \otimes \mathcal{K}$ of the target system $\mathcal{H}$ and the meter system $\mathcal{K}$, and the operators involved are (local) self-adjoint operators $A_{1}$ and $A_{2}$ on the respective Hilbert spaces.
Observing that the operators
\begin{equation}
\tilde{A}_{1} := A_{1} \otimes I, \quad \tilde{A}_{2} := I \otimes A_{2}
\end{equation}
strongly commute with each other on the composite Hilbert space $\mathcal{H} \otimes \mathcal{K}$, and that their spectral measures respectively read
\begin{equation}
E_{\tilde{A}_{1}} := E_{A_{1}} \otimes I, \quad E_{\tilde{A}_{2}} := I \otimes E_{A_{2}},
\end{equation}
the previous argument leads to the existence of a unique two-dimensional product spectral measure $E_{A_{1}} \otimes E_{A_{2}} := E_{\tilde{A}_{1},\tilde{A}_{2}}$
satisfying the operator equality
\begin{align}
\left( E_{A_{1}} \otimes E_{A_{2}} \right)(\Delta_{1} \times \Delta_{2})
&= E_{\tilde{A}_{1}}(\Delta_{1}) E_{\tilde{A}_{2}}(\Delta_{2})\nonumber \\
&= (E_{A_{1}}(\Delta_{1}) \otimes I) (I \otimes E_{A_{2}}(\Delta_{2})) \nonumber \\
&= E_{A_{1}}(\Delta_{1}) \otimes E_{A_{2}}(\Delta_{2}), \quad \Delta_{1}, \Delta_{2} \in \mathfrak{B}.
\end{align}
Here, the left-most hand side denotes the two-dimensional spectral measure defined as in \eqref{eq:product_spectral_measure}, while the right-most hand side denotes the tensor product of the self-adjoint operators $E_{A_{1}}(\Delta_{1})$ and $E_{A_{2}}(\Delta_{2})$ for each $\Delta_{1}, \Delta_{2} \in \mathfrak{B}$.
As we have seen in the previous argument, this gives rise to a complex measure,
\begin{align}
\Delta \mapsto \langle \Phi^{\prime}, \left(E_{A_{1}} \otimes E_{A_{2}}\right)(\Delta)\, \Phi \rangle, \quad \Delta \in \mathfrak{B}^{2},
\end{align}
for a given selection of a pair $|\Psi\rangle, |\Phi \rangle \in \mathcal{H} \otimes \mathcal{K}$
of vectors of the composite system, and the map,
\begin{equation}\label{def:prob_measrue_A_entangled}
\Delta \mapsto \mu_{A_{1},A_{2}}^{\Phi}(\Delta) := \frac{\langle\Phi, \left(E_{A_{1}} \otimes E_{A_{2}}\right)(\Delta)\, \Phi\rangle}{\|\Phi\|^{2}}, \quad \Delta \in \mathfrak{B}^{2},
\end{equation}
(here, we have slightly abused the notation on the l.~h.~s. by writing $A_{n}$ in place of $\tilde{A}_{n}$ for each $n = 1, 2$) provides a probability measure describing the probabilistic behaviour of the outcomes of the ideal local measurements simultaneously performed on each system in the state $|\Phi\rangle \in \mathcal{H} \otimes \mathcal{K}$.
In passing, we note that in the case where the state $|\Phi\rangle$ happens to be a direct product state $|\Phi\rangle = |\phi_{1} \otimes \phi_{2} \rangle$,
the induced joint-probability distribution of the two local observables \eqref{def:prob_measrue_A_entangled} becomes the product measure of the two probability measures associated with $A_{1}$ and $A_{2}$,
\begin{align}\label{eq:sim_prob_measrue_dps}
\mu_{A_{1},A_{2}}^{\phi_{1} \otimes \phi_{2}}(\Delta_{1} \times \Delta_{2}) = \mu_{A_{1}}^{\phi_{1}}(\Delta_{1}) \cdot \mu_{A_{2}}^{\phi_{2}}(\Delta_{2}), \quad \Delta_{1}, \Delta_{2} \in \mathfrak{B},
\end{align}
indicating that the measurement outcomes of each local measurement $A_{1}$ and $A_{2}$
are statistically independent ({\it i.e.}, $\mu_{A_{1},A_{2}}^{\phi_{1} \otimes \phi_{2}} = \mu_{A_{1}}^{\phi_{1}} \otimes \mu_{A_{2}}^{\phi_{2}}$).
On the other hand, if one chooses the state $|\Phi\rangle$ to be an entangled state ({\it i.e.}, those states in $\mathcal{H} \otimes \mathcal{K}$ that are not direct product states), the joint-probability distribution \eqref{def:prob_measrue_A_entangled} is no more a product measure of those associated to the local observables in general. In the language of physics, this implies that the local measurements performed on each remote system may have some correlation if the state of the composite system happens to be entangled, and this is widely considered to be one of the most intriguing properties of quantum mechanics.
Of course, statistical independence between the target and the meter systems is useless for the purpose of our measurement, and we naturally need an entangled state $|\Phi\rangle$ in order to retrieve any meaningful information of the former system out of the measurement of the latter.
\paragraph{Sum and Product of Local Observables}
As for the sum and product of a pair of local observables, we note that a direct application of Lemma~\ref{lem:st_strong_commutative} leads to
\begin{equation}
A_{1} \otimes I = \int_{\mathbb{R}} a_{1}\ d\left(E_{A_{1}} \otimes E_{A_{2}}\right)(a_{1},a_{2}), \quad I \otimes A_{2} = \int_{\mathbb{R}} a_{2}\ d\left(E_{A_{1}} \otimes E_{A_{2}}\right)(a_{1},a_{2}),
\end{equation}
and subsequently
\begin{align}
\overline{A_{1} \otimes I + I \otimes A_{2}} &= \int_{\mathbb{R}^{2}} (a_{1} + a_{2})\ d\left(E_{A_{1}} \otimes E_{A_{2}}\right)(a_{1}, a_{2}), \\
A_{1} \otimes A_{2} &= \int_{\mathbb{R}^{2}} a_{1}a_{2}\ d\left(E_{A_{1}} \otimes E_{A_{2}}\right)(a_{1}, a_{2}),
\end{align}
as expected.
\subsection{Unconditioned Measurement}\label{sec:ups_II_ups}
Now that we have recalled the necessary mathematical concepts and results, we shall embark on our main analysis.
The target of our analysis is the probability measure describing the behaviour of the outcome of the composite observable $I \otimes X$ on $|\Psi^{g}\rangle$, which may be rewritten in terms of that of the local observable $X$ on the mixed state $\psi^{g}$ as
\begin{align}\label{eq:pdf_x_outcome}
\mu_{I \otimes X}^{\Psi^{g}}(\Delta)
= \mu_{X}^{\psi^{g}}(\Delta), \quad \Delta \in \mathfrak{B},
\end{align}
where the last definition $\mu_{X}^{\psi^{g}}(\Delta) := \mathrm{Tr}[E_{X}(\Delta)\psi^{g}] / \mathrm{Tr}[\psi^{g}]$ is merely a straightforward extension of probability measures \eqref{def:prob_measrue_A} for density operators (for the proof of the equality \eqref{eq:pdf_x_outcome}, just replace $X$ with $E_{X}(\Delta)$ in \eqref{eq:exp_x_outcome}).
\paragraph{Main Objective of this Subsection}
The primary interest of our study is now to investigate how the information of the target system is encoded into the profile of the outcome of the meter system \eqref{eq:pdf_x_outcome} through the interaction.
As in the previous subsection, we assume without loss of generality that the meter observable $Y$ coupled with the target observable $A$ to yield the von Neumann interaction \eqref{intro:von_Neumann_interaction} is given by $Y=P$. The main objective of the passage is to demonstrate the following proposition as an answer to this question. The results, which shall be shortly demonstrated, form the bases we rely on in conducting our further study.
\begin{proposition}[Unconditioned Measurement II.a]\label{prop:UCM_II}
In the context of the UM scheme, let $Y=P$ be fixed for definiteness, and let $|\phi\rangle \in \mathcal{H}$ and $|\psi\rangle \in \mathcal{K}$ respectively be the initial states of the target and the meter systems. Then, the probability measure \eqref{eq:pdf_x_outcome} for both the choice $X = Q, P$ reads
\begin{equation}\label{eq:outcome_prob_mod_01}
\begin{split}
\mu_{Q}^{\psi^{g}}
&= \mu_{Q}^{\psi} \ast \mu_{(gA)}^{\phi}, \\
\mu_{P}^{\psi^{g}}
&= \mu_{P}^{\psi},
\end{split}
\qquad g \in \mathbb{R},
\end{equation}
in which the resultant profile of the measurement outcomes of $X$ after the interaction can be exclusively written by the convolution of the initial profiles of both the target and the meter systems.
\end{proposition}
\noindent
Specifically, the interaction causes the change only in the profile of the outcome of the observable $X$ conjugate to $Y$, in which the initial profile of the target system acts upon that of the meter system through convolution of measures. On the other hand, the profile of $X$ for the same choice as $Y$ is left untouched. The proposition can be readily demonstrated by observing that the change of the spectral measure of the measuring observables $(I \otimes X)$ with respect to the unitary operator $U(g) := e^{-igA \otimes P}$ is provided by
\begin{align}
U(-g)\, E_{I \otimes Q}(\Delta)\, U(g) &= E_{\overline{(I \otimes Q + gA \otimes I)}}(\Delta), \quad \Delta \in \mathfrak{B}, \\
U(-g)\, E_{I \otimes P}(\Delta)\, U(g) &= E_{(I \otimes P)}(\Delta), \quad \quad \Delta \in \mathfrak{B},
\end{align}
in the Heisenberg picture (they are respectively direct consequences of \eqref{eq:weyl_analogue01} and \eqref{eq:weyl_analogue02}), and that the probability distribution dictating the probabilistic behaviour of the sum of two simultaneously measurable observables is described by the convolution of both the individual profiles of the observables involved (which is in parallel to the well-known result for random variables in classical probability theory). However, in the main passages that follow, we intend to provide a more elementary and straightforward demonstration. As a corollary to this, one equivalently has:
\begin{corollary}[Unconditioned Measurement II.b]\label{cor:UCM_II}
Under the same condition as above, the result \eqref{eq:outcome_prob_mod_01} can also be rewritten as
\begin{equation}\label{eq:outcome_prob_mod_02}
\begin{split}
\mu_{(g^{-1}Q)}^{\psi^{g}}
&= \mu_{(g^{-1}Q)}^{\psi} \ast \mu_{A}^{\phi}, \\
\mu_{(gP)}^{\psi^{g}}
&= \mu_{(gP)}^{\psi},
\end{split}
\qquad g \in \mathbb{R}^{\times},
\end{equation}
by rescaling the outcome by the interaction parameter.
\end{corollary}
\noindent
The two different manners \eqref{eq:outcome_prob_mod_01} and \eqref{eq:outcome_prob_mod_02} of describing the effect of the interaction correspond to the two possible ways of combining the interaction parameter $g$ in the unitary group as
\begin{equation}\label{eq:combining_of_the_interaction}
U(g) := e^{-i(gA)\otimes P} = e^{-i A\otimes (gP)}.
\end{equation}
Combining the interaction parameter $g$ and the target observable $A$ (the former) corresponds to the scaling of the target observable $A \to gA$, whereas combining $g$ and the meter observable $P$ (the latter) corresponds to the scaling of the pair of the meter observables $\{Q, P\} \to \{g^{-1}Q, gP\}$.
Note that the pair of scaled observables $\{g^{-1}Q, gP\}$ for $g \in \mathbb{R}^{\times}$ still satisfies the Weyl relations \eqref{eq:weyl01} and \eqref{eq:weyl02}.
Later on, we shall be investigating how one could recover the information of the target system $\mu_{A}^{\phi}$ based on the results that we obtained here. Incidentally, one finds that probing either the strong or the weak region of the interaction parameter proves itself useful for this purpose, and the equalities \eqref{eq:outcome_prob_mod_01} and \eqref{eq:outcome_prob_mod_02} shall serve as the respective starting points for analysing the weak and the strong UM schemes.
\paragraph{Preliminary Observation}
For our purpose, we first consider the case where the target observable $A$ has a finite point spectrum $\sigma(A) = \{ a_{1}, \dots, a_{N} \}$, $N \in \mathbb{N}^{\times}$. Writing the spectral decomposition of $A$ as \eqref{eq:spect_decomp_fin} and applying \eqref{eq:int_op_fin}, one finds that the composite state after the interaction reads
\begin{align}\label{eq:interaction_finite}
|\Psi^{g}\rangle
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes e^{-iga_{n}P} \right) |\phi \otimes \psi \rangle \nonumber \\
&= \sum_{n = 1}^{N} \left( |\Pi_{a_{n}}\phi\rangle \otimes |e^{-iga_{n}P} \psi\rangle \right), \quad g \in \mathbb{R}.
\end{align}
It then follows that
\begin{align}\label{eq:importantidentity}
\mu_{Q}^{\psi^{g}}(\Delta)
&= \frac{\| (I \otimes E_{Q}(\Delta)) \Psi^{g}\|^{2}}{\|\Psi^{g}\|^{2}} \nonumber \\
&= \sum_{m = 1}^{N}\sum_{n = 1}^{N} \frac{\langle \phi, \Pi_{a_{m}}\Pi_{a_{n}}\phi \rangle}{\|\phi\|^{2}} \cdot \frac{\langle e^{-iga_{m}P}\psi, E_{Q}(\Delta) e^{-iga_{n}P} \psi\rangle}{\|\psi\|^{2}} \nonumber \\
&= \sum_{n = 1}^{N} \frac{\| \Pi_{a_{n}}\phi\|^{2}}{\|\phi\|^{2}} \cdot \frac{\| E_{Q}(\Delta)e^{-iga_{n}P}\psi\|^{2}}{{\|\psi\|^{2}}} \nonumber \\
&= \sum_{n = 1}^{N} \mu_{A}^{\phi}(\{a_{n}\}) \cdot \mu_{Q}^{\psi}(\Delta - ga_{n}) \nonumber \\
&= \sum_{n = 1}^{N} \mu_{A}^{\phi}(\{a_{n}\}) \cdot \int_{\mathbb{R}} \mu_{Q}^{\psi}(\Delta - ga)\ d\delta_{a_{n}}(a) \nonumber \\
&= \int_{\mathbb{R}} \mu_{Q}^{\psi}(\Delta - ga) \ d\mu_{A}^{\phi}(a), \quad g \in \mathbb{R}, \quad \Delta \in \mathfrak{B},
\end{align}
where we have used the operator equality%
\footnote{
This is a direct result of \eqref{eq:weak_weyl}.
}
$e^{itP}E_{Q}(\Delta)e^{-itP} = E_{Q}(\Delta - t)$, $\Delta \in \mathfrak{B}$ in the third to last equality, and have applied \eqref{eq:prob_meas_fin_spec_obs} to obtain the last equality.
\paragraph{Description of the Measurement Outcome}
Returning to the general case, where the target observable $A$ is now arbitrary, we may conjecture from \eqref{eq:importantidentity} that
\begin{equation}\label{eq:ups_prob_dens_func}
\mu_{Q}^{\psi^{g}}(\Delta) = \int_{\mathbb{R}} \mu_{Q}^{\psi}(\Delta - ga) \ d\mu_{A}^{\phi}(a), \quad g \in \mathbb{R}, \quad \Delta \in \mathfrak{B}
\end{equation}
generally holds, which indeed turns out to be true; it can be shown straightforwardly in the general framework of functional analysis and measure and integration theory.
From \eqref{eq:ups_prob_dens_func}, we see that the probability measure describing the behaviour of the measurement outcome of $Q$ on the (mixed) state $\psi^{g}$ after the interaction can be explicitly given by those of the initial states of both the meter and the system. Speaking in an intuitive way, each value $a \in \sigma(A)$ of the spectrum of $A$ causes a translation $\mu_{Q}^{\psi}(\Delta) \mapsto \mu_{Q}^{\psi}(\Delta-ga)$, $\Delta \in \mathfrak{B}$ to the probability measure of the initial meter state while keeping its `shape' of the profile intact, and each of these effects is all added over, weighted by the original probability $\mu_{A}^{\phi}$ of the target observable $A$.
Parallel to this, we remark that the ideal measurement of the observable $X=P$ after the von Neumann interaction would result in
\begin{equation}
\mu_{P}^{\psi^{g}} = \mu_{P}^{\psi}, \quad g \in \mathbb{R},
\end{equation}
which states that the interaction does not alter the profile of the measurement of $X=P$ at all. This can be readily shown by changing $Q$ to $P$ in \eqref{eq:importantidentity}, and by applying the operator equality $e^{itP}E_{P}(\Delta)e^{-itP} = E_{P}(\Delta)$, $\Delta \in \mathfrak{B}$.
\paragraph{Scaling of Measures and Density Functions}
For later arguments, it proves convenient to rewrite our previous result \eqref{eq:ups_prob_dens_func} in terms of convolution of measures after introducing
some notations.
Let $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ be a complex measure, and define a parametrised family $\{\mu_{t}\}_{t \in \mathbb{R}}$ of complex measures by
\begin{equation}\label{def:measure_scaling}
\mu_{t}(B) :=
\begin{cases}
\mu(t^{-1}\Delta), & \quad t \in \mathbb{R}^{\times}, \\
\mu(\mathbb{R}^{n}) \cdot \delta_{0}(\Delta), & \quad t = 0,
\end{cases}
\qquad \Delta \in \mathfrak{B}^{n}.
\end{equation}
Note that this definition is well-defined, for the continuity of the map $x \mapsto tx$ implies its Borel-measurability, hence $t^{-1}\Delta \in \mathfrak{B}^{n}$ for $\Delta \in \mathfrak{B}^{n}$. The coefficient $\mu(\mathbb{R}^{n})$ multiplied to the delta measure for $t=0$ is to keep the total evaluation $\mu_{t}(\mathbb{R}^{n}) = \mu(\mathbb{R}^{n})$ constant for all $t \in \mathbb{R}$.
Intuitively speaking, this parametrisation allows us to narrow down the profile of a given complex measure $\mu$ while keeping its total evaluation $\mu_{t}(\mathbb{R}^{n}) = \mu(\mathbb{R}^{n})$ intact, so that it `tends' in an intuitive way to the delta measure (weighted by its total evaluation $\mu(\mathbb{R}^{n})$) as $t \to 0$. To help visualise this, suppose that $\mu$ is absolutely continuous and write $\rho := d\mu/d\beta^{n}$ for simplicity. One then finds
\begin{align}
\mu(t^{-1}\Delta)
&= \int_{(t^{-1}\Delta)} \rho(x)\ d\beta^{n}(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} \chi_{\Delta}(tx) \rho(x)\ d\beta^{n}(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} \chi_{\Delta}(x) \cdot |t|^{-n}\rho\left(\frac{x}{t}\right)\ d\beta^{n}(x), \nonumber \\
&= \int_{\Delta} \rho_{t}(x)\ d\beta^{n}(x), \quad \Delta \in \mathfrak{B}^{n},\ t \in \mathbb{R}^{\times},
\end{align}
where we have introduced the scaling
\begin{equation}\label{def:function_scaling}
f_{t}(x) := {|t|^{-n}}\, f\left(\frac{x}{t}\right), \qquad t \in \mathbb{R}^{\times},
\end{equation}
of any given integrable function $f \in L^{1}(\mathbb{R}^{n})$ by $t \in \mathbb{R}^{\times}$. This implies that $\mu_{t}$ is also absolutely continuous for each $t \in \mathbb{R}^{\times}$ by definition, and that its density is given by $\rho_{t}$, {\it i.e.},
\begin{equation}\label{eq:density_scaling}
\frac{d\mu_{t}}{d\beta^{n}} = \left(\frac{d\mu}{d\beta^{n}}\right)_{t}, \quad t \in \mathbb{R}^{\times},
\end{equation}
where the l.~h.~s. is the density of the scaled probability measure $\mu_{t}$, and the r.~h.~s. is the density of the original probability measure $\mu$ scaled by $t$ as in \eqref{def:function_scaling}. In the special case where $\mu$ is a probability measure, one may intuitively see that the parametrisation \eqref{def:function_scaling} takes any non-negative integrable function with the total integral of unity ({\it i.e.}, a probability density function) to the `delta function' in the limit $t \to 0$.
\paragraph{Von Neumann Interaction and Convolution}
Now, note here that for each $t \in \mathbb{R}^{\times}$, the probability measure $\mu_{t}$ is nothing but the image measure \eqref{def:Bildmass} of $\mu$ with respect to the map $x \mapsto tx$ ({\it i.e.}, multiplication by $t$). With the help of the change of variables formula for image measures \eqref{eq:Transformationsformel_Bildmass}, one confirms that the equality
\begin{equation}\label{eq:Transformationsformel_Skalierung}
\int_{\mathbb{R}^{n}}f(x)\ d\mu_{t}(x) = \int_{\mathbb{R}^{n}}f(tx)\ d\mu(x), \quad t \in \mathbb{R}
\end{equation}
holds for all $f$ that is integrable with respect to $\mu$. This allows us to rewrite \eqref{eq:ups_prob_dens_func} in terms of convolution as
\begin{align}\label{eq:outcome_prob01}
\mu_{Q}^{\psi^{g}}
&= \mu_{Q}^{\psi} \ast \left(\mu_{A}^{\phi}\right)_{g}, \quad g \in \mathbb{R}.
\end{align}
Alternatively, by scaling $\Delta \to g\Delta$ in \eqref{eq:ups_prob_dens_func},
one finds from the definition that
\begin{equation}\label{eq:outcome_prob02}
\left(\mu_{Q}^{\psi^{g}}\right)_{g^{-1}} = \left(\mu_{Q}^{\psi}\right)_{g^{-1}} \ast \mu_{A}^{\phi}, \quad g \in \mathbb{R}^{\times},
\end{equation}
which is another way to describe how the von Neumann type interaction causes a change in the profile of the meter observable $X=Q$.
\paragraph{Scaling of Observables}
We make a short digression at this point to seek for the physical meaning of the two findings \eqref{eq:outcome_prob01} and \eqref{eq:outcome_prob02}, which we have just acquired.
To prepare for our argument, we first introduce some notations regarding scaling of spectral measures, in parallel to that of complex measures as we have done before. Let $E : \mathfrak{B}^{n} \to L(\mathcal{H})$ be an $n$-dimensional spectral measure on the Hilbert space $\mathcal{H}$, and define a parametrised family $\{E_{t}\}_{t \in \mathbb{R}}$ of spectral measures by
\begin{equation}
E_{t}(\Delta) :=
\begin{cases}
E(t^{-1}\Delta), & \quad t \in \mathbb{R}^{\times}, \\
E_{0}(\Delta), & \quad t = 0,
\end{cases}
\qquad \Delta \in \mathfrak{B}^{n}.
\end{equation}
Here, we have introduced the `delta spectral measure' $E_{0}$ centred at $0 \in \mathbb{R}^{n}$, defined by
\begin{equation}\label{def:delta_spectral_meas_0}
E_{0}(\Delta) :=
\begin{cases}
I, & \quad 0 \in \Delta, \\
0, & \quad 0 \notin \Delta,
\end{cases}
\qquad \Delta \in \mathfrak{B}^{n}.
\end{equation}
Incidentally, for the one-dimensional case ($n=1$), the delta spectral measure $E_{0}$ centred at the origin is nothing but the spectral measure accompanying the zero operator $0$ on $\mathcal{H}$.
We next confirm some basic facts regarding scaling of observables and their accompanying spectral measures. Let $E_{A}$ be the spectral measure of a self-adjoint operator $A : \mathcal{H} \supset \mathrm{dom}(A) \to \mathcal{H}$. The goal is to specify the spectral measure of the scaled self-adjoint operator $tA$, ($t \in \mathbb{R}$) and to show that
\begin{equation}\label{eq:spectral_measure_scaled_obs}
E_{(tA)} = \left(E_{A}\right)_{t}, \quad t \in \mathbb{R},
\end{equation}
where the l.~h.~s. is the desired spectral measure accompanying the scaled operator $tA$, whereas the r.~h.~s. is the spectral measure accompanying the operator $A$ scaled by $t$. To see this, first observe the following equality
\begin{align}
\langle \phi, (tA) \phi \rangle
&= \int_{\mathbb{R}} ta\ d\mu_{A}^{\phi}(a) \nonumber \\
&= \int_{\mathbb{R}} a\ d(\mu_{A}^{\phi})_{t}(a) \nonumber \\
&= \int_{\mathbb{R}} a\ d\langle \phi, (E_{A})_{t}(a) \phi \rangle, \quad t \in \mathbb{R}
\end{align}
for the choice $|\phi\rangle \in \mathrm{dom}(A)$, where we have used \eqref{eq:Transformationsformel_Skalierung} to obtain the second to last equality. Applying the polarisation identity \eqref{def:polarisation_identity} for $T = tA$, one then has
\begin{equation}
\langle \phi^{\prime}, (tA) \phi \rangle = \int_{\mathbb{R}} a\ d\langle \phi^{\prime}, (E_{A})_{t}(a) \phi \rangle, \quad t \in \mathbb{R}
\end{equation}
for any $|\phi\rangle, |\phi^{\prime}\rangle \in \mathrm{dom}(A)$. Observing that the domain of a self-adjoint operator is dense in $\mathcal{H}$ by definition, one may continuously extend the above equality on $|\phi^{\prime}\rangle \in \mathcal{H}$, based on which the uniqueness of the spectral measure leads to the desired result \eqref{eq:spectral_measure_scaled_obs}.
Returning to our main line of arguments, we first observe that the equality \eqref{eq:spectral_measure_scaled_obs} leads to
\begin{equation}
\mu_{(tA)}^{\phi} = \left(\mu_{A}^{\phi}\right)_{t}, \quad t \in \mathbb{R},
\end{equation}
which states that the probability measure describing the ideal measurement outcome of the scaled observable $tA$ on the state $|\phi\rangle \in \mathcal{H}$ coincides with that of the original observable $A$ scaled by $t$. Armed with this result, one may reformulate our previous findings \eqref{eq:outcome_prob01} and \eqref{eq:outcome_prob02} respectively as
\begin{equation}\label{eq:outcome_prob_mod_01_proof}
\left\{
\begin{split}
\mu_{Q}^{\psi^{g}}
&= \mu_{Q}^{\psi} \ast \mu_{(gA)}^{\phi}, \\
\mu_{P}^{\psi^{g}}
&= \mu_{P}^{\psi},
\end{split}
\qquad g \in \mathbb{R},\right.
\end{equation}
and
\begin{equation}\label{eq:outcome_prob_mod_02_proof}
\left\{
\begin{split}
\mu_{(g^{-1}Q)}^{\psi^{g}}
&= \mu_{(g^{-1}Q)}^{\psi} \ast \mu_{A}^{\phi}, \\
\mu_{(gP)}^{\psi^{g}}
&= \mu_{(gP)}^{\psi},
\end{split}
\qquad g \in \mathbb{R}^{\times},\right.
\end{equation}
where we have also explicitly written down the profile of the outcome of the measurement of $X=P$. This completes our proof for Proposition~\ref{prop:UCM_II} and Corollary~\ref{cor:UCM_II}.
\subsection{Recovery of the Target Profile}
We now consider the inverse problem of what we have discussed so far, that is, we argue how one can recover the probability measure $\mu_{A}^{\phi}$ of the target observable $A$ from the probability measure $\mu_{Q}^{\psi^{g}}$ obtained through the measurement of $Q$ on the meter system. Following the same line in the previous section, one finds it useful to probe either the strong or the weak region of the interaction for this purpose, which we shall see below one by one.
\subsubsection{Strong Unconditioned Measurement}\label{sec:Strong_Unconditioned_Measurement}
We first concentrate on \eqref{eq:outcome_prob_mod_02} (or equivalently \eqref{eq:outcome_prob02}), and observe that the problem of recovering the desired probability measure reduces to the problem of `deconvolution', where one wishes to find the solution $\mu := \mu_{A}^{\phi}$ of the equation of the form
\begin{equation}\label{eq:invprob}
\nu_{\mathrm{out}} = \nu_{\mathrm{in}} \ast \mu,
\end{equation}
having knowledge and control over both the `input' $\nu_{\mathrm{in}} := \mu_{(g^{-1}Q)}^{\psi}$ and `output' $\nu_{\mathrm{out}} := \mu_{(g^{-1}Q)}^{\psi^{g}}$ on their respective sides.
Whilst there is rich literature on the topic of deconvolution, we take a specific approach to the solution in order to make our arguments simple.
\paragraph{Main Objective of this Passage}
A quick observation leads us to a na{\"i}ve expectation that, if one could attune the input so that $\nu_{\mathrm{in}}$ may become a multiplicative identity (in our case, it is the delta measure $\delta_{0}$ centred at the origin), or in the case where this is impossible, if one gradually approximates the input close enough to it, then, one may obtain the desired solution $\mu$ directly as the measured output $\nu_{\mathrm{out}} \to \delta_{0} \ast \mu = \mu$. One of the typical manners in which we attain such gradual approximation would be to fix the initial state $\psi$ and taking the strong limit $g^{-1} \to 0$ ($g \to \pm \infty$) of the interaction parameter, so that $\nu_{\mathrm{in}} = (\mu_{Q}^{\psi})_{g^{-1}}$ `tends' towards the desired identity $\delta_{0}$ in an intuitive manner (recall \eqref{def:measure_scaling} and \eqref{def:function_scaling}). The main objective of this passage is to confirm that this idea is indeed valid, and thus to state it in a mathematically rigorous way.
As it becomes apparent through the line of discussions below, there are some certain mathematical hurdles that must be overcome to achieve this objective. In order to avoid much intricacies, we shall impose certain condition to the choice of the target observable, and present our main result in the following way:
\begin{proposition}[Strong Unconditioned Measurement]
In the context of the UM scheme, suppose that
\begin{enumerate}
\item the target observable $A$ admits description by density functions,
\item the initial profile $\mu_{Q}^{\psi}$ of the meter observable $Q$ on the state $|\psi\rangle$ is compactly supported%
\footnote{We say that a complex measure $\nu$ has a compact support if there exists a compact subset $K \subset \mathbb{R}$ for which the restriction of the variation $|\nu|$ on the complement $|\nu||_{K^{c}} = 0$ is a zero measure.}.
\end{enumerate}
Then, the scaled profile of $Q$ after the interaction converges to the desired target in the strong limit of interaction
\begin{equation}
\lim_{g \to \pm \infty} \left\| \mu_{(g^{-1}Q)}^{\psi^{g}} - \mu_{A}^{\phi} \right\| = 0
\end{equation}
with respect to the total variation norm (or, equivalently the $L^{1}$-norm) for any choice of the initial states $|\phi\rangle \in \mathcal{H}$.
\end{proposition}
\noindent
The remainder of this passage is devoted to its demonstration.
\paragraph{Preliminary Observations}
Let us make a preliminary observation following the above idea. The first thing we realise is that, in general, we cannot prepare the input $\nu_{\mathrm{in}}$ so that its profile may exactly coincide with the multiplicative identity $\delta_{0}$. To see this quickly, first recall
that the realisable input probability measures $\nu_{\mathrm{in}}$ are exactly those that are absolutely continuous with respect to the Lebesgue-Borel measure. Since the delta measure $\delta_{0}$ does not belong to the space $L^{1}(\mathfrak{B})$, one concludes that it is impossible to prepare the input in such a way that $\nu_{\mathrm{in}} = \delta_{0}$ holds.
An alternative approach to this problem may be to consider a sequence of inputs $(\nu_{\mathrm{in}})_{n}$ that tends to the delta measure $\delta_{0}$ in hope that the resultant sequence of multiplicative products $(\nu_{\mathrm{out}})_{n} := (\nu_{\mathrm{in}})_{n} \ast \mu$ also converges towards the desired solution $\mu$ in the limit. Indeed, if one could only construct a sequence $(\nu_{\mathrm{in}})_{n}$ so that
\begin{equation}\label{eq:input_convergence}
\lim_{n \to \infty} \|(\nu_{\mathrm{in}})_{n} - \delta_{0}\| = 0,
\end{equation}
under the total variation norm, one concludes from the evaluation
\begin{align}\label{eq:continuity_convolution}
\|(\nu_{\mathrm{in}})_{n} \ast \mu - \mu\| = \|((\nu_{\mathrm{in}})_{n} - \delta_{0}) \ast \mu \| \leq \|(\nu_{\mathrm{in}})_{n} - \delta_{0}\| \cdot \| \mu \|
\end{align}
that the outcome tends to the desired solution
\begin{equation}\label{eq:output_convergence}
\lim_{n \to \infty} \|(\nu_{\mathrm{out}})_{n} - \mu\| = 0
\end{equation}
in the limit. Unfortunately, however, one immediately realises that this idea also fails, since in general there is no such sequence $(\nu_{\mathrm{in}})_{n}$ that meets the condition \eqref{eq:input_convergence} in the first place, for indeed, since the space $L^{1}(\mathfrak{B})$ of absolutely continuous complex measures is a topologically closed subset of the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$, a sequence in $L^{1}(\mathfrak{B})$ never converges to an element outside of $L^{1}(\mathfrak{B})$ with respect to the total variation norm.
\paragraph{Discussion on the possible Approaches}
From the quick overview of our current situation, we learn that the problem at hand is to do with the \emph{topology} we have given to the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$.
Namely, the topology induced from the total variation norm is too strong (fine) for our convenience.
A fundamental cure for this would thus be to equip the space with a weaker (coarser) topology on $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ such that, at least, it may allow us to construct sufficiently abundant sequences (or nets, in general) of the `inputs' in $L^{1}(\mathfrak{B})$ that converges towards $\delta_{0}$, and that the sequence of the resulting `outputs' ({\it i.e.}, the multiplicative product \eqref{eq:invprob}) would subsequently converge towards the desired solution in the limi
\footnote
A straightforward candidate for such a topology would be the weak-$\ast$ topology based on the identification \eqref{eq:RMK-thm} by the Riesz-Markov-Kakutani representation theorem, namely, the initial topology with respect to the family of all algebraic linear functionals of the form $\mu \mapsto \int_{\mathbb{R}} f d\mu$, where $f \in C_{0}(\mathbb{R})$.
One eventually finds that the norm topology of the total variation is nothing but the strong topology with respect to the identification, which implies that the weak-$\ast$ topology is strictly weaker than the topology we currently have at hand. Moreover, direct application of the dominated convergence theorem and Fubini's theorem reveals that the convergence of a sequence of probability measures $\nu_{n} \to \delta_{0}$ implies $\nu_{n} \ast \mu \to \mu$ (both the convergence is meant in weak-$\ast$), which is a much cleaner result than what we have seen in the main paragraphs. As an example of such a sequence (net) of probability measures converging towards $\delta_{0}$, one finds that the scaling $\nu_{t}$ \eqref{def:measure_scaling} of a given probability measure $\nu$ is typical. In fact, the scaling becomes a continuous parametrisation from $\mathbb{R}$ to $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ under the topology, which is also a welcome property.
}.
However, since this strategy, while being desirable, presupposes moderate familiarity with the mathematical branch of general topology, which the authors have deemed to be beyond the scope of this paper, an alternative approach to the problem without explicit exposure to it would be favourable (possibly at the cost of generality, while hopefully having the merit of being mathematically less demanding). In this paper, this would be accomplished by introducing an auxiliary concept of `approximate identities', whose definition would be shortly presented.
In essence, we focus only on the convergence of the output in the total variation norm, based on the observation that, even though there is \emph{no} sequence of the input that converges to the delta measure \eqref{eq:input_convergence}, there are certain conditions in which the sequence of the output \emph{do} converge towards the desired solution \eqref{eq:output_convergence}. As a preliminary observation to this approach, observe that the output $\nu_{\mathrm{out}}$ also necessarily lies in $L^{1}(\mathfrak{B})$%
\footnote{To see this, recall that the output can be written as a multiplicative product of two probability measures with one of which being absolutely continuous, and that the space $L^{1}(\mathfrak{B})$ of absolutely continuous complex measures is an ideal in $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$.},
and by recalling that $L^{1}(\mathfrak{B})$ is closed under the topology induced by the total variation norm, one finds that the candidates of the solution $\mu$ towards which the sequence of outputs could ever converge are only those that also lie in $L^{1}(\mathfrak{B})$. Based on this inspection, in what follows, we shall only treat the case in which \emph{the target observable $A$ admits a description by density functions},
which is to say that the solutions $\mu = \mu_{A}^{\phi}$ are always guaranteed to lie in $L^{1}(\mathfrak{B})$, is assumed.
\paragraph{Approximate Identities}
The convolution algebra $L^{1}(\mathfrak{B}) \cong L^{1}(\mathbb{R}^{n})$, contrasted to the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$, is non-unital.
In order to compensate the inconvenience arising from the lack of a multiplicative identity, a weaker concept is often used in analysing problems involving algebras.
In this paper, we call a family $\{e_{t}\}_{t>0}$ of elements of $L^{1}(\mathbb{R}^{n})$ an \emph{approximate identity}, if for every element $f\in L^{1}(\mathbb{R}^{n})$, the convolution $e_{t} \ast f$ converges to $f$ in the topology induced by the $L^{1}$-norm, {\it i.e.},
\begin{equation}\label{eq:appident}
\lim_{t \to 0} \|e_{t} \ast f - f\|_{1} = 0, \quad f \in L^{1}(\mathbb{R}^{n}).
\end{equation}
Before we move on to the construction of an example, we collect some necessary terminologies. Recall that the support of a function $f: \mathbb{R}^{n} \to \mathbb{K}$ is a subset of $\mathbb{R}^{n}$ defined by
\begin{equation}
\mathrm{supp}(f) := \overline{\{x \in \mathbb{R}^{n} : f(x) \ne 0\}},
\end{equation}
where the overline on a set denotes its topological closure. A support of a function $f: \mathbb{R}^{n} \to \mathbb{K}$ is said to be compact if $\mathrm{supp}(f)$ is bounded.
Now, let $\eta \in L^{1}(\mathbb{R}^{n})$ be any integrable function possessing a compact support with the total integration of unity,
\begin{equation}
\int_{\mathbb{R}^{n}} \eta(x)\ dx = 1.
\end{equation}
With this, consider a family $\{\eta_t\}_{t \in \mathbb{R}^{\times}}$ of scaled functions defined as in \eqref{def:function_scaling},
which preserve the total integration of unity for all $t \in \mathbb{R}^{\times}$.
One may then intuitively expect that $\eta_t$ tends to the `delta function' in the limit $t \to 0$ and can be used for an approximate identity,
\begin{equation}\label{ineq:approx_ident}
\lim_{t \to 0} \|\eta_t \ast f - f \|_{1} = 0,
\end{equation}
for all $f \in L^{1}(\mathbb{R}^{n})$.
To confirm that this is indeed the case, observe the inequality
\begin{align}
\|\eta_t \ast f - f \|_{1}
&:= \int_{\mathbb{R}^{n}} \left| \left( \int_{\mathbb{R}^{n}} \eta_t(y) f(x-y) \ d\beta^{n}(y) \right) - f(x) \right|\ d\beta^{n}(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} \left| \int_{\mathbb{R}^{n}} \eta_t(y) (f(x-y) - f(x))\ d\beta^{n}(y) \right|\ d\beta^{n}(x) \nonumber \\
&\leq \int_{\mathbb{R}^{n}} |\eta(y)| \left( \int_{\mathbb{R}^{n}} |f(x-ty) - f(x)|\ d\beta^{n}(x) \right)\ d\beta^{n}(y) \nonumber \\
&= \int_{\mathbb{R}^{n}} |\eta(y)| \cdot \|\tau_{(-ty)}f- f\|_{1}\ d\beta^{n}(y),
\end{align}
where $\tau_{a}$ is the translation operator defined by
\begin{equation}\label{eq:translation}
\tau_{a}f(x) := f(x + a).
\end{equation}
Recalling that $\lim_{a \to 0} \|\tau_{a}f- f\|_{1} = 0$ for any $f \in L^{1}(\mathbb{R}^{n})$, we see that for any $\epsilon > 0$, there exists a $\delta >0$ for which $a \in K_{\delta}(0) := \{x \in \mathbb{R}^{n} : |x| < \delta \}$ leads to $\|\tau_{a}f- f\|_{1} < \epsilon$. By taking $|t|$ small enough so that $\mathrm{supp}(\eta) \subset K_{t^{-1}\delta}(0)$, we find that the r.~h.~s of the above inequality is less than $\epsilon$. This shows that the family defined by
\begin{equation}
e_{t} := \eta_{(\pm t)}, \quad t > 0,
\end{equation}
makes a simple example of approximate identities (here, the meaning of the subscript on both sides of the equation is not to be confused, where the subscript on the l.~h.~s. indicates an index of the elements of the convolution algebra $L^{1}(\mathbb{R}^{n})$, whereas that on the r.~h.~s. indicates the scaling parameter of an integrable function $\eta$ defined in \eqref{def:function_scaling}).
Obviously, the construction of such approximate identities is highly non-unique, and one may attain it in various different ways.
\paragraph{Realisation of Approximate Identities}
Our observation so far revealed that, as long as the target profile $\mu \in L^{1}(\mathfrak{B})$ is absolutely continuous, by considering the family of inputs $\{(\nu_{\mathrm{in}})_{t}\}_{t > 0}$
in such a way that it makes an approximate identity in $L^{1}(\mathfrak{B})$, the resulting family of outputs $(\nu_{\mathrm{out}})_{t} := (\nu_{\mathrm{in}})_{t} \ast \mu$ would successfully converge to the desired solution
\begin{equation}
\lim_{t \to 0} \|(\nu_{\mathrm{out}})_{t} - \mu \| = 0
\end{equation}
in the $L^{1}$-norm (or equivalently, in the total variation norm)%
\footnote{We note again that the subscripts $t$ used here is meant to be an index, and not to be confused with that denoting scaling of complex measures.}.
We are now interested in the construction of such approximate identities for our current situation.
To this, we first observe that, since the profile of the input $\nu_{\mathrm{in}} = \mu_{(g^{-1}Q)}^{\psi}$ in our case is exclusively determined by the choice of the interaction parameter $g$ and the initial state $|\psi\rangle \in \mathcal{K}$ of the meter system, the problem reduces to finding a sequence of the pair
$(g, |\psi\rangle)_{t}$, $t>0$ that makes the input an approximate identity. As an example of such a construction, we first fix the initial state $|\psi\rangle$ and observe that the density of the input is given by
\begin{equation}
\frac{d\mu_{(g^{-1}Q)}^{\psi}}{d\beta} = \left(\frac{d\mu_{Q}^{\psi}}{d\beta}\right)_{g^{-1}}, \quad g \in \mathbb{R}^{\times},
\end{equation}
where we have used our previous result \eqref{eq:density_scaling}. Then, choosing $|\psi\rangle$ so that the density of $\mu_{Q}^{\psi}$ may be compactly supported, one realises that taking the strong limit of the interaction $g^{-1} \to 0$ (or equivalently $g \to \pm \infty$) yields the desired result. In turn, we fix the interaction parameter $g \in \mathbb{R}^{\times}$ and choose a sequence of initial states that makes the corresponding probability measures an approximate identity. Since the scaling of an approximate identity by $g^{-1}$ is still an approximate identity, one achieves another example of such a construction.
One thus finds a general guiding principle for the construction of an approximate identity to be the combination of the two manoeuvres, namely, either
\begin{itemize}
\item by taking the strong limit of the interaction $g^{-1} \to 0$,
\item by narrowing down the profile of the probability measures to the delta measure (symbolically $\mu_{Q}^{\psi} \to \delta_{0}$) by changing the meter state $|\psi\rangle \in \mathcal{K}$.
\end{itemize}
In order to explicitly see how these work together, choose a sequence of initial states $|\psi\rangle$, $|\psi_{h}\rangle \in \mathcal{K}$, $h >0$, such that the density of the initial profile $\mu_{Q}^{\psi}$ is compactly supported and that the parametrisation corresponds to its scaling
\begin{equation}\label{eq:approx_ident_state}
\frac{d\mu_{Q}^{\psi_{h}}}{d\beta} = \left( \frac{d\mu_{Q}^{\psi}}{d\beta} \right)_{h},
\end{equation}
which makes itself an approximate identity as $h \to 0$
(one may easily construct such a sequence in the special case in which the meter system is described in the Schr{\"o}dinger representation of the CCR%
\footnote{One may choose any wave-function $\psi \in L^{2}(\mathbb{R})$ with compact support,
and define
\begin{equation}\label{eq:approx_ident_state_Sch}
\psi_{(h)}(x) := |h|^{-1/2}\psi\left( \frac{x}{h} \right).
\end{equation}
Here, the braces among the subscript $h$ to denote the index is merely employed in order to avoid confusion with that denoting scaling of a function \eqref{def:function_scaling}. One then readily finds that this qualifies as an example of the desired family \eqref{eq:approx_ident_state}.}).
Then, observing that the scaling of it by $g^{-1}$ is
\begin{equation}\label{eq:ups_g_and_h}
\left(\frac{d\mu_{Q}^{\psi_{h}}}{d\beta}\right)_{g^{-1}} = \left(\frac{d\mu_{Q}^{\psi}}{d\beta}\right)_{hg^{-1}},
\end{equation}
one finds that it is indeed an approximate identity that tends to the delta in the limit as $hg^{-1} \to 0$ together.
\paragraph{Concluding Remarks}
In conclusion, we see that the UM scheme allows us to recover the information of the target system and its observable $A$, not only in the form of expectation values described earlier, but also in the form of probability measures $\mu_{A}^{\phi}$. This is accomplished by taking the limit of either narrowing the profile of the probability measure $\mu_{Q}^{\psi}$ of the meter system, or intensifying the interaction parameter $g \to \pm \infty$, or otherwise by appropriately balancing both contributions and having $hg^{-1} \to 0$ as a whole. In this sense, we may say that intensifying the interaction parameter has an equivalent role to narrowing the profile of the probability measure of the meter.
It may thus appear reasonable that, also in this respect, the von Neumann measurement scheme is sometimes referred to as the `strong measurement' or the `sharp measurement'.
\subsubsection{Weak Unconditioned Measurement}\label{sec:ups_II_wups}
We shall see next how the measurement outcome of the UM scheme behaves locally around $g=0$ in terms of probability measures. Specifically, we are interested in the (higher-order) derivatives of the map
\begin{equation}\label{eq:param_to_outcome}
\mathbb{R} \to L^{1}(\mathfrak{B}), \ g \mapsto \mu_{Q}^{\psi^{g}},
\end{equation}
which is now a map from the real line $\mathbb{R}$ to the space of complex measures $L^{1}(\mathfrak{B}) \subset \mathbf{M}_{\mathbb{C}}(\mathfrak{B})$.
\paragraph{Main Objective of this Passage}
The main objective of this passage is to first compute the derivatives of the map \eqref{eq:param_to_outcome} at the origin $g=0$, and subsequently argue how one may reconstruct the profile of the probability measure $\mu_{A}^{\phi}$ of our interest from the information obtained. However, as one realises in the line of discussion that follows, this involves certain mathematical intricacies. In order to avoid any difficulties and complication that may arise, we impose some restrictions to the configuration of the target and meter systems, and thus obtain the following two propositions, the first of which shall be demonstrated in the main passages below.
\begin{proposition}[Outcome of the Weak Unconditioned Measurement]\label{prop:WUCM_II}
In the context of the UM scheme, suppose that
\begin{enumerate}
\item the target profile $\mu_{A}^{\phi}$ is compactly supported,
\item the density of $\mu_{Q}^{\psi}$ belongs to the Schwartz space $d\mu_{Q}^{\psi}/d\beta \in \mathscr{S}(\mathbb{R})$.
\end{enumerate}
Then, the map \eqref{eq:param_to_outcome} is arbitrarily many times strongly differentiable in the $L^{1}$-norm (or, equivalently, in the total variation norm), and its derivatives at $g=0$ reads
\begin{equation}\label{eq:main_result_wupsm}
\left. \frac{d^{n}}{dg^{n}} \mu_{Q}^{\psi^{g}} \right|_{g=0} = \mathbb{E}[A^{n};\phi] \cdot (-D)^{n} \mu_{Q}^{\psi}, \quad n \in \mathbb{N}_{0},
\end{equation}
where $D$ denotes the operation uniquely specified through the relation
\begin{equation}\label{def:diff_measure}
d(D \nu) / d\beta := D (d\nu / d\beta),
\end{equation}
by differentiating the density of absolutely continuous complex measures $\nu \in L^{1}(\mathfrak{B})$ whose density $d\nu/d\beta \in \mathscr{S}(\mathbb{R})$ lies in the Schwartz space.
\end{proposition}
\noindent
Note that compactness of the support of $\mu_{A}^{\phi}$ implies the existence of all the higher-order moments $|\, \mathbb{E}[A^{n};\phi]\,| < \infty$ of the observable $A$, and that the Schwartz space is closed under the operation of differentiation ({\it i.e.}, $D^{n} (d\nu / d\beta) \in \mathscr{S}(\mathbb{R})$), hence both sides of \eqref{eq:main_result_wupsm} is well-defined.
Operationally, the above proposition implies that one may obtain not only the expectation value ($n=1$) of $\mu_{A}^{\phi}$, as we have found by the operator level analysis \eqref{eq:ups_weak_I} conducted in the previous section, but also its higher-order moments
\begin{equation}\label{eq:higher-order_moments_A}
\mathbb{E}[A^{n};\phi] = \int_{\mathbb{R}} a^{n}\ d\mu_{A}^{\phi}(a), \quad n \in \mathbb{N}_{0},
\end{equation}
by probing the local behaviour of the interaction around $g=0$. Incidentally, one might expect that one could recover the full profile of the original probability measure $\mu_{A}^{\phi}$ by knowing enough numbers of its higher-order moments, which in fact turns out to be positive under our assumption.
\begin{proposition}[Weak Unconditioned Measurement]
Let $A$ be self-adjoint and $|\phi\rangle \in \mathcal{H}$ for which the probability measure $\mu_{A}^{\phi}$ is compactly supported. Given another compactly supported probability measure $\mu$ on $(\mathbb{R}, \mathfrak{B})$ such that all their higher moments
\begin{equation}
\mathbb{E}[A^{n};\phi] = \int_{\mathbb{R}} a^{n}\ d\mu(a), \quad n \in \mathbb{N}_{0},
\end{equation}
coincide with those of $\mu_{A}^{\phi}$, then the two probability measures agree $\mu = \mu_{A}^{\phi}$. In other words, one may uniquely reconstruct the probability measure $\mu_{A}^{\phi}$ of the target system by knowing all the higher moments of $A$ by means of the weak UM.
\end{proposition}
\begin{proof}
In fact, this is one instance of the famous problems collectively called the classical \emph{moment problem} \cite{Hausdorff_1921_a,Hausdorff_1921_b}. We provide a sketch of the proof for our specific case at hand, and to this, we first observe that knowing all the higher-order moments \eqref{eq:higher-order_moments_A} is equivalent to knowing the integral $\int p(a)\, d\mu_{A}^{\phi}(a)$ of all polynomials $p \in P(K)$ on some compact subset $K \subset \mathbb{R}$ on which $\mu_{A}^{\phi}$ is supported. Now, choose a compact subset $K \subset \mathbb{R}$ that contains the support of both $\mu_{A}^{\phi}$ and $\mu$, {\it i.e.}, $\mu|_{K^{c}} = \mu_{A}^{\phi}|_{K^{c}} = 0$, and observe that the space of continuous functions on $K$ trivially coincide with that of continuous functions on $K$ that vanishes at infinity $C(K) = C_{0}(K)$. We thus have $C(K)^{\prime} = C_{0}(K)^{\prime} \cong \mathbf{M}_{\mathbb{C}}(\mathfrak{B}|_{K})$ by the Riesz-Markov-Kakutani representation theorem. Since the space of polynomials $P(K)$ is dense in $C(K)$ with respect to the supremum norm ({\it cf.} Stone-Weierstra{\ss} approximation theorem), one concludes that $\int_{\mathbb{R}} p(a)\, d\mu(a) = \int p(a)\, d\mu_{A}^{\phi}(a)$, $p \in P(K)$ implies $\mu = \mu_{A}^{\phi}$.
\end{proof}
\paragraph{Preliminary Observation}
We now begin our analysis. To provide some preliminary observation to this problem, we start by observing that the target of our study would be the following formal expression
\begin{equation}\label{eq:wups_prob_formal}
\left. \frac{d}{dg} \mu_{Q}^{\psi^{g}} \right|_{g=0} := \lim_{g \to 0} \frac{\mu_{Q}^{\psi^{g}} - \mu_{Q}^{\psi}}{g},
\end{equation}
in which we leave aside, just for now, all the inherent subtleties that will shortly become apparent regarding the operation of taking the limit.
Now, since the numerator of the r.~h.~s. of the above formula can be written as
\begin{equation}
\mu_{Q}^{\psi^{g}} - \mu_{Q}^{\psi} = \mu_{Q}^{\psi} \ast \left( \left(\mu_{A}^{\phi}\right)_{g} - \delta_{0} \right),
\end{equation}
one finds that the analysis of \eqref{eq:wups_prob_formal} reduces to the study of the formal expression of the form
\begin{equation}\label{eq:wups_form}
\nu_{\mathrm{out}}^{\prime}(0) := \left. \frac{d}{dt}\nu_{\mathrm{out}}(t) \right|_{t=0} = \lim_{t \to 0}\ \nu_{\mathrm{in}} \ast \frac{\mu_{t} - \delta_{0}}{t},
\end{equation}
where $\mu, \nu_{\mathrm{in}} \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ are probability measures (the latter being absolutely continuous), $\nu_{\mathrm{out}}(t) := \nu_{\mathrm{in}} \ast \mu_{t}$, and the subscript on $\mu_{t}$ denotes the scaling defined in \eqref{def:measure_scaling}.
In studying \eqref{eq:wups_form}, one might find it a decent starting point to focus on the formal expression (the right component of the above convolution)
\begin{equation}\label{eq:diff_measure_scaling}
\lim_{t \to 0}\ \frac{\mu_{t} - \delta_{0}}{t} =: \mu_{0}^{\prime}.
\end{equation}
From this, one realises that our problem is nothing but the differentiability of the map $t \mapsto \mu_{t}$ at the origin $t=0$ (recall that we have defined $\mu_{0} := \delta_{0}$ for any probability measure $\mu$),
and thus have symbolically written the limit of the above expression by $\mu_{0}^{\prime}$, temporarily leaving aside the question of its existence and well-definedness just as before. It would then be tempting to expect
\begin{equation}\label{eq:diff_conv}
\nu_{\mathrm{out}}^{\prime}(0) = \nu_{\mathrm{in}} \ast \mu_{0}^{\prime},
\end{equation}
which should resolve our main problem fairly nicely.
\paragraph{A Formal Computation of the Derivative}
Guided by the above na{\"i}ve observation, we are naturally led to consider what the derivative of the map $t \to \mu_{t}$ at $t=0$ for a given probability measure $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ would look like.
As a first step, suppose for simplicity that $\mu$ is absolutely continuous, and denote its density by $\eta := d\mu/d\beta$. Armed with our previous findings $\eta_{t} = d\mu_{t}/d\beta$, $t \in \mathbb{R}^{\times}$ regarding scaling of measures and that of its densities (see \eqref{eq:density_scaling}), we then intend to formally obtain
\begin{equation}
\mu_{0}^{\prime} = \lim_{t \to 0} \mu_{t}^{\prime}
\end{equation}
in view of density functions, by first computing its derivative at $t>0$ and then taking the limit $t \to 0$.
Now, assuming suitable differentiability and integrability conditions for the density $\eta$, one computes the derivative of the map $t \mapsto \eta_{t}$ at $t > 0$ as
\begin{align}
\lim_{h \to 0} \frac{\eta_{t + h} - \eta_{t}}{h}
&= -\frac{1}{t^{2}} \eta\left(\frac{x}{t}\right) - \frac{x}{t^{3}}(D\eta)\left(\frac{x}{t}\right) \nonumber \\
&= - D \left( \frac{1}{t}\frac{x}{t}\eta\left(\frac{x}{t}\right) \right) \nonumber \\
&= - D\left(x\eta\right)_{t}, \quad t > 0,
\end{align}
where $D := d/dx$ was the usual operation of differentiation. Then, one might be tempted to formally proceed as
\begin{align}
\lim_{t \to 0} D\left(x\eta\right)_{t}
&= D \left[ \lim_{t \to 0} \left(x\eta\right)_{t} \right] \nonumber \\
&= D \left[ \left( \int_{\mathbb{R}} x\eta\ d\beta \right) \cdot \delta_{0} \right] \nonumber \\
&= \mathbb{E}[x;\mu] \cdot D \delta_{0},
\end{align}
where we have used \eqref{def:measure_scaling} in the second equality.
The above argument implies that the derivative of the map $t \to \mu_{t}$ at the origin would appear as
\begin{equation}\label{eq:derivative_scaled_measure}
\mu_{0}^{\prime} = \mathbb{E}[x;\mu] \cdot (- D) \delta_{0},
\end{equation}
which is the `derivative of the delta measure' weighted by the expectation value of the original probability measure $\mu$.
As for the general case in which the original probability measure $\mu$ is now not necessarily absolutely continuous,
we may conjecture that, since the r.~h.~s. of \eqref{eq:derivative_scaled_measure} does not depend on the absolute continuity of the original probability measure $\mu$, the same result should hold even in the general case as well.
\paragraph{Discussion on the possible Approaches}
While we have conducted a very formal discussion above, the result in fact turns out to be true and can be made mathematically fully rigorous in the framework of the \emph{theory of generalised functions (distributions)}. In fact, it turns out that the derivative $\mu_{0}^{\prime}$ that appears in \eqref{eq:derivative_scaled_measure} is no longer a member of the space $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ of complex measures%
\footnote{Incidentally, one may recall that the (higher-order) derivatives of the delta distribution appears in several branches of physics, one of the most familiar of which being presumably the theory of electromagnetism. The derivative of the delta distribution $D \delta_{0}$ is among the most well-known example of a distribution that cannot be expressed by a complex measure. In order to provide an intuitive reasoning with the tools at hand, let $\varphi$ be a smooth function with compact support ({\it i.e}, a test function) satisfying $(D\varphi)(0) = 1$. As a concrete example, one may take $\varphi(x) := x\varphi_{0}(x)$ with
\begin{equation}
\varphi_{0}(x) :=
\begin{cases}
e^{-\frac{1}{1-x^{2}}} & (|x| < 1) \\
0 &(|x| \geq 1).
\end{cases}
\end{equation}
Defining a sequence of test functions by $\varphi_{n}(x) := n^{-1}\varphi(nx)$, $n \in \mathbb{N}^{\times}$, observe that the dominated convergence theorem necessarily implies $\lim_{n \to \infty} \int_{\mathbb{R}} \varphi_{n} d\mu = 0$ for any complex measure $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B})$. On the other hand, with the help of an auxiliary smooth density function $\rho$ to symbolically express the delta distribution by the limit of its scaling $\delta_{0} = \lim_{t \to 0} \rho_{t}$, one may formally compute the integral of $\varphi_{n}$ weighted by the `density' $D \delta_{0}$ as
\begin{align}\label{eq:distributional_delta_not_a_measure}
\int_{\mathbb{R}} \varphi_{n}(x)\ (D\delta_{0})(x) d\beta(x)
&= \lim_{t \to 0} \left( \int_{\mathbb{R}} \varphi_{n}(x)\ (D\rho_{t})(x) d\beta(x) \right) \nonumber \\
&= \lim_{t \to 0} \left( -\int_{\mathbb{R}} (D\varphi_{n})(x)\ \rho_{t}(x) d\beta(x) \right) \nonumber \\
&= -\int_{\mathbb{R}} (D\varphi_{n})(x)\ \delta_{0}(x) d\beta(x),
\end{align}
where we have used integration by parts to obtain the second equality. This implies $\lim_{n \to \infty} \int_{\mathbb{R}} \varphi_{n} (D\delta_{0})d\beta = \lim_{n \to \infty} -(D\varphi_{n})(0) = -1$, which would lead to a contradiction if $(D\delta_{0})$ were to be expressed by a complex measure.
},
and accordingly the framework in which we have been working so far ({\it i.e.}, the space of complex measures) is insufficient for our analysis.
For further study of the weak UM scheme, a preferable approach would thus be to expand our framework by introducing the space of distributions. While this method has a great merit in being able to conduct our analysis with decent generality (and in fact, distributions have their role, not just in this subsection, but also later in studying the quasi-joint-probability distributions in Section~\ref{sec:ps_II} and \ref{sec:qp_qo}), at the same time, it has a drawback in that it would be rather mathematically demanding, especially since the theory of distributions is build up on the results of general topology.
In view of this, an alternative approach to the problem without direct exposure to the theory of distributions would be favourable. To this end, recalling the idea employed in the previous subsection, we concentrate only on the differentiability of the multiplicative product \eqref{eq:wups_form}, setting aside the intricacies involving that of the map $t \mapsto \mu_{t}$ we have seen above. To see what we mean, we first expect, by combining \eqref{eq:diff_conv} and \eqref{eq:derivative_scaled_measure}, that the derivative of the map $t \mapsto \nu_{\mathrm{out}}(t)$ at the origin be written as
\begin{equation}
\nu_{\mathrm{out}}^{\prime}(0) = \mathbb{E}[x;\mu] \cdot \left( \nu_{\mathrm{in}} \ast (- D) \delta_{0} \right).
\end{equation}
Now, assuming suitable differentiability condition of the density $\rho_{\mathrm{in}} := d\nu_{\mathrm{in}}/d\beta$ of the imput $\nu_{\mathrm{in}}$ as a starting point, we employ an auxiliary smooth density function $\eta$ to symbolically express the delta distribution by the limit of its scaling $\delta_{0} = \lim_{t \to 0} \eta_{t}$ (a similar technique is used in \eqref{eq:distributional_delta_not_a_measure}) and formally obtain the `density' of the convolution $\nu_{\mathrm{in}} \ast D \delta_{0}$ as
\begin{align}
\left( \rho_{\mathrm{in}} \ast D\delta_{0}\right)(x)
&= \left( \rho_{\mathrm{in}} \ast D\left( \lim_{t \to 0} \eta_{t} \right) \right)(x) \nonumber \\
&= \lim_{t \to 0} \int_{\mathbb{R}} \rho_{\mathrm{in}}(x-y) \left(D\eta_{t}\right)(y)\ d\beta(y) \nonumber \\
&= \lim_{t \to 0} \int_{\mathbb{R}} (D\rho_{\mathrm{in}})(x-y) \eta_{t}(y)\ d\beta(y) \nonumber \\
&= (D\rho_{\mathrm{in}})(x).
\end{align}
Introducing the notation $D\nu_{\mathrm{in}}$ as defined in \eqref{def:diff_measure}, we thus obtain
\begin{align}\label{eq:diff_ups_output}
\nu_{\mathrm{out}}^{\prime}(0) = \mathbb{E}[x;\mu] \cdot (-D) \nu_{\mathrm{in}}.
\end{align}
The basic idea is that, while we have seen that the distributional derivative of the delta $D\delta_{0}$ does not allow itself to be expressed by a complex measure, the distributional derivative $D\nu_{\mathrm{in}}$ of some probability measure $\nu_{\mathrm{in}}$ might belong to the space $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$ of complex measures%
\footnote{
As one may expect, the distributional derivative $D \nu$ of an arbitrary complex measure $\nu$ can be made well-defined by extending our framework into the theory of generalised functions. In general, the derivative derivative $D \nu$ is a distribution itself (as we have seen for the special case $\nu = \delta_{0}$), but not necessarily a complex measure anymore.
}.
If we could moreover find a condition for which the differentiability \eqref{eq:diff_ups_output} is valid with respect to the norm topology of the total variation ({\it i.e.}, strongly differentiable), we could develop a line of argument that is totally confined in the space $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$, without referring to the theory of distributions at all.
\paragraph{On the Main Results}
One finds below that the the above idea is indeed valid. To this end, we assume
\begin{itemize}
\item The probability measure $\mu$ has compact support.
\item The density of $\nu_{\mathrm{in}}$ belongs to the Schwartz space $d\nu_{\mathrm{in}}/d\beta \in \mathscr{S}(\mathbb{R})$.
\end{itemize}
Under the above two conditions, we demonstrate below that the map $t \mapsto \nu_{\mathrm{out}}(t)$ is in fact arbitrarily many times strongly differentiable, and that its higher-order derivatives read
\begin{equation}\label{eq:upsm_prob_diff_gen}
\nu_{\mathrm{out}}^{(n)}(t) = ((-D)^{n}\nu_{\mathrm{in}}) \ast (x^{n} \odot \mu)_{t}, \quad t \in \mathbb{R},\ n \in \mathbb{N}_{0},
\end{equation}
which in particular implies
\begin{equation}\label{eq:upsm_prob_diff_0}
\nu_{\mathrm{out}}^{(n)}(0) = \mathbb{E}[x^{n};\mu] \cdot (-D)^{n} \nu_{\mathrm{in}}, \quad n \in \mathbb{N}_{0}
\end{equation}
at the origin $t=0$. Here, $D^{n} \nu_{\mathrm{in}}$ denotes the signed measure defined in \eqref{def:diff_measure}, and the signed measure $x^{n} \odot \mu$ is defined in \eqref{def:Mass_mit_Dichte}. Note that our two conditions above, namely, the compactness of the support of $\mu = x^{0} \odot \mu$ and the density of $\nu_{\mathrm{in}} = (-D)^{0}\nu_{\mathrm{in}}$ belonging to the Schwartz space, are true not only for $n=0$, but for all $n \in N_{0}$. Note also that compactness of the support of $\mu$ guarantees the finiteness of all its higher-order moments $|\, \mathbb{E}[x^{n};\mu] \,| < \infty$, $n \in \mathbb{N}_{0}$. Applying \eqref{eq:upsm_prob_diff_0} to our physical situation by letting $\mu = \mu_{A}^{\phi}$ and $\nu_{\mathrm{in}} = \mu_{Q}^{\psi}$ would prove Proposition~\ref{prop:WUCM_II}.
\begin{proof}[Proof of our Main Result]
For demonstration, we provide a sketch of the proof by mathematical induction. One may readily confirm by definition that the above statement is trivially true for $n=0$. Now, assuming that the statement is true for $n \in N_{0}$, we rewrite $\tilde{\nu}_{\mathrm{in}} := (-D)^{n}\nu_{\mathrm{in}}$, $\tilde{\mu} := x^{n} \odot \mu$ and $\tilde{\nu}_{\mathrm{out}}(t) := \tilde{\nu}_{\mathrm{in}} \ast \tilde{\mu}_{t}$ for better readability.
Now, recalling that the convolution algebra $L^{1}(\mathfrak{B})$ is an ideal in the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$, one finds that $\tilde{\nu}_{\mathrm{out}}(t)$ is absolutely continuous for all $t \in \mathbb{R}$ (in passing, one moreover finds that the density of $\tilde{\nu}_{\mathrm{out}}(t)$ is also a Schwartz function), and that its density $\tilde{\rho}_{\mathrm{out}}(t) := d\tilde{\nu}_{\mathrm{out}}(t)/d\beta$ is given by
\begin{equation}
\tilde{\rho}_{\mathrm{out}}(t)(x) = \int_{\mathbb{R}} \tilde{\rho}_{\mathrm{in}}(x - ty)\ d\tilde{\mu}(y), \quad t \in \mathbb{R},
\end{equation}
where $\tilde{\rho}_{\mathrm{in}}$ denotes the density of $\tilde{\nu}_{\mathrm{in}}$ (see \eqref{eq:density_ac_cmeas} for this result).
In order to prove the strong differentiability of the map $t \mapsto \tilde{\nu}_{\mathrm{out}}(t)$, we work in the space of density functions. We start by demonstrating the point-wise differentiability of the map $t \mapsto \tilde{\rho}_{\mathrm{out}}(t)$, and to this end, we fix $t_{0}, x \in \mathbb{R}$ and observe
\begin{align}\label{eq:main_diff_out_prob_point-wise}
\left( \tilde{\rho}_{\mathrm{out}}^{\prime}(t_{0})\right) (x)
&:= \lim_{t \to t_{0}} \frac{\tilde{\rho}_{\mathrm{out}}(t)(x) - \tilde{\rho}_{\mathrm{out}}(t_{0})(x)}{t - t_{0}} \nonumber \\
&= \lim_{t \to 0} \int_{\mathbb{R}} \frac{\tilde{\rho}_{\mathrm{in}}(x - ty) - \tilde{\rho}_{\mathrm{in}}(x - t_{0}y)}{t - t_{0}} \ d\tilde{\mu}(y) \nonumber \\
&= \int_{\mathbb{R}} y(-D\tilde{\rho}_{\mathrm{in}})(x - t_{0}y) \ d\tilde{\mu}(y),
\end{align}
where the exchange of the limit and integration in the second equality, while we shall omit any details of its proof, is essentially a consequence of the dominated convergence theorem. Next, we return to its strong differentiability ({\it i.e.}, differentiability with respect to the $L^{1}$-norm). To this end, we assume $t_{0} < t$ without loss of generality and recall the mean-value theorem, which state that there exists a $t_{1} \in ]t_{0}, t[$ such that
\begin{equation}
\frac{\tilde{\rho}_{\mathrm{out}}(t)(x) - \tilde{\rho}_{\mathrm{out}}(t_{0})(x)}{t - t_{0}} = \left( \tilde{\rho}_{\mathrm{out}}^{\prime}(t_{1})\right) (x)
\end{equation}
holds. Then, one has
\begin{align}
&\left\|\frac{\tilde{\rho}_{\mathrm{out}}(t)(x) - \tilde{\rho}_{\mathrm{out}}(t_{0})(x)}{t - t_{0}} - \tilde{\rho}_{\mathrm{out}}^{\prime}(t_{0}) \right\|_{1} \nonumber \\
&\qquad = \int_{\mathbb{R}} \left| \int_{\mathbb{R}} y(-D\tilde{\rho}_{\mathrm{in}})(x - t_{1}y) - y(-D\tilde{\rho}_{\mathrm{in}})(x - t_{0}y) \ d\tilde{\mu}(y) \right|\ d\beta(x) \nonumber \\
&\qquad \leq \int_{\mathbb{R}} |y| \cdot \left\|\tau_{(-t_{1}y)}(D\tilde{\rho}_{\mathrm{in}})- \tau_{(-t_{0}y)}(D\tilde{\rho}_{\mathrm{in}})\right\|_{1}\ d\tilde{\mu}(y),
\end{align}
where the exchange of the order of integration in the last inequality is guaranteed to hold (Fubini's theorem), and the translation operator $\tau_{a}$ is defined in \eqref{eq:translation}. Compactness of the support of $\tilde{\mu}$ together with an analogous argument made in \eqref{ineq:approx_ident} implies that the r.~h.~s. of the above inequality tends to $0$ as $t \to 0$, which completes our proof for strong differentiability. We thus have by \eqref{eq:main_diff_out_prob_point-wise}
\begin{align}
\rho_{\mathrm{out}}^{(n+1)}(t)
&= \tilde{\rho}_{\mathrm{out}}^{\prime}(0) \nonumber \\
&= (-D\tilde{\rho}_{\mathrm{in}}) \ast (x \odot \tilde{\mu})_{t} \nonumber \\
&= ((-D)^{n+1}\rho_{\mathrm{in}}) \ast (x^{n+1} \odot \mu)_{t}, \quad t \in \mathbb{R}
\end{align}
and
\begin{align}
\rho_{\mathrm{out}}^{(n+1)}(0)
&= ((-D)^{n+1}\rho_{\mathrm{in}}) \ast (x^{n+1} \odot \mu)_{0} \nonumber \\
&= \mathbb{E}[x^{n+1};\mu] \cdot (-D)^{n+1} \nu_{\mathrm{in}},
\end{align}
where we have used \eqref{def:measure_scaling} and $(x^{n+1} \odot \mu)(\mathbb{R}) = \mathbb{E}[x^{n+1};\mu]$ in the last equality. This completes our whole proof.
\end{proof}
\begin{comment}
\paragraph{Main Results}
One finds below that the the above idea is indeed valid. To this end, we first introduce the space
\begin{equation}
\mathscr{D}(\mathbb{R}) := \left\{ f \in C^{\infty}(\mathbb{R}) : f \text{ is compactly supported}\right\}
\end{equation}
of test functions, and assume that the density of $\nu_{\mathrm{in}}$ belongs to
\begin{equation}
\rho_{\mathrm{in}} := \frac{d\nu_{\mathrm{in}}}{d\beta} \in \mathscr{D}(\mathbb{R}).
\end{equation}
For extensive use later, we note here that the space $\mathscr{D}(\mathbb{R})$ is closed under the operation of differentiation $D$, {\it i.e.}, $f \in \mathscr{D}(\mathbb{R}) \Rightarrow Df \in \mathscr{D}(\mathbb{R})$.
Moreover, we assume that the probability measure $\mu$ has compact support%
\footnote{A signed measure $\mu$ is said to have compact support if there exists a compact subset $K \subset \mathbb{R}$ for which the restriction of $\mu$ on the complement $\mu|_{K^{c}} = 0$ is a zero measure.},
which in particular implies that all the higher-order moments are finite $|\,\mathbb{E}[x^{n};\mu]\,| < \infty$, $n \in \mathbb{N}$.
Now, recalling that the convolution algebra $L^{1}(\mathfrak{B})$ is an ideal in the measure algebra $\mathbf{M}_{\mathbb{C}}(\mathfrak{B})$, one finds that $\nu_{\mathrm{out}}(t) := \nu_{\mathrm{in}} \ast \mu_{t}$ is absolutely continuous for all $t \in \mathbb{R}$, and that the density of the output $\rho_{\mathrm{out}}(t) := d\nu_{\mathrm{out}}(t)/d\beta$ is written as
\begin{equation}
\rho_{\mathrm{out}}(t)(x) = \int_{\mathbb{R}} \rho_{\mathrm{in}}(x - ty)\ d\mu(y), \quad t \in \mathbb{R}
\end{equation}
(see \eqref{eq:density_ac_cmeas}). Then, in order to prove the strong differentiability of the output \eqref{eq:wups_form}, we work in the space of density functions, and by fixing $t_{0}, x \in \mathbb{R}$, we first observe
\begin{align}\label{eq:main_diff_out_prob_point-wise}
\lim_{t \to t_{0}} \frac{\rho_{\mathrm{out}}(t)(x) - \rho_{\mathrm{out}}(t_{0})(x)}{t - t_{0}}
&= \lim_{t \to 0} \int_{\mathbb{R}} \frac{\rho_{\mathrm{in}}(x - ty) - \rho_{\mathrm{in}}(x - t_{0}y)}{t - t_{0}} \ d\mu(y) \nonumber \\
&= \int_{\mathbb{R}} y(-D\rho_{\mathrm{in}})(x - t_{0}y) \ d\mu(y),
\end{align}
where the exchange of the limit and integration in the second equality is, while we shall omit any details of its proof, essentially a consequence of the dominated convergence theorem. Observe that this in nothing but the point-wise convergence of the fraction in the left-most hand side for each $x \in \mathbb{R}$.
Combining the result \eqref{eq:main_diff_out_prob_point-wise} and the mean-value theorem, one finds that there is an integrable function dominating $|\rho_{\mathrm{out}}(t) - \rho_{\mathrm{out}}(t_{0})/t - t_{0}|$ for all $t \in U_{0}$ for some bounded neighbourhood $U_{0}$ of $t_{0}$. The dominated convergence theorem thus implies the strong differentiability of the map $t \mapsto \rho_{\mathrm{out}}(t)$ on all $t \in \mathbb{R}$, whose derivative $\rho_{\mathrm{out}}^{\prime}(t) \in L^{1}(\mathbb{R})$ reads
\begin{equation}\label{eq:single_diff_ups_output_01}
\left(\rho_{\mathrm{out}}^{\prime}(t)\right)(x) = \int_{\mathbb{R}} y(-D\rho_{\mathrm{in}})(x - ty) \ d\mu(y),
\end{equation}
which, in another form of writing, is to say
\begin{equation}\label{eq:single_diff_ups_output_02}
\nu_{\mathrm{out}}^{\prime}(t) = (-D\nu_{\mathrm{in}}) \ast (x \odot \mu)_{t},
\end{equation}
where $D \nu_{\mathrm{in}}$ denotes the signed measure that is uniquely specified through the relation $d(D \nu_{\mathrm{in}}) / d\beta := D (d\nu_{\mathrm{in}} / d\beta)$, and the signed measure $y \odot \mu$ is defined as in \eqref{def:Mass_mit_Dichte}).
Taking $t=0$, this in particular implies
\begin{equation}
\rho_{\mathrm{out}}^{\prime}(0) = \mathbb{E}[x;\mu] \cdot (-D) \rho_{\mathrm{in}},
\end{equation}
which completes our main demonstration.
\paragraph{Higher-order Differentiability}
In passing, we note that our condition moreover assures infinite-order strong differentiability of the map $t \mapsto \nu_{\mathrm{out}}(t)$ on all $t \in \mathbb{R}$, and that the higher-order derivatives read
\begin{equation}
\nu_{\mathrm{out}}^{(n)}(t) = ((-D)^{n}\nu_{\mathrm{in}}) \ast (x^{n} \odot \mu)_{t}, \quad n \in \mathbb{N}_{0}
\end{equation}
which in particular implies
\begin{equation}
\nu_{\mathrm{out}}^{(n)}(0) = \mathbb{E}[x^{n};\mu] \cdot (-D)^{n} \nu_{\mathrm{in}}, \quad n \in \mathbb{N}_{0}
\end{equation}
at the origin $t=0$.
We provide a sketch of the proof by mathematical induction. We have already seen that this is true for $n=0$ and $n=1$. Assuming that this is true for $n \in \mathbb{N}^{\times}$, we let $\tilde{\nu}_{n} := (-D)^{n}\nu_{\mathrm{in}}$, $\tilde{\mu}_{n} := x^{n} \odot \mu$ and observe that the derivative of the output
\begin{equation}
\nu_{\mathrm{out}}^{(n)}(t) = \tilde{\nu}_{n} \ast (\tilde{\mu}_{n})_{t}
\end{equation}
can be written as a convolution of $\tilde{\nu}_{n} \in L^{1}(\mathfrak{B})$ satisfying $ d\tilde{\nu}_{n} / d\beta \in \mathscr{D}(\mathbb{R})$ and a compactly supported signed measure $\tilde{\mu}_{n}$. Since this provides the same condition as we have assumed in demonstrating the strong differentiability of the map $t \mapsto \nu_{\mathrm{out}}(t)$ before, an analogous argument also leads to the strong differentiability of the map $t \mapsto \nu_{\mathrm{out}}^{(n)}(t)$ with its derivative being
\begin{align}
\nu_{\mathrm{out}}^{(n+1)}(t)
&= (-D\tilde{\nu}_{n}) \ast (x \odot \tilde{\mu}_{n})_{t} \nonumber \\
&= ((-D)^{n+1}\nu_{\mathrm{in}}) \ast (x^{n+1} \odot \mu)_{t},
\end{align}
where we have used our result \eqref{eq:single_diff_ups_output_02} in the first equality. Taking $t=0$, we find that
\begin{align}
\nu_{\mathrm{out}}^{(n+1)}(t)
&= ((-D)^{n+1}\nu_{\mathrm{in}}) \ast \left( \tilde{\mu}_{n}(\mathbb{R}) \cdot \delta_{0} \right) \nonumber \\
&= \mathbb{E}[x^{n};\mu] \cdot (-D)^{n+1}\nu_{\mathrm{in}},
\end{align}
where we have used
\begin{equation}
\tilde{\mu}_{n}(\mathbb{R}) = \int_{\mathbb{R}}x^{n}\ d\mu(x) = \mathbb{E}[x^{n};\mu],
\end{equation}
whose integration exists due to the fact that $\mu$ is compactly supported. This completes our demonstration.
\end{comment}
\newpage
\begin{comment}
\paragraph{Preparations}
To avoid unnecessary complication,
we assume that the initial state of the meter is a Schwartz function $\psi \in \mathscr{S}(\mathbb{R}) \subset L^{1}(\mathbb{R})$ (rather than assuming the compactness of its support as we did before).
We argue below that the information of the combination $A$ and $|\phi\rangle$ of the target system is encoded into the differential coefficients of $\rho^{\psi^{g}}(x)$ at $g=0$ in such a way that the
`$k$th differential coefficient' is proportional to the `$k$th moment' $\mathbb{E}[A^{k};\phi]$ of the target observable $A$ of interest, given $|\phi\rangle \in \mathrm{dom}(A^{k})$, $k \in \mathbb{N}_{0}$.
\begin{equation}
\left(\mu_{A}^{\phi}\right)_{0}^{\prime} = \mathbb{E}[A;\phi] \cdot (D\delta_{0})
\end{equation}
\begin{align}
\left. \frac{d}{dg} \mu_{Q}^{\psi^{g}} \right|_{g=0}
&= \mu_{Q}^{\psi} \ast \left( \mathbb{E}[A;\phi] \cdot (D\delta_{0}) \right) \nonumber \\
&=
\end{align}
\paragraph{Preliminary Facts}
Before we begin our analysis, we provide a note on some useful mathematical facts.
Let $f \in L^{1}(\mathbb{R}^{n})$ such that $x_{i}f \in L^{1}(\mathbb{R}^{n})$, $i = 1, \dots, n$, and $g \in \mathscr{S}(\mathbb{R}^{n})$. One then has in general,
\begin{align}\label{eq:diff_conv}
\left. \frac{d}{dt} (f_{t} \ast g)(x) \right|_{t=0}
&= \left. \frac{d}{dt} \left( \int_{\mathbb{R}^{n}} f(y)g(x - ty)\ dy \right) \right|_{t=0} \nonumber \\
&= \int_{\mathbb{R}^{n}} f(y)\left(\left. \frac{d}{dt} g(x - ty)\right|_{t=0} \right) \ dy \nonumber \\
&= \int_{\mathbb{R}^{n}} f(y) \left\langle (-\nabla g^{*}) (x), y \right\rangle \ dy \nonumber \\
&= -\left\langle \nabla g^{*} (x), \left( \int_{\mathbb{R}^{n}} yf(y)\ dy \right) \right\rangle,
\end{align}
where we have used the gradient
$\nabla g := (\partial_{1}g, \dots, \partial_{n}g)$,
the scalar product
\begin{align}\label{eq:scalar_prod}
\langle x, y \rangle := \sum_{i=1}^{n} x_{i}^* y_{i}, \qquad x, y \in \mathbb{K}^{n},
\end{align}
and the notation,
\begin{equation}
\int_{\mathbb{R}^{n}} xf(x)\ dx := \left( \int_{\mathbb{R}^{n}} x_{i}f(x)\ dx, \dots, \int_{\mathbb{R}^{n}} x_{n}f(x)\ dx \right).
\end{equation}
Here, the exchange of integration and differentiation (the second equality in \eqref{eq:diff_conv}) is guaranteed by the result of measure and integration theory (a consequence of the dominated convergence theorem).
In the case $n=1$, one may also readily evaluate the higher-order derivatives as
\begin{align}\label{eq:derscale}
\left. \frac{d^{k}}{dt^{k}} (f_{t} \ast g)(x) \right|_{t=0}
&= \int_{\mathbb{R}^{n}} f(y)\left(\left. \frac{d^{k}}{dt^{k}} g(x - ty)\right|_{t=0} \right) \ dy \nonumber \\
&= (-1)^{k} \frac{d^{k}}{dx^{k}}g^{*}(x) \int_{\mathbb{R}} y^{k}f(y)\ dy,
\end{align}
for $x^{k}f \in L^{1}(\mathbb{R})$.
\paragraph{Local Behaviour of $\rho^{\psi^{g}}$ and the Expectation Value}
Returning to our line of argument, suppose that $|\phi\rangle \in \mathrm{dom}(A^{k})$, $k \in \mathbb{N}_{0}$ and $\psi \in \mathscr{S}(\mathbb{R})$
which implies $\rho^{\psi} \in \mathscr{S}(\mathbb{R})$. From \eqref{eq:outcome_prob01} and \eqref{eq:derscale}, we then have
\begin{align}\label{eq:ups_weak_II}
\left. \frac{d^{k}}{dg^{k}}\rho^{\psi^{g}}(x) \right|_{g=0}
&= \left. \frac{d^{k}}{dg^{k}} (\rho^{\phi}_g \ast \rho^{\psi})(x) \right|_{g=0}\nonumber \\
&= (-1)^{k} \frac{d^{k}}{dx^{k}} \rho^{\psi}(x) \int_{\mathbb{R}} a^{k}\ \rho^{\phi}(a)\, da \nonumber \\
&= \mathbb{E}[A^{k};\phi] \cdot \left(-\frac{d}{dx}\right)^{k} \rho^{\psi}(x).
\end{align}
This shows that the `$k$th differential coefficients' of the output probability density function at $g=0$ are proportional to the `$k$th moments' of the target observable $A$ of interest.
\paragraph{Connection to the Operator level Analysis}
We can also find the connection between the result on the operator level analysis \eqref{eq:ups_weak_I} and that on the probability level analysis \eqref{eq:ups_weak_II} for $k = 1$ by observing
\begin{align}
\left. \frac{d}{dg} \mathbb{E}[I \otimes X;\Psi^{g}] \right|_{g=0}
&= \left. \frac{d}{dg} \left( \int_{\mathbb{R}} x\ \rho^{\psi^{g}}(x)dx \right) \right|_{g=0} \nonumber \\
&= \int_{\mathbb{R}} x\ \left( \left. \frac{d}{dg}\rho^{\psi^{g}}(x) \right|_{g=0} \right)dx \nonumber \\
&= \mathbb{E}[A;\phi] \cdot \int_{\mathbb{R}} x\ (-\frac{d}{dx} \rho^{\psi}) dx \nonumber \\
&= \mathbb{E}[A;\phi],
\end{align}
where we have used integration by parts to obtain the last equality, which is possible in view of $x\rho^{\psi}(x) \in \mathscr{S}(\mathbb{R})$.
Observe that the family $\{\rho^{\psi}_{g^{-1}}\}_{g^{-1} > 0}$ gives an approximate identity under the assumption that the probability density function $\rho^{\psi}$ describing the behaviour of the measurement outcome of the initial meter state is compactly supported.
One then sees from \eqref{eq:appident} that the scaled probability density function obtained by the measurement of $I \otimes \hat{x}$ tends to the desired probability density function $\rho^{\phi}$ in the limit $g \to \infty$ as
\begin{align}
\lim_{g^{-1} \to 0} \rho^{\psi^{g}}_{g^{-1}}
&= \lim_{g^{-1} \to 0} \rho^{\phi} \ast \rho^{\psi}_{g^{-1}}
= \rho^{\phi},
\end{align}
with respect to the $L^{1}$-norm. To see the connection of this result to \eqref{eq:ups_op_recovery}, one simply considers
\begin{equation}
\mathbb{E}[I \otimes Q; \Psi^{g}] = \int_{\mathbb{R}} x\ \rho^{\psi^{g}}(x) \, dx, \quad g \in \mathbb{R},
\end{equation}
from which one finds
\begin{align}
\lim_{g^{-1} \to 0} \frac{\mathbb{E}[I \otimes Q; \Psi^{g}]}{g}
&= \lim_{g^{-1} \to 0} \int_{\mathbb{R}} \frac{x}{g}\ \rho^{\psi^{g}}(x) \, dx \nonumber \\
&= \lim_{g^{-1} \to 0} \int_{\mathbb{R}} x\ \rho^{\psi^{g}}_{g^{-1}}(x) \, dx \nonumber \\
&= \int_{\mathbb{R}} x\ \left( \lim_{g^{-1} \to 0} \rho^{\psi^{g}}_{g^{-1}} \right)(x) \, dx \nonumber \\
&= \int_{\mathbb{R}} x\ \rho^{\phi}(x) \, dx \nonumber \\
&= \mathbb{E}[A;\phi].
\end{align}
Note that the exchange of the limit and integration (the third equality) is a consequence of the fact that $\rho^{\psi^{g}}_{g^{-1}}$ converges with respect to the $L^{1}$-norm.
\end{comment}
\newpage
\section{Conditioned Measurement I: In Terms of Conditional Expectations}\label{sec:ps_I}
We shall next embark on our study of the measurement scheme that we call the \emph{conditioned measurement} (CM) scheme.
As the name indicates, the CM scheme involves \emph{conditioning}, where one employs the measurement of another observable on the target system on top of the UM scheme studied earlier. The CM scheme can be understood as a natural generalisation of the \emph{post-selected measurement scheme}, which has recently been attracting much attention of several groups among the physics community. While the post-selected measurement scheme
itself has been practiced for quite a while, it has caught a renewed interest since Aharonov {\it et al.}~reintroduced it with the term \emph{weak measurement} which in particular applies to the post-selected measurement in the weak limit, along with the complex quantity termed \emph{weak value} purported to be measured by it. Two sections starting from here is devoted to the analysis on the CM scheme, and by following the same line as that of the former unconditioned counterpart, we start by examining the measurement scheme in terms of conditional expectations (Section~\ref{sec:ps_I}), and subsequently in terms of conditional probabilities (Section~\ref{sec:ps_II}).
\paragraph{Organisation of this Section}
The contents of this section is organised as follows. We first provide a concise summary of some of the necessary mathematical concepts that provides us the tools for conducting the analysis. We then make a brief review on the CM scheme from a relatively general framework, and make some comments on the technique of employing conditioning (or post-selection, as a special case) in precision measurements, whose alleged advantages has recently become the topic of intensive debate. We shall then investigate how one could reclaim the information of the configuration of the target system from the the measured outcomes, and to this end, we concentrate on the behaviour of the conditional expectation of the meter observable around the weak limit $g=0$ of the interaction parameter. In parallel to the unconditional case, we call this procedure the \emph{weak conditioned measurement scheme} in this paper. We finally close this section by introducing the concept of \emph{conditional quasi-expectations} of a quantum observable given another (not necessarily simultaneously measurable) observable, as a generalisation to that of the standard conditional expectations, and examine some of their notable properties.
\subsection{Reference Materials}
In this subsection, we shall briefly recall the necessary mathematical definitions and results regarding the formal mathematical description of conditioning.
\subsubsection{Conditioning}
The essence of the CM scheme lies in the conditioning of the outcomes of a measurement of an observable $X$ of the meter system $\mathcal{K}$ by that of an additional observable $B$ of the target system $\mathcal{H}$. The quantity of interest is then the \emph{conditional expectation} of $X$ given $B$, in contrast to the UM scheme described in Section~\ref{sec:ups_I}, where the quantity of interest was the mere (unconditional) expectation value of $X$.
\paragraph{Conditional Expectation given a Sub-$\sigma$-algebra}
Since one may find the general definition of conditional expectations to be rather involved, we start by some preliminary discussion in order to ease the introduction.
Let $(\mathbb{R}^{n}, \mathfrak{B}^{n}, \mu)$ be a probability space, and let $f: \mathbb{R}^{n} \to \mathbb{R}$ be $\mu$-integrable. Given a Borel set $B \in \mathfrak{B}^{n}$ with non-vanishing probability $\mu(B) \neq 0$, one defines the \emph{conditional expectation of $f$ given the measurable set $B \in \mathfrak{B}^{n}$} by the real number
\begin{equation}\label{def:cond_exp_set_elementary}
\mathbb{E}[f|B] := \frac{\int_{B}f(x)\ d\mu(x)}{\mu(B)}.
\end{equation}
Now, let $\mathbb{R}^{n} = \cup_{i=1}^{N} B_{i}$, $B_{i} \in \mathfrak{B}^{n}$ be a decomposition of $\mathbb{R}^{n}$ into finite numbers of mutually disjoint Borel sets, and let $\mathfrak{E} := \{B_{i}\}_{i=1,\dots,N}$ denote their collection. We then define
\begin{align}\label{def:sub_algebra_elementary}
\mathfrak{A}
&:= \sigma(\mathfrak{E})
= \left\{\bigcup_{i \in I} B_{i} : I \subset \{1, \dots, N\}\right\}
\end{align}
to be the sub-$\sigma$-algebra of $\mathfrak{B}^{n}$ generated by $\mathfrak{E}$. Assuming $\mu(B_{i}) \neq 0$ for all $i=1, \dots, N$, this gives rise to an $\mathfrak{A}$-$\mathfrak{B}$ measurable function
\begin{equation}\label{def:cond_exp_elementary}
\mathbb{E}[f|\mathfrak{A}](x) := \sum_{i = 1}^{N} \mathbb{E}[f|B_{i}] \cdot \chi_{B_{i}}(x),
\end{equation}
where each $\chi_{B_{i}}$ is the characteristic function of the subset $B_{i}$. Observing that each element $A \in \mathfrak{A}$ can be expressed by a union of elements of $\mathfrak{E}$, one has
\begin{align}
\int_{A}f(x)\ d\mu(x)
&= \sum_{B_{i} \subset A} \mathbb{E}[f|B_{i}] \cdot \mu(B_{i}) \nonumber \\
&= \int_{A} \mathbb{E}[f|\mathfrak{A}](x)\ d\mu|_{\mathfrak{A}}(x), \quad \forall A \in \mathfrak{A},
\end{align}
where $\mu|_{\mathfrak{A}}$ denotes the restriction of the probability measure $\mu$ on the sub-$\sigma$-algebra $\mathfrak{A}$.
Guided by this observation, the conditional expectation of an integrable function $f$ given a sub-$\sigma$-algebra $\mathfrak{A}$ is defined in the following manner:
\begin{definition*}[Conditional expectation given a sub-$\sigma$-algebra]
Let $(\mathbb{R}^{n}, \mathfrak{B}^{n}, \mu)$ be a probability space. For a sub-$\sigma$-algebra $\mathfrak{A} \subset \mathfrak{B}^{n}$ and a $\mu$-integrable function $f: \mathbb{R}^{n} \to \mathbb{R}$, the conditional expectation of $f$ given $\mathfrak{A}$, denoted as $\mathbb{E}[f|\mathfrak{A}]$, is defined as a $\mu|_{\mathfrak{A}}$-integrable function satisfying
\begin{equation}\label{def:cond_exp}
\int_{A} f(x)\ d\mu(x) = \int_{A} \mathbb{E}[f|\mathfrak{A}](x)\ d\mu|_{\mathfrak{A}}(x), \quad \forall A \in \mathfrak{A}.
\end{equation}
The conditional expectation $\mathbb{E}[f|\mathfrak{A}]$ exists, and is unique $\mu|_{\mathfrak{A}}$-a.e.
\end{definition*}
\noindent
To see the validity of the definition, first observe that the l.~h.~s. of \eqref{def:cond_exp} defines a complex measure $A \mapsto (f \odot \mu)(A)$, $A \in \mathfrak{A}$. Since $(f \odot \mu)|_{\mathfrak{A}} \ll \mu|_{\mathfrak{A}}$, the Radon-Nikod{\'y}m theorem leads to the existence and uniqueness $\mu|_{\mathfrak{A}}$-a.e. of the conditional expectation
\begin{equation}
\mathbb{E}[f|\mathfrak{A}] := \frac{d(f \odot \mu)|_{\mathfrak{A}}}{d\mu|_{\mathfrak{A}}},
\end{equation}
which is nothing but the Radon-Nikod{\'y}m derivative (density) of the restriction $(f \odot \mu)|_{\mathfrak{A}}$ with respect to the restriction $\mu|_{\mathfrak{A}}$. Note that the conditional expectation is defined as a \emph{function} (or more precisely, an equivalent class of functions) rather than a mere number. The elementary definition \eqref{def:cond_exp_elementary}
mentioned earlier is in fact a special case of the above general definition, in which the sub-$\sigma$-algebra concerned is given by \eqref{def:sub_algebra_elementary}. The conditional expectation $\mathbb{E}[f|\mathfrak{A}]$ serves as the, so to speak, best approximation of the original function $f$ by measurable functions defined on the coarser%
\footnote{
Given two $\sigma$-algebras $\mathfrak{A} \subset \mathfrak{B}$, $\mathfrak{A}$ is said to be smaller or \emph{coarser} than $\mathfrak{B}$, and on the other hand, $\mathfrak{B}$ is said to be larger or \emph{finer} than $\mathfrak{A}$.
}
$\sigma$-algebra $\mathfrak{A} \subset \mathfrak{B}^{n}$.
\paragraph{Conditional Expectation given another Function}
We next recall the definition of the conditional expectation given another real measurable function. As above, we first provide an introductory argument. Let $(\mathbb{R}^{n}, \mathfrak{B}^{n}, \mu)$ be a probability space, and let $f: \mathbb{R}^{n} \to \mathbb{R}$ be $\mu$-integrable. Given another measurable function $g: \mathbb{R}^{n} \to \mathbb{R}$, suppose that the probability of obtaining the outcome $y \in \mathbb{R}$ of $g$ is non-vanishing $\mu(g^{-1}(y)) \neq 0$. In a similar manner as before, one may define the conditional expectation of $f$ given the outcome $y$ of $g$ as
\begin{equation}\label{def:cond_exp_outcome}
\mathbb{E}[f|g = y] := \mathbb{E}[f|g^{-1}(y)] = \frac{\int_{g^{-1}(y)}f(x)\ d\mu(x)}{\mu(g^{-1}(y))},
\end{equation}
where we have just replaced $B = g^{-1}(y)$ in \eqref{def:cond_exp_set_elementary}.
It is now tempting to construct a function $y \mapsto \mathbb{E}[f|g = y]$ that maps each of the possible outcomes of $g$ to the corresponding conditional expectation. Assuming that the function $g$ only takes a finite number of distinct outcomes $\{y_{i}\}_{i=1, \dots, N}$, $y_{i} \in \mathbb{R}$, one accordingly obtains a decomposition $\mathbb{R}^{n} = \cup_{i=1}^{N} g^{-1}(y_{i})$ of $\mathbb{R}^{n}$ into a finite number of mutually disjoint Borel sets.
Assuming moreover that $\mu(g^{-1}(y_{i})) \neq 0$ for all $i$, one obtains a well-defined measurable function
\begin{equation}
\mathbb{E}[f|g] : \mathbb{R} \to \mathbb{R},\quad y \mapsto \mathbb{E}[f|g = y],
\end{equation}
called the \emph{conditional expectation of $f$ given $g$}.
To see how this relates to the previous definition of the conditional expectation given a sub-$\sigma$-algebra, consider a general situation in which one is given a set $X$ (without a $\sigma$-algebra), a measurable space $(Y,\mathfrak{A})$ and a function $g: X \to Y$. The collection
\begin{equation}\label{def:initial_sigma_algebra}
\mathcal{I}(g) := g^{-1}(\mathfrak{A}) := \{g^{-1}(A) : A \in \mathfrak{A}\}
\end{equation}
makes itself into a $\sigma$-algebra, called the \emph{initial $\sigma$-algebra} on $X$ with respect to $g$, and it is the coarsest $\sigma$-algebra on $\mathbb{R}^{n}$ for which the map $g$ is measurable.
In the above situation, we take $(Y,\mathfrak{A}) = (\mathbb{R},\mathfrak{B}^{1})$ and define
\begin{equation}
\mathcal{I}(g) := g^{-1}(\mathfrak{B}^{1}) = \sigma\left( \mathfrak{E}\right),
\end{equation}
where we have let $\mathfrak{E} := \{g^{-1}(y_{i})\}_{i=1,\dots,N}$.
Now, since we have assumed that $\mu(g^{-1}(y_{i})) \neq 0$ for all $i$, the conditional expectation of $f$ given $\mathcal{I}(g)$ can be expressed as
\begin{equation}
\mathbb{E}[f|\mathcal{I}(g)] = \mathbb{E}[f|\sigma(\mathfrak{E})] = \sum_{i = 1}^{N} \mathbb{E}[f|g^{-1}(y_{i})] \cdot \chi_{g^{-1}(y_{i})},
\end{equation}
where the last equality is due to \eqref{def:cond_exp_elementary} by replacing $B_{i} = g^{-1}(y_{i})$. It is then fairly straightforward to see that the conditional expectations $\mathbb{E}[f|\mathcal{I}(g)]$, $\mathbb{E}[f|g]$ and the conditioning function $g$ are related to one another through the commutative diagram,
\begin{equation}\label{diagram:cond_exp}
\xymatrix{
(\mathbb{R}^{n}, \mathcal{I}(g))
\ar[rd]_{\mathbb{E}[f|\mathcal{I}(g)]}
\ar[r]^{g}
& (\mathbb{R}, \mathfrak{B}^{1})
\ar[d]^{\mathbb{E}[f|g]}\\
& (\mathbb{R}, \mathfrak{B}^{1})
}
\end{equation}
where each of the functions is measurable. In this sense, the function $\mathbb{E}[f|g]$ is understood to be nothing but the factorisation of $\mathbb{E}[f|\mathcal{I}(g)]$ by $g$. The validity of such observation for the general case is guaranteed by the following Factorisation Theorem.
\begin{theorem*}[Factorisation Theorem]
Let $X$ be a non-empty set, and let $\mathcal{I}(g) := g^{-1}(\mathfrak{A})$ be the initial $\sigma$-algebra of a map $g:X \to (Y,\mathfrak{B})$. A function $h: (X,\mathcal{I}(g)) \to (\mathbb{R},\mathfrak{B}^{1})$ is measurable if and only if there exists a measurable function $\tilde{h}: (Y,\mathfrak{B}) \to (\mathbb{R},\mathfrak{B}^{1})$ that makes the diagram
\begin{equation}
\xymatrix{
(X, \mathcal{I}(g))
\ar[rd]_{h}
\ar[r]^{g}
& (Y, \mathfrak{B})
\ar[d]^{\tilde{h}}\\
& (\mathbb{R}, \mathfrak{B}^{1})
}
\end{equation}
commute.
\end{theorem*}
\noindent
By letting $(Y,\mathfrak{B}) = (\mathbb{R},\mathfrak{B}^{1})$ and $h = \mathbb{E}[f|\mathcal{I}(g)]$, this guarantees the existence of the function $\mathbb{E}[f|g] := \tilde{h}$ that makes the desired diagram commute, even for the general case.
As for the integrability of the conditional expectation $\mathbb{E}[f|g]$, we first observe that the probability of obtaining the outcome of $g$ in a Borel set $B \in \mathfrak{B}^{1}$ is dictated by the probability measure
\begin{equation}
g(\mu)(B) := \mu(g^{-1}(B)), \quad B \in \mathfrak{B}^{1},
\end{equation}
which is nothing but the image measure of $\mu$ with respect to $g$ (see \eqref{def:Bildmass} for its definition and properties). One thus sees by the formula
\begin{align}
\int_{\mathbb{R}} \mathbb{E}[f|g]\ dg(\mu)
&= \sum_{i=1}^{N} \mathbb{E}[f|g](y_{i}) \cdot g(\mu)(\{y_{i}\}) \nonumber \\
&= \sum_{i=1}^{N} \mathbb{E}[f|g = y_{i}] \cdot \mu(g^{-1}(y_{i})) \nonumber \\
&= \sum_{i=1}^{N} \int_{g^{-1}(y_{i})}f(x)\ d\mu(x) \nonumber \\
&= \int_{\mathbb{R}^{n}}f(x)\ d\mu(x),
\end{align}
that the function $\mathbb{E}[f|g]$ is $g(\mu)$-integrable, and its expectation value coincides with the expectation value of $f$ under $\mu$, which is what one naturally expects.
Guided by the above observation, the conditional expectation of an integrable function $f$ given another measurable function $g$ is defined in the following manner:
\begin{definition*}[Conditional expectation given a measurable function]
Let $(\mathbb{R}^{n}, \mathfrak{B}^{n}, \mu)$ be a probability space, and let $f : \mathbb{R}^{n} \to \mathbb{R}$ be $\mu$-integrable. The conditional expectation of $f$ given a measurable function $g : \mathbb{R}^{n} \to \mathbb{R}$, denoted as $\mathbb{E}[f|g]$, is defined as a $g(\mu)$-integrable function that makes the diagram
\begin{equation}\label{def:conditional_expectation}
\xymatrix{
\left(\mathbb{R}^{n}, \mathcal{I}(g), \mu|_{\mathcal{I}(g)}\right)
\ar[rd]_{\mathbb{E}[f|\mathcal{I}(g)]}
\ar[r]^{g}
& \left(\mathbb{R}, \mathfrak{B}^{1}, g(\mu)\right)
\ar[d]^{\mathbb{E}[f|g]}\\
& \left(\mathbb{R}, \mathfrak{B}^{1}\right)
}
\end{equation}
commute. Its existence and uniqueness $g(\mu)$-a.e. is known to be guaranteed.
\end{definition*}
\noindent
Note that integrability of $\mathbb{E}[f|\mathcal{I}(g)]$ is due to the change of variables formula \eqref{eq:Transformationsformel_Bildmass} for image measures, and its uniqueness $g(\mu)$-a.e. is immediate by definition.
Based on the above definition, let $\mathbb{E}[f|g]$ be (a representative of) the conditional expectation of $f$ given $g$. We write
\begin{equation}
\mathbb{E}[f|g = y] := \mathbb{E}[f|g](y)
\end{equation}
to denote the \emph{conditional expectation of $f$ given the outcome $y$ of $g$}. Note that this definition is dependent on the choice of the representative and
may admit ambiguity. Indeed, for the choice $y \in \mathbb{R}$ for which the probability of obtaining the outcome of $g$ in $\{y\}$ is vanishing: $g(\mu)(\{y\}) = \mu(g^{-1}(\{y\})) = 0$, one sees that $\mathbb{E}[f|g = y]$ is \emph{indefinite} and may take \emph{any} real number.
As exemplified in here, the conditional expectation $\mathbb{E}[f|g]$ of $f$ given $g$ is appropriate to be viewed as an equivalent class of integrable functions, rather than a function alone.
\paragraph{Conditioning by Simultaneously Measurable Observables}
As in the previous section, we occasionally denote the Borel sets on $\mathbb{R}^{n}$ by $\Delta \in \mathfrak{B}^{n}$ in place of $B$ for better understanding and readability, especially in the context of quantum theory, where the confusion of the notation of $B$ with that of an operator may become a concern.
Let $A$ and $B$ be a pair of simultaneously measurable observables on a quantum system $\mathcal{H}$. We have seen that this yields a probability measure $\mu_{A,B}^{\phi}$ on $(\mathbb{R}^{2}, \mathfrak{B}^{2})$ ({\it cf.} \eqref{def:prob_measrue_A_simul}), which is interpreted as the joint-probability distribution describing the outcomes of a simultaneous measurement of $A$ and $B$ performed on the quantum system in the state $|\phi\rangle \in \mathcal{H}$. Letting $f(a,b)=\pi_{A}(a,b) := a$ and $g(a,b) = \pi_{B}(a,b) := b$ describe the measurement outcomes of each of the observables $A$ and $B$,
we shall briefly see below how the previous discussions on conditioning fits in the context of quantum mechanics.
For our purpose, assume $|\phi\rangle \in \mathrm{dom}(A)$ so that the projection $\pi_{A}(a,b) = a$ may be integrable
\begin{align}
\int_{\mathbb{R}^{2}} \pi_{A}(a,b)\ d\mu_{A,B}^{\phi}(a,b)
&= \int_{\mathbb{R}^{2}} a\ d\mu_{A}^{\phi}(a) \nonumber \\
&= \mathbb{E}[A;\phi],
\end{align}
with respect to the probability measure $\mu_{A,B}^{\phi}$. Observing that the image measure of $\mu_{A,B}^{\phi}$ with respect to the second projection
\begin{equation}
\pi_{B}\left(\mu_{A,B}^{\phi}\right)(\Delta_{B}) := \mu_{A,B}^{\phi}(\mathbb{R} \times \Delta_{B}) = \mu_{B}^{\phi}(\Delta), \quad \Delta_{B} \in \mathfrak{B}
\end{equation}
is nothing but the probability measure describing the outcome of $B$, we define the \emph{conditional expectation} $\mathbb{E}[A|B ; \phi]$ of an observable $A$ given $B$ on the state $|\phi\rangle $ as the (equivalence class of) $\mu_{B}^{\phi}$-integrable function(s)
\begin{align}\label{def:conditional_expectation_of_observables}
\mathbb{E}[A|B ; \phi] := \mathbb{E}[\pi_{A}|\pi_{B}],
\end{align}
where the r.~h.~s. is the conditional expectation of $\pi_{A}$ given $\pi_{B}$ under the probability measure $\mu_{A,B}^{\phi}$. Under the same assumption, we analogously define the conditional expectation of an observable $A$ given the outcome $b$ of an observable $B$ on the state $|\phi\rangle \in \mathrm{dom}(A)$ by
\begin{equation}
\mathbb{E}[A|B = b ; \phi] := \mathbb{E}[A|B ; \phi](b).
\end{equation}
We note again that the last definition incorporates some ambiguity, in which the number $\mathbb{E}[A|B = b; \phi]$ is not well-defined in the case where the probability that the measurement of $B$ yields the outcome $b$ is vanishing.
\subsection{Conditioned Measurement}
The CM scheme incorporates the measurements of \emph{two} observables, where the experimenter measures one local observable on the meter system and the other on the target system. In this paper, we generally define the CM scheme as the act of measuring the conditional expectation
\begin{equation}\label{def:post-selected_measurement}
\mathbb{E}[X|B ; \Psi^{g}] := \mathbb{E}[I \otimes X|B \otimes I ; \Psi^{g}]
\end{equation}
of an observable for the choice of either $X = Q$ or $X = P$ of the meter system given another observable $B$ of the target system.
Here, for better readability, we have made a little abuse of notation by writing $X$ instead of $I \otimes X$ and $B$ for $B \otimes I$. We emphasise again that the conditional expectation \eqref{def:post-selected_measurement} is defined as an \emph{equivalence class of functions} that are integrable with respect to the probability measure
\begin{equation}
\mu_{B}^{\phi^{g}} := \mu_{B \otimes I}^{\Psi^{g}},
\end{equation}
which describes the behaviour of the outcome of the measurement of the local observable $B$ on the target system. Here, we have introduced the density matrix
\begin{equation}\label{def:targetstate}
\phi^{g} := \mathrm{Tr}_\mathcal{K}[|\Psi^{g}\rangle\langle\Psi^{g}|]
\end{equation}
on the target system defined in a parallel manner as in \eqref{def:meterstate}.
For its well-definedness, we note the following statement for reference.
\begin{proposition}[Well-definedness of the Conditional Expectation]\label{prop:def_CM_I}
In the context of the CM scheme, let
\begin{enumerate}
\item If $X = Y$: $|\phi\rangle \in \mathcal{H}$, $|\psi\rangle \in \mathrm{dom}(X)$
\item If $X \neq Y$: $|\phi\rangle \in \mathrm{dom}(A)$, $|\psi\rangle \in \mathrm{dom}(X)$
\end{enumerate}
be the choice of the initial states of the target and meter systems. Then, the conditional expectation $\mathbb{E}[X|B ; \Psi^{g}]$ is well-defined for all range of the interaction parameter $g \in \mathbb{R}$.
\end{proposition}
\begin{proof}
For demonstration, we shall only refer to Proposition~\ref{prop:UCM_I} that guarantees the integrability of the outcomes of the measurement of $X$ ({\it i.e.}, $|\,\mathbb{E}[I \otimes X; \Psi^{g}]\,| < \infty$) for all range of $g \in \mathbb{R}$, given the conditions assumed.
\end{proof}
\paragraph{Post-selected Measurement}
As a special subclass of this measurement scheme, we prepare the term \emph{post-selected measurement scheme} to refer to the case where the conditioning observable $B = |\phi^{\prime}\rangle\langle\phi^{\prime}|$ happens to be a projection on some one-dimensional subspace of $\mathcal{H}$ spanned by some normalised vector $|\phi^{\prime}\rangle \in \mathcal{H}$, and in such a case, the act of conditioning will be occasionally referred to as the \emph{post-selection}. It is also a common practice found in various literatures to call the state $|\phi\rangle$ prepared prior to the measurement the \emph{initial} or the \emph{pre-selected} state, and the normalised vector $|\phi^{\prime}\rangle$ spanning the image of the one-dimensional projection $B = |\phi^{\prime}\rangle\langle\phi^{\prime}|$ the \emph{final} or the \emph{post-selected} state.
\subsubsection{Topic: `Amplification Technique' by Conditioning}\label{sec:psI_amplification}
It is widely known that, in general, the range of conditional expectation may exceed the (unconditional) expectation value, {\it i.e.}, for some clever choice of the conditioning observable $B$ and its outcome $b \in \mathbb{R}$, one has
\begin{equation}\label{eq:amplification_by_conditioning}
\big\vert \, \mathbb{E}\left[X; \Psi^{g} \right] \big\vert \leq \big\vert\, \mathbb{E}\left[X | B=b ; \Psi^{g} \right] \big\vert
\end{equation}
with non-vanishing probability. Clearly, this property should prove itself useful in some certain situations.
While this property has occasionally been utilised in experiments, it has recently caught wide attention due to the reports on the success of application in precision measurements, including the experimental detection of the spin-Hall effect of light (SHEL) in 2008 \cite{Hosten_2008}, and the detection of an ultra-sensitive beam deflection in a Sagnac interferometer in 2009 \cite{Dixon_2009}. The experiments have effectively utilised the technique of conditioning (or post-selection) to yield an enhancement (or `amplification') of an extremely small beam displacement to the extent that it is large enough to overcome various technical imperfections (noise level), and eventually realising significant detection of such tiny effects. In this context,
this technique has often been referred to as the `weak value amplification' or as `Aharonov-Albert-Vaidman effect' of amplification \cite{Aharonov_1988}.
\paragraph{Review of the Recent theoretical Analyses}
Extensive theoretical analyses have been conducted in recent years from various viewpoints
on the technical advantages of the technique of post-selection over the conventional unconditioned counterpart.
Some of them addressed the question of signal amplification and its limit, where
one asks the question as to what extent one can amplify the signal \cite{Koike_2011} and how one could achieve the optimisation \cite{Susa_2012}; the question of the existence of the limit of amplification will be addressed shortly in a more general framework.
As far as the authors are aware of, the first sound analytic result appeared around 2012 \cite{Nakamura_2012}, in which the limit to the amplification rate, as well as the signal-to-noise ratio has been explicitly presented. The computation was conducted for a special case where the observable $A$ fulfils the condition $A^2 = I$ and the meter wave functions were assumed to be of Gaussian states, which we shall also address in a relatively more general setting later in this section, and also in Appendix~\ref{sec:PSM}.
Others focused on the statistical loss which occurs due to the post-selection and examine the feasibility of improving the parameter estimation of the coupling constant $g$ by post-selection based on estimation theory (for a concise review on the topic form this point of view, see \cite{Knee_2014}). The result is that the post-selection statistically deteriorates the quality of estimation, both in the case where ideal noiseless experiments can be performed \cite{Tanaka_2013}, and also in some case where certain types of fully-known or controllable noise are present \cite{Knee_2013_1,Knee_2013_2,Ferrie_2013}.
In an attempt to address the question of how the post-selection technique, while being statistically inferior to the unconditioned case, could be advantageous in realistic experiments, the authors have conducted a theoretical analysis on post-selected measurement in the presence of some intractable `measurement uncertainty', a relatively modern concept in metrology to express \emph{unknown} or \emph{uncontrollable} source of technical imperfections \cite{Lee_2014}. It was then found that, while post-selection suffers from statistical deterioration, in certain cases the amplification effect becomes favourable in overcoming the unknown/uncontrollable source of technical imperfections one could not completely eliminate through `noise hunting', which accordingly cannot be reduced from statistical reiteration. This suggests that the post-selection technique should be understood as the \emph{practice of taking advantage of the trade-off relation} between the reduced contribution from intractable source of measurement uncertainty due to its signal amplification effect, and the statistical deterioration caused by the decrease in success probability.
\subsubsection{Topic: `Limit of Amplification' in Terms of Essential Suprema}
In what follows, we provide a somewhat general result regarding the question of `limit of amplification' by conditioning, which has been one of the hottest topics among the study of the technical advantages in employing conditioning in experiments. A typical way to address this problem is to ask oneself, to what extent one could enlarge the conditional expectation $\mathbb{E}[X|B ; \Psi^{g}]$ by choosing an appropriate conditioning observable $B$ and its outcome $b \in \sigma(B)$ with non-vanishing probability. By recalling the definition of essential supremum of a function \eqref{def:ess_sup}, one realises that the question is equivalent to asking to what extent one could make the essential supremum of the conditional expectation
\begin{equation}
\big\|\,\mathbb{E}[X|B ; \Psi^{g}]\,\big\|_{\infty}
\end{equation}
large by the choice of the conditioning observable $B$.
\paragraph{Preliminaries}
To prepare for our arguments, we first observe some basic facts regarding absolute continuity and essential suprema.
\begin{lemma}
Let $(X,\mathfrak{A},\mu)$ be a probability space, and let $\nu: \mathfrak{A} \to \mathbb{C}$ be a complex measure. Then, the following conditions are equivalent:
\begin{enumerate}
\item $\nu \ll \mu$.
\item $|\nu| \ll \mu$.
\item There exists a non-negative number $M \in [0,\infty]$ such that
\begin{equation}\label{eq:abs_cont_cond}
|\nu(A)| \leq |\nu|(A) \leq M \cdot \mu(A)
\end{equation}
holds for all $A \in \mathfrak{A}$.
\end{enumerate}
In such a cases, the Radon-Nikod{\'y}m derivative $d\nu/d\mu$ exists by the Radon-Nikod{\'y}m theorem, and its essential supremum $\|d\nu/d\mu\|_{\infty}$
gives the smallest of such $M$ that satisfies \eqref{eq:abs_cont_cond}.
\end{lemma}
\begin{proof}
For the equivalence of the condition $(i) \Leftrightarrow (ii)$, the reader is referred to any textbooks on measure and integration theory. We already know from the Reference Material in Section~\ref{sec:ups_II_pre} that $|\nu(A)| \leq |\nu|(A)$, $A \in \mathfrak{A}$. The implication $(ii) \Rightarrow (iii)$ is then trivial by simply taking $M=\infty$. The converse $(iii) \Rightarrow (ii)$ is also immediate by the definition of absolute continuity. Now that we have proved the equivalence of the three conditions, we move on to the demonstration of the final statement. To this end, first observe the evaluation
\begin{align}
|\nu(A)|
&= \left| \int_{A} \frac{d\nu}{d\mu}\ d\mu \right| \nonumber \\
&\leq \int_{A} \left| \frac{d\nu}{d\mu} \right| d\mu \leq \left\|\frac{d\nu}{d\mu}\right\|_{\infty} \cdot \mu(A).
\end{align}
Combining this with the minimality of the variation $|\nu|$, one sees that the choice $M = \|d\nu/d\mu\|_{\infty}$ of the upper bound satisfies \eqref{eq:abs_cont_cond}. Now, suppose that there exists a non-negative number $0 \leq M < \|d\nu/d\mu\|_{\infty}$ satisfying \eqref{eq:abs_cont_cond}. Then, by definition of the essential supremum, there exists a measurable set $A$ satisfying $0 < \mu(A)$ and $M < |d\nu/d\mu||_{A}$ (just take $A := \{x \in X : M < |d\nu/d\mu|(x) \}$), hence
\begin{equation}
|\nu|(A) = \int_{A} \left| \frac{d\nu}{d\mu} \right|\ d\mu > M \cdot \mu(A),
\end{equation}
which contradicts the minimality of $|\nu|$.
\end{proof}
\noindent
As a corollary to this, the following observation is of special interest.
\begin{corollary}[Conditional Expectations and Essential Suprema]\label{cor:cond_exp_and_ess_sup}
Let $(X, \mathfrak{A}, \mu)$ be a probability space, $f: X \to \mathbb{R}$ be $\mu$-integrable, and $\mathfrak{B} \subset \mathfrak{A}$ be a sub-$\sigma$-algebra. Then the evaluation
\begin{equation}
\|\, \mathbb{E}[f|\mathfrak{B}]\,\|_{\infty} \leq \|f\|_{\infty}
\end{equation}
holds. As a direct consequence, if moreover a measurable function $g: X \to \mathbb{R}$ is given, the evaluation
\begin{equation}
\|\, \mathbb{E}[f|g]\,\|_{\infty} \leq \|f\|_{\infty}
\end{equation}
naturally holds.
\end{corollary}
\begin{proof}
First recall that the conditional expectation $\mathbb{E}[f|\mathfrak{B}]$ is nothing but the Radon-Nikod{\'y}m derivative of the complex measure $f \odot \mu$ with respect to the restriction $\mu|_{\mathfrak{B}}$. Letting $\nu := f \odot \mu$ and replacing $\mu$ by $\mu|_{\mathfrak{B}}$ in the above Lemma, one finds
\begin{equation}
|\nu(A)| = \left| \int_{A} f\ d\mu \right| \leq \|f\|_{\infty} \cdot \mu(A),
\end{equation}
hence
\begin{equation}
\|d\nu/d\mu\|_{\infty} = \|\, \mathbb{E}[f|\mathfrak{B}]\,\|_{\infty} \leq \|f\|_{\infty},
\end{equation}
which was to be demonstrated.
\end{proof}
\noindent
In casual language, this is to say that each value of the conditional expectation of $f$ never exceeds the maximum number that $f$ takes under a given probability measure, which is a result that should be intuitively clear. As a direct application of the result in the context of quantum measurement of a pair of simultaneously measurable observables $A$ and $B$, this reduces to the following.
\begin{corollary}\label{cor:cond_exp_obs_ess_sup}
Given a pair of strongly commuting self-adjoint operators $A$ and $B$ and a fixed state $|\phi\rangle \in \mathrm{dom}(A)$, the essential supremum of the conditional expectation of $A$ given $B$ is never greater than
\begin{equation}
\|\, \mathbb{E}[A|B;\phi] \,\|_{\infty} \leq \|A\|_{\infty}^{\phi},
\end{equation}
where $\|A \|_{\infty}^{\phi} := \|a\|_{\infty}$ denotes the essential supremum of the measurable function $a \mapsto a$ under the probability measure $\mu_{A}^{\phi}$ describing the behaviour of the outcome of the measurement of $A$ on the state $|\phi\rangle$. If $A$ happens to be bounded, its operator norm%
\footnote{For a bounded operator $X$, recall that the \emph{operator norm} of $X$ is defined by
\begin{equation}
\|X\| := \sup\{ \|X\phi\| : \|\phi\| = 1 \}.
\end{equation}
}
$\|A\|$ becomes the universal ({\it i.e.}, state independent) upper bound of $\|A \|_{\infty}^{\phi}$, hence
\begin{equation}
\|\, \mathbb{E}[A|B;\phi] \,\|_{\infty} \leq \|A \|_{\infty}^{\phi} \leq \| A\| < \infty
\end{equation}
holds for all $|\phi\rangle \in \mathcal{H}$.
\end{corollary}
\begin{proof}
The former part of the statement is immediate by Corollary~\ref{cor:cond_exp_and_ess_sup}. For the latter part, we first recall that the \emph{numerical range} of a self-adjoint operator $X$ is defined as
\begin{equation}
W(X) := \{ \langle \psi, X \psi \rangle : |\psi\rangle \in \mathrm{dom}(X), \|\psi\|^{2} = 1 \},
\end{equation}
which is nothing but the collection of all possible expectation values of $X$. Now, a direct application of the Cauchy-Schwarz inequality leads to
\begin{equation}
\left|\, \mathbb{E}[X;\phi] \,\right| \leq \|X\|, \quad \mathbb{E}[X;\phi] \in W(X),
\end{equation}
for bounded $X$, and by recalling the basic relation $\sigma(X) \subset \overline{W(X)}$, where the overline on $W(X)$ denotes its topological closure, one concludes
\begin{align}
\| X \|_{\infty}^{\phi}
&\leq \sup\{|x| : x \in \sigma(X)\} \nonumber \\
&\leq \sup\{|x| : x \in \overline{W(X)}\} \leq \|X\|,
\end{align}
which was to be demonstrated.
\end{proof}
\noindent
The latter part of the statement is to say that conditional expectations of a bounded observable has a universal upper bound given by its operator norm, which is also a result that should be intuitively clear.
\paragraph{On the `Limit of Amplification' by Conditional Measurement}
As a direct application of the above corollary to our problem, we obtain the main result of this passage.
\begin{proposition}[Amplification by Conditioning]\label{prop:limit_of_amplification}
Under the framework of the CM scheme, the essential supremum of the conditional expectation of $X$ given $B$ is never greater than that of the UM scheme of $X$
\begin{equation}\label{ineq:limit_of_amplification_01}
|\, \mathbb{E}[X|B = b;\Psi^{g}]\,| \leq \|\, \mathbb{E}[X|B;\Psi^{g}]\,\|_{\infty} \leq \|X\|_{\infty}^{\psi^{g}},
\end{equation}
where $\|X\|_{\infty}^{\psi^{g}}:= \|x\|_{\infty}$ denotes the essential supremum of $x$ under the probability measure $\mu_{X}^{\psi^{g}}$ describing the behaviour of the outcome of the local measurement $X$ on the meter system. In other words, $\|X\|_{\infty}^{\psi^{g}}$ gives the (conditioning-observable-independent) upper bound to the extent the conditional expectation can be `amplified' by means of conditioning%
\footnote{Recall the inherent subtlety when we use the expression $\mathbb{E}[X|B = b;\Psi^{g}]$. The left most inequality in \eqref{ineq:limit_of_amplification_01} should thus be understood to hold $\mu_{B}^{\phi^{g}}$-a.e.}.
\end{proposition}
\noindent
In physical terms, this is to say that the extent one may `amplify' the conditional expectation $\mathbb{E}[X|B ; \Psi^{g}]$ by means of changing the conditioning observable $B$ is predetermined by $\|X\|_{\infty}^{\psi^{g}}$. This is one general form to answer the question of the existence of the limit of `amplification' by conditioning.
As the next step, one might eventually be interested in seeking for the condition under which $\|X\|_{\infty}^{\psi^{g}}$ is bounded from above, even if we could freely choose the initial state $|\phi\rangle$ of the target system. This would create a universal upper bound of $\|\, \mathbb{E}[X|B;\Psi^{g}]\,\|_{\infty}$ that is indifferent to both the initial and final configurations of the target system ({\it i.e.}, the choice of the initial target state $|\phi\rangle$ and the conditioning observable $B$). As we have learned from the discussions above, this would typically be the case when there exists a subspace $U(g,\psi) \subset \mathcal{H} \otimes \mathcal{K}$, for fixed $g \in \mathbb{R}$ and $|\psi\rangle \in \mathrm{dom}(X)$, such that $|\Psi^{g}\rangle \in U(g,\psi)$ for all $|\phi\rangle \in \mathrm{dom}(A)$, and that the restriction of $I \otimes X$ on $U(g,\psi)$ is bounded.
\begin{proposition}[Limit of Amplification by Conditioning]\label{prop:limit_of_amplification_typ}\label{prop:lim_amp_example}
Under the framework of the CM scheme, let both the interaction parameter $g \in \mathbb{R}$ and the initial meter state $|\psi\rangle \in \mathrm{dom}(X)$ be fixed, and suppose that the target observable $A$ has a spectrum $\sigma(A) = \{a_{1}, \dots, a_{N}\}$, $N \in \mathbb{N}^{\times}$ of finite cardinality. Then, the following facts hold:
\begin{enumerate}
\item The density operator $\psi^{g}$ of the meter system \eqref{def:meterstate} can be written as a probabilistic mixture of a finite number of projection operators (pure states) supported on the finite-dimensional (at most $N$-dimensional) subspace
\begin{equation}\label{eq:fin_dim}
\mathcal{K}(g,\psi) := \mathrm{span}(\{ |e^{-iga_{1}Y} \psi \rangle, \dots, |e^{-iga_{N}Y} \psi \rangle\}),
\end{equation}
which is independent of the initial choice $|\phi\rangle \in \mathcal{H}$ of the target state.
\item The restriction $X|_{\mathcal{K}(g,\psi)}$ of the meter observable $X$ on the subspace \eqref{eq:fin_dim} is bounded, and thus its operator norm
\begin{equation}
\|\, \mathbb{E}[X|B;\Psi^{g}]\,\|_{\infty} \leq \left\|X|_{\mathcal{K}(g,\psi)}\right\| < \infty
\end{equation}
provides a finite universal upper bound to the conditional expectation that is independent of the configuration of the target system ({\it i.e.}, the choice of the initial state $|\phi\rangle \in \mathcal{H}$ and that of the conditioning observable $B$).
\end{enumerate}
\end{proposition}
\begin{proof}
Under the above condition, first observe that
\begin{align}
|\Psi^{g}\rangle
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} \otimes e^{-iga_{n}Y} \right) | \phi \otimes \psi \rangle \nonumber \\
&= \sum_{n = 1}^{N} \left( \Pi_{a_{n}} |\phi\rangle \otimes |e^{-iga_{n}Y}\psi\rangle \right),
\end{align}
where we have used \eqref{eq:int_op_fin}. One readily finds from the above formula that the density operator
\begin{equation}
\psi^{g} = \mathrm{Tr}_{\mathcal{H}}\left[ |\Psi^{g}\rangle \langle\Psi^{g}| \right],
\end{equation}
defined as in \eqref{def:meterstate}, can indeed be written as a probabilistic mixture of a finite number of projection operators (pure states) supported on the subspace \eqref{eq:fin_dim}.
We then recall that any operator $X$ defined on a finite-dimensional Hilbert space are necessarily bounded, and thus observe that the current problem at hand reduces to the situation of Corollary~\ref{cor:cond_exp_obs_ess_sup}.
\end{proof}
\noindent
In physical terms, this is to say that there exists a finite limit $\left\|X|_{\mathcal{K}(g,\psi)}\right\| < \infty$ to the extent one may `amplify' the conditional expectation $\mathbb{E}[X|B ; \Psi^{g}]$ by means of only changing the configuration of the target system (namely, by changing either or both the conditioning observable $B$ and the initial state $|\phi\rangle \in \mathcal{H}$ of the target system). Specifically, the evaluation
\begin{equation}\label{eq:limit_of_amp_explicit}
\big|\, \mathbb{E}[X|B = b; \Psi^{g}] \,\big| \leq \left\|X|_{\mathcal{K}(g,\psi)}\right\| < \infty
\end{equation}
holds for all $b \in \sigma(B)$ up to a set of probability zero, and the upper bound $\left\|X|_{\mathcal{K}(g,\psi)}\right\|$ does not depend on the choice of $B$ nor $|\phi\rangle$.
Naturally, if one could change either the interaction parameter $g$ or the initial state $|\psi\rangle$ of the meter system alongside, the above result is no more valid.
\subsection{Recovery of the Target Profile}
Parallel to the study of the UM scheme, we are now interested in the information of the target system which is to be extracted from the CM scheme.
Following the line of arguments for the UM scheme, we are specifically interested in investigating the local behaviour of the outcome of the CM scheme around $g=0$, {\it i.e.}, the \emph{weak conditioned measurement}, in which the target of our analysis is the map
\begin{equation}
g \mapsto \mathbb{E}[X | B; \Psi^{g}]
\end{equation}
from the interaction parameter $g$ to the conditional expectation of $X$ given $B$, which was in general defined as a map from the real line to an equivalent class of functions. To this end, we first conduct a preliminary observation.
\subsubsection{Preliminary Observation}\label{sec:ps_I_wps}
Since the definition of the conditional expectation is given in a rather abstract way, the conditional expectation \eqref{def:post-selected_measurement} in general does not admit an explicit expression by vectors and operators (in contrast to the UM case \eqref{prop:UCM_I_formula}, which always admits such an explicit expression). In view of this, it would be sometimes helpful if one could find a condition for which the conditional expectation \eqref{def:post-selected_measurement} of our interest may be explicitly written down. We first point out that this will be indeed the case given that the spectrum of the conditioning observable $B$ has finite cardinality.
Now, let
\begin{equation}\label{eq:spectr_decomp_B_fin}
B = \sum_{n=1}^{N} b_{n} \,\Pi_{b_{n}}
\end{equation}
be the spectral decomposition of $B$, where $\sigma(B) = \{b_{1}, \dots, b_{N}\}$ is any enumeration of its eigenvalues, and $\Pi_{b} := E_{B}(\{b\})$, $b \in \sigma(B)$ denotes the unique projection on the eigenspace associated to it. It is then fairly straightforward to see by definition that the conditional expectation of $X$ given $B$ is explicitly given by
\begin{align}\label{eq:cond_exp_B_fin}
&\mathbb{E}[X|B = b ; \Psi^{g}] \nonumber \\
& \quad = \begin{cases}
{\mathbb{E}\left[\Pi_{b} \otimes X; \Psi^{g} \right] / \left\|(\Pi_{b} \otimes I)\Psi^{g}\right\|^{2}}, & \quad (b \in \sigma(B),\ \left\|(\Pi_{b} \otimes I)\Psi^{g}\right\|^{2} \neq 0), \\
\text{indefinite}, & \quad (\text{else}).
\end{cases}
\end{align}
Here, recall that conditional expectations are defined as an equivalence class of functions, and hence its value for the outcome $b$ of the measurement of the observable $B$ such that the probability of observing it is vanishing, is indefinite by definition.
The study of the weak CM scheme then reduces to the analysis of the map
\begin{equation}\label{def:wpsm_special}
g \mapsto \mathbb{E}[X | B = b; \Psi^{g}]
\end{equation}
for each $b \in \sigma(B)$ such that the probability of observing it is non-vanishing. Since this is a map from the real line to itself ({\it i.e.}, a function), it should be a much more familiar and straightforward object to deal with.
\paragraph{Objective of this Passage}
In what follows, we will be discussing the differentiability of the function \eqref{def:wpsm_special} at the point $g=0$.
To this end, first observe that the choice of $b \in \sigma(B)$ for which the probability of observing it is non-vanishing is dependent on $g$. Hence, for each $b \in \sigma(B)$, we must first guarantee its well-definedness, at least on some neighbourhood of $g=0$.
Fortunately, this is indeed the case for the choice $b \in \sigma(B)$ such that the probability of finding it on the initial state $|\phi\rangle$ of the target system $\mathbb{E}\left[ \Pi_{b} \otimes I; \Psi^{0} \right] = \|\Pi_{b} \phi \|^{2} \neq 0$ is non-vanishing, due to continuity of the function $g \mapsto \mathbb{E}\left[ \Pi_{b} \otimes I; \Psi^{g} \right]$. The main objective of this passage is to demonstrate the following statement.
\begin{proposition}[Differentiability of the Conditional Expectation: Preliminary]\label{prop:diff_cond_exp}
Suppose that the conditioning observable $B$ has spectrum of finite cardinality, and moreover let $|\phi\rangle \in \mathrm{dom}(A)$, $|\psi\rangle \in \mathcal{D} \subset \mathrm{dom}(X)$ (the subspace $\mathcal{D}$ is defined as in \eqref{eq:weyl_subspace}) be assumed, so that the conditional expectation $\mathbb{E}[X|B;\Psi^{g}]$ is well-defined for all range of $g \in \mathbb{R}$. Then for $b \in \sigma(B)$ such that $\|\Pi_{b} \phi \|^{2} \neq 0$, the conditional expectation $\mathbb{E}[X|B = b;\Psi^{g}]$ is well-defined on some neighbourhood of $g=0$. It is moreover differentiable with respect to $g$ at the origin, for which the differential coefficient reads
\begin{align}\label{eq:wpsm_diff}
&\left. \frac{d}{dg} \mathbb{E}[X|B = b; \Psi^{g}] \right|_{g=0} \nonumber \\
& \qquad = 2\,\mathrm{Re} \left[\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}\right] \cdot \mathbb{CV}_{\mathrm{A}}[X,Y; \psi] +
2\,\mathrm{Im}\left[\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}\right] \cdot \mathbb{CV}_{\mathrm{S}}[X,Y; \psi].
\end{align}
\end{proposition}
\noindent
Here, we have introduced the quantities,
\begin{align}\label{def:q_covariance}
\mathbb{CV}_{\mathrm{S}}[X,Y; \psi] &:= \mathbb{E}[\{X,Y\}/2; \psi] - \mathbb{E}[X; \psi]\mathbb{E}[Y; \psi], \\
\mathbb{CV}_{\mathrm{A}}[X,Y; \psi] &:= \mathbb{E}[[X,Y]/(2i); \psi],
\end{align}
occasionally called the symmetric and anti-symmetric (quantum) covariance%
\footnote{
Note that in the case where the two observables coincide $X=Y$, the symmetric quantum covariance reduces to the familiar variance,
\begin{equation}\label{def:variance}
\mathbb{CV}_{\mathrm{S}}[X,X; \psi] = \mathbb{V}[X;\psi] := \mathbb{E}[X^{2};\phi] - \mathbb{E}[X;\phi]^{2},
\end{equation}
which is reminiscent of the familiar result in classical probability theory, whereas the anti-symmetric covariance reduces to null $\mathbb{CV}_{\mathrm{A}}[X,X; \psi] = 0$.
}
of $X$ and $Y$ on the state $|\psi\rangle \in \mathcal{D}$, respectively, where $\{X,Y\} := XY + YX$ denotes the anti-commutator (not to be confused with the braces denoting sets).
\begin{proof}
Throughout the proof, we choose $b \in \sigma(B)$ such that $\|\Pi_{b} \phi \|^{2} \neq 0$. Then, it is fairly straightforward to see that the map
\begin{equation}\label{eq:cond_exp_val}
\mathbb{E}\left[X | B=b ; \Psi^{g} \right] = \frac{\mathbb{E}\left[ \Pi_{b} \otimes X; \Psi^{g} \right]}{\mathbb{E}\left[ \Pi_{b} \otimes I; \Psi^{g} \right]}, \quad g \in U_{0},
\end{equation}
is well-defined on some neighbourhood $U_{0}$ around the origin $g=0$. It then follows directly from the expression \eqref{eq:cond_exp_val} that the differentiability of both the numerator and the denominator of the r.~h.~s. gives a sufficient condition for the conditional expectation $\mathbb{E}[X|B = b; \Psi^{g}]$ to be differentiable. In order to simplify our notations, we assume in the following that all the vectors $|\phi\rangle$ and $|\psi\rangle$, respectively representing the initial quantum states of the target and the meter system, are normalised. Since the proof is rather lengthy, we divide it into several parts.
\paragraph{Leibniz Rule}
To prepare for our arguments, we first recall some basic facts. Let $F, G: U \to \mathcal{H}$ be a map from an open subset $U \subset \mathbb{R}$ of the real line to a Hilbert space $\mathcal{H}$. If both maps $F$ and $G$ are strongly differentiable at $t_{0} \in U$, the inner product $t \mapsto \langle F(t), G(t) \rangle$ is differentiable at $t_{0} \in U$, and the derivative satisfies the Leibniz rule,
\begin{align}\label{thm:leibniz_rule}
\left. \frac{d}{dt} \langle F(t), G(t) \rangle \right|_{t=t_{0}}
&:= \lim_{u \to 0} \frac{\langle F(u + t_{0}), G(u + t_{0}) \rangle - \langle F(t_{0}), G(t_{0}) \rangle}{u} \nonumber \\
&= \lim_{u \to 0} \frac{\langle F(u + t_{0}) - F(t_{0}), G(u + t_{0}) \rangle + \langle F(t_{0}), G(u + t_{0}) - G(t_{0}) \rangle}{u} \nonumber \\
&= \left\langle \frac{dF(t_{0})}{dt} , G(t_{0}) \right\rangle + \left\langle F(t_{0}), \frac{dG(t_{0})}{dt} \right\rangle.
\end{align}
\paragraph{Differentiability of the Numerator}
To prove the differentiability of the numerator of \eqref{eq:cond_exp_val} and obtain its derivative,
we first introduce two auxiliary maps $F(g) := |\Psi^{g}\rangle$ and $G_{X}(g) := (\Pi_{b} \otimes X) F(g)$, by which we rewrite the numerator
\begin{equation}\label{def:psm_numerator}
g \mapsto \mathbb{E}\left[ \Pi_{b} \otimes X; \Psi^{g} \right] = \langle F(g), G_{X}(g) \rangle
\end{equation}
in terms of their inner products. From the Leibniz rule, one sees that the desired result can be immediately obtained once the differentiability of both the maps $F(g)$ and $G_{X}(g)$ are proven and their derivatives are given.
As for the strong differentiability of the map $g \mapsto F(g)$, one readily finds by Stone's theorem on one-parameter unitary groups that the condition
\begin{equation}\label{eq:strong_diff_F}
|\phi\rangle \in \mathrm{dom}(A), \quad |\psi\rangle \in \mathcal{D} \subset \mathrm{dom}(Y)
\end{equation}
would suffice, in which case the derivative is given by
\begin{equation}\label{eq:diff_num_1}
\frac{dF(0)}{dg} = -i(A\otimes Y) |\phi\otimes\psi\rangle.
\end{equation}
As for the map $g \mapsto G_{X}(g)$, we first observe that it is written as
\begin{equation}
G_{X}(g) = (\Pi_{b} \otimes I)(I \otimes X) F(g).
\end{equation}
Due to the boundedness (continuity) of the operator $(\Pi_{b} \otimes I)$, strong differentiability of the vector-valued map
$
g \mapsto (I \otimes X) F(g)
$
would give a sufficient condition for $G_{X}(g)$ to be strongly differentiable, which one readily proves under the condition
\begin{equation}
|\phi\rangle \in \mathrm{dom}(A), \quad |\psi\rangle \in \mathcal{D} \subset \mathrm{dom}(XY) \cap \mathrm{dom}(Y)
\end{equation}
by imitating the arguments we have made starting from \eqref{eq:diff_2_start} with the help of the relation \eqref{eq:weak_weyl_analogue01}. Now that the strong differentiability of both the maps $g \mapsto F(g)$, $G_{X}(g)$ are proven, one finds from the closedness of the self-adjoint operator $(\Pi_{b} \otimes X)$ that
\begin{align}\label{eq:diff_num_2}
\frac{dG_{X}(0)}{dg}
&= (\Pi_{b} \otimes X) \frac{dF(0)}{dg} \nonumber \\
&= -i(\Pi_{b} \otimes X)(A\otimes Y)|\phi\otimes\psi\rangle.
\end{align}
Given the results \eqref{eq:diff_num_1} and \eqref{eq:diff_num_2}, the Leibniz rule leads to the desired differentiability of the numerator \eqref{def:psm_numerator}, in which one computes its derivative as
\begin{align}\label{eq:diff_numerator}
\left. \frac{d}{dg} \mathbb{E}\left[ \Pi_{b} \otimes X; \Psi^{g} \right] \right|_{g=0}
&= \left\langle \frac{dF(0)}{dg} , (\Pi_{b} \otimes X) F(0) \right\rangle + \left\langle F(0), (\Pi_{b} \otimes X) \frac{dF(0)}{dg} \right\rangle \nonumber \\
&= 2\,\mathrm{Re}\left[ \left\langle F(0), (\Pi_{b} \otimes X) \frac{dF(0)}{dg} \right\rangle \right] \nonumber \\
&= 2\,\mathrm{Re}\left[ -i\left\langle F(0) , (\Pi_{b} \otimes X)(A \otimes Y) F(0) \right\rangle \right] \nonumber \\
&= 2\,\mathrm{Im}\left[ \langle \phi , \Pi_{b}A \phi \rangle \langle \psi, XY \psi \rangle \right] \nonumber \\
&= 2\,\mathrm{Re}\left[ \langle \phi , \Pi_{b}A \phi \rangle \right] \cdot \mathbb{E}[[X,Y]/(2i);\psi] \nonumber \\
&\qquad + 2\,\mathrm{Im}\left[ \langle \phi , \Pi_{b}A \phi \rangle \right] \cdot \mathbb{E}[\{X,Y\}/2; \psi],
\end{align}
where we have used the operator equality
\begin{equation}
XY = \frac{\{X,Y\}}{2} + i\frac{[X,Y]}{2i}
\end{equation}
valid on the subspace $\mathcal{D}$.
\paragraph{Differentiability of the Denominator}
The proof for the differentiability of the denominator $\mathbb{E}\left[ \Pi_{b} \otimes I; \Psi^{g} \right]$ goes essentially the same as that for the numerator, where one readily proves its differentiability at $g=0$ under the condition $|\phi\rangle \in \mathrm{dom}(A)$, $|\psi\rangle \in \mathrm{dom}(Y)$, in which case the derivative reads
\begin{equation}\label{eq:diff_denominator}
\left. \frac{d}{dg} \mathbb{E}[ \Pi_{b} \otimes I; \Psi^{g} ] \right|_{g=0} = 2\,\mathrm{Im}\left[ \langle \phi , \Pi_{b}A \phi \rangle \right] \cdot \mathbb{E}[Y; \psi],
\end{equation}
by formally replacing $X$ with $I$ in \eqref{eq:diff_numerator}.
\paragraph{Final Result}
Combining the above two results \eqref{eq:diff_numerator} and \eqref{eq:diff_denominator}, one concludes that, given the choice $|\phi\rangle \in \mathrm{dom}(A)$ and $b \in \sigma(B)$ with $\|\Pi_{b}\phi\|^{2} \neq 0$ of the target configuration, and $|\psi\rangle \in \mathcal{D}$ for the meter system, the conditional expectation $\mathbb{E}[X|B = b; \Psi^{g}]$ is indeed differentiable at $g=0$. Its derivative can then be evaluated based on the classical result of calculus (the quotient rule for derivative) as
\begin{align}
&\left. \frac{d}{dg} \mathbb{E}[X|B = b; \Psi^{g}] \right|_{g=0} \nonumber \\
&\quad = \frac{\left. \frac{d}{dg} \mathbb{E}\left[ \Pi_{b} \otimes X; \Psi^{g} \right]\right|_{g=0} \cdot \mathbb{E}\left[ \Pi_{b} \otimes I; \Psi^{0} \right] - \mathbb{E}\left[ \Pi_{b} \otimes X; \Psi^{0} \right] \cdot \left. \frac{d}{dg} \mathbb{E}\left[\Pi_{b} \otimes I; \Psi^{g} \right]\right|_{g=0} }{\mathbb{E}\left[\Pi_{b} \otimes I; \Psi^{0} \right]^{2}} \nonumber \\
&\quad = 2\,\mathrm{Re} \left[\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}\right] \cdot \mathbb{E}[[X,Y]/(2i); \psi] \nonumber \\
& \qquad + 2\,\mathrm{Im}\left[\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}\right] \cdot \left( \mathbb{E}[\{X,Y\}/2; \psi] - \mathbb{E}[X; \psi]\mathbb{E}[Y; \psi] \right) \nonumber \\
&\quad = 2\,\mathrm{Re} \left[\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}\right] \cdot \mathbb{CV}_{\mathrm{A}}[X,Y; \psi] + 2\,\mathrm{Im}\left[\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}\right] \cdot \mathbb{CV}_{\mathrm{S}}[X,Y; \psi].
\end{align}
We have thus verified our desired statement \eqref{eq:wpsm_diff}.
\end{proof}
\subsubsection{Conditional Quasi-expectations of Quantum Observables}\label{sec:psI_cond_quasi_exp}
Now that we have computed the derivative of the map \eqref{def:wpsm_special} for the special case, we are now interested in the case in which the conditioning observable $B$ is general, and wish to specify the limit of the formal expression
\begin{equation}
\lim_{g \to 0} \frac{\mathbb{E}[X|B; \Psi^{g}] - \mathbb{E}[X|B; \Psi^{0}]}{g}
\end{equation}
and the topology in which the convergence is meant. From the result of Proposition~\ref{prop:diff_cond_exp}, one might naturally conjecture that the limit is given by
\begin{equation}
2\,\mathrm{Re} f \cdot \mathbb{CV}_{\mathrm{A}}[X,Y; \psi] + 2\,\mathrm{Im}f \cdot \mathbb{CV}_{\mathrm{S}}[X,Y; \psi]
\end{equation}
with a `function' $f$ defined formally as
\begin{equation}\label{def:cond_quasi_exp_prelim}
f(b) :=
\frac{\langle \phi, \Pi_{b}A\phi \rangle}{\|\Pi_{b}\phi\|^{2}}.
\end{equation}
In order to make this observation a precise mathematical statement, we first introduce a convenient concept.
\paragraph{Conditional Quasi-expectations}
Observing that in the case where $A$ and $B$ are simultaneously measurable, the function \eqref{def:cond_quasi_exp_prelim} is nothing but the conditional expectation of $A$ given $B$. In general, however, the target observable and the conditioning observable $B$ need not be simultaneously observable. We thus wish to define a quantum analogue of conditional expectations of an observable $A$ given another observable $B$, well-defined even for the pair that are not necessarily simultaneously measurable. To this end, we first fix a non-zero vector $|\phi\rangle \in \mathrm{dom}(A)$ and consider a complex measure
\begin{equation}\label{def:nu_temp}
\nu(\Delta) := \langle \phi, E_{B}(\Delta)A\phi \rangle/\|\phi\|^{2},
\quad \Delta \in \mathfrak{B},
\end{equation}
where $E_{B}$ is the unique spectral measure accompanying $B$.
Now, a direct application of the Cauchy-Schwarz inequality leads to
\begin{equation}
| \langle \phi, E_{B}(\Delta)A\phi \rangle | \leq \| E_{B}(\Delta)\phi\| \cdot \|A\phi\|,
\end{equation}
by which one finds the absolute continuity $\nu \ll \mu_{B}^{\phi}$, where $\mu_{B}^{\phi}(\Delta) := \| E_{B}(\Delta) \phi\|^{2}/\|\phi\|^{2}$ as usual. This allows us to define the Radon-Nikod{\'y}m derivative
\begin{equation}\label{def:cond_quas-exp}
\mathbb{E}[A|B;\phi] := d\nu/d\mu_{B}^{\phi}.
\end{equation}
By definition, it is the unique $\mu_{B}^{\phi}$-integrable (equivalence class of) function(s) that satisfies
\begin{equation}
\langle \phi, E_{B}(\Delta)A\phi \rangle / \|\phi\|^{2} = \int_{\Delta} \mathbb{E}[A|B = b;\phi]\ d\mu_{B}^{\phi}(b), \quad \Delta \in \mathfrak{B},
\end{equation}
and as such,
\begin{equation}
\mathbb{E}[A;\phi] = \int_{\mathbb{R}} \mathbb{E}[A|B = b;\phi]\ d\mu_{B}^{\phi}(b)
\end{equation}
holds in particular. Incidentally, when the state $|\phi_{a}\rangle \in \mathrm{dom}(A)$ happens to be an eigenvector of $A$ with the eigenvalue $a$, the map
\begin{equation}
\mathbb{E}[A|B;\phi_{a}] = a
\end{equation}
becomes a constant function independent of the choice of the conditioning observable $B$. The map $\mathbb{E}[A|B ;\phi]$ thus shares properties similar to the conditional expectations, and in the special case in which $A$ and $B$ happens to be simultaneously measurable, it actually reduces to the standard conditional expectation. However, as one finds shortly below, it can be shown by reductio ad absurdum that the map $\mathbb{E}[A|B ;\phi]$ may not admit itself to be understood as a standard conditional expectation in the case where the pair of observables concerned does not admit coexistence. These preliminary observations may tempt one to call the map \eqref{def:cond_quas-exp} a \emph{conditional quasi-expectation} of $A$ given $B$.
\paragraph{Arbitrariness to Conditional Quasi-expectations}
As one may immediately notice, there exists an arbitrariness to the way one may define conditional quasi-expectations. For example, one may just define the complex conjugate of the complex measure \eqref{def:nu_temp} as
\begin{equation}
\nu^{*}(\Delta) = \langle \phi, AE_{B}(\Delta)\phi \rangle/\|\phi\|^{2}
\end{equation}
and introduce the Radon-Nikod{\'y}m derivative as
\begin{equation}
\mathbb{E}^{*}[A|B;\phi] := d\nu^{*}/d\mu_{B}^{\phi} = \mathbb{E}[A|B;\phi]^{*}.
\end{equation}
One may conduct analogous reasoning to verify that the function $\mathbb{E}^{*}[A|B;\phi]$ also satisfies properties similar to the usual conditional expectations, and that both definitions coincide when the pair of $A$ and $B$ happens to be simultaneously measurable. One may even consider a complex linear combination of $\mathbb{E}[A|B;\phi]$ and its complex conjugate to define
\begin{align}\label{def:cond_quasi-exp_alpha}
\mathbb{E}^{\alpha}[A|B;\phi]
&:= \frac{1+\alpha}{2} \cdot \mathbb{E}[A|B;\phi] + \frac{1-\alpha}{2} \cdot \mathbb{E}^{*}[A|B;\phi] \nonumber \\
&= \mathrm{Re}\left[ \mathbb{E}[A|B;\phi] \right] + \alpha i\, \mathrm{Im}\left[ \mathbb{E}[A|B;\phi] \right] , \quad \alpha \in \mathbb{C},
\end{align}
for example, so that $\mathbb{E}^{1}[A|B;\phi] = \mathbb{E}[A|B;\phi]$ and $\mathbb{E}^{-1}[A|B;\phi] = \mathbb{E}^{*}[A|B;\phi]$. In fact, it reveals that there exists a multitude of potential candidates for possible definitions of such `conditional quasi-expectations', all sharing desirable properties mentioned earlier. We shall be returning to this problem in a more general framework of quasi-joint-probabilities of quantum observables in Section~\ref{sec:qp_qo}, but for our purpose and the scope of this paper, it suffices to concentrate only on the family \eqref{def:cond_quasi-exp_alpha} for definiteness, and we thus introduce:
\begin{definition*}[Conditional Quasi-expectation of $A$ given $B$]
Let $A$ and $B$ be self-adjoint operators on a Hilbert space $\mathcal{H}$, and let $E_{B}$ be the spectral measure of $B$. For a given state $|\phi\rangle \in \mathrm{dom}(A)$, we call the family of complex linear combinations of the Radon-Nikod{\'y}m derivatives \eqref{def:cond_quasi-exp_alpha}
the complex-parametrised family of conditional quasi-expectations of $A$ given $B$. They are, by definition, a (family of) complex function(s) defined on the spectrum $\sigma(B)$.
\end{definition*}
\noindent
Note, by definition, that each member $\mathbb{E}^{\alpha}[A|B;\phi]$, $\alpha \in \mathbb{C}$, of the family of conditional quasi-expectations is integrable with respect to the probability measure $\mu_{B}^{\phi}$, and its total integration coincides with the expectation value $\mathbb{E}[A;\phi]$ of $A$.
If the conditioning observable $B$ happens to possess spectrum with finite cardinality, so that its spectral decomposition reads \eqref{eq:spectr_decomp_B_fin}, the conditional quasi-expectation admits an expression by operators and vectors as
\begin{equation}\label{eq:quasi-cond_exp_fin}
\mathbb{E}[A|B = b;\phi] =
\begin{cases}
\langle \phi, \Pi_{b}A\phi \rangle / \|\Pi_{b}\phi\|^{2}, \quad &(b \in \sigma(B),\ \|\Pi_{b}\phi\|^{2} \neq 0), \\
\text{indefinite}, \quad &(\text{else})
\end{cases}
\end{equation}
and
\begin{equation}
\mathbb{E}^{\alpha}[A|B;\phi] = \frac{1+\alpha}{2} \cdot \mathbb{E}[A|B;\phi] + \frac{1-\alpha}{2} \cdot \mathbb{E}[A|B;\phi]^{*}
\end{equation}
if explicitly written out.
\paragraph{Conditional Quasi-expectations, Two-state Values and the Weak Value}
Incidentally, if the conditioning observable happens to be a projection $B=|\phi^{\prime}\rangle\langle\phi^{\prime}|$ on a one-dimensional subspace of $\mathcal{H}$ spanned by a unit vector $|\phi^{\prime}\rangle$ ({\it i.e.}, a post-selection), the conditional quasi-expectation of $A$ given the outcome $B=1$ reads
\begin{align}\label{eq:qce_and_tsv}
\mathbb{E}^{\alpha}[A|B = 1;\phi]
&= \frac{1+\alpha}{2} \cdot \frac{\langle \phi^{\prime}, A\phi \rangle}{\langle \phi^{\prime}, \phi \rangle} + \frac{1-\alpha}{2} \cdot \frac{\langle \phi, A\phi^{\prime} \rangle}{\langle \phi, \phi^{\prime} \rangle},
\end{align}
given that the probability of finding the outcome $1$ of $B$ is non-vanishing $\mu_{B}^{\phi}(\{1\}) = |\langle \phi^{\prime}, \phi \rangle|^{2} \neq 0$. Specifically for the choice $\alpha = 1$, this reduces to
\begin{equation}\label{def:weak_value}
\mathbb{E}[A|B = 1;\phi] = \frac{\langle \phi^{\prime}, A\phi \rangle}{\langle \phi^{\prime}, \phi \rangle} =: A_{w},
\end{equation}
The value $A_{w}$ is widely referred to as Aharonov's \emph{weak value} \cite{Aharonov_1964,Aharonov_1988} of $A$ for the pair of the pre-selected state $|\phi \rangle \in \mathrm{dom}(A)$ and the post-selected state $|\phi^{\prime} \rangle \in \mathcal{H}$. Historically, the weak value is said to have been originally introduced as a hypothetical value of an observable $A$ assigned to a quantum \emph{process} from the pre-selected to the post-selected state, generalising the common practice of solely assigning values to a single static \emph{state} in the standard framework of quantum mechanics. Following this philosophy, the value \eqref{eq:qce_and_tsv} termed the \emph{two-state value} \cite{Morita_2013} of $A$ under the respective selections of states was recently introduced in an attempt to generalise the idea of the weak value and to find out the possible form of a quantity of an observable specified by two quantum states. An application of the generalised Gleason's theorem revealed that, under certain desirable conditions, the most general form of the values of an observable $A$ that can be assigned to the two specification of the quantum states $|\phi\rangle \in \mathrm{dom}(A)$, $|\phi^{\prime}\rangle \in \mathcal{H}$ satisfying $\langle \phi^{\prime}, \phi\rangle \neq 0$ is given by \eqref{eq:qce_and_tsv} with a parameter $\alpha \in \mathbb{C}$ representing the ambiguity inherent to it.
\paragraph{Essential Supremum of Conditional Quasi-expectations}
While conditional quasi-expectations and the standard conditional expectations share various properties in common, the non-commutative nature of quantum observables results in some interesting distinctions between the two concepts. In this paper, as an example, we shall focus on the remarkable difference in the behaviour of their essential suprema. Now, as one recalls from Corollary~\ref{cor:cond_exp_obs_ess_sup}, for a pair of \emph{simultaneously measurable} observables $A$ and $B$ and a fixed state $|\phi\rangle \in \mathrm{dom}(A)$, the essential supremum of the conditional expectation $\left\|\, \mathbb{E}[A|B;\phi] \,\right\|_{\infty}$ is never greater than the essential supremum $\|A\|_{\infty}^{\phi}$ of the measurable function $a \mapsto a$ under the probability measure $\mu_{A}^{\phi}$. If $A$ happens to be bounded, the operator norm $\|A\|$ gives the state independent universal upper bound to $\|A\|_{\infty}^{\phi}$, which in turn also naturally becomes an upper bound to the conditional expectation $\mathbb{E}[A|B;\phi]$.
However, in general, this property is no longer preserved when $A$ and $B$ fail to be simultaneously measurable. There are several possible ways to express this discrepancy, but for brevity, we formulate it in the following manner.
To this end, we first prepare a terminology. In this paper, we say that an observable $A$ on $\mathcal{H}$ is \emph{non-trivial} if $A$ is not a scalar multiple of the identity operator $tI$, $t \in \mathbb{R}$, or equivalently, if $A$ has a spectrum $\sigma(A)$ of cardinality not less than $2$. Note that the non-triviality of $A$ automatically implies $\mathrm{dim}(\mathcal{H}) \geq 2$, where $\mathrm{dim}(\mathcal{H})$ denotes the dimension of the Hilbert space $\mathcal{H}$. Since trivial operators strongly commute with any other self-adjoint operators, the function $\mathbb{E}^{\alpha}[A|B;\phi]$ always become an authentic conditional expectation, revealing itself to be a constant function always taking its unique eigenvalue $\mathbb{E}^{\alpha}[A|B;\phi] = t$, whose case is not interesting for our purpose. Hence, we shall from now on confine ourselves to the case where $A$ is non-trivial.
\begin{proposition}[Essential Supremum of Conditional Quasi-expectations]\label{prop:lim_amp_quasi_cond}
Let $A$ be a non-trivial observable, $|\phi\rangle \in \mathrm{dom}(A)$ a vector that is not an eigenvector of $A$, and let $\alpha \in \mathbb{C}$ be any choice of the ambiguity parameter of the conditional quasi-expectation.
Then, for any non-negative number $0 \leq M < \infty$, there exists a self-adjoint operator $B$ (not-necessarily simultaneously measurable with $A$) such that the essential supremum of the conditional quasi-expectation of $A$ given $B$ is not less than
\begin{equation}
M \leq \left\|\, \mathbb{E}^{\alpha}[A|B;\phi] \,\right\|_{\infty}.
\end{equation}
Specifically, one may always choose such conditioning observable $B = |\phi^{\prime}\rangle\langle\phi^{\prime}|$ to be a projection onto a one-dimensional subspace of $\mathcal{H}$ spanned by some unit vector $|\phi^{\prime}\rangle \in \mathcal{H}$.
\end{proposition}
\begin{proof}
It suffices to prove that, one may always adjust the choice of the conditioning observable $B = |\phi^{\prime}\rangle\langle\phi^{\prime}|$ so that the conditional quasi-expectation
\begin{equation}
\mathbb{E}^{\alpha}[A|B = 1;\phi] =
\begin{cases}
c, \quad (c \in \mathbb{C}), &(\alpha \neq 0) \\
r, \quad (r \in \mathbb{R}), &(\alpha = 0)
\end{cases}
\end{equation}
may take any complex number for the choice $\alpha \neq 0$, and any real number for the choice $\alpha = 0$, while maintaining the probability of observing it to be non-vanishing $\mu_{B}^{\phi}(\{1\}) > 0$.
The proof is a direct corollary of Proposition~\ref{prop:tsv} that follows immediately.
\end{proof}
\noindent
In particular, this result is to say that one may always choose a conditioning observable $B$ such that the essential supremum $\left\|\, \mathbb{E}[A|B;\phi] \,\right\|_{\infty}$ of the conditional quasi-expectation exceeds $\|A\|_{\infty}^{\phi}$, which is never possible for standard conditional expectations defined for a pair of simultaneously measurable observables.
This `amplification of conditional quasi-expectations' is a noteworthy property of quantum mechanics, and the oft-discussed `amplification of weak values' could be understood as its special case.
\begin{proposition}[Range of the Two-state Value]\label{prop:tsv}
Let $A$ be a non-trivial observable on $\mathcal{H}$, and let $|\phi\rangle \in \mathrm{dom}(A)$ be a pre-selected state that is not an eigenvector of $A$. Then, the two-state value of $A$ under the pre-selected state $|\phi\rangle$ may take any complex number in the case $\alpha \neq 0$, and in turn any real number in the case $\alpha = 0$, given an appropriate choice of the post-selected state $|\phi^{\prime}\rangle \in \mathcal{H}$.
\end{proposition}
\begin{proof}
For simplicity, we only provide the proof of the statement for the specific choice $\alpha = 1$ of the ambiguity parameter without loss of generality.
Now, before we go into the main part of the proof, we first observe that, for a non-trivial self-adjoint operator $A$ and a normalised vector $|\phi\rangle \in \mathrm{dom}(A)$, there exists a normalised vector $|\chi\rangle \in \mathcal{H}$ orthogonal to $|\phi\rangle$ such that
\begin{equation}\label{eq:tsv_prop_lem}
A|\phi\rangle = \mathbb{E}[A;\phi] \cdot |\phi\rangle + \left\| \left( A - \mathbb{E}[A;\phi] \right) \phi \right\| \cdot |\chi\rangle
\end{equation}
holds%
\footnote{Note that in the case $|\phi\rangle \in \mathrm{dom}(A^{2})$, the equality \eqref{eq:tsv_prop_lem} is equivalent to
\begin{equation}
A|\phi\rangle = \mathbb{E}[A;\phi] \cdot |\phi\rangle + \sqrt{\mathbb{V}[A;\phi]} \cdot |\chi\rangle,
\end{equation}
since one has $\left\| \left( A - \mathbb{E}[A;\phi] \right) \phi \right\|^{2} = \langle \phi, \left( A - \mathbb{E}[A;\phi] \right)^{2} \phi \rangle = \mathbb{V}[A;\phi]$ with the variance defined as in \eqref{def:variance}.}.
To see this, we first consider the case
\begin{equation}
A|\phi\rangle = \mathbb{E}[A;\phi] \cdot |\phi\rangle,
\end{equation}
that is, when $|\phi\rangle$ is an eigenvector of $A$. Then, by choosing any normalised state $|\chi\rangle$ satisfying $\langle \chi, \phi \rangle = 0$ (the existence of such $|\chi\rangle$ is guaranteed by the fact $\mathrm{dim}(\mathcal{H}) \geq 2$), one finds that the above equality is fulfilled. Next, suppose that $A|\phi\rangle \neq \mathbb{E}[A;\phi] \cdot |\phi\rangle$. Then, by defining
\begin{equation}
|\chi\rangle := \frac{\left( A - \mathbb{E}[A;\phi] \right) |\phi\rangle}{\left\| \left( A - \mathbb{E}[A;\phi] \right) \phi \right\|},
\end{equation}
one indeed learns that $\|\chi\| = 1$ and $\langle \chi, \phi \rangle = 0$ as stated.
Armed with this fact and by fixing such $|\chi\rangle$, we choose the post-selected state as
\begin{equation}
|\phi^{\prime}\rangle = \frac{1}{c^{*}} |\phi\rangle + |\chi\rangle
\end{equation}
with a free parameter $c \in \mathbb{C}^{\times}$.
One then finds
\begin{align}
\mathbb{E}^{1}[A|B = 1;\phi]
&= \frac{\langle \phi^{\prime}, A\phi \rangle}{\langle \phi^{\prime}, \phi \rangle} \nonumber \\
&= \mathbb{E}[A;\phi] \cdot \frac{\langle\phi^{\prime}, \phi\rangle}{\langle\phi^{\prime}, \phi\rangle} + \left\| \left( A - \mathbb{E}[A;\phi] \right) \phi \right\| \cdot \frac{\langle\phi^{\prime}, \chi\rangle}{\langle\phi^{\prime}, \phi\rangle} \nonumber \\
&= \mathbb{E}[A;\phi] + c \left\| \left( A - \mathbb{E}[A;\phi] \right) \phi \right\|.
\end{align}
This shows that, for the choice of an initial state $|\phi\rangle \in \mathrm{dom}(A)$ that is not an eigenvector of $A$ (which is always possible due to the non-triviality of $A$), the weak value (hence, also the two-state value) may indeed take any complex number by adjusting the free parameter $c$ appropriately.
\end{proof}
\noindent
The difference between (standard) conditional expectations and conditional quasi-expectations in the behaviour of their essential suprema makes it clear that, conditional quasi-expectations are not conditional expectations in the classical sense. This provides an indirect proof for the fact that, in general, \emph{the `joint behaviour' of the outcomes of the pair of (generally non-commuting) quantum observables $A$ and $B$ does not allow itself to be described by probability spaces}. This would be accounted for in depth in Section~\ref{sec:ps_II} and \ref{sec:qp_qo} shortly.
\subsubsection{Weak Conditioned Measurement}
Armed with our newly introduced concept of conditional quasi-expectations \eqref{def:cond_quasi-exp_alpha} of a quantum observable given another (not necessarily simultaneously measurable) quantum observable, we shall summarise our findings regarding the first-order local behaviour of the conditional expectation at the origin. Combining Proposition~\ref{prop:diff_cond_exp} and \eqref{eq:quasi-cond_exp_fin}, one is naturally tempted to conjecture that:
\begin{proposition}[Weak Conditioned Measurement]
Let $A$ and $B$ be self-adjoint operators defined on the target system $\mathcal{H}$, and let the respective initial states $|\phi\rangle \in \mathrm{dom}(A)$, $|\psi\rangle \in \mathcal{D}$ be fixed. Then, the conditional expectation $\mathbb{E}[X|B;\Psi^{g}]$ is well-defined for all range of $g \in \mathbb{R}$, and the limit converges to
\begin{align}\label{eq:wpsm_diff_02}
\left. \frac{d}{dg} \mathbb{E}[X|B; \Psi^{g}] \right|_{g=0}
&:= \lim_{g \to 0} \frac{\mathbb{E}[X|B; \Psi^{g}] - \mathbb{E}[X|B; \Psi^{0}]}{g} \nonumber \\
&= 2\,\mathrm{Re} \left[ \mathbb{E}[A|B; \phi] \right] \cdot \mathbb{CV}_{\mathrm{A}}[X,Y; \psi] \nonumber \\
& \qquad + 2\,\mathrm{Im}\left[ \mathbb{E}[A|B; \phi] \right] \cdot \mathbb{CV}_{\mathrm{S}}[X,Y; \psi]
\end{align}
point-wise $\mu_{B}^{\phi}$-almost everywhere.
\end{proposition}
\noindent
While we have explicitly proved the above statement only in the special case where $B$ has spectrum of finite cardinality, the same statement indeed holds for general $B$, although we do not go into the technical details for its demonstration. One may thus understand the process of the weak CM scheme as the practice of measuring (the real and imaginary parts of) the conditional quasi-expectation $\mathbb{E}[A|B; \phi]$ of the target system. This result is to be compared with the unconditioned counterpart, in which one may extract the standard (unconditional) expectation $\mathbb{E}[A; \phi]$ by means of the weak UM scheme from the first-order differential coefficient of the measurement outcomes.
\paragraph{Topic: Conditional Quasi-expectation as the Merkmal for Amplification}
Under the above conditions, Taylor's theorem states that one has the following first-order expansion of the conditional expectation
\begin{align}\label{eq:cexp_expansion}
\mathbb{E}[X|B; \Psi^{g}]
&= \mathbb{E}[X;\psi] \nonumber \\
&\quad + g \cdot \big( 2\,\mathrm{Re} \left[ \mathbb{E}[A|B; \phi] \right] \cdot \mathbb{CV}_{\mathrm{A}}[X,Y; \psi] + 2\,\mathrm{Im}\left[ \mathbb{E}[A|B; \phi] \right] \cdot \mathbb{CV}_{\mathrm{S}}[X,Y; \psi] \big) \nonumber \\
& \qquad + o(g),
\end{align}
where $o(g)$ (Landau symbol) denotes a member of the class of functions satisfying the asymptotic property
\begin{equation}
\lim_{g \to 0} \frac{o(g)}{|g|} = 0,
\end{equation}
and the equality \eqref{eq:cexp_expansion} is understood to hold $\mu_{B}^{\phi}$-almost everywhere.
The above fact purports that the conditional quasi-expectation $\mathbb{E}[A|B; \phi]$ gives the (best first-order) indicator on the degree of `amplification' of the conditional expectation $\mathbb{E}[X|B; \Psi^{g}]$ one may attain by means of choosing the conditioning observable $B$ on the target system. Colloquially speaking, if one hopes to gain large amplification effect by conditioning, the first place one should look for is its conditional quasi-expectation, and one may hopefully achieve it by adjusting the conditioning observable $B$ so that the conditional quasi-expectation $\mathbb{E}[A|B; \phi]$ becomes large enough. However, note here that while the conditional quasi-expectation (for non-trivial $A$, and in addition, for the choice of the initial state $|\phi\rangle \in \mathrm{dom}(A)$ that is not an eigenvector of $A$) admits arbitrary large amplification by a suitable choice of the conditioning observable $B$ (Proposition~\ref{prop:lim_amp_quasi_cond}), the classical conditional expectation $\mathbb{E}[X|B; \Psi^{g}]$ may have an upper bound depending on its configuration (Proposition~\ref{prop:limit_of_amplification}). This generally suggests that the discrepancies between the full-order behaviour of $\mathbb{E}[X|B; \Psi^{g}]$ and its first-order approximation becomes larger (in other words, the higher-order terms $o(g)$ becomes more significant) as one adjusts the choice of the conditioning observable $B$ so that the conditional quasi-expectation may become larger. As for the higher-order terms, although we shall omit details, we note that one may also prove higher-order differentiability of the conditional expectation by placing stricter conditions for the choice of both the initial states of the target system and the meter system, and subsequently compute higher-order derivatives through analogous procedure as demonstrated above.
In order to confirm this observation with a concrete model, we have included in Appendix~\ref{sec:PSM} an analytic example where we compute the conditional expectation $\mathbb{E}[X|B; \Psi^{g}]$ for the special case in which the conditioning observable $B = |\phi^{\prime}\rangle\langle\phi^{\prime}|$ is a projection onto a one-dimensional subspace spanned by a unit vector $|\phi^{\prime}\rangle \in \mathcal{H}$ ({\it i.e.}, the post-selected measurement scheme), and moreover the target observable $A$ is dichotomic. One shall indeed find the existence of the limit of `amplification' of the conditional expectation by the `weak value amplification', and various other general properties alongside that we found in the discussions throughout this section.
\newpage
\section{Conditioned Measurement II: In Terms of Conditional Probabilities}\label{sec:ps_II}
In Section~\ref{sec:ups_II}, we have elaborated the study of the UM scheme conducted in the preceding Section~\ref{sec:ups_I} in terms of probabilities. In this section, we follow the same line and intend to refine our analysis for the conditioned counterpart.
\paragraph{Preliminary Observations}
As one may recall, we have seen in Section~\ref{sec:ups_I} and Section~\ref{sec:ups_II} that, by means of the UM scheme, one could extract the information of the target system in both the form of the expectation value $\mathbb{E}[A;\phi]$ and the probability measure $\mu_{A}^{\phi}$, the former by looking at the expectation value of the meter observable $X$ conjugate to $Y$, whereas the latter by focusing at the probability measure of it, and they were obtained by either inspecting the strong region $g \to \pm \infty$ of the interaction or by probing its local behaviour at $g=0$, both in parallel manners.
Now, as for the conditioned case, while we have not looked into the strong region $g \to \pm \infty$ of the interaction parameter, our analysis on the local behaviour conducted in Section~\ref{sec:ps_I} revealed that the first-order derivative of the expectation value for the choice $X = Q, P$ both contain potions (real and imaginary parts) of the conditional quasi-expectation $\mathbb{E}[A|B;\phi]$.
By comparing this result to the unconditioned case, one may come to a na{\"i}ve conjecture that both the expectation value and the conditional quasi-expectation of an observable $A$ has some quality in common. Namely, since the CM scheme incorporate conditioning, one may speculate that the conditional quasi-expectation $\mathbb{E}[A|B;\phi]$ may be interpreted as some form of a `conditional average' with respect to an underlying `probability distribution' of some kind.
\paragraph{Quasi-joint-probability Distributions in Quantum Mechanics}
A quick observation on our previous result \eqref{eq:wpsm_diff} reveals that, the full description of the CM scheme must incorporate the information of the measurement outcomes of \emph{both} the choice $X=Q, P$ of the meter observables, which is in contrast to the unconditioned case where we may concentrate only on the analysis of the probability distribution describing the outcome of a \emph{single} observable $X$ that is conjugate to $Y$.
In view of this, it would thus be natural to consider some form of a `joint-distribution' describing the measurement outcome of both the observables $Q$ and $P$. However, as we have seen in Section~\ref{sec:sim_meas_obs}, and also from an indirect proof by observing the difference of conditional (quasi)-expectations in their behaviour regarding essential suprema that, by definition, only a pair of observables that are simultaneously measurable admits a description by joint-probability distributions in the classical sense and, unfortunately, the pair $\{Q, P\}$ of observables of our present interest does not fall into this category.
On account of this, there have been various attempts to construct some alternative form of `joint-distributions' for pairs of (generally non-commuting) quantum observables that possess convenient or desirable properties in describing the behaviour of both their outcomes.
The Wigner-Ville distribution (WD distribution) \cite{Wigner_1932,Ville_1948}, which purports to describe the `joint behaviour' of the otherwise incompatible pair of observables $\hat{x}$ and $\hat{p}$ on the normalised wave-function $\psi \in L^{2}(\mathbb{R})$, symbolically defined by
\begin{equation}\label{def:Winger_Distr}
W^{\psi}(x,p) := \frac{1}{\pi} \int_{\mathbb{R}} \psi^{*}(x + y) \psi(x - y)e^{2ipy}\ d\beta(y),
\end{equation}
and the Kirkwood-Dirac distribution (KD distribution) \cite{Kirkwood_1933, Dirac_1945}, which on the other hand allows itself to be defined for arbitrary pair of observables $A$ and $B$, symbolically defined by
\begin{equation}\label{def:Kirkwood-Dirac_Distr}
K_{A,B}^{\phi}(a,b) := \frac{\langle \phi, b\rangle\langle b, a \rangle\langle a, \phi \rangle}{\|\phi\|^{2}}
\end{equation}
with the symbolical decomposition $A = \int_{\mathbb{R}} a\ d|a\rangle\langle a|$, $B = \int_{\mathbb{R}} b\ d|b\rangle\langle b|$, are among the most well-known classical proposals. The former allows negative numbers to be assigned, whereas the latter even admits complex numbers. Despite their queerness, they both retain some properties that one finds common in the standard ({\it i.e.,} real and non-negative) joint-probability distributions, {\it e.g.}, that they both have total integration of unity, and that the marginals coincide with the probability distribution describing the behaviour of the remaining observable, and in this sense, they are occasionally referred to as \emph{quasi-joint-probability} (QJP) distributions of the specific pairs of observables.
\paragraph{Quasi-joint-probability Distributions and Conditional Quasi-expectations}
Now, as some may expect, conditional quasi-expectations are closely related to the notion of quasi-joint-probabilities in quantum mechanics. Indeed, a quick observation reveals that, given a symbolical spectral decomposition $B = \int_{\mathbb{R}} b\ d|b\rangle\langle b|$ of the conditioning observable, the complex-parametrised conditional quasi-expectation \eqref{def:cond_quasi-exp_alpha} for the choice $\alpha =1$ coincides with the, so to speak, `conditional average' of $A$ given the outcome $b$ of $B$ under the Kirkwood-Dirac distribution, as one finds under the formal computation
\begin{equation}
\mathbb{E}^{1}[A | B=b;\phi] = \frac{\langle \phi, b\rangle\langle b, A\phi \rangle}{|\langle b, \phi \rangle|^{2}} = \frac{\int_{\mathbb{R}} a\ K_{A,B}^{\phi}(a,b)d\beta(a)}{\int_{\mathbb{R}} K_{A,B}^{\phi}(a,b)d\beta(a)}.
\end{equation}
As for the Wigner-Ville distribution, pure realness of its values might lead one to think that this is in some form related to the parametrised conditional quasi-expectation for the choice $\alpha = 0$. Indeed, one confirms under the formal computation
\begin{align}
\mathbb{E}^{0}[\hat{p} | \hat{x}=x;\phi]
&:= \frac{1}{2} \left[ \frac{\langle x, \hat{p} \psi \rangle}{\langle x, \psi \rangle} + \left(\frac{\langle x, \hat{p} \psi \rangle}{\langle x, \psi \rangle}\right)^{*} \right] \nonumber \\
&= |\langle x, \psi \rangle|^{-2} \cdot \frac{\left[ \langle \psi, x\rangle\langle x, \hat{p} \psi \rangle + (\langle \psi, x\rangle\langle x, \hat{p} \psi \rangle)^{*}\right]}{2} \nonumber \\
&= |\langle x, \psi \rangle|^{-2} \cdot \left. (-i)^{-1}\frac{d}{dy} \left( \frac{\langle \psi, e^{-iy\hat{p}} x\rangle\langle x, e^{-iy\hat{p}} \psi \rangle}{2} \right)\right|_{y=0} \nonumber \\
&= |\langle x, \psi \rangle|^{-2} \cdot \left. (-i)^{-1} \frac{d}{dy}\left( \frac{\psi^{*}(x+y)\psi(x-y)}{2} \right) \right|_{y=0} \nonumber \\
&= |\langle x, \psi \rangle|^{-2} \cdot \frac{1}{2\pi} \int_{\mathbb{R}} p \psi^{*}(x + y) \psi(x - y)e^{2ipy}\ d\beta(y) \nonumber \\
&= \frac{\int_{\mathbb{R}} p\ W^{\psi}(x,p)d\beta(p)}{\int_{\mathbb{R}} W^{\psi}(x,p) d\beta(p)},
\end{align}
that the conditional quasi-expectation of $\hat{p}$ given $\hat{x}=x$ for the choice $\alpha = 0$ coincides with the, again so to speak, `conditional average' of the momentum $
\hat{p}$ given the outcome $x$ of the position $\hat{x}$ under the Wigner-Ville distribution.
\paragraph{Conditioned Measurement}
The above observation is instructive in guiding the direction of our analysis. Indeed, it would be natural to expect that the measurement of the meter system in view of QJP distributions of the pair of observables $\{Q,P\}$ would allow us to extract the information of the target system in the form that is `akin' to it, {\it i.e.}, one might hope to obtain a QJP distributions of the target system, of which `conditional average' coincides with the conditional quasi-expectations $\mathbb{E}^{\alpha}[A|B = b]$ of our interest.
Guided by this formal argument and heuristic observation, in this section, we shall be analysing the CM scheme in terms of quasi-probabilities, or more specifically, in terms of `conditional' quasi-probabilities.
Now, as our previous arguments (in particular, those developed in Section~\ref{sec:ups_II_wups}) indicate, analysis directly on the level of probabilities is better suited to be performed in the space of generalised functions, rather than density functions or measures, if one is to conduct it with decent mathematical rigour and generality. This becomes especially crucial when introducing `quasi-joint-probabilities' of a pair of (generally not necessarily simultaneously measurable) quantum observables, which is one of the main themes of this paper, and thus examined in depth in the next Section~\ref{sec:qp_qo}. However, since the present authors have judged the theory of generalised functions to be beyond the scope of this paper as a tool for analysis, we shall be working exclusively in the space of complex measures and density functions as usual. While this treatment comes with some unavoidable compromise on generality of the results and loss of transparency of the line of arguments, we hope that we may still convey the essence of the contents.
\paragraph{Conditioned Measurement in View of the WV Distributions}
In this section, the target of our interest for our measurement is the QJP distribution of the pair of observables $Q$ and $P$ on the meter, and we shall study how one may extract information of the configuration of the target system from this viewpoint. Now, as one may realise from the two concrete classical proposals given above (namely, WV distribution and KD distribution), there exist an indefiniteness/arbitrariness to the choice of such distributions, and by its very nature, one may equally conduct the analysis in view of any of one's own selection.
In this section, we shall be analysing the CM scheme exclusively in terms of the Wigner-Ville distribution.
The primary reason for our choice is merely based on its degree of familiarity in the physics community, and as mentioned above, the choice is essentially arbitrary. One may naturally conduct the same type of analysis in view of another type of quasi-probability distribution ({\it e.g.,} the Kirkwood-Dirac type) in a similar manner and obtain analogous results, or may treat them collectively from a more general viewpoint (more to this in Section~\ref{sec:qp_qo}).
\subsection{Reference Materials}
As usual, we first make a brief review on the basic concepts and facts that are used in our later discussion.
\subsubsection{Conditional Probabilities}
We first introduce some basic definitions and results on the topic of conditioning of probability measures and some intricacies inherent to it.
Let $(X, \mathfrak{A}, \mu)$ be a probability space, and let $B \in \mathfrak{A}$ such that $\mu(B) \neq 0$. For $A \in \mathfrak{A}$, we define the \emph{conditional probability of $A$ given $B$} by the number
\begin{equation}
\mu(A|B) := \frac{\mu(A \cap B)}{\mu(B)}, \quad A \in \mathfrak{A}.
\end{equation}
It is immediate that the map $\mu(\,\cdot\,|B)$ is itself a probability measure satisfying the relation
\begin{equation}\label{def:cond_prob_elementary}
\mu(A \cap B) = \mu(A|B) \cdot \mu(B), \quad A \in \mathfrak{A}.
\end{equation}
\paragraph{Conditional Probability given a Sub-$\sigma$-Algebra}
We now intend to generalise the elementary definition above to suit our further needs. In parallel to the manner we have done for conditional expectations in the previous section, let $\mathfrak{B} \subset \mathfrak{A}$ be a sub-$\sigma$-algebra, and for each measurable set $A \in \mathfrak{A}$, we define the \emph{conditional probability of $A$ given $\mathfrak{B}$} by
\begin{equation}\label{def:cond_prob}
\mu(A|\mathfrak{B}) := \mathbb{E}[\chi_{A} | \mathfrak{B}], \quad A \in \mathfrak{A},
\end{equation}
where $\chi_{A}$ is the characteristic function \eqref{def:characteristic_function} of $A$. For fixed $A \in \mathfrak{A}$, note that by definition, the conditional probability \eqref{def:cond_prob} is understood as an equivalence class of a \emph{family} of $\mu|_{\mathfrak{B}}$-integrable functions by identifying those that are indistinguishable under the given probability measure $\mu|_{\mathfrak{B}}$.
In the simplest case where $\mathfrak{B} = \sigma(B) = \{\emptyset, B, B^{c}, X\}$ given some $B \in \mathfrak{A}$, the conditional probability $\mu(A|\sigma(B))$ satisfies
\begin{align}
\mu(A \cap B)
&= \int_{B} \chi_{A}\ d\mu \nonumber \\
&= \int_{B} \mathbb{E}[\chi_{A} | \sigma(B)]\ d\mu|_{\sigma(B)}
= \mu(A|\sigma(B)) \cdot \mu(B), \quad A \in \mathfrak{A}.
\end{align}
This clarifies the relation between the general definition \eqref{def:cond_prob} and the elementary definition \eqref{def:cond_prob_elementary}.
\paragraph{Conditional Probability given a Function}
Now, under the condition above, instead of being given a sub-$\sigma$-algebra, suppose that one is given a measurable function $g: X \to Y$ for conditioning. We thus define
\begin{equation}
\mu(A|g) := \mu(A|\mathcal{I}(g)), \quad A \in \mathfrak{A},
\end{equation}
to be the conditional probability of $A$ given $g$, where $\mathcal{I}(g)$ is the initial $\sigma$-algebra of $g$ (see \eqref{def:initial_sigma_algebra} for its definition), and also introduce
\begin{equation}
\mu(A| g = y) := \mu(A|g)(y), \quad y \in \mathbb{R},
\end{equation}
of which notation involves subtlety regarding the choice of the representative, in parallel to the situation of conditional expectations we have seen earlier.
\paragraph{Conditional Probabilities as Equivalent Classes of Functions}
Given a probability space $(X, \mathfrak{A}, \mu)$ and a sub-$\sigma$-algebra $\mathfrak{B} \subset \mathfrak{A}$, the conditional probability $\mu(\,\cdot\, |\mathfrak{B})$ satisfies properties analogous to those of probability measures, namely
\begin{enumerate}
\item $\mu(\emptyset |\mathfrak{B}) = 0$, $\mu(X |\mathfrak{B}) = 1$,
\item $\mu( A |\mathfrak{B}) \geq 0, \quad A \in \mathfrak{A}$,
\item for any sequence $(A_{n})_{n \geq 1}$ of pairwise disjoint subsets of $X$, the equality
\begin{equation}
\mu\left(\left. \bigcup_{n=1}^{\infty} A_{n} \right| \mathfrak{B} \right) = \sum_{n=1}^{\infty} \mu(A_{n} | \mathfrak{B})
\end{equation}
holds.
\end{enumerate}
However, the key distinction to be noted between the usual probability measures is that, the above (in)equalities are guaranteed to hold \emph{almost everywhere}, since by definition, conditional probabilities are equivalent classes of functions. It is thus of natural interest whether we could raise the limitation by dropping `validity almost everywhere', which one may occasionally find troublesome.
\paragraph{Transition Kernels}
To this end, we first recall the definition of transition kernels. Let $(X, \mathfrak{A})$ and $(Y, \mathfrak{B})$ be measurable spaces. We say that a map $K : X \times \mathfrak{B} \to [0,\infty]$ that satisfies the conditions
\begin{enumerate}
\item the map $x \mapsto K(x, B)$ is $\mathfrak{A}$-measurable for every $B \in \mathfrak{B}$,
\item the map $B \mapsto K(x, B)$ is a measure on $(Y, \mathfrak{B})$ for every $x \in X$,
\end{enumerate}
a \emph{transition kernel} from $(X, \mathfrak{A})$ into $(Y, \mathfrak{B})$. A transition kernel is said to be \mbox{($\sigma$-)finite} if the map $B \mapsto K(x, B)$ is \mbox{($\sigma$-)finite} for all $x \in X$. If $K$ is normalised to unity $K(x, Y) = 1$ for all $x \in X$, we say that $K$ is a \emph{transition probability kernel}.
Given a $\sigma$-finite transition kernel $K : X \times \mathfrak{B} \to [0, \infty]$ from $(X, \mathfrak{A})$ into $(Y, \mathfrak{B})$ and a function $f \in \mathcal{M}^{+}(\mathfrak{B})$, the integral
\begin{equation}
(Kf)(x) := \int_{Y} f(y) K(x,dy), \quad x \in X
\end{equation}
defines a function $Kf \in \mathcal{M}^{+}(\mathfrak{A})$. On the other hand, given a measure $\mu$ on $(X, \mathfrak{A})$, the integral
\begin{equation}
(\mu K)(B) := \int_{X} K(x,B)\ d\mu(x), \quad B \in \mathfrak{B}
\end{equation}
defines a measure $\mu K$ on $(Y,\mathfrak{B})$. Associative law is valid, which is to say that
\begin{align}
\mu(Kf)
&:= \int_{X} \left( \int_{Y} f(y) K(x,dy) \right) d\mu(x) \nonumber \\
&= \int_{Y} f(y) \left( \int_{Y} K(x,dy)\ d\mu(x) \right)
=: (\mu K)f
\end{align}
holds. The following theorem is of much use.
\begin{theorem*}[Transition Kernels into Measures on Product Spaces]
Let $K : X \times \mathfrak{B} \to [0,\infty]$ be a $\sigma$-finite transition kernel from $(X, \mathfrak{A})$ into $(Y, \mathfrak{B})$, and let $\mu$ be a measure on $(X, \mathfrak{A})$. Then, there exists a measure $\pi$ on the product space $(X \times Y,\ \mathfrak{A} \otimes \mathfrak{B})$ that satisfies
\begin{align}
\int_{X \times Y} f(x,y)\ d\pi(x,y) := \int_{X} \int_{Y} f(x,y) K(x,dy)\ d\mu(x)
\end{align}
for all $f \in \mathcal{M}^{+}(\mathfrak{A} \otimes \mathfrak{B})$. If, moreover, both $\mu$ and $K$ happens to be finite, then $\pi$ is the unique finite measure on the product space satisfying
\begin{equation}
\pi(A \times B) = \int_{A} K(x,B)\ d\mu(x), \quad A \in \mathfrak{A},\ B \in \mathfrak{B}.
\end{equation}
\end{theorem*}
\noindent
This provides us a convenient way to construct a measure on the product spaces given a transition kernel and a measure.
\paragraph{Conditional Probability Distributions}
We now return to our main line of arguments, and first introduce the definition of conditional probability measures.
\begin{definition*}[Conditional Probability Measure]
Let $(X, \mathfrak{A}, \mu)$ be a probability space, and let $\mathfrak{B} \subset \mathfrak{A}$ be a sub-$\sigma$-algebra. We call a transition probability kernel $K : X \times \mathfrak{B} \to [0, 1]$ a conditional probability measure (or a regular version) of the conditional probability $\mu(\,\cdot\, |\mathfrak{B})$ given $\mathfrak{B}$, if the map $x \mapsto K(x, A)$ happens to be a representative of $\mu( A |\mathfrak{B})$ for all $A \in \mathfrak{A}$, namely
\begin{equation}
K(\,\cdot\,, A) \in \big[\, \mu(A |\mathfrak{B}) \,\big], \quad A \in \mathfrak{A}
\end{equation}
holds, where the brackets around an element denote its equivalence class. If such a transition probability kernel exists, we customarily denote it with the same notation $\mu(\,\cdot\, |\mathfrak{B})$, and its images are in turn denoted as
\begin{equation}
K(x,A) = \mu(A |\mathfrak{B})(x) = \mu_{x}(A), \quad x \in X, A \in \mathfrak{A}
\end{equation}
interchangeably, depending on the aesthetics of the formula in which it should appear.
\end{definition*}
\noindent
The presence of conditional probability measures allows us to readily make a connection between conditional expectations (defined previously in \eqref{def:cond_exp}) and averages with respect to conditional probabilities under consideration.
\begin{proposition*}[Conditional Expectations as Averages over Conditional Probability Measures]
Let $(X, \mathfrak{A}, \mu)$ be a probability space, $\mathfrak{B} \subset \mathfrak{A}$ be a sub-$\sigma$-algebra, and suppose that the conditional probability $\mu(\,\cdot\, |\mathfrak{B})$ has a conditional probability measure. Then, for every $\mu$-integrable function $f$, the map
\begin{equation}
x \mapsto \int_{X} f(x^{\prime})\ d\mu_{x}(x^{\prime}) \in \big[\, \mathbb{E}[f|\mathfrak{B}] \,\big]
\end{equation}
is a representative of the conditional expectation of $f$ given $\mathfrak{B}$.
\end{proposition*}
\noindent
We note that conditional probability measures do not necessarily exist for general measure spaces. However, fortunately for us, the case $(X, \mathfrak{A}) = (\mathbb{R}^{n}, \mathfrak{B}^{n})$ that we are interested in is known to always admit it.
\paragraph{Conditional Probability Distributions}
Given a probability space $(X, \mathfrak{A}, \mu)$ and a sub-$\sigma$-algebra $\mathfrak{B} \subset \mathfrak{A}$, suppose that a measurable map $f: (X,\mathfrak{A}) \to (X^{\prime},\mathfrak{A}^{\prime})$ is moreover given. In parallel to what we have seen for conditional expectations, this allows us to define an equivalence class of functions
\begin{equation}
\mu(f \in A^{\prime} | \mathfrak{B}) := \mu(f^{-1}(A^{\prime}) | \mathfrak{B})
\end{equation}
for all $A^{\prime} \in \mathfrak{A}^{\prime}$. Then, a transition probability kernel $K : X \times \mathfrak{A}^{\prime} \to [0,1]$ from $(X, \mathfrak{B})$ into $(X^{\prime},\mathfrak{A}^{\prime})$ satisfying
\begin{equation}
K(\,\cdot\, , A^{\prime}) \in \big[\, \mu(f \in A^{\prime} | \mathfrak{B}) \,\big], \quad A^{\prime} \in \mathfrak{A}^{\prime}
\end{equation}
is called a \emph{conditional probability distribution of $f$ given $\mathfrak{B}$}. Likewise, given another measurable map $g: (X,\mathfrak{A}) \to (Y^{\prime},\mathfrak{B}^{\prime})$, a transition probability kernel $K : X \times \mathfrak{A}^{\prime} \to [0,1]$ from $(X, \mathcal{I}(g))$ into $(X^{\prime},\mathfrak{A}^{\prime})$ satisfying
\begin{equation}
K(\,\cdot\, , A^{\prime}) \in \big[\, \mu(f \in A^{\prime} | \mathcal{I}(g)) \,\big], \quad A^{\prime} \in \mathfrak{A}^{\prime}
\end{equation}
is called a \emph{conditional probability distribution of $f$ given $g$}. Such transition probability kernels do not necessarily exist in general, but as above, the case $(X^{\prime},\mathfrak{A}^{\prime}) = (\mathbb{R}^{n}, \mathfrak{B}^{n})$ that we are interested in is known to always admit it.
\paragraph{Conditioning in Quantum Measurements}
Under the context of quantum measurements, let $A$ and $B$ be a pair of simultaneously measurable observables. Given a joint-probability distribution $\mu_{A,B}^{\phi}$ of $A$ and $B$ on some quantum state $|\phi\rangle \in \mathcal{H}$, we introduce
\begin{equation}
\mu_{A}^{\phi}(\Delta_{A} | B) := \mu_{A,B}^{\phi}(\pi_{A} \in \Delta_{A} | \pi_{B}), \quad \Delta_{A} \in \mathfrak{B}^{1},
\end{equation}
where $\pi_{A}(a,b) = a$ and $\pi_{B}(a,b) = b$ are measurable functions (projections) respectively representing the behaviour of the measurement outcomes of $A$ and $B$.
Accordingly, we define the \emph{conditional probability distribution of $A$ given $B$} to be a transition probability kernel $K : \mathbb{R} \times \mathfrak{B}^{1} \to [0,1]$ that satisfies
\begin{equation}
K(\,\cdot\,,\Delta_{A}) \in \big[\, \mu_{A}^{\phi}(\Delta_{A} | B) \,\big], \quad \Delta_{A} \in \mathfrak{B}^{1},
\end{equation}
which, as guaranteed above, is known to always exist. The values of the conditional probability distribution of $A$ given $B$ are in turn denoted interchangeably by
\begin{equation}
\mu_{A}^{\phi}(\Delta_{A} | B = b) = \mu_{B=b}^{\phi}(A \in \Delta_{A}) = \mu_{A}^{\phi}(\Delta_{A} | B)(b) := K(b, \Delta_{A}),
\end{equation}
depending on the context.
\subsubsection{Fourier Transformation}
We next recall the basic definitions and properties of the Fourier transformation.
For convenience, we first introduce the renormalised $n$-dimensional Lebesgue-Borel measure on $\mathbb{R}^{n}$ by
\begin{equation}\label{def:renormalised_LB_measure}
dm_{n} := (2\pi)^{-n/2} d\beta^{n}.
\end{equation}
Accordingly, in this section we employ the renormalised $L^{p}$-norm and the convolution defined by the renormalised Lebesgue-Borel measure,
\begin{gather}
\|f\|_{p} := \left( \int_{\mathbb{R}^{n}} |f(x)|^{p}\ dm_{n}(x)\right)^{1/p}, \quad f \in L^{p}(\mathbb{R}^{n}), \\
(f \ast g)(x) := \int_{\mathbb{R}^{n}} f(x-y)g(y)\ dm_{n}(y), \quad f, g \in L^{1}(\mathbb{R}^{n}).
\label{eq:convol}
\end{gather}
For brevity, we occasionally write $dm_{1} = dm$ whenever there is no risk for confusion.
Now, for a function $f \in L^{1}(\mathbb{R}^{n})$, recall that the functions $\hat{f}, \check{f} : \mathbb{R}^{n} \to \mathbb{C}$ defined by
\begin{align}
\hat{f}(q) &:= \int_{\mathbb{R}^{n}} e^{-i\langle q, x \rangle}f(x)\ dm_{n}(x), \\
\check{f}(q) &:= \int_{\mathbb{R}^{n}} e^{i\langle q, x \rangle}f(x)\ dm_{n}(x),
\end{align}
with the scalar product $\langle q, x \rangle := \sum_{k=1}^{n} q_{k}x_{k}$ of two real vectors in $\mathbb{R}^{n}$, are respectively called the \emph{Fourier transform} and the \emph{inverse Fourier transform} of $f$.
The $\mathbb{C}$-linear map $\mathscr{F}$ that maps $f$ to its Fourier transform $\hat{f}$ is called the \emph{Fourier transformation}. It is known that the Fourier transformation is injective, {\it i.e.} $\hat{f} = \hat{g}$ implies $f = g$. For $f, g \in L^{1}(\mathbb{R}^{n})$,
the following properties under the convolution \eqref{eq:convol}, scaling \eqref{def:function_scaling}, and translation \eqref{eq:translation},
\begin{align}
\widehat{(f \ast g)} &= \hat{f} \cdot \hat{g}, \\
\widehat{(f_{t})}(q) &= \hat{f}(tq), \quad t \neq 0, \label{eq:Scaling_Fourier_translation} \\
\widehat{(\tau_{a}f)}(t) &= e^{i\langle a,x\rangle}\hat{f}(t),\quad a \in \mathbb{R}, \label{eq:Fourier_translation}
\end{align}
respectively, are basic.
The Fourier transformation $\mathscr{F}$ plays particularly well on the subspace $\mathscr{S}(\mathbb{R}^{n}) \subset L^{1}(\mathbb{R}^{n})$, where it becomes a linear bijection of $\mathscr{S}(\mathbb{R}^{n})$ onto $\mathscr{S}(\mathbb{R}^{n})$, whose inverse is given by the inverse Fourier transformation (recall, on the other hand, that one does not necessarily have $\hat{f} \in L^{1}(\mathbb{R}^{n})$ for $f \in L^{1}(\mathbb{R}^{n})$ in general). One then has
\begin{align}
\partial^{\gamma}(\mathscr{F}f) &= (-i)^{|\gamma|}\mathscr{F}(x^{\gamma}f), \quad \gamma \in \mathbb{N}^{n}_{0}, \label{eq:fourier_diff_01} \\
\mathscr{F}(\partial^{\gamma}f) &= i^{|\gamma|}q^{\gamma}\mathscr{F}f, \quad \gamma \in \mathbb{N}^{n}_{0},\label{eq:fourier_diff_02}
\end{align}
for $f \in \mathscr{S}(\mathbb{R}^{n})$, where we have used the multi-index $\gamma := (\gamma_{1}, \dots, \gamma_{n}) \in \mathbb{N}^{n}_{0}$ as in \eqref{def:use_alpha} and
introduced the shorthand $|\gamma| := \gamma_{1} + \cdots + \gamma_{n}$.
\subsubsection{Wigner-Ville Distribution}
In order to make our line of arguments self-contained in the framework of density functions $L^{1}(\mathfrak{B})$, we assume $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ throughout this passage.
Given such $\psi$, we define a complex function $\omega^{\psi} \in L^{1}(\mathbb{R}^{2}) \cap L^{2}(\mathbb{R}^{2})$ by
\begin{align}\label{def:func_V}
\tilde{\omega}^{\psi}(x,y) &:= \psi^{*}(x - y/2) \psi(x + y/2),
\end{align}
and evaluate its total integration as
\begin{align}\label{eq:V_tot_int}
\int_{\mathbb{R}^{2}} \tilde{\omega}^{\psi}(x,y)\ dm_{2}(x,y)
&= \int_{\mathbb{R}^{2}} \psi^{*}(x - y/2) \psi(x + y/2)\ dm_{2}(x,y) \nonumber \\
&= \left( \int_{\mathbb{R}} \psi^{*}(x) \ dm(x) \right) \left( \int_{\mathbb{R}} \psi(y) \ dm(y) \right) \nonumber \\
&= \left| \int_{\mathbb{R}} \psi(x) \ dm(x) \right|^{2}.
\end{align}
Whenever the total integration \eqref{eq:V_tot_int} is non-vanishing, we introduce
\begin{equation}\label{eq:omega_norm}
\omega^{\psi}(x,y) := \frac{\tilde{\omega}^{\psi}(x,y)}{\int_{\mathbb{R}^{2}} \tilde{\omega}^{\psi}(x,y)\ dm_{2}(x,y)},
\end{equation}
to denote its normalisation.
On the other hand, if we consider the Fourier transform of $\tilde{\omega}^{\psi}(x,y)$ with respect to its second parameter $y$,
\begin{align}
\tilde{W}^{\psi}(x,p)
&:= \int_{\mathbb{R}} e^{-ipy} \tilde\omega^{\psi}(x,y)\ dm(y) \nonumber \\
&= \int_{\mathbb{R}} \psi^{*}(x + y/2) \psi(x - y/2)e^{ipy}\ dm(y),
\end{align}
we readily find that it is a real function,
\begin{align}
\left(\tilde{W}^{\psi}(x,p)\right)^{*}
&= \tilde{W}^{\psi}(x,p),
\end{align}
whose marginals are given by
\begin{equation}
\begin{split}
\label{eq:wigner_merginals}
\int_{\mathbb{R}} \tilde{W}^{\psi}(x,p)\ dm(x) &= |\hat{\psi}(p)|^{2}, \\
\int_{\mathbb{R}} \tilde{W}^{\psi}(x,p)\ dm(p) &= |\psi(x)|^{2}.
\end{split}
\end{equation}
Applying Plancherel's theorem, one finds that $\tilde{W}^{\psi} \in L^{1}(m_{2})$, and thus its total integration reads
\begin{align}\label{eq:wigner_tot_int}
\int_{\mathbb{R}^{2}} \tilde{W}^{\psi}(x,p)\ dm_{2}(x,p) &= \|\psi\|_{2}^{2}.
\end{align}
If the total integration \eqref{eq:wigner_tot_int} is non-vanishing, which is equivalent to the condition $\psi \neq 0$, the real quasi-probability density function denoted by
\begin{equation}\label{def:Wigner-Ville_QPD}
W^{\psi}(x,p) := \frac{\tilde{W}^{\psi}(x,p)}{\int_{\mathbb{R}^{2}} \tilde{W}^{\psi}(x,p)\ dm_{2}(x, p)}
\end{equation}
is called the \emph{Wigner-Ville distribution} on $\psi$. As we have seen in \eqref{eq:wigner_merginals},
the WV distribution
possesses useful properties for our analysis, namely, that its marginals yield the probability density function describing the behaviour of the measurement outcomes of the respective observables $\hat{x}$ and $\hat{p}$ on the state $\psi$, which is to say that
\begin{equation}
\begin{split}
\int_{\mathbb{R}^{2}} W^{\psi}(x,p)\, dm(x)
&= \rho_{\hat{p}}^{\psi}(p), \\
\int_{\mathbb{R}^{2}} W^{\psi}(x,p)\, dm(p)
&= \rho_{\hat{x}}^{\psi}(x),
\end{split}
\end{equation}
\begin{comment}
subsequently leads to
\begin{equation}
\begin{split}
\int_{\mathbb{R}^{2}} x\ W^{\psi}(x,p) dm_{2}(x,p)
&= \mathbb{E}[\hat{x};\psi], \\
\int_{\mathbb{R}^{2}} p\ W^{\psi}(x,p) dm_{2}(x,p)
&= \mathbb{E}[\hat{p};\psi],
\end{split}
\end{equation}
\end{comment}
if explicitly written down. Thus, the choice $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ defines a complex measure
\begin{equation}\label{def:qpm_for_WV}
\mu_{Q,P}^{\psi} := W^{\psi} \odot \beta^{2}
\end{equation}
on the measurable space $(\mathbb{R}, \mathfrak{B}^{2})$ that satisfies
\begin{equation}
\begin{split}\label{eq:WV_qualifies_as_QJP}
\mu_{Q,P}^{\psi}(\mathbb{R} \times \Delta)
&= \mu_{P}^{\psi}(\Delta), \\
\mu_{Q,P}^{\psi}(\Delta \times \mathbb{R})
&= \mu_{Q}^{\psi}(\Delta),
\end{split}
\quad \Delta \in \mathfrak{B}.
\end{equation}
For our later argument we note that, since the functions $\omega^{\psi}(x,y)$ and $W^{\psi}(x,y)$ are mapped to one another by Fourier transformation, they just represent the same contents seen from different viewpoints, and are thus essentially the same object.
\subsection{Conditioned Measurement}\label{sec:ps_II_ps}
We are now interested in simultaneously measuring the probability measure of $B$ on the target system and a QJP distribution of $Q$ and $P$ on the meter system. This should be possible since every local measurements can be simultaneously performed on separate systems, and this leads to an existence of a joint distribution of the probability measure of $B$ on one side, and a QJP distribution of $Q$ and $P$ on the other. Throughout this section, for definiteness, we exclusively treat the special case in which the meter state is described by the one-dimensional Schr{\"o}dinger representation of the CCR $\{ L^{2}(\mathbb{R}), \mathscr{S}(\mathbb{R}), \{\hat{x}, \hat{p}\}\}$ and choose $Y=\hat{p}$ without loss of generality.
\subsubsection{Conditioning over Quasi-probabilities}\label{sec:CM_II_cond_over_QP}
Since we are now dealing with complex measures, the definitions for conditioning must be suitably expanded accordingly. To this end, we first prepare a terminology:
\begin{definition*}[Quasi-probabilities]
Let $(X,\mathfrak{A})$ be a measurable space. We call a complex measure $\nu$ on $(X,\mathfrak{A})$ satisfying the normalisation condition $\nu(X) = 1$ a quasi-probability measure, and accordingly the triplet $(X,\mathfrak{A}, \nu)$, a quasi-probability space.
\end{definition*}
\noindent
If the underlying space is given by $(X,\mathfrak{A}) = (\mathbb{R}^{n}, \mathfrak{B}^{n})$, and the quasi-probability measure $\nu$ happens to be absolutely continuous, we call its density $d\nu/d\beta^{n} \in L^{1}(\mathbb{R}^{n})$, which is in general a complex function that has the total integration of unity, a \emph{quasi-probability density function}. According to the definition, note that the usual ({\it i.e.,} real and non-negative) probability measures and density functions are special members of the respective families of quasi-probability measures and density functions. In analogy to the standard probability spaces, given a quasi-probability space $(X,\mathfrak{A},\nu)$ and a $\nu$-integrable function $f$, we occasionally denote the total integration by
\begin{equation}
\mathbb{E}[f;\nu] := \int_{X} f\ d\nu,
\end{equation}
and call it the \emph{quasi-expectation value of $f$ under $\nu$}.
\paragraph{Quasi-joint-probabilities}
As a special subclass of quasi-probability measures, we say that a quasi-probability measure $\nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ \emph{qualifies as a QJP distribution} of the observables $A_{1}, \dots, A_{n}$ on the state $|\phi\rangle \in \mathcal{H}$, if it satisfies
\begin{equation}\label{def:qjpm}
\nu(\underbrace{\mathbb{K} \times \cdots \times \mathbb{K} \times \stackrel{k\text{-th}}{\Delta} \times \mathbb{K} \times \cdots \times \mathbb{K}}_{n}) = \mu_{A_{k}}^{\phi}(B), \quad \Delta \in \mathfrak{B}(\mathbb{K})
\end{equation}
for all $1 \leq k \leq n$. In parallel to it, we prepare the term \emph{QJP density function} for those $\nu$ that are absolutely continuous%
\footnote{Here, we occasionally admit complex parameters to describe outcomes of each observable $A_{k}$ for formal completeness. Accordingly, the r.~h.~s. of the above formula is understood as the probability measure induced by the two-dimensional spectral measure of $A_{k}$ seen as a normal operator ({\it cf.} spectral theorem for normal operators).}.
One confirms from \eqref{eq:WV_qualifies_as_QJP} that, for the choice of the quantum state $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$, the quasi-probability measure \eqref{def:qpm_for_WV} qualifies as a QJP distribution for the pair of observables $Q$ and $P$. It should be intuitively straightforward to see by the formal arguments made in the introduction that the Kirkwood-Dirac distribution also qualifies as a QJP distribution of the pair of observables under consideration.
\paragraph{Conditional Quasi-expectations}
We next intend to introduce analogous definitions regarding conditioning on quasi-probability measure spaces $(X,\mathfrak{A},\nu)$. To this end, we make some very important remarks on the different properties between standard probability measures and quasi-probability measures. Recall that we have made extensive use of the Radon-Nikod{\'y}m theory for defining conditional expectations and conditional probabilities. In applying the theory, first note that positiveness of the measure $\mu$ is necessary in order for the Radon-Nikod{\'y}m derivative $d\nu/d\mu$ of some complex measure $\nu \ll \mu$ to be well-defined. Hence, conditioning by a sub-$\sigma$-algebra $\mathfrak{B} \subset \mathfrak{A}$ must be such that the restriction $\nu|_{\mathfrak{B}}$ becomes a measure. The second fact to notice is that, for a $\nu$-integrable function $f$, the complex measure on the sub-$\sigma$-algebra defined by
\begin{equation}
B \mapsto (f\odot\nu)(B) := \int_{B} f\ d\nu, \quad B \in \mathfrak{B}
\end{equation}
is not necessarily absolutely continuous with respect to the restriction $\nu|_{\mathfrak{B}}$, in contrast to that of positive measures. With these in mind, we hereby define:
\begin{definition*}[Conditional Quasi-expectation]
Let $(X,\mathfrak{A}, \nu)$ be a quasi-probability space, and $\mathfrak{B} \subset \mathfrak{A}$ a sub-$\sigma$-algebra such that the restriction $\nu|_{\mathfrak{B}}$ becomes a probability measure ({\it i.e.}, real and non-negative). For a $\nu$-integrable function $f$ such that $f \odot \nu \ll \nu|_{\mathfrak{B}}$, we define the conditional quasi-expectation of $f$ given $\mathfrak{B}$
\begin{equation}
\mathbb{E}[f|\mathfrak{B}] := \frac{d (f \odot \nu)}{d(\nu|_{\mathfrak{B}})}
\end{equation}
by the Radon-Nikod{\'y}m derivative of the complex measure $f \odot \nu$ with respect to the measure $\nu|_{\mathfrak{B}}$.
\end{definition*}
\noindent
Given another measurable function $g: X \to \mathbb{R}$ such that the above conditions are fulfilled for the initial $\sigma$-algebra $\mathfrak{B} = \mathcal{I}(g)$, we define $\mathbb{E}[f|g]$ and any other relevant notations such as $\mathbb{E}[f|g = y]$ {\it etc.} in an analogous manner to those defined for standard probability measures.
\paragraph{Conditional Quasi-probabilities}
We then intend introduce a complex analogue of conditional probabilities defined for quasi-probability measures.
\begin{definition*}[Quasi-Conditional Probabilities]
Let $(X,\mathfrak{A}, \nu)$ be a quasi-probability space, and $\mathfrak{B} \subset \mathfrak{A}$ a sub-$\sigma$-algebra such that the restriction $\nu|_{\mathfrak{B}}$ becomes a probability measure. For a measurable set $A \in \mathfrak{A}$, we define
\begin{equation}\label{def:cond_quasi-prob}
\nu(A|\mathfrak{B}) := \mathbb{E}[\chi_{A} | \mathfrak{B}], \quad A \in \mathfrak{A},
\end{equation}
to be the quasi-conditional probability of $A$ given $\mathfrak{B}$, whenever $\chi_{A} \odot \nu \ll \nu|_{\mathfrak{B}}$, where $\chi_{A}$ is the characteristic function of $A$.
Likewise, given a measurable function $f: (X,\mathfrak{A}) \to (X^{\prime},\mathfrak{A}^{\prime})$, we introduce
\begin{equation}
\nu(f \in A^{\prime} | \mathfrak{B}) := \nu(f^{-1}(A^{\prime}) | \mathfrak{B}), \quad A^{\prime} \in \mathfrak{A}^{\prime},
\end{equation}
whenever the r.~h.~s. is well-defined.
If, instead of being given a sub-$\sigma$-algebra, one is given a measurable function $g: X \to Y$ for conditioning, we define
\begin{equation}
\nu(A|g) := \nu(A|\mathcal{I}(g)), \quad A \in \mathfrak{A},
\end{equation}
where $\mathcal{I}(g)$ is the initial $\sigma$-algebra of $g$, whenever, as usual, the r.~h.~s. is well-defined.
\end{definition*}
\noindent
The conditional quasi-probability $\nu(\,\cdot\, | \mathfrak{B})$ satisfies properties analogous to those of quasi-probability measures, namely
\begin{enumerate}
\item $\nu(\emptyset |\mathfrak{B}) = 0$, $\nu(X |\mathfrak{B}) = 1$,
\item $\nu( A |\mathfrak{B}) \in \mathbb{C}, \quad A \in \mathfrak{A}$,
\item for any sequence $(A_{n})_{n \geq 1}$ of pairwise disjoint subsets of $X$, the equality
\begin{equation}
\nu\left(\left. \bigcup_{n=1}^{\infty} A_{n} \right| \mathfrak{B} \right) = \sum_{n=1}^{\infty} \nu(A_{n} | \mathfrak{B})
\end{equation}
holds,
\end{enumerate}
whenever every component above is well-defined.
In parallel to conditional probabilities, the validity of the (in)equalities above are significant only in the sense of $\nu|_{\mathfrak{B}}$-a.e.
\paragraph{Conditional Quasi-probability Measures}
We now expand the definition of transition kernels to fit into the theory of complex measures. Let $(X, \mathfrak{A})$ and $(Y, \mathfrak{B})$ be measurable spaces. We say that a map $K : X \times \mathfrak{B} \to \mathbb{C}$ that satisfies the conditions
\begin{enumerate}
\item the map $x \mapsto K(x, B)$ is $\mathfrak{A}$-measurable for every $B \in \mathfrak{B}$,
\item the map $B \mapsto K(x, B)$ is a complex measure on $(Y, \mathfrak{B})$ for every $x \in X$,
\end{enumerate}
a \emph{complex transition kernel} from $(X, \mathfrak{A})$ into $(Y, \mathfrak{B})$. If a complex transition kernel $K$ satisfies $K(x,Y) = 1$ for all $x \in X$, we call such $K$ a \emph{transition quasi-probability kernel}. The following analogous result is of use.
\begin{proposition}[Complex Transition Kernels into Complex Measures on Product Spaces]\label{prop:ctk_to_cm}
Let $K : X \times \mathfrak{B} \to \mathbb{C}$ be a complex transition kernel from $(X, \mathfrak{A})$ into $(Y, \mathfrak{B})$, and let $\mu$ be a measure on $(X, \mathfrak{A})$. Then, there exists a complex measure $\pi$ on the product space $(X \times Y,\ \mathfrak{A} \otimes \mathfrak{B})$ that satisfies
\begin{align}
\int_{X \times Y} f(x,y)\ d\pi(x,y) := \int_{X} \int_{Y} f(x,y) K(x,dy)\ d\mu(x)
\end{align}
for all $f$, whenever the integration on the r.~h.~s. is well-defined. In particular, the complex measure $\pi$ satisfies
\begin{equation}
\pi(A \times B) = \int_{A} K(x,B)\ d\mu(x), \quad A \in \mathfrak{A},\ B \in \mathfrak{B}.
\end{equation}
\end{proposition}
\noindent
Armed with the above concepts, we thus introduce:
\begin{definition*}[Conditional Quasi-Probability Measure]
Let $(X, \mathfrak{A}, \nu)$ be a quasi-probability space, and let $\mathfrak{B} \subset \mathfrak{A}$ be a sub-$\sigma$-algebra such that the restriction $\nu|_{\mathfrak{B}}$ becomes a probability measure, and that $\nu(A|\mathfrak{B})$ is well-defined for all $A \in \mathfrak{A}$. We call a transition quasi-probability kernel $K : X \times \mathfrak{B} \to \mathbb{C}$ a conditional quasi-probability measure of the conditional quasi-probability $\nu(\,\cdot\, |\mathfrak{B})$, if the map $x \mapsto K(x, A)$ happens to be a representative of $\nu( A |\mathfrak{B})$ for all $A \in \mathfrak{A}$, namely
\begin{equation}
K(\,\cdot\,, A) \in \big[\, \nu(A |\mathfrak{B}) \,\big], \quad A \in \mathfrak{A}
\end{equation}
holds, where the brackets around an element denote its equivalence class. If such a transition quasi-probability kernel exists, we customarily denote it with the same notation $\nu(\,\cdot\, |\mathfrak{B})$, and its images are in turn interchangeably denoted by
\begin{equation}
K(x,A) = \nu(A |\mathfrak{B})(x) = \nu_{x}(A), \quad x \in X, A \in \mathfrak{A},
\end{equation}
depending on the aesthetics of the formula in which it should appear.
\end{definition*}
\noindent
As above, such transition quasi-probability kernels do not exist in general, while the case $(X, \mathfrak{A}) = (\mathbb{R}^{n}, \mathfrak{B}^{n})$ is known to always admit it.
We then have:
\begin{proposition*}[Conditional Quasi-expectations as Averages over Conditional Quasi-probability Measures]
Let $(X, \mathfrak{A}, \nu)$ be a quasi-probability space, $\mathfrak{B} \subset \mathfrak{A}$ be a sub-$\sigma$-algebra such that the restriction $\nu|_{\mathfrak{B}}$ becomes a probability measure, and suppose that the conditional quasi-probability $\nu(\,\cdot\, |\mathfrak{B})$ has a conditional quasi-probability measure. Then, for every $\nu$-integrable function $f$, the map
\begin{equation}
x \mapsto \int_{X} f(x^{\prime})\ d\nu_{x}(x^{\prime}) \in \big[\, \mathbb{E}[f|\mathfrak{B}] \,\big]
\end{equation}
is a representative of the conditional quasi-expectation of $f$ given $\mathfrak{B}$.
\end{proposition*}
\paragraph{Conditional Probability Distributions}
On a quasi-probability space $(X, \mathfrak{A}, \nu)$, suppose that a measurable map $f: (X,\mathfrak{A}) \to (X^{\prime},\mathfrak{A}^{\prime})$ is moreover given. Choosing a sub-$\sigma$-algebra $\mathfrak{B} \subset \mathfrak{A}$ such that the restriction $\nu|_{\mathfrak{B}}$ is a measure, this allows us to define an equivalence class of functions
\begin{equation}
\nu(f \in A^{\prime} | \mathfrak{B}) := \nu(f^{-1}(A^{\prime}) | \mathfrak{B})
\end{equation}
for all $A^{\prime} \in \mathfrak{A}^{\prime}$, whenever they are well-defined. Then, a transition quasi-probability kernel $K : X \times \mathfrak{A}^{\prime} \to [0,1]$ from $(X, \mathfrak{B})$ into $(X^{\prime},\mathfrak{A}^{\prime})$ satisfying
\begin{equation}
K(\,\cdot\, , A^{\prime}) \in \big[\, \mu(f \in A^{\prime} | \mathfrak{B}) \,\big], \quad A^{\prime} \in \mathfrak{A}^{\prime}
\end{equation}
is called a \emph{conditional quasi-probability distribution of $f$ given $\mathfrak{B}$}. Likewise, given another measurable map $g: (X,\mathfrak{A}) \to (Y^{\prime},\mathfrak{B}^{\prime})$ such that the restriction of $\nu$ over its initial $\sigma$-algebra $\mathcal{I}(g)$ is a measure, a transition probability kernel $K : X \times \mathfrak{A}^{\prime} \to [0,1]$ from $(X, \mathcal{I}(g))$ into $(X^{\prime},\mathfrak{A}^{\prime})$ satisfying
\begin{equation}
K(\,\cdot\, , A^{\prime}) \in \big[\, \mu(f \in A^{\prime} | \mathcal{I}(g)) \,\big], \quad A^{\prime} \in \mathfrak{A}^{\prime}
\end{equation}
is called a \emph{conditional quasi-probability distribution of $f$ given $g$}.
\subsubsection{Conditioned Measurement via the WV Distributions}
Now that we have prepared the necessary concepts and results, we may embark on our analysis. By measuring $B$ locally on the target system on one side, and a specific QJP distribution of $Q$ and $P$ locally on the meter system on the other, we obtain a quasi-probability distribution that describes the joint behaviour of the target system and the meter system. If, by haps ({\it e.g.} by choosing the right initial state $|\psi\rangle \in \mathcal{K}$) the QJP distribution of $Q$ and $P$ on the meter admits representation by a complex measure, the total quasi-probability distribution of both the target and the meter system also admits representation by a complex measure.
We thus generally define the CM scheme as an act of measuring the conditional quasi-probability distribution of the `joint outcome' of $Q$ and $P$ of the meter system given the outcome of the conditioning observable $B$ on the target system.
\paragraph{WV Distribution}
To demonstrate our point with an example, we shall from now on exclusively concentrate on the \emph{Wigner-Ville distribution} for our choice of the QJP distribution of $Q$ and $P$ for definiteness. Since the choice $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ of the initial meter state allows the WV distribution to be described by quasi-probability measures on $(\mathbb{R}^{2}, \mathfrak{B}^{2})$, we assume such special choice throughout this passage in order to remain contained in the framework of measure and integration theory (so that we may not have to deal with the theory of generalised functions).
In this subsection, the CM scheme is studied in view of the WV distribution. We first start by transcribing the CM scheme, which was initially introduced in terms of vectors and operators on Hilbert spaces, into the description by quasi-probability density functions on $\mathbb{R}^{2}$. It is then found that the transcription allows a much simpler expression in view of its Fourier transform (rather than the WV distribution itself), in which the description of the meter system after the interaction is given precisely by the convolution of the configuration of both the meter and the target system, quite analogous to the case of the UM scheme that we have previously seen. This allows us to extract the information of the target system either by means of deconvolution discussed earlier (specifically by constructing an approximate identity on the meter system), or by probing the behaviour of the distribution around the origin $g=0$. We shall then investigate the properties of the information of the target system we have just obtained, and find that this qualifies as a `conditional quasi-probability distribution of $A$ given $B$', of which the average has a connection to the conditional quasi-expectation of $A$ given $B$ introduced earlier.
\paragraph{Preliminary Observation}
As a preliminary observation, we start by assuming that the target observable $A$ has a spectrum consisting of a finite number of eigenvalues $\sigma(A) = \{a_{1}, \dots, a_{N}\}$ so that its spectral decomposition reads \eqref{eq:spect_decomp_fin}. For the ease of arguments, we further assume that the conditioning observable $B$ also has a spectrum consisting of a finite number of eigenvalues $\sigma(B) = \{b_{1}, \dots, b_{M}\}$, that every eigenvalue of $B$ is degenerate, {\it i.e.}, $\Pi_{b_{m}} = |b_{m}\rangle\langle b_{m}|$ for some normalised vectors $|b_{m}\rangle \in \mathcal{H}$ for all $1 \leq m \leq M$, and moreover that $\|\Pi_{b}\phi\|^{2} \neq 0$ for all $b \in \sigma(B)$. As for the state preparation, let $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ be a wave-function of the meter system with normalisation $\|\psi\|_{2} = 1$ so that the WV distribution can be represented by a quasi-probability density function, and we also let the initial selection $|\phi\rangle \in \mathcal{H}$ of the target system be normalised $\|\phi\| = 1$.
\paragraph{Computing the WV Distribution}
We are now interested in measuring the WV distribution of the meter system given the outcome of $B$ on the target system.
Since both the measurements are local measurements performed on the respective systems, this should be statistically equivalent to measuring the WV distribution for the meter state
\begin{equation}\label{def:mixed_state_given_b}
\psi_{B=b}^{g} := \mathrm{Tr}_{\mathcal{H}}\left[ \left|\Psi_{B=b}^{g}\right\rangle\left\langle\Psi_{B=b}^{g} \right| \right]
\end{equation}
for all $b \in \sigma(B)$, where
\begin{equation}\label{def:state_given_b}
\left|\Psi_{B=b}^{g}\right\rangle := \frac{(\Pi_{b} \otimes I) |\Psi^{g}\rangle}{\| (\Pi_{b} \otimes I) \Psi^{g} \|^{2}}
\end{equation}
is the, so-to-speak, `conditional' meter state%
\footnote{
Naturally, \eqref{def:state_given_b} and \eqref{def:mixed_state_given_b} are nothing but the state one would expect when the ideal measurement of $B$ yielded the outcome $b \in \sigma(B)$, if one adopted the standard von Neumann projection postulate.
} given the outcome $b$ of $B$.
Our analysis thus reduces to computing the WV distribution of the density operator $\psi_{B=b}^{g}$ for each of the outcomes $b \in \sigma(B)$.
In our case, in which we assume that the eigenvalues of $B$ are all degenerate, the density operator \eqref{def:mixed_state_given_b} in fact becomes a pure state, of which representation by wave-functions reads
\begin{align}\label{def:pure_state_given_b}
\psi_{B=b}^{g}(x)
&= \sum_{n = 1}^{N} \frac{\langle b, \Pi_{a_{n}}\phi\rangle}{\langle b, \phi\rangle} \left( e^{-iga_{n}\hat{p}} \psi \right)(x) \nonumber \\
&= \sum_{n=1}^{N} \frac{\langle b, \Pi_{a_{n}}\phi\rangle}{\langle b, \phi\rangle} \psi(x - ga_{n}) \nonumber \\
&= \sum_{n=1}^{N} \frac{\langle b, E_{A}(\{a_{n}\})\phi\rangle}{\langle b, \phi\rangle} \int_{\mathbb{R}} \psi(x - ga)\ d\delta_{a_{n}}(a) \nonumber \\
&= \int_{\mathbb{R}} \psi(x - ga)\ d\nu_{b}(a), \quad g \in \mathbb{R},
\end{align}
where we have used \eqref{eq:interaction_finite} to obtain the first equality. Here, we have introduced an auxiliary quasi-probability measure
\begin{equation}\label{def:aux_qpm_b_a}
\nu_{b}^{\phantom{*}}(\Delta) := \frac{\langle b, E_{A}(\Delta) \phi\rangle}{\langle b, \phi\rangle}, \quad b \in \sigma(B),\ \Delta \in \mathfrak{B},
\end{equation}
defined by means of the spectral measure $E_{A}$ of $A$, the initial state $|\phi\rangle \in \mathcal{H}$ of the target system, and the outcome $b \in \sigma(B)$ of the conditioning observable, and have used a result analogous to \eqref{eq:prob_meas_fin_spec_obs} in the last equality. One then finds that the WV distribution of the meter wave-function \eqref{def:pure_state_given_b} reads
\begin{align}\label{eq:wigner_post}
&W^{\psi^{g}_{B=b}}(x,p) \nonumber \\
&:= \int_{\mathbb{R}} \left(\psi_{B=b}^{g}(x + y/2)\right)^{*} \psi_{B=b}^{g}(x - y/2) e^{ipy}\ dm(y) \nonumber \\
&= \int_{\mathbb{R}} \left( \int_{\mathbb{R}} \psi^{*}(x - ga_{1}^{\prime} + y/2)\ d\nu_{b}^{*}(a_{1}^{\prime}) \right) \left( \int_{\mathbb{R}} \psi(x - ga_{2}^{\prime} - y/2)\ d\nu_{b}^{\phantom{*}}(a_{2}^{\prime}) \right) e^{ipy}\ dm(y) \nonumber \\
&= \int_{\mathbb{R}} \left( \int_{\mathbb{R}^{2}} \psi^{*}(x - ga_{1}^{\prime} + y/2) \psi(x - ga_{2}^{\prime} - y/2)\ d\left(\nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(a_{1}^{\prime},a_{2}^{\prime}) \right) e^{ipy}\ dm(y),
\end{align}
where we have introduced the product measure $\nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}}$ of $\nu_{b}^{\phantom{*}}$ and its complex conjugate%
\footnote{
For a pair of complex measures $\mu$ and $\nu$, by observing that $\mu \ll |\mu|$ and $\nu \ll |\nu|$, we define the \emph{product complex measure} of $\mu$ and $\nu$ by
\begin{equation}
\mu \otimes \nu := \left( \frac{d\mu}{d|\mu|} \cdot \frac{d\nu}{d|\nu|} \right) \odot \left( |\mu| \otimes |\nu| \right).
\end{equation}
By definition, product complex measures share properties similar to those of product measures \eqref{def:product_measure}, and an analogue of Fubini's theorem holds. Product complex measures reduce to the usual product measures when both of the components happen to be finite measures.
}
in the last equality.
In order to gain a better view of our findings, let us now change variables according to the linear transformation,
\begin{align}\label{eq:lin_trans_T_2i}
\left(
\begin{array}{l}
a_{1} \\
a_{2}
\end{array}
\right)
= T
\left(
\begin{array}{l}
a_{1}^{\prime} \\
a_{2}^{\prime}
\end{array}
\right),
\qquad
T := \left(
\begin{array}{cc}
1/2 & 1/2 \\
-1/2 & 1/2
\end{array}
\right).
\end{align}
Since $T \in \mathrm{GL}(2,\mathbb{R})$ belongs to the general linear group, for indeed $\det T = 1/2$, note that this transformation is invertible, {\it i.e.}, it is a linear automorphism. We then introduce the quasi-probability measure
\begin{align}\label{def:quasi_prob_A}
\mu_{A}^{\phi}(\Delta|B = b)
&:= T\left(\nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(\Delta) \nonumber \\
&:= \left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(T^{-1}\Delta), \quad \Delta \in \mathfrak{B}^{2},\ b \in \sigma(B)
\end{align}
defined on the measurable space $(\mathbb{R}^{2}, \mathfrak{B}^{2})$ as the image measure ({\it cf}.~see \eqref{def:Bildmass} for the definition of image measures) of the product complex measure with respect to the automorphism $T$ (we shall be shortly returning to the properties of the quasi-probability measure \eqref{def:quasi_prob_A} and the righteousness of its notation). Then, due to the change of variables formula \eqref{eq:Transformationsformel_Bildmass}, one may rewrite our previous findings \eqref{eq:wigner_post} by letting $a_{1}^{\prime} = a_{1} - a_{2}$ and $a_{2}^{\prime} = a_{1} + a_{2}$ as
\begin{align}
&W^{\psi^{g}_{B=b}}(x,p) \nonumber \\
&= \int_{\mathbb{R}} \left( \int_{\mathbb{R}^{2}} \psi^{*}(x - g(a_{1} - a_{2}) + y/2) \right. \nonumber \\
&\qquad \qquad \left. \phantom{\int_{\mathbb{R}}} \times \psi(x - g(a_{1} + a_{2}) - y/2)\ d\mu_{A}^{\phi}(a_{1}, a_{2}|B= b) \right) e^{ipy}\ dm(y) \nonumber \\
&= \int_{\mathbb{R}^{2}} \left( \int_{\mathbb{R}} \psi^{*}((x - ga_{1}) + (y + 2ga_{2})/2) \right. \nonumber \\
&\qquad \qquad \left. \phantom{\int_{\mathbb{R}}} \times \psi((x - ga_{1}) - (y + 2ga_{2})/2) e^{ipy}\ dm(y) \right) d\mu_{A}^{\phi}(a_{1},a_{2}|B= b) \nonumber \\
&= \int_{\mathbb{R}^{2}} e^{-i2ga_{2}p} \left( \int_{\mathbb{R}} \psi^{*}((x - ga_{1}) + y/2) \right. \nonumber \\
& \phantom{e^{-i2ga_{2}p}} \qquad \qquad \left. \phantom{\int_{\mathbb{R}}} \times \psi((x - ga_{1}) - y/2) e^{ipy}\ dm(y) \right) d\mu_{A}^{\phi}(a_{1},a_{2}|B= b)\nonumber \\
&= \int_{\mathbb{R}^{2}} e^{-i2ga_{2}p} W^{\psi}(x - ga_{1},p)\ d\mu_{A}^{\phi}(a_{1},a_{2}|B= b),
\end{align}
where the change of the order of the integration in the second equality is guaranteed by the Fubini's theorem.
For later convenience, we introduce the complex number $a \in \mathbb{C}$ defined by $a := a_{1} + i a_{2}$ by identifying $\mathbb{C} \cong \mathbb{R}^{2}$ in a usual manner, and write
\begin{equation}\label{eq:psm_in_W}
W^{\psi^{g}_{B=b}}(x,p) = \int_{\mathbb{C}} e^{-i2ga_{2}p} W^{\psi}(x - ga_{1},p)\ d\mu_{A}^{\phi}(a|B= b), \quad g \in \mathbb{R}.
\end{equation}
To sum up, here we have learned how the CM scheme may be rewritten in terms of quasi-probability measures, in which the WV distribution of the initial meter wave-function $\psi$ is acted upon by the quasi-probability measure \eqref{def:quasi_prob_A} of the target system to yield the final WV distribution of the meter wave-function $\psi_{B=b}^{g}$.
\paragraph{Changing the Viewpoint through Fourier Transformation}
One finds below that the transcription \eqref{eq:psm_in_W} of the CM scheme admits a much simpler expression when described in terms of the inverse Fourier transform \eqref{eq:omega_norm} of the WV distribution,
rather than the WV distribution itself. Introducing the (yet to be normalised) function $\tilde{\omega}^{\psi^{g}_{B=b}}(x,y)$ uniquely specified through the relation\begin{equation}\label{def:W_to_w_g}
W^{\psi^{g}_{B=b}}(x,p) = \int_{\mathbb{R}} e^{-ipy} \tilde{\omega}^{\psi^{g}_{B=b}}(x,y)\ dm(y), \quad g \in \mathbb{R}
\end{equation}
({\it cf.}, injectivity of the Fourier transformation), the goal of this small paragraph is to show that our finding \eqref{eq:psm_in_W} is equivalent to
\begin{align}\label{eq:psm_in_V_pre}
\tilde{\omega}^{\psi^{g}_{B=b}}(x,y)
&= \int_{\mathbb{C}} \tilde{\omega}^{\psi}(x - ga_{1},y - 2ga_{2})\ d\mu_{A}^{\phi}(a|B = b), \quad g \in \mathbb{R},
\end{align}
which is essentially nothing but the convolution of the initial profile $\tilde{\omega}^{\psi}$ of the meter state by that of the two-dimensional quasi-probability measure $\Delta \mapsto \mu_{A}^{\phi}(\Delta|B = b)$ scaled by $g$. If, moreover, the total integration of $\tilde{\omega}^{\psi}$ happens to be non-vanishing, we may renormalise both sides of the above equality to obtain
\begin{equation}\label{eq:psm_in_V}
\omega^{\psi^{g}_{B=b}}(x,y)
= \int_{\mathbb{C}} \omega^{\psi}(x - ga_{1},y - 2ga_{2})\ d\mu_{A}^{\phi}(a|B = b), \quad g \in \mathbb{R},
\end{equation}
for later use%
\footnote{
Here, note that we have used the general property of convolutions
\begin{equation}
\int_{\mathbb{R}^{n}} (f \ast g)\ d\beta^{n} = \int_{\mathbb{R}^{n}} f\ d\beta^{n} \cdot \int_{\mathbb{R}^{n}} g\ d\beta^{n},\quad f, g \in L^{1}(\mathbb{R}^{n})
\end{equation}
regarding integration.
}.
Observe here the analogy between the unconditioned case \eqref{eq:outcome_prob01}: in both cases, the profile of the `output' of the meter is given by the convolution of the profile of the `input' of the meter and that of the target system scaled by $g$.
To verify our statement, one may simply repeat the previous argument to obtain the result directly, but it is actually easier to demonstrate that the Fourier transforms of the two sides of the above equality coincide. Indeed, the Fourier transform of the l.~h.~s. is nothing but $W^{\psi^{g}_{B=b}}$, which is just the definition \eqref{def:W_to_w_g}. As for the r.~h.~s., one has
\begin{align}
&\int_{\mathbb{R}} e^{-ipy} \left( \int_{\mathbb{C}} \tilde{\omega}^{\psi}(x - ga_{1},y - 2ga_{2})\ d\mu_{A}^{\phi}(a|B = b) \right) dm(y) \nonumber \\
&\qquad = \int_{\mathbb{C}} \left( \int_{\mathbb{R}} e^{-ipy}\, \tilde{\omega}^{\psi}(x - ga_{1},y - 2ga_{2})\ dm(y)\right) d\mu_{A}^{\phi}(a|B = b) \nonumber \\
&\qquad = \int_{\mathbb{C}} e^{-i2ga_{2}p}\, W^{\psi}(x - ga_{1},p)\ d\mu_{A}^{\phi}(a|B = b),
\end{align}
where the exchange of the order of the integration (the first equality) is guaranteed by Fubini's theorem, and the last equality is due to \eqref{eq:Fourier_translation}. Combining the above two results and by observing \eqref{eq:psm_in_W}, the injectivity of the Fourier transformation leads to the desired statement.
We emphasise again that both \eqref{eq:psm_in_W} and \eqref{eq:psm_in_V} represent the same contents seen from different viewpoints.
\subsection{Recovery of the Target Profile}
We are now interested in how one may recover the profile $\Delta \mapsto \mu_{A}^{\phi}(\Delta|B = b)$ of the target system for each $b \in \sigma(B)$ through CM scheme.
As one may expect, the procedure essentially goes analogously to that of the recovery of the probability measure $\mu_{A}^{\phi}$ in the case of the UM scheme demonstrated in Section~\ref{sec:ups_II_ups}.
Recalling the techniques employed there, and by introducing the rescaling
\begin{equation}
\upsilon^{\psi}(x,y) := 2^{-1} \omega^{\psi}(x,2y),
\end{equation}
for the ease of discussion,
one may readily rewrite \eqref{eq:psm_in_V} into
\begin{equation}\label{eq:psm_outcome_prob01}
\upsilon^{\psi^{g}_{B=b}}
= \upsilon^{\psi} \ast \left( \mu_{A}^{\phi}(\, \cdot \, | B=b) \right)_{g}, \quad g \in \mathbb{R},
\end{equation}
or equivalently
\begin{equation}\label{eq:psm_outcome_prob02}
\upsilon^{\psi^{g}_{B=b}}_{g^{-1}}
= \upsilon^{\psi}_{g^{-1}} \ast \mu_{A}^{\phi}(\, \cdot \, | B=b), \quad g \in \mathbb{R}^{\times}
\end{equation}
where the subscript on the respective quasi-probability measures/density functions denotes the scaling \eqref{def:measure_scaling} and \eqref{def:function_scaling}, just as we have done for the case of the UM scheme (see \eqref{eq:outcome_prob01} and \eqref{eq:outcome_prob02}). In parallel to the case of the UM case, these two expressions \eqref{eq:psm_outcome_prob01} and \eqref{eq:psm_outcome_prob02} correspond to the manner in which one combines the interaction parameter \eqref{eq:combining_of_the_interaction}, where the former corresponds to the scaling of the target observable $A \to gA$, whereas the latter corresponds to the scaling of the pair of the meter observables $\{Q, P\} \to \{g^{-1}Q, gP\}$ ({\it cf.} \eqref{eq:outcome_prob_mod_01} and \eqref{eq:outcome_prob_mod_02}).
\subsubsection{Strong Conditioned Measurement}
We now intend to recover the quasi-probability measure $\Delta \mapsto \mu_{A}^{\phi}(\Delta|B = b)$ by making use of the latter expression \eqref{eq:psm_outcome_prob02}. The idea and the procedure are essentially the same as those we have employed in the unconditional case, namely, we manipulate both the interaction parameter $g \in \mathbb{R}^{\times}$ and the initial meter state $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ so that the scaling $\upsilon^{\psi}_{g^{-1}}$ of the inverse Fourier transform of the WV distribution tends towards the delta measure $\delta_{0}$ centred at the origin $0 \in \mathbb{R}^{2}$.
\paragraph{Recovery of the Conditional Quasi-joint-probability}
For the same reason discussed in Section~\ref{sec:Strong_Unconditioned_Measurement}, we assume throughout this passage:
\begin{itemize}
\item The target observable $A$ admits description by density functions.
\item The total integration of $\tilde{\omega}^{\psi}$ is non-vanishing.
\end{itemize}
The first condition guarantees that the quasi-probability measure $\Delta \mapsto \mu_{A}^{\phi}(\Delta|B = b)$, $\Delta \in \mathfrak{B}^{2}$ is absolutely continuous for all $b \in \sigma(B)$, of which density we shall write
\begin{equation}
\rho_{A}^{\phi}(\,\cdot\, |B = b) := \frac{d\mu_{A}^{\phi}(\,\cdot\,|B = b)}{d\beta^{2}}.
\end{equation}
The last condition is necessary in order to assure the well-definedness of $\omega^{\psi}$.
Then, one sees from an analogous argument that we have previously made in Section~\ref{sec:Strong_Unconditioned_Measurement} that, if one adjusts the pair of $g \in \mathbb{R}^{\times}$ and $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ so that $\upsilon^{\psi}_{g^{-1}}$ makes itself an approximate identity in $L^{1}(\mathbb{R}^{2})$, one may let the product of the convolution ({\it i.e.}, the `outcome') converge towards the desired target
\begin{align}
\upsilon^{\psi^{g}_{B=b}}_{g^{-1}} \ \to \ \rho_{A}^{\phi}(\,\cdot\, |B = b)
\end{align}
with respect to the $L^{1}$-norm. A typical way to construct such an approximate identity is to start by preparing a compactly supported wave-function $\psi$, which automatically guarantees $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$, and to consider a family $\{\psi_{(h)}\}_{h > 0}$ of the initial meter state defined as in \eqref{eq:approx_ident_state_Sch}. One then finds
\begin{align}
\tilde{\omega}^{\psi_{(h)}}(x,y)
&:= \psi^{*}_{(h)}(x - y/2) \psi_{(h)}(x + y/2) \nonumber \\
&= |h|^{-1}\psi^{*}\left(\frac{x - y/2}{h}\right) \psi\left(\frac{x + y/2}{h}\right) \nonumber \\
&= |h| \cdot \tilde{\omega}^{\psi}_{h}(x,y),
\end{align}
and hence by observing that the above equality has total integration of $|h| \int_{\mathbb{R}^{2}} \tilde{\omega}^{\psi} dm_{2}$, its normalisation becomes
\begin{equation}
\omega^{\psi_{(h)}}(x,y) = \omega^{\psi}_{h}(x,y).
\end{equation}
This should further lead to
\begin{equation}
\upsilon^{\psi_{(h)}}_{g^{-1}} = \upsilon^{\psi}_{hg^{-1}},
\end{equation}
when scaled by $g^{-1}$. With the initial $\omega^{\psi}$ (or equivalently $\upsilon^{\psi}$) being compactly supported, one then sees that this indeed makes an example of an approximate identity, and we may thus achieve our objective by either narrowing the wave-function $h \to 0$, by intensifying the interaction $g^{-1} \to 0$ ($g \to \pm \infty$) or by appropriately balancing both manoeuvres and letting $hg^{-1} \to 0$ altogether.
\subsubsection{Weak Conditioned Measurement}
We shall next investigate how the map
\begin{equation}\label{eq:parameter_to_omega}
g \mapsto \upsilon^{\psi^{g}_{B=b}}(x,y)
\end{equation}
behaves locally around $g=0$, and discuss what information of the target configuration one might reveal through it. In parallel to the case of the UM scheme discussed in Section~\ref{sec:ups_II_wups}, one finds below that the information of the configuration of the target system is encoded into the differential coefficients of the above map at $g=0$, and that by knowing all the higher-order derivatives, one may fully recover the quasi-probability measure $\Delta \mapsto \mu_{A}^{\phi}(\Delta|B = b)$ of our interest.
\paragraph{Main Objective}
Throughout the following passage, we assume the following.
\begin{itemize}
\item The quasi-probability measure $\Delta \mapsto \mu_{A}^{\phi}(\Delta|B = b)$ has a compact support.
\item The total integration of $\tilde{\omega}^{\psi}$ is non-vanishing, and its normalisation belongs to the Schwartz space $\omega^{\psi} \in \mathscr{S}(\mathbb{R}^{2})$.
\end{itemize}
These requirements are imposed primarily for the same reason as we have previously discussed in analysing the weak UM scheme in Section~\ref{sec:ups_II_wups} (which, in short, is to say that we do not wish to get involved in the theory of generalised functions). A sufficient condition for the first and second assumptions would be to respectively require that the spectral measure $E_{A}$ be compactly supported, and that $\psi \in \mathscr{S}(\mathbb{R})$.
Under such conditions, the main objective of this passage is to demonstrate the following Proposition:
\begin{proposition}[Weak Conditioned Measurement]\label{prop:WPSM}
Under the above conditions, the map \eqref{eq:parameter_to_omega} is arbitrarily many times strongly differentiable on all the real line $\mathbb{R}$, and its $n$th derivatives at the origin $g=0$ reads
\begin{align}
\left. \frac{d^{n}}{dg^{n}} \upsilon^{\psi^{g}_{B=b}} \right|_{g=0} = \sum_{|\gamma| = n} \mathbb{E} \left[a^{\gamma} ; \mu_{A}^{\phi}(\,\cdot\,|B = b) \right] \cdot (- D)^{\gamma}\upsilon^{\psi}, \quad n \in \mathbb{N}_{0}.
\end{align}
Here, $\gamma = (\gamma_{1}, \gamma_{2}) \in \mathbb{N}_{0}^{2}$ is a multi-index introduced in \eqref{def:use_alpha}, and the `quasi-moments' under the quasi-probability measure $\mu_{A}^{\phi}(\,\cdot\,|B = b)$ is defined by
\begin{equation}\label{def:moments_qjpm}
\mathbb{E} \left[a^{\gamma} ; \mu_{A}^{\phi}(\,\cdot\,|B = b) \right] := \int_{\mathbb{C}} a_{1}^{\gamma_{1}}a_{2}^{\gamma_{2}}\ d\mu_{A}^{\phi}(a|B = b),
\end{equation}
in its explicit form, where we understand $a = a_{1} + i a_{2} \in \mathbb{C}$.
\end{proposition}
\begin{proof}
Since the assumptions and reasonings are essentially the same as those provided for the unconditioned counterpart, we shall avoid reiteration and provide a rough sketch of the proof. In order to avoid clumsiness of notation, we write $\upsilon := \upsilon^{\psi}$, $\upsilon[g] := \upsilon^{\psi^{g}_{B=b}}$ and $\mu := \mu_{A}^{\phi}(\,\cdot\,|B = b)$ for simplicity, and denote by $\upsilon^{(n)}[g]$ the $n$th derivative of the map $g \mapsto \upsilon[g]$.
We first prove that the $n$th derivative of $\upsilon[g]$ reads
\begin{equation}\label{eq:n_wcd}
\upsilon^{(n)}[g](x) = \sum_{|\gamma| = n} \int_{\mathbb{R}^{2}} (- D)^{\gamma}\upsilon(x - ga) a^{\gamma}\ d\mu(a).
\end{equation}
As above, we argue by mathematical induction. The case $n=0$ is trivial. Suppose that the statement is true for $n \in \mathbb{N}_{0}$. Then, one may compute its point-wise derivative as
\begin{align}
\frac{d}{dg} \upsilon^{(n)}[g](x)
&= \sum_{|\gamma| = n} \int_{\mathbb{R}^{2}} \left( \frac{d}{dg} (- D)^{\gamma}\upsilon(x - ga) \right) a^{\gamma}\ d\mu(a) \nonumber \\
&= \sum_{|\gamma| = n} \int_{\mathbb{R}^{2}} \left( \sum_{i=1}^{2} a_{i} (-D)_{i}(- D)^{\gamma}\upsilon(x - ga) \right) a^{\gamma}\ d\mu(a) \nonumber \\
&= \sum_{|\gamma| = n + 1} \int_{\mathbb{R}^{2}} (- D)^{\gamma}\upsilon(x - ga) a^{\gamma}\ d\mu(a),
\end{align}
and subsequently prove its strong differentiability by employing the same technique as above. This completes our first step of the proof.
Now, by taking $g=0$ of \eqref{eq:n_wcd}, we observe
\begin{align}
\upsilon^{(n)}[0](x)
&= \sum_{|\gamma| = n} \int_{\mathbb{R}^{2}} (- D)^{\gamma}\upsilon(x - 0a) a^{\gamma}\ d\mu(a) \nonumber \\
&= \sum_{|\gamma| = n} \int_{\mathbb{R}^{2}} a^{\gamma}\ d\mu(a) \cdot (- D)^{\gamma}\upsilon(x) \nonumber \\
&= \sum_{|\gamma| = n} \mathbb{E}[a^{\gamma}; \nu] \cdot (- D)^{\gamma}\upsilon(x),
\end{align}
which completes our proof.
\end{proof}
\noindent
One then immediately obtains the following corollary by applying the Stone-Weierstra{\ss} approximation theorem and the Riesz-Markov-Kakutani representation theorem.
\begin{corollary}[Recovery of the Target Profile by Weak Conditioned Measurement]
The weak CM scheme ({\it i.e.}, the knowledge of all the `quasi-moments' \eqref{def:moments_qjpm}) allows us to uniquely specify the quasi-probability measure $\mu_{A}^{\phi}(\,\cdot\,|B = b)$ of our interest.
\end{corollary}
\noindent
Compare these results to those obtained in the case of the weak UM scheme described in Section~\ref{sec:ups_II_wups}.
\subsection{Profile of the Target System}
We have so far investigated how the CM scheme can be transcribed into the language of conditional quasi-probabilities, rather than in terms of mere conditional expectations. As a result, we found that the measurement outcome after the interaction incorporates two components: one being the profile of the meter system in the form of the WV distribution and the other being the that of the target system in the form of the quasi-probability measure $\mu_{A}^{\phi}(\, \cdot \, | B=b)$ defined in \eqref{def:quasi_prob_A}. Specifically, in view of the (scaled) inverse Fourier transform of the WV distribution, we found that the manner in which the two components interact with each other admits a simple description by convolution \eqref{eq:psm_in_V}, which is quite analogous to the unconditioned case. Based on our findings, we have thus analysed how one may recover the profile $\mu_{A}^{\phi}(\, \cdot \, | B=b)$ by means of both the strong and weak CM schemes, whose procedures are also quite analogous to the unconditioned counterpart. We are now interested in the properties of the quasi-probability measure $\mu_{A}^{\phi}(\, \cdot \, | B=b)$ we have obtained, which should be expected to convey some information of the target system.
\paragraph{Quasi-joint-probability Distribution of a Pair of Observables}
By means of either the strong or weak CM scheme, we have so far obtained the family of quasi-probability measures $\mu_{A}^{\phi}(\, \cdot \, | B=b)$ for all $b \in \sigma(B)$.
Allowing it to extend on the whole real line, one may construct a complex transition kernel by
\begin{equation}\label{def:quasi_trans_kern}
\mu_{A}^{\phi}(\Delta_{A} | B=b) :=
\begin{cases}
\mu_{A}^{\phi}(\Delta_{A} | B=b), & (b \in \sigma(B)) \\
\text{indefinite}, & (b \notin \sigma(B))
\end{cases}
, \quad \Delta_{A} \in \mathfrak{B}(\mathbb{C})
\end{equation}
from the space $(\mathbb{R}, \mathfrak{B}^{1})$ of the measurement outcomes of $B$ into $(\mathbb{C}, \mathfrak{B}(\mathbb{C}))$. For definiteness, we assign to each $b \notin \sigma(B)$ any quasi-probability measure, so that \eqref{def:quasi_trans_kern} defines a \emph{transitional quasi-probability kernel} as a whole.
This allows us to construct a quasi-probability measure $\mu_{A,B}^{\phi}$ on the product space $(\mathbb{C} \times \mathbb{R}, \mathfrak{B}(\mathbb{C}) \otimes \mathfrak{B}^{1})$, by combining the transition quasi-probability kernel \eqref{def:quasi_trans_kern} and the probability measure $\mu_{B}^{\phi}$, that satisfies
\begin{equation}\label{def:quasi_joint_prob_meas}
\mu_{A,B}^{\phi}(\Delta_{A} \times \Delta_{B}) = \int_{\Delta_{B}} \mu_{A}^{\phi}(\Delta_{A} | B=b)\ d\mu_{B}^{\phi}(b), \quad \Delta_{A} \in \mathfrak{B}(\mathbb{C}),\ \Delta_{B} \in \mathfrak{B}^{1},
\end{equation}
whose existence is guaranteed by Proposition~\ref{prop:ctk_to_cm}.
The target of our analysis in this passage is the quasi-probability measure \eqref{def:quasi_joint_prob_meas}. As one may expect from the notation employed, we shall shortly see that this qualifies as a QJP distribution of the target observable $A$ and the conditioning observable $B$.
\begin{proposition}[Quasi-joint-probability Distribution]
Under the definitions above, the quasi-probability measure $\mu_{A,B}^{\phi}$ qualifies as a QJP distribution of $A$ and $B$ in the sense of \eqref{def:qjpm}, namely
\begin{align}
\begin{split}
\mu_{A,B}^{\phi}(\Delta_{A} \times \mathbb{R}) &= \mu_{A}^{\phi}(\Delta_{A}), \quad \Delta_{A} \in \mathfrak{B}(\mathbb{C}), \\
\mu_{A,B}^{\phi}(\mathbb{C} \times \Delta_{B}) &= \mu_{B}^{\phi}(\Delta_{B}), \quad \Delta_{B} \in \mathfrak{B}(\mathbb{R})
\end{split}
\end{align}
holds. Here, $\mu_{A}^{\phi}$ denotes the probability measure on $(\mathbb{C}, \mathfrak{B}(\mathbb{C}))$ generated by the two-dimensional spectral measure associated to $A$ understood as a normal operator, whereas $\mu_{B}^{\phi}$ denotes the probability measure on $(\mathbb{R}, \mathfrak{B})$ generated by the one-dimensional spectral measure associated to the self-adjoint operator $B$.
\end{proposition}
\begin{proof}
We start by demonstrating that the marginal of the quasi-probability measure $\mu_{A,B}^{\phi}$ of the first term coincides with the probability measure $\mu_{A}^{\phi}$ on $(\mathbb{C}, \mathfrak{B}(\mathbb{C}))$ generated by the spectral measure of $A$ (seen as a normal operator). To this end, we first observe
\begin{align}\label{eq:marginal_A}
\mu_{A,B}^{\phi}(\Delta_{A} \times \mathbb{R})
&= \int_{\mathbb{R}} \mu_{A}^{\phi}(\Delta_{A} | B=b)\ d\mu_{B}^{\phi}(b) \nonumber \\
&= \sum_{b \in \sigma(B)} \left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(T^{-1}\Delta_{A}) \cdot |\langle b , \phi \rangle|^{2},
\end{align}
where we have used \eqref{def:quasi_prob_A} in the last equality.
Now, in order to proceed further, we then maintain that the measure
\begin{equation}
\Delta \mapsto \mu(\Delta) := \sum_{b \in \sigma(B)} \left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(\Delta) \cdot |\langle b , \phi \rangle|^{2}, \quad \Delta \in \mathfrak{B}(\mathbb{C})
\end{equation}
is essentially the same object as the continuous $\mathbb{C}$-linear map defined by
\begin{equation}
I_{\mathrm{diag}}(f) := \int_{\mathbb{R}} f(a,a)\ d\mu_{A}^{\phi}(a), \quad f \in C_{0}(\mathbb{C})
\end{equation}
in the sense of the Riesz-Markov-Kakutani representation theorem.
The proof can be carried out in several ways, but for the sake of simplicity, we rather take an elementary approach.
Observing that any two measures on a product space $(X \times Y,\, \mathfrak{A} \otimes \mathfrak{B})$ coincides with each other if they coincide on the subset $\mathfrak{A} \ast \mathfrak{B} \subset \mathfrak{A} \otimes \mathfrak{B}$ (see \eqref{def:product_set_of_sigma_algebras} for the definition), one proceeds as
\begin{align}
\int_{\mathbb{C}}\chi_{\Delta_{1}}(a_{1})\chi_{\Delta_{2}}(a_{2})\ d\mu(a)
&= \mu(\Delta_{1} \times \Delta_{2})\nonumber \\
&= \sum_{b \in \sigma(B)} \left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(\Delta_{1} \times \Delta_{2}) \cdot |\langle b , \phi \rangle|^{2} \nonumber \\
&= \sum_{b \in \sigma(B)} \langle \phi, E_{A}(\Delta_{1}) b\rangle \langle b, E_{A}(\Delta_{2}) \phi \rangle \nonumber \\
&= \langle \phi, E_{A}(\Delta_{1})E_{A}(\Delta_{2}) \phi \rangle \nonumber \\
&= \int_{\mathbb{R}} \chi_{\Delta_{1}}(a)\chi_{\Delta_{2}}(a)\ d\mu_{A}^{\phi}(a) \nonumber \\
&= I_{\mathrm{diag}}(\chi_{\Delta_{1}}\chi_{\Delta_{2}}),
\end{align}
which proves
\begin{equation}
\mu \cong I_{\mathrm{diag}}.
\end{equation}
Armed with the findings, we return to our original problem \eqref{eq:marginal_A} and finally obtain
\begin{align}
\sum_{b \in \sigma(B)} \left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(T^{-1}\Delta_{A}) \cdot |\langle b , \phi \rangle|^{2}
&= I_{\mathrm{diag}}(\chi_{(T^{-1}\Delta_{A})}) \nonumber \\
&= \int_{\mathbb{R}} \chi_{(T^{-1}\Delta_{A})}(a,a)\ d\mu_{A}^{\phi}(a) \nonumber \\
&= \int_{\mathbb{R}} \chi_{\Delta_{A}}(a,0)\ d\mu_{A}^{\phi}(a) \nonumber \\
&= \langle \phi, E_{A, 0}(\Delta_{A}) \phi \rangle \nonumber \\
&= \mu_{A}^{\phi}(\Delta_{A}), \quad \Delta_{A} \in \mathfrak{B}(\mathbb{C}),
\end{align}
where $E_{A, 0}$ denotes the product spectral measure of the one-dimensional spectral measure $E_{A}$ of $A$ (as a self-adjoint operator) and that of the $0$ operator $E_{0}$ ({\it i.e.,} the `delta spectral measure' \eqref{def:delta_spectral_meas_0} centred at the origin), and the last equality is due to the observation that the two-dimensional spectral measure $\tilde{E}_{A}$ of $A$ as a normal operator coincides with the product spectral measure $\tilde{E}_{A} = E_{A, 0}$ introduced above. This completes our proof for the marginal of the first term.
It now remains to compute the marginal of $\mu_{A,B}^{\phi}$ of the second term, which one carries out as
\begin{align}
\mu_{A,B}^{\phi}(\mathbb{C} \times \Delta_{B})
&= \int_{\mathbb{C} \times \mathbb{R}} \chi_{\Delta_{B}}(b)\ d\mu_{A,B}^{\phi}(a,b) \nonumber \\
&= \int_{\mathbb{C} \times \mathbb{R}} \chi_{\Delta_{B}}(b) \mu_{A}^{\phi}(\mathbb{C} | B=b)\ d\mu_{B}^{\phi}(b) \nonumber \\
&= \int_{\mathbb{C} \times \mathbb{R}} \chi_{\Delta_{B}}(b)\ d\mu_{B}^{\phi}(b) \nonumber \\
&= \mu_{B}^{\phi}(\Delta_{B}), \quad \Delta_{B} \in \mathfrak{B},
\end{align}
where the second equality is due to the definition of $\mu_{A,B}^{\phi}$, and the third equality is due to the fact that $\mu_{A}^{\phi}(\mathbb{C} | B=b) = 1$ is normalised to unity ({\it i.e.}, a quasi-probability measure) for all $b \in \sigma(B)$.
\end{proof}
\noindent
As for the relation between the QJP distribution $\mu_{A,B}^{\phi}$ and the transition quasi-probability kernel $(b,\Delta_{A}) \mapsto \mu_{A}(\Delta_{A}|B=b)$, one immediately has the following corollary by construction.
\begin{corollary}
The transition quasi-probability kernel \eqref{def:quasi_trans_kern} is a conditional quasi-probability distribution of $A$ given $B$ under the QJP distribution $\mu_{A,B}^{\phi}$.
\end{corollary}
\paragraph{Conditional Quasi-expectation of $A$ given $B$}
It is now tempting to investigate how the `conditional average' of the QJP distribution $\mu_{A,B}^{\phi}$ relates to the conditional quasi-expectation $\mathbb{E}^{\alpha}[A|B;\phi]$ we have introduced earlier in \eqref{def:cond_quasi-exp_alpha}.
\begin{proposition}[Conditional Average of the Quasi-joint-probability Distribution]
Under the definitions above, the conditional average of $A$ given $B$ under the QJP distribution $\mu_{A,B}^{\phi}$ reads
\begin{align}\label{eq:cond_quas-exp_as_conditional_average}
\int_{\mathbb{C}} a\ d\mu_{A}^{\phi}(a | B=b)
&= \mathbb{E}^{i}[A | B = b ;\phi],
\end{align}
where the r.~h.~s. is the member of the complex-parametrised sub-family of conditional quasi-expectations of $A$ given $B$ introduced in \eqref{def:cond_quasi-exp_alpha} for the purely imaginary choice $\alpha = i$ of the parameter.
\end{proposition}
\begin{proof}
For the demonstration, let $b \in \sigma(B)$.
One then has
\begin{align}
\int_{\mathbb{C}} a\ d\mu_{A}^{\phi}(a | B=b)
&= \int_{\mathbb{R}^{2}} (a_{1} + i a_{2})\ dT\left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(a_{1}, a_{2}) \nonumber \\
&= \int_{\mathbb{C}} \frac{a + a^{\prime}}{2} + i\frac{a - a^{\prime}}{2} \ d\left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(a^{\prime}, a) \nonumber \\
&= \int_{\mathbb{C}} \frac{1 + i}{2} a + \frac{1 - i}{2} a^{\prime} \ d\left( \nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}} \right)(a, a^{\prime}) \nonumber \\
&= \frac{1 + i}{2} \cdot \frac{\langle b, A \phi \rangle}{\langle b, \phi \rangle} + \frac{1 - i}{2} \cdot \left( \frac{\langle b, A \phi \rangle}{\langle b, \phi \rangle} \right)^{*} \nonumber \\
&= \mathbb{E}^{i}[A | B = b ;\phi],
\end{align}
where the second equality is due to the change of variables formula \eqref{eq:Transformationsformel_Bildmass} for image measures.
\end{proof}
\paragraph{Obtaining Conditional Quasi-probability Distribution by Conditioned Measurement}
We now realise that the CM scheme, in view of conditional quasi-probabilities, can be regarded as a \emph{method of obtaining conditional quasi-probability distributions of the target observable $A$ given the conditioning observable $B$}, and that it implies the existence of QJP distributions of a pair of (generally not necessarily simultaneously measurable) quantum observables lying underneath. Moreover, we have seen a connection between the concept of conditional quasi-expectations and the `conditional average' of the QJP distributions, which is reminiscent of the familiar relation between classical conditional expectations and conditional average of probability measures.
While we have conducted an analysis for the special case in which both $A$ and $B$ happen to possess spectra of finite cardinalities (and that $B$ is degenerate), we note that one may suitably generalise the results obtained here by introducing appropriate mathematical tools and some little more advanced mathematical languages.
\newpage
\section{Quasi-probabilities of Quantum Observables}\label{sec:qp_qo}
By studying the both the UM and CM schemes in depth throughout the preceding four sections, we have so far naturally arrived, by a purely bottom-up construction, at the concept of \emph{quasi-joint-probability} (QJP) of an arbitrary pair of quantum observables. While such an operational way of demonstration has its own merit of being solid and down to earth, it has an apparent downside in that the line of argument lacks transparency and that the whole structure may become obscure on occasions. In this section, we will be conducting a top-down study on the topic
as a complement to the analyses made in the preceding sections.
\paragraph{Organisation of this Section}
In this section, we first devote several pages to introducing some mathematical tools for our analysis as usual. We then propose a general prescription for the construction of QJP distributions of a given pair of quantum observables, and observe their basic properties. Since it is difficult to perform a general analysis on the whole class of all possible candidates of QJP distributions with full mathematical rigour due to the limited framework and tools available, for our demonstration we shall mostly concentrate on a special sub-family of such distributions parametrised by a single complex number, hopefully without loss of too much essence. We finally close this section by observing where the bottom-up line of discussion performed in Section~\ref{sec:ps_II} fits in this more general framework.
\subsection{Reference Materials}
As usual, we first prepare some necessary mathematical tools for reference. As a generalisation to those defined on integrable functions, we now introduce Fourier transforms of complex measures.
\subsubsection{Fourier Transform of Complex Measures}
Analogous to the manner in which we have defined Fourier transforms of elements of $L^{1}(\mathbb{R}^{n})$ (namely, the density functions), one may define Fourier transforms of complex measures. Given a complex measure $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{A})$ on a measurable space $(X, \mathfrak{A})$, we define the Fourier transform and the inverse Fourier transform of $\mu$, respectively by the functions
\begin{align}
\hat{\mu}(q) &:= \int_{X} e^{-i\langle q, x \rangle}\ d\mu(x), \label{def:FT_cm} \\
\check{\mu}(q) &:= \int_{X} e^{i\langle q, x \rangle}\ d\mu(x), \label{def:IFT_cm}
\end{align}
where $\langle q, x \rangle := \sum_{k=1}^{n} q_{k}x_{k}$ denotes the scalar product on $\mathbb{R}^{n}$ as usual. Note that the functions $\hat{\mu}$, $\check{\mu}$ are well-defined, for indeed $|\hat{\mu}(q)| \leq \int_{X} |e^{-i\langle q, x \rangle}|\, d|\mu|(x) = \|\mu\| < \infty$ for all $q \in \mathbb{R}^{n}$, where $|\mu|$ and $\|\mu\|$ are respectively the variation and the total variation of $\mu$ (a similar evaluation holds for $\check{\mu}$).
\paragraph{Basic Properties}
To see how this newly introduced definition of Fourier transforms relates to that of integrable functions introduced earlier, let $L^{1}(\mathfrak{B}^{n}) \subset \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ be the sub-algebra of absolutely continuous complex measures with respect to $m_{n}$, where $m_{n}$ denotes the renormalised $n$-dimensional Lebesgue-Borel measure on $(\mathbb{R}^{n},\mathfrak{B}^{n})$ defined in \eqref{def:renormalised_LB_measure}. Choosing $\mu \in L^{1}(\mathfrak{B}^{n})$ and letting $\rho := d\mu/dm_{n}$ be the Radon-Nikod{\'y}m derivative of $\mu$, one finds by a direct application of \eqref{eq:Mass_mit_Dichte_AC} that
\begin{align}\label{eq:FT_measure_vs_density}
\hat{\mu}(q)
&:= \int_{\mathbb{R}^{n}} e^{-i\langle q, x \rangle}\ d\mu(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} e^{-i\langle q, x \rangle} \rho(x)\ dm_{n}(x)
=: \hat{\rho}(q),
\end{align}
holds as expected. An analogous relation holds for the inverse Fourier transform as well. The $\mathbb{C}$-linear map $\mathscr{F}$ that maps a complex measure into its Fourier transform is called the \emph{Fourier transformation}. In parallel to that defined for integrable functions, the Fourier transformation on the measure algebra is injective, {\it i.e.}, $\hat{\mu} = \hat{\nu}$ implies $\mu = \nu$. For $\mu, \nu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$, the
properties
\begin{align}
\widehat{(\mu \ast \nu)} &= \hat{\mu} \cdot \hat{\nu}, \\
\widehat{(\mu_{t})}(q) &= \hat{\mu}(tq), \quad t \neq 0, \\
\widehat{(\tau_{a}\mu)}(t) &= e^{i\langle a,x\rangle}\hat{\mu}(t),\quad a \in \mathbb{R},
\end{align}
are basic, in which one sees how the Fourier transformation behaves under the convolution \eqref{def:convolution_measure}, scaling \eqref{def:measure_scaling}, and translation
\begin{equation}\label{def:translation_measure}
(\tau_{a}\mu)(B) := \mu(B + a), \quad a \in \mathbb{R}^{n},
\end{equation}
respectively.
\paragraph{Linear Transformation}
Let $T$ be a linear operator on $\mathbb{R}^{n}$ ({\it i.e.}, an $n \times n$ real matrix), and let $\mu \in \mathbf{M}_{\mathbb{C}}(\mathfrak{B}^{n})$ be a complex measure. We then define the \emph{linear transform}
\begin{equation}\label{def:lin_trans_measure}
\mu_{T}(B) := \mu(T^{-1}B), \quad B \in \mathfrak{B}^{n},
\end{equation}
of $\mu$ with respect to $T$ by its image measure. By definition, one readily finds the validity of the product rule
\begin{equation}\label{eq:product_rule_image_measure_lin}
(\mu_{T})_{S} = \mu_{(ST)}
\end{equation}
for a pair of linear operators $S$ and $T$ on $\mathbb{R}^{n}$, and that
\begin{equation}
\int_{X} f(x)\ d\mu_{T}(x)
= \int_{X} f(Tx)\ d\mu(x)
\end{equation}
by the change of variables formula \eqref{eq:Transformationsformel_Bildmass}, whenever the integration exists. Note that the familiar scaling $\mu_{t}$, $t \neq \mathbb{R}$ defined in \eqref{def:measure_scaling}, and the translation $\tau_{a}\mu$, $a \in \mathbb{R}^{n}$ defined in \eqref{def:translation_measure} are respectively special cases of the linear transform of $\mu$ with respect to $T = tI$ and $T = I - a$, where $I$ denotes the identity operator. In such a cases, note also that the linear operators involved are automorphisms, hence members of the general linear group $\mathrm{GL}(n;\mathbb{R})$.
In relation to the Fourier transformation, one finds
\begin{align}\label{eq:FT_and_Lin_Trans}
(\mathscr{F}\mu_{T})(q)
&:= \int_{\mathbb{R}^{n}} e^{-i\langle q,x\rangle}\ d\mu_{T}(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} e^{-i\langle q,Tx\rangle}\ d\mu(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} e^{-i\langle T^{*}q,x\rangle}\ d\mu(x) \nonumber \\
&= (\mathscr{F}\mu)(T^{*}q),
\end{align}
where $T^{*}$ denotes the adjoint (in this case, the transpose $T^{*} = T^{t}$) of the Matrix $T$.
\paragraph{Complex Conjugate}
We finally review how the Fourier transform behaves regarding the operation of taking the complex conjugate of a complex measure. To this end, let $\mu$ be a complex measure on $(\mathbb{R}^{n}, \mathfrak{B}^{n})$, and define the complex conjugate of $\mu$ by $\mu^{*}(\Delta) := \mu(\Delta)^{*}$, $\Delta \in \mathfrak{B}^{n}$ in a natural manner. One then readily finds
\begin{align}\label{eq:FT_cc_Measure}
(\mathscr{F}\mu^{*})(q)
&:= \int_{\mathbb{R}^{n}} e^{-i\langle q,x\rangle}\ d\mu^{*}(x) \nonumber \\
&= \left( \int_{\mathbb{R}^{n}} e^{-i\langle -q,x\rangle}\ d\mu(x) \right)^{*} \nonumber \\
&= (\mathscr{F}\mu)^{*}(-q) \nonumber \\
&= (\mathscr{F}\mu)^{\dagger}(q),
\end{align}
where $f^{\dagger}(x) := f^{*}(-x)$ denotes the \emph{involution} of a function $f$.
\paragraph{Differentiation}
We finally make a brief note on the basic results regarding differentiability and derivatives of a Fourier transform of a complex measure at the origin.
\begin{lemma}\label{lem:FT_CM_diff_exp}
Let $\mu$ be a complex measure on $(\mathbb{R}^{n},\mathfrak{B}^{n})$, and let $\gamma \in \mathbb{N}_{0}^{n}$ be a multi-index. If the integration
\begin{equation}
\int_{\mathbb{R}^{n}} x^{\gamma^{\prime}}\ d\mu(x),
\end{equation}
exists for all $0 \leq \gamma^{\prime} \leq \gamma$, then the derivative $D^{\gamma}\hat{\mu}$ of the Fourier transform of $\mu$ exists at the origin, in which case the derivative reads
\begin{equation}
\left(D^{\gamma}\hat{\mu}\right)(0) = (-i)^{|\gamma|}\int_{\mathbb{R}^{n}} x^{\gamma}\ d\mu(x).
\end{equation}
\end{lemma}
\begin{proof}
One readily computes
\begin{align}
\left(D^{\gamma}\hat{\mu}\right)(0)
&:= \left. D^{\gamma} \left( \int_{\mathbb{R}^{n}} e^{-i\langle q, x\rangle}\ d\mu(x) \right) \right|_{q=0} \nonumber \\
&= \int_{\mathbb{R}^{n}} \left( \left. D^{\gamma} e^{-i\langle q, x\rangle} \right|_{q=0} \right) d\mu(x) \nonumber \\
&= (-i)^{|\gamma|}\int_{\mathbb{R}^{n}} x^{\gamma}\ d\mu(x),
\end{align}
where the second equality (exchange of the differentiation and integration) is a consequence of the dominated convergence theorem.
\end{proof}
\noindent
Compare this result to that for Schwartz functions \eqref{eq:fourier_diff_01}.
\subsection{Quasi-joint-probabilities of a Combination of Quantum Observables}
We now intend to provide a general prescription for defining a QJP distribution of a combination of generally not necessarily simultaneously measurable quantum observables.
\subsubsection{Preliminary Observations}
In this passage, we conduct some formal discussions on the topic of QJP distributions of a combination of quantum observables. Since rigorous treatment requires advanced mathematical tools that is beyond the scope of this paper, we first conduct a formal and intuitive argument to obtain the essence of the idea.
Now, before we embark on our main objective, we first recall a basic theorem regarding strong commutativity of $A$ and $B$ and that of their unitary operators.
\begin{theorem*}
Let $A$ and $B$ be self-adjoint. Then, the following conditions are equivalent.
\begin{enumerate}
\item The operators $A$ and $B$ strongly commute with each other.
\item The operators $e^{isA}$ and $e^{itB}$ commute with each other for all $s,t \in \mathbb{R}$, namely
\begin{equation}\label{eq:strong_commutativity_vs_FT}
e^{itA}e^{isB} = e^{isB}e^{itA}, \quad s,t \in \mathbb{R}
\end{equation}
holds.
\end{enumerate}
\end{theorem*}
\noindent
This familiar theorem builds the starting point of our discussion that follows.
\paragraph{Fourier Transform of Product Spectral Measures}
Recall that the joint behaviour of the outcomes of an ideal measurement of a pair of simultaneously measurable observables $A$ and $B$ is governed by the product spectral measure $E_{A,B}$ of their respective spectral measures $E_{A}$, $E_{B}$ introduced earlier in \eqref{eq:product_spectral_measure}. An important observation here is to see that the `Fourier transform' of the product spectral measure $E_{A,B}$ is nothing but the product \eqref{eq:strong_commutativity_vs_FT} of the parametrised unitary operators
\begin{align}
(\mathscr{F}E_{A,B})(s,t)
&:= \int_{\mathbb{R}^{2}} e^{-i(\langle s, a\rangle + \langle t, b\rangle)}\ dE_{A,B}(a,b) \nonumber \\
&= e^{-i\overline{(sA + tB)}} \nonumber \\
&= \lim_{N \to \infty} ( e^{-isA/N} e^{-itB/N} )^{N} \nonumber \\
&= e^{-isA}e^{-itB},
\end{align}
where the overline on the essentially self-adjoint operator $sA + tB$ denotes its unique self-adjoint extension as usual, and the second equality is due to the familiar Trotter formula.
\paragraph{Hashed Operators}
We now consider a pair of \emph{arbitrary} (not necessarily strongly-commuting) self-adjoint operators $A$ and $B$. Guided by the above observation, we formally introduce
\begin{align}\label{def:decent_mixture}
\#(s,t) :=\, \text{a `decent' mixture of the disintegrated components of } e^{-isA} \text{ and } e^{-itB}
\end{align}
for the pair of $A$ and $B$. Example of such mixtures of the disintegrated components of the unitary operators are given by:
\begin{align}\label{eq:mixture_examples}
\#(s,t) =
\begin{cases}
e^{-isA}e^{-itB}, \\
e^{-itB}e^{-isA}, \\
\Pi_{k = 1}^{N} e^{-isA/L_{k}}e^{-itB/M_{k}}, \qquad \left(\sum_{k=1}^{N}L_{k}^{-1} = 1,\ \sum_{k=1}^{N}M_{k}^{-1} = 1\right), \\
\left( e^{-isA/N} e^{-itB/N} \right)^{N}, \\
e^{-i\overline{(sA + tB)}} = \lim_{N \to \infty} \left( e^{-isA/N} e^{-itB/N} \right)^{N}, \\
\textit{etc.},
\end{cases}
\end{align}
or even any linear combinations of them.
The term `decent' is intended to express a mathematical condition as to what qualifies as a reasonable `mixture' to meet our purpose. However, we do not intend to discuss its precise mathematical definition here, for it is beyond the scope of this paper. In this paper, the `parametrised family of operators' $\#(s,t)$ shall occasionally be referred to as \emph{hashed operators} of the unitary operators, in a rather casual manner. Due to the commutativity of the unitary operators for a simultaneously measurable pair, the hashed operator $\#(s,t) = e^{-isA}e^{-itB}$ is always unique, while on the other hand, hashed operators admit variety for non-commutative pairs.
Now, given a hashed operator $\#$ of $A$ and $B$, we then introduce the collection of all parametrised operators of the form
\begin{equation}\label{def:QSM_FT}
\hat{\mathfrak{M}}_{A,B} := \left\{ \# : \# \text{ is a hashed operator of $A$ and $B$ defined as in \eqref{def:decent_mixture} } \right\}.
\end{equation}
As we have seen above, in the case in which $A$ and $B$ are simultaneously measurable, the above collection in fact consists of only one trivial element
\begin{equation}
\hat{\mathfrak{M}}_{A,B} = \left\{ e^{-isA}e^{-itB} \right\},
\end{equation}
due to the strong commutativity of the two operators.
On the other hand, one readily observes that the cardinality of $\hat{\mathfrak{M}}_{A,B}$ is always greater than unity if the pair of observables $A$ and $B$ fails to strongly commute.
\begin{lemma}\label{lem:card_hash_op}
The cardinality of the collection $\hat{\mathfrak{M}}_{A,B}$ is equal to unity if and only if the observables $A$ and $B$ strongly commute with each other. Otherwise, the cardinality is always greater than unity.
\end{lemma}
\paragraph{Distributions generated by Hashed Operators}
We now consider the inverse Fourier transform of all the elements of the hashed operators $\hat{\mathfrak{M}}_{A,B}$, and thus formally introduce
\begin{equation}\label{def:QJSD}
\mathfrak{M}_{A,B} := \left\{\mathscr{F}^{-1} \# : \# \in \hat{\mathfrak{M}}_{A,B} \right\},
\end{equation}
without any consideration of the mathematical intricacies involved in its well-definedness. By the injectivity of the Fourier transformation, one intuitively expects that the collection reduces to the single element
\begin{equation}
\mathfrak{M}_{A,B} = \left\{ E_{A,B} \right\}
\end{equation}
when the operators $A$ and $B$ strongly commute with each other, which should be nothing but the original product spectral measure of $A$ and $B$. On the other hand, Lemma~\ref{lem:card_hash_op} implies that the cardinality of the collection $\mathfrak{M}_{A,B}$ is always greater than unity in the case where $A$ and $B$ are not simultaneously measurable. Although being possibly highly non-unique, each element of the collection $\mathfrak{M}_{A,B}$ defined for non-commuting pairs retains similar properties to those of the standard product spectral measures. Incidentally, choosing any element $\Pi \in \mathfrak{M}_{A,B}$ of the collection, the `total integration' reduces to the unit $I$, as one finds under the formal computation
\begin{align}\label{eq:tot_int_joint_spect_dist}
\int_{\mathbb{K}^{2}} \Pi(a,b)\ dm_{2}(a,b)
&= \int_{\mathbb{K}^{2}} e^{-i (\langle 0,a\rangle + \langle 0,b\rangle)} \Pi(a,b)\ dm_{2}(a,b) \nonumber \\
&= \left( \mathscr{F}\Pi \right)(0,0) \nonumber \\
&= \#(0,0) = I,
\end{align}
where $\#$ is the hashed operator whose inverse Fourier transform is the element $\Pi = \mathscr{F}^{-1}\#$ of our choice. As for the marginals, by formally introducing
\begin{equation}
\Pi_{B}(b) := \int_{\mathbb{K}} \Pi(a,b)\ dm(a),
\end{equation}
one observes under a formal computation that
\begin{align}
\left(\mathscr{F}\Pi_{B}\right) (t)
&= \int_{\mathbb{K}} e^{-i \langle t,b\rangle} \left( \int_{\mathbb{K}} \Pi(a,b)\ dm(a) \right) dm(b) \nonumber \\
&= \int_{\mathbb{K}^{2}} e^{-i (\langle 0,a\rangle + \langle t,b\rangle)} \Pi(a,b)\ dm_{2}(a,b) \nonumber \\
&= \#(0,t) \nonumber \\
&= e^{-itB} \nonumber \\
&= \left(\mathscr{F}E_{B}\right) (t).
\end{align}
The injectivity of the Fourier transformation $\mathscr{F}$ leads us to conclude that the marginal $\Pi_{B} = E_{B}$ is essentially the same object as the original spectral measure governing the probabilistic behaviour of the outcomes of $B$. By a parallel argument, one also finds that the marginal
\begin{equation}
\Pi_{A}(a) := \int_{\mathbb{K}} \Pi(a,b)\ dm(b)
\end{equation}
is nothing but $\Pi_{A} = E_{A}$. These properties are naturally found common in product spectral measures defined for strongly commuting pairs of self-adjoint operators, although each $\Pi(a,b)$ is not necessarily a projection, or may not be even positive. This tempts us to introduce the term \emph{quasi-joint-spectral distributions} of a pair of observables, which can be understood as a generalisation of the concept of spectral measures or POVMs.
\begin{definition*}[Quasi-joint-spectral Distribution of a Pair of Quantum Observables]
Let $A$ and $B$ be self-adjoint operators on $\mathcal{H}$. We call an element of $\mathfrak{M}_{A,B}$ a quasi-joint-spectral distribution of the pair of observables $A$ and $B$.
The cardinality of the collection $\mathfrak{M}_{A,B}$ is equal to unity if and only if $A$ and $B$ strongly commute with each other. Otherwise, it is always greater than unity.
\end{definition*}
\noindent
In the case where the observables $A$ and $B$ happen to strongly commute with each other, we specifically call the unique element of $\mathfrak{M}_{A,B}$ the \emph{joint-spectral distribution} of $A$ and $B$, which is nothing but the product spectral measure $E_{A,B}$ of the pair in standard terminology.
We note that the terminologies introduced above are non-standard, and are to be used only in this paper.
\subsubsection{Quasi-joint-probability Distributions}
Although the study on the precise definitions and properties of the family of quasi-joint-spectral distributions would be of mathematical interest in its own right, we shall refrain from going further due to the limited mathematical tools available. Instead, we turn to a more elementary object to ease our discussion.
Now, given a quasi-joint-spectral distribution $\Pi \in \mathfrak{M}_{A,B}$ of $A$ and $B$, we fix a specific quantum state $|\phi\rangle \in \mathcal{H}$, and consider a distribution of the form formally defined by
\begin{align}\label{def:QJP_explicit}
p(a,b)
&:= \frac{\langle \phi, \Pi(a,b) \phi \rangle}{\|\phi\|^{2}} \nonumber \\
&= \frac{\langle \phi, (\mathscr{F}^{-1}\#)(a,b) \phi \rangle}{\|\phi\|^{2}} \nonumber \\
&= \left( \mathscr{F}^{-1} \frac{\langle \phi, \#(\,\cdot\, , \,\cdot\,) \phi \rangle}{\|\phi\|^{2}} \right)(a,b),
\end{align}
where $\#$ is the hashed operator of which inverse Fourier transform $\Pi = \mathscr{F}^{-1}\#$ is the quasi-joint-spectral distribution under consideration.
Since the distribution $p$ is `scalar valued', it should be a much more feasible object to deal with than the `operator valued' distribution $\Pi$ introduced earlier.
We thus introduce the collection
\begin{equation}\label{def:QJP_FT}
\hat{\mathfrak{M}}_{A,B}^{\phi} := \left\{ \frac{\langle \phi, \#(s,t) \phi \rangle}{\|\phi\|^{2}} : \# \in \hat{\mathfrak{M}}_{A,B}, |\phi\rangle \in \mathcal{H} \right\}
\end{equation}
of all distributions generated by the hashed operators of the parametrised unitary operators give a fixed state, and in turn formally define
\begin{equation}\label{def:QJP}
\mathfrak{M}_{A,B}^{\phi} := \left\{\mathscr{F}^{-1} u : u \in \hat{\mathfrak{M}}_{A,B}^{\phi} \right\},
\end{equation}
by their inverse Fourier transforms. We thus summarise as:
\begin{definition*}[Quasi-joint-probability Distribution of a Pair of Quantum Observables]
Let $A$ and $B$ be self-adjoint operators on $\mathcal{H}$, and let $|\phi\rangle \in \mathcal{H}$. We call an element of $\mathfrak{M}_{A,B}^{\phi}$ a quasi-joint-probability (QJP) distribution of the pair of observables $A$ and $B$ on $|\phi\rangle$. The cardinality of the collection $\mathfrak{M}_{A,B}^{\phi}$ is equal to unity for every choice of the vector $|\phi\rangle$ if and only if the observables $A$ and $B$ strongly commute with each other. Otherwise, there exists a vector $|\phi\rangle$ such that the cardinality is greater than unity.
\end{definition*}
\noindent
In the case where the observables $A$ and $B$ happen to strongly commute with each other, we specifically call the unique element of $\mathfrak{M}_{A,B}^{\phi}$ the joint-probability distribution of $A$ and $B$ on $|\phi\rangle$, which is nothing but the probability measure $\mu_{A,B}^{\phi}$ of the pair introduced in \eqref{def:prob_measrue_A_simul}.
Given a hashed operator $\#$ of the parametrised unitary groups and a quantum state $|\phi\rangle \in \mathcal{H}$, we call an element $p \in \mathfrak{M}_{A,B}^{\phi}$ specified by
\begin{equation}
(\mathscr{F}p)(s,t) = \frac{\langle \phi, \#(s,t) \phi \rangle}{\|\phi\|^{2}},
\end{equation}
the QJP distribution generated by $\#$ and $|\phi\rangle$.
Our choice of the denomination of the elements of $p \in \mathfrak{M}_{A,B}^{\phi}$ is due to the fact that they retain similar properties to those of classical joint-probability distributions.
Indeed, the `total integration' reduces to
\begin{align}
\int_{\mathbb{K}^{2}} p(a,b)\ dm_{2}(a,b)
&= \int_{\mathbb{K}^{2}} e^{-i (\langle 0,a\rangle + \langle 0,b\rangle)} p(a,b)\ dm_{2}(a,b) \nonumber \\
&= \left( \mathscr{F}p \right)(0,0) \nonumber \\
&= \frac{\langle \phi, \#(0,0) \phi\rangle}{\|\phi\|^{2}} = 1,
\end{align}
where $\#$ is the hashed operator that, together with $|\phi\rangle$, generates $p$. As for the marginals, by introducing the marginal distribution formally defined by
\begin{equation}
p_{B}(b) := \int_{\mathbb{K}} p(a,b)\ dm(a),
\end{equation}
one observes through a formal computation that
\begin{align}
\left(\mathscr{F}p_{B}\right) (t)
&= \int_{\mathbb{K}} e^{-i \langle t,b\rangle} \left( \int_{\mathbb{K}} p(a,b)\ dm(a) \right) dm(b) \nonumber \\
&= \int_{\mathbb{K}^{2}} e^{-i (\langle 0,a\rangle + \langle t,b\rangle)} p(a,b)\ dm_{2}(a,b) \nonumber \\
&= \frac{\langle \phi, \#(0,t) \phi\rangle}{\|\phi\|^{2}} \nonumber \\
&= \frac{\langle \phi, e^{-itB} \phi\rangle}{\|\phi\|^{2}} \nonumber \\
&= \left(\mathscr{F}\mu_{B}^{\phi}\right) (t).
\end{align}
The injectivity of the Fourier transformation $\mathscr{F}$ leads us to conclude that the distribution $p_{B}(b)$ is essentially the same object as the probability measure $\mu_{B}^{\phi}$ describing the probabilistic behaviour of the outcomes of $B$. By a parallel argument, one also finds that the marginal
\begin{equation}
p_{A}(b) := \int_{\mathbb{K}} p(a,b)\ dm(b)
\end{equation}
is nothing but $p_{A} = \mu_{A}^{\phi}$.
Before we proceed further, we make notes on some mathematical intricacies involved in their definitions for the interested.
\paragraph{Mathematical Remarks}
One may notice some subtleties inherent to the definition of $\mathfrak{M}_{A,B}^{\phi}$. The first problem might be the domain of the definition of the inverse Fourier transformation: while the Fourier transform of a complex measure $\mu$ is a function, in regard that it does not necessarily lie in $\hat{\mu} \notin L^{1}(\mathbb{R}^{n})$, its inverse Fourier transform may not be well-defined, even in the case where $A$ and $B$ strongly commute with each other. This can be temporarily remedied by understanding the inverse Fourier transform of an element $u \in \hat{\mathfrak{M}}_{A,B}^{\phi}$ to be the unique complex measure $\mu$ such that $u = \hat{\mu}$ holds, which should be a reasonable treatment due to the injectivity of the Fourier transformation. This provides a sufficient cure in the case where the pair of observables strongly commutes.
On the other hand, another problem arises in the case in which the pair of self-adjoint operators fails to strongly commute: it might happen that, for some element $u \in \hat{\mathfrak{M}}_{A,B}^{\phi}$, there is no complex measure $\mu$ such that its Fourier transform coincides with $u = \hat{\mu}$.
A straightforward and more fundamental cure for this would be to expand our framework into that of generalised functions, specifically, by embedding the space of complex measures into that of tempered distributions.
Indeed, since the Fourier transformation is a bijection on the space of tempered distributions, by understanding that each of the elements of $\hat{\mathfrak{M}}_{A,B}^{\phi}$ to be a tempered distribution, its inverse Fourier transform itself always exists as a tempered distribution.
In consideration of this, since we do not wish to get involved with the theory of generalised functions, we shall be exclusively dealing with those elements $u \in \hat{\mathfrak{M}}_{A,B}^{\phi}$ for which there exists a complex measure $\mu$ satisfying $u = \hat{\mu}$, and understand the element $\mu := \mathscr{F}^{-1}u \in \mathfrak{M}_{A,B}^{\phi}$ to be the complex measure. To this end, we introduce:
\begin{definition*}[Representation by Quasi-probability Measures]
Under the above situation, let $p \in \mathfrak{M}_{A,B}^{\phi}$ be a QJP distribution of $A$ and $B$, and let $u \in \hat{\mathfrak{M}}_{A,B}^{\phi}$ be an element such that $p = \mathscr{F}^{-1}u$. We say that the QJP distribution $p$ admits representation by a quasi-probability measure, if there exists a quasi-probability measure $\mu$ on $\mathbb{K}^{2}$ such that $u = \hat{\mu}$ holds, and understand the QJP distribution $p = \mu$ to be the quasi-probability measure.
\end{definition*}
\noindent
A similar concern arises for the definition of quasi-joint-spectral distributions $\Pi = \mathscr{F}^{-1}\#$ defined as inverse Fourier transforms of hashed operators $\#$ of the unitary operators $e^{-isA}$ and $e^{-itB}$. Parallel to the `scalar valued' case seen above, quasi-joint-spectral distributions $\Pi$ are better understood as an object generalising the concept of spectral measures (or POVMs), in the sense that, while spectral measures $E$ (or POVMs) yield probability measures $\langle \phi, E(\,\cdot\,)\phi\rangle/\|\phi\|^{2}$ when combined with a vector $|\phi\rangle$, quasi-joint-spectral distributions $\Pi$ yield generalised functions, symbolically denoted by $\langle \phi, \Pi(a,b)\phi\rangle/\|\phi\|^{2}$. In this respect, quasi-joint-spectral distributions are to be understood as elements of the space of \emph{operator valued (tempered) distributions} (OVDs), which should serve as a generalisation to that of POVMs.
We also note that the methods introduced above in defining QJSDs admit a straightforward generalisation in defining them, not only for a \emph{pair} ($N=2$) of quantum observables as presented above, but also for \emph{arbitrary combinations} $(N \geq 2)$ of quantum observables, or even for \emph{arbitrary combinations of POVMs}. Also, one may readily generalise the discussion for defining QJP distributions, not just for pure states as presented above by sandwiching the QJSPs by kets and bras, but also for mixed states by taking the trace of the product of QJSPs and density operators.
\subsection{Complex-parametrised Sub-families}
Since we have decided to confine ourselves in the framework of complex measures rather than that of generalised functions due to our restricted mathematical tools available, we would mostly refrain from treating the general cases, and shall concentrate on a special sub-families of QJP distributions of a pair of quantum observables $A$ and $B$.
\subsubsection{Additive Sub-family}\label{sec:additive_subfamily}
As a simple example of QJP distributions admitting representation by quasi-probability measures, we observe:
\begin{lemma}
Let $A$ and $B$ be self-adjoint operators on $\mathcal{H}$, and consider the hashed operator of either of the form
\begin{equation}
\#(s,t) =
\begin{cases}
e^{-itA}e^{-isB} \\
e^{-isB}e^{-itA}
\end{cases}, \quad s,t \in \mathbb{R}.
\end{equation}
Then, the QJP distributions generated by $\#$ and any choice of the vector $|\phi\rangle \in \mathcal{H}$ admit representation by quasi-probability measures.
\end{lemma}
\begin{proof}
We provide the proof for the first case without loss of generality. Observe that the complex measure
\begin{equation}\label{def:c_qpm_1}
\Delta \mapsto \nu(\Delta,\Delta_{A}) := \frac{\langle \phi, E_{B}(\Delta)E_{A}(\Delta_{A})\phi \rangle}{\|\phi\|^{2}}, \quad \Delta \in \mathfrak{B}
\end{equation}
is absolutely continuous with respect to $\mu_{B}^{\phi}$ for all fixed $\Delta_{A} \in \mathfrak{B}$. This allows us to construct a transition quasi-probability kernel by taking the Radon-Nikod{\'y}m derivative of the above complex measure with respect to $\mu_{B}^{\phi}$. A direct application of Proposition~\ref{prop:ctk_to_cm} with $f(a,b) := e^{-i(as + bt)}$ then leads to the desired statement.
\end{proof}
\noindent
This inspires us to introduce the complex linear combinations of the above two distributions.
We hereby consider the hashed operators of the form
\begin{equation}\label{def:hash_add_cparam}
\#_{\mathrm{add}}^{\alpha}(s,t) := \frac{1 + \alpha}{2} e^{-itB}e^{-isA} + \frac{1 - \alpha}{2}e^{-isA}e^{-itB}, \quad s, t \in \mathbb{R},\ \alpha \in \mathbb{C},
\end{equation}
and observe that the QJP distributions induced by them naturally admit representation by quasi-probability measures.
\begin{corollary}
The QJP distributions generated by the hashed operators of the form \eqref{def:hash_add_cparam} and $|\phi\rangle \in \mathcal{H}$ admits representation by quasi-probability measures.
\end{corollary}
\noindent
In this paper, we call the above sub-family of QJP distributions the \emph{additive complex-parametrised sub-family} of QJP distributions of $A$ and $B$ on $|\phi\rangle$ (or simply, the additive sub-family, for short).
\subsubsection{Convolutive Sub-family}
One realises below that another class of QJP distributions parametrised by a complex number can be introduced.
We hereby consider the hashed operators of the form
\begin{equation}\label{def:hash_conv_cparam}
\#_{\mathrm{cnv}}^{\alpha}(s,t) := e^{-i\langle s, (1 -\alpha)/2 \rangle A}e^{-itB}e^{-i\langle s, (1 + \alpha)/2 \rangle A}, \quad s \in \mathbb{C}, t \in \mathbb{R},\ \alpha \in \mathbb{C}
\end{equation}
where $\langle s, \alpha \rangle := s_{1} \alpha_{1} + s_{2} \alpha_{2}$ denotes the inner product of
\begin{equation}\label{eq:complex_as_vector}
s = s_{1} + i s_{2}, \quad \alpha = \alpha_{1} + i\alpha_{2},
\end{equation}
each of them understood as real vectors of $\mathbb{R}^{2} \cong \mathbb{C}$, and introduce the \emph{convolutive complex-parametrised sub-family} of QJP distributions of $A$ and $B$ on $|\phi\rangle$ (or simply, the convolutive sub-family, for short) by those elements of $\mathfrak{M}_{A,B}^{\phi}$ that are generated by the hashed operators of the form \eqref{def:hash_conv_cparam} and $|\phi\rangle$.
\paragraph{Linear Transformation}
It is of natural interest to find out the condition as to when an element of the convolutive sub-family admits representation by quasi-probability measures. Obviously, the choice $\alpha = \pm 1$ admits it, since they are also members of the additive sub-family introduced earlier. As for the other choices of the complex parameter $\alpha \in \mathbb{C}$, we first introduce an auxiliary distribution defined by
\begin{equation}\label{def:FT_gen_dist}
\tilde{u}(s,t) := \frac{\langle \phi, e^{-is_{1} A}e^{-itB}e^{-i s_{2} A} \phi \rangle}{\|\phi\|^{2}}, \quad s \in \mathbb{C}, t \in \mathbb{R},
\end{equation}
where $s = s_{1} + is_{2} \in \mathbb{C}$, $s_{1}, s_{2} \in \mathbb{R}$ is defined as \eqref{eq:complex_as_vector}. Once there exists a quasi-probability measure $\tilde{\mu}$ such that its Fourier transform coincides with $\mathscr{F}\tilde{\mu} = \tilde{u}$, one finds below that every member of the convolutive sub-family is a linear transform of the quasi-probability measure $\tilde{\mu}$, hence themselves admit representation by quasi-probability measures. To see this, we first introduce the matrix
\begin{align}\label{eq:lin_trans_T}
T_{\alpha} := \left(
\begin{array}{cc}
(1 - \alpha_{1})/2 & (1 + \alpha_{1})/2 \\
-\alpha_{2}/2 & \alpha_{2}/2
\end{array}
\right),
\end{align}
defined for each complex number $\alpha = \alpha_{1} + i \alpha_{2} \in \mathbb{C}$, $\alpha_{1}, \alpha_{2} \in \mathbb{R}$. The Fourier transform $\mathscr{F} \tilde{\mu}_{(T_{\alpha} \times I)}$ of the linear transform of the quasi-probability measure $\tilde{\mu}$ with respect to the operator
\begin{equation}\label{def:two_lin_trans}
(T_{\alpha} \times I)(a,b) := (T_{\alpha}a, b), \quad a \in \mathbb{C}, b \in \mathbb{R},
\end{equation}
reads
\begin{align}
\left( \mathscr{F} \tilde{\mu}_{(T_{\alpha} \times I)} \right)(s,t)
&= \tilde{u}(T_{\alpha}^{*}s, t) \nonumber \\
&= \frac{\langle \phi, e^{-i\langle s, (1 -\alpha)/2 \rangle A}e^{-itB}e^{-i\langle s, (1+\alpha)/2 \rangle A} \phi \rangle}{\|\phi\|^{2}},
\end{align}
where we have used \eqref{eq:FT_and_Lin_Trans} in the first equality.
We thus have:
\begin{lemma}[Transformation between Parameters]\label{lem:Trans_b_Param}
Let $A$ and $B$ be self-adjoint operators on $\mathcal{H}$, and let $|\phi\rangle \in \mathcal{H}$.
\begin{enumerate}
\item If the inverse Fourier transform of the auxiliary distribution \eqref{def:FT_gen_dist} admits representation by a quasi-probability measure, then all the members of the convolutive sub-family admit representation by quasi-probability measures.
\item Under the above situation, let $T_{\alpha}$ be the linear transformation defined for every choice of the complex parameter $\alpha \in \mathbb{C}$ as in \eqref{eq:lin_trans_T}, and let $\tilde{u}$ be the auxiliary distribution \eqref{def:FT_gen_dist} and $\tilde{\mu}$ be the quasi-probability measure such that $\mathscr{F}\tilde{\mu} = \tilde{u}$. Then, every member of the convolutive sub-family can be described as the linear transform of $\tilde{\mu}$ as
\begin{equation}\label{eq:alpha_generation}
\mu^{\phi,\alpha}_{\mathrm{cnv}} = \tilde{\mu}_{(T_{\alpha} \times I)},
\end{equation}
where $\mu^{\phi,\alpha}_{\mathrm{cnv}}$ denotes the quasi-probability measure generated by the hashed operator of the form \eqref{def:hash_conv_cparam}, and $T_{\alpha} \times I$ is the operator defined as in \eqref{def:two_lin_trans}.
\end{enumerate}
\end{lemma}
\paragraph{Representation by Quasi-probability Measures}
Now, observing that the determinant of the linear transform $T_{\alpha}$ reads
\begin{equation}
\det T_{\alpha} = \mathrm{Im}\, \alpha/2,
\end{equation}
one finds that the transformation $\mu \mapsto \mu_{(T_{\alpha} \times I)}$ is invertible if and only if $\alpha \in \mathbb{C} \setminus \mathbb{R}$, for indeed $T_{\alpha} \in \mathrm{GL}(2;\mathbb{R}) \Leftrightarrow \alpha \in \mathbb{C}\setminus\mathbb{R}$. The product rule \eqref{eq:product_rule_image_measure_lin} then reveals that, one may move from one member of the convolutive sub-family to another by a sequential application of the transformations as
\begin{equation}\label{scheme:trans_param}
\mu^{\alpha} \xrightarrow{\quad T_{\alpha}^{-1} \times I \quad} \tilde{\mu} \xrightarrow{\quad T_{\alpha^{\prime}} \times I \quad} \mu^{\alpha^{\prime}}
\end{equation}
for the choice $\alpha \in \mathbb{C}\setminus\mathbb{R}$ and $\alpha^{\prime} \in \mathbb{C}$. Combining Lemma~\ref{lem:Trans_b_Param} with the above observation, one concludes:
\begin{corollary}[Representation by Quasi-probability Measures]\label{cor:rep_by_qpm_condition}
The following conditions are equivalent.
\begin{enumerate}
\item The inverse Fourier transform of the distribution $\tilde{u}$ defined in \eqref{def:FT_gen_dist} admits representation by quasi-probability measures.
\item A member of the convolutive sub-family for the choice of the parameter $\alpha \in \mathbb{C}\setminus\mathbb{R}$ admits representation by quasi-probability measures.
\item Every member of the convolutive sub-family admits representation by quasi-probability measures.
\end{enumerate}
\end{corollary}
\paragraph{Explicit Computation of the Members of the convolutive Sub-family}
We shall provide an explicit example of the case in which every member of the convolutive complex-parametrised sub-family admits representation by quasi-probability measures.
\begin{proposition}\label{prop:construction_of_qjp}
Let $A$ and $B$ self-adjoint, and suppose that $B$ has spectrum $\sigma(B)$ of finite cardinality and that it is non-degenerate
\begin{equation}
B = \sum_{b \in \sigma(B)} b \cdot |b \rangle\langle b|.
\end{equation}
For a quantum state $|\phi\rangle \in \mathcal{H}$ such that the probability of finding the outcomes of $B$ is non-vanishing $|\langle b, \phi \rangle|^{2} \neq 0$ for all its eigenvalues $b \in \sigma(B)$, every member of the convolutive sub-family of QJP distributions admits representation by quasi-probability measures.
\end{proposition}
\begin{proof}
Corollary~\ref{cor:rep_by_qpm_condition} purports that it suffices to construct the quasi-probability measure $\tilde{\mu}$ that satisfies $\mathscr{F}\tilde{\mu} = \tilde{u}$, where $\tilde{u}$ is the auxiliary distribution \eqref{def:FT_gen_dist}.
Now, under the above conditions, let $b \in \sigma(B)$ and $\Delta_{A} \in \mathfrak{B}^{1}$ be fixed, and introduce the Radon-Nikod{\'y}m derivative
\begin{align}
\nu_{b}^{\phantom{*}}(\Delta_{A})
&:= (d\nu(\,\cdot\,,\Delta_{A}) / d\mu_{B}^{\phi})(b) \nonumber \\
&= \frac{\langle b, E_{A}(\Delta_{A})\phi\rangle}{\langle b, \phi\rangle},
\end{align}
where the complex measure $\nu(\,\cdot\,,\Delta_{A})$ was defined in \eqref{def:c_qpm_1}. For every fixed $b \in \sigma(B)$, this defines a quasi-probability measure $\Delta_{A} \mapsto \nu_{b}^{\phantom{*}}(\Delta_{A})$, which is in fact nothing but a slight generalisation of the quasi-probability measure \eqref{def:aux_qpm_b_a} previously introduced in Section~\ref{sec:ps_II}. Defining the product complex measure
\begin{equation}
K(b,\,\cdot\,) := \nu(b,\,\cdot\,)^{*} \otimes \nu(b,\,\cdot\, ),
\end{equation}
on the product space $\mathbb{R}^{2} \cong \mathbb{C}$ for each $b \in \sigma(B)$, we intend to extend the domain of the variable $b$ to the whole real line to make a transition quasi-probability kernel from $(\mathbb{R},\mathfrak{B})$ into $(\mathbb{C},\mathfrak{B}(\mathbb{C}))$ by defining, for example,
\begin{equation}
\tilde{K}(b,\Delta_{A}) :=
\begin{cases}
K(b,\Delta_{A}), & (b \in \sigma(B)) \\
\delta_{0}(\Delta_{A}) , & (b \notin \sigma(B)),
\end{cases}
\end{equation}
where $\delta_{0}$ is the delta measure centred at the origin (for the extension into $\mathbb{R} \setminus \sigma(B)$, we could have assigned any quasi-probability measure so that the extension makes a transition quasi-probability kernel as a whole).
Letting $\tilde{\mu}$ denote the quasi-probability measure on the product space $\mathbb{C}\times\mathbb{R}$ defined by $\tilde{K}$ and $\mu_{B}^{\phi}$ by means of
\begin{equation}
\int_{\mathbb{C} \times \mathbb{R}} f(a,b)\ d\tilde{\mu}(a,b) = \int_{\mathbb{R}} \int_{\mathbb{C}} f(a,b) \tilde{K}(b,da)\ d\mu_{B}^{\phi}(b),
\end{equation}
(see \eqref{prop:ctk_to_cm}), we maintain that $(\mathscr{F}\tilde{\mu})(s,t) = \langle \phi, e^{-is_{1} A}e^{-itB}e^{-i s_{2} A} \phi \rangle / \|\phi\|^{2}$.
To see this, just let $f(a,b) := e^{-i\langle s,a\rangle}e^{-itb}$ above and compute
\begin{align}
\int_{\mathbb{R}} \int_{\mathbb{C}} e^{-i\langle s,a\rangle}e^{-itb} \tilde{K}(b,da)\ d\mu_{B}^{\phi}(b)
&= \sum_{b \in \sigma(B)} \int_{\mathbb{C}} e^{-i\langle s,a\rangle}e^{-itb} K(b,da)\cdot \frac{|\langle b, \phi \rangle|^{2}}{\|\phi\|^{2}} \nonumber \\
&= \sum_{b \in \sigma(B)} e^{-itb} \frac{\langle \phi, e^{-is_{1}A} b\rangle \langle b, e^{-is_{2}A}\phi\rangle}{\|\phi\|^{2}} \nonumber \\
&= \frac{\langle \phi, e^{-is_{1} A}e^{-itB}e^{-i s_{2} A} \phi \rangle}{\|\phi\|^{2}},
\end{align}
which was to be demonstrated. We have thus achieved a concrete construction of the quasi-probability measure, whose Fourier transform is the distribution $\tilde{u}$ defined in \eqref{def:FT_gen_dist}.
\end{proof}
\noindent
In passing, we note that, by comparing the transformation matrices \eqref{eq:lin_trans_T} and \eqref{eq:lin_trans_T_2i}, one finds that the quasi-probability measure $\mu_{A,B}^{\phi}$ obtained in the preceding Section~\ref{sec:ps_II} defined as in \eqref{def:quasi_joint_prob_meas} is nothing but the member of the convolutive complex-parametrised sub-family
\begin{equation}\label{eq:qjpm_2i_revealed}
\mu_{A,B}^{\phi} = \mu^{\phi,i}_{\mathrm{cnv}}
\end{equation}
for the purely imaginary choice $\alpha = i$ of the complex parameter.
\subsubsection{Qualification as Quasi-joint-probability Distributions}
Although we have provided a formal discussion to the problem, it yet remains to be confirmed by a rigorous treatment that every member of either the additive or the convolutive complex-parametrised sub-family of QJP distributions of a pair of quantum observables indeed qualifies as what its name indicates itself to be. Without loss of generality, we only provide the demonstration for the convolutive sub-family, since the proof for the additive subfamily is essentially the same.
\begin{proposition}[Qualification as Quasi-joint-probability Distributions]
Let $A$ and $B$ be self-adjoint, $|\phi\rangle \in \mathcal{H}$, and suppose that there exists a quasi-probability measure $\mu$ on the product space $\mathbb{C} \times \mathbb{R}$ such that
\begin{equation}
\left( \mathscr{F} \mu \right)(s,t) = \frac{\langle \phi, \#_{\mathrm{cnv}}^{\alpha}(s,t) \phi \rangle}{\|\phi\|^{2}}
\end{equation}
holds for some $\alpha \in \mathbb{C}$. Then, $\mu$ qualifies as a QJP distribution of $A$ and $B$ on $|\phi\rangle$, in the sense that \eqref{def:qjpm} holds.
\end{proposition}
\begin{proof}
We first observe a general result regarding marginals of complex measures and Fourier transformations. Let $\mu$ be a complex measure on the product space $\mathbb{R}^{m} \times \mathbb{R}^{n}$, and define the marginal of $\mu$ by
\begin{align}
\mu_{2} : \Delta \mapsto \mu_{2}(\Delta) := \mu(\mathbb{R}^{m} \times \Delta), \quad \Delta \in \mathfrak{B}^{n},
\end{align}
which is itself a complex measure on $(\mathbb{R}^{n},\mathfrak{B}^{n})$. One then observes
\begin{align}
\hat{\mu}_{2}(p)
&:= \int_{\mathbb{R}^{n}} e^{-i\langle p,y \rangle}\ d\mu_{2}(y) \nonumber \\
&= \int_{\mathbb{R}^{(m+n)}} e^{-i\langle 0,x \rangle} e^{-i\langle p,y \rangle}\ d\mu(x,y) \nonumber \\
&= \hat{\mu}(0,p),
\end{align}
where the second equality is due to the change of variables formula \eqref{eq:Transformationsformel_Bildmass} for the image measure $\mu_{2} = \pi_{2}(\mu)$, where $\pi_{2}(x,y) = y$, $x \in \mathbb{R}^{m}$, $y \in \mathbb{R}^{n}$ is the projection on the second variable.
Applying this fact to our situation as
\begin{equation}
\left( \mathscr{F} \mu \right)(0,t)
:= \frac{\langle \phi, e^{-itB} \phi \rangle}{\|\phi\|^{2}} = \left( \mathscr{F} \mu_{B}^{\phi} \right)(t),
\end{equation}
one readily finds
\begin{equation}
\mu(\mathbb{C} \times \Delta) = \mu_{B}^{\phi}(\Delta), \quad \Delta \in \mathfrak{B}^{1},
\end{equation}
by the injectivity of the Fourier transformation.
One may also demonstrate $\mu(\Delta \times \mathbb{R}) = \mu_{A}^{\phi}(\Delta)$, $\Delta \in \mathfrak{B}(\mathbb{K})$ by an analogous reasoning, which completes our proof.
\end{proof}
\subsubsection{Relation to other known Proposals}
We demonstrate below, in passing, that the complex-parametrised sub-families of the QJP distributions of a pair of quantum observables serve as generalisations to the other well known proposals of quasi-probability distributions.
\paragraph{Kirkwood-Dirac Distribution}
We first note that the Kirkwood-Dirac distribution, introduced in \eqref{def:Kirkwood-Dirac_Distr} in a formal manner, can be given a mathematically rigorous definition within our framework, and that it belongs to both the additive and convolutive sub-families of the QJP distributions for the choice $\alpha = 1$.
\begin{definition*}[Kirkwook-Dirac Quasi-joint-probability Distribution]
Let $A$ and $B$ be self-adjoint on $\mathcal{H}$, and let $|\phi\rangle \in \mathcal{H}$. We call the member of the additive/convolutive sub-family of the QJP distributions of the pair of observables $A$ and $B$ for the choice $\alpha = 1$, the Kirkwook-Dirac QJP distribution of the pair.
\end{definition*}
\noindent
To see how this definition can be justified, observe the following formal chain of expressions
\begin{align}
\int_{\mathbb{R}^{2}} e^{-i(as + bt)} K_{A,B}^{\phi}(a,b)\ dm_{2}(a,b)
&= \int_{\mathbb{R}^{2}} e^{-i(sa + tb)} \frac{\langle \phi, b\rangle\langle b, a \rangle\langle a, \phi \rangle}{\|\phi\|^{2}}\ dm_{2}(a,b) \nonumber \\
&= \frac{\langle \phi, e^{-itB}e^{-isA} \phi \rangle}{\|\phi\|^{2}},
\end{align}
where $K_{A,B}^{\phi}$ is the formal definition of the Kirkwood-Dirac distribution introduced in \eqref{def:Kirkwood-Dirac_Distr}. The injectivity of the Fourier transformation leads to the desired statement.
\paragraph{Wigner-Ville Distribution}
We next note that the Wigner-Ville distribution, introduced in \eqref{def:Wigner-Ville_QPD}, is also a special member of the convolutive sub-family of QJP distributions.
\begin{proposition}[Wigner-Ville Distribution]
Let $\{ L^{2}(\mathbb{R}), \mathscr{S}(\mathbb{R}), \{\hat{x}, \hat{p}\}\}$ denote the one-dimensional Schr{\"o}dinger representation of the CCR. Then, for the choice $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ of the wave-function, the member of the convolutive sub-family of the QJP distributions of the canonically conjugate pair $\hat{p}$ and $\hat{x}$ admits representation by quasi-probability measures for the choice $\alpha = 0$, which we denote by $\mu_{\mathrm{cnv}}^{\psi,0}$. The quasi-probability measure $\mu_{\mathrm{cnv}}^{\psi,0}$ is absolutely continuous, and its Radon-Nikod{\'y}m derivative with respect to the renormalised two-dimensional Lebesgue-Borel measure reads
\begin{equation}
\left( d\mu_{\mathrm{cnv}}^{\psi,0} / dm_{2} \right)(p,x) := W^{\psi}(x,p),
\end{equation}
where the r.~h.~s. is the Wigner-Ville distribution introduced in \eqref{def:Wigner-Ville_QPD}.
\end{proposition}
\begin{proof}
Observe that the condition $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$ guarantees the integrability $W^{\psi} \in L^{1}(\mathbb{R}^{2})$ of the WV distribution, based on which we compute
\begin{align}
&\int_{\mathbb{R}^{2}} e^{-i(sp + tx)} W^{\psi}(x,p)\ dm_{2}(x,p) \nonumber \\
&\qquad := \int_{\mathbb{R}^{2}} e^{-i(sp + tx)} \left( \int_{\mathbb{R}} \psi^{*}(x + y/2) \psi(x - y/2)e^{ipy}\ dm(y) \right)dm_{2}(x,p) \nonumber \\
&\qquad = \int_{\mathbb{R}} e^{-itx} \psi^{*}(x + s/2) \psi(x - s/2)\ dm(x) \nonumber \\
&\qquad = \int_{\mathbb{R}} e^{-itx} (e^{is\hat{p}/2}\psi)^{*}(x) (e^{-is\hat{p}/2}\psi)(x)\ dm(x) \nonumber \\
&\qquad = \langle e^{is\hat{p}/2}\psi, e^{-it\hat{x}} e^{-is\hat{p}/2}\psi\rangle \nonumber \\
&\qquad = \mathscr{F}\left( \mu_{\mathrm{cnv}}^{\psi,0} \right).
\end{align}
Combining \eqref{eq:FT_measure_vs_density} and the injectivity of the Fourier transformation, one arrives at the desired statement.
\end{proof}
\subsection{Some General Properties}
We next observe some general properties of QJP distributions. We first provide some discussion regarding the operation of taking the complex conjugate, and subsequently seek for the condition for their realness.
\subsubsection{Complex Conjugate}
We are interested in the complex conjugate of QJP distributions of a pair of observables $A$ and $B$ on $|\phi\rangle$. To this, let $\#$ be a hashed operator of the unitary operators $e^{-isA}$, $e^{-itB}$, and let $|\phi\rangle \in \mathcal{H}$ be such that the QJP distribution generated by them admits representation by a quasi-probability measure $\mu$. By applying \eqref{eq:FT_cc_Measure}, one readily finds that the Fourier transform of the complex conjugate $\mu^{*}$ reads
\begin{align}\label{eq:FT_cc}
(\mathscr{F}\mu^{*})(s,t)
&= (\mathscr{F}\mu)^{*}(-s,-t) \nonumber \\
&= \frac{\langle \#(-s,-t) \phi,\phi\rangle}{\|\phi\|^{2}} \nonumber \\
&= \frac{\langle \phi, \#(-s,-t)^{*} \phi\rangle}{\|\phi\|^{2}}, \quad s, t \in \mathbb{K}
\end{align}
where $\#(s,t)^{*}$ denotes the `adjoint' of the hashed operator. Since the `involution' $\#(-s,-t)^{*}$ is itself a hashed operator of $e^{-isA}$ and $e^{-itB}$, one concludes that the complex conjugate $\mu^{*}$ is again a QJP distribution of the pair of observables $A$ and $B$, and that it is precisely the distribution generated by the `involution' of the original hashed operator.
One also specifically finds that the sub-family of the QJP distributions $\mathfrak{M}_{A,B}^{\phi}$ that admit representations by quasi-probability measures is closed under the operation of taking the complex conjugate.
Parallel to this, by observing that the left most hand side of \eqref{eq:FT_cc} can be written as
\begin{align}
(\mathscr{F}\mu^{*})(s,t) = \frac{\langle \phi,(\mathscr{F}\Pi^{*})(s,t) \phi \rangle}{\|\phi\|^{2}}, \quad s, t \in \mathbb{K}
\end{align}
where $\Pi = \mathscr{F}^{-1}\#$ is the quasi-joint-spectral distribution, one concludes the validity of the equality
\begin{align}\label{eq:FT_cc_SpectDist}
(\mathscr{F}\Pi^{*})(s,t)
&= \#(-s,-t)^{*} \nonumber \\
&= (\mathscr{F}\Pi)(-s,-t)^{*} \nonumber \\
&=: (\mathscr{F}\Pi)^{\dagger}(s,t), \quad s, t \in \mathbb{K},
\end{align}
where $(\mathscr{F}\Pi)^{\dagger}$ denotes the `involution' (observe the analogy between \eqref{eq:FT_cc_Measure}).
This shows that the `adjoint' $\Pi^{*}$ of the quasi-joint-spectral distribution of $A$ and $B$ is again a quasi-joint-spectral distribution of the pair, and that it is precisely the inverse Fourier transform of the `involution' of the original hashed operator.
\paragraph{Complex-parametrised Sub-families}
Armed with our findings, one may explicit compute the complex conjugate of the elements of both the additive and the convolutive sub-families, and see that the sub-families are also closed under the operation of taking the complex conjugate. Indeed, if we respectively introduce
\begin{equation}
\Pi_{\mathrm{add}}^{\alpha} := \mathscr{F}^{-1} \#_{\mathrm{add}}^{\alpha}, \qquad \Pi_{\mathrm{cnv}}^{\alpha} := \mathscr{F}^{-1} \#_{\mathrm{cnv}}^{\alpha},
\end{equation}
for the members of the additive and convolutive sub-families, by observing that the `involution' of the hashed operators read
\begin{align}
\#_{\mathrm{add}}^{\alpha}(-s,-t)^{*} &= \#_{\mathrm{add}}^{-\alpha^{*}}(s,t), \\
\#_{\mathrm{cnv}}^{\alpha}(-s,-t)^{*} &= \#_{\mathrm{cnv}}^{-\alpha}(s,t),
\end{align}
one finds
\begin{align}
(\Pi_{\mathrm{add}}^{\alpha})^{*} &= \Pi_{\mathrm{add}}^{-\alpha^{*}}, \\
(\Pi_{\mathrm{cnv}}^{\alpha})^{*} &= \Pi_{\mathrm{cnv}}^{-\alpha}.
\end{align}
This provide explicit formulae for the computation of the complex conjugate of the members of the sub-families, and one specifically finds from it that both the sub-families are closed under the operation of taking the complex conjugate as promised.
Now, fixing $|\phi\rangle \in \mathcal{H}$ of one's choice, one observes:
\begin{lemma}[Complex Conjugate: Additive Sub-family]
Let $\mu_{\mathrm{add}}^{\alpha}$ denote the QJP distribution of $A$ and $B$ generated by $\#_{\mathrm{add}}^{\alpha}$ and $|\phi\rangle$. Then, its complex conjugate reads
\begin{equation}
\left(\mu_{\mathrm{add}}^{\phi,\alpha}\right)^{*} = \mu_{\mathrm{add}}^{\phi,-\alpha^{*}}.
\end{equation}
\end{lemma}
\begin{lemma}[Complex Conjugate: Convolutive Sub-family]
Suppose that the member of the convolutive sub-family admits representation by the quasi-probability measure $\mu_{\mathrm{cnv}}^{\phi,\alpha}$ for the choice $\alpha \in \mathbb{C}$. Then, the member for the choice $-\alpha \in \mathbb{C}$ also admits representation by quasi-probability measures, and the equality
\begin{equation}
\left(\mu_{\mathrm{cnv}}^{\phi,\alpha}\right)^{*} = \mu_{\mathrm{cnv}}^{\phi,-\alpha}.
\end{equation}
holds.
\end{lemma}
\subsubsection{Realness of the QJP Distributions}
One may naturally be interested in the condition as to when the quasi-joint-spectral distribution $\Pi = \mathscr{F}^{-1}\#$ becomes `self-adjoint' so that the resulting QJP distribution, symbolically denoted by $p(a,b) = \langle \phi, \Pi(a,b) \phi \rangle/\|\phi\|^{2}$, is also `real' for any choice of the vector $|\phi\rangle \in \mathcal{H}$. While the task of finding the explicit condition for which $\Pi(a,b)$ becomes `self-adjoint' seems at first non-trivial, the problem becomes significantly tractable if one considers its Fourier transform. Indeed, combining \eqref{eq:FT_cc_SpectDist} with the injectivity of the Fourier transform, one concludes that $\Pi = \Pi^{*}$ is `self-adjoint' if and only if its Fourier transform (namely, the hashed operator) $\# = \#^{\dagger}$ is a `self-involution'. Examples of such `self-involutive' hashed operators are provided by
\begin{align}
\#(s,t) =
\begin{cases}
e^{-itB/2}e^{-isA}e^{-itB/2}, \\
e^{-isA/2}e^{-itB}e^{-isA/2}, \\
\frac{1}{2}\left( e^{-itB}e^{-isA} + e^{-isA}e^{-itB} \right), \\
e^{-itB/L_{N}}e^{-isA/M_{N}} \cdots e^{-itB/L_{1}}e^{-isA/M_{1}}e^{-itB/L_{1}} \cdots e^{-isA/M_{N}}e^{-itB/L_{N}}, \\
e^{-i\overline{(sA + tB)}} = \lim_{N \to \infty} \left( e^{-isA/N} e^{-itB/N} \right)^{N}, \\
\textit{etc.},
\end{cases}
\end{align}
where $\sum_{k=1}^{N}M_{k}^{-1} = 1$, $\sum_{k=1}^{N}L_{k}^{-1} = 1$ in the third example. Colloquially speaking, hashed operators in which the disintegrated components of the unitary operators appear `symmetrically' provide straightforward examples.
As for our concrete examples, one finds:
\begin{corollary}[Condition for Realness]
A member of either the additive or convolutive sub-families of QJP distributions of $A$ and $B$ for the choice $\alpha = 0$ is always real.
\end{corollary}
\subsection{Conditioned Measurement Revisited}
We finally investigate how the CM scheme described in Section~\ref{sec:ps_II} fits into our general framework of quasi-joint-probabilities of quantum observables. What we see below is that the CM scheme is essentially \emph{a measurement scheme for measuring QJP distributions of an arbitrary pair of quantum observables}. As before, since the tools for the analysis of the most general cases are beyond the scope of this paper, we shall exclusively concentrate on the subfamily of quasi-joint-probabilities parametrised by a single complex number.
Without loss of generality, we only provide below a demonstration for the convolutive sub-family for simplicity.
\subsubsection{Short Introduction}
We now intend to construct a measurement scheme for obtaining the member of the convolutive sub-family of the QJP distributions for arbitrary choices of the parameter $\alpha \in \mathbb{C}$.
As for the problem, let us first recall that the quasi-probability measure \eqref{def:quasi_joint_prob_meas} obtained in Section~\ref{sec:ps_II} was nothing but the member for the choice of the parameter $\alpha = i$ (see \eqref{eq:qjpm_2i_revealed} for the discussion). In fact, as we have seen before, once we know the member of the subfamily for the parameter $\alpha \in \mathbb{C}\setminus\mathbb{R}$, we may compute all other members of the complex parameters by sequentially applying linear transformations as depicted in \eqref{scheme:trans_param}.
Hence, the knowledge of the distribution for the choice $\alpha = i$, obtained by means of the CM scheme in view of the WV distribution, actually suffices for our purpose.
Even so, one might be interested in how one could measure the QJP distribution for some specific parameter in a more direct manner. This should also provide a much more transparent view of the measurement scheme described in Section~\ref{sec:ps_II} from a more general viewpoint, which may be beneficial in its own right.
\paragraph{Model and Assumption}
Throughout this subsection, we let $A$ denote an observable on the target system $\mathcal{H}$, and assume that the meter system $\mathcal{K}$ is described by the one-dimensional Schr{\"o}dinger representation of the CCR $\{ L^{2}(\mathbb{R}), \mathscr{S}(\mathbb{R}), \{\hat{x}, \hat{p}\}\}$ for simplicity.
As usual, we prepare the two systems into their respective initial states $|\phi\rangle \in \mathcal{H}$, $|\psi\rangle \in \mathcal{K}$, and let them interact under the unitary operator $e^{-igA \otimes Y}$, $g \in \mathbb{R}$, for which we choose $Y=\hat{p}$ for definiteness, and let $|\Psi^{g}\rangle$ denote the state of the composite system after the interaction.
Since we intend to confine ourselves within the framework of complex measures, we place several conditions throughout this passage, so that, given a conditioning observable $B$ on the target system $\mathcal{H}$, all the members of the convolutive sub-family of the QJP distributions of $A$ and $B$ on $|\phi\rangle$ admits representation by quasi-probability measures.
\subsubsection{Conditioned Measurement Revisited}
In the previous section, the choice of the QJP distribution we intend to measure on the meter system was the WV distribution, which we found to be nothing but the member of our convolutive sub-family of the QJP distributions of the canonically conjugate pair of observables $A=\hat{p}$, $B=\hat{x}$ for the choice of the parameter $\alpha = 0$. The result was that, one could obtain the member of the convolutive sub-family of the QJP distributions for arbitrary pairs of quantum observables for the choice $\alpha = i$.
Motivated by this finding, it is then natural to conjecture that a different choice of the meter QJP distribution results in different choice of the target QJP distribution.
\paragraph{Meter QJP}
The starting point would be to find the equivalent object to $\omega^{\psi}$ for the other choices of the parameter $\alpha \in \mathbb{C}$. To this end, we first assume $\psi \in L^{1}(\mathbb{R}) \cap L^{2}(\mathbb{R})$, and introduce the function
\begin{equation}
\tilde{\omega}^{\psi,\alpha_{1}}(x,y) := \psi^{*}\left(x - y (1 - \alpha_{1})/2\right) \psi\left(x + y (1 + \alpha_{1})/2 \right)
\end{equation}
and also its Fourier transform
\begin{equation}
\tilde{W}^{\psi,\alpha_{1}}(x,p) := \int_{\mathbb{R}} e^{-ipy} \tilde{\omega}^{\psi,\alpha_{1}}(x,y)\ dm(y),
\end{equation}
where we let $\alpha = \alpha_{1} + i\alpha_{2}$, $\alpha_{1}, \alpha_{2} \in \mathbb{R}$.
Needless to say, the function $\tilde{\omega}^{\psi,0} = \tilde{\omega}^{\psi}$ for the choice $\alpha_{1} = 0$ reduces to the original function introduced in \eqref{def:func_V}, and thus $\tilde{W}^{\psi,0} = \tilde{W}^{\psi}$ is nothing but the (yet-to-be-normalised) WV distribution. By computing the Fourier transform
\begin{align}\label{eq:FT_W_w_alpha}
\left(\mathscr{F}\tilde{W}^{\psi,\alpha_{1}}\right)(q,y)
&= \int_{\mathbb{R}^{2}} e^{-i(qx + yp)}\left( \int_{\mathbb{R}} e^{-ipy} \tilde{\omega}^{\psi,\alpha_{1}}(x,y)\ dm(y) \right)dm_{2}(x,p) \nonumber \\
&= \int_{\mathbb{R}} e^{-iqx} \tilde{\omega}^{\psi,\alpha_{1}}(x,-y)\ dm(x) \nonumber \\
&= \int_{\mathbb{R}} e^{-iqx} \psi^{*}\left(x + y (1 - \alpha_{1})/2\right) \psi\left(x - y (1 + \alpha_{1})/2 \right)\ dm(x) \nonumber \\
&= \left\langle \psi, e^{-i\langle y, (1 -\alpha_{1})/2 \rangle \hat{p}} e^{-iq\hat{x}}e^{- i\langle y, (1 + \alpha_{1})/2 \rangle \hat{p}} \psi \right\rangle,
\end{align}
one concludes from the injectivity of the Fourier transformation that the normalisation
\begin{align}
W^{\psi,\alpha_{1}}
&:= \tilde{W}^{\psi,\alpha_{1}} / \|\psi\|^{2} \nonumber \\
&= \left( d\mu_{\mathrm{cnv}}^{\psi,\alpha_{1}} / dm_{2} \right)
\end{align}
is nothing but the Radon-Nikod{\'y}m derivative of the member of the convolutive sub-family of the QJP distributions of the canonically conjugate pair $A = \hat{p}$, $B = \hat{x}$ for the choice of the real parameter $\alpha_{1} \in \mathbb{R}$.
\paragraph{Rescaling}
For simplicity of the argument, we only treat the case for the choice $\alpha \in \mathbb{C} \setminus \mathbb{R}$, and for later convenience, we introduce the function
\begin{equation}
\tilde{\upsilon}^{\psi,\alpha}\left(x,y \right) :=
|2/\alpha_{2}|^{-1}\tilde{\omega}^{\psi,-\alpha_{1}}(x,(2/\alpha_{2})y), \quad (\alpha_{2} \neq 0)
\end{equation}
for a given choice of the parameter $\alpha \in \mathbb{C}\setminus\mathbb{R}$ (note the minus sign for the real part $\alpha_{1} := \mathrm{Re}\, \alpha$ in the definition).
Its Fourier transform then reads
\begin{align}\label{eq:FT_upsilon_alpha}
\int_{\mathbb{R}} e^{-iqx} \tilde{\upsilon}^{\psi,\alpha}(x,-y)\ dm(x)
&= \left\langle \psi, e^{-i\langle (2/\alpha_{2})y, (1 +\alpha_{1})/2 \rangle \hat{p}} e^{-iq\hat{x}}e^{- i\langle (2/\alpha_{2})y, (1 - \alpha_{1})/2 \rangle \hat{p}} \psi \right\rangle \nonumber \\
&= \left\langle \psi, e^{-i\langle y, (1 +\alpha_{1})/\alpha_{2} \rangle \hat{p}} e^{-iq\hat{x}}e^{- i\langle y, (1 - \alpha_{1})/\alpha_{2} \rangle \hat{p}} \psi \right\rangle,
\end{align}
where we have combined the second and the last equality of \eqref{eq:FT_W_w_alpha}, and applied the result \eqref{eq:Scaling_Fourier_translation}.
\paragraph{QJP of the `conditional' Meter State}
The next step is to compute the function $\tilde{\upsilon}^{\psi^{g}_{b},\alpha}$ for the `conditional' meter state $\psi^{g}_{b} := \psi^{g}_{B=b}$ introduced in \eqref{def:pure_state_given_b}. What we find below is that, parallel to the findings in Section~\ref{sec:ps_II}, the resulting function $\tilde{\upsilon}^{\psi^{g}_{b},\alpha}$ is provided by the convolution of the initial profiles of both the meter and the target configurations.
As above, we assume, for the ease of demonstration, that both the target and the conditioning observables $A$ and $B$ have spectra of finite cardinality, that $B$ is degenerate, and the probability of finding the outcomes of $B$ is non-vanishing $|\langle b, \phi \rangle|^{2} \neq 0$ for all its eigenvalues $b \in \sigma(B)$. Since the essence of the demonstration is substantially the same as those provided in Section~\ref{sec:ps_II}, we proceed by sketching the proofs.
In computing the function of our interest, we first compute its Fourier transform to observe
\begin{align}
&\int_{\mathbb{R}} e^{-iqx} \tilde{\upsilon}^{\psi^{g}_{b},\alpha}(x,-y)\ dm(x) \nonumber \\
&\quad = \left\langle \psi^{g}_{b}, e^{-i\langle y, (1 +\alpha_{1})/\alpha_{2} \rangle \hat{p}} e^{-iq\hat{x}}e^{- i\langle y, (1 - \alpha_{1})/\alpha_{2} \rangle \hat{p}} \psi^{g}_{b} \right\rangle \nonumber \\
&\quad = \int_{\mathbb{R}^{2}} \left\langle e^{-igs_{1}\hat{p}} \psi, e^{-i\langle y, (1 +\alpha_{1})/\alpha_{2} \rangle \hat{p}} e^{-iq\hat{x}}e^{- i\langle y, (1 - \alpha_{1})/\alpha_{2} \rangle \hat{p}} e^{-igs_{2}\hat{p}}\psi \right\rangle d(\nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}})(s_{1},s_{2}),
\end{align}
where we have used \eqref{eq:FT_upsilon_alpha} in the first equality, and where $\nu_{b}$ is the quasi-probability measure introduced in \eqref{def:aux_qpm_b_a}.
We next change variables of the above equality according to the linear transformation
\begin{align}
\left(
\begin{array}{l}
a_{1} \\
a_{2}
\end{array}
\right)
= T_{\alpha}
\left(
\begin{array}{l}
s_{1} \\
s_{2}
\end{array}
\right),
\qquad
T_{\alpha} := \left(
\begin{array}{cc}
(1 - \alpha_{1})/2 & (1 + \alpha_{1})/2 \\
-\alpha_{2}/2 & \alpha_{2}/2
\end{array}
\right)
\end{align}
by substituting
\begin{align}
s_{1} = a_{1} - \frac{1+\alpha_{1}}{\alpha_{2}}a_{2}, \quad s_{2} = a_{1} + \frac{1-\alpha_{1}}{\alpha_{2}}a_{2},
\end{align}
to find
\begin{align}
&\int_{\mathbb{R}} e^{-iqx} \tilde{\upsilon}^{\psi^{g}_{b},\alpha}(x,-y)\ dm(x) \nonumber \\
& \quad = \int_{\mathbb{R}^{2}} \left\langle \psi, e^{-i\langle y - ga_{2}, (1 +\alpha_{1})/\alpha_{2} \rangle \hat{p}} e^{-i(q\hat{x} - ga_{1}I)}e^{- i\langle y - ga_{2}, (1 - \alpha_{1})/\alpha_{2} \rangle \hat{p}} e^{-igs\hat{p}}\psi \right\rangle\ d\mu_{A}^{\phi,\alpha}(a|B=b) \nonumber \\
& \quad = \int_{\mathbb{R}^{2}} \left( \int_{\mathbb{R}} e^{-iqx} \tilde{\upsilon}^{\psi,\alpha}(x - ga_{1},-(y - ga_{2}))\ dm(x) \right) d\mu_{A}^{\phi,\alpha}(a|B=b) \nonumber \\
& \quad = \int_{\mathbb{R}} e^{-iqx} \left(\int_{\mathbb{R}^{2}} \tilde{\upsilon}^{\psi,\alpha}(x - ga_{1},-(y - ga_{2}))\ d\mu_{A}^{\phi,\alpha}(a|B=b) \right)dm(x),
\end{align}
where $\mu_{A}^{\phi,\alpha}(\Delta|B=b) := (\nu_{b}^{*} \otimes \nu_{b}^{\phantom{*}})(T_{\alpha}^{-1}\Delta)$, $\Delta \in \mathfrak{B}^{2}$ is the image measure, and we have combined \eqref{eq:FT_upsilon_alpha} with \eqref{eq:Fourier_translation} to obtain the second equality.
One thus concludes from the injectivity of the Fourier transformation that
\begin{equation}
\tilde{\upsilon}^{\psi^{g}_{b},\alpha}(x,y) = \int_{\mathbb{R}^{2}} \tilde{\upsilon}^{\psi,\alpha}(x - ga_{1},y - ga_{2})\ d\mu_{A}^{\phi,\alpha}(a|B=b),
\end{equation}
as promised.
\paragraph{Recovery of the Target QJP}
As for the recovery of the target information $\Delta \mapsto \mu_{A}^{\phi,\alpha}(\Delta|B=b)$, $\Delta \in \mathfrak{B}(\mathbb{C})$, one may resort to the familiar techniques we have discussed so far in depth, namely, one may recover the full profile by either probing the strong or the weak region of the interaction parameter. Once we obtained $\mu_{A}^{\phi,\alpha}(\,\cdot\,|B=b)$ for all $b \in \sigma(B)$, one may extend the domain of $b \in \sigma(B)$ to the whole real line $\mathbb{R}$ in a consistent manner, making it a transition quasi-probability kernel. This allows us to construct the QJP $\mu_{A,B}^{\phi,\alpha}$ of the pair of $A$ and $B$ in a manner described in Proposition~\ref{prop:ctk_to_cm} that satisfies
\begin{equation}
\mu_{A,B}^{\phi,\alpha}(\Delta_{A} \times \Delta_{B}) = \int_{\Delta_{B}} \mu_{A}^{\phi,\alpha}(\Delta_{A} | B=b)\ d\mu_{B}^{\phi}(b), \quad \Delta_{A} \in \mathfrak{B}(\mathbb{C}),\ \Delta_{B} \in \mathfrak{B}^{1}.
\end{equation}
A close look on the proof of Proposition~\ref{prop:construction_of_qjp} leads one to conclude that the QJP obtained here
\begin{equation}
\mu_{A,B}^{\phi,\alpha} = \mu_{\mathrm{cnv}}^{\phi,\alpha},
\end{equation}
is in fact nothing but the member of the convolutive sub-family for the choice $\alpha \in \mathbb{C} \setminus \mathbb{R}$, and that $\mu_{A}^{\phi,\alpha}(\,\cdot\,|B=b)$ is the conditional quasi-probability distribution of $A$ given $B=b$.
\newpage
\section{Application: Interpretation of Aharonov's Weak Value}\label{sec:app}
As an application of the findings on the QJP distributions of quantum observables, we now focus on the geometric structure that the QJP distributions induce in the space of quantum observables. Specifically, by drawing an analogy between the result of classical probability theory, we provide a geometric and statistical interpretation of Aharonov's weak value as `orthogonal projection' and `conditional average', respectively.
\subsection{Reference Materials}\label{sec:app_ref}
As usual, we start by preparing some necessary materials that become useful for our analysis. The main objective of this subsection is to obtain a geometric understanding of conditional expectations in classical probability theory.
\subsubsection{$L^{2}$-Theory of Conditional Expectations}
\paragraph{$L^{p}$-spaces for finite Measures}
Let $\mu$ be a finite measure on a measurable space $(X,\mathfrak{A})$, {\it i.e.}, $\mu(X) < \infty$, and let $1 \leq p < q \leq \infty$. By defining $r > 0$ satisfying $\frac{1}{r} = \frac{1}{p} - \frac{1}{q} $, a direct application of H{\"o}lder's inequality yields
\begin{equation}
\|f\|_{p} \leq \|f\|_{q} \cdot \|1\|_{r} = \|f\|_{q} \cdot |\mu(X)|^{1/r} < \infty
\end{equation}
for $f \in L^{q}(\mu)$. The following Lemma is worth of special notice.
\begin{lemma}
Let $\mu$ be a finite measure on a measurable space $(X,\mathfrak{A})$. Then, for any $1 \leq p \leq q \leq \infty$, the relation
\begin{equation}
L^{q}(\mu) \subset L^{p}(\mu)
\end{equation}
holds.
\end{lemma}
\noindent
Specifically, for probability spaces, note that one has the evaluation $\|f\|_{p} \leq \|f\|_{q}$ for the choice of the parameters $1 \leq p \leq q \leq \infty$.
\paragraph{Conditional Expectations for square-integrable Functions}
Now, consider a probability space $(\mathbb{R}^{n}, \mathfrak{B}^{n}, \mu)$, and let $\mathfrak{A} \subset \mathfrak{B}^{n}$ be a sub-$\sigma$-algebra. Since every square-integrable function $f \in L^{2}(\mu) \subset L^{1}(\mu)$ is integrable due to the above Lemma, its conditional expectation $\mathbb{E}[f|\mathfrak{A}] := d (f \odot \mu)|_{\mathfrak{A}} / d \mu|_{\mathfrak{A}} \in L^{1}(\mu|_\mathfrak{A})$ is well-defined, where $(f \odot \mu)|_{\mathfrak{A}}$ and $\mu|_\mathfrak{A}$ denotes the restriction of the respective (complex) measures on the sub-$\sigma$-algebra. Now, observe that, for any square-integrable function $g \in L^{2}(\mu|_{\mathfrak{A}})$, the equality
\begin{align}\label{eq:cond_exp_orth_proj_adj}
\langle g, f \rangle
&:= \int_{\mathbb{R}^{n}} g^{*}(x) f(x)\ d\mu(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} g^{*}(x)\ d(f \odot \mu)(x) \nonumber \\
&= \int_{\mathbb{R}^{n}} g^{*}(x) \frac{d (f \odot \mu)|_{\mathfrak{A}}}{d \mu|_{\mathfrak{A}}}(x)\ d \mu|_{\mathfrak{A}}(x) \nonumber \\
&=: \langle g, \mathbb{E}[f|\mathfrak{A}] \rangle
\end{align}
holds by the definition of the Radon-Nikod{\'y}m derivative. Specifically, note that this leads to the fact that the conditional expectation $\mathbb{E}[f|\mathfrak{A}] \in L^{2}(\mu|_\mathfrak{A})$ of a square-integrable function $f \in L^{2}(\mu)$ is again square-integrable.
\paragraph{Conditioning as Projection}
Another important observation to make from the above equality is that, the act of conditioning
\begin{equation}
\mathbb{E}[\,\cdot\,|\mathfrak{A}] : L^{2}(\mu) \to L^{2}(\mu|_{\mathfrak{A}}), \quad f \mapsto \mathbb{E}[f|\mathfrak{A}]
\end{equation}
that takes a $\mu$-square-integrable function to its conditional expectation, is an \emph{orthogonal projection}. To see this, first observe that linearity $\mathbb{E}[af + bg| \mathfrak{A}] = a\mathbb{E}[f| \mathfrak{A}] + b\mathbb{E}[g| \mathfrak{A}]$, $f, g \in L^{2}(\mu)$, $a, b \in \mathbb{C}$ follows immediately by definition (naturally, equality is only valid $\mu|_{\mathfrak{A}}$-almost everywhere). Now, since $L^{2}(\mu|_{\mathfrak{A}})$ is itself a complex Hilbert space, it is a topologically closed subspace of the larger complex Hilbert space $L^{2}(\mu)$. By recalling that there is a one-to-one correspondence between closed subspaces and orthogonal projections in Hilbert spaces, let $P(\,\cdot\,|\mathfrak{A}) : L^{2}(\mu) \to L^{2}(\mu|_\mathfrak{A})$ denote the unique orthogonal projection associated with it. By observing that
\begin{equation}
\langle g, f \rangle = \langle g, P(f |\mathfrak{A}) \rangle
\end{equation}
holds for all $g \in L^{2}(\mu|_\mathfrak{A})$ and $f \in L^{2}(\mu)$ by definition of orthogonal projections, one realises that the equality \eqref{eq:cond_exp_orth_proj_adj} combined with the non-degenerateness of inner products leads to
\begin{equation}\label{eq:cond_exp_orth_proj}
P(f |\mathfrak{A}) = \mathbb{E}[f|\mathfrak{A}], \quad f \in L^{2}(\mu).
\end{equation}
We summarise the results as follows.
\begin{proposition}(Orthogonal Projection and Conditional Expectation)
Let $\mu$ be a probability measure on $(\mathbb{R}^{n},\mathfrak{B}^{n})$, and let $\mathfrak{A} \subset \mathfrak{B}^{n}$ be a sub-$\sigma$-algebra. Then, the unique orthogonal projection $P(\,\cdot\,|\mathfrak{A}) : L^{2}(\mu) \to L^{2}(\mu|_\mathfrak{A})$ associated with the subspace $L^{2}(\mu|_\mathfrak{A})$ is provided by the conditional expectation
\begin{equation}
P(f|\mathfrak{A}) = \mathbb{E}[f|\mathfrak{A}],
\end{equation}
where $f \in L^{2}(\mu)$.
\end{proposition}
\noindent
We here see how the geometric concept of orthogonality relates to the statistical concept of conditioning in $L^{2}$-spaces.
\subsubsection{Conditioning as Optimal Approximation}
The geometric property mentioned above leads to several important interpretation of conditional expectations. One of the prominent characteristics of orthogonal projections is the validity of the \emph{Pythagorean identity}
\begin{equation}\label{eq:pythagorean_identity_classical}
\|f - g\|_{2}^{2} = \|f - \mathbb{E}[f|\mathfrak{A}]\|_{2}^{2} + \|\mathbb{E}[f|\mathfrak{A}] - g\|_{2}^{2}, \quad f \in L^{2}(\mu),\ g \in L^{2}(\mu|_{\mathfrak{A}}),
\end{equation}
where $\|\cdot\|_{2}$ denotes the standard $L^{2}$-norm introduced earlier in \eqref{def:Lp_norm}.
An immediate consequence of the above Pythagorean identity is the following equality
\begin{equation}
\|f - \mathbb{E}[f|\mathfrak{A}]\|_{2} = \min_{g \in L^{2}(\mu|_{\mathfrak{A}})} \|f - g\|_{2},\quad f \in L^{2}(\mu),
\end{equation}
which states that the optimal $\mu|_{\mathfrak{A}}$-square-integrable function one can find in approximating a function $f$ is explicitly provided by the conditional expectation of $f$ given $\mathfrak{A}$, and the positive-definiteness of the $L^{2}$-norm shows that the optimum is unique $\mu|_{\mathfrak{A}}$-a.e.
\subsection{Statistical Interpretation of Geometric Structures}
Now that we have reviewed the geometric interpretation of conditional expectations in classical probability theory, we shall begin our main analysis.
\subsubsection{Preliminary Observation}
In classical probability theory, probability measures equip the space of square-integrable functions with a geometry, {\it i.e.,} an inner product defined by \eqref{def:L2-inner_product}, which we reiterate for the readers' convenience as
\begin{equation}\label{def:classical_geometry}
\langle g, f \rangle_{\mu} := \int g^{*}f\ d\mu
\end{equation}
(here, we have also explicitly written the probability measure $\mu$ under consideration for clarity).
In the context of classical physics in which observables are represented by functions, a probability measure defines quantities on a given pair of square-integrable classical observables interpreted as \emph{correlations} or \emph{covariances} between them.
\paragraph{`Correlations' in Quantum Theory}
In quantum mechanics, observables are represented by self-adjoint operators on Hilbert spaces, and the statistics of the system are in turn represented by vectors of Hilbert spaces. In order to see how the two distinct frameworks of classical and quantum theory on correlations play together, first let $A$ and $B$ be a pair of simultaneously measurable bounded quantum observables on a Hilbert space $\mathcal{H}$, $|\psi\rangle \in \mathcal{H}$ a vector, and introduce
\begin{equation}\label{def:quantum_cor_sim}
\langle B, A \rangle_{\psi} := \frac{\langle B\psi, A\psi \rangle}{\|\psi\|^{2}}.
\end{equation}
One readily sees that, for the present case, the above geometry induced by the vector $|\psi\rangle$ is in accordance with the classical theory. Indeed, the unique product spectral measure $E_{A,B}$ of $A$ and $B$ admits a unique representation of the observables given by
\begin{equation}
A = \int_{\mathbb{R}^{2}} A(a,b)\ dE_{A,B}(a,b), \quad B = \int_{\mathbb{R}^{2}} B(a,b)\ dE_{A,B}(a,b)
\end{equation}
with $A(a,b) = a$, $B(a,b) = b$, and moreover defines a joint-probability measure $\mu_{A,B}^{\psi}$ of the pair (see \eqref{def:prob_measrue_A_simul}) on the state $|\psi\rangle \in \mathcal{H}$. It is then straightforward to see the validity of the equality
\begin{align}\label{eq:quantum_cor_sim_stat}
\langle B, A \rangle_{\psi}
&= \int_{\mathbb{R}^{2}} B^{*}(a,b)A(a,b)\ d\mu_{A,B}^{\psi}(a,b) \nonumber \\
&= \langle B, A \rangle_{\mu_{A,B}^{\psi}},
\end{align}
based on which one obtains a statistical interpretation of the geometry \eqref{def:quantum_cor_sim} as the correlation or covariance between the observables in the classical sense.
\paragraph{Non-commutative Case}
On the other hand, the problem is not so straightforward for the case where the pair $A$ and $B$ does not admit simultaneous measurability. This is essentially to do with the lack of the unique product spectral measure of the pair. As we have seen in the previous Section~\ref{sec:qp_qo}, the non-commutative analogues of product spectral measures are the quasi-joint-spectral distributions (QJSDs) defined as the inverse Fourier transforms of the hashed unitary groups \eqref{def:QJSD}. The non-uniqueness of the QJSDs for the non-commuting case generally leads to the non-uniqueness of the representation of operators and vectors by functions and quasi-probability distributions. Specifically, given a choice of a QJSD $\Pi_{A,B}$ for a pair of generally non-commuting observables $A$ and $B$, one obtains a functional representation of operators as
\begin{equation}
A = \int_{\mathbb{R}^{2}} A(a,b)\ d\Pi_{A,B}(a,b), \quad B = \int_{\mathbb{R}^{2}} B(a,b)\ d\Pi_{A,B}(a,b)
\end{equation}
with $A(a,b) = a$, $B(a,b) = b$ (fortunately, as for this specific case, all representations coincide irrespective of the choice of the QJSD), and also a representation of quantum states $|\psi\rangle \in \mathcal{H}$ by QJP distributions $p_{A,B}^{\psi}(a,b)$ defined as in \eqref{def:QJP_explicit}.
Guided by a straightforward analogy, one realises that the quantity formally defined by
\begin{equation}\label{def:quasi_covariance_general}
\llangle B, A \rrangle_{\psi} := \int_{\mathbb{K}^{2}} B^{*}(a,b)A(a,b)\ p_{A,B}^{\psi}(a,b)\ dm_{2}(a,b)
\end{equation}
defines various different geometries between quantum observables dependent on the choice of the QJSDs. This implies that, parallel to the classical case, QJSDs serves as a bridge that offers a `statistical' interpretation of the (non-unique) geometric structures that can be introduced in the space of quantum observables.
\subsubsection{Geometry induced by a specific Sub-family of QJSD}
The general treatment involving the entire class of QJSDs makes extensive use of the theory of generalised functions and its operator valued analogue (operator valued distributions), which may be far beyond the scope of this paper. In this paper, mainly in order to confine our argument in the theory of complex measures and its operator valued analogue, we concentrate on the specific choice of the QJSD of a pair of quantum observables, namely, to the additive sub-family (introduced in Section~\ref{sec:additive_subfamily}) for the choice $-1 \leq \alpha \leq 1$ of the complex parameter, hopefully without essential loss of generality. We shall moreover confine ourselves to bounded operators for simplicity, but the general treatment including unbounded operators is also possible without any essential alteration.
\paragraph{Sesquilinear Forms}
We are interested in introducing geometries in the space $L(\mathcal{H})$ of all bounded linear operators on a Hilbert space $\mathcal{H}$ given a fixed state $|\psi\rangle \in \mathcal{H}$.
To this end, let $\alpha \in \mathbb{C}$ be a complex number, and define
\begin{equation}\label{def:alpha_sesquilinear_form}
\llangle Y, X \rrangle_{\psi,\alpha} := \frac{1 + \alpha}{2} \cdot \frac{\langle Y \psi, X \psi \rangle}{\|\psi\|^{2}} + \frac{1 - \alpha}{2} \cdot \frac{\langle X^{*} \psi, Y^{*} \psi \rangle}{\|\psi\|^{2}}, \quad X, Y \in L(H).
\end{equation}
As described earlier, this is just one possible straightforward way to extend the geometry \eqref{def:quantum_cor_sim} so that it can be defined even for non-commuting observables. One readily sees that this satisfies
\begin{enumerate}
\item $\llangle Y + Y^{\prime}, X + X^{\prime} \rrangle_{\psi,\alpha} = \llangle Y, X \rrangle_{\psi,\alpha} + \llangle Y, X^{\prime} \rrangle_{\psi,\alpha} + \llangle Y^{\prime}, X \rrangle_{\psi,\alpha} + \llangle Y^{\prime}, X^{\prime} \rrangle_{\psi,\alpha}$,
\item $\llangle bY, aX \rrangle_{\psi,\alpha} = \overline{b}a\llangle Y, X \rrangle_{\psi,\alpha}$
\end{enumerate}
for any $X, X^{\prime}, Y, Y^{\prime} \in L(\mathcal{H})$ and $a, b \in \mathbb{C}$, hence it defines a \emph{sesquilinear form} on $L(\mathcal{H})$. By definition, one has
\begin{equation}
\llangle Y^{*}, X^{*} \rrangle_{\psi,\alpha} = \llangle X, Y \rrangle_{\psi,-\alpha}.
\end{equation}
By observing moreover that
\begin{equation}
\llangle Y, X \rrangle_{\psi,\alpha}^{*} = \llangle X, Y \rrangle_{\psi,\alpha^{*}}, \quad X, Y \in L(\mathcal{H})
\end{equation}
holds, the sesquilinear form \eqref{def:alpha_sesquilinear_form} is \emph{symmetric} (Hermitian) if the given parameter $\alpha \in \mathbb{R}$ is real. If one takes $X =Y$, this reads
\begin{equation}\label{def:eval}
\llangle X, X \rrangle_{\psi, \alpha} = \frac{1 + \alpha}{2} \cdot \frac{\|X\psi\|^{2}}{\|\psi\|^{2}} + \frac{1 - \alpha}{2} \cdot \frac{\|X^{*}\psi\|^{2}}{\|\psi\|^{2}}, \quad X \in L(\mathcal{H}).
\end{equation}
Specifically for the choice $-1 \leq \alpha \leq 1$ of the parameter, note that the above quantity happens to be always \emph{positive}, hence becomes a \emph{semi-norm}, for which we introduce the notation
\begin{equation}
\|X\|_{\psi,\alpha} := \llangle X, X \rrangle_{\psi, \alpha}^{\frac{1}{2}}, \quad X \in L(\mathcal{H}),\ -1 \leq \alpha \leq 1.
\end{equation}
Note that
\begin{equation}
\|X\|_{\psi,\alpha}^{2} \leq \frac{1 + \alpha}{2} \cdot \|X\|^{2} + \frac{1 - \alpha}{2} \cdot \|X^{*}\|^{2} = \|X\|^{2},
\end{equation}
where $\|X\|$ denotes the usual operator norm of $X \in L(\mathcal{H})$, which shows that the semi-norm $\|\cdot\|_{\psi,\alpha}$ induces a topology on $L(\mathcal{H})$ coarser than that induced by the usual operator norm $\|\cdot\|$.
\paragraph{Statistical Interpretation}
In classical theory, the natural geometry \eqref{def:classical_geometry} introduced on the space of observables admits statistical interpretation as correlations by means of probability measures. Parallel to this, we next intend to provide a statistical representation of the sesquilinear form \eqref{def:alpha_sesquilinear_form} for the quantum case, and this is achieved by means of QJP distributions. As mentioned earlier, we let $\mu_{\mathrm{add}}^{\psi,\alpha}$ denote a quasi-probability measure satisfying
\begin{equation}\label{def:sec_app_additive}
(\mathscr{F}\mu_{\mathrm{add}}^{\psi,\alpha})(s,t) = \frac{1 + \alpha}{2} \cdot \frac{\langle \psi, e^{-itB}e^{-isA}\psi\rangle}{\|\psi\|^{2}} + \frac{1 - \alpha}{2} \cdot \frac{\langle \psi, e^{-isA}e^{-itB}\psi\rangle}{\|\psi\|^{2}}, \quad \alpha \in \mathbb{C},
\end{equation}
which are namely members of the additive complex-parametrised sub-family of QJP distributions of $A$ and $B$ introduced in Section~\ref{sec:additive_subfamily}.
One readily sees from a direct application of Lemma~\ref{lem:FT_CM_diff_exp} that the integration of polynomial functions reduces to
\begin{align}
\int_{\mathbb{R}^{2}} b^{m}a^{n}\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b)
&= \left. (i\partial_{t})^{m} (i\partial_{s})^{n} (\mathscr{F}\mu_{\mathrm{add}}^{\psi,\alpha})(s,t) \right|_{(s,t) = (0,0)} \nonumber \\
&= \llangle B^{m},A^{n} \rrangle_{\psi,\alpha}.
\end{align}
One may also readily obtain a generalisation of this observation to continuous functions.
Indeed, according to the Stone-Weierstra{\ss} approximation theorem, since continuous functions $f, g$ defined on a compact space admit uniform approximations by polynomial functions as
\begin{align}
f(a) = \sum_{n=0}^{\infty} f_{n}a^{n}, \quad g(b) = \sum_{n=0}^{\infty} g_{n}b^{n},
\end{align}
one has
\begin{align}
\int_{\mathbb{R}^{2}} g^{*}(b)f(a)\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b)
&= \sum_{n,m=0}^{\infty} g_{m}^{*}f_{n}\int_{\mathbb{R}^{2}} b^{m}a^{n}\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b) \nonumber \\
&= \sum_{n,m=0}^{\infty} g_{m}^{*}f_{n} \llangle B^{m},A^{n} \rrangle_{\psi,\alpha} \nonumber \\
&= \left\llangle \sum_{m}^{\infty}g_{m}B^{m}, \sum_{n}^{\infty}f_{n}A^{n} \right\rrangle_{\psi,\alpha} \nonumber \\
&= \llangle g(B),f(A) \rrangle_{\psi,\alpha}.
\end{align}
This observation can be summarised as:
\begin{lemma}[Statistical Representation of Sesquilinear Forms]\label{lem:stat_rep_sesq_form}
Let $\mu_{\mathrm{add}}^{\psi,\alpha}$ be a member of the additive complex-parametrised sub-family of QJP distributions of the pair of observables $A, B \in L(\mathcal{H})$ defined as in \eqref{def:sec_app_additive}. Then, for any continuous functions $f$ and $g$ defined on the real line $\mathbb{R}$, the equality
\begin{equation}
\llangle g(B), f(A) \rrangle_{\psi,\alpha} = \int_{\mathbb{R}^{2}} g^{*}(b)f(a)\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b)
\end{equation}
holds.
\end{lemma}
\noindent
We have thus obtained a convenient representation of the sesquilinear form by integration with respect to QJP distributions, which offers a `statistical' interpretation to the geometry as `correlations' of a pair of generally non-commuting quantum observables.
\paragraph{Topic: Quasi-covariances (Quantum Covariances)}
As a natural extension to the classical notion of covariances, we may introduce the term `quasi-covariance' (or `quantum-covariance') of a pair of quantum observables under a given QJP distribution $p_{A,B}^{\psi}(a,b)$ for the quantity formally defined by
\begin{align}\label{def:quasi_cov}
\int_{\mathbb{K}^{2}} (a - \mathbb{E}[A;\psi])(b - \mathbb{E}[B;\psi]) p_{A,B}^{\psi}(a,b)\ dm_{2}(a,b),
\end{align}
whenever the integration exists.
Specifically, from an immediate application of the above Lemma, one may readily compute the quantum-covariances with respect to the additive subgroup $\mu_{\mathrm{add}}^{\psi,\alpha}(a,b)$ of the QJP distributions as
\begin{align}\label{eq:quantum_cov_alpha}
\mathbb{CV}^{\alpha}[A,B;\psi]
&:= \int_{\mathbb{R}^{2}} (a - \mathbb{E}[A;\psi])(b - \mathbb{E}[B;\psi])\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b) \nonumber \\
&= \llangle A - \mathbb{E}[A;\psi], B - \mathbb{E}[B;\psi] \rrangle_{\psi,\alpha} \nonumber \\
&= \mathbb{CV}_{\mathrm{S}}[A,B;\psi] + \alpha i \,\mathbb{CV}_{\mathrm{A}}[A,B;\psi],
\end{align}
where $\mathbb{CV}_{\mathrm{S}}[A,B;\psi]$ and $\mathbb{CV}_{\mathrm{A}}[A,B;\psi]$ are respectively the symmetric and anti-symmetric quantum covariances introduced in \eqref{def:q_covariance}.
\subsubsection{The Hilbert Space of Bounded Operators given a fixed State}
In classical theory, observables were described by functions, whereas in quantum theory, observables become self-adjoint operators. In order to conduct an analogue of the $L^{2}$-theory for quantum observables, it reveals for our purpose that it is convenient not just to deal with self-adjoint operators, but rather to consider the collection $L(\mathcal{H})$ of all bounded linear operators defined on the Hilbert space $\mathcal{H}$.
\paragraph{Identification}
In classical theory, recall that we made identification of observables that cannot be distinguished in view of the probability measure $\mu$ by introducing the equivalence relation $f \sim g \Leftrightarrow f = g \text{ $\mu$-a.e.}$, and slimmed down the space of functions into quotient spaces (see Section~\ref{sec:usp_I_MI}).
We intend to follow the same path for the quantum case, and
to this, we introduce the subspace
\begin{equation}
Z_{\psi}(H) := \{ X \in L(H) : X|\psi\rangle = X^{*}|\psi\rangle = 0 \}
\end{equation}
and define the $\mathbb{C}$-linear quotient space
\begin{equation}
L_{\psi}(\mathcal{H}) := L(H) / Z_{\psi}(H)
\end{equation}
by identifying those operators for which the action of both themselves and their adjoints are indistinguishable on the state $|\psi\rangle$. In other words, this is to say that we identify two operators $X, Y \in L(\mathcal{H})$ by the equivalence relation
\begin{equation}
X \sim Y \Leftrightarrow
X|\psi\rangle = Y|\psi\rangle \text{ and } X^{*}|\psi\rangle = Y^{*}|\psi\rangle.
\end{equation}
One readily sees that the sesquilinear form \eqref{def:alpha_sesquilinear_form} passes to the quotient, and we thus obtain a sesquilinear form
\begin{equation}\label{def:alpha_sesquilinear_form_quotient}
\llangle [Y]_{\psi}, [X]_{\psi} \rrangle_{\psi,\alpha} := \llangle Y, X \rrangle_{\psi,\alpha}, \quad [X]_{\psi}, [Y]_{\psi} \in L_{\psi}(\mathcal{H})
\end{equation}
on the quotient space $L_{\psi}(\mathcal{H})$. Whenever there is no risk of confusion, we shall mostly denote equivalence classes by their representatives for simplicity of notation.
Note also that the \emph{involution} $\ast : L(\mathcal{H}) \to L(\mathcal{H}),\ X \mapsto X^{*}$ that takes a bounded linear operator to its adjoint is also well-defined on the quotient space.
\paragraph{Hilbert Space of Operators}
We have already seen that the original sesquilinear form \eqref{def:alpha_sesquilinear_form} becomes positive and symmetric for the choice $-1 \leq \alpha \leq 1$ of the parameter. Based on the identification above, the sesquilinear form \eqref{def:alpha_sesquilinear_form_quotient} on the quotient space $L_{\psi}(\mathcal{H})$ becomes \emph{positive definite}, which is to say that
\begin{enumerate}
\item $\llangle X, X \rrangle_{\psi,\alpha} \geq 0, \quad X \in L_{\psi}(\mathcal{H})$,
\item $\llangle X, X \rrangle_{\psi,\alpha} = 0 \quad \Leftrightarrow \quad X = 0$
\end{enumerate}
for the choice $-1 \leq \alpha \leq 1$. This makes \eqref{def:alpha_sesquilinear_form_quotient} an \emph{inner product} on $L_{\psi}(\mathcal{H})$ for $-1 \leq \alpha \leq 1$, allowing us to define the norm
\begin{equation}\label{def:norm_quotient}
\|X\|_{\psi,\alpha} := \llangle X, X \rrangle_{\psi, \alpha}^{\frac{1}{2}}, \quad X \in L_{\psi}(\mathcal{H}),\ -1 \leq \alpha \leq 1.
\end{equation}
One moreover proves by rudimentary technique that the space is in fact complete with respect to the norm. We thus have the following result.
\begin{proposition}[Hilbert Space of Operators]
For a fixed $|\psi\rangle \in \mathcal{H}$ and the choice $-1 \leq \alpha \leq 1$ of the parameter, the ordered pair $(L_{\psi}(\mathcal{H}), \llangle\, \cdot\, , \, \cdot\, \rrangle_{\psi, \alpha})$ defines a complex Hilbert Space.
\end{proposition}
\noindent
This convenient property greatly facilitates our further argument. Hence, in what follows, we will be treating only those geometries associated with the specific choice $-1 \leq \alpha \leq 1$ of the complex parameter.
\paragraph{Sub-algebra generated by an Observable}
We next introduce an important subspace of $L(\mathcal{H})$.
Given a bounded self-adjoint operator $A \in L(\mathcal{H})$, we prepare a special symbol
\begin{equation}\label{def:subspace_generated_by_normal_operators}
\mathfrak{E}(A) := \left\{ f(A) : f \text{ is a continuous function on $\sigma(A)$} \right\}
\end{equation}
for the $\mathbb{C}$-linear subspace of $L(\mathcal{H})$ consisting of all operators defined by means of the functional calculus \eqref{def:func_calc}. By definition, one proves that
\begin{equation}
\left\| f(A) \right\| = \| f \|_{\infty},
\end{equation}
where the l.~h.~s. is the operator norm of $f(A)$, and the r.~h.~s. is the supremum norm of the continuous function $f$ (note that the spectrum $\sigma(A)$ of a bounded self-adjoint operator $A$ is compact, hence any continuous function defined on the spectrum is necessarily bounded). Moreover, it is easy to see that $f(A)^{*} = f^{*}(A)$ holds, where the l.~h.~s. denotes the adjoint of the linear operator $f(A)$, whereas the r.~h.~s. denotes the operator induced by the complex conjugate $f^{*}$ of the original function $f$. This implies that all the operators of the form $f(A)$ are normal%
\footnote{Recall that a bounded operator $N \in L(\mathcal{H})$ is \emph{normal} if and only if $\|N\psi\| = \|N^{*}\psi\|$ holds for all $|\psi\rangle \in \mathcal{H}$.},
and that the space $\mathfrak{E}(A)$ is closed under the operation of taking the adjoint. Moreover, one sees that any two operators $f(A), g(A) \in \mathfrak{E}(A)$ commute with each other $f(A)g(A) = (fg)(A) = g(A)f(A)$, and that the space $\mathfrak{E}(A)$ thus makes itself into a commutative $C^{*}$-algebra. We call the space $\mathfrak{E}(A)$ the sub-algebra \emph{generated by $A$}.
\paragraph{Identification}
An immediate observation one makes is that the sesquilinear form \eqref{def:alpha_sesquilinear_form} is independent of the choice of the parameter $\alpha \in \mathbb{C}$ on the space $\mathfrak{E}(A)$. Indeed, for any choice of a pair of continuous functions $f$, $g$ defined on the spectrum $\sigma(A)$, the equality
\begin{align}\label{def:sesquilinear_form_normal}
\llangle g(A), f(A) \rrangle_{\psi, \alpha}
&= \frac{\langle g(A)\psi, f(A)\psi \rangle}{\|\psi\|^{2}} \nonumber \\
&= \int_{\mathbb{R}} g^{*}(a)f(a)\ d\mu_{A}^{\psi}(a)
= \langle g, f \rangle_{\mu_{A}^{\psi}}
\end{align}
holds, where the right-most hand side denotes the standard inner product introduced on the space of square-integrable complex functions.
Following the same line of arguments we have made in the previous discussion of this section, we next intend to identify those operators that are not distinguishable in view of a given state $|\psi\rangle \in \mathcal{H}$. To this, we introduce the subspace
\begin{equation}
Z_{\psi}(A) :=
\{ N \in \mathfrak{E}(A) : N|\psi\rangle = N^{*}|\psi\rangle = 0 \},
\end{equation}
and define the space
\begin{equation}\label{def:subspace_generated_by_normal_operators_quotient}
\mathfrak{E}_{\psi}(A) := \overline{\mathfrak{E}(A) / Z_{\psi}(A)}
\end{equation}
by identifying those normal operators for which the action of both themselves and their adjoints on the state $|\psi\rangle$ are indistinguishable. Here, the overline on the quotient space $\mathfrak{E}(A) / Z_{\psi}(A) \subset L_{\psi}(\mathcal{H})$ denotes its topological closure with respect to the topology on the superset $L_{\psi}(\mathcal{H})$ induced by the norm $\|\cdot\|_{\psi,\alpha}$ \eqref{def:norm_quotient}. Note here that, as a set, the closure $\mathfrak{E}_{\psi}(A)$ is independent of the choice of the parameter $-1 \leq \alpha \leq 1$, since all the norms $\|\cdot\|_{\psi,\alpha}$ coincide on the subspace $\mathfrak{E}(A) / Z_{\psi}(A)$. Moreover, one may readily check that all the inner products $\llangle \,\cdot\, , \,\cdot\, \rrangle_{\psi,\alpha}$ also coincide for the pair of elements of the closure $\mathfrak{E}_{\psi}(A)$ for any choice of the parameter $-1 \leq \alpha \leq 1$.
Now, since by definition the space $\mathfrak{E}_{\psi}(A)$ is a closed subspace of the complex Hilbert space, it is itself a complex Hilbert space. By denoting the restriction of the inner product as $\llangle \,\cdot\,, \,\cdot\, \rrangle_{\psi} := \llangle \,\cdot\,, \,\cdot\, \rrangle_{\psi,\alpha}|_{\mathfrak{E}_{\psi}(A)}$, which does not depend on the choice of the parameter as we have mentioned above, we have:
\begin{lemma}
For a fixed $|\psi\rangle \in \mathcal{H}$, the ordered pair $(\mathfrak{E}_{\psi}(A), \llangle\, \cdot\, , \, \cdot\, \rrangle_{\psi})$ defines a complex Hilbert space.
\end{lemma}
\noindent
The next Lemma is of special interest for our purpose.
\begin{lemma}
Let $A \in L(\mathcal{H})$ be self-adjoint, $|\psi\rangle \in \mathcal{H}$, and let $\mathfrak{E}_{\psi}(A)$ be defined as in \eqref{def:subspace_generated_by_normal_operators_quotient}. Then, there exists a unique unitary operator $\Phi : L^{2}(\mu_{A}^{\psi}) \mapsto \mathfrak{E}_{\psi}(A)$ such that
\begin{equation}
\Phi(f) = [f(A)]_{\psi}
\end{equation}
holds for every continuous function $f$ on $\sigma(A)$, where the r.~h.~s. denotes the equivalent class of $f(A)$, which in turn is a bounded linear operator defined by means of the functional calculus \eqref{def:func_calc}.
\end{lemma}
\begin{proof}
We will construct the map $\Phi$ by continuous linear extension. To this, first recall that, since $\sigma(A)$ is compact, the space
$C(\sigma(A))$ of all continuous functions on $\sigma(A)$ is dense in $L^{2}(\mu_{A}^{\psi})$. Since the map
\begin{equation}
\tilde{\Phi} : L^{2}(\mu_{A}^{\psi}) \supset C(\sigma(A)) \to \mathfrak{E}_{\psi}(A), \quad f \mapsto f(A),
\end{equation}
is an isometry $\|\tilde{\Phi}(f)\|_{\psi} := \| f(A) \psi \|= \|f\|_{2}$ from a dense subspace of a normed space to a Banach space, there exists a unique isometric extension $\Phi : L^{2}(\mu_{A}^{\psi}) \to \mathfrak{E}_{\psi}(A)$. By construction, one may also prove the surjectivity of $\Phi$, hence $\Phi$ is unitary.
\end{proof}
\noindent
This is to say that the Hilbert spaces $\mathfrak{E}_{\psi}(A) \cong L^{2}(\mu_{A}^{\psi})$ are unitarily isomorphic, and that $\Phi$ gives an embedding of the space of square-integrable functions $L^{2}(\mu_{A}^{\psi})$ into the space of bounded operators on a Hilbert space. For simplicity of notation, we occasionally denote the image of a square-integrable function $f \in L^{2}(\mu_{A}^{\psi})$ by $f(A) := \Phi(f)$. A word of caution is to be made here for the possible confusion for the notation used. Here, the notation $f(A) := \Phi(f)$ is meant to denote (a representative of) the equivalence class of bounded linear operators, whereas the notation $f(A)$ is usually used to represent (generally unbounded) linear operator defined by means of the functional calculus \eqref{def:func_calc}. The relation between the two different notations can be understood in the following way. For $f \in L^{2}(\mu_{A}^{\psi})$, let $T \in \Phi(f)$ be a representative of the equivalence class of bounded operators, and let $f(A)$ be a (generally unbounded) operator defined by means of the functional calculus \eqref{def:func_calc}, and note that $|\psi\rangle \in \mathrm{dom}(f(A))$ by definition. Then, we have $T|\psi\rangle = f(A)|\psi\rangle$.
\subsection{Interpretation of Conditional Quasi-expectations}
Now that we have sufficiently prepared our tools, the most important among which is the embedding
\begin{equation}\label{eq:embedding_l2}
\Phi: L^{2}(\mu_{A}^{\psi}) \cong \mathfrak{E}_{\psi}(A) \subset L_{\psi}(\mathcal{H})
\end{equation}
of the $L^{2}$-space of functions into that of bounded linear operators on a Hilbert space, we next focus on orthogonal projections and `conditioning' with respect to QJP distributions.
\subsubsection{Geometric Interpretation of Conditional Quasi-expectations}
Recall that, with each closed subspace of a Hilbert space, a unique \emph{orthogonal projection} is associated. In what follows, we are interested in the orthogonal projection of an observable $A$ onto the subspace $\mathfrak{E}_{\psi}(B)$ generated by another observable $B$, and see that this provides a geometric interpretation of the conditional quasi-expectation $\mathbb{E}^{\alpha}[A|B; \psi]$ introduced earlier in \eqref{def:cond_quasi-exp_alpha}.
To this, let $B \in L(\mathcal{H})$ be self-adjoint, $|\psi\rangle \in \mathcal{H}$, $-1 \leq \alpha \leq 1$, and let $\mathfrak{E}_{\psi}(B)$ be the space generated by $B$, which is a closed subspace of the Hilbert spaces $(L_{\psi}(\mathcal{H}), \llangle \,\cdot\,, \,\cdot\,\rrangle_{\psi,\alpha}$). As a closed subspace of a Hilbert space, there exists a unique orthogonal projection
\begin{equation}\label{def:orth_proj}
P_{\alpha}(\ \cdot \ |B ; \psi): L_{\psi}(\mathcal{H}) \to \mathfrak{E}_{\psi}(B),\ X \mapsto P_{\alpha}(X|B;\psi)
\end{equation}
associated to $\mathfrak{E}_{\psi}(B)$. Recalling the relation between orthogonal projections and conditional expectations in classical probability theory (see Section~\ref{sec:app_ref}), it is natural to conjecture that an analogous relation holds for the quantum case. To this, let $A \in L(\mathcal{H})$ be self-adjoint, and consider the projection $P_{\alpha}(A |B ; \psi)$ of $A$ onto $\mathfrak{E}_{\psi}(B)$. We have seen in Section~\ref{sec:ps_I} that the conditional quasi-expectations $\mathbb{E}^{\alpha}[A|B; \psi] \in L^{1}(\mu_{B}^{\psi})$ introduced in \eqref{def:cond_quasi-exp_alpha} serve as possible candidates of quantum analogues of conditional expectations that can even be defined for non-commuting pair of quantum observables. Since, the observable $A$ we consider here is bounded, one may prove that the conditional quasi-expectation $\mathbb{E}^{\alpha}[A|B; \psi] \in L^{2}(\mu_{B}^{\psi})$ is in fact square-integrable. By letting $\mathbb{E}^{\alpha}[A|B; \psi]$ denote both the square-integrable function and its image by the unitary map $\Phi : L^{2}(\mu_{B}^{\psi}) \to \mathfrak{E}_{\psi}(B)$, it is natural to conjecture the validity of the equality $P_{\alpha}(A|B;\psi) = \mathbb{E}^{\alpha}[A|B; \psi]$, where the l.~h.~s. is the orthogonal projection of $A$ onto the space $\mathfrak{E}_{\psi}(B)$ with respect to the (parameter-dependent) inner product $\llangle \,\cdot\,, \,\cdot\,\rrangle_{\psi,\alpha}$, whereas the r.~h.~s. denotes the image of the (parameter-dependent) conditional quasi-expectation of $A$ given $B$ by the unitary map $\Phi$.
\begin{proposition}[Orthogonal Projection and Conditional Quasi-Expectation]\label{prop:orth_proj_cond_qe}
Let $A, B \in L(\mathcal{H})$ be self-adjoint, $|\psi\rangle \in \mathcal{H}$ and $-1 \leq \alpha \leq 1$. Then, the conditional quasi-expectation $\mathbb{E}^{\alpha}[A|B; \psi]$ introduced in \eqref{def:cond_quasi-exp_alpha} is $\mu_{B}^{\psi}$-square-integrable, which could thus be identified with the equivalence class of bounded operators
\begin{equation}
\mathbb{E}^{\alpha}[A|B; \psi] := \Phi(\mathbb{E}^{\alpha}[A|B; \psi])
\end{equation}
by means of the embedding $\Phi : L^{2}(\mu_{B}^{\psi}) \to \mathfrak{E}_{\psi}(B)$ defined in \eqref{eq:embedding_l2}. Then, the orthogonal projection of $A$ onto the subspace $\mathfrak{E}_{\psi}(B)$ generated by $B$, defined in \eqref{def:orth_proj}, reads
\begin{equation}
P_{\alpha}(A|B;\psi) = \mathbb{E}^{\alpha}[A|B; \psi],
\end{equation}
which is to say that orthogonal projections are equivalent to conditional quasi-expectations.
\end{proposition}
\begin{proof}
Let $f \in L^{2}(\mu_{B}^{\psi})$, and let $f(B) := \Phi(f)$ denote the embedding of $f$ into the space $\mathfrak{E}_{\psi}(B)$. By definition of orthogonal projections, one readily finds
\begin{equation}\label{eq:orth_proj_proof_1}
\llangle f(B), A \rrangle_{\psi,\alpha} = \llangle f(B), P_{\alpha}(A|B;\psi) \rrangle_{\psi}.
\end{equation}
On the other hand, observe that
\begin{align}
\langle f, \mathbb{E}^{\alpha}[A|B; \psi] \rangle_{\mu_{B}^{\psi}}
&:= \int_{\mathbb{R}} f^{*}(b) \cdot \mathbb{E}^{\alpha}[A|B = b; \psi]\ d\mu_{B}^{\psi}(b) \nonumber \\
&= \frac{1+\alpha}{2} \int_{\mathbb{R}} f^{*}(b)\ d\frac{\langle \psi,E_{B}(b)A\psi\rangle}{\|\psi\|^{2}} + \frac{1-\alpha}{2} \int_{\mathbb{R}} f^{*}(b)\ d\frac{\langle \psi, AE_{B}(b)\psi\rangle}{\|\psi\|^{2}} \nonumber \\
&= \frac{1+\alpha}{2} \cdot \frac{\langle f(B)\psi, A\psi\rangle}{\|\psi\|^{2}} + \frac{1-\alpha}{2} \cdot \frac{\langle A^{*}\psi, f(B)^{*}\psi\rangle}{\|\psi\|^{2}} \nonumber \\
&=: \llangle f(B), A \rrangle_{\psi,\alpha}
\end{align}
Combining this with the unitarity of the embedding $\Phi : L^{2}(\mu_{B}^{\psi}) \to \mathfrak{E}_{\psi}(B)$, one thus has
\begin{equation}\label{eq:orth_proj_proof_2}
\llangle f(B), A \rrangle_{\psi,\alpha} = \llangle f(B), \mathbb{E}^{\alpha}[A|B; \psi] \rrangle_{\psi}.
\end{equation}
The positive-definiteness of the inner product applied to the two results \eqref{eq:orth_proj_proof_1} and \eqref{eq:orth_proj_proof_2} proves our statement.
\end{proof}
\noindent
Just as we have seen for the classical case, this result provides a geometric interpretation of conditional quasi-expectations as orthogonal projections. As a corollary, one has a geometric interpretation of Aharonov's weak value.
\begin{corollary}[Geometric Interpretation of Aharonov's Weak Value]\label{prop:orth_proj_wv}
Under the same conditions, let
\begin{equation}
A_{w}(B) := \Phi(A_{w})
\end{equation}
denote the embedding of the Aharonov's weak value $A_{w} : = \mathbb{E}^{1}[A|B; \psi]$ introduced in \eqref{def:weak_value}. Then, the weak value
\begin{equation}
A_{w}(B) = P_{1}(A|B;\psi) = \mathbb{E}^{1}[A|B; \psi]
\end{equation}
could be interpreted as the orthogonal projection of $A$ onto the subspace $\mathfrak{E}_{\psi}(B)$ generated by $B$.
\end{corollary}
\paragraph{Topic: Weak Value as Optimal Approximation}
As a direct consequence of Proposition~\ref{prop:orth_proj_wv} (specifically, Corollary~\ref{prop:orth_proj_wv}), we note an interesting result regarding conditional quasi-expectations (specifically, the weak value) and optimal approximation. As orthogonal projections, observe that conditional quasi-expectations furnish the optimal proxy function for $A$ minimising the distance
\begin{equation}
\|A - \mathbb{E}^{\alpha}[A|B; \psi] \|_{\psi,\alpha} = \min_{f} \|A - f(B) \|_{\psi,\alpha}
\end{equation}
from an observable $A$ to the space of normal observables $\mathfrak{E}_{\psi}(B)$ generated by another observable $B$. An equivalent expression to this is the equality
\begin{equation}
\|A - f(B) \|_{\psi,\alpha} = \|A - \mathbb{E}^{\alpha}[A|B; \psi] \|_{\psi,\alpha} + \|\mathbb{E}^{\alpha}[A|B; \psi] - f(B) \|_{\psi,\alpha},
\end{equation}
which is nothing but the `Pythagorean identity' valid for orthogonal projections%
\footnote{Specifically, observing that $\mathbb{E}^{\alpha}[A|B; \psi] = \mathrm{Re}A_{w}(B) + i \alpha \mathrm{Im}A_{w}(B)$, we have $\mathbb{E}^{\alpha}[A|B; \psi] = \mathrm{Re}A_{w}(B)$ for the choice $\alpha = 0$. This gives
\begin{align}
\| \left( A - \mathrm{Re}A_{w}(B) \right) \psi \|
&= \|A - \mathrm{Re}A_{w}(B) \|_{\psi,0} \nonumber \\
&\leq \|A - f(B) \|_{\psi,0} \nonumber \\
&= \| \left( A - f(B) \right) \psi \|, \quad f \in L^{2}(\mu_{B}^{\psi}),\ f \text{ is real},
\end{align}
as a special case. This specific form is known by \cite{Hall_2001,Johansen_2004}, although proven from a different perspective than directly utilising the geometric observation made in this paper.}
in Hilbert spaces.
The interpretation of conditional quasi-expectations as orthogonal projections provide the core geometric observations why the weak value appears as the optimal choice for the proxy functions in the novel uncertainty relations for approximation/estimation \cite{Lee_2016}.
\begin{figure}
\includegraphics[width=160mm]{projection.eps}
\caption{Geometric relations among the operators involved for the choice $\alpha = 1$. The left illustrates how the operator $A$ is projected onto the subspace of normal operators $\mathfrak{E}_{\psi}(B)$ generated by $B$, with the center line representing the space of self-adjoint operators $\{ f(B)\}$. The right elaborates the projection onto the space $\{ f(B)\}$, where
now the center line represents the space of constant functions $\{ f(B) = \hbox{const.} \}$ (more precisely, functions proportional to the identity operator $I$) including $f(B) = \langle A\rangle$.
}
\end{figure}
\subsubsection{Statistical Interpretation of Conditional Quasi-expectations}
In classical probability theory, conditional expectations not only admitted geometric interpretation as orthogonal projections, but also statistical interpretation as conditioned averages. We next seek to provide a quantum analogue of this observation, namely, to provide a statistical interpretation of the conditional quasi-expectations as `conditional averages' with respect to QJP distributions. To this, we first introduce a general term:
\begin{definition*}[Conditional Quasi-expectation of Quantum Observables]
Let $\mu \in \mathfrak{M}_{A,B}^{\psi}$ be a QJP distribution of a pair of quantum observables $A$ and $B$ on the state $|\psi\rangle \in \mathcal{H}$, such that is admits representation by quasi-probability measures, and suppose that the expectation value $\mathbb{E}[\pi_{A};\mu]$ exists. Denoting the measurable functions representing the behaviour of the measurement outcomes of $A$ and $B$ by $\pi_{A}(a,b) = a$ and $\pi_{B}(a,b) = b$, respectively, we then define the conditional quasi-expectation of $A$ given $B$ under the QJP distribution $\mu$ by
\begin{equation}
\mathbb{E}[A|B;\mu] := \mathbb{E}[\pi_{A}|\pi_{B}; \mu],
\end{equation}
where the definition of the r.~h.~s. is given in Section~\ref{sec:CM_II_cond_over_QP}, whenever the Radon-Nikod{\'y}m derivatives concerned exist.
\end{definition*}
\noindent
We next see that the definition of conditional quasi-expectations agree with those introduced earlier \eqref{def:cond_quasi-exp_alpha}.
\begin{proposition}[Statistical Interpretation of Conditional Quasi-expectations]\label{prop:stat_int_cond_quasi_exp}
Let $A, B \in L(\mathcal{H})$ be self-adjoint, $|\psi\rangle \in \mathcal{H}$, and let $\mu_{\mathrm{add}}^{\psi,\alpha}$ be a member of the additive complex-parametrised sub-family of the QJP distributions of $A$ and $B$ for the choice $\alpha \in \mathbb{C}$ defined as in \eqref{def:sec_app_additive}.
Then, the conditional quasi-expectation $\mathbb{E}[A|B; \mu_{\mathrm{add}}^{\psi,\alpha}]$ of $A$ given $B$ under $\mu_{\mathrm{add}}^{\psi,\alpha}$ is well-defined, which reads
\begin{align}
\mathbb{E}^{\alpha}[A|B;\psi] = \mathbb{E}[A|B; \mu_{\mathrm{add}}^{\psi,\alpha}],
\end{align}
where the l.~h.~s. is the $\alpha$-parametrised conditional quasi-expectation introduced earlier in \eqref{def:cond_quasi-exp_alpha}.
\end{proposition}
\begin{proof}
We first demonstrate the well-definedness of $\mathbb{E}[A|B; \mu_{\mathrm{add}}^{\psi,\alpha}]$, and to this, let
\begin{align}
\nu(\Delta)
&:= \int_{\pi_{B}^{-1}(\Delta)} \pi_{A}(a,b)\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b) \nonumber \\
&= \int_{\mathbb{R}\times\Delta} a\ d\mu_{\mathrm{add}}^{\psi,\alpha}(a,b) \nonumber \\
&= \frac{1+\alpha}{2} \cdot \frac{\langle \psi, E_{B}(\Delta)A\psi\rangle}{\|\psi\|^{2}} + \frac{1-\alpha}{2} \cdot \frac{\langle \psi, AE_{B}(\Delta)\psi\rangle}{\|\psi\|^{2}}.
\end{align}
Since $\nu \ll \mu_{B}^{\psi}$, the Radon-Nikod{\'y}m derivative $\mathbb{E}[A|B; \mu_{\mathrm{add}}^{\psi,\alpha}] := d\nu/d\mu_{B}^{\psi}$ exists. The validity of the equality $\mathbb{E}^{\alpha}[A|B;\psi] = \mathbb{E}[A|B; \mu_{\mathrm{add}}^{\psi,\alpha}]$ is immediate by definition.
\end{proof}
\noindent
This result provides a statistical interpretation of conditional quasi-expectations as `conditional averages' of an observable $A$ given another observable $B$ with respect to the QJP distributions concerned. As a corollary, one also has a statistical interpretation of Aharonov's weak value.
\begin{corollary}[Statistical Interpretation of Aharonov's Weak Value]
Under the same conditions, let $A_{w} : = \mathbb{E}^{1}[A|B; \psi]$ denote the Aharonov's weak value introduced in \eqref{def:weak_value}. Then, the weak value
\begin{equation}
A_{w}(B) = \mathbb{E}[A|B; \mu_{\mathrm{add}}^{\psi,1}] = \mathbb{E}^{1}[A|B; \psi]
\end{equation}
admits interpretation as the `conditional averages' of an observable $A$ given another observable $B$ with respect to the additive subfamily of QJP distributions for the choice $\alpha = 1$.
\end{corollary}
\newpage
\section{Summary and Discussion}\label{sec:sc}
Now we present a recap of our results obtained in this paper before going over to our discussions.
\subsection{Summary}
The underlying motivation for our study was to obtain a coherent understanding to the formalism of quasi-joint-probabilities (QJP) of quantum observables, and to find the interpretation of Aharonov's weak value within this framework. The main body, starting from Section 2 to 7, was devoted to the discussion of three logical groups of mutually interrelated topics, namely (i) an heuristic construction of QJP distributions (Section~\ref{sec:ups_I} to \ref{sec:ps_II}), (ii) formal definition of QJP distributions (Section~\ref{sec:qp_qo}), and (iii) its application to the interpretation of the weak value (Section~\ref{sec:app}).
Each of these sections will be summarised concisely below.
\subsubsection{QJP: Heuristic Construction}
Four sections starting from Section~\ref{sec:ups_I} to Section~\ref{sec:ps_II} were devoted to some careful analyses on the quantum measurement models that we called the unconditioned and the conditioned measurement (UM and CM) schemes. By inspecting each of the measurement models in terms of statistical averages and raw probability distributions, we confirmed that one may obtain the desired information of the target system by either looking into the strong or weak regions of the intensity of the interaction parameter. Specifically, we saw that the study on the CM scheme naturally lead us to the concept of quasi-joint-probability (QJP) distributions of generally non-commuting pair of observables.
\paragraph{Section~\ref{sec:ups_I} (UM I)}
Section~\ref{sec:ups_I} was devoted to a review on the UM scheme from a standard operator-centric approach.
The quantity of interest was the statistical average of the meter observable after the interaction, from which we confirmed the well known fact that the information of both the target observable $A$ and the target state $|\phi\rangle$ can be retrieved in the form of the expectation value $\mathbb{E}[A;\phi]$ of $A$. The expectation value $\mathbb{E}[A;\phi]$ was shown to be obtained from the measurement outcome of the meter observable, either by probing the strong limit $g \to \pm \infty$ or the weak limit $g \to 0$ of the interaction parameter.
\paragraph{Section~\ref{sec:ups_II} (UM II)}
In Section~\ref{sec:ups_II}, we took a closer look at the UM scheme in the level of probabilities, where the quantity of interest was now not just the statistical average but the `raw' probability measure describing the probabilistic behaviour of the measurement outcomes of the meter observable. We saw that the outcome of the meter observable after the interaction was given by a convolution of both the initial profiles of the target and the meter observables. As for the retrieval of the target information, we found that, in a parallel manner to the previous section, the full profile of the target observable
can be reclaimed by either probing the strong or the weak limit of the interaction.
\paragraph{Section~\ref{sec:ps_I} (CM I)}
In Section~\ref{sec:ps_I}, we conducted an analysis of the conditioned measurement scheme in the operator level, where the quantity of interest became the \emph{conditional expectation} of the meter observable under another given conditioning observable $B$ of the target system. Some relevant topics, including a review and comments on the recent theoretical analyses on the alleged technical advantages of employing conditioning for precision measurements were presented, along with a measure theoretic approach to the possible limit for `amplification' by conditioning, and a systematical method (with an example) to analytically evaluate the conditional expectation in the case where $A$ has a spectrum consisting of finite points. As for the retrieval of the target information, we exclusively studied the behaviour of the meter outcome in the weak region of the interaction parameter, and observed that the obtained value can be understood as a quantum analogue of conditional expectations, which we termed \emph{conditional quasi-expectations}, of the target observable $A$ given the conditioning observable $B$, to which Aharonov's weak value belongs as a special case. It was also revealed that there exists some qualitative difference on the properties between the classical conditional expectations and the quantum analogue discussed here.
\paragraph{Section~\ref{sec:ps_II} (CM II)}
In Section~\ref{sec:ps_II}, the study of the conditioned measurement scheme was given a probabilistic approach, where the quantity of interest now became the \emph{QJP distribution} of a pair of canonically conjugate observables on the meter system, conditioned by the outcome of the conditioning observable $B$ of the target system. For definiteness, this was accomplished in view of the Wigner-Ville distribution, which was primarily chosen as a convenient realisation among the various candidates of the quasi-probability distributions of the canonically conjugate pair that may be naturally associated with the quantum state of the meter system.
It was then argued that, in parallel to the UM case, one can recover the information of the target system by examining either the strong or the weak region of the interaction parameter, and that the information obtained can be understood as a quantum analogue of conditional probabilities, which we termed \emph{conditional quasi-probabilities}, of the target observable $A$ given the conditioning observable $B$ on the initial state $|\phi\rangle$.
We then found that the conditional quasi-probability shares similar properties with the classical counterpart, while it admits complex values unlike the classical one. We subsequently confirmed that, given $B$, the `statistical average' of the conditional quasi-probability of $A$ coincides with the conditional quasi-expectation of $A$ obtained in the preceding section. This is precisely the same as the relation between classical conditional probabilities and conditional expectations.
\subsubsection{QJP: Formal Definition}
Inspired by the heuristic arguments from the bottom-up and operational analyses given in the preceding four sections, in Section~\ref{sec:qp_qo} we provided the top-down discussion on QJP distributions defined for arbitrary pairs of generally non-commutative quantum observables.
\paragraph{Section~\ref{sec:qp_qo} (QJP of Quantum Observables)}
Based on the results of the spectral theorem for self-adjoint/normal operators on Hilbert spaces and their Fourier transforms, we proposed a general prescription for defining distributions describing the `joint behaviour' of a pair of generally non-commuting quantum observables, which serves as a natural generalisation to that defined for a pair of simultaneously measurable observables. We then observed that the QJP defined this way for a non-commutative pair of observables admits arbitrariness, that is, there exists a multitude of candidates that all share in common certain desirable properties to be qualified as QJP.
We subsequently concentrated on a special sub-family of the class of QJP distributions parametrised by a single complex number for the ease of further discussions, such that it includes both the Wigner-Ville type and the Kirkwood-Dirac type of QJP distributions which are among the most familiar examples considered in the literature. We then summarised our results obtained up to Section~\ref{sec:ps_II} from a relatively aerial viewpoint gained here, and discussed where the heuristic arguments and observations in the foregoing sections find their places in this broader framework.
\subsubsection{Application}
As the final topic, we gave an example of application of our observations on QJP distributions of quantum observables.
\paragraph{Section~\ref{sec:app} (Application: Interpretation of the Weak Value)}
To discuss where the mathematical observations on QJP distributions may find their use, we studied on the quantum analogue of `correlations' (inner products) that can be defined even for a pair of non-commuting observables. As is well known, due to the non-commutative nature of quantum observables, there is no unique way to introduce a `natural inner product' on the space of quantum observables. We showed that the ambiguity of the possible geometries that can be introduced on the space corresponds precisely to the ambiguity of the definition of QJP distributions, and that the QJP distributions provides a convenient representation of the geometries in terms of `integration' (statistics).
We then concentrated on a special sub-family of all possible QJP distributions parametrised by a single complex number and observed that
the geometric concept of \emph{orthogonal projection} may be endowed with a statistical interpretation as \emph{conditioning}.
This fact is analogous to the classical case, while the difference lying in the fact that, for the quantum case, there could be multiple orthogonal projections due to the non-uniqueness of the inner product.
The main finding is that, Aharonov's weak value may be understood as a special realisation of the possible orthogonal projections of a quantum observable $A$ onto the space of all normal operators generated by another observable $B$, and at the same time, as a conditioning of $A$ when the outcomes of $B$ is given. The former is a geometric interpretation of the weak value, while the latter is its statistical interpretation, but since QJP distributions tie them together, both interpretations are equivalent.
\subsection{Discussion}
Since the advent of quantum theory founded nearly a century ago,
non-commutativity of quantum observables has undoubtedly been one of the major sources of troubles we face when we try to interpret their measurement outcomes in a sensible manner.
This has naturally led to various attempts of `quasi-classical' interpretation of quantum observables in terms of commuting quantities familiar to us in classical theory.
Wigner, Weyl and Moyal were among the prominent figures who have made much contribution in this effort, bearing most notably the theory of Wigner-Weyl transform \cite{Weyl_1927} and Weyl-Groenewold-Moyal product \cite{Groenewold_1946, Moyal_1949}. In particular, the theory of
Wigner-Weyl transform provides an invertible mapping between functions defined on a phase space and operators on a Hilbert space, in which the mapping from functions to operators is called the Weyl transform, whereas the inverse is called the Wigner transform.
It is notable in this respect that the Wigner-Ville distributions arise as the Wigner transform of density operators, and from this follows the fact that
the expectation values of quantum observables can be expressed as the statistical average by integration of their Wigner transforms with respect to the Wigner-Ville distribution defined on the phase space.
Viewed from the broader context of these quasi-classical transforms, the mathematical methods developed in this paper may be understood as another functional analytic approach to this problem. Recall that, in functional analysis, a map that assigns a ring of functions onto a \emph{commutative} sub-algebra of the algebra of quantum observables is known as the functional calculus, which in turn is known to be uniquely represented by a spectral measure. The family of quasi-joint-spectral distributions (QJSDs) introduced in Section~\ref{sec:qp_qo} are non-commutative analogues of spectral measures, which induce maps that assign functions to generally \emph{non-commutative} sets of quantum observables. Due to the possible non-commutativity of the chosen combination of observables, QJSDs are in general highly non-unique, and this leads to various candidates of quasi-classical transforms. In fact, the Wigner-Weyl transform can be understood as a special case in this framework, namely, the quasi-classical transform corresponding to the member of our complex parametrised convolutive sub-family of QJSDs mentioned in the text for the particular choice $\alpha = 0$. The method of `hashing' presented in this paper thus exemplifies a procedure for constructing a broad class of candidates of quasi-classical transforms.
The method of quasi-classical transforms, to which the Wigner-Weyl transform belongs as a special case, not only offers a statistical interpretation of the behaviour of a combination of non-commuting quantum observables, but also sheds new light on the physical analysis in quantum mechanics pertaining to that process. It should be obvious that one can draw an analogy to various concepts and results in classical probability theory when one considers the quantum counterparts obtained by this method, which allows for an intuitive treatment of the latter based on the geometric structure present in the probability theory. Besides, transformation of Hilbert space operators into functions or quasi-probability distributions has its own technical merit in the mathematical analysis, since familiar results in measure and integration theory, including various convergence theorems, integral inequalities and representation theorems, are readily available.
One of the direct applications taking advantage of these properties is the geometric/statistical interpretation of the weak value discussed in Section~\ref{sec:app}.
There, we found that the weak value can be regarded as one of the possible quantum analogues of conditional expectations, which are indeed fundamental quantities in quantum mechanics as much as the standard conditional expectations are in classical probability theory.
This interpretation also leads to novel inequalities of uncertainty relations for approximation and estimation which are capable of treating both the position-momentum inequality and the time-energy inequality \cite{Lee_2016} within a unified framework.
Finally, we wish to note that, in any conditioned quantum measurement such as the weak measurement, non-commutative observables must be dealt with in one way or another in the context of probability theory when one tries to make sense of the measurement outcome.
Given this, we expect that our method of quasi-classical transforms, which is established on a rigorous mathematical basis, may offer
a fundamental and practical scheme in which issues involving measurement results of non-commuting observables are analysed.
\newpage
\section*{Acknowledgment}
The authors appreciate Professor A. Hosoya and S. Tanimura for helpful discussions and insightful comments.
This work was supported in part by JSPS KAKENHI No.~25400423, No.~26011506, and by the Center for the Promotion of Integrated Sciences (CPIS) of SOKENDAI.
\vfill\pagebreak
\bibliographystyle{unsrt}
|
\section{Introduction}
Service brokering is a vivid research area where recent publications with several
approaches can be found targeting various contexts such as cloud service marketplaces and cloud federations.
\textit{Cloud service marketplaces} are platforms which act as a mediator between the service requesters and service providers. These platforms discover and store services in a service repository and allow service requesters to browse, select and interact with services via a \textit{service broker} component. They provide a unified view of the cloud service descriptions to the service requester and in some cases additional functionality such as unified monitoring, billing, and enhanced single sign on services \cite{appexchange}, \cite{googlemarketplace}, \cite{awsMarketplace}, \cite{thatmann2012towards}.
\textit{Cloud federations} are defined as inter-cloud organisations which comprise a set of autonomous and heterogeneous clouds \cite{grozev2014inter}. Initial discussions on cloud federations bring up the question on how the service requester can keep control of the selected clouds. For this, it is suggested that the service requester should be able to set requirements to be fulfilled while the federation distributes the deployment on several clouds. A cloud federation should "respect end-to-end SLAs" (service level agreements) \cite{cloudFed2014}. To ensure this, a service broker is essential.
One of the key actions of a service broker is \textit{service matchmaking}, that is, returning one or more suitable cloud service offers from the \textit{service providers} which fulfil the requirements of a \textit{service requester}. Cloud service offers are any computing resources which are provided on application, system and infrastructure levels on the cloud; respectively SaaS, PaaS and IaaS. These vary in functionality, quality metrics, and other non-functional properties such as legal aspects. Moreover, the exact same application deployed on different cloud infrastructures are only different in non-functional aspects. In this setting, service matchmaking is a complex problem for which the scope of matching, the detail level of results, and the service descriptions vary. The distinguishing aspects of the services can be formally described in \textit{service descriptions} using \textit{service description languages} such as Linked-USDL\footnote{Linked Unified Service Description Language} \cite{linkedUSDLSLA}, SMI\footnote{Service Measurement Index} \cite{smi2014v2}, OWL-S\footnote{Semantic Markup for Web Services} \cite{martin2004owl} and SDL-NG\footnote{Next Generation Service Description Language Framework} \cite{slawikdomain} in ongoing research projects.
We apply the \textit{Information Systems Research Framework} \cite{hevner2004design} as our research methodology. Our work primarily targets cloud service marketplaces context. Therefore, we first evaluate existing approaches against two essential requirements of service matchmaking on service marketplaces: the ability to handle incomplete knowledge and to take the service requester's perspective into account. Constraint-based approaches presented in the papers \cite{mobedpour2013user}, \cite{kritikos2009mixed} and \cite{bacciu2010adaptive} align at best with these requirements. Following this, we evaluate the approaches which support the essential requirements with respect to the cloud service descriptions. These approaches cover the most types of properties which result from our analysis of the service descriptions in general and in our service description language\cite{slawikdomain}. Further examination reveals two improvement opportunities: list-typed quality of service (QoS) properties as well as explicit handling of preferences for lower or higher property values.
The goal of this paper is to address these issues by explicitly handling the preferences for QoS parameters and by adding support for list typed QoS parameters. Our solution builds on the idea of using constraint programming to solve the service matchmaking problem. Therefore, we present constraint models and a prototype implementation using constraint solvers which also allows fuzzy service requests.
This paper is organised as follows: The next section describes the service matchmaking problem, the analysis of the types of properties in service descriptions, and evaluates existing approaches against the requirements. Section 3 presents the constraint models, Section 4 presents their implementation using the Java Constraint Programming API (JSR-331)\footnote{Java Specification Request 331}. Section 5 evaluates our approach and presents next steps.
\section{Problem Definition}\label{sec:prob}
The requirements for service matchmakers can be defined on two levels: 1. the process as a whole, 2. the core matchmaking functionality.
The requirements for the process depend on the application context. The application context can be cloud service marketplaces, automated service composition and inter-cloud. Service marketplaces position the service matchmaker as an assistant to a service requester who is a person. Therefore, the service matchmakers in this group build up the request step by step, consider the priorities of the service requester and categorise the results as very good, good and satisfactory (\emph{R1: Service Requester Priorities} and \emph{R2: Comprehensive Results with Matching Degrees}).
The automatic service composition context requires that the optimal service is found without user interaction and that the over-constrained requests are automatically adjusted till a service description matches the service request. For the inter-cloud context, several ways to handle the application brokering are discussed: SLA-based, trigger-action and directly managed \cite{grozev2014inter}. In the directly managed fashion, the service requester handles the deployment on multiple clouds and therefore no broker is involved. In the first two ways, a broker is involved but the service requester does not take the final decision. We argue that in a practical context the service requesters have to take the final decisions with assistance of the broker similar to service marketplaces, as they are most often liable for it - both legally and economically.
The core matchmaking functionality can be examined from three aspects: the scope of matching, incomplete knowledge and fuzziness, and the types of properties identified in the service descriptions.
The scope of matching covers non-functional and functional properties. Functional properties matching is implemented as pre- and post-condition matching and/or API matching, which rather targets the automated service composition context and software developers. The non-functional requirements of the service requester such as interoperability, quality metrics, and legal aspects are handled as QoS matching (\emph{R3:QoS Matching}).
Service matchmakers must deal with missing values, since the service requester might not be sure about all the QoS constraints and the service provider might not supply information for all the QoS properties of the service (\emph{R4: Incomplete Knowledge}). Moreover, the service requesters should be able to define their priorities and fuzzy values with \textit{variational scope} for the constraints (\emph{R5: Fuzziness}).
Variational scope \cite{platenius2013survey} can be introduced to a service matching approach in three ways:
(i)The matching of service description parameters to requirements with a certain amount of tolerance---assuming most requesters would accept a service with a value slightly different than the specified value, (ii) The parameters are specified with fuzzy terms such as "good" or "very fast", (iii)The requirements are specified with their level of importance to the
requester with qualifiers such as mandatory and optional.
\subsection{Analysis of QoS Properties in Service Descriptions}\label{sec:typeqos}
\begin{table}[htp]
\begin{minipage}{\textheight}
\begin{tabular}
{p{15mm}p{14mm}p{14mm}p{14mm}p{12mm}} \toprule
\emph{QoS Property} & \emph{App. X aaS by Provider \#1} &\emph{App. X aaS by Provider \#2}&\emph{App. X aaS by Provider \#3}&\emph{Service \newline Request}\\\midrule
Version &5.5 & 5.6 & 5.6& $=5.6?$\\
Response time & $<120ms$ & $<200ms$ & $<400ms$ & $<300ms$ \\
Storage in Free Version & 0GB& 15GB & 20GB & $>5GB$\\
Availability & $>99.99\%$& $>99.95\%$ & $>99.95\%$& $>99\%$ \\
Establishment Year &2010 & 2005 & 2012& $>2009$\\
Pricing &per dyno-hour & per number of requests & per hour & per hour \\
Compatible Browsers & Explorer, Chrome, Firefox & Explorer, Chrome, Firefox, Safari & Explorer, Safari & Explorer, Firefox, Safari \\ \bottomrule
\hline \end{tabular} \end{minipage}
\caption{An Example Service Matchmaking Problem} \label{tab:example}
\end{table}
This paper analyses the service description
languages/non-functional properties frameworks and parameters commonly used in SLA description languages, especially SMI \cite{siegel2012cloud} and
CRF \footnote{Cloud Requirements Framework}\cite{repschlaeger2012cloud} and the SDL-NG\cite{slawikdomain} developed as part of the project TRESOR\footnote{Trusted Ecosystem for Standardized and Open cloud-based Resources}\cite{tresor2014}.
Table \ref{tab:example} shows examples for different types of properties in three service descriptions and a service request. The example describes the non-functional properties of a database application deployed on different cloud infrastructures by different cloud providers. The service request comprise the constraints on service properties on the right most column of Table \ref{tab:example}.
For \textit{version}, the specification must be equal to the service request which is a numeric value, we name these \textit{discrete numeric value} and \textit{discrete value matching}. For \textit{response time}, the service description guarantees an upper limit. It can be assumed that no discrete value matching will be performed and the lower limit of the request can be ignored. We will refer to these as
\textit{low-value preferred properties}. For \textit{storage in free version}, higher values are preferred, only interval matching with a lower limit is needed and the service description guarantees a lower limit. Similarly, the upper limit of the request can be ignored. We name these \textit{high-value preferred properties}. For the low-value preferred and high-value preferred properties \textit{interval matching} is applied with the assumptions stated above. In addition, for some properties some service requesters prefer higher values and some lower values as in the case of \textit{establishment year}. We will call this \textit{requester defined preference}. The requester defines an upper or lower limit, and it will be matched to the service value or range e.g. $51<x<200$ or 100+. An example for this is the number of employees in a service provider profile. \textit{Pricing} is an example for an enumeration, the service
specification can only take one of the values in the predefined list. In the case of feature lists both the service request and the service specification can take multiple values from the list as given in the
\textit{compatible browsers} example.
Based on the types of properties analysed above, we define the subproblems of service matching as follows: discrete value matching, feature list matching, interval matching, and discrete value matching with soft constraints (\textit{R6: QoS Matching Data Types Coverage}).
\section{Related Work}
Some approaches do not take incomplete knowledge (R1) into account in the service descriptions and service requests \cite{yu2007efficient}, \cite{sarang2012clustering}. D'Mello et al. present an approach \cite{d2008semantic} which compares the services with each other without requester's constraints.
Table \ref{tab:treReqTable} examines QoS Matching data types coverage in related work. Although some subproblems are identified and addressed, existing approaches have shortcomings. All three approaches suggested in \cite{kritikos2008evaluation}, \cite{mukhija2007qos} and \cite{mobedpour2013user} assume that the properties are either low-value or high-value preferred, although there are properties for which the service requester might be searching for exact values. The approach presented by Kritikos et al. \cite{kritikos2008evaluation}, \cite{kritikos2009mixed} improve the algorithm presented by Ruiz et al. \cite{ruiz2005improving} with the advanced categorisation of results. Moreover, they do not consider enumerations and fuzziness. None of them support feature lists leaving this subproblem out.
\begin{table}[htp]
\begin{minipage}{6cm}
\begin{tabular}
{p{16mm}p{9mm}p{5mm}p{5mm}p{7mm}p{5mm}p{10mm}} \toprule
\emph{Research Work} & \emph{Features} &&&&& Maturity Level \\\midrule
Matchmaker & Discrete numeric data & Enumer- ation & Intervals & Fuzziness & Feature Lists & \\
\cite{mobedpour2013user} & yes & no & yes & yes& no& prototype \\
\cite{bacciu2010adaptive}, \cite{mukhija2007qos} & yes & yes & yes & yes \footnote{false negatives for super matches} & no & prototype and theory \footnote{fuzzy not in prototype} \\
\cite{kritikos2008evaluation},\cite{kritikos2009mixed}, \cite{kritikos2007semantic}& yes & no & yes & no &no&prototype\\
\cite{wang2009qos} &no & no & no& yes&no&theory \\
\bottomrule
\end{tabular}
\end{minipage}
\caption{Data Types of Properties and Their Handling in Related Work} \label{tab:treReqTable}
\end{table}
The interval matching approach suggested in \cite{mukhija2007qos} determines if the property in question is high-value preferred or low-value preferred based on the values that the service requester specified: If the most preferred value of the service requester is smaller than the least
preferred value, than the property is assumed to be low-value preferred. Firstly, this sets the prerequisite that the service requester knows which values are better. However, this might not be the case. Secondly, the service requester specifies both an upper limit and a lower limit in all cases. If the property is low-value preferred, the service requester should not be prompted to specify a lower limit. Besides, the BV calculation would not work if the service requester specifies the same value for the most preferred and the least preferred values.
The trapezoidal fuzzy numbers approach \cite{bacciu2010adaptive} would deliver faulty results for low-value preferred properties and high-value preferred properties. However, it can be applied to the cases where the properties do not have broadly accepted tendencies.
\section{Constraint Models}\label{sec:cpModels}
This section presents the constraint models which address the subproblems defined in Section \ref{sec:typeqos}.
Bockmayr and Hooker \cite{cpElement} define element constraints as follows:
"$element(i, l, v)$ expresses that the $i-th$ variable in a
list of variables $l = [x 1 , . . . , x n ]$ takes the value $v, i.e., x i = v.$"
We use element constraints to model the service matchmaking problem.
\subsection{Model \#1}
Fig. \ref{fig:matrix1} illustrates the first constraint model developed in this paper. Each row in the matrix contains the values of
service specifications for the QoS property which is on the left most column.
For each row, an element constraint is defined which adds the condition:
\begin{flalign}
qvalues[i]"operator"qosdemand[i-1]
\end{flalign}
to the constraint solving problem. The operators can be adjusted from the list of available operators according to the specific purpose of the CSP
utilising the model. The first row in the qvalues matrix is the array of service ids. For this reason, the element constraints begin with the
second row. Note that the QoS request and service specifications are modelled as Java integer arrays, but not as variables, since the values for
those are fixed and the variable the model sets as unknown is the index variable.
\begin{figure*}[htpb]
\centering
\captionsetup{justification=centering,margin=2cm}
\includegraphics[width=17cm]{images/matrixGeneric.pdf}
\caption{Constraint Solving Problem Model as a Matrix}
\label{fig:matrix1}
\end{figure*}
\subsection{Model \#2}
The second constraint model takes another perspective to the service matching problem. Its main difference to
the first model is that it takes properties as JSR-331 variables whose domain is an array consisting of the values from the service specifications. This means each row in the matrix is defined as a variable:
\begin{lstlisting}[frame=single,language=Java,caption=Discrete Value Matching Model]
problem.variable(”property1”,qvalues[1]);
problem.variable(”property2”,qvalues[2]);
\end{lstlisting}
The resulting CSP searches for appropriate values of the property variables and the index variable. To ensure the integrity of a service description,
additional constraints are needed, since a Java array cannot get a JSR-331 variable as an index and the index of a Java array cannot be tracked by the
JSR-331. These constraints state that if serviceId has a certain value, the property value can have only one value from its domain, which is the
value in domain[serviceId].
\noindent\begin{minipage}{.45\textwidth}
\begin{equation}\label{eq:1}
\begin{aligned}
p:serviceId=x,\\
q:property1=domain[x]\\
p \equiv q
\end{aligned}
\end{equation}
\end{minipage}\hfill%
The relation can be expressed with logical equation since $p$ and $q$ are either both true or both false to achieve a result true.
With the element constraints, it was possible to do this without additional constraints, however element constraints are only available as hard constraints.
\section{Implementation}\label{sec:impl}
The prototype implementation uses JSR-331\cite{jsr331} with Choco Solver \cite{Choco} to implement the constraint models. Discrete value matching with hard constraints, interval matching for negative and positive
tendencies and feature list matching is realised using Model \#1. Discrete value matching with soft constraints
is realised using Model \#2. The models are described in Section \ref{sec:cpModels}. The implementation source can be found in our repository. \footnote{\url{https://github.com/TU-Berlin-SNET/cloud-service-matcher}}
Our implementation employs four methods which model the problem differently: \textit{buildModel} for exact matching with only hard constraints,
\textit{buildModelSoftAsBool} for matching with soft Boolean constraints, \textit{buildModelSoftDifference} for matching with soft constraints according to the difference
between values of the service offer and request, and buildSimpleModelDifference which is an enhanced version of buildModelSoftDifference.
\subsection{Discrete Value Matching with Hard Constraints}\label{subs:discH}
For discrete value matching with hard constraints, the implementation makes use of Model \#1 which is described in Section \ref{sec:cpModels}. Each row in the matrix contains the discrete numeric values of service specifications. For each row there is an element constraint which adds the condition qvalues[i]=qosdemand[i-1] to the CSP.
For example, indexVar=1 is in the solution set since qvalues[1][1]=2 equals to qosdemand[0].
\begin{lstlisting}[frame=single,language=Java,caption=Discrete Value Matching Model]
Var indexVar = matching
.variable("serviceIndex", 0, serviceIndexMax);
for (int j = 1; j < qosdemand.length; j++) {
matching.postElement(qvalues[j],
indexVar, "=", qosdemand[j - 1]);
}
\end{lstlisting}
For interval matching only the operator has to be changed: for properties with positive tendency $>=$ and for negative tendency $<=$.
\subsection{Discrete Value Matching with Soft Constraints}\label{sec:discreteSoft}
\textit{buildModelSoftAsBool} introduces fuzziness with the third option of variational scopes described in Section \ref{sec:prob}. It creates the negation of the element constraints defined in \textit{buildModel} described in Section \ref{subs:discH}. In contrast to hard constraints,
these are not posted, instead an optimisation objective is defined using them. If the constraint is satisfied, the constraint method returns 1, if not 0. If the service specification value is not equal to the service request value the value 1 is then multiplied by the weight for the QoS parameter that the service requester specified.
The violation is calculated for each element constraint and then added to the violation sum.
The solver returns the service index with the minimum violation sum which is the optimisation objective.
The condition checks if the service specification value exactly matches the requirement value and if not adds up to the violation sum. In some cases, this is not enough since the requester might specify an approximate value for a requirement and the results would be still fulfilling even if they are slightly different
than the requirement which is described as fuzziness with variational scope's first option above. To provide this kind of fuzziness, the optimisation objective must be the difference between the service specification value and the requirement value.
In this case, the optimisation objective can be defined as:
\begin{equation}
\begin{aligned}\label{eq:2}
i=qoSPropertyId\\
j=serviceId\\
|S_{ij}-Q_{i}|
\end{aligned}
\end{equation}
The coding experiments in \textit{buildModelSoftDifference} with element constraints as soft constraints showed that if the element constraints are not posted, then the variable \textit{serviceIndex} is not constrained, so they were not
effective. For this reason, the element constraints were removed from the problem and linear constraints were added to ensure service id and service value
bindings. In other words, for defining the
optimisation as the difference and getting consistent results, the Model \#2 was designed (see Section \ref{sec:cpModels}).
Getting only one solution is not suitable for the service matching problem, since it does not provide all, if there are equally optimal solutions.
As a workaround, the service matcher uses the CP solver to find all the solutions as a list and
orders them according to their values for violation from minimum to maximum. This way, it can be seen if there are some solutions with the same violation value
and appropriately evaluated.
\subsection{Feature List Matching}\label{sec:featureLM}
For feature list type of constraints, a constraint solving problem per constraint must be defined.
This time, the index variable shows the elements where in the feature list QoS specification the values match with the feature list QoS constraint. We create and post a new element constraint and the default solution logger lists the values for \textit{indexVar} and \textit{var} which satisfy the constraint.
The matching degree is calculated based on the size of the solution set and further explained in Section \ref{ssubs:rankingF}. This implementation handles all the items in the required list equally.
\begin{lstlisting}[frame=single,language=Java, caption= Feature List Matching Example\label{lst:featureExample} ]
String[] browsers = { "explorer", "firefox", "chrome",
"safari", "opera" };
int[] providedBrowsers = { 1, 2, 0 };
int[] requiredBrowsers = { 0, 2, 3 };
INFO:
providedIndex[1] query[2]
providedIndex[2] query[0]
matching provided browser value:0 name:explorer
matching provided browser value:2 name:chrome
\end{lstlisting}
Listing \ref{lst:featureExample} shows that the solution set has two elements: $providedBrowsers[1]=2$ and $query=2$, $providedBrowsers[2]=0$ and $query=0$. The matching degree
is calculated based on the size of the solution set and further explained in Section \ref{ssubs:rankingF}.
\subsubsection{Ranking for Feature List Matching}\label{ssubs:rankingF}
\paragraph*{Matching Degree}
The Feature List Constraint contains the number codes for a list of required items. Accordingly, the Feature List QoS Specification contains the list of number
codes that the service offers for that property.
$P$ is the set of provided browsers. $R$ is the set of requested browsers. $S$ is the set of solutions, and can be described as the intersection of provided and requested sets.
\begin{flalign}
S=P \cap R \nonumber \\
P=\emptyset \implies NOSPEC \nonumber \\
S=\emptyset \implies FAIL \nonumber \\
\left\vert{R}\right\vert > \left\vert{S}\right\vert \implies PARTIAL \\
\left\vert{R}\right\vert = \left\vert{S}\right\vert, \left\vert{P}\right\vert = \left\vert{S}\right\vert \implies EXACT \nonumber \\
\left\vert{R}\right\vert = \left\vert{S}\right\vert, \left\vert{P}\right\vert > \left\vert{S}\right\vert \implies SUPER \nonumber
\end{flalign}
An example for this type of QoS property is the list of compatible browsers.
\paragraph*{Ranking Rules}
The ranking rules define how many points a service description gets according to the matching degree of its QoS specification. These are defined in the \textit{Evaluator} classes.
For example, an \textit{ExactEvaluator} gives 2 points to the QoS specification. These rules can be changed at the corresponding Evaluator without touching other
parts of the code. Table \ref{tab:matchingdegreeF} shows the scheme for the ranking rules. At the time of writing, soft constraints calculate the violation based on the
weights and the ranking for the hard constraints add scores for each matching QoS specification independently.
The final score of a service can be calculated as
$(sum Of Scores) - (sum Of Violations)$ as also shown below.
\begin{flalign}
\sum_{i=1}^{j} s_i - \sum_{i=1}^{k} v_i \nonumber \\
j=\mbox{number of hard QoS constraints} \nonumber \\
k=\mbox{number of soft QoS constraints} \\
s_i= \mbox{score for the QoS Specification} \nonumber \\
v_i=\mbox{violation for the QoS Specification} \nonumber
\end{flalign}
\begin{table}[htpb
\begin{tabular}{llllll}
\toprule
Provided & Requested & Solutions & Matching Degree & Ranking Rules \\
\midrule
0,1,4 & 0,1 & 0,1 & SUPER & 3 points\\
0,1 & 0,1 & 0,1 & EXACT & 2 points\\
0,4 & 0,1 & 0 & PARTIAL & 1 point\\
2,3 & 0,1 & none & FAIL & 0 points\\
none & 0,1 & none & NOSPEC & 0 points\\ \bottomrule
\end{tabular}
\caption{Matching Degree Examples and Ranking Rules}\label{tab:matchingdegreeF}
\end{table}
\vspace*{-\baselineskip}
\section{Evaluation}
This section describes the goal-free comparison of our approach with other processes.
\begin{table}
\begin{tabular}{p{35mm}p{10mm}p{30mm}} \toprule
Description & Design \newline Artifact& Evaluation Method \\\midrule
process of matching with its inputs and outputs & method & goal-free comparison with other processes \cite{scriven1991evaluation} \\
our implementation of the process & instantiation & testing \\
\bottomrule
\end{tabular}
\caption{Evaluation Methods}
\label{tab:evalMethodsTable}
\end{table}
Our solution is designed analysing QoS properties in service descriptions, therefore it addresses \textit{R3:QoS Matching}. Moreover, it supports the most frequent combinations of the data types in its core matching functionality addressing all subproblems in Table \ref{tab:treReqTable}.
We address \textit{R4: Incomplete Knowledge} and \textit{R5:Fuzziness} by allowing service requesters to define both hard and soft constraints. We provide two types of soft constraints: (i)the equality of discrete values in the specification and the request as Boolean with weights also addressing \textit{R1: Service Requester Priorities}, (ii)the distance of the value in the specification to the specified value.
\textit{R2: Comprehensive Results with Matching Degrees} is addressed by the service matcher since it includes all evaluations of service specifications within the service descriptions which is easily accessible if needed. The implementation for the exact matching of two intervals is left for future
work, since the priority was feature lists due to the TRESOR project context.
This paper contributes to the constraint-based service matching methods suggested in \cite{kritikos2009mixed}, \cite{mobedpour2013user} and the approaches developed in the Dino project \cite{mukhija2007qos},\cite{bacciu2010adaptive} by developing a better picture of the properties to be matched and diagnosing various assumptions made in the definitions of the preferences of the service requesters. It supports the view that models the service matching problem as an optimisation problem and that the use of
constraint solvers is especially suitable for the implementation of soft constraints. It challenges the views which implicitly assume that the QoS properties are either low-value preferred or high-value preferred.
\section{Conclusion and Future Work}
In this paper, we analyse the requirements for service matchmaking approaches on the process and on the core functionality levels. We identify that the application context has a defining effect on how the service requester interacts with the system and how the results are further categorised. This serves as a tool to evaluate each service matchmaker in its context. On the core functionality level, we analyse the QoS properties and identify the subproblems discrete value matching, feature list matching, interval matching, and discrete value matching with soft constraints. Our prototype implementation provides solutions for all these subproblems. We suggest that the low-value, high-value, and neutral preferences for QoS properties are explicitly stated when documenting the target properties for matchmaking functionalities.
As future work, a case study might be useful to identify additional requirements in a practical context.
Moreover, we will look into the specific requirements of intercloud application brokering and extend the service matchmaker accordingly.
\section*{Acknowledgment}
This work is supported by the Horizon 2020 EU funded Integrated project
CYCLONE\footnote{cyclone-project.eu}, grant number 644925.
|
\section{Introduction} \label{sec:intro}
Searches for the annihilation products of dark matter (DM) are now testing significant portions of the theoretically motivated parameter space for weakly interacting massive particles.
The rapid progress of indirect DM searches can be attributed to a large number of astrophysical probes that have become available over the last decade \citep[reviewed by][]{Gaskins:2016cha}.
Among these, the Large Area Telescope (LAT) on board \textit{Fermi} has played a vital role due to its full-sky coverage, sensitivity, and energy range relevant for DM searches at the electroweak scale.
LAT data have been used for numerous DM searches involving a variety of astrophysical objects, including dwarf spheroidal satellite galaxies (dSphs) of the Milky Way (MW).
dSphs are especially promising targets for indirect DM searches due to their (1) substantial DM content \citep[e.g.,][]{mat98, Simon:2007dq} and proximity, (2) distribution over a range of Galactic latitudes, including regions with low diffuse foreground emission, and (3) dearth of non-thermal production mechanisms.
No {$\gamma$-ray}\xspace signal has been conclusively associated with dSphs, either individually or as a population, and the corresponding upper limits have been used to set competitive constraints on DM annihilation \citep[summarized by][]{Charles:2016pgz}.
For example, a joint analysis of 15 dSphs with 6 years of LAT data excluded DM particles annihilating at the canonical thermal relic cross section in some annihilation channels for DM masses up to 100 GeV \citep{ack15}.
Although the non-DM {$\gamma$-ray}\xspace emission from dSphs is expected to be low, no empirical measurement and few quantitative estimates for this contribution have been previously available.
dSphs have old stellar populations \citep[e.g.,][]{2012ApJ...753L..21B,2014ApJ...789..147W} and low gas content \citep{2009ApJ...696..385G,2014ApJ...795L...5S}, and therefore contain few sites for non-thermal radiation from cosmic-ray (CR) interactions.
However, their ancient stellar populations might include small populations of {$\gamma$-ray}\xspace-emitting millisecond pulsars (MSPs), which have characteristic ages of several Gyr based upon their measured spin periods and period derivatives ($\tau \equiv P / 2 \dot{P}$).
MSPs are luminous sources that account for nearly half of LAT-detected pulsars.
In addition, 25 MW globular clusters, which have similar-age stellar populations to dSphs, have been detected by the collective emission of their MSP populations \citep{abd10,Hooper:2016rap}.
MSPs exhibit hard spectral indices $\sim1.5$ and spectral cut-offs around 3~GeV \citep{abd13,cho14}.
As a consequence, their intensity peaks in the GeV range, where the LAT sensitivity is highest.
The characteristic spectral shape of MSPs is also similar to that of the Galactic Center excess, and many authors have investigated the contribution of MSPs to that signal \citep[e.g.,][]{Abazajian:2012pn,Brandt:2015,bar15,Lee:2015fea,hoo15}.
In this Letter, we estimate the conventional astrophysical emission intrinsic to dSphs, focusing on MSPs, and evaluate the potential for confusion with DM annihilation signatures at GeV energies.
\section{MSP Formation Mechanisms and the Adopted Strategy}\label{sec:formation}
MSPs are neutron stars that have been spun up to millisecond spin periods via mass accretion from a binary companion \citep{alp82}.
Different MSP formation mechanisms are thought to dominate in various stellar environments.
The classic ``primordial'' channel begins with a close stellar binary born from a single gas cloud with an extreme mass ratio between the two stars.
These binary systems remain bound after the supernova explosion of the more massive star.
In a later phase, during mass transfer from the companion star to the neutron star, the binary becomes visible as a low-mass X-ray binary (LMXB), and eventually as a MSP accompanied by a low-mass white dwarf \citep{Zwart:2011py}.
For dense systems with high stellar encounter rates, the MSP formation rate is increased by an additional mechanism: the gravitational capture of neutron stars into binary systems \citep{1998MNRAS.301...15D}.
This ``dynamical'' channel is expected to dominate in globular clusters \citep{Hui:2010vt}.
In galaxy disks, the number of LMXBs scales linearly with stellar mass of the host galaxy \citep{gil04}, suggesting that primordial formation is dominant.
dSphs also have low-density stellar environments, and therefore, it is likely that most LMXBs in dSphs have a primordial origin.
While any direct comparison between the number of LMXBs and MSPs is uncertain, due to the different lifetimes of these two evolutionary stages, it is reasonable to expect the population of field MSPs to scale with stellar mass of the host galaxy if it does so for the progenitor systems.
Thus far, \citet{2005MNRAS.364L..61M} detected five LMXBs in a deep \textit{Chandra} survey of the Sculptor dSph, implying that MSPs may also be present.
No pulsars have been found yet in Ursa Minor, Draco, or Leo~I with searches at 350~MHz using the Green Bank Radio Telescope \citep{Rubio-Herrera:2013}.
Our strategy to predict the {$\gamma$-ray}\xspace emission of MSPs in dSphs will be the following.
First, we construct the {$\gamma$-ray}\xspace luminosity function (LF) of MSPs in the MW disk (Section~\ref{sec:calculation}).
Second, we scale this LF from the stellar mass of the MW to the stellar masses of dSphs and compute the cumulative emission of the MSP population (Section~\ref{sec:dsphs}).
This approach should yield upper limits on the MSP emission because it implicitly assumes that MSPs formed within dSphs do not escape, and that the ages of MSPs in the MW disk and dSphs are similar.
\subsection{Neutron Star Escape}
Neutron stars receive a ``kick'' in the supernova events in which they originate.
Predicted kick velocity distributions vary widely in the literature: \citet{Hooper:2013nhl} and \citet[][for long lived MSPs]{Cordes:1997my} find 10--50 km s$^{-1}$, \citet{hob05} and \citet{tos99} find $\sim85\pm13$ km s$^{-1}$, while \citet{lyn98} find a higher velocity range $\sim130\pm30$ km s$^{-1}$.
If unbound, MSPs would leave these systems in $t_{\rm esc} \sim 10^8 (R/{\rm 1~kpc}) (10~{\rm km s}^{-1} /v_{\rm kick})$ yr, i.e., much shorter than the typical stellar ages of dSphs.
The kick-velocity estimates above are larger than the typical stellar velocity dispersion of dSphs, $\sim 10$ km s$^{-1}$ in dSphs \citep{mcc12}, but are comparable to the escape velocity.
In particular, the enclosed masses of classical dwarfs within 600 pc from the center are (2--7)$ \times 10^7$ M$_{\odot}$ \citep{Walker:2007ju}, corresponding to an escape velocity $v_{\rm esc}=\sqrt{2GM_{600}/r_{600}} \sim 30$ km s$^{-1} \sim v_{\rm kick}$.
This suggests that a fraction of MSPs should be retained within the virial radii of dSphs, which are assumed to be at kpc distances with a contained mass of $\sim10^{8-9}$ M$_{\odot}$ \citep[see also the calculations of ][]{Dehnen:2005fs}.\footnote{MSPs with larger kick velocities would attain higher-eccentricity orbits and this may provide a means to constrain the mass profile of dSphs beyond the radii of their normal stellar populations.}
An MSP orbiting at 600~pc from the center of a dSph located 80~kpc away would span an apparent $0\fdg4$ angular size, which is within the 95\% confinement radius of LAT \texttt{Pass8 FRONT} events at 2~GeV.\footnote{\url{https://www.slac.stanford.edu/exp/glast/groups/canda/lat\_Performance.htm}}
Depending on the distance to a given dSph and the photon event class used for the analysis, a slight signal extension might be present.
\subsection{Stellar Population Comparison}
Star formation in most MW dSphs ended several Gyr ago \citep[e.g.,][]{2012ApJ...753L..21B,2014ApJ...789..147W}, whereas the MW disk remains active.
The formation of LMXBs, and accordingly MSPs, is thought to peak most strongly in the first $0.5$~Gyr after star formation \citep{fra08}, suggesting that MSPs in dSphs are potentially older and less luminous \citep{Hooper:2016rap} than the average Galactic MSP.
However, other factors might increase the MSP population in dSphs.
Neutron star production could be enhanced in low metallicity systems relative to the Galactic field by $\sim20$\% \citep{Ivanova:2007bu}.
Also, two dSphs host their own globular clusters: five in Fornax \citep{2016A&A...590A..35D} and one in Eridanus~II \citep{crn16}.
The abundance of MSPs might be higher in these systems due to more frequent stellar encounters.
\section{Derivation of a Stellar-Mass-normalized MSP Luminosity Function}\label{sec:calculation}
\subsection{MSP Sample}\label{sec:sample}
\begin{figure*}
\includegraphics[width=.5\textwidth]{f1.pdf}
\includegraphics[width=.5\textwidth]{f2.pdf}
\caption{\textit{Left}: Luminosity (0.1--100~GeV) distribution of LAT detected MSPs calculated for 2PC and ATNF reported distances. \textit{Right}: Estimated LAT survey completeness fraction as a function of {$\gamma$-ray}\xspace luminosity (0.1--100~GeV). The red curve and gray band indicate the median and inner 68\% interval of the completeness fraction across the MC realizations. \label{fig:complete}}
\end{figure*}
Our pulsar sample consists primarily of MSPs in the Second LAT Pulsar Catalog \citep[2PC;][]{abd13}, although all LAT-detected MSPs were considered.\footnote{\url{https://confluence.slac.stanford.edu/display/GLAMCOG/Public+List+of+LAT-Detected+Gamma-Ray+Pulsars}\label{fn:lat_msps}}
Pulsars residing within globular clusters, as well as those lacking distance measurements, were both excluded, leaving 66 MSPs in our sample.
The luminosities of Galactic MSPs often have large uncertainties because accurate distance measurement is difficult.
Distance measurements based on dispersion measure suggest that for roughly 75\% of the directions in the sky the accuracy is no better than a factor of 1.5 to 2 \citep{sch12}.
We compare luminosities calculated with distances from 2PC and from the Australian Telescope National Facility Pulsar Catalog \citep[ATNF;][]{man05}\footnote{\url{http://www.atnf.csiro.au/people/pulsar/psrcat/}} in Figure \ref{fig:complete}.
\subsection{Incompleteness Correction}\label{sec:completeness}
As a first step towards constructing the Galactic MSP LF we perform a Monte Carlo (MC) incompleteness correction and apply it to our MSP sample.
In our model, we assume an exponential spatial distribution for MSPs in the MW disk:
\begin{equation}
\rho(R,z)\propto e^{(-R/R_0)}e^{(-|z|/z_0)},
\label{eqn:mw_dist}
\end{equation}
\noindent where $R$ is the radial distance from the Galactic center and $z$ is the vertical scale height above the Galactic plane.
While other authors model the radial distribution of Galactic MSPs using a Gaussian density profile \citep{fau10}, \citet{gre13} find that their results are fairly insensitive to the selected radial law.
We make no attempt to model a special population of MSPs in the Galactic bulge; our population of interest is field MSPs.
To account for systematic uncertainties both in the spatial distribution of Galactic MSPs, as well as the effective selection function of the LAT pulsar survey, we repeat the following MC procedure many times with different input parameter sets.
First, following the MSP population model of \citet{gre13}, we draw radial and vertical scale lengths from log-normal distributions defined by $R_0=3^{+3}_{-1}$~kpc and $z_0=0.6^{+0.6}_{-0.3}$~kpc, respectively (see their Table~3).
Using these scale lengths, we generate $10^7$ MSPs at random locations consistent with the spatial distribution of Equation \ref{eqn:mw_dist} and assuming azimuthal symmetry in Galactocentric coordinates.
We then determine the apparent Galactic coordinates $(l, b)$ of each MSP as viewed from the Sun's position at $(R,\phi,z) = (8.5\text{ kpc},0\text{ rad}, 20\text{ pc})$.
Each MSP $(l,b)$ coordinate is mapped to an effective flux detection threshold using the sensitivity curve in Figure~17 of 2PC \citep{abd13}, which is expressed as a function of Galactic latitude.
This direction-dependent flux threshold partially accounts for variations in the intensity of diffuse Galactic emission.
We model systematic uncertainty in the 2PC selection function by drawing a direction-dependent flux threshold curve between the 10\% and 90\% percentile sensitivity range for each realization of the MSP population (sampled from a uniform distribution).
The survey completeness is evaluated as the number of detectable MSPs at a given luminosity divided by the total number of simulated MSPs.
After calculating the detection efficiency for $10^5$ sets of spatial parameters and flux detection thresholds, we arrive at the completeness function shown in Figure \ref{fig:complete}.
\subsection{{$\gamma$-ray}\xspace Luminosity Function}\label{sec:lumfunction}
\begin{figure}
\centering
\includegraphics[width=.5\textwidth]{f3.pdf}
\caption{MSP {$\gamma$-ray}\xspace luminosity function (0.1--100~GeV) normalized to the stellar mass of the MW.
Data points include a incompleteness correction applied to LAT MSP sample assuming 2PC distances.
Error bars correspond to statistical uncertainty associated with the finite number of LAT-detected MSPs.
The shaded gray band represents the $1\sigma$ statistical uncertainty on the broken power-law fit to these data and dashed gray lines represent the systematic uncertainty envelope (distances to LAT-detected MSPs, spatial distribution of Galactic MSPs, and effective selection function of LAT pulsar catalog). The blue dashed curve represents the best-fit LF of \citet{hoo15} normalized to the cumulative luminosity of our best-fit LF.
\label{fig:l2dndl}}
\end{figure}
We derive the MSP {$\gamma$-ray}\xspace LF using a Markov Chain Monte Carlo (MCMC) assuming a broken power law parametrization.
During the fit, the median incompleteness correction (Section~\ref{sec:completeness}) is applied to the MSPs in each luminosity bin of our sample.
After determining the LF posterior, we normalize to unit stellar mass, assuming a MW stellar mass of $7\times10^{10}M_\sun$ \citep{mal96}.
The resultant LF is shown in Figure \ref{fig:l2dndl}.
To estimate the systematic uncertainty, we re-fit the LF pairing luminosities calculated from 2PC and ATNF distances with the upper and lower limits of the incompleteness correction.
These four bracketing cases establish the upper and lower systematic envelope on the LF.
At high luminosity, uncertainty on the LF is driven by the small number of extremely luminous Galactic MSPs, since the LAT census of such sources is expected to be largely complete.
Systematic uncertainty dominates at low luminosity, due in part to the large and imperfectly known incompleteness correction that must be applied.
Our model predicts that the main contribution comes from MSPs with luminosities $\sim10^{33}~\text{erg}~\text{s}^{-1}$.
Both our incompleteness correction and LF are in reasonable agreement with \citet{hoo15}; their best-fit LF for field MSPs peaks around $3 \times 10^{33}~\text{erg}~\text{s}^{-1}$ in $L_{\gamma}^2 dN/dL_{\gamma}$.
\section{Cumulative MSP Emission Towards dSphs}\label{sec:dsphs}
Our target sample, summarized in Table~\ref{tbl:dsph_flux}, includes both spectroscopically confirmed dSphs and recently reported dSph candidates, which we collectively refer to as ``dSphs" for simplicity.
Since our model requires the stellar mass of the host galaxy as an input, we include only targets with published stellar masses.
We first discuss the emission from MSPs within dSphs, followed by the MW foreground contribution.
\subsection{Internal MSP Emission of dSphs}
\begin{figure*}
\centering
\includegraphics[width=1.\textwidth]{f4.pdf}
\caption{Predicted flux distributions illustrating the effect of Poisson fluctuations (i.e., ``shot noise'') in the number of MSPs towards four representative dSphs.
Green histograms represent the contribution from MSPs within each dSph, and blue histograms represent the contribution from MW foreground MSPs along the same line of sight (integrated within a $1\fdg$ radius of the dSph location).
Bins labeled ``underflow'' represent the fraction of realizations without a MSP flux contribution (no MSPs with $L_{0.1-100~\rm{GeV}}>10^{31}~\text{erg}~\text{s}^{-1}$).
For comparison, the dashed black vertical line indicates the typical LAT sensitivity at high Galactic latitudes ($2\sigma$ upper limit).
\label{fig:pois}}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.\textwidth]{f5.pdf}
\caption{Same as Figure~\ref{fig:pois}, but representing predicted MSP flux distributions with complementary cumulative distribution functions.
\label{fig:survival}}
\end{figure*}
For each of the 30 dSphs in our sample, we scale the LF of Galactic MSPs (Section~\ref{sec:lumfunction}) to the dSph stellar mass.
The expectation value for the cumulative MSP luminosity is found by integrating the LF, $dN/dL_\gamma$:
\begin{equation}
\langle L_{\gamma,{\rm MSP}} \rangle = \frac{M_{*, {\rm dSph}}}{M_{*, {\rm MW}}} \int_{L_{\rm min}}^{L_{\rm max}} L_\gamma \frac{dN}{dL_\gamma} dL_\gamma,
\end{equation}
where $L_{\rm min}=10^{31}~\text{erg}~\text{s}^{-1}$ and $L_{\rm max}=10^{36}~\text{erg}~\text{s}^{-1}$.
dSphs have small enough stellar populations that the number of luminous MSPs in a particular dSph may exhibit substantial ``shot noise''.
To model this effect, we evaluate the cumulative MSP flux including Poisson fluctuations in the number of MSPs within multiple luminosity intervals.
We repeat this process $10^4$ times for each dSph to construct a PDF of the expected flux.
As shown in Figures~\ref{fig:pois} and \ref{fig:survival}, the spread in cumulative MSP flux for a low stellar mass dSph, such as Segue 1, spans several order of magnitude, whereas high stellar mass dSphs, such as Fornax and Sculptor, have sharply peaked flux PDFs.
The cumulative MSP flux in an ultra-faint dSph is likely dominated by a handful of luminous sources, but the relative impact of such statistical variations diminishes with increasing stellar mass.
\subsection{Galactic Foreground MSP Emission}
Using the same MW MSP population model discussed in Section~\ref{sec:sample}, we also estimate the probability of confusion due to the chance alignment of foreground MSPs along the same sightlines to the target dSphs.
We begin by randomly drawing luminosities from the best-fit LF (normalized to the stellar mass of the MW).
Next, the luminosity values are assigned to random MSP locations within the MW according to the fiducial spatial model described in Section~\ref{sec:completeness}.
We then compute the cumulative flux within a $1\fdg0$ radius (comparable to the LAT PSF for 1~GeV {$\gamma$ rays}\xspace) around the location of each dwarf in our sample.
The resulting foreground MSP flux distributions for four representative dSphs are shown in Figures~\ref{fig:pois} and \ref{fig:survival}.
\subsection{Results and Discussion}\label{sec:results}
For each of the 30 dSphs in our sample, we compute the expectation value and uncertainty on the cumulative MSP flux by adding in quadrature
\begin{enumerate}
\item statistical uncertainty due to the finite number of LAT-detected Galactic MSPs,
\item systematic uncertainty in the luminosities of Galactic MSPs due to distance uncertainties,
\item systematic uncertainty in the detection efficiency of Galactic MSPs due to their unknown spatial distribution and the effective flux threshold of the LAT pulsar catalog, and
\item uncertainty in the stellar mass of the host dSph.
\end{enumerate}
\noindent We then compute flux upper limits taking into account Poisson fluctuations in the number of luminous MSPs (using the best-fit LF). While stellar mass uncertainty is dominant for select dSphs, we find that systematic uncertainty, together with Poisson fluctuations, have the greatest impact on the predicted signals from both ultra-faint and classical dSPhs. The results of these calculations, summarized in Table~\ref{tbl:dsph_flux}, suggest that high stellar mass dSphs are likely to host modest MSP populations.
Consistent with our initial assumptions, we find the mean number of predicted MSPs to be higher than a few in all dSphs where LMXBs or LMXB candidates have been detected.
However, even for the largest classical dSph, Fornax, the predicted MSP flux $> 500$~MeV is $6.0^{+9.7}_{-4.4}\times10^{-12}$~ph~cm$^{-2}$~s$^{-1}$, which is about an order of magnitude below the typical flux upper limits obtained at high Galactic latitudes after six years of the LAT survey, $\sim10^{-10}$ ph cm$^{-2}$ s$^{-1}$ \citep{ack15}.
Perhaps more importantly, Draco and Ursa Minor, two classical dSphs with large and well constrained J-factors \citep[e.g.,][]{Geringer-Sameth:2014yza}, have estimated MSP fluxes another order of magnitude fainter than Fornax.
The MSP emission of ultra-faint dSphs, such as Segue 1, is typically below $10^{-13}$~ph~cm$^{-2}$~s$^{-1}$ and likely beyond the reach of {$\gamma$-ray}\xspace telescopes in the foreseeable future.
To more easily compare {$\gamma$-ray}\xspace signals from MSPs and DM annihilation, in Figure~\ref{fig:sum_fig} we show the expected MSP flux versus J-factor for the host dSph.
The potential for confusion is greatest between models with similar {$\gamma$-ray}\xspace spectra, so we consider a DM model that produces a {$\gamma$-ray}\xspace energy flux peaking at $\sim1$~GeV, namely $\chi\chi \rightarrow b \bar{b}$ with a DM particle mass of 25 GeV.
This DM model is also compatible with DM interpretations of the Galactic Center excess \citep[e.g.,][and references therein]{Calore:2014xka}.
In addition to being below the current LAT sensitivity threshold, we predict that MSP emission in each of the dSphs is also below the annihilation signal expected from this DM model for an annihilation cross section at the thermal relic value $\langle \sigma v \rangle = 3 \times 10^{-26} {\rm cm}^3 {\rm s}^{-1}$.
Figure~\ref{fig:sum_fig} also demonstrates that the GeV emission from MSPs and from DM annihilation follow different scaling relations across the dSph population.
The targets with the largest J-factors are mainly nearby ultra-faint dSphs with $M_{*} < 10^{4} M_{\odot}$, whereas we argue that dSphs with $M_{*} \sim 10^{7} M_{\odot}$ are most likely to host detectable {$\gamma$-ray}\xspace MSP populations.
This distinction offers another means to distinguish a putative DM signal from conventional astrophysical sources (e.g., using a joint-likelihood analysis).
We find that Galactic foreground MSP emission is potentially a non-negligible source of confusion for a small fraction of dSphs, even at the current LAT sensitivity.
However, for a typical target, there is no foreground MSP with luminosity exceeding $L_{0.1-100~\rm{GeV}}=10^{31}~\text{erg}~\text{s}^{-1}$ in $\approx90\%$ of model realizations.
For dSphs with $M_{*} > 10^{4} M_{\odot}$, the contribution from MSPs within the dSph is predicted to exceed the foreground in most cases (e.g., Fornax, Sculptor, Draco).
Otherwise, the MSP emission internal the dSph is likely to be sub-dominant (e.g., Segue~1).
Several LAT analyses have already employed empirical background modeling methods to partially account for sub-threshold point sources coincident with dSphs \citep[e.g.,][]{2011PhRvL.107x1303G,Ackermann:2013yva,Carlson:2014nra}.
\begin{figure*}
\centering
\includegraphics[width=1.\textwidth]{f6.pdf}
\caption{Expected {$\gamma$-ray}\xspace flux versus J-factor.
Blue points indicate expectation values for the predicted MSP emission in 30 MW dSphs and dSph candidates.
Blue vertical error bars show 3$\sigma$ upper limits on the predicted flux due to Poisson fluctuations alone.
Additional uncertainty is attributed to uncertainty on the LF of MW MSPs (see Section~\ref{sec:results}).
This contribution spans the same logarithmic interval for all targets due to the assumed linear scaling with stellar mass, and is represented by the rightmost set of black error bars.
J-factor uncertainties are shown for kinematically confirmed dSphs only. The predicted strengths of the MSP and DM annihilation (red line) signals have different dependence on the J-factor. This distinction offers another means to distinguish a putative DM signal from conventional astrophysical sources (e.g., using a joint-likelihood analysis). The gray shaded band represents the typical $2\sigma$ upper limit derived in high Galactic latitude blank fields after 6 years of the LAT survey. The red curve represents a DM annihilation model that is consistent with both DM interpretations of the Galactic Center Excess and the characteristic spectral shape of MSPs. \label{fig:sum_fig}}
\end{figure*}
\section{Cosmic-ray Induced Emission in dSphs}
The non-thermal emission in dSphs related to CR interactions is highly uncertain due to our present lack of knowledge regarding the acceleration and confinement of CRs in such environments.
Still, it is possible to obtain an estimate using the quasi-linear scaling relation between {$\gamma$-ray}\xspace luminosity and tracers of the star formation rate \citep[SFR;][]{Ackermann:2012vca}.
Considering the most massive dSph, Fornax \citep[${\rm SFR} \sim 10^{-4}$~M$_{\odot}$~yr$^{-1}$;][]{2012A&A...544A..73D}, and using the SMC as the nearest point of comparison with $L_{0.1-100~{\rm GeV}} \sim 1 \times 10^{37}$~erg~s$^{-1}$ and ${\rm SFR} \sim 10^{-1}$~M$_{\odot}$~yr$^{-1}$ \citep{2004AJ....127.1531H}, the predicted CR-induced luminosity of Fornax is $L_{0.1-100~{\rm GeV}} \sim 1 \times 10^{34}$~erg~s$^{-1}$.
This estimate is comparable to the expected MSP luminosity, but should be regarded as an upper limit considering the low gas content of dSphs \citep[${\rm M}_{\rm HI} / {\rm M}_{*} \lesssim 10^{-2}$;][]{2014ApJ...795L...5S} compared to the SMC \citep[${\rm M}_{\rm HI} / {\rm M}_{*} \sim 1$;][]{mcc12}.
In addition, the interval between CR acceleration events in dSphs might be comparable to the CR confinement time, in which case the CR population may not reach an equilibrium state.
Following \citet{2015ApJ...805L...2C}, the rates of core-collapse and type Ia supernovae for a galaxy with a constant SFR are $\approx 10^{-2}$~yr$^{-1}$ and $\approx 10^{-13}$~yr$^{-1}$~M$_{\odot}^{-1}$, respectively.
Applying these rates to Fornax \citep[$M_{*} \approx 2 \times 10^{7}$~M$_{\odot}$;][]{mcc12}, the expected supernova rate is $\sim1$~Myr$^{-1}$.
By comparison, the CR confinement time in the (much larger) MW at GeV energies is $\sim10^{7}$~yr.
\citet{Chen:2015zcw} proposed that afterglows from tidal disruption events might also enhance {$\gamma$-ray}\xspace emission in dSphs, although the duty cycle of such events is likely to be low.
\section{Conclusions} \label{sec:summary}
dSphs have commonly been regarded as astrophysically ``clean'' targets for which the detection of excess {$\gamma$ rays}\xspace would constitute compelling evidence for particle DM.
However, conventional astrophysical emission must be present at some level, and in this Letter we predict the contribution from MSPs.
Under the assumption that MSPs in both the MW disk and in dSphs originate mainly from primordial binary systems (in contrast with globular clusters), we scale the LF of Galactic MSPs to the stellar masses of dSphs to quantify their MSP populations.
We estimate that MSP emission within the highest stellar mass dSphs, Fornax and Sculptor, is a factor $\sim10$ below the current LAT sensitivity threshold (Figure~\ref{fig:sum_fig}).
The MSP emission within ultra-faint dSphs (including targets with the largest J-factors) is several orders of magnitude lower.
Moreover, for a DM particle of mass 25~GeV annihilating to $b$ quarks at the thermal relic cross section (consistent with DM interpretations of the Galactic Center excess), the expected {$\gamma$-ray}\xspace emission due to DM exceeds that of the MSP population in all of the dSphs considered here.
At the current LAT sensitivity, the more likely source of confusion is a Galactic foreground MSP along the same line of sight, although the probability of this alignment is typically $\lesssim10\%$ per target dSph (Figure~\ref{fig:survival}).
The LAT sensitivity to DM annihilation in dSphs is anticipated to improve by nearly an order of magnitude over the coming decade due to increasing LAT exposure, more precise J-factor measurements from deep spectroscopy, and additional dSph targets discovered in optical surveys such as LSST \citep{Charles:2016pgz}.
These forecasts are based on a combined likelihood analysis weighting dSphs by their J-factors, as appropriate for DM searches.
Since we do not expect that many nearby and high stellar mass dSphs remain to be discovered, the sensitivity to the MSP contribution may not improve as quickly as for DM signals.
\section{Acknowledgements}
We are especially indebted to Andy Strong for the insights provided by his GALPLOT code runs. We acknowledge helpful discussions with Andrea Albert, Brandon Anderson, John Beacom, Alessandro Bressan, Alessandro Cuoco, Alex Drlica-Wagner, German Arturo Gomez-Vargas, Dan Hooper, Matthew Kerr, Tim Linden, Dimitry Malyshev, Nestor Mirabal, Pasquale Serpico, David Smith, and Piero Ullio.
We thank the anonymous referee for constructive suggestions.
\begin{table*}
\centering
\caption{Estimated GeV Contribution from MSPs in MW dSphs \label{tbl:dsph_flux}}
\begin{tabular}{l l c c c c c c c c c c}
\tableline\tableline
Galaxy\tablenote{dSphs marked with a star, $^*$, contain globular clusters. A dagger, $^\dagger$, indicates that LMXB or LMXB candidates have been detected.} & D(kpc) & $(l,b)$ & $\log_{10}\left(\frac{M_*}{M_\odot}\right)$ & \multicolumn{5}{c}{\underline{$\quad{}\quad{}\log_{10}\left(\text{Flux} >500 \text{MeV} \left[\frac{\text{ph}}{\text{cm}^{2} \text{s}^{1}}\right]\right)\tablenote{Column 1: flux expectation value calculated from the best-fit LF. The uncertainty on the expectation value is computed by adding in quadrature the $1\sigma$ statistical ($^{+0.18}_{-0.12}$ dex) and systematic ($^{+0.67}_{-0.26}$ dex) uncertainty on the LF together with stellar mass uncertainty. Columns 2, 3: 2$\sigma$ and $3\sigma$ upper limits on the predicted flux due to Poisson fluctuations in the number of MSPs. Columns 4, 5: 1$\sigma$ and 2$\sigma$ upper limits on the expected flux from MW foreground MSPs ($>10^{31}$~erg~s$^{-1}$) along the line of sight (integrated within a $1\fdg0$ radius of the dSph location). A dash, $-$, denotes zero predicted MSPs with $L_{0.1-100~{\rm GeV}} > 10^{31}$~erg~s$^{-1}$ at the given confidence level.}\quad{}\quad{}$}} & $\langle\text{N}_{\rm MSP}\rangle$\tablenote{Predicted number of MSPs with $L_{0.1-100~{\rm GeV}} > 10^{31}$~erg~s$^{-1}$ computed from the best-fit LF} & $\log_{10}\left(\text{J} \left[\frac{\text{GeV}^2}{\text{cm}^5}\right]\right)$\tablenote{For dSphs that are kinematically confirmed to be DM dominated, measured J-factors with uncertainties are quoted from \citet{ack15}. For the remaining targets, we use a predicted J-factor following the distance scaling relation proposed by \citet{drl15}.} & Ref. \\[1ex]
& & & & Mean & \multicolumn{2}{c}{Internal} & \multicolumn{2}{c}{Foreground} & & & \\
& & & & & $2\sigma$ & $3\sigma$ & $1\sigma$ & $2\sigma$ & & &\\[.5ex]
\tableline
Segue 1 &$ 23 $&$( 220 , 50 )$&$ 2.53 _{- 0.20 }^{+ 0.38 }$&$ -14.47 _{- 0.35 }^{+ 0.79 }$&$ - $&$ -12.57 $&$ - $&$ -12.12 $&$ 0.004 $&$ 19.5 \pm 0.29 $& 1,5 \\
Tucana III &$ 25 $&$( 315 , -56 )$&$ 2.90 _{- 0.05 }^{+ 0.05 }$&$ -14.18 _{- 0.29 }^{+ 0.69 }$&$ - $&$ -12.08 $&$ -13.4 $&$ -11.74 $&$ 0.009 $&$ 19.3 $& 4 \\
Reticulum II &$ 32 $&$( 266 , -50 )$&$ 3.41 _{- 0.03 }^{+ 0.30 }$&$ -13.88 _{- 0.29 }^{+ 0.69 }$&$ -13.45 $&$ -11.74 $&$ -13.85 $&$ -11.88 $&$ 0.03 $&$ 19.3 $& 3,6 \\
Ursa Major II &$ 32 $&$( 152 , 37 )$&$ 3.73 _{- 0.23 }^{+ 0.23 }$&$ -13.56 _{- 0.37 }^{+ 0.73 }$&$ -12.86 $&$ -11.53 $&$ -13.76 $&$ -12.00 $&$ 0.06 $&$ 19.3 \pm 0.28 $& 2,5 \\
Willman I &$ 38 $&$( 159 , 57 )$&$ 3.00 _{- 0.22 }^{+ 0.39 }$&$ -14.44 _{- 0.36 }^{+ 0.79 }$&$ - $&$ -12.23 $&$ - $&$ -12.19 $&$ 0.01 $&$ 19.1 \pm 0.31 $& 1,5 \\
Coma Ber. &$ 44 $&$( 242 , 84 )$&$ 3.68 _{- 0.22 }^{+ 0.22 }$&$ -13.89 _{- 0.36 }^{+ 0.73 }$&$ -13.22 $&$ -11.83 $&$ - $&$ -12.32 $&$ 0.05 $&$ 19.0 \pm 0.25 $& 2,5 \\
Tucana IV &$ 48 $&$( 313 , -55 )$&$ 3.34 _{- 0.06 }^{+ 0.08 }$&$ -14.30 _{- 0.29 }^{+ 0.69 }$&$ -13.89 $&$ -12.13 $&$ -13.34 $&$ -11.83 $&$ 0.02 $&$ 18.7 $& 4 \\
Grus II &$ 53 $&$( 351 , -52 )$&$ 3.53 _{- 0.05 }^{+ 0.04 }$&$ -14.20 _{- 0.29 }^{+ 0.69 }$&$ -13.65 $&$ -12.09 $&$ -12.95 $&$ -11.88 $&$ 0.04 $&$ 18.7 $& 4 \\
Tucana II &$ 58 $&$( 328 , -52 )$&$ 3.48 _{- 0.14 }^{+ 1.01 }$&$ -14.33 _{- 0.32 }^{+ 1.22 }$&$ -13.81 $&$ -12.16 $&$ -13.08 $&$ -11.81 $&$ 0.03 $&$ 18.8 $& 3,6 \\
Bootes I &$ 66 $&$( 358 , 70 )$&$ 4.45 _{- 0.06 }^{+ 0.09 }$&$ -13.47 _{- 0.29 }^{+ 0.70 }$&$ -12.46 $&$ -11.74 $&$ - $&$ -12.06 $&$ 0.3 $&$ 18.2 \pm 0.22 $& 1,5 \\
Indus I &$ 69 $&$( 347 , -42 )$&$ 2.90 _{- 0.22 }^{+ 0.22 }$&$ -15.06 _{- 0.36 }^{+ 0.72 }$&$ - $&$ -12.94 $&$ -12.58 $&$ -11.57 $&$ 0.009 $&$ 18.3 $& 3,6 \\
Draco\tablenotemark{$\dagger$} &$ 76 $&$( 86 , 35 )$&$ 5.51 _{- 0.10 }^{+ 0.10 }$&$ -12.53 _{- 0.30 }^{+ 0.70 }$&$ -11.87 $&$ -11.32 $&$ -13.61 $&$ -11.86 $&$ 3.7 $&$ 18.8 \pm 0.16 $& 2,5 \\
Ursa Minor &$ 76 $&$( 105 , 45 )$&$ 5.73 _{- 0.20 }^{+ 0.20 }$&$ -12.31 _{- 0.35 }^{+ 0.72 }$&$ -11.73 $&$ -11.17 $&$ -12.97 $&$ -11.91 $&$ 6.1 $&$ 18.8 \pm 0.19 $& 2,5 \\
Sculptor\tablenotemark{$\dagger$} &$ 86 $&$( 288 , -83 )$&$ 6.59 _{- 0.21 }^{+ 0.21 }$&$ -11.56 _{- 0.36 }^{+ 0.72 }$&$ -11.22 $&$ -10.78 $&$ - $&$ -12.13 $&$ 44.2 $&$ 18.6 \pm 0.18 $& 2,5 \\
Sextans &$ 86 $&$( 244 , 42 )$&$ 5.84 _{- 0.20 }^{+ 0.20 }$&$ -12.31 _{- 0.35 }^{+ 0.72 }$&$ -11.76 $&$ -11.20 $&$ -13.63 $&$ -11.86 $&$ 7.9 $&$ 18.4 \pm 0.27 $& 2,5 \\
Horologium I &$ 87 $&$( 271 , -55 )$&$ 3.38 _{- 0.13 }^{+ 0.25 }$&$ -14.78 _{- 0.31 }^{+ 0.74 }$&$ -14.34 $&$ -12.55 $&$ -14.11 $&$ -11.87 $&$ 0.03 $&$ 18.4 $& 3,6 \\
Reticulum III &$ 92 $&$( 274 , -46 )$&$ 3.30 _{- 0.15 }^{+ 0.13 }$&$ -14.91 _{- 0.32 }^{+ 0.70 }$&$ - $&$ -12.68 $&$ -13.39 $&$ -11.77 $&$ 0.02 $&$ 18.2 $& 4 \\
Phoenix II &$ 95 $&$( 324 , -60 )$&$ 3.45 _{- 0.11 }^{+ 0.19 }$&$ -14.79 _{- 0.31 }^{+ 0.72 }$&$ -14.34 $&$ -12.63 $&$ -13.57 $&$ -11.98 $&$ 0.03 $&$ 18.4 $& 3,6 \\
Ursa Major I &$ 97 $&$( 159 , 54 )$&$ 4.28 _{- 0.13 }^{+ 0.13 }$&$ -13.97 _{- 0.31 }^{+ 0.70 }$&$ -13.00 $&$ -12.21 $&$ - $&$ -12.39 $&$ 0.2 $&$ 18.3 \pm 0.24 $& 2,5 \\
Carina &$ 105 $&$( 260 , -22 )$&$ 5.63 _{- 0.09 }^{+ 0.11 }$&$ -12.69 _{- 0.30 }^{+ 0.70 }$&$ -12.07 $&$ -11.53 $&$ -12.47 $&$ -11.72 $&$ 4.8 $&$ 18.1 \pm 0.23 $& 1,5 \\
Hercules &$ 132 $&$( 29 , 37 )$&$ 4.57 _{- 0.14 }^{+ 0.14 }$&$ -13.95 _{- 0.32 }^{+ 0.71 }$&$ -12.98 $&$ -12.31 $&$ -12.47 $&$ -11.61 $&$ 0.4 $&$ 18.1 \pm 0.25 $& 2,5 \\
Fornax\tablenotemark{$\dagger$*} &$ 147 $&$( 237 , -66 )$&$ 7.39 _{- 0.14 }^{+ 0.14 }$&$ -11.22 _{- 0.32 }^{+ 0.70 }$&$ -11.04 $&$ -10.74 $&$ - $&$ -12.13 $&$ 278.7 $&$ 18.2 \pm 0.21 $& 2,5 \\
Leo IV &$ 154 $&$( 265 , 57 )$&$ 3.93 _{- 0.15 }^{+ 0.15 }$&$ -14.72 _{- 0.32 }^{+ 0.71 }$&$ -13.88 $&$ -12.79 $&$ - $&$ -11.97 $&$ 0.1 $&$ 17.9 \pm 0.28 $& 2,5 \\
Canes Ven. II &$ 160 $&$( 114 , 83 )$&$ 3.90 _{- 0.20 }^{+ 0.20 }$&$ -14.79 _{- 0.35 }^{+ 0.72 }$&$ -13.94 $&$ -12.83 $&$ - $&$ -12.26 $&$ 0.09 $&$ 17.9 \pm 0.25 $& 2,5 \\
Columba I &$ 182 $&$( 232 , -29 )$&$ 3.79 _{- 0.07 }^{+ 0.13 }$&$ -15.01 _{- 0.29 }^{+ 0.70 }$&$ -14.24 $&$ -13.04 $&$ -13.09 $&$ -11.78 $&$ 0.07 $&$ 17.6 $& 4 \\
Indus II &$ 214 $&$( 354 , -37 )$&$ 3.69 _{- 0.14 }^{+ 0.16 }$&$ -15.25 _{- 0.32 }^{+ 0.71 }$&$ -14.55 $&$ -13.21 $&$ -12.37 $&$ -11.57 $&$ 0.06 $&$ 17.4 $& 4 \\
Canes Ven. I &$ 218 $&$( 74 , 80 )$&$ 5.48 _{- 0.09 }^{+ 0.09 }$&$ -13.48 _{- 0.30 }^{+ 0.70 }$&$ -12.81 $&$ -12.24 $&$ - $&$ -11.98 $&$ 3.4 $&$ 17.7 \pm 0.26 $& 2,5 \\
Leo II &$ 233 $&$( 220 , 67 )$&$ 6.07 _{- 0.13 }^{+ 0.13 }$&$ -12.94 _{- 0.31 }^{+ 0.70 }$&$ -12.46 $&$ -11.95 $&$ - $&$ -12.37 $&$ 13.3 $&$ 17.6 \pm 0.18 $& 2,5 \\
Leo I\tablenotemark{$\dagger$} &$ 254 $&$( 226 , 49 )$&$ 6.69 _{- 0.13 }^{+ 0.13 }$&$ -12.40 _{- 0.31 }^{+ 0.70 }$&$ -12.08 $&$ -11.68 $&$ - $&$ -12.20 $&$ 55.6 $&$ 17.7 \pm 0.18 $& 2,5 \\
Eridanus II\tablenotemark{*} &$ 330 $&$( 250 , -52 )$&$ 4.92 _{- 0.07 }^{+ 0.09 }$&$ -14.40 _{- 0.29 }^{+ 0.70 }$&$ -13.54 $&$ -12.91 $&$ -14.96 $&$ -11.86 $&$ 0.9 $&$ 17.3 $& 3,6 \\[1ex]
\tableline\tableline
\end{tabular}
\tablerefs{(1) \citet{wol10}, (2) \citet{kir13}, (3) \citet{bec15}, (4) \citet{drl15_2}, (5) \citet{ack15}, (6) \citet{drl15}}
\end{table*}
|
\section{Introduction}\label{intro}
An equivalent form of the L{\'e}vy-Leblond equation \cite{LevyLeblond:1967zz} was proposed in \cite{Ajaib:2015uha} and it was shown to be consistent with standard quantum mechanical results. The L{\'e}vy-Leblond equation is the analogue of the Dirac equation and describes spin 1/2 particles in the non-relativistic limit. In references \cite{Ajaib:2015uha} and \cite{Ajaib:2015eer} it was shown that the equivalent form of L{\'e}vy-Leblond equation can be employed to solve the step potential problem and the finite potential barrier problem. It was also shown that this equation is the non-relativistic limit of the Dirac equation and the Pauli Hamiltonian can be obtained from this equation by requiring it to be locally invariant.
In this paper, we present this equation with two dimensional nilpotent matrices and derive it from the (2+1) dimensional Dirac equation. We further illustrate its applications by solving it for a Coulomb potential in (3+1) dimensions when the nilpotent matrices are 4 dimensional. We show that the known expression for the quantized energy levels of the hydrogen atom is obtained from this equation. {The novelty of the approach employed herein is that the spectrum of the hydrogen atom is derived from the L{\'e}vy-Leblond equation which takes into account the spin of the particle in the non-relativistic limit.}
The paper is organized as follows: In section \ref{sec:2dmatrices} we discuss the equivalent form of the L{\'e}vy-Leblond equation in (1+1) dimensions and (2+1) dimensions when the nilpotent matrices are two dimensional.
In section \ref{sec:hamiltonian} we consider the equation in (3+1) dimensions with 4 dimensional nilpotent matrices and propose a Hamiltonian for this equation. In section \ref{sec:hydrogen} we solve this equation for the Coulomb potential and derive the quantized energy levels of a hydrogen-like atom. We conclude in section \ref{conclude}.
\section{Two Dimensional Matrices}\label{sec:2dmatrices}
In this section we introduce the equivalent form of the equivalent form of the L{\'e}vy-Leblond equation where the nilpotent matrices are 2 dimensional. It was shown in \cite{LevyLeblond:1967zz, Ajaib:2015uha} that the Schr{\"o}dinger equation can be derived from a first order equation similar to the manner in which the Klein Gordon equation can be derived from the Dirac equation. The nilpotent matrices considered in \cite{LevyLeblond:1967zz, Ajaib:2015uha} were four dimensional. In this section we consider the nilpotent matrices to be 2 dimensional. In (1+1) dimensions the equivalent form of the L{\'e}vy-Leblond equation is given by \cite{Ajaib:2015uha}
\begin{eqnarray}
-i \partial_z \psi = (i \eta \partial_t + \eta^\dagger m) \psi
\label{eq:2dmat}
\end{eqnarray}
where the matrix $\eta$ is a 2$\times$2 nilpotent matrix given by
\begin{eqnarray}
\eta = \frac{\sigma_1-i \sigma_2}{\sqrt{2}} =
{\sqrt{2}}
\left(
\begin{array}{cc}
0 & 0 \\
1 & 0 \\
\end{array}
\right)
\label{matrix:eta2d}
\end{eqnarray}
Following the procedure presented in \cite{Ajaib:2015uha}, we can show that the probability current in this case as well is given by
\begin{eqnarray}
J &=& \psi^\dagger ( \eta+\eta^{\dagger} ) \psi \\
\rho &=& \psi^\dagger \eta^\dagger \eta \psi
\end{eqnarray}
where $\eta+\eta^{\dagger}=\sqrt{2} \ \sigma_1$ and $\eta^\dagger \eta=I+\sigma_3$. In momentum space, equation (\ref{eq:2dmat}) is given by
\begin{eqnarray}
p_z = (i \eta \partial_t + \eta^\dagger m) \psi
\label{eq:2dmatp}
\end{eqnarray}
The eigenvectors of the momentum operator are given by
\begin{eqnarray}
e_{1,2} &=&
\left(
\begin{array}{c}
\pm \sqrt{\frac{E}{m}} \\
1 \\
\end{array}
\right)
\label{matrix:ev2d}
\end{eqnarray}
which correspond to eigenvalues $\pm p_z=\pm \sqrt{2Em}$, respectively. Note that, in contrast to the equation with four dimensional matrices \cite{Ajaib:2015uha, Ajaib:2015eer}, the spin of the particle is not taken into account by equation (\ref{eq:2dmat}). The author has checked that the step potential problem and the finite step potential problems solved with equation (\ref{eq:2dmat}) yield results that are consistent with standard quantum mechanical results as in the case of four dimensional matrices \cite{Ajaib:2015uha, Ajaib:2015eer}.
The (2+1) dimensional version of the L{\'e}vy-Leblond equation for 2$\times$2 matrices, in momentum space, is given by
\begin{eqnarray}
\mu_i p_i = ( \eta E + \eta^\dagger m)
\label{eq:2dmat2p1}
\end{eqnarray}
where $\mu_1=I$ and $\mu_2=i\sigma_3$.
We can show that equation (\ref{eq:2dmat2p1}) is the non-relativistic limit of the Dirac equation in (2+1) dimensions. Consider the following form of the (2+1) dimensional Dirac equation in momentum space
\begin{eqnarray}
\gamma_i p_i = ( \sigma_1 E +i\sigma_2 m) \psi
\label{eq:2dmat2p1de}
\end{eqnarray}
where $\gamma_1=I$ and $\gamma_2=i\sigma_3$. The above equation yields the dispersion relation of a massive relativistic particle in 2D. As in reference \cite{Ajaib:2015eer}, we can substitute $\sigma_1=(\eta+\eta^\dagger)/\sqrt{2}$ and $-i\sigma_2=(\eta-\eta^\dagger)/\sqrt{2}$ and apply the non-relativistic limit $E-m \simeq E'$ and $E+m \simeq 2m$ to obtain equation (\ref{eq:2dmat2p1}) from (\ref{eq:2dmat2p1de}).
Note also that in the limit $m=0$, equation (\ref{eq:2dmat2p1de}) reduces to the Dirac equation for massless fermions
\begin{eqnarray}
E=\sigma_i p_i
\end{eqnarray}
which, as an example, is employed to describe massless fermions in condensed matter systems such as graphene.
\section{The Hamiltonian}\label{sec:hamiltonian}
In this section we present the Hamiltonian corresponding to the equivalent form of the L{\'e}vy-Leblond equation with four dimensional matrices and discuss the constants of motion. The (3+1) dimensional version of equation is given by \cite{Ajaib:2015uha, Ajaib:2015eer}
\begin{eqnarray}
-i \gamma_i \partial_i \psi = (i \eta \partial_t + \eta^\dagger m) \psi
\label{mse-eq-3d}
\end{eqnarray}
where $\gamma_i$ are the Dirac gamma matrices and $\eta=(\gamma_0+i \gamma_5)/\sqrt{2}$. One of the issues in obtaining the Hamiltonian of equation (\ref{mse-eq-3d}) is that the matrix $\eta$ is singular. Recently, a Hamiltonian was proposed in \cite{Sobhani:2016xao} and we adopt a different approach herein. In order to obtain the Hamiltonian we replace $\eta \rightarrow \eta^\prime=\eta-\epsilon \eta^\dagger$ and analyze the limit $\epsilon \rightarrow 0$. We thereby obtain the following Hamiltonian for equation (\ref{mse-eq-3d})
\begin{eqnarray}
H=\eta^{\prime -1} (-i \gamma_i \partial_i-m \eta^{\prime \dagger})
\label{eq:hamiltonian}
\end{eqnarray}
where $\eta^\prime=\eta-\epsilon \eta^\dagger$ and we choose $\hbar=c=1$. In the limit $\epsilon \rightarrow 0$ two of the eigenvalues of the Hamiltonian in (\ref{eq:hamiltonian}) are finite where as two approach infinity
\begin{eqnarray}
E_{1,2} &=& \frac{\vec{p}^2}{2m} \\
E_{3,4} &=& -\frac{\vec{p}^2}{2m}+\frac{m}{\epsilon}
\end{eqnarray}
The Hamiltonian yields the two finite energy states in addition to negative energy states with an infinite part. The infinity associated with the negative energy states can be interpreted as the ``sea" of filled negative energy states. For the negative energy states we can define the renormalized energy as
\begin{eqnarray*}
E^\prime_{3,4}=E_{3,4}-\frac{m}{\epsilon}= -\frac{\vec{p}^2}{2m}
\end{eqnarray*}
The Hamiltonian (\ref{eq:hamiltonian}) is not hermitian however the eigenvalues of the operator are real.
Interestingly, the Hamiltonian (\ref{eq:hamiltonian}) commutes with the total angular momentum operator $\vec{J}=\vec{L}+1/2 \vec{\Sigma}$ and the operators $J^2, \ J_z$ and $K$, i.e.,
\begin{eqnarray*}
[H,J^2] &=& 0 \\
\left[H,J_z\right] &=& 0 \\
\left[H,K\right] &=& 0
\end{eqnarray*}
and the operator $K$ also commutes with the total angular momentum operators $J^2$ and $J_z$. The operator $K$ is given by
\begin{eqnarray}
K &=& i \gamma_5 \gamma_0 (\vec{\Sigma}.\vec{J}-\frac{1}{2} I ) \\
&=& i \gamma_5 \gamma_0 (\vec{\Sigma}.\vec{L}+ I ) \nonumber \\
&=&
i\left(
\begin{array}{cc}
0 & \vec{\sigma}.\vec{L}+I \\
-\vec{\sigma}.\vec{L}-I & 0
\end{array}
\right)
\label{matrix-k}
\end{eqnarray}
where $\vec{J}=\vec{L}+1/2 \vec{\Sigma}$. We can construct simultaneous eigenfunctions of the mutually commuting operators $H$, $J^2$, $J_z$ and $K$. The corresponding eigenvalues of these operators are denoted by $E$, $j(j+1)$, $m_j$ and $-\kappa$. We consider the following four component wave function as the simultaneous eigenfunction of these operators
\begin{eqnarray}
\psi=
\left(
\begin{array}{c}
\psi_A \\
\psi_B
\end{array}
\right)
=
\left(
\begin{array}{c}
g(r) Y^{j,m_j}_{l_A}(\theta, \phi) \\
i f(r) Y^{j,m_j}_{l_B}(\theta, \phi)
\end{array}
\right)
\equiv
\left(
\begin{array}{c}
g(r) Y_{A} \\
i f(r) Y_{B}
\end{array}
\right)
\label{matrix-2}
\end{eqnarray}
and for the angular part $Y^{j,m_j}_{l_A, l_B}(\theta, \phi)$ we consider the case $\theta=0$ \cite{Sakurai:2014}
\begin{eqnarray}
Y^{j,m_j}_{l=j\mp 1/2}(\theta=0,\phi)=\sqrt{\frac{j+1/2}{4\pi}} \left(
\begin{array}{c}
\pm \delta_{m,1/2} \\
\delta_{m,-1/2}
\end{array}
\right)
\label{eq:angular-part}
\end{eqnarray}
where $l_A=j+1/2$ and $l_B=j-1/2$. {We choose $\theta=0$ because the effect of the pseudo-scalar operator $\vec{\sigma}.\vec{r}/r$ on $Y^{j,m_j}_{l}$ is independent of $\theta$ \cite{Sakurai:2014}}. The eigenvalues of the operator $K$ are given by
\begin{eqnarray}
K \psi &=& - \kappa \psi
\end{eqnarray}
Since $J^2=K^2-1/4 I$, the eigenvalues of the two operators are related as $\kappa=\pm(j+1/2)$. Plugging in for $K$ yields the following equations
\begin{eqnarray}
\vec{\sigma}.\vec{L} \psi_A &=& - i \kappa \psi_B-\psi_A \label{eq:sla}\\
\vec{\sigma}.\vec{L} \psi_B &=& i \kappa \psi_A-\psi_B \label{eq:slb}
\end{eqnarray}
In addition we have the following eigenvalue equations
\begin{eqnarray}
\vec{J}^2 \psi_{A,B} &=& j(j+1) \psi_{A,B} \\
J_z \psi_{A,B} &=& j_z \psi_{A,B}
\end{eqnarray}
\section{Solution for the Coulomb Potential and the Hydrogen-like Atom}\label{sec:hydrogen}
In this section, we study the problem of an electron bound to a nucleus by a Coulomb potential for a hydrogen-like atom (For the analysis with the Dirac equation and further details the reader is referred to \cite{Sakurai:2014, Sakurai:1967, Greiner:2000}). For the case of a Coulomb potential, the Hamiltonian is given by
\begin{eqnarray}
H=\eta^{\prime -1} (-i \gamma_i \partial_i-m \eta^{\prime \dagger})+ V(r)
\end{eqnarray}
where $V(r)=-Z\alpha/r$, $\alpha\approx 1/137$ is the fine structure constant and $Z$ is the atomic number of the atom. Since $\psi$ is an eigenstate of the Hamiltonian
\begin{eqnarray*}
H \psi &=& E \psi \\
\eta^{\prime -1} (-i \gamma_i \partial_i-m \eta^{\prime \dagger}) \psi +V(r) \psi &=& E \psi \\
( \gamma_i p_i-m \eta^{\prime \dagger})\psi &=& (E-V(r))\eta^{\prime} \psi
\end{eqnarray*}
where $p_i=-i \partial_i$. We therefore obtain
\begin{eqnarray}
\gamma_i p_i \psi = ( \eta^\prime (E-V(r)) + \eta^{\prime \dagger} m) \psi
\label{eq-3d}
\end{eqnarray}
\begin{eqnarray}
\vec{\sigma}.\vec{p}
\left(
\begin{array}{c}
\psi_B \\
-\psi_A
\end{array}
\right)
= \frac{1}{\sqrt{2}}
\left(
\begin{array}{cc}
a^\prime (E-V+m) & i a (E-V-m) \\
i a (E-V-m) & -a^\prime (E-V+m)
\end{array}
\right)
\left(
\begin{array}{c}
\psi_A \\
\psi_B
\end{array}
\right)
\label{matrix-2b}
\end{eqnarray}
where $a^\prime=1-\epsilon$ and $a=1+\epsilon$. For brevity, we write
\begin{eqnarray}
\vec{\sigma}.\vec{p}
\left(
\begin{array}{c}
\psi_B \\
-\psi_A
\end{array}
\right)
=
\left(
\begin{array}{cc}
h_1 & i h_2 \\
i h_2 & -h_1
\end{array}
\right)
\left(
\begin{array}{c}
\psi_A \\
\psi_B
\end{array}
\right)
\label{matrix-3}
\end{eqnarray}
where
\begin{eqnarray}
h_1(r)=a^\prime/\sqrt{2} (E-V(r)+m) \label{eq:h1} \\
h_2(r)=a/\sqrt{2} (E-V(r)-m) \label{eq:h2}
\end{eqnarray}
The operator $\vec{\sigma}.\vec{p}$ can be written in terms of the radial and angular operators as
\begin{eqnarray}
\vec{\sigma}.\vec{p} = \frac{1}{r} \frac{\vec{\sigma}.\vec{r}}{r}\left( -i r \frac{\partial}{\partial r} + i \vec{\sigma}.\vec{L} \right)
\label{eq:sigdp}
\end{eqnarray}
The operator $\vec{\sigma}.\vec{r}/r$ is a pseudo scalar and changes the parity of the state, i.e.
\begin{eqnarray}
\frac{\vec{\sigma}.\vec{r}}{r} Y_A=-Y_B
\label{eq:paritych}
\end{eqnarray}
with $(\vec{\sigma}.\vec{r}/{r})^2=1$. We are interested in the effect of the operator $\vec{\sigma}.\vec{r}/r$ on $Y^{j,m_j}_{l}$ and due to its pseudo-scalar nature its effect on $Y^{j,m_j}_{l}$ is independent of $\theta$ \cite{Sakurai:2014}. So we choose $\theta=0$ for the angular part and employ the expression given in equation (\ref{eq:angular-part}) for the analysis. Plugging (\ref{eq:sigdp}) in (\ref{matrix-3}) we obtain the following two equations
\begin{eqnarray}
\frac{1}{r} \frac{\vec{\sigma}.\vec{r}}{r}\left( -i r \frac{\partial}{\partial r} + i \vec{\sigma}.\vec{L} \right) \psi_B = h_1 \psi_A +i h_2 \psi_B \\
-\frac{1}{r} \frac{\vec{\sigma}.\vec{r}}{r}\left( -i r \frac{\partial}{\partial r} + i \vec{\sigma}.\vec{L} \right) \psi_A = i h_2 \psi_A - h_1 \psi_B
\end{eqnarray}
Plugging in $\psi_A=g(r) Y_{A}$ and $\psi_B= i f(r) Y_{B}$ and using equations (\ref{eq:angular-part}), (\ref{eq:sla}), (\ref{eq:slb}), and (\ref{eq:paritych}) results in the following equations
\begin{eqnarray}
\frac{\partial f}{\partial r}+\frac{1}{r} f+h_2 f+h_1 g +\frac{\kappa}{r} g =0 \\
\frac{\partial g}{\partial r}+\frac{1}{r} g-h_2 g-h_1 f +\frac{\kappa}{r} f =0
\end{eqnarray}
The above equations are obtained for the $m=+1/2$ case. The analysis below also holds for the $m=-1/2$ which yields the similar results. Next, plugging in $f(r)=F(r)/r$ and $g(r)=G(r)/r$ and using (\ref{eq:h1}) and (\ref{eq:h2}) we obtain the following equations
\begin{eqnarray}
\frac{\partial F}{\partial r}+\left( q_1 + \frac{q_2}{r} \right) F+\left( p_1 + \frac{p_2}{r} +\frac{\kappa }{r} \right) G =0
\label{eq:Fr} \\
\frac{\partial G}{\partial r}-\left( q_1 + \frac{q_2}{r} \right) G+\left(-p_1 - \frac{p_2}{r} +\frac{\kappa }{r} \right) F =0
\label{eq:Gr}
\end{eqnarray}
Here we have defined the following constants
\begin{eqnarray}
p_1=\frac{a}{\sqrt{2}} (E+m), \ \
p_2=\frac{a}{\sqrt{2}} {Z\alpha} \\
q_1=\frac{a^\prime}{\sqrt{2}} (E-m), \ \
q_2=\frac{a^\prime}{\sqrt{2}} {Z\alpha}
\end{eqnarray}
We postulate series solutions of (\ref{eq:Fr}) and (\ref{eq:Gr}) of the form
\begin{eqnarray}
F(r)=e^{-\lambda r} \sum\limits_{n=0}^{\infty}a_n r^{s+n} \label{eq:series-1}\\
G(r)=e^{-\lambda r} \sum\limits_{n=0}^{\infty}b_n r^{s+n}
\label{eq:series-2}
\end{eqnarray}
Plugging (\ref{eq:series-1}) and (\ref{eq:series-2}) in (\ref{eq:Fr}) and (\ref{eq:Gr}) we obtain the following equations for the coefficients of the two series
\begin{eqnarray}
q_2 a_{n + 1} + (n + 1) a_{n+1} + q_1 a_{n} +
s a_{n+1} - \lambda a_{n} + p_2 b_{n+1} + \kappa b_{n+1} + p_1 b_{n}=0
\label{eq:rec-1} \\
-p_2 a_{n+1} + \kappa a_{n+1} - p_1 a_{n} -
q_2 b_{n+1} + (n + 1) b_{n+1} - q_1 b_{n} +
s b_{n+1} - \lambda b_{n}=0
\label{eq:rec-2}
\end{eqnarray}
For $n=-1$, the above equations are given as follows
\begin{eqnarray}
q_1 a_{-1} + (q_2 + s) a_{0} +
p_1 b_{-1} + (p_2 + \kappa) b_{0} &=& \lambda a_{-1} \\
p_1 a_{-1} + p_2 a_{0} + (q_1 + \lambda) b_{-1} + q_2 b_{0} &=&
\kappa a_{0} + s b_{0}
\end{eqnarray}
Setting $a_{-1}=b_{-1}=0$ yields
\begin{eqnarray}
(q_2 + s) a_{0} + (p_2 + k) b_{0} &=& 0 \\
p_2 a_{0} + q_2 b_{0} &=& \kappa a_{0} + s b_{0}
\end{eqnarray}
The solution of the above equations is
\begin{eqnarray}
s &=& \pm \sqrt{\kappa^2+q_2^2-p_2^2 }
\end{eqnarray}
For the wave function to be normalizable we choose the positive sign of the square root. Furthermore, the series of $F(r)$ and $G(r)$ must terminate at some $n=n^{\prime}$ for the state to be normalizable. This implies that the coefficients $a_i=b_i=0$ for $i=n^\prime+1$ and we obtain the following relation
\begin{eqnarray}
b_{n^\prime}=\frac{\sqrt{q_1^2-p_1^2}-q_1}{p_1} a_{n^\prime}
\label{eq:bnp}
\end{eqnarray}
where we have chosen $\lambda=\sqrt{q_1^2-p_1^2}$. Next we solve the recursion relations (\ref{eq:rec-1}) and (\ref{eq:rec-2}) for $n=n^\prime-1$
\begin{eqnarray}
(q_1 - \lambda) a_{n^\prime-1} + (q_2 + n^\prime + s) a_{n^\prime} +
p_1 b_{n^\prime-1} + (p_2 + k) b_{n^\prime}=0 \label{eq:np1}\\
-p_1 a_{n^\prime-1} + (-p_2 + k) a_{
n^\prime} - (q_1 + \lambda) b_{n^\prime-1} + (-q_2 + n^\prime + s) b_{n^\prime}=0 \label{eq:np2}
\end{eqnarray}
\noindent
Multiplying (\ref{eq:np1}) by $1/(\lambda-q_1)$ and (\ref{eq:np2}) by $1/p_1$ and subtracting we obtain the following equation
\begin{eqnarray}
p_1((q_2 + n^\prime + s) a_{n^\prime} + (p_2 + k) b_{n^\prime})) + (
q_1 - \lambda)((-p_2 + k) a_{n^\prime} + (-q_2 + n^\prime + s) b_{n^\prime})=0
\end{eqnarray}
Taking the limit $\epsilon \rightarrow 0$ ($a=a^\prime=1$, $s=\kappa=j+1/2$) and using equation (\ref{eq:bnp}) and $\lambda=\sqrt{-2 E m}$ we obtain the relation for the energy level
\begin{eqnarray}
E = -\frac{m Z^2 \alpha^2}{2 n^2}
\label{eq:energy-level}
\end{eqnarray}
where $n=n^\prime + s=n^\prime+j+1/2=n^\prime+l+1$
is the principal quantum number. The above equation is the known expression for the energy level of a hydrogen-like atom. For the hydrogen atom $Z=1$. Note that the parameter $s$ has to be positive and since $s=\kappa$, only $\kappa=+(j+1/2)$ is relevant. The functions $f(r)$ and $g(r)$ are therefore given by
\begin{eqnarray}
f(r)=e^{-\sqrt{-2Em} r} \ r^{\kappa-1} \sum\limits_{m=0}^{\infty}a_m r^{m}
= e^{-\frac{m Z \alpha}{n} r} \ r^{l} \sum\limits_{m=0}^{\infty}a_m r^{m}
\label{eq:series-3}\\
g(r)=e^{-\sqrt{-2Em} r} \ r^{\kappa-1} \sum\limits_{m=0}^{\infty}b_m r^{s+m}
= e^{-\frac{m Z \alpha}{n} r} \ r^{l} \sum\limits_{m=0}^{\infty}b_m r^{m}
\label{eq:series-4}
\end{eqnarray}
The ground state wave function ($n^\prime=0, \ \kappa=1, \ j=1/2$) of the Hydrogen atom can be written as
\begin{eqnarray}
\psi_{gd}=
N \frac{1}{\sqrt{4\pi}} e^{-Z r/a_{B}}\left(
\begin{array}{c}
g(r) \chi_s \\
-i f(r) \ {\vec{\sigma}}.{\hat{{r}}} \ \chi_s
\end{array}
\right)
\label{matrix-3b}
\end{eqnarray}
where $a_B=1/\alpha m$ is the Bohr's radius and
\begin{eqnarray}
\vec{\sigma}.\hat{r}
=
\left(
\begin{array}{cc}
\cos\theta & e^{-i\phi} \sin\theta \\
e^{i\phi} \sin\theta & -\cos\theta
\end{array}
\right)
\end{eqnarray}
\begin{eqnarray}
\chi_s=
\left(
\begin{array}{c}
1 \\
0 \\
\end{array}
\right) \,
\mathrm{or}
\,
\left(
\begin{array}{c}
0\\
1 \\
\end{array}
\right)
\label{matrix-5}
\end{eqnarray}
for the spin quantum number $m_s=+1/2$ and $m_s=-1/2$. For $m_s=+1/2$ the wave function is given by
\begin{eqnarray}
\psi_{gd}= N \frac{1}{\sqrt{4\pi}} e^{-Z r/a_{B}}
\left(
\begin{array}{c}
1 \\
0 \\
-i d_0 \cos\theta \\
-i d_0 \sin\theta e^{i \phi}\\
\end{array}
\right)
\label{matrix-4}
\end{eqnarray}
For $m_s=-1/2$
\begin{eqnarray}
\psi_{gd}= N \frac{1}{\sqrt{4\pi}} e^{-Z r/a_{B}}
\left(
\begin{array}{c}
0 \\
1 \\
-i d_0 \sin\theta e^{-i \phi} \\
i d_0 \cos\theta \\
\end{array}
\right)
\label{matrix-4b}
\end{eqnarray}
where $d_0=a_0/b_0=\frac{2-\sqrt{2} Z \alpha}{2+\sqrt{2} Z \alpha}$. The normalization constant is given by
\begin{eqnarray}
N=2\sqrt{\pi}\left(\frac{Z}{a_B}\right)^{3/2} \frac{2+\sqrt{2}{Z \alpha}}{\sqrt{2+Z^2\alpha^2}}
\end{eqnarray}
\section{Conclusion} \label{conclude}
We presented the equivalent form of the L{\'e}vy-Leblond equation with two dimensional nilpotent matrices and showed that in (2+1) dimensions it can be obtained from the Dirac equation in the non-relativistic limit. In (3+1) dimensions we also proposed a Hamiltonian for this equation with four dimensional nilpotent matrices and showed that the quantized energy level of the hydrogen atom are obtained when the equation is solved for a Coulomb potential. We also derived the ground state wave function for spin up and down electron for a hydrogen-like atom. {The novelty of this approach is that the spin of the electron is taken into account in the non-relativistic limit to obtain the spectrum of the hydrogen atom}. This analysis further illustrates the application of this equation which allows for additional insights into a problem corresponding to the spin of the particle.
\section{Acknowledgments}
The author would like to thank Fariha Nasir, Warren Siegel and Mansoor Ur Rehman for useful discussions and suggestions. The author is also grateful to Amer Iqbal for bringing reference \cite{LevyLeblond:1967zz} to the authors attention.
|
\section{Introduction}
On-shell approaches play a central role in many state-of-the-art calculations in perturbative gauge theories.
Since only physical degrees of freedom appear on-shell, they enable to build observables in terms of the simplest but meaningful physical building blocks.
This is especially advantageous for massless particles with spin, such as gluons, where the focus on the two physical helicities
removes the need to introduce gauge redundancies, removing at the same time intricate cancellations among large numbers of Feynman diagrams.
In on-shell approaches, the Lagrangian and Feynman rules of a theory tend to occupy a secondary role, if any.
It is therefore crucial to develop a conceptual understanding, directly in the language that is used in calculations,
of the phenomena that are traditionally understood from the Lagrangian.
In this paper, we discuss a direct connection
between the high-energy behavior of the $S$-matrix of a theory and the running of coupling constants and renormalization of local operators.
We will build on recent developments in the context of
the dilatation operator in $\mathcal{N}=4$ super Yang-Mills (SYM) \cite{Zwiebel:2011bx,Wilhelm:2014qua,Nandan:2014oga,Koster:2014fva,Brandhuber:2014pta,Brandhuber:2015boa,Loebbert:2015ova,Brandhuber:2016fni} and other work based on generalized unitarity \cite{ArkaniHamed:2008gz,Huang:2012aq,Cheung:2015aba}, which we will extend to arbitrary weakly coupled field theories.
Our main physical idea will be the notion that large logarithms signaling the running of couplings originate from states which propagate over a ``long distance'' in an appropriate metric,
making them effectively on-shell.
Quantitatively, we will consider form factors, which are matrix elements between an operator and on-shell states:
\begin{equation}
\FO{p_1,\ldots,p_n;\mu} \equiv \langle p_1,\ldots, p_n | \mathcal{O}|0\rangle\,, \label{form_factor}
\end{equation}
where $\mu$ is the renormalization scale.
Such form factors figure prominently in effective-theory descriptions of weak processes including Higgs production and decay, see e.g.~\cite{Schmidt:1997wr,Anastasiou:2015ema}.
They convert the scale dependence of the local operator $\mathcal{O}$ into a physically measurable energy dependence of its decay products.
The key fact for us will be that the energy dependence and phase are tied to each other, as can be seen from the imaginary part
acquired by the logarithms for timelike momentum invariants ($p^2>0$)
due to Feynman's prescription $p^2\to p^2+i0$:
\begin{equation}
\log \left(\frac{-p^2}{\mu^2}\right) \equiv
\log \left(\frac{|p^2|}{\mu^2}\right)-i\pi \quad\Rightarrow\quad
p^\mu \frac{\partial}{\partial p^\mu} \log \left(\frac{-p^2}{\mu^2}\right) = -\frac{2}{\pi} \Im \log \left(\frac{-p^2}{\mu^2}\right)\,.
\label{logs_imaginary_part}
\end{equation}
This is interesting because, as understood from conventional unitarity and the optical theorem,
imaginary parts originate physically from the long time propagation of intermediate on-shell states.
This suggests that the scale dependence of a process can be understood directly from the
propagation of on-shell particles.
In this paper, we propose a precise quantitative relationship, which we will verify in a number of classic examples.
This paper is organized as follows.
In section \ref{sec: S matrix as dilatation operator}, we expand on the ideas sketched above, deriving a relation between the $S$-matrix and the dilatation operator, and we set up our notations.
In section \ref{sec: one-loop application}, we apply these ideas at one-loop level.
We calculate the $\beta$-functions and anomalous dimensions of various composite operators in pure Yang-Mills, perturbative QCD and $\mathcal{N}=4$ SYM.
In section \ref{sec: towards higher loops}, we extend our study to several features that appear at higher loop orders, in particular the mixing of operators of different lengths.
We conclude with a summary of our results and an outlook on future directions in section \ref{sec: Summary}.
\section{The \texorpdfstring{$S$}{S}-matrix and the dilatation operator}
\label{sec: S matrix as dilatation operator}
In this section, we derive a concrete formula, eq.~(\ref{cute_eigenvalue_equation}),
which instantiates the above general ideas, and we set up the notations we will use to test it.
The main first step is to connect the phase and energy dependence of form factors. This connection stems from analyticity.
The trick is to use a complex scale transformation to relate a form factor to its complex conjugate.
We start from a kinematic configuration where all momenta $p_i$ are outgoing,
so that all Mandelstam invariants are positive (timelike): $s_{ij\ldots k}=(p_i+p_j+\ldots+p_k)^2>0$.
The form factor is not real because the Feynman prescription adds a small positive imaginary part to all invariants: $s_{I}\mapsto s_{I}+i0$.
But it can be related to its conjugate by an analytic continuation in which all the invariants are rotated along a large circle in the complex plane,
with a common phase, as illustrated in fig.~\ref{fig: half-circle}.
Such a rotation is generated by the dilatation operator $D$:
\begin{equation}
F(p_1,\ldots,p_n) \to F(p_1\operatorname{e}^{i\alpha},\ldots,p_n\operatorname{e}^{i\alpha}) = \operatorname{e}^{i\alpha D} F(p_1,\ldots,p_n)\,,\quad\mbox{where}\quad D\equiv \sum_i p_i^\mu \frac{\partial}{\partial p_i^\mu}\,.
\end{equation}
We do not expect any singularity until the angle reaches $\pi$, where all energies are reversed.
(This is easily proved in perturbation theory, where the Feynman parameter representation contains denominators
of the form $\big(\sum_j c_j m_j^2-\sum_J c_J s_{J}-i0\big)$ with all $c_j,c_J$ positive. Taking all $s_J$ to have the same phase $\operatorname{e}^{2i\alpha}$, the first singularity is at $\alpha=\pi$.)
At this point, the invariants are back to the original ones but on the ``wrong'' side of the cut,
giving the conjugate form factor. Thus,
\begin{equation}
F = \operatorname{e}^{-i\pi D}F^*\,, \label{continuation_from_D}
\end{equation}
where $F^*$ is the form factor computed using anti-time-ordered propagators.
The second fundamental equation we will need is a version of the optical theorem. The conventional optical theorem
expresses unitarity of the $S$-matrix: $SS^\dagger=1$,
where the product contains a phase-space integral over intermediate $n$-particle states summed over all $n$.
Formally using the physical interpretation of a form factor as a small perturbation to the $S$-matrix,
$\delta S=i\mathcal{F}$, using the calligraphic font here to distinguish the operator $\mathcal{F}$ from its matrix elements to outgoing states $F$,
unitarity becomes $\mathcal{F}= S \mathcal{F}^\dagger S$. For vacuum initial states, this reduces to
\begin{equation}
F = S F^*\,.
\end{equation}
In this note, we will mostly rely on the imaginary part of this relation to one-loop order, which is easily verified from the Cutkowski rules.
The diagrams which contribute to the product $SF^*$ originate by drawing a cut through form factor diagrams, as depicted for example in fig.~\ref{fig: one-loop double cut} below. The massless scattering amplitudes contained in $S$ then join the cut to the final states.
We note that the other side of the cut involves a complex conjugate amplitude, as is typically the case for Cutkowski rules.
Combining the two relations above gives
\begin{equation}
\operatorname{e}^{-i\pi D}F^*=S F^*\,. \label{cute_eigenvalue_equation}
\end{equation}
This will be the central equation in this paper. We will read it as follows:
{\it the dilatation operator is minus the phase of the $S$-matrix, divided by $\pi$.}\footnote{Strictly speaking, we are omitting a CPT transformation here,
whose necessity can be seen for example using the commutation relation with the Hamiltonian $H$. We thank Amit Sever for this observation.}
\begin{figure}[htbp]
\begin{center}
\begin{tikzpicture}
[ axis/.style={->,black},
continuation/.style={->,black,thick},
line/.style={black,thick},
discontinuity/.style={black}]
\draw[axis] (-4,0) -- (4,0) node[anchor=west]{};
\draw[axis] (0,-1) -- (0,4) node[anchor=south]{};
\draw[continuation] ++(175:3) arc (175:5:3);
\draw[discontinuity,decoration={zigzag},decorate] (-0.6,0) -- (-4,0);
\draw[discontinuity,decoration={zigzag},decorate] (0.6,0) -- (4,0);
\node[anchor=south west] at (3,0) {$F$};
\node[anchor=north west] at (-3.7,0) {$F$};
\node[anchor=south west] at (-3.7,0) {$F^*$};
\node[anchor=south west] at (60:3) {$\operatorname{e}^{-i\pi D}$};
\draw (4,3.5) -- (3.5,3.5) -- (3.5,4.0);
\node[anchor=south west] at (3.5,3.5) {$p_i^0$};
\end{tikzpicture}
\end{center}
\caption{Analytic continuation from the complex conjugate using a complex scale transformation.}
\label{fig: half-circle}
\end{figure}
The dilatation operator is of course closely related to renormalization group evolution. Indeed, at high energies,
by dimensional analysis, $F$ can depend only on dimensionless ratios $s_{I}/\mu^2$, and therefore
$D\simeq-\mu\partial_\mu$. Starting from the renormalization group equation
\begin{equation}
\left[\mu\partial_\mu +\beta(g^2)\frac{\partial}{\partial g^2} + \gamma_\mathcal{O} - \gamma_{\rm IR} \right]F = 0\,, \label{eq: RG equation mu}
\end{equation}
one hence obtains
\begin{equation}
DF=\left(\gamma_\mathcal{O} - \gamma_{\rm IR}+\beta(g^2)\frac{\partial}{\partial g^2}\right)F\,. \label{rg_equation}
\end{equation}
It will be important that ultraviolet and infrared divergences both contribute to the energy dependence of form factors. Their relative sign is simply a convention
which ensures that the naturally large logarithms $\log(\mu_{\rm UV}^2/p^2)$ and $\log(p^2/\mu_{\rm IR}^2)$ come with the same sign
when their renormalization scales are treated independently.
Logarithms of momentum-independent masses will be discussed briefly in section \ref{eq: comment on masses} but do not fundamentally affect the discussion.
Inserting (the complex conjugate of) (\ref{rg_equation}) into (\ref{cute_eigenvalue_equation})
yields a relation between the renormalization group coefficients $\gamma_\mathcal{O}$, $\gamma_{\rm IR}$, $\beta$ and the S-matrix.
Let us focus on the leading approximation to this otherwise exact relation.
It is useful to restrict to so-called minimal form factors, which are non-vanishing in the free-theory limit.
The $\beta$-function term can then be neglected. Writing $S=1+i\mathcal{M}$ and
inserting (\ref{rg_equation}) into (\ref{cute_eigenvalue_equation}) then gives to leading non-trivial order:
\begin{equation}
\left(\gamma_\mathcal{O}^{(1)} - \gamma_{\rm IR}^{(1)}\right) \langle p_1,\ldots,p_n| \mathcal{O}|0\rangle^{(0)}
= -\frac{1}{\pi} \langle p_1,\ldots,p_n| \mathcal{M}\otimes \mathcal{O}|0\rangle^{(0)}\,, \label{eq: gamma UV plus IR}
\end{equation}
where $\mathcal{M}$ is the tree-level $2{\to}2$ $S$-matrix, and
the convolution, to be defined shortly, represents the phase-space integral over intermediate two-particle states
in the product $\mathcal{M} F^*$.
Note that we have dropped the complex conjugation sign, as the tree-level form factors are naturally real.
In order to use the above equation to extract anomalous dimensions, the infrared contributions must be subtracted.
The key fact is that these depend only on the external particles but not on $\mathcal{O}$. This makes it possible to construct infrared-safe ratios.
This is particularly simple in the special case $n=2$, where one can put the stress-tensor in the denominator.
It has vanishing anomalous dimension in any theory. Ignoring again the $\beta$-function, this gives
\begin{equation}
\gamma_{\mathcal{O}} = D \log \frac{\langle p_1,p_2|\mathcal{O}|0\rangle}{\langle p_1,p_2|T^{\mu\nu}|0\rangle}
=-\frac{1}{\pi} 2\Im \log \frac{\langle p_1,p_2|\mathcal{O}|0\rangle}{\langle p_1,p_2|T^{\mu\nu}|0\rangle}\,, \label{eq: IR safe ratio}
\end{equation}
which gives rise to the more practical one-loop equation
\begin{equation}
\gamma_\mathcal{O}^{(1)} \langle p_1,p_2| \mathcal{O}|0\rangle^{(0)}
= -\frac{1}{\pi} \langle p_1,p_2| \mathcal{M}\otimes \mathcal{O}|0\rangle^{(0)}+\frac{1}{\pi} \langle p_1,p_2| \mathcal{O}|0\rangle^{(0)} \frac{\langle p_1,p_2|\mathcal{M}\otimes T^{\mu\nu}|0\rangle^{(0)}}{\langle p_1,p_2|T^{\mu\nu}|0\rangle^{(0)}}\,.
\label{eq: gammaO practical}
\end{equation}
This equation is new.
Multiple examples and applications will be given in sections \ref{sec: one-loop application} and \ref{sec: towards higher loops}.
Note that the ratio in \eqref{eq: gammaO practical} does not depend on the indices on $T^{\mu\nu}$ because the infrared divergences are blind to these.
At higher loops and in the presence of a $\beta$-function, the imaginary part of the above logarithm is still useful and detects
the anomalous dimensions and coupling dependence of the form factor averaged over the half-circle of fig.~\ref{fig: half-circle}.
The anomalous dimensions of marginal and relevant operators are of particular physical interest due to
their relation to the $\beta$-functions of corresponding running couplings.
For example, in Yang-Mills theory,
the anomalous dimension of the Lagrangian density is a derivative of the $\beta$-function \cite{KlubergStern:1974rs,Grinstein:1988wz}:
\begin{equation}
\gamma_{\mathcal{L}} = g^2 \frac{\partial}{\partial g^2} \left(\frac{\beta(g^2)}{g^2}\right)\,. \label{eq:gamma vs beta function}
\end{equation}
The two are therefore essentially equivalent, making it possible
to use the preceding formulas to obtain $\beta$-functions. The multi-coupling case will be discussed further in section \ref{sec: towards higher loops}.
Note that the arguments above are valid in any space-time dimension.
In the following, we will restrict ourselves to four dimensions though.
\subsection{Notations: form factors and spinor-helicity variables}
Form factors provide a map between on-shell states and local operators.
In a free theory, they are just polynomials in the momenta. For example, for a free scalar
\begin{equation}
i^n \langle 1_\phi | \partial^{\mu_1}\cdots \partial^{\mu_n} \phi|0\rangle = p_1^{\mu_1}\cdots p_1^{\mu_n}\,.
\end{equation}
In general, for final state with multiple particles, there is a one-to-one correspondence between such polynomials and local operators modulo equations of motion.
Note that we use an abbreviated notation where the bra $\langle 1_h|$ denotes a particle of type $h$ with momentum $p_1$.
When dealing with particles with spin, it is useful to use variables which can absorb the phase ambiguities
of their polarization vectors and spinors. In four dimensions, this is nicely achieved by the so-called spinor-helicity variables.
These are defined by splitting a null four-momentum into two Weyl spinors:
\begin{equation}
p_j^{\alpha{\dot{\alpha}}} \equiv p_j^\mu \sigma_\mu^{\alpha{\dot{\alpha}}}= \lambda_j^\alpha \tilde\lambda_j^{{\dot{\alpha}}} \,,
\label{eq: spinor helicity}
\end{equation}
where $(\sigma_\mu)^{\alpha{\dot{\alpha}}}$ are the four-dimensional ($2\times 2$) Pauli matrices.
The two helicity polarizations of a gluon can be parametrized explicitly in terms of the spinors, see for example \cite{Henn:2014yza}.
The important fact is that the physics is invariant if spinors and antispinors are rotated by opposite phases,
provided the external states are simultaneously rotated according to their helicity:
\begin{equation}
\lambda_j\to \lambda_j \operatorname{e}^{i\alpha_j}\,,\quad
\tilde\lambda_j\to \tilde\lambda_j \operatorname{e}^{-i\alpha_j}\,,\quad
\langle j_-| \to \operatorname{e}^{2i\alpha_j}\langle j_-|\,,\quad
\langle j_+| \to \operatorname{e}^{-2i\alpha_j}\langle j_+|\,. \label{eq:little group}
\end{equation}
This is called little-group scaling because the same phases would arise from a rotation along the propagation axis of particle $j$.
Thus, form factors are polynomials in the spinor-helicity variables with a specific little-group weight for each particle.
This fixes the form of form factors for the self-dual and anti-self-dual parts of the field strength and fermion fields,
\begin{equation}\begin{aligned}
\langle 1_-|F^{\alpha\beta}|0\rangle \equiv \lambda_1^{\alpha}\lambda_1^{\beta}\,,\quad
\langle 1_+|\bar{F}^{\dot\alpha\dot\beta}|0\rangle \equiv \tilde{\lambda}_1^{{\dot{\alpha}}}\tilde{\lambda}_1^{{\dot{\beta}}}\,,\quad
\langle 1_{{\bar{\psi}}}| \psi^{\alpha}|0\rangle \equiv \lambda_1^{\alpha}\,,\quad
\langle 1_{{\psi}}| \bar{\psi}^{\dot\alpha}|0\rangle \equiv \tilde{\lambda}_1^{{\dot{\alpha}}}\,,
\end{aligned}\end{equation}
where the state $\langle 1_{{\psi}}|$ is a Weyl fermion of positive helicity.
We follow conventions where the basic Lorentz invariant combinations are the brackets
\begin{equation}
s_{ij}=2p_i{\cdot}p_j = \langle i\,j\rangle[j\,i]\,,\quad\mbox{where}\quad
\langle i\,j\rangle \equiv \epsilon_{\alpha\beta}\lambda_i^\alpha\lambda_j^\beta\,,\quad
[i\,j] \equiv \epsilon_{\dot\alpha\dot\beta}\tilde\lambda_i^{\dot\alpha}\tilde\lambda_j^{\dot\beta}\,,
\end{equation}
with the Mandelstam invariant $s_{ij}>0$ when the invariant is timelike, as is the case for two outgoing particles.
For outgoing momenta, there is the complex conjugation relation $\lambda_i=(\tilde\lambda_i)^*$.
Like its name suggests, the $S$-matrix $S=1+i\mathcal{M}$ is an operator,
which in particular can act on the polynomial states produced by minimal form factors. This action, which we denote as a convolution,
is simply the on-shell phase-space integral:
\begin{eqnarray}
\langle 12| \mathcal{M}\otimes F|0\rangle^{(0)} \equiv \frac{1}{16\pi} \sum_{h_{1'},h_{2'}}
\int \frac{\operatorname{d}\!\Omega}{4\pi} \langle 12| \mathcal{M}| 1'_{h_{1'}}2'_{h_{2'}}\rangle^{(0)}
\langle 1'_{h_{1'}}2'_{h_{2'}}| F|0\rangle^{(0)}\,, \label{eq: phase space}
\end{eqnarray}
where the sum is over all intermediate helicity states.
The following elegant phase-space parametrization using spinors will be useful: one simply rotates the spinors as \cite{Zwiebel:2011bx}
\begin{equation}
\left(\begin{array}{c} \lambda_1'\\ \lambda_2'\end{array}\right)=
\left(\begin{array}{cc} \cos\theta & -\sin\theta \operatorname{e}^{i\phi}\\\sin\theta \operatorname{e}^{-i\phi}&\cos\theta\end{array}\right)
\left(\begin{array}{c} \lambda_1\\ \lambda_2\end{array}\right)\,, \label{rotation}
\end{equation}
together with the complex conjugate rotation for the conjugate spinors $\tilde{\lambda}_1'$ and $\tilde{\lambda}_2'$.
It is easy to verify that $p_1'+p_2'=p_1+p_2$.
In a center-of-mass frame where $p_1$ and $p_2$ are back-to-back along the $z$-axis, this reduces to a standard parametrization of spinors in terms of
polar half-angle $\theta$ and azimuthal angle $\phi$. The advantage is that, being covariant, this can be used in any frame.
The integration measure is simply
\begin{equation}
\label{eq: one-loop measure}
\int \frac{\operatorname{d}\!\Omega}{4\pi} \equiv
\int \frac{\operatorname{d}\!\phi}{2\pi}\int_0^{\frac{\pi}{2}}2\cos\theta\sin\theta \operatorname{d}\!\theta\,.
\end{equation}
Finally, following general practice in the amplitudes community, we will use crossing symmetry liberally
and often express $S$-matrix elements in a notation where momenta and other quantum numbers are outgoing:
\begin{equation}
(-1)^{n_{{\bar{\psi}}}}
\langle 1_{h_1}2_{h_2}| \mathcal{M} | 3_{h_3}4_{h_4}\rangle \equiv \langle 1_{h_1}2_{h_2}\bar{4}_{-h_4}\bar{3}_{-h_3}|\mathcal{M}|0\rangle\equiv \mathcal{M}_{1_{h_1}2_{h_2}\bar{4}_{-h_4}\bar{3}_{-h_3}}\,,
\label{eq: crossing}
\end{equation}
where $\bar{j}$ means minus the momentum $p_j$ (with $\lambda_{\bar{j}}=\lambda_j, \tilde{\lambda}_{\bar{j}}=-\tilde{\lambda}_{j}$).
Reversing the order of fields upon crossing is a useful convention which ensures the proper minus signs for fermion loops.
An additional minus sign counts the number of negative-helicity fermions ${\bar{\psi}}$ in the initial state.
\section{Application: Yang-Mills theory at one-loop}
\label{sec: one-loop application}
To compute all one-loop anomalous dimensions in Yang-Mills theory, the major ingredient will be
the on-shell four-gluon amplitude, given by the famous Parke-Taylor expression:
\begin{equation}
\label{eq: Parke-Taylor formula}
\mathcal{M}^{abcd}_{1^-2^-3^+4^+} = -2g^2 \langle 12\rangle^4 \left[ \frac{f^{abe}f^{cde}}{\l12\rangle\l23\rangle\l34\rangle\l41\rangle}+\frac{f^{ace}f^{bde}}{\l13\rangle\l32\rangle\l24\rangle\l41\rangle}\right]\,.
\end{equation}
For other helicity choices, one simply replaces $\l12\rangle^4$ by $\langle ij\rangle^4$, where $i$ and $j$ are the two negative-helicity gluons;
the four-gluon tree amplitude vanishes if there are not exactly two negative-helicity gluons.
We will mostly need the case where the initial state is a color-singlet gluon pair, in which case the formula simplifies as the first term vanishes:
\begin{equation}
\mathcal{M}^{abcd}_{1^-2^-3^+4^+} \delta^{cd}= -2g^2C_A \delta^{ab} \frac{\langle 12\rangle^4}{\langle 13\rangle\l32\rangle\l24\rangle\l41\rangle}\,. \label{eq: Parke-Taylor formula_trace}
\end{equation}
Here, $C_A$ denotes the Casimir in the adjoint representation, which is $N_c$ for gauge group SU(N${}_c$).
Before using it, let us briefly comment on various ways to obtain eq.~(\ref{eq: Parke-Taylor formula_trace}), which of course include
direct Feynman diagram calculation \cite{Srednicki:2007qs,Elvang:2013cua}. It is also a special case of the celebrated MHV $n$-point amplitude,
now understood from a large number of viewpoints including Berends-Giele \cite{Berends:1987me} and BCFW recursion \cite{Britto:2004ap,Britto:2005fq}, properties of self-dual Yang-Mills \cite{Rosly:1996vr}, the twistor string \cite{Witten:2003nn}, etc.
In fact, the above formula is a direct consequence of basic physical principles, specifically its little-group properties and classical small-angle limits.
The key point is that the little-group scaling (\ref{eq:little group}) implies that the amplitude can be written as $\frac{\langle 34\rangle^2}{\langle 12\rangle^2}$ times a rational function $G(s,t,u)$. Since a tree amplitude cannot have a squared denominator such as $1/\langle 12\rangle^2$, $G$ needs to be proportional to $s=\l12\rangle[21]$, and since it needs to be dimensionless and only massless poles can appear in its denominator,
the most general possibility is $G=c_1 \frac{s}{t} + c_2 \frac{s}{u}$. In the small-angle limit $t\to 0$,
the amplitude has to reproduce the Coulomb-like attractive potential $\mathcal{M}\to -2g^2f^{a_1a_4b}f^{a_2a_3b}\tfrac{s}{t}$,
and similarly at $u\to 0$, which fixes $c_1=c_2=-2g^2C_A$. This reproduces eq.~(\ref{eq: Parke-Taylor formula_trace}) using spinor identities.
The absence of polynomial ambiguities for massless particles with spin is a generic consequence of little-group scaling \cite{Weinberg:1965rz,Benincasa:2007xk}.
Plugging in the explicit values for the rotated spinors in eq.~(\ref{rotation}),
\begin{equation}
\langle 1'2'\rangle=\l12\rangle,\quad \l12'\rangle=\l1'2\rangle=\l12\rangle\cos\theta,\quad \l1'1\rangle=\l12\rangle\sin\theta \operatorname{e}^{i\phi},\quad \l2'2\rangle=\l12\rangle\sin\theta \operatorname{e}^{-i\phi}\,, \label{eq: rotated spinor products}
\end{equation}
one thus evaluates using the amplitude (\ref{eq: Parke-Taylor formula_trace}):
\begin{equation}
\label{eq: M++++}
\langle 1_-^a2_-^b| \mathcal{M}^{(0)}|1_-'^c2_-'^d\rangle \delta^{cd} = 2g^2C_A \delta^{ab} \frac{1}{\cos^2\theta\sin^2\theta}\,.
\end{equation}
For $+-$ pairs, one simply inserts either $\cos^4\theta$ or $\sin^4\theta \operatorname{e}^{\pm4i\phi}$ into the numerator, respectively, depending on whether $1$ and $1'$ have the
same or opposite helicity; in the latter case, the sign of the phase is given by the helicity of $1'$.
\subsection{One-loop \texorpdfstring{$\beta$}{beta}-function}
\label{sec: YM eleven thirds}
\begin{figure}
\centering
\begin{subfigure}[c]{0.3\textwidth}
\centering
$
\settoheight{\eqoff}{$\times$}%
\setlength{\eqoff}{0.5\eqoff}%
\addtolength{\eqoff}{-12.0\unitlength}%
\raisebox{\eqoff}{%
\fmfframe(2,2)(2,2){%
\begin{fmfchar*}(40,20)
\fmfright{vq}
\fmfleft{vp2,vp1}
\fmf{dbl_plain_arrow,tension=1.2}{vq,v1}
\fmf{plain_arrow,tension=0.5,left=0.7,label=$\scriptstyle 2'\,,$,l.d=15}{v1,v2}
\fmf{plain_arrow,tension=0.5,right=0.7,label=$\scriptstyle 1'\,,$,l.d=15}{v1,v2}
\fmf{phantom_smallcut,left=0.7,tension=0}{v1,v2}
\fmf{phantom_smallcut,right=0.7,tension=0}{v1,v2}
\fmf{plain_arrow}{v2,vp1}
\fmf{plain_arrow}{v2,vp2}
\fmfv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(0)}$,label.dist=0}{v1}
\fmfv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(0)}$,label.dist=0}{v2}
\fmffreeze
\fmfcmd{pair vertq, vertpone, vertptwo, vertpthree, vertpL, vertone, verttwo; vertone = vloc(__v1); verttwo = vloc(__v2); vertq = vloc(__vq); vertpone = vloc(__vp1); vertptwo = vloc(__vp2);}
\fmfiv{label=$\scriptstyle \mathcal{L}$}{vertq}
\fmfiv{label=$\scriptstyle 1-$}{vertpone}
\fmfiv{label=$\scriptstyle 2-$}{vertptwo}
\fmfiv{label=$\scriptstyle -$,l.d=20,l.a=-120}{vertone}
\fmfiv{label=$\scriptstyle -$,l.d=20,l.a=+120}{vertone}
\fmfiv{label=$\scriptstyle +$,l.d=20,l.a=-60}{verttwo}
\fmfiv{label=$\scriptstyle +$,l.d=20,l.a=+60}{verttwo}
\end{fmfchar*}%
}}%
$ \qquad
\caption{\textcolor{white}{.}}
\label{subfig: L}
\end{subfigure}
\begin{subfigure}[c]{0.3\textwidth}
\centering
$
\settoheight{\eqoff}{$\times$}%
\setlength{\eqoff}{0.5\eqoff}%
\addtolength{\eqoff}{-12.0\unitlength}%
\raisebox{\eqoff}{%
\fmfframe(2,2)(2,2){%
\begin{fmfchar*}(40,20)
\fmfright{vq}
\fmfleft{vp2,vp1}
\fmf{dbl_plain_arrow,tension=1.2}{vq,v1}
\fmf{plain_arrow,tension=0.5,left=0.7,label=$\scriptstyle 2'\,,$,l.d=15}{v1,v2}
\fmf{plain_arrow,tension=0.5,right=0.7,label=$\scriptstyle 1'\,,$,l.d=15}{v1,v2}
\fmf{phantom_smallcut,left=0.7,tension=0}{v1,v2}
\fmf{phantom_smallcut,right=0.7,tension=0}{v1,v2}
\fmf{plain_arrow}{v2,vp1}
\fmf{plain_arrow}{v2,vp2}
\fmfv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(0)}$,label.dist=0}{v1}
\fmfv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(0)}$,label.dist=0}{v2}
\fmffreeze
\fmfcmd{pair vertq, vertpone, vertptwo, vertpthree, vertpL, vertone, verttwo; vertone = vloc(__v1); verttwo = vloc(__v2); vertq = vloc(__vq); vertpone = vloc(__vp1); vertptwo = vloc(__vp2);}
\fmfiv{label=$\scriptstyle T$}{vertq}
\fmfiv{label=$\scriptstyle 1-$}{vertpone}
\fmfiv{label=$\scriptstyle 2+$}{vertptwo}
\fmfiv{label=$\scriptstyle +$,l.d=20,l.a=-120}{vertone}
\fmfiv{label=$\scriptstyle -$,l.d=20,l.a=+120}{vertone}
\fmfiv{label=$\scriptstyle -$,l.d=20,l.a=-60}{verttwo}
\fmfiv{label=$\scriptstyle +$,l.d=20,l.a=+60}{verttwo}
\end{fmfchar*}%
}}%
$ \qquad
\caption{\textcolor{white}{.}}
\label{subfig: T gluons}
\end{subfigure}
\begin{subfigure}[c]{0.3\textwidth}
\centering
$
\settoheight{\eqoff}{$\times$}%
\setlength{\eqoff}{0.5\eqoff}%
\addtolength{\eqoff}{-12.0\unitlength}%
\raisebox{\eqoff}{%
\fmfframe(2,2)(2,2){%
\begin{fmfchar*}(40,20)
\fmfright{vq}
\fmfleft{vp2,vp1}
\fmf{dbl_plain_arrow,tension=1.2}{vq,v1}
\fmf{plain_arrow,tension=0.5,left=0.7,label=$\scriptstyle 2'\,,$,l.d=15}{v1,v2}
\fmf{plain_arrow,tension=0.5,right=0.7,label=$\scriptstyle 1'\,,$,l.d=15}{v1,v2}
\fmf{phantom_smallcut,left=0.7,tension=0}{v1,v2}
\fmf{phantom_smallcut,right=0.7,tension=0}{v1,v2}
\fmf{plain_arrow}{v2,vp1}
\fmf{plain_arrow}{v2,vp2}
\fmfv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(0)}$,label.dist=0}{v1}
\fmfv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(0)}$,label.dist=0}{v2}
\fmffreeze
\fmfcmd{pair vertq, vertpone, vertptwo, vertpthree, vertpL, vertone, verttwo; vertone = vloc(__v1); verttwo = vloc(__v2); vertq = vloc(__vq); vertpone = vloc(__vp1); vertptwo = vloc(__vp2);}
\fmfiv{label=$\scriptstyle T$}{vertq}
\fmfiv{label=$\scriptstyle 1-$}{vertpone}
\fmfiv{label=$\scriptstyle 2+$}{vertptwo}
\fmfiv{label=$\scriptstyle {\psi}$,l.d=20,l.a=-120}{vertone}
\fmfiv{label=$\scriptstyle {\bar{\psi}}$,l.d=20,l.a=+120}{vertone}
\fmfiv{label=$\scriptstyle {\bar{\psi}}$,l.d=20,l.a=-60}{verttwo}
\fmfiv{label=$\scriptstyle {\psi}$,l.d=20,l.a=+60}{verttwo}
\end{fmfchar*}%
}}%
$ \qquad
\caption{\textcolor{white}{.}}
\label{subfig: T fermions}
\end{subfigure}
\caption{Different contributions to the anomalous dimension $\gamma_\mathcal{L}$ of the Lagrangian density and thus the Yang-Mills $\beta$-function.
This requires form factors for both the Lagrangian density (a) and stress tensor (b), with matter fields (c) contributing only to the latter.}
\label{fig: one-loop double cut}
\end{figure}
The Yang-Mills $\beta$-function is now given, according to the infrared-safe ratio in eq.~(\ref{eq: IR safe ratio}),
by acting with the above tree amplitude on the form factors
for the Lagrangian density $\mathcal{L}\equiv -G^a_{\mu\nu}G^{\mu\nu\,a}/(4g^2)$ and the stress tensor $T^{\alpha\beta,\dot\alpha\dot\beta}$.
At tree level, for each of these form factors, there is a unique polynomial in spinors that one can write down with the correct dimension, Lorentz indices,
and little-group phases:
\begin{equation}
\begin{aligned}
\langle 1^a_-2^b_-| \mathcal{L} |0\rangle &=\tfrac12\delta^{ab} \l12\rangle^2\,,
\\
\langle 1^a_- 2^b_+| T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle &= 2\delta^{ab}\lambda_1^\alpha\lambda_1^\beta \tilde{\lambda}_2^{\dot\alpha}\tilde{\lambda}_2^{\dot\beta}\,.
\end{aligned}
\end{equation}
The overall normalizations are physically meaningful and will be discussed shortly for the latter case, but they play no role for the present discussion.
To evaluate the imaginary part of the corresponding one-loop form factors,
we substitute the tree amplitude (\ref{eq: M++++}) into the phase-space integral in eq.~(\ref{eq: phase space})
as depicted in fig.~\ref{fig: one-loop double cut}\subref{subfig: L},\subref{subfig: T gluons}:
\begin{subequations}
\begin{alignat}{1}
\langle 1^a_-2^b_-|\mathcal{M} \otimes \mathcal{L}|0\rangle^{(0)} &= \frac{2g^2C_A}{16\pi} \int \frac{\operatorname{d}\!\Omega}{4\pi}\frac{1}{\cos^2\theta\sin^2\theta} \left(\tfrac12\delta^{ab}\l1'2'\rangle^2\right)\,,
\label{eq: Matrix element L}
\\
\langle 1^a_-2^b_+|\mathcal{M}\otimes T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}
&= \frac{2g^2C_A}{16\pi} \int \frac{\operatorname{d}\!\Omega}{4\pi}
\frac{
1}
{\cos^2\theta\sin^2\theta}
\Big(2\delta^{ab}\lambda'{}_1^\alpha\lambda'{}_1^\beta\tilde{\lambda}'{}_2^{\dot{\alpha}}\tilde{\lambda}'{}_2^{\dot{\beta}}\,\cos^4\theta\\[-0.2\baselineskip]\nonumber
&\hspace{0.32\textwidth}+ 2\delta^{ab}\tilde{\lambda}'{}_1^{\dot\alpha}\tilde{\lambda}'{}_1^{\dot\beta}\lambda'{}_2^\alpha\lambda'{}_2^\beta\,\sin^4\theta \operatorname{e}^{4i\phi}
\Big)
\,.
\end{alignat}\end{subequations}
Note that the tree form factors are evaluated with the rotated spinors (\ref{rotation}) parametrizing the two intermediate states in the cut.
The two terms in the last line come from the two possible intermediate helicities, of which only one is shown in fig.~\ref{subfig: T gluons}.
A priori, they look quite complicated; expanding out the first gives
\begin{equation}
\begin{aligned}
\lambda'{}_1^\alpha\lambda'{}_1^\beta\tilde{\lambda}'{}_2^{\dot{\alpha}}\tilde{\lambda}'{}_2^{\dot{\beta}} &=
(\lambda_1\cos\theta-\lambda_2\sin\theta \operatorname{e}^{i\phi})^\alpha
(\lambda_1\cos\theta-\lambda_2\sin\theta \operatorname{e}^{i\phi})^\beta
\\ &\hspace{5mm}\times (\tilde{\lambda}_2\cos\theta+\tilde{\lambda}_1\sin\theta \operatorname{e}^{i\phi})^{\dot\alpha} (\tilde{\lambda}_2\cos\theta+\tilde{\lambda}_1\sin\theta \operatorname{e}^{i\phi})^{\dot\beta}\,.
\end{aligned}
\end{equation}
However, the key is that ultimately the spinor structure is fixed by little-group weights, which are enforced by the azimuthal angle integration.
Indeed, we see that all terms with non-vanishing phases are killed by the $\phi$ integration!
Dropping these, the integral becomes simply proportional to the tree form factor,
as anticipated below eq.~(\ref{eq: gammaO practical}).
Hence, the ratio does not depend on the spinor indices and
\begin{equation} \label{eq: stress tensor YM}
\frac{\langle 1^a_-2^b_+|\mathcal{M} \otimes T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}{\langle 1^a_-2^b_+|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}
= \frac{2g^2C_A}{16\pi} \int_0^{\frac{\pi}{2}} 2\sin\theta\cos\theta\operatorname{d}\!\theta \frac{\cos^8\theta+\sin^8\theta}{\cos^2\theta\sin^2\theta}\,.
\end{equation}
Now we can observe that the divergences in the collinear limits $\theta\to 0,\pi/2$ cancel precisely against those in the Lagrangian density
in (\ref{eq: gammaO practical}), yielding, as anticipated, a convergent integral:
\begin{align}
\gamma^{(1)}_{\mathcal{L}} &\equiv -\frac{1}{\pi} \left(
\frac{\langle 1^a_-2^b_-|\mathcal{M} \otimes \mathcal{L}|0\rangle^{(0)}}{\langle 1^a_-2^b_-|\mathcal{L}|0\rangle^{(0)}} -
\frac{\langle 1^a_-2^b_+|\mathcal{M}\otimes T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}{\langle 1^a_-2^b_+|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}
\right)
\nonumber\\ &= -\frac{2g^2C_A}{16\pi^2} \int_0^{\frac{\pi}{2}} 2\sin\theta\cos\theta\operatorname{d}\!\theta \left(\frac{1}{\cos^2\theta\sin^2\theta}-\frac{\cos^8\theta+\sin^8\theta}{\cos^2\theta\sin^2\theta}\right)
\nonumber\\ &= -\frac{2g^2}{16\pi^2} \times \frac{11C_A}{3}\,. \label{eq: beta function YM}
\end{align}
Using the relation between the running of the Yang-Mills Lagrangian and the $\beta$-function
quoted earlier, eq.~(\ref{eq:gamma vs beta function}), we have therefore obtained the one-loop $\beta$-function:
\begin{equation}
\beta(g^2)= -\frac{2g^4}{16\pi^2}\times \frac{11C_A}{3}\,, \qquad \mbox{where}\qquad \beta(g^2)\equiv \mu\partial_\mu g^2(\mu) \,.
\end{equation}
This is in perfect agreement with the famous result, including, of course, the sign!
This example confirms that one-loop anomalous dimensions can be obtained as suitable differences between
eigenvalues of the tree-level $S$-matrix, or more precisely, of minus the phase of the $S$-matrix divided by $\pi$.
In the case above, the scattering phase is positive ($\mathcal{M}>0$ in eq.~(\ref{eq: M++++})), which is attributed to the attractive nature
of the interaction between opposite color charges (the scattering phase represents, roughly, minus the interaction energy).
This is the reason in this framework for the famous negative sign of the $\beta$-function.
More precisely, the reason is that the attraction is felt more strongly in the $s$-wave state (Lagrangian density) than in the $d$-wave state (stress tensor).
\subsection{Matter-field contributions} \label{ssec: matter}
It is instructive to see how the method works in the presence of fermions and scalars coupled to the Yang-Mills field.
Naively, since the Yang-Mills part of the Lagrangian density has no tree-level coupling to matter, one might worry that its anomalous dimension would be insensitive to these.
However, the infrared structure of the theory is modified and this is detected by the stress tensor in the denominator of the IR-safe ratio (\ref{eq: IR safe ratio}).
In QED, this would be the only contribution.
To find out how the stress tensor couples to fermions and scalars, one could
construct the stress tensor following the Noether procedure and apply standard Feynman rules.
We use a shortcut exploiting the symmetries of the problem.
The overall normalization (at least, relative to the gluon contribution)
will be important. It is fixed physically by requiring that the expectation value of the stress tensor in a state returns its momentum \cite{Peskin:1995ev}:
\begin{equation}
\langle 1_\Phi | T^{\alpha\beta,\dot\alpha\dot\beta}|1_\Phi\rangle = 2p_1^{\alpha\dot\alpha}p_1^{\beta\dot\beta} = \langle 1_\Phi\bar{1}_{\bar{\Phi}}| T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle\,,
\end{equation}
where in the second step we used crossing symmetry. Thus, the forward limit $p_2\to -p_1$ of the form factor is fixed.
For fermions there is an analogous equation, but one needs to be mindful of the
sign in the crossing relation (\ref{eq: crossing}) for each ${\bar{\psi}}$ in the initial state, so the condition is
\begin{equation}
\langle 1_{{\bar{\psi}}} | T^{\alpha\beta,\dot\alpha\dot\beta}|1_{{\bar{\psi}}}\rangle = 2p_1^{\alpha\dot\alpha}p_1^{\beta\dot\beta} = -\langle 1_{{\bar{\psi}}} \bar{1}_{{\psi}}| T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle\,.
\end{equation}
The other constraint is that the stress tensor is conserved: it must be orthogonal to $(p_1+p_2)$.
For scalars, as is well-known, this leaves an ambiguity which can be removed by imposing tracelessness (equivalent to symmetry in the spinor indices).
For both scalars and fermions, there is then a unique polynomial satisfying these constraints and little-group scaling:
\begin{equation}
\label{eq: form factor of energy momentum YM with fermions}
\begin{aligned}
\langle 1_{\bar\Phi} 2_{\Phi}|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle &= \tfrac13\big(
p_1^{\alpha{\dot{\alpha}}}p_1^{\beta{\dot{\beta}}}+p_2^{\alpha{\dot{\alpha}}}p_2^{\beta{\dot{\beta}}}
-p_1^{\alpha{\dot{\alpha}}}p_2^{\beta{\dot{\beta}}}-p_1^{\beta{\dot{\alpha}}}p_2^{\alpha{\dot{\beta}}}-p_1^{\alpha{\dot{\beta}}}p_2^{\beta{\dot{\alpha}}}-p_1^{\beta{\dot{\beta}}}p_2^{\alpha{\dot{\alpha}}}\big)\,,\\
\langle 1_{{\bar{\psi}}} 2_{{\psi}}|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle &=
\tfrac{1}{2}
\big(\lambda_1^\alpha\lambda_1^\beta \tilde{\lambda}_1^{\dot\alpha}\tilde{\lambda}_2^{\dot\beta}+
\lambda_1^\alpha\lambda_1^\beta \tilde{\lambda}_1^{\dot\beta}\tilde{\lambda}_2^{\dot\alpha}-
\lambda_1^\alpha\lambda_2^\beta \tilde{\lambda}_2^{\dot\alpha}\tilde{\lambda}_2^{\dot\beta}-
\lambda_1^\beta\lambda_2^\alpha \tilde{\lambda}_2^{\dot\alpha}\tilde{\lambda}_2^{\dot\beta}\big)\,.
\end{aligned}\end{equation}
In accordance with eq.~(\ref{eq: gammaO practical}), we now convolute these form factors with annihilation amplitudes into two gluons,
as illustrated in fig.~\ref{fig: one-loop double cut}\subref{subfig: T fermions} for fermions.
The relevant tree amplitudes are all concisely encoded in an $\mathcal{N}=4$ supersymmetric expression
using Nair's $\mathcal{N}=4$ on-shell superspace \cite{Nair:1988bq}, which generalizes the amplitude (\ref{eq: Parke-Taylor formula_trace}) to
\begin{equation} \label{eq: Nair superamplitude}
\mathcal{M}^{abcd}_{1234}\delta^{cd}=-2g^2C_A\delta^{ab}
\frac{\delta^8(\sum_{i=1}^{4}\lambda_i \tilde{\eta}_i)}{\l13\rangle\l32\rangle\l24\rangle\langle 41\rangle}\equiv
-2g^2C_A\delta^{ab}\frac{\prod_{A=1}^4 \sum_{1\leq i<j\leq4}\langle ij\rangle \tilde{\eta}_i\tilde{\eta}_j}{\l13\rangle\l32\rangle\l24\rangle\langle 41\rangle}\,.
\end{equation}
To insert a negative-helicity gluon, fermion, scalar or positive-helicity fermion on site $j$, one extracts, respectively,
four, three, two or one powers of $\tilde{\eta}_j$, giving the required amplitudes:
\begin{equation}
\label{eq: annihilation amplitude}
\begin{aligned}
\langle 1^a_-2^b_+|\mathcal{M}|1_{\Phi}'2_{\bar\Phi}'\rangle &=
-2g^2n_sT_s \delta^{ab} \frac{\l11'\rangle^2\l12'\rangle^2}{\langle 11'\rangle\l12'\rangle\l21'\rangle\l22'\rangle}=2g^2n_sT_s \delta^{ab} \frac{\cos^2\theta\sin^2\theta \operatorname{e}^{2i\phi}}{\cos^2\theta\sin^2\theta}\,,
\\
\langle 1^a_-2^b_+|\mathcal{M}|1_{{\bar{\psi}}}'2_{{\psi}}'\rangle &=
-2g^2(2n_f)T_f \delta^{ab} \frac{\l11'\rangle\l12'\rangle^3}{\langle 11'\rangle\l12'\rangle\l21'\rangle\l22'\rangle}=-4g^2n_fT_f \delta^{ab} \frac{\cos^3\theta\sin\theta \operatorname{e}^{i\phi}}{\cos^2\theta\sin^2\theta}\,,
\\
\langle 1^a_-2^b_+|\mathcal{M}|1_{{\psi}}'2_{{\bar{\psi}}}'\rangle &=
-2g^2(2n_f)T_f \delta^{ab} \frac{\l11'\rangle^3\l12'\rangle}{\langle 11'\rangle\l12'\rangle\l21'\rangle\l22'\rangle}=-4g^2n_fT_f \delta^{ab} \frac{\cos\theta\sin^3\theta \operatorname{e}^{3i\phi}}{\cos^2\theta\sin^2\theta}\,,
\end{aligned}
\end{equation}
and $\langle 1^a_-2^b_+|\mathcal{M}|1_{\bar\Phi}'2_{\Phi}'\rangle=\langle 1^a_-2^b_+|\mathcal{M}|1_{\Phi}'2_{\bar\Phi}'\rangle$.
Here, anticipating the contraction with the stress tensor, we have re-weighted the color-adjoint amplitude (\ref{eq: Nair superamplitude}) in accordance
to $n_s$ complex scalars and $n_f$ Dirac fermions (and thus $(2n_f)$ Weyl fermions) in representations where $\Tr[T^aT^b]=T_{s,f}\delta^{ab}$, with $T_F=\tfrac12$ in the fundamental representation.
The final step is to integrate this over phase space, weighted by the tree form factors in eqs.~(\ref{eq: form factor of energy momentum YM with fermions})
evaluated with the rotated spinors (\ref{rotation}).
Again, most terms drop out upon azimuthal integration, leaving, as expected, a result proportional to the tree form factor:
\begin{align}\label{eq: stress tensor QCD}
\frac{\langle 1^a_-2^b_+|\mathcal{M} \otimes T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}{
\langle 1^a_-2^b_+|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}} &\simeq
\frac{2g^2}{16\pi} \int_0^{\frac{\pi}{2}}
\frac{2\sin\theta\cos\theta \operatorname{d}\!\theta}{\cos^2\theta\sin^2\theta}\Big[C_A(\cos^8\theta+\sin^8\theta)
\\ & \hspace{10mm} +2 n_fT_f(\cos^6\theta\sin^2\theta +\sin^6\theta\cos^2\theta)+ 2n_sT_s \cos^4\theta\sin^4\theta\Big]\,. \nonumber
\end{align}
As a simple check, one can plug in the matter content of $\mathcal{N}=4$ SYM
(two adjoint Dirac fermions and three complex scalars: $n_f{=}2$, $n_s{=}3$, $T_f{=}T_s{=}C_A$),
and see that the bracket reduces to $C_A(\cos^2\theta+\sin^2\theta)^4=C_A$. This reproduces
the integrand for the Lagrangian density in eq.~(\ref{eq: Matrix element L}), as required by supersymmetry
since the stress tensor and Lagrangian density are in the same supermultiplet. The vanishing of the $\beta$-function in $\mathcal{N}=4$ is thus automatic in this formalism
and can be used as a simple check on the algebra.
For other theories, replacing the subtraction in eq.~(\ref{eq: beta function YM}) by (\ref{eq: stress tensor QCD}) and integrating, we reproduce
the well-known one-loop result for general matter content:
\begin{eqnarray}
\beta(g^2)&=& -\frac{2g^4}{16\pi^2}b_0\,,\qquad b_0\equiv \frac{11}{3}C_A -\frac{4}{3}n_f T_f -\frac{1}{3} n_s T_s\,.
\end{eqnarray}
\def\fermm_F{{\bar{\psi}}_F}
\def\fermp_F{{\psi}_F}
\def\fermm_{\bar F}{{\bar{\psi}}_{\bar F}}
\def\fermp_{\bar F}{{\psi}_{\bar F}}
In a theory with fermion masses like QCD, the running of mass parameters is also interesting.
At energies much higher than the masses (the situation where ``running" is meaningful), we expect this question
to be answerable within the massless theory. Writing a Dirac fermion as a combination
of positive- and negative-helicity fundamental Weyl fermions $\Psi=(\fermp_F,\fermm_F)$ and complex conjugate $\bar\Psi=(\fermm_{\bar F},\fermp_{\bar F})$,
the minimal form factor for the mass operator $\bar{\Psi}\Psi=\bar{\Psi}_a\Psi^a=(\fermm_{\bar F}{}_a\fermm_F^a+\fermp_{\bar F}{}_a\fermp_F^a)$
is $\langle 1_{\fermm_F}2_{\fermm_{\bar F}}|\bar{\Psi}\Psi|0\rangle=\l12\rangle$.
The required scattering amplitudes between fundamental and antifundamental fermions, for same and opposite helicity respectively, are then
\begin{equation}
\langle 1_{\fermm_F}2_{\fermm_{\bar F}}|\mathcal{M}| 1'_{\fermm_F}2'_{\fermm_{\bar F}}\rangle =
\frac{2g^2C_F}{\cos^2\theta\sin^2\theta}\,,
\qquad
\langle 1_{\fermm_F}2_{\fermp_{\bar F}}|\mathcal{M}| 1'_{\fermm_F}2'_{\fermp_{\bar F}}\rangle=\frac{2g^2C_F\cos^4\theta}{\cos^2\theta\sin^2\theta}\,,
\end{equation}
where the fundamental Casimir is $C_F=\frac{N_c^2-1}{2N_c}$ for gauge group SU(N$_c$). The positive signs again reflect the attractive gauge interaction.
We also need the pair production amplitude $\langle 1_{\fermm_F}2_{\fermp_{\bar F}}|\mathcal{M}|1'_-2'_+\rangle$, equal to minus the complex conjugate of (\ref{eq: annihilation amplitude}).
Armed with these and the above stress-tensor form factors for gluons and fermions, we compute
\begin{align}
\gamma^{(1)}_{\bar\Psi\Psi} &\equiv -\frac{1}{\pi} \left(
\frac{\langle 1_{\fermm_F}2_{\fermm_{\bar F}}|\mathcal{M} \otimes \bar\Psi\Psi|0\rangle^{(0)}}{\langle 1_{\fermm_F}2_{\fermm_{\bar F}}| \bar\Psi\Psi|0\rangle^{(0)}} -
\frac{\langle 1_{\fermm_F}2_{\fermp_{\bar F}}|\mathcal{M}\otimes T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}{\langle 1_{\fermm_F}2_{\fermp_{\bar F}}|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}
\right)
\nonumber\\ &= -\frac{2g^2C_F}{16\pi^2} \int_0^{\frac{\pi}{2}} 2\sin\theta\cos\theta\operatorname{d}\!\theta \left(\frac{1}{\cos^2\theta\sin^2\theta}-\frac{\cos^6\theta+\sin^6\theta}{\cos^2\theta\sin^2\theta}\right)
= -\frac{6g^2C_F}{16\pi^2}\,. \label{eq: fermion mass running}
\end{align}
Again, this is in agreement with the standard result, confirming that running-mass effects at short distances can be computed using unitarity with massless states.
\subsubsection{Comments on masses}
\label{eq: comment on masses}
Our discussions so far have been restricted to the $S$-matrix of strictly massless particles -- the dilatation operator $D$ is only defined on the massless $S$-matrix!
We believe that this is not a significant restriction. Rather, we believe it is entirely consistent with conventional applications of the renormalization group, where
a particle is either regarded as heavy and integrated out, or as light, in which case its mass is neglected.
These two effective descriptions are connected by so-called matching regions where the masses are important,
but which do not produce the kind of large logarithms that the renormalization group resums and which are the focus of this paper.
The running of relevant operators such as QCD masses can be correctly calculated within the massless theory in the high-energy regime,
as we have just explicitly verified.
With massive particles, one can get in addition momentum-independent logarithms.
For example, a massive tadpole integral\footnote{%
Such logarithms can appear from any integral with an explicit mass, not necessarily of tadpole topology.} gives
\begin{equation}
\int \frac{\mu^{2\varepsilon}\operatorname{d}\!^{4{-}2\varepsilon}l}{i (2\pi)^{4{-}2\varepsilon}} \frac{1}{l^2-m^2}= \frac{m^2}{16\pi^2}\left(\frac{1}{\varepsilon}+\log\frac{\mu^2}{m^2}+\ldots\right)\,. \label{eq: massless log}
\end{equation}
It is common in textbook presentations of the renormalization group to focus on ultraviolet divergences and therefore include such logarithms when computing $\beta$-functions.
Yet these logarithms lack an imaginary part and so they cannot be detected by unitarity.
Does this mean that the unitarity method is incomplete? We believe no, because the renormalization group can answer two distinct questions.
The first type of question regards the running of \emph{bare} couplings as a function of the short-distance cutoff. This is clearly of importance to lattice practitioners, for example.
The above momentum-independent logarithms are then clearly relevant (and possibly also power divergences as well as details of the short-distance dynamics).
The second type of question regards the optimal coupling to use to minimize large logarithms, for example in the perturbative calculation of a cross-section at a given energy scale.
This is the typical question of interest to collider physicists.
Momentum-independent logarithms are then clearly \emph{not} relevant: once the bare couplings have been tuned to cancel $\log(\mu)$ for one observable, the same tuning removes it from any physical observable.
Our conclusion is that unitarity, by throwing away the logarithms (\ref{eq: massless log}),\footnote{%
In massless contexts with evanescent operators, a similar distinction between physically observable logarithms versus bare ultraviolet divergences (poles in dimensional regularization) is also important \cite{Bern:2015xsa}.} correctly answers the second type of question.
\subsection{Twist-two operators and partial-wave amplitudes}
\def\theta{\theta}
A pleasant feature of the unitarity approach is that the $S$-matrix for just a few basic processes controls the anomalous dimension of essentially any operator.
Let us here discuss those operators which can be accessed using a color-singlet pair of partons, as considered so far.
Let us ignore spin for a moment and consider for simplicity two-particle form factors for a complex scalar.
Tree form factors are polynomial in $p_1^\mu,p_2^\mu$.
Factors of $(p_1{+}p_2)^\mu$ represent uninteresting total derivatives; these can be projected out by considering the forward case $p_2=-p_1$.
As $p_1^2=0$, only traceless tensors then survive.
Thus, the interesting polynomials of order $j$ represent operators of spin $j$ and dimension $j+2$ (the plus two is because any external on-shell parton carries dimension 1).
These are the form factors of twist-two operators:
\begin{equation}
\langle 1_\Phi \bar{1}_{\bar\Phi}|\mathcal{O}_m| 0\rangle = p_1^{\mu_1}\cdots p_1^{\mu_m} \quad\Leftrightarrow\quad
\mathcal{O}_m = i^n \,\bar\Phi \partial^{\mu_1}\cdots \partial^{\mu_m}\Phi\,.
\end{equation}
Let us act on these polynomial with the tree-level $S$-matrix.
Note that, even though this action is originally derived assuming that all particles carry positive energy, in the spinor
parametrization (\ref{rotation}) the phase-space integrals can be seamlessly continued to the forward case $p_2=-p_1$.
The rotated form factor in this parametrization is then a multiple of itself, since
\begin{equation}
p_1'^{\alpha{\dot{\alpha}}} \equiv \lambda_1'^\alpha\tilde{\lambda}_1'^{{\dot{\alpha}}} =\lambda_1^\alpha\tilde{\lambda}_1^{{\dot{\alpha}}}(\cos\theta-\sin\theta \operatorname{e}^{i\phi})(\cos\theta+\sin\theta \operatorname{e}^{-i\phi})
= p_1^{\alpha{\dot{\alpha}}}(\cos(2\theta)-i\sin(2\theta)\sin\phi).
\end{equation}
Using that the $S$-matrix for scalars does not depend on the azimuthal part of the scattering angle,
the latter can be integrated out immediately.
Temporarily rescaling the scattering angle $2\theta\to \theta$, we find that this
produces Legendre polynomials:
\begin{equation}
\int_0^{2\pi}\frac{\operatorname{d}\!\phi}{2\pi}\big(\!\cos\theta-i\sin\theta\sin\phi\big)^m=P_m(\cos\theta).
\end{equation}
For two complex scalars, the basic unitarity relation (\ref{eq: gamma UV plus IR}) thus becomes
\begin{equation}
\gamma^{(1)}_{\mathcal{O}_m} - \gamma^{(1)}_{\textrm{IR}} = -\frac{1}{16\pi^2} \int_0^{\pi} \frac{\sin\theta \operatorname{d}\!\theta}{2} \mathcal{M}^{(0)}(\cos\theta) P_m(\cos\theta) \equiv -\frac{1}{\pi} a_m^{(0)}\,.
\end{equation}
We recognize $a_m$ as the partial-wave amplitude with angular momentum $m$, which at leading order can be identified
with the phase of the $S$-matrix, normalized as $S_m=1+i\mathcal{M}_m = \operatorname{e}^{ia_m}$.
Thus, anomalous dimension are indeed minus the phase of the $S$-matrix, divided by $\pi$, as expected from eq.~\eqref{cute_eigenvalue_equation},
and two-particle states with definite angular momentum map to twist-two operators.
Let us apply this to a few examples. First consider twist-two operators with two identical complex scalars $Z$ in $\mathcal{N}=4$ SYM:
$\mathcal{O}_m=\Tr[Z\partial^+_mZ]$, where $m$ is even. Because it is in the same multiplet, the tree amplitude is the same as in eq.~(\ref{eq: Matrix element L}). Using the stress tensor
to subtract the infrared divergence via eq.~(\ref{eq: stress tensor QCD}), the formula becomes
\begin{equation}
\gamma^{(1)}_{\mathcal{O}_m} = -\frac{2g^2N_c}{16\pi^2} \int_0^{\pi} \frac{2\operatorname{d}\!\theta}{\sin\theta} \left(P_m(\cos\theta)-P_m(1)\right) = \frac{g^2N_c}{16\pi^2}\times 8S_1(m)\,,
\end{equation}
where $S_1(m)=\sum_{i=1}^m \tfrac{1}{i}$ denotes the harmonic sum. This is precisely the know result \cite{Kotikov:2000pm}.
In pure Yang-Mills, the similar partial-wave analysis requires partial waves for particles with spin.
These are more complicated than Legendre polynomials but the spinor parametrization provides a straightforward way to proceed.
Let us first record a formula for the evolution of an arbitrary operator which can decay to two particles at tree level,
which follows by combining the unitarity relation (\ref{eq: gammaO practical}), the matrix element (\ref{eq: M++++})
and the stress-tensor eigenvalue (\ref{eq: stress tensor QCD}):
\begin{equation}
\begin{aligned}
\gamma_\mathcal{O}^{(1)} \l1_-2_-| \mathcal{O}|0\rangle^{(0)} &= \frac{g^2C_A}{16\pi^2} \int_0^{2\pi} \frac{\operatorname{d}\!\phi}{2\pi}\int_0^{\frac{\pi}{2}}\frac{4\operatorname{d}\!\theta}{\cos\theta\sin\theta}
\left[\begin{array}{r}\big(\cos^8\theta+\sin^8\theta\big)\l1_-2_-| \mathcal{O}|0\rangle^{(0)}\\-\l1_-'2_-'| \mathcal{O}|0\rangle^{(0)}\end{array}\right],
\\
\gamma_\mathcal{O}^{(1)} \l1_-2_+| \mathcal{O}|0\rangle^{(0)} &= \frac{g^2C_A}{16\pi^2} \int_0^{2\pi} \frac{\operatorname{d}\!\phi}{2\pi}\int_0^{\frac{\pi}{2}}\frac{4\operatorname{d}\!\theta}{\cos\theta\sin\theta}
\left[\begin{array}{r}
\big(\cos^8\theta+\sin^8\theta\big)\l1_-2_+| \mathcal{O}|0\rangle^{(0)}\\
-\cos^4\theta\l1_-'2_+'| \mathcal{O}|0\rangle^{(0)}\\
-\sin^4\theta \operatorname{e}^{4i\phi}\l1_+'2_-'| \mathcal{O}|0\rangle^{(0)}\end{array}\right].
\end{aligned}\label{mainresult}
\end{equation}
In Yang-Mills theory, the leading-twist operators can be either in the vector-like Lorentz representation
$(\tfrac{m}{2},\tfrac{m}{2})$, or in chiral representations $(\frac{m}{2}+1,\frac{m}{2}-1)$ with $m\geq 1$.
Focusing on the former, which control the energy dependence of unpolarized parton distribution functions and are associated with
the polynomials ($m\geq 2$)\footnote{%
Note that even though $\lambda_2^\alpha \tilde{\lambda}_2^{{\dot{\alpha}}}\simeq -\lambda_1^\alpha \tilde{\lambda}_1^{{\dot{\alpha}}}$
has been used to simplify the form factor in the forward limit, we have not used the stronger condition $\tilde{\lambda}_2^{{\dot{\alpha}}}\simeq -\tilde{\lambda}_1^{{\dot{\alpha}}}$
to eliminate $\tilde\lambda_2$ because the phase-space integral using eq.~(\ref{rotation}) produces additional little-group phases that do not preserve this relation.
}
\begin{equation}
\langle 1_-\bar{1}_+|\mathcal{O}_{gg,m}|0\rangle= (\lambda_1^1\tilde{\lambda}_2^{\dot1})^{2} (\lambda_1^1\tilde{\lambda}_1^{\dot 1})^{m-2},
\end{equation}
this gives
\begin{multline}
\gamma_{gg,m}^{(1)} =
\frac{g^2C_A}{16\pi^2} \int_0^{2\pi} \frac{\operatorname{d}\!\phi}{2\pi}\int_0^{\frac{\pi}{2}}\frac{4\operatorname{d}\!\theta}{\cos\theta\sin\theta}
\big[
\cos^8\theta+\sin^8\theta\\ \qquad \qquad \qquad \,\,\,\,
- \cos^4\theta(\cos\theta{-}\sin\theta \operatorname{e}^{i\phi})^{m{+}2}(\cos\theta{+}\sin\theta \operatorname{e}^{-i\phi})^{m{-}2}
\\
-\sin^4\theta \operatorname{e}^{4i\phi}(\cos\theta{-}\sin\theta \operatorname{e}^{i\phi})^{m{-}2}(\cos\theta{+}\sin\theta \operatorname{e}^{-i\phi})^{m{+}2}
\big]\,.
\label{eq: YM evolution}
\end{multline}
We have checked for several values of $m$ that this reproduces precisely the moments of the DGLAP parton evolution
equation in Yang-Mills theory,
\begin{equation}
\gamma_{gg,m} = -\int_0^1 \operatorname{d}\! x x^{m{-}1} P_{gg}(x)\,,\quad
P_{gg}^{(1)}(x) = \frac{2g^2C_A}{16\pi^2}\left[
2\frac{1+x^4+(1-x)^4}{x(1-x)_+} +\frac{11}{3}\delta(1{-}x)\right]\,, \label{eq: DGLAP}
\end{equation}
as expected from the standard relation between twist-two operators and parton distribution functions \cite{Peskin:1995ev}.
Therefore, the tree-level scattering phases in Yang-Mills theory are indeed
the same as the anomalous dimensions of twist-two operators.
It would be nice to find a more direct mathematical map between eqs.~(\ref{eq: YM evolution}) and (\ref{eq: DGLAP}).
At higher loops, we warn the reader that since $S$ is a matrix, its \emph{phase} (as defined from its eigenvalues) need not agree with the phase of $2{\to}2$ $S$-matrix elements! Rather, when evaluating the product $\mathcal{M}F^*$ in the unitarity formula (\ref{cute_eigenvalue_equation}),
as shown in fig.~\ref{fig: two-loop cuts} one sees that $2{\to}3$ amplitudes and higher also contribute to the anomalous dimension of twist-two operators.
According to our main equation (\ref{cute_eigenvalue_equation}), anomalous dimensions are then
obtained by comparing this product $\mathcal{M}F^*$ with $-i(\operatorname{e}^{-i\pi D}-1)F^*$.
Using the dilatation operator $D$ given in eq.~(\ref{rg_equation}), one sees that at two-loops
this removes terms proportional to either the square of one-loop anomalous dimensions or to the one-loop $\beta$-function.
\def0.8{0.8}
\begin{figure}[tbp]
\centering
\begin{subfigure}[t]{0.3\textwidth}
\centering
$
\settoheight{\eqoff}{$\times$}%
\setlength{\eqoff}{0.5\eqoff}%
\addtolength{\eqoff}{-12.0\unitlength}%
\raisebox{\eqoff}{%
\fmfframe(2,2)(2,2){%
\begin{fmfchar*}(40,20)
\fmfright{vq}
\fmfleft{vp2,vp1}
\fmf{dbl_plain_arrow,tension=1.2}{vq,v1}
\fmf{plain_arrow_smallcut,left=0.7,tension=0.5}{v1,v2}
\fmf{plain_arrow_smallcut,right=0.7,tension=0.5}{v1,v2}
\fmf{plain_arrow,tension=1.2}{v2,vp1}
\fmf{plain_arrow,tension=1.2}{v2,vp2}
\fmfv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(0)}$,label.dist=0}{v1}
\fmfv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(1)}$,label.dist=0}{v2}
\fmffreeze
\fmfcmd{pair vertq, vertpone, vertptwo, vertpthree, vertpL, vertone, verttwo; vertone = vloc(__v1); verttwo = vloc(__v2); vertq = vloc(__vq); vertpone = vloc(__vp1); vertptwo = vloc(__vp2);}
\fmfiv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(0)}$,label.dist=0}{vertone}
\fmfiv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(1)}$,label.dist=0}{verttwo}
\fmfiv{decor.shape=circle,decor.filled=0,decor.size=19,label=$\scriptstyle \mathcal{M}^{(1)}$,label.dist=0}{verttwo}
\fmfdraw
\fmfiv{label=$\scriptstyle \mathcal{O}$}{vertq}
\fmfiv{label=$\scriptstyle 1$}{vertpone}
\fmfiv{label=$\scriptstyle 2$}{vertptwo}
\end{fmfchar*}%
}}%
$
\caption{\textcolor{white}{.}}
\label{fig: one two-loop double cut}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
\centering
$
\settoheight{\eqoff}{$\times$}%
\setlength{\eqoff}{0.5\eqoff}%
\addtolength{\eqoff}{-12.0\unitlength}%
\raisebox{\eqoff}{%
\fmfframe(2,2)(2,2){%
\begin{fmfchar*}(40,20)
\fmfright{vq}
\fmfleft{vp2,vp1}
\fmf{dbl_plain_arrow,tension=1.2}{vq,v1}
\fmf{plain_arrow,left=0.7,tension=0.25}{v1,v2}
\fmf{plain_arrow,right=0.7,tension=0.25}{v1,v2}
\fmf{phantom_smallcut,left=0.7,tension=0.25}{v1,v2}
\fmf{phantom_smallcut,right=0.7,tension=0.25}{v1,v2}
\fmf{plain_arrow,tension=1.2}{v2,vp1}
\fmf{plain_arrow,tension=1.2}{v2,vp2}
\fmfv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(0)}$,label.dist=0}{v2}
\fmfv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(1)*}$,label.dist=0}{v1}
\fmffreeze
\fmfdraw
\fmfcmd{pair vertq, vertpone, vertptwo, vertpthree, vertpL, vertone, verttwo; vertone = vloc(__v1); verttwo = vloc(__v2); vertq = vloc(__vq); vertpone = vloc(__vp1); vertptwo = vloc(__vp2);}
\fmfiv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(0)}$,label.dist=0}{verttwo}
\fmfiv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(1)*}$,label.dist=0}{vertone}
\fmfiv{decor.shape=circle,decor.filled=0,decor.size=19,label=$\scriptstyle F^{(1)*}$,label.dist=0}{vertone}
\fmfiv{label=$\scriptstyle \mathcal{O}$}{vertq}
\fmfiv{label=$\scriptstyle 1$}{vertpone}
\fmfiv{label=$\scriptstyle 2$}{vertptwo}
\end{fmfchar*}%
}}%
$
\caption{\textcolor{white}{.}}
\label{fig: other two-loop double cut}
\end{subfigure}
\begin{subfigure}[t]{0.3\textwidth}
\centering
$
\settoheight{\eqoff}{$\times$}%
\setlength{\eqoff}{0.5\eqoff}%
\addtolength{\eqoff}{-12.0\unitlength}%
\raisebox{\eqoff}{%
\fmfframe(2,2)(2,2){%
\begin{fmfchar*}(40,20)
\fmfright{vq}
\fmfleft{vp2,vp1}
\fmf{dbl_plain_arrow,tension=1.2}{vq,v1}
\fmf{plain_arrow,left=0.7,tension=0.25}{v1,v2}
\fmf{plain_arrow,right=0.7,tension=0.25}{v1,v2}
\fmf{phantom_smallcut,left=0.7,tension=0.25}{v1,v2}
\fmf{phantom_smallcut,right=0.7,tension=0.25}{v1,v2}
\fmf{plain_arrow,tension=0,tag=1}{v1,v2}
\fmf{phantom_smallcut,tension=0}{v1,v2}
\fmf{plain_arrow,tension=1.2}{v2,vp1}
\fmf{plain_arrow,tension=1.2}{v2,vp2}
\fmfv{decor.shape=circle,decor.filled=30,decor.size=24,label=$\scriptstyle F^{(0)}$,label.dist=0}{v1}
\fmfv{decor.shape=circle,decor.filled=10,decor.size=24,label=$\scriptstyle \mathcal{M}^{(0)}$,label.dist=0}{v2}
\fmffreeze
\fmfcmd{pair vertq, vertpone, vertptwo, vertpthree, vertpL, vertone, verttwo; vertone = vloc(__v1); verttwo = vloc(__v2); vertq = vloc(__vq); vertpone = vloc(__vp1); vertptwo = vloc(__vp2);}
\fmfiv{label=$\scriptstyle \mathcal{O}$}{vertq}
\fmfiv{label=$\scriptstyle 1$}{vertpone}
\fmfiv{label=$\scriptstyle 2$}{vertptwo}
\fmfipath{p[]}
\fmfiset{p1}{vpath1(__v1,__v2)}
\end{fmfchar*}%
}}%
$
\caption{\textcolor{white}{.}}
\label{fig: two-loop p1+p2 triple cut}
\end{subfigure}
\caption{Cut diagrams which contribute to the product $\mathcal{M}F^*$ at two-loop order.
}
\label{fig: two-loop cuts}
\end{figure}
\subsection{General operators at one-loop level and the \texorpdfstring{$\mathcal{N}=4$}{N=4} spin chain}
We conclude this section by discussing general operators at one-loop level.
The main issue is the cancellation of infrared divergences for multiple external partons.
In principle, one could use again matrix elements of the stress tensor, but since they are coupling-constant suppressed this is not so convenient.
At one-loop, the tight structure of infrared divergences however makes this unnecessary.
The one-loop infrared anomalous dimension (defined by the renormalization group equation (\ref{eq: RG equation mu}) for the IR- and UV-renormalized form factor)
takes a very specific form in any gauge theory, see for example \cite{Sterman:2002qn,Becher:2009cu}:
\begin{equation}
\label{eq: general structure of one-loop IR}
\gamma^{(1)}_{\textrm{IR}}(\{p_i\};\mu)=\frac{g^2}{4\pi^2}\sum_{i<j} T_i^aT_j^a \log \frac{\mu^2}{-s_{ij}} + \sum_i \gamma^{\rm coll.}_i\,,
\end{equation}
where $T_i^a$ denotes the gauge-group generator acting on particle $i$.
The fact that infrared divergences obey a renormalization group equation stems, of course,
from the general Wilsonian principle that disparate energy scales decouple;
we refer the reader to \cite{Feige:2014wja} for a recent explicit proof and further references.
Note that, in contrast to the ultraviolet case, infrared anomalous dimensions can depend explicitly on $\log\mu^2$ (at most linearly to any loop order),
reflecting double-logarithmic divergences from modes that are simultaneously soft and collinear.
The first term in \eqref{eq: general structure of one-loop IR}, coming from soft wide-angle radiation, can be identified with the integral over the
$1/(\sin^2\theta\cos^2\theta)$ term in eq.~\eqref{eq: Matrix element L},
which is the squared matrix element one would get from an integral over real radiation.
Therefore, the general one-loop dilatation operator (encoding all one-loop anomalous dimensions)
in an arbitrary gauge theory with matter contains a double-sum term,
from the sum over unitarity cuts and soft contribution to the anomalous dimension, together with a single-sum term accounting for remaining hard-collinear divergences:
\begin{align}
\label{eq: gamma general length}
\gamma^{(1)}_\mathcal{O}\langle p_1,\ldots,p_n| \mathcal{O}|0\rangle^{(0)}
=& -\frac{1}{\pi} \langle p_1,\ldots,p_n| \sum_{i<j}\left(\mathcal{M}^{2{\leftarrow}2}_{ij}+\frac{2g^2T_i^a T_j^a}{\sin^2\theta\cos^2\theta}\right)\otimes \mathcal{O}|0\rangle^{(0)}\\&{}{}+ \langle p_1,\ldots,p_n| \mathcal{O}|0\rangle^{(0)}
\times\sum_{i=1}^n \gamma^{\rm coll.}_i\,.\nonumber
\end{align}
Here, $\mathcal{M}_{ij}$ denotes the $2{\to}2$ amplitude acting on the final-state particles $i$ and $j$.
The two-body phase-space integral, which is represented by the convolution sign and
defined in eqs.~(\ref{eq: phase space})-(\ref{eq: one-loop measure}), is absolutely convergent for each term.
In QCD, matching and integrating the explicit expressions for the stress-tensor subtractions in eqs.~(\ref{eq: stress tensor QCD}) and (\ref{eq: fermion mass running}),
we get the one-loop collinear anomalous dimensions $\gamma^{\rm coll.}_g = -\frac{g^2b_0}{16\pi^2}$ and $\gamma^{\rm coll.}_\psi =-\frac{3g^2C_F}{16\pi^2}$,
again in agreement with standard results \cite{Becher:2009cu}.
The equality of the one-loop $\beta$-function and the collinear anomalous dimension (and the coefficient of $\delta(1{-}x)$ in eq.~(\ref{eq: DGLAP}))
can be attributed, in this framework,
to the simplicity of the Lagrangian form factor (\ref{eq: Matrix element L}), which exactly matches the soft (classical) infrared divergences.
An interesting special case of this formula is the planar limit of $\mathcal{N}=4$ super Yang-Mills.
In the planar limit, we consider single-trace operators, and $i$ and $j$ must be color-adjacent. Thus, we set
$j=i+1$ with $n+1\equiv 1$ identified following the cyclic invariance of the trace, and $T_i^aT_{i{+}1}^a\to -\tfrac{N_c}{2}$.
In this model, all states lie within one supermultiplet and are conveniently labelled
by polynomials in superspinors $\lambda_i,\tilde\lambda_i,\tilde{\eta}_i$ as in eq.~(\ref{eq: Nair superamplitude}); summing over internal helicities, one finds that
the supermomentum-conserving $\delta$-function simply forces the $\tilde{\eta}'$ to rotate like the $\tilde\lambda'$, so the right-hand side here will be evaluated with rotated superspinors
(\ref{rotation}). Finally, the planar $2{\to}2$ amplitude is equal to the first term in eq.~(\ref{eq: Parke-Taylor formula}), which is larger than
the color-singlet amplitude (\ref{eq: Parke-Taylor formula_trace}) by a factor $\cos^2\theta$.
Substituting it into the above formula, we thus get\footnote{In the planar limit, it is conventional to not symmetrize in the two cut particles;
it can be verified that after symmetrization, using that $\cot\theta{+}\tan\theta{=}\frac{1}{\cos\theta\sin\theta}$,
the IR subtraction is exactly as in eq.~(\ref{eq: gamma general length}) with $\gamma^{\rm coll.}_{\mathcal{N}=4}=0$.}
\begin{align}
\gamma^{(1)}_\mathcal{O} \langle 1, \ldots, n|\mathcal{O}|0\rangle^{(0)} =
\frac{4g^2N_c}{16\pi^2} \sum_{i=1}^n \int_0^{2\pi} \frac{\operatorname{d}\!\phi}{2\pi}\int_0^{\frac{\pi}{2}} \operatorname{d}\!\theta \cot\theta \left(
\begin{array}{l}\phantom{+}
\langle 1,\ldots, i,i{+}1,\ldots, n|\mathcal{O}|0\rangle^{(0)} \\ -
\langle 1,\ldots, i',(i{+}1)',\ldots, n|\mathcal{O}|0\rangle^{(0)}\end{array}\right)\,.
\end{align}
This formula is precisely the one written down by Zwiebel for the one-loop dilatation operator in planar $\mathcal{N}=4$ SYM \cite{Zwiebel:2011bx},
which in some way led to this work.
In various subsectors, the expression above reduces for example to the Hamiltonian of the integrable SU(2) or SL(2) Heisenberg spin chain,
revealing the integrability of the theory \cite{Minahan:2002ve,Beisert:2003yb}.
As far as we know, the original motivation of \cite{Zwiebel:2011bx} was based on symmetries: the one-loop dilatation operator and tree-level four-point $S$-matrix being both completely fixed by Yangian symmetry up to a multiplicative constant, they may be proportional to each other.
This was then understood more directly from generalized unitarity \cite{Wilhelm:2014qua}.
In this paper, we have derived this formula using conventional unitarity and given a quantitative
extension to an arbitrary weakly coupled field theory, eq.~(\ref{eq: gamma general length}).
In large $N_c$-QCD, we thus expect that upon substituting the appropriate quark and gluon $2{\to}2$ tree amplitudes as in eq.~(\ref{mainresult}),
the formula will reproduce the one-loop dilatation operator from ref.~\cite{Beisert:2004fv}.
It would also be interesting to specialize the formula to the Standard Model and compare with the dimension-six anomalous dimensions, see for example
ref.~\cite{Alonso:2013hga};
certain qualitative features, such as zeros that are not obvious from Feynman diagrams, are nicely explained from unitarity and
on-shell tree-level helicity conservation rules \cite{Cheung:2015aba}.
\section{Length-changing effects and towards higher loops: Yukawa theory}
\label{sec: towards higher loops}
Let us now look at Yukawa theory, where we will encounter several new effects looking at operators of higher length and at higher loops.
These include mixing between operators of different lengths and the cancellation of logarithms between different cuts.
For illustration, it will be sufficient to consider a theory with one real scalar and one Weyl fermion,
with interaction Lagrangian
\begin{equation}
\label{eq: interaction Lagrangian}
\mathcal{L}_{\rm int}=
-\lambda\mathcal{O}_\lambda -y \mathcal{O}_y\quad\text{with}\quad\mathcal{O}_\lambda=\frac{1}{4!}\phi^4\quad\text{and}\quad\mathcal{O}_y=\frac{1}{2}(\psi\psi\phi+\text{h.c.})\,.
\end{equation}
The minimal form factors of the operators $\mathcal{O}_\lambda$ and $\mathcal{O}_y$ are
\begin{equation} \label{eq: form factor}
\langle 1_\phi2_\phi3_\phi4_\phi| \mathcal{O}_\lambda|0\rangle = 1\,,\quad
\langle 1_{{\bar{\psi}}}2_{{\bar{\psi}}}3_\phi| \mathcal{O}_y|0\rangle = \l12\rangle\,,\quad
\langle 1_{{\psi}}2_{{\psi}}3_\phi| \mathcal{O}_y|0\rangle = [12]\,.
\end{equation}
Correspondingly, the elemental scattering amplitudes are
\begin{equation}
\label{eq: Yukawa theory elemental scattering amplitudes}
\begin{aligned}
\mathcal{M}_{1_\phi2_\phi3_\phi4_\phi} &= -\lambda\,,\qquad
\mathcal{M}_{1_{{\bar{\psi}}}2_{{\bar{\psi}}}3_\phi} &= -y\l12\rangle\,, \qquad
\mathcal{M}_{1_{{\psi}}2_{{\psi}}3_\phi} &= -y[12]\,.
\end{aligned}
\end{equation}
One can check that the relative signs between the latter two amplitudes is consistent with unitarity,
so that $\langle 1_\phi|2_\fermm3_{\bar{\psi}}\rangle\l2_\fermm3_{\bar{\psi}}|1_\phi\rangle\geq0$ as it should, using the crossing relation
(\ref{eq: crossing}).
Other amplitudes can be obtained using the factorization on poles:
\begin{equation}
\label{eq: Yukawa theory scattering amplitudes}
\begin{aligned}
\mathcal{M}_{1_{{\bar{\psi}}}2_{{\psi}}3_\phi4_\phi} &= y^2\left(\frac{\l13\rangle}{\l23\rangle}+\frac{\l14\rangle}{\l24\rangle}\right)\,, \\
\mathcal{M}_{1_{{\psi}}2_\fermp3_{{\bar{\psi}}}4_{{\bar{\psi}}}} &= y^2\frac{\l34\rangle}{\l12\rangle}\,, \\
\mathcal{M}_{1_{{\psi}}2_{{\psi}}3_{{\psi}}4_{{\psi}}} &= 3y^2 \frac{[34]}{\l12\rangle}\,, \\
\mathcal{M}_{1_{{\psi}}2_{{\psi}}3_\phi4_\phi5_\phi}&= y\lambda \frac{1}{\l12\rangle}-y^3\left( \frac{\l35\rangle}{\l13\rangle\l25\rangle} + \mbox{5 permutations of (345)}\right)\,.
\end{aligned}
\end{equation}
The signs of these amplitudes will be significant since in this approach they ultimately determine the sign of the anomalous dimension;
they are fixed for example by the factorization of trees $\langle 12|\mathcal{M}|34\rangle \to \langle 12|\mathcal{M}|i\rangle \frac{1}{\l12\rangle[12]} \langle i|\mathcal{M}|34\rangle$ in the limit where $p_i=(p_1+p_2)$ becomes null.
From the above scattering amplitudes and form factors, we will calculate the anomalous-dimension matrix
\begin{equation}
\left(\mu\frac{\partial}{\partial \mu} + \sum_{a=y,\lambda}\beta(a) \frac{\partial}{\partial a}\right)
\left(\begin{array}{c} \mathcal{O}_y\\ \mathcal{O}_\lambda\end{array}\right)
=\left(\begin{array}{cc} \gamma_{yy}&\gamma_{y\lambda}\\ \gamma_{\lambda y}&\gamma_{\lambda\lambda}\end{array}\right)
\left(\begin{array}{c} \mathcal{O}_y\\ \mathcal{O}_\lambda\end{array}\right)\,.
\end{equation}
From it one can then get $\beta$-functions,
using a generalization of the relation (\ref{eq:gamma vs beta function}) that we used in the Yang-Mills case.
We briefly recall its derivation \cite{KlubergStern:1974rs}.
First we note that we have normalized the operators so that their form factors (\ref{eq: form factor}) restricted to zero total momentum
are precisely the derivatives of the $S$-matrix with respect to the corresponding coupling:
\begin{equation}
\mathcal{F}_a=-\frac{\partial}{\partial a} \mathcal{M} \,.
\end{equation}
One now considers the RG equation for the UV (not IR)-renormalized amplitude and form factor (that is, contrary to what was done so far in this paper,
here we consider independent ultraviolet and infrared renormalization scales):
\begin{equation}
\left(\mu_{\UV}\frac{\partial}{\partial \mu_{\UV}} + \sum_{b=y,\lambda}\beta(b) \frac{\partial}{\partial b}\right)\mathcal{M}=0\,,
\qquad
\left(\mu_{\UV}\frac{\partial}{\partial \mu_{\UV}} + \sum_{b=y,\lambda}\beta(b) \frac{\partial}{\partial b}\right)\mathcal{F}_a=-\sum_{b=y,\lambda}\gamma_{ab}\mathcal{F}_b\,.
\end{equation}
Deriving the first equation with respect to the coupling and comparing with the second equation gives the desired relation:
\begin{equation}
\frac{\partial}{\partial a} \beta(b)=\gamma_{ab}\,,\qquad a,b= \lambda \, \mbox{ or }\, y\,. \label{eq: beta from gamma}
\end{equation}
\subsection{IR structure and diagonal elements}
The diagonal (length-preserving) elements of the mixing matrix can be calculated straightforwardly using the by-now familiar procedure
of the preceding section: we act on form factors with the $2{\to}2$ tree amplitudes,
employing the stress tensor to remove the infrared (and collinear) contributions. There are no new subtleties but if anything it is instructive to carry through this exercise.
Using the crossing relation (\ref{eq: crossing}) and the spinor products (\ref{eq: rotated spinor products}), the matrix elements
we will need are easily obtained from \eqref{eq: Yukawa theory scattering amplitudes}:
\begin{equation}\begin{aligned}
\langle 1_\phi2_\phi| \mathcal{M}|1'_\fermm2'_{\psi}\rangle&=-\langle 1_\fermm2_{\psi}|\mathcal{M}|1'_\phi2'_\phi\rangle^*= y^2\left(\frac{\cos\theta}{\sin\theta}-\frac{\sin\theta}{\cos\theta}\right)\operatorname{e}^{-i\phi}\,,
\\
\langle 1_\fermm2_{\bar{\psi}} |\mathcal{M}|1'_\fermm2'_{\bar{\psi}}\rangle&=\langle 1_\fermm2_{\psi} |\mathcal{M}|1'_\fermm2'_{\psi}\rangle= -\langle 1_\fermm2_{\psi} |\mathcal{M}|1'_\fermp2'_{\bar{\psi}}\rangle=-y^2\,,
\\
\langle 1_\fermm2_{\bar{\psi}} |\mathcal{M}|1'_\fermp2'_{\psi}\rangle &=-3y^2\,,
\qquad
\langle 1_\fermm2_\phi |\mathcal{M}|1'_\fermm2'_\phi\rangle= -y^2\frac{1+\cos^2\theta}{\cos\theta}\,. \label{eq: Yukawa rotated amplitudes}
\end{aligned}\end{equation}
Multiplying the first by the tree form factor for the stress tensor to fermions, given in eq.~(\ref{eq: form factor of energy momentum YM with fermions}),
and performing the azimuthal integrals then gives
\begin{equation}\begin{aligned}
2\gamma^{\rm coll.}_{\phi} &\equiv \frac{1}{\pi}
\frac{\langle 1_\phi2_\phi|\mathcal{M}\otimes T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}{\langle 1_\phi2_\phi|T^{\alpha\beta,\dot\alpha\dot\beta}|0\rangle^{(0)}}
\\ &= \frac{1}{16\pi^2}\int_0^{\frac{\pi}{2}} 2\cos\theta\sin\theta\operatorname{d}\!\theta
\left(-\frac14\lambda \big(1+\cos(4\theta)\big)+ 6y^2 \cos^2(2\theta)\right)
= \frac{2y^2}{16\pi^2}
\\
\Rightarrow \gamma^{\rm coll.}_\phi&=\frac{y^2}{16\pi^2}\,,
\end{aligned}\label{eq: collinear phi}\end{equation}
where we have included also the scalar ($\lambda$ term) and antifermion (factor of 2) in the cut.
The $\lambda$ term integrates to zero: there are no one-loop IR divergences in pure $\phi^4$ theory, as expected.
Here we remark that, even though the form factors (\ref{eq: form factor of energy momentum YM with fermions}) contain many terms,
because they are fixed by symmetry the algebra is highly redundant and just the coefficient
of one term, for example $p_1^{\alpha{\dot{\alpha}}}p_1^{\beta{\dot{\beta}}}$, is enough to determine the anomalous dimension.
Considering similarly the form factor for two fermions, we find
\begin{equation}\begin{aligned}
2\gamma^{\rm coll.}_{\psi} &= \frac{1}{16\pi^2}\int_0^{\frac{\pi}{2}} 2\cos\theta\sin\theta\operatorname{d}\!\theta \left(
2y^2\cos^2(2\theta)-y^2\cos(4\theta)
\right)
= \frac{y^2}{16\pi^2}
\quad
\Rightarrow \gamma^{\rm coll.}_\psi=\frac{\tfrac12y^2}{16\pi^2}\,,
\end{aligned}\end{equation}
where we have used the second, fourth and fifth of the amplitudes in eq.~(\ref{eq: Yukawa rotated amplitudes}).
With the infrared contributions under control, we can now calculate the diagonal matrix elements,
which is particularly trivial for the $\phi^4$ vertex correction since the four-scalar amplitude is just a constant so
each matrix element gives a factor $-\lambda/(16\pi^2)$:
\begin{align}
\gamma^{(1)}_{\lambda\lambda} &= 4\gamma^{\rm coll.}_\phi
-\frac{1}{\pi}
\frac{\langle 1_\phi2_\phi3_\phi4_\phi|\left(\mathcal{M}_{12}+\mathcal{M}_{13}+\mathcal{M}_{14}+\mathcal{M}_{23}+\mathcal{M}_{24}+\mathcal{M}_{34}\right) \otimes \mathcal{O}_\lambda|0\rangle^{(0)}}{\langle 1_\phi2_\phi3_\phi4_\phi|\mathcal{O}_\lambda|0\rangle^{(0)}}
\nonumber \\
&= \frac{4y^2}{16\pi^2} + \frac{6\lambda}{16\pi^2} \,.
\label{one_loop_gamma_ll}
\end{align}
The $y^2$ term comes entirely from the collinear divergences, following eq.~(\ref{eq: gamma general length}).
In $\phi^4$ theory, it would be absent and, reassuringly, the relation (\ref{eq: beta from gamma}) would give
\begin{equation}
\beta(\lambda) = \frac{3\lambda^2}{16\pi^2} \qquad \mbox{($\phi^4$ theory)}\,,
\end{equation}
which is of course the standard result.
For the Yukawa vertex renormalization, we have some angular integrals to do,
involving the third, sixth and two permutations of the seventh term in (\ref{eq: Yukawa rotated amplitudes}):
\begin{align}
\label{one_loop_gamma_yy}
\gamma^{(1)}_{yy} &= 2\gamma^{\rm coll.}_{{\bar{\psi}}}+\gamma^{\rm coll.}_\phi
-\frac{1}{\pi}
\frac{\langle 1_{{\bar{\psi}}}2_{{\bar{\psi}}}3_\phi|\left(\mathcal{M}_{12}+\mathcal{M}_{13}+\mathcal{M}_{23}\right) \otimes \mathcal{O}_y|0\rangle^{(0)}}{\langle 1_{{\bar{\psi}}}2_{{\bar{\psi}}}3_\phi|\mathcal{O}_y|0\rangle^{(0)}}
\\ &= \frac{2y^2}{16\pi^2}-\frac{1}{16\pi^2} \int_0^{\frac{\pi}{2}}2\cos\theta\sin\theta\operatorname{d}\!\theta\left(-4y^2-2y^2(1+\cos^2\theta)-2y^2(1+\sin^2\theta)\right)
= \frac{12y^2}{16\pi^2} \,. \nonumber
\end{align}
Note that, even though the Yukawa interaction between identical fermions is often said to be attractive, the matrix elements here are mostly negative,
thus leading to a positive anomalous dimension and a positive contribution to the $\beta$-function.
The difference is because the conventional statement applies to non-relativistic massive fermions while we are looking here at
the ultrarelativistic case where the amplitude involves a helicity flip and is quite different.
\subsection{Length-increasing effects: Yukawa coupling contributing to \texorpdfstring{$\phi^4$}{phi to the fourth}} \label{ssec: increasing}
We now turn to some novel effects not discussed earlier -- at one-loop we can also have length-increasing mixing,
for example between the operators $\mathcal{O}_\lambda$ and $\mathcal{O}_y$.
In terms of the unitarity method, this will involve the $2{\to}3$ amplitude acting on the minimal form factor,
as well as the $2{\to}2$ scattering acting on the \emph{non-minimal} form factor:
\begin{equation}
\label{eq: length increasing mixing}
\begin{aligned}
\gamma^{(1)}_{y\lambda} &= -\frac{1}{\pi}
\frac{\langle 1_\phi2_\phi3_\phi4_\phi|\left(\mathcal{M}^{2\leftarrow2}_{12}+\mathcal{M}^{2\leftarrow2}_{13}+\mathcal{M}^{2\leftarrow2}_{14}+\mathcal{M}^{2\leftarrow2}_{23}+\mathcal{M}^{2\leftarrow2}_{24}+\mathcal{M}^{2\leftarrow2}_{34}\right) \otimes \mathcal{O}_y|0\rangle^{(0)}}{\langle 1_\phi2_\phi3_\phi4_\phi|\mathcal{O}_\lambda|0\rangle^{(0)}}\\
&\phantom{{}=} -\frac{1}{\pi}
\frac{\langle 1_\phi2_\phi3_\phi4_\phi|\left(\mathcal{M}^{3\leftarrow2}_{123}+\mathcal{M}^{3\leftarrow2}_{124}+\mathcal{M}^{3\leftarrow2}_{134}+\mathcal{M}^{3\leftarrow2}_{234}\right) \otimes \mathcal{O}_y|0\rangle^{(0)}}{\langle 1_\phi2_\phi3_\phi4_\phi|\mathcal{O}_\lambda|0\rangle^{(0)}}\,.
\end{aligned}
\end{equation}
The subscripts on the amplitude indicate the final state partons to which it is connected.
From the second line we get a rather simple $\lambda y$ term (see the last amplitude in eq.~(\ref{eq: Yukawa theory scattering amplitudes})),
but for $y^3$ contributions there will be a non-trivial interplay between the two lines.
However, the sum of terms should give a polynomial since it is the form factor of a local operator.
Anticipating that this will require cancellations, here we organize the terms into cuts of Feynman diagrams (since Feynman diagrams make locality manifest).
For example, consider the three cuts of the fermion box with on-shell scalars $p_1,p_2,p_3$ shown in fig.~\ref{fig: cancellation in pictures}.
The first cut is related to $\mathcal{M}_{12}^{2\leftarrow 2}$ multiplied by the non-minimal form factor
\begin{equation}
\langle 1_\fermp2_\fermm3_\phi4_\phi| \mathcal{O}_y|0\rangle = -y\left( \frac{\langle 13\rangle}{\langle 23\rangle}+\frac{\langle 14\rangle}{\langle 24\rangle}-\frac{[23]}{[13]}-\frac{[24]}{[14]}\right).
\end{equation}
The sign of this expression can be verified by noting that for zero total momentum this is equal to minus
the $y$ derivative of the $2{\to}2$ $S$-matrix element given in the first line of eq.~(\ref{eq: Yukawa theory scattering amplitudes}).
The first cut comes from the first term in the amplitude (\ref{eq: Yukawa theory scattering amplitudes}) multiplied by the first term in the form factor:
\begin{equation}
\langle 1_\phi2_\phi| \mathcal{M}|1'_\fermp2'_{\bar{\psi}}\rangle_{\text{first term}}\langle 1_{\psi}'2_{\bar{\psi}}'3_\phi4_\phi| \mathcal{O}_y|0\rangle_{\text{first term}}=y^3\frac{\langle 2'1\rangle}{\langle 1' 1\rangle}\frac{\langle 1' 3\rangle}{\langle 2' 3\rangle}\,.
\end{equation}
In the parametrization \eqref{rotation}, the phase-space integral reads
\begin{equation}
\begin{aligned}
\int \frac{\operatorname{d}\!\Omega}{4\pi}\frac{\langle 2'1\rangle}{\langle 1' 1\rangle}\frac{\langle 1' 3\rangle}{\langle 2' 3\rangle}&=\int_0^{2\pi}\frac{\operatorname{d}\!\phi}{2\pi} \int_0^{\frac{\pi}{2}}\operatorname{d}\!\theta
\, 2\cos^2\theta \frac{\l13\rangle\cos\theta-\operatorname{e}^{i\phi}\l23\rangle\sin\theta }{\l13\rangle\sin\theta+\operatorname{e}^{i\phi} \l23\rangle\cos\theta}\,,
\end{aligned}
\end{equation}
where we have dropped the prefactor $-\frac{y^3}{16\pi^2}$ in the relation to the anomalous dimension.
A good way to perform the $\phi$ integral is as a contour integral over $z=\operatorname{e}^{i\phi}$ along the unit circle, which allows us to use Cauchy's residue theorem, obtaining
\begin{equation}
\label{eq: result double-cut two-particle channel}
\begin{aligned}
- \int_0^{\frac{\pi}{2}}\operatorname{d}\!\theta \,2\cos^2\theta \left(\frac{\cos\theta }{\sin\theta }-\frac{1}{\cos\theta\sin\theta} \Theta\left(1-\left|\frac{\langle 13\rangle\sin\theta}{\langle 23\rangle\cos\theta}\right|\right) \right)=1+\log \frac{s_{23}}{s_{13}+s_{23}}\,.
\end{aligned}
\end{equation}
The step function $\Theta$ arises from whether the pole from the denominator is inside the unit circle.
The result from the double cut in the other two-particle channel can be obtained by replacing $1\leftrightarrow3$ in \eqref{eq: result double-cut two-particle channel}.
In this way, we have accounted for 2 out of the $2\times 6\times 2\times 4=96$ terms in the first line of eq.~(\ref{eq: length increasing mixing})
(one of the 2's is from exchanging ${\bar{\psi}}$ and ${\psi}$).
We now consider the double cut in the three-particle channel.
We require both the three-point form factors in \eqref{eq: form factor} and the five-particle amplitude in the last line of \eqref{eq: Yukawa theory scattering amplitudes}.
As before, we will focus on one particular term corresponding to the third diagram of fig.~\ref{fig: cancellation in pictures}:
\begin{equation}
\begin{aligned}
\langle 1_\phi2_\phi3_\phi| \mathcal{M}|1'_\fermm2'_{\bar{\psi}}\rangle_{\text{first $y^3$ term}}\, \langle 1_{\bar{\psi}}'2_{\bar{\psi}}'4_\phi| \mathcal{O}_y|0\rangle=
\frac{\langle 1 3\rangle}{\langle 1' 1\rangle\langle 2' 3\rangle}\langle 1'2'\rangle \, .
\end{aligned}
\end{equation}
We want to parametrize $p_1'$ and $p_2'$ as a rotation of suitable basis spinors.
In contrast to the cases above, we cannot simply take the base vectors to be external ones. Instead, we choose
\begin{equation}
p_a=p_1 \frac{s_{123}}{s_{12}+s_{13}}\,, \qquad p_b=p_2+p_3-p_1 \frac{s_{23}}{s_{12}+s_{13}}\,,
\end{equation}
which are both on-shell and satisfy $p_a+p_b=p_1+p_2+p_3$. Corresponding spinors are
\begin{equation}
\lambda_a=\lambda_1\sqrt{\frac{s_{123}}{s_{12}+s_{13}}}\, ,\qquad \lambda_b=([12]\lambda_2+[13]\lambda_3)\frac{1}{\sqrt{s_{12}+s_{13}}}\, .
\end{equation}
We then find
\begin{equation}
\begin{aligned}
\int \frac{\operatorname{d}\!\Omega}{4\pi}\frac{\langle 1'2'\rangle\langle 1 3\rangle}{\langle 1' 1\rangle\langle 2' 3\rangle}
&=\int_0^{2\pi}\frac{\operatorname{d}\! \phi}{2\pi} \int_0^{\frac{\pi}{2}}\operatorname{d}\!\theta \frac{2\cos\theta}{\sin\theta+\operatorname{e}^{i\phi}\cos\theta \frac{[12]\l23\rangle}{\l13\rangle\sqrt{s_{123}}}}\,,
\end{aligned}
\end{equation}
and, doing the $\phi$ integral again using Cauchy's theorem, we obtain
\begin{equation}
\label{eq: double-cut three-particle channel}
\begin{aligned}
\int_0^{\frac{\pi}{2}}\operatorname{d}\!\theta \,2\frac{\cos\theta}{\sin\theta}\Theta\left(1-\left|\frac{[12]\langle 23\rangle\cos\theta}{\langle 13\rangle\sqrt{s_{123}}\sin\theta}\right|\right) = \log \left(\frac{\left(s_{12}+s_{13}\right) \left(s_{13}+s_{23}\right)}{s_{12} s_{23}}\right)\, .
\end{aligned}
\end{equation}
Summing (\ref{eq: result double-cut two-particle channel}), its image under $1\leftrightarrow3$ and \eqref{eq: result double-cut two-particle channel} to get the three cuts in fig.~\ref{fig: cancellation in pictures} finally gives
\begin{equation}
\label{eq: cancellation of logarithms}
\left(1+\log \frac{s_{23}}{s_{13}+s_{23}}\right)+ \left(1+\log \frac{s_{12}}{s_{12}+s_{13}}\right)+ \log \left(\frac{\left(s_{12}+s_{13}\right) \left(s_{13}+s_{23}\right)}{s_{12} s_{23}}\right)=2\,.
\end{equation}
As expected, the dependence on the kinematic variables has cancelled!
Restoring the factor $-y^3/(16\pi^2)$ and multiplying by 48 then gives the $y^3$ term in the anomalous dimension.
As already mentioned, there is also a simpler piece proportional to $y\lambda$, which comes only from the second line of eq.~(\ref{eq: length increasing mixing}) and involves the comparatively simpler
amplitude given in the last line of \eqref{eq: Yukawa theory scattering amplitudes}. In total, we thus get
\begin{equation}
\gamma_{y\lambda}=-\frac{96y^3}{16\pi^2}+\frac{8y\lambda}{16\pi^2}\,.
\end{equation}
Of course, the first term could have been obtained much more easily by extracting the ultraviolet-divergent part of the fermion box diagram. But this examples shows how,
through a non-trivial interplay between $S$-matrix elements and form factors responsible from the cancellation of logarithms (\ref{eq: cancellation of logarithms}),
the ultraviolet properties of the theories are also encoded in on-shell amplitudes with finite momentum.
Since all cuts ended up being computed by residues using Cauchy's formula, we can track the cancellations to the fact that the residues on triple cuts agree
regardless of the order in which the propagators are cut.
Physically, this is a consequence of the factorization of amplitudes and form factors on their poles.
Understanding how to systematize such cancellations would be of great help for applications to the dilatation operator at lengths $\geq 3$ at higher-loops,
especially in gauge theories where the comparative simplicity of on-shell amplitudes adds a practical advantage to the method.
\begin{figure}[htbp]
\centering
$
\begin{aligned}
\FDinline[boxplus,big,bslashdoublecut,fourlabels,labelone=p_1,labeltwo=p_2,labelthree=p_3,labelfour=p_4]
\end{aligned}
\qquad
+
\quad
\begin{aligned}
\FDinline[boxplus,big,slashdoublecut,fourlabels,labelone=p_1,labeltwo=p_2,labelthree=p_3,labelfour=p_4]
\end{aligned}
\qquad
+
\quad
\begin{aligned}
\FDinline[boxplus,big,doublecut,fourlabels,labelone=p_1,labeltwo=p_2,labelthree=p_3,labelfour=p_4]
\end{aligned}
$
\caption{Three cuts of the box integral among which logarithms cancel, see eq.~(\ref{eq: cancellation of logarithms}).}
\label{fig: cancellation in pictures}
\end{figure}
\subsection{Length-decreasing effects: a simple two-loop contribution}
Finally, we consider the length-decreasing mixing of $\mathcal{O}_\lambda$ into $\mathcal{O}_y$.
An important feature is that such mixing is not possible at one-loop: it would require a cut in a massless channel (with a $2{\to} 1$ amplitude on one side),
which of course is kinematically impossible.
Therefore, the first length-decreasing effects occur at two-loops, through a $3{\to} 2$ amplitude integrated over a 3-particle cut.
There exist efficient modern techniques to deal with such two-loop cut integrals, notably by using integration-by-parts techniques and so-called reverse unitarity, see for example \cite{Anastasiou:2002yz}.
Here, in line with previous examples, we adopt a low-tech approach and parametrize directly the angular integrals.
A price to pay is that we have to use different parametrizations for different terms in the amplitude.
We use eq.~(5.23) and preceding ones from \cite{Zwiebel:2011bx}:
\begin{equation}
\begin{aligned}
{\lambda_1'}^\alpha&=\lambda_1^\alpha \cos\theta_2-\operatorname{e}^{i \phi } \lambda_2^\alpha \cos\theta_1 \sin\theta_2 \,,\\
{\lambda_2'}^\alpha&=\lambda_1^\alpha \sin\theta_2 \cos\theta_3+\operatorname{e}^{i \phi } \lambda_2^\alpha \left(\cos\theta_1 \cos\theta_2 \cos\theta_3-\operatorname{e}^{i \rho } \sin\theta_1 \sin\theta_3\right)\,,\\
{\lambda_3'}^\alpha&=\lambda_1^\alpha \sin\theta_2 \sin\theta_3+\operatorname{e}^{i \phi } \lambda_2^\alpha \left(\cos\theta_1 \cos\theta_2 \sin\theta_3+\operatorname{e}^{i \rho } \sin\theta_1 \cos\theta_3\right)\,.
\end{aligned}
\end{equation}
This has a simple physical interpretation in terms of collinearly splitting $p_2$ into two daughters with momentum fractions $\cos^2\theta_1$ and $\sin^2\theta_1$,
followed by applying the rotation (\ref{rotation}) on two different pairs.
A nice feature is that the propagators in the first $y^3$ term in eq.~(\ref{eq: Yukawa theory scattering amplitudes}) become elementary trigonometric functions:
\begin{equation}
\label{eq: amplitude 3 to 2}
\langle 1_{{\psi}}2_{{\psi}}|\mathcal{M}^{2{\leftarrow}3}|1'_\phi2'_\phi3'_\phi\rangle\big|_{\text{first $y^3$ term}}
=-\frac{y^3}{\l12\rangle}(\operatorname{e}^{i \rho } \tan\theta_1 \cot\theta_2 \csc\theta_2 \cot\theta_3+\cot^2\theta_2+1)\,.
\end{equation}
In order to calculate the phase-space integral, we also require the measure factor, given as $(-1/\pi)$ times the phase-space volume.
It is given by
\begin{equation}
\label{eq: measure 3 to 2}
-\frac{s_{12}}{(4\pi)^4}\operatorname{d}\! \mu\quad \text{with} \quad\operatorname{d}\!\mu=2\sin\theta_1\cos\theta_1\operatorname{d}\!\theta_1\,4\sin^3\theta_2\cos\theta_2\operatorname{d}\!\theta_2\,2\sin\theta_3\cos\theta_3\operatorname{d}\!\theta_3 \frac{\operatorname{d}\!\rho}{2\pi}\frac{\operatorname{d}\!\phi}{2\pi}\,.
\end{equation}
In order to check the normalization of \eqref{eq: measure 3 to 2}, we compute
\begin{align}
\int\operatorname{d}\!\mu&=\int_0^{\frac{\pi}{2}}2\sin\theta_1\cos\theta_1\operatorname{d}\!\theta_1\int_0^{\frac{\pi}{2}}4\sin^3\theta_2\cos\theta_2\operatorname{d}\!\theta_2\int_0^{\frac{\pi}{2}}2\sin\theta_3\cos\theta_3\operatorname{d}\!\theta_3 \int_0^{2\pi}\frac{\operatorname{d}\!\rho}{2\pi}\int_0^{2\pi}\frac{\operatorname{d}\!\phi}{2\pi}\nonumber\\
&=1\,,
\label{eq: measure 3 to 2 integrated}
\end{align}
and we compare this to the discontinuity of the sunrise integral
\begin{equation}
\begin{aligned}
-\frac{1}{\pi}\FDinline[sunrise,twolabels,labelone=p_1,labeltwo=p_2,triplecut] &= -\frac{2}{\pi}\Im\left[\frac{1}{(4\pi)^{4-2\varepsilon}} \frac{\Gamma(1-\varepsilon)^3\Gamma(1+2\varepsilon)}{2\varepsilon (1-2\varepsilon)\Gamma(3-3\varepsilon)} (-s_{12}) \left(-\frac{s_{12}}{\mu^2} \right)^{-2\varepsilon}\right]
=-\frac{s_{12}}{(4\pi)^4} \,,
\end{aligned}
\end{equation}
finding perfect agreement.
Integrating the amplitude \eqref{eq: amplitude 3 to 2} against the measure \eqref{eq: measure 3 to 2} with an additional symmetry factor of $\tfrac{1}{3!}$ for the three-particle phase space, we thus find
\begin{equation}
-\frac{s_{12}}{(4\pi)^4}\frac{1}{3!}\int\operatorname{d}\! \mu\langle 1_{{\psi}}2_{{\psi}}|\mathcal{M}^{2{\leftarrow}3}|1'_\phi2'_\phi3'_\phi\rangle\big|_{\text{first $y^3$ term}}\l1_\phi'2_\phi'3_\phi'3_\phi|\mathcal{O}_\lambda|0\rangle=-\frac{2y^3}{3!(4\pi)^4}\langle 1_{{\psi}}2_{{\psi}}3_\phi| \mathcal{O}_y|0\rangle\,.
\end{equation}
The contributions from the five permutations are identical. Finally, the contribution from the term in the amplitude proportional to $\lambda y$ can be integrated trivially, as it does not depend on the phase-space parameters.
Adding all seven terms, we find
\begin{equation}
\gamma_{\lambda y}=-\frac{2y^3}{(4\pi)^4}+\frac{1}{6}\frac{y\lambda}{(4\pi)^4}\,.
\end{equation}
\subsection{Summary}
In total, we find
\begin{equation}
\left(\begin{array}{cc} \gamma_{yy}&\gamma_{y\lambda}\\ \gamma_{\lambda y}&\gamma_{\lambda\lambda}\end{array}\right)
=
\frac{1}{16\pi^2} \left(\begin{array}{cc} 12y^2+O(y^4)& -96y^3+8y\lambda+O(y^5) \\
-\frac{2y^3}{16\pi^2}+\frac{1}{6}\frac{y\lambda}{16\pi^2}+O(y^5)& 6\lambda+4y^2+O(y^4)\end{array}\right)\,, \label{eq: Yukawa anomalous dimension result}
\end{equation}
where for simplicity we quote the errors in the technically natural power counting $\lambda\sim y^2$.
All of these entries are one-loop except for the lower-diagonal one, $\gamma_{\lambda y}$, for which we have included the two-loop contribution which is leading.
Integrating the relation (\ref{eq: beta from gamma}) between anomalous dimensions and $\beta$-function, $\gamma_{ab}=\partial_a \beta(b)$,
in particular yields the one-loop $\beta$-functions
\begin{equation}
\beta^{(1)}(y) = \frac{1}{16\pi^2}\left( 4y^3\right),\qquad
\beta^{(1)}(\lambda) = \frac{1}{16\pi^2}\left( -24y^4 +4y^2\lambda + 3\lambda^2\right)\,,
\end{equation}
which is the standard textbook result for the considered theory of one Weyl fermion and one real scalar; see for example \cite{Srednicki:2007qs}, up to minor modifications to reflect our matter content.
The computed two-loop entry also yields some simple two-loop contributions
\begin{equation}
\beta^{(2)}(y) = \frac{1}{(16\pi^2)^2}\left(-2y^3\lambda+ \frac{1}{12}y\lambda^2+ \mbox{undetermined terms proportional to $y^5$}\right)\,,
\end{equation}
which can be compared for example with eq.\ (3.3) of \cite{Machacek:1983fi}, finding perfect agreement.
This demonstrates in a non-trivial way the correct handling of length-changing effects in the dilatation operator
by the proposed unitarity relation (\ref{cute_eigenvalue_equation}): $SF^*=\operatorname{e}^{-i\pi D}F^*$.
It is noteworthy that the relation between $\beta$-function and anomalous dimension is \emph{overconstrained}: the $4y^2\lambda$ term in $\beta(\lambda)$
is encoded in two different matrix elements of eq.~(\ref{eq: Yukawa anomalous dimension result}).
Since we obtained anomalous dimensions effectively as eigenvalues of the $S$-matrix, this must be viewed as a constraint satisfied by the $S$-matrix.
In fact, looking at the calculation, the $8y\lambda$ term in $\gamma_{y\lambda}$ is obtained from the $2{\to} 3$ amplitude in the last line of eq.~(\ref{eq: Yukawa theory scattering amplitudes}),
whereas the equivalent $4y^2$ term in $\gamma_{\lambda\lambda}$ comes from the collinear anomalous dimension (\ref{eq: collinear phi}), itself obtained from the $2{\to}2$ amplitudes acting on the stress tensor. It was not a-priori obvious why these $S$-matrix elements should be related, so it would be interesting to investigate such relations further.
It is interesting to see also that certain two-loop calculations in this section are actually simpler than one-loop calculations.
This is because, as presently formulated, calculating the anomalous dimension of high-length operators requires dealing with
a multi-scale problem (see eq.~(\ref{eq: cancellation of logarithms})) and so the number of legs has a strong impact on the complexity.
\section{Discussion and conclusion}
\label{sec: Summary}
In this paper, we have proposed a simple relation between the $S$-matrix of a theory at high energies and its dilatation operator:
\begin{equation}
\operatorname{e}^{-i\pi D} F^* = S F^*\,. \label{eq: main equation conclusion}
\end{equation}
In essence, this states that the time evolution from asymptotic past to future, as encoded by the $S$-matrix, is equivalent to following a
half-circle generated by a complex scale transformation as shown in fig.~\ref{fig: half-circle}.
This means that {\it the dilatation operator is minus the phase of the $S$-matrix, divided by the circumference of the half-circle} ($\pi$).
At one-loop in Yang-Mills theory, this provides a surprisingly efficient way to calculate the $\beta$-function of the theory.
Starting with the famous Parke-Taylor tree-level amplitude for scattering four on-shell gluons in eq.~(\ref{eq: Parke-Taylor formula}),
and performing elementary operations such as integrating over the two-body phase space, one reproduces the
famous result proportional to $-11C_A/3$ in section \ref{sec: YM eleven thirds}. In particular, as usual with on-shell methods, only physical on-shell gluon states enter the calculation.
Furthermore, the sign is directly tied to the positive sign of the amplitude, itself stemming from the attractive force between opposite charges.
The extension to QCD including masses is discussed in subsection \ref{ssec: matter} and poses no significant problem.
We also found a correspondence between twist-two anomalous dimensions and the phase of $2{\to}2$ angular-momentum partial waves,
and obtained a novel formula, eq.~(\ref{eq: gamma general length}), for the one-loop dilatation operator in any gauge theory.
A pleasant feature is that the one-loop anomalous dimensions of all operators
are generated by the same building blocks, the $2{\to}2$ tree amplitudes of the theory.
Of course, in QCD the technology to calculate twist-two anomalous dimensions and $\beta$-functions is already very
well developed: for example, three-loop anomalous dimensions have been known for some time \cite{Vogt:2004mw,Moch:2004pa}
and the four- and even five-loop $\beta$-function are now known \cite{vanRitbergen:1997va,Czakon:2004bu,Baikov:2016tgj}.
On the other hand, for more general operators such as dimension-six operators in the Standard Model effective theory,
one-loop results have only been obtained recently \cite{Alonso:2013hga}.
An advantage of the present method is that it treats all operators of the theory on the same footing, which could help automation in this context.
In addition, certain qualitative features such as helicity selection rules are automatically manifest \cite{Cheung:2015aba}.
As mentioned in the main text, the present method could also be advantageous at higher loops in the context of theories with extended symmetries,
since the symmetries of the $S$-matrix are naturally maintained (including integrability in planar $\mathcal{N}=4$ SYM).
We have also investigated Yukawa theory at one-loop and beyond, confirming the general validity of the approach.
This is a phenomenologically important theory which
allows us to study effects which generically will be present at higher loops in any theory.
Of course, since on-shell amplitudes in this case are not simpler than the corresponding Feynman diagrams,
we did not expect a significant advantage to using this method.
The main new effect is that, while the $\beta$-function of Yang-Mills could be determined using an operator which decays
to two partons at tree level (the gluon density ${\rm Tr}\,[G^2]$), measuring the couplings in Yukawa theory requires more external legs interacting together, which makes the problem multi-scale and causes individual cuts to be more complicated.
The simplest case where this occurs is the length-increasing mixing at one-loop studied in section \ref{ssec: increasing}; here, logarithms cancel non-trivially between cuts (see eq.~\ref{eq: cancellation of logarithms}).
We expect such effects to be generic for higher-twist operators in any theory beyond one-loop,
and a formalism where such canceling transcendental functions could be discarded in individual cuts would greatly simplify calculations.
We note however that for twist-two operators and $\beta$-functions in QCD, the problem is always single-scale and such difficulties are absent.
In this work, we have taken the renormalization group equation as an input, but it is interesting to ask if it could be derived in an on-shell framework using physical
principles like unitarity of the $S$-matrix. For example, there might be a recursive way to construct the scale dependence of the amplitudes and form factors on each side of a cut.
In general, the formalism exposes interesting relationships between form factors and the $S$-matrix, and it would be fascinating to study this interplay
in explicit examples at two loops and higher.
\section*{Acknowledgments}
We thank David McGady, Mike Trott and Florian Loebbert for discussion.
M.W.\ thanks Florian Loebbert, Christoph Sieg and Gang Yang for collaboration on a related project.
M.W.\ was supported in part by DFF-FNU through grant number DFF-4002-00037.
S.C.H.'s research was partly funded by the Danish National Research Foundation (DNRF91).
Both authors acknowledge the kind hospitality of NORDITA during the program ``Aspects of Amplitudes,'' where parts of this work were carried out.
\bibliographystyle{JHEP}
|
\section{Introduction}
Throughout, we will denote by $\mathcal{B}(\mathcal{H})$ the complex Banach algebra of all bounded linear operators on a complex Hilbert space $\mathcal{H}$.
Recall that an operator $T\in \mathcal{B}(\mathcal{H})$ is said to be :
\begin{itemize}
\item positive if $ <Tx,x>\geq 0$ for all $x\in \mathcal{H}$
\item self-adjoint if $T=T^*$
\item isometry if $T^*T=I$, which equivalent to the condition $\|Tx\|=\|x\|$ for all $x\in \mathcal{H}$
\item normal if $T^*T=TT^*$
\item unitary $T^*T=TT^*=I$ (i.e $T$ is an onto isometry)
\item quasinormal if $T[T^*T]=[T^*T]T$
\item binormal if $[T^*T][TT^*]=[TT^*][T^*T]$ [\ref{3}]
\item subnormal if $T$ has a normal extension
\item hyponormal if $T^*T\geq TT^*$, which equivalent to the condition $\|T^*x\|\leq\|Tx\|$ for all $x\in \mathcal{H}$ [\ref{12}]
\item $M$-hyponormal if $T^*T\geq MTT^*$, where $1\leq M$ which equivalent to the condition $\|T^*x\|\leq M \|Tx\|$ for all $x\in \mathcal{H}$ [\ref{16}]
\item $p$-hyponormal if $(T^*T)^p\leq (TT^*)^p$, where $0<p\leq 1$ [\ref{1}]
\item class $A$ if $|T|^2\leq |T^2|$, where $|T|=(T^*T)^{\frac{1}{2}}$
\item paranormal if $\|Tx\|^2\leq\|T^2x\|\|x\|$ for all $x\in \mathcal{H}$ [\ref{17}]
\item $k$-hyponormal if $\|Tx\|^k\leq \|T^kx\|\|x\|^{k-1}$ for all $x\in \mathcal{H}$, where $ k\geq 2$
\item $*$-paranormal if $\|T^*x\|^2\leq\|T^2x\|\|x\|$ for all $x\in \mathcal{H}$ [\ref{18}]
\item quasi $*$-paranormal if $\|T^*Tx\|^2\leq \|T^3x\|\|Tx\|$ for all $x\in \mathcal{H}$ [\ref{19}]
\item $\log$-hyponormal if $T$ invertible and satisfies $\log(T^*T)\geq \log(TT^*)$ [\ref{13}]
\item $p$-quasihyponormal if $T^*[(T^*T)^p-(TT^*)^p]T\geq 0$ , where $0<p\leq 1$ [\ref{2}]
\item normoloid if $\|T\| =r(T)$
\item quasinilpotent if $r(T)= 0$ , where $r(T)=\lim\|T^n\|^{\frac{1}{n}}$
\end{itemize}
We can notice that $T$ is hyponormal if $T$ is $p$-hyponormal with $p=1$ and that $p$-hyponormal is $q$-hyponormal for every $0<q\leq p$ from L\"{o}wner-Heinz inquality [\ref{10}].
Also we can notice that $T$ is paranormal if $T$ is $k$-hyponormal with $k=2$.\\
It known that invertible $p$-hyponormal is $\log$-hyponormal, we may regard $\log$-hyponormal operator as $0$-hyponormal [\ref{13}].\\
It is well known that for any operators $A,B$ and $C$, $ A^*A-2\lambda B^*B+\lambda^{2}C^{*}C\geq 0$ for all $\lambda>0$, if and only if $||Bx||^{2}\leq ||Ax||||Cx||\,\mbox{for all}\,x\in H.$
Thus we have\\
\begin{itemize}
\item $T$ is quasi $*$-paranormal if and only if $T^*[(T^*)^2T^2-2\lambda TT^*+\lambda^2]T\geq 0$, for all $\lambda>0.$
\item $T$ is $*$-paranormal if and only if $(T^*)^2T^2-2\lambda TT^*+\lambda^2\geq 0$ , for all $\lambda>0$.\\
\end{itemize}
Thus every $*$-paranormal is quasi $*$-paranormal and we have the following implications:\\
quasinormal $\Longrightarrow$ binormal \\
self-adjoint $\Longrightarrow$ normal $\Longrightarrow$ quasinormal $\Longrightarrow$ subnormal $\Longrightarrow$ hyponormal $\Longrightarrow$ $*$-paranormal $\Longrightarrow$ quasi $*$-paranormal\\
hyponormal $\Longrightarrow$ $p$-hyponormal $\Longrightarrow$ $p$-quasihyponormal $\Longrightarrow$ class $A$ $\Longrightarrow$ paranormal\\
invertible $p$-hyponormal $\Longrightarrow$ $\log$-hponormal $\Longrightarrow$ paranormal.\\
For a scalar $\lambda$, two operators $A$ and $B$ in $\mathcal{B}(\mathcal{H})$ are said be $\lambda$-commute if $AB=\lambda BA$. Recently many authors have studied this equation for several classes of operators, for example\\
\begin{itemize}
\item In [\ref{9}] authors have proved that if an operator in $\mathcal{B}(\mathcal{H})$ $\lambda$-commutes with a
compact operator, then this operator has a non-trivial hyperinvariant subspace
\item In [\ref{7}] Conway and Prajitura characterized the closure and the interior of the set of operators that $\lambda$-commute with
a compact operator
\item In [\ref{15}] Zhang ,Ohawada and Cho have studied the properties of operators that $\lambda$-commute with
a paranormal operator
\item In [\ref{4}] Brooke, Busch and Pearson showed that if $AB$ is not quasinilpotent, then $|\lambda| =1$,
and if $A$ or $B$ is self-adjoint then $\lambda\in\mathbb{R}$
\item In [\ref{14}] Yang and Du gave a simple proofs and generalizations of this results, particulary if $AB$ is bounded below if and only if both $A$ and $B$ are bounded below
\item In [\ref{11}] Schmoeger generalized this results
to hermitian or normal elements of a complex Banach algebra
\item In [\ref{5}] Cho, Duggal, Harte and Ota generalized some Schmoeger's results and they
gave the new characterization of a commutativity of Banach space operators.\\
\end{itemize}
The aim of this paper is to study the situation for binormal, $M$-hyponormal, quasi $*$-paranormal operators. Again other related results are also given.\\
We denote the range and the kernel of $T$ by $R(T)$ and $N(T)$ respectively.
\section{ Main results }
We begin with the following result.
\begin{lem}[\ref{19}]\label{lem1}
Let $T \in \mathcal{B}(\mathcal{H})$ be quasi $*$-paranormal and $M$ be a closed $T$-invariant subspace of $\mathcal{H}$. Then $T_{|M}$ is also quasi $*$-paranormal.
\end{lem}
\begin{lem}
Let $A\in \mathcal{B}(\mathcal{H})$ be quasi $*$-paranormal operator. If $A$ is quasinilpotent, then $A=0$
\end{lem}
\begin{pr}
Suppose that $A\neq0$, we have for any operator $ T\in \mathcal{B}(\mathcal{H})$,
$\|T^*T\|=\|TT^*\|=\|T\|^2=\|T^*\|^2$.
Firstly, by induction we prove that $A$ is normaloid.\\
For $n=3$,
since $A\in \mathcal{B}(\mathcal{H})$ is quasi $*$-paranormal, then
$\|A\|^4=[\|A\|^2]^2=\|A^*A\|^2\leq\|A^3\|\|A\|$, whence
$\|A\|^3\leq\|A^3\|\leq \|A\|^3$. From where $\|A^3\|= \|A\|^3$.\\
It is now assumed for an integer $n\geq 3$ and any integer k such that $3\leq k\leq n$,
we have $\|A^k\|= \|A\|^k$.
we know that $\|A^{n+1}\|\leq\|A\|^{n+1}$, other hand We have
\begin{eqnarray*}
[\|A\|^{n-1}]^4 & =& [\|A^{n-1}\|^2]^2\\
& = & [\sup_{\|x\|=1}<A^{n-1}x,A^{n-1}x>]^2\\
& = & [\sup_{\|x\|=1}<A^*A[A^{n-2}x],A^{n-2}x>]^2 \\
& \leq & \sup_{\|x\|=1} [\|A^*A[A^{n-2}x]\|]^2[\|A^{n-2}x\|]^2\\
& \leq & \sup_{\|x\|=1} [\|A^3[A^{n-2}x]\|\|A[A^{n-2}x]\|[\|A^{n-2}x\|]^2]\\
& \leq& \sup_{\|x\|=1} [\|A^{n+1}x]\|\|A^{n-1}x\|[\|A^{n-2}x\|]^2]\\
& \leq & \sup_{\|x\|=1} [\|A^{n+1}x]\|\sup_{\|x\|=1}\|A^{n-1}x\|\sup_{\|x\|=1}\|A^{n-2}x\|^2]\\
& \leq & \|A^{n+1}\|\|A^{n-1}\|\|A^{n-2}\|^2\\
& \leq & \|A^{n+1}\|\|A\|^{n-1}\|A\|^{2(n-2)}
\end{eqnarray*}
whence $\|A^{n+1}\| \geq \|A\|^{4(n-1)-(n-1)-2(n-2)}=\|A\|^{n+1}$,
thus $\|A^{n+1}\|=\|A\|^{n+1}$. Since $A$ is normaloid we have $r(A)=\|A\|$.\\
But $A$ is quasinilpotent, then
$r(A)=0$, therefore $\|A\|=0$ and this is contradiction.
\end{pr}
\begin{coro}
Let $A\in \mathcal{B}(\mathcal{H})$ be $*$-paranormal operator. If $A$ is quasinilpotent, then $A=0$
\end{coro}
\begin{thm}
Let $A,B\in \mathcal{B}(\mathcal{H})$ satisfy $AB=\lambda BA\neq 0,\> \lambda\in\mathbb{C}$. Let $A$ be hyponormal and $B$ be quasi $*$-paranormal. If $A$ is invertible or 0 is an isolated point of $\sigma(A)$, then $|\lambda|=1$.
\end{thm}
\begin{pr}
It is clear that if $A$ is invertible then $|\lambda|=1$.\\
Suppose that 0 is isolated point of $\sigma(A)$.
Since $\mathcal{H}=\overline{R(A^*)} \oplus N(A)$ we can decompose $A=\left(
\begin{array}{cc}
A_1 & 0 \\
A_2 & 0 \\
\end{array}
\right)$
on $\overline{R(A^*)} \oplus N(A)$.\\
Since $A$ is hyponormal
$$\left(\begin{array}{cc}
A_1^*A_ 1+A_2^*A_2& 0 \\
0 & 0 \\
\end{array}
\right)=A^*A\geq
AA^*=\left(
\begin{array}{cc}
A_1A_1^* & A_1A_2^* \\
A_2A_1^* & A_2A_2^* \\
\end{array}
\right)$$
it holds $ A_2A_2^*\leq 0$
and hence $A_2 =0$, we have $$A=\left(
\begin{array}{cc}
A_1 & 0 \\
0 & 0 \\
\end{array}
\right),
$$
Consequently $\sigma(A)=\sigma(A_1)\cup\{0\}$.\\
We conclude that if $A_1$ is not invertible then 0 is isolated point of $\sigma(A_1)$ and since $A_1$ is hyponormal, then $0\in\sigma_p(A_1)$, a contradiction and hence $A_1$ is invertible.\\
Let $B=\left(
\begin{array}{cc}
B_1 & B_2 \\
B_3 & B_4 \\
\end{array}
\right)$
on $\overline{R(A^*)} \oplus N(A)$.
By $\lambda\neq 0$ and $AB=\lambda BA$, it holds $A_1B_2=B_3A_1=0$. Since $A_1$ is invertible, we have $B_2=B_3=0$. Hence, $$B=\left(
\begin{array}{cc}
B_1 & 0 \\
0 & B_4 \\
\end{array}
\right).$$
Therefore $\overline{R(A^*)}$ is invariant for $B$ and hence $B_1=B_{|\overline{R(A^*)}}$ is quasi $*$-paranormal by lemma$\>$\ref{lem1}.
Since
\begin{eqnarray*}
AB &=& \left(\begin{array}{cc} A_1 & 0 \\ 0 & 0 \\
\end{array}
\right)\left(\begin{array}{cc}
B_1 & 0 \\
0 & B_4 \\
\end{array}
\right)\\
&=&
\left(\begin{array}{cc}
A_1B_1 & 0 \\
0 & 0\\
\end{array}
\right)\\
&=&
\lambda BA\\
&= & \lambda \left(\begin{array}{cc}
B_1 & 0 \\
0 & B_4 \\
\end{array}
\right)\left(
\begin{array}{cc}
A_1 & 0 \\
0 & 0 \\
\end{array}
\right)\\
&= & \lambda \left(\begin{array}{cc}
B_1A_1 & 0 \\
0 & 0 \\
\end{array}
\right).
\end{eqnarray*}
Whence $A_1B_1=\lambda B_1A_1$, then $B_1=\lambda A_1^{-1}B_1A_1$.\\
If $r(B_1)= 0$ then $B_1$ is quasi $*$-paranormal and quasinilpotent, then $B_1=0$ by lemma 2.2, therefore AB=0 and it's a contradiction because $AB\neq 0$. From where $r(B_1)\neq 0$ and consequently $|\lambda|=1$.
\end{pr}
\begin{coro}
Let $A,B\in \mathcal{B}(\mathcal{H})$ satisfy $AB=\lambda BA\neq 0,\> \lambda\in\mathbb{C}$. Let $A$ be hyponormal and $B$ be $*$-paranormal operator. If $A$ is invertible or 0 is an isolated point of $\sigma(A)$, then $|\lambda|=1$.
\end{coro}
\begin{thm}
Let $A$ be quasinormal operator and $B$ normal such that $AB=\lambda BA\neq 0,\;\; \lambda\in\mathbb{C}$.
If $|\lambda|=1$ then $AB$ is quasinormal operator.
\end{thm}
\begin{pr}
Assume that $AB=\lambda BA\neq 0$, then $B^*A^*=\bar{\lambda}A^*B^*$. Since $B$ and $\lambda B$ are normal operators by Fuglede-Putnam theorem, condition $AB=\lambda BA$ imply that $BA^*=\lambda A^*B$ and $AB^*=\bar{\lambda}B^*A$. Now have
\begin{eqnarray*} AB[(AB)^*AB] &=&[AB][B^*A^*AB]\\
&=&[\lambda BA]B^*A^*AB\\
&=&\lambda B[AB^*]A^*AB\\
&=&\lambda B[\bar{\lambda}B^*A]A^*AB\\
&=&|\lambda|^2 [BB^*][AA^*A]B\\
&=& [B^*B][A^*AA]B\\
&=& B^*[BA^*]AAB\\
&=& B^*[\lambda A^*B]AAB\\
&=& B^*A^*[\lambda BA]AB\\
&=& B^*A^*[AB]AB\\
&=& [(AB)^*AB] AB
\end{eqnarray*}
Therefore $AB$ is quasinormal.\\
\end{pr}
\begin{thm}
Let $A$ be binormal operator and $B$ normal such that $AB=\lambda BA\neq 0, \;\; \lambda\in\mathbb{C}$
if $|\lambda|=1$ then $AB$ is binormal operator.
\end{thm}
\begin{pr}
Because $B$ and $\lambda B$ are normal operators by Fuglede-Putnam theorem, condition $AB=\lambda BA$ imply that
$BA^*=\lambda A^*B$ and $AB^*=\bar{\lambda}B^*A$. Therfore
\begin{eqnarray*}AB(AB)^*(AB)^*AB &=& A[BB^*]A^*B^*A^*AB\\
&=& A[B^*B]A^*B^*A^*AB\\
&=& [AB^*]BA^*[B^*A^*]AB\\
&=& [\bar{\lambda}B^*A]BA^*[\bar{\lambda}A^*B^*]AB\\
&=& (\bar{\lambda})^2B^*[AB]A^*A^*[B^*A]B\\
&=& (\bar{\lambda})^2B^*[\lambda BA]A^*A^*[\frac{1}{\bar{\lambda}}AB^*]B\\
&=& |\lambda|^2B^*B[AA^*A^*A]B^*B\\
&=& B^*B[A^*AAA^*]B^*B\\
&=& B^*[BA^*]AA[A^*B^*]B\\
&=& B^*[\lambda A^*B]AA[\frac{1}{\bar{\lambda}}B^*A^*]B\\
&=& \frac{\lambda}{\bar{\lambda}} B^*A^*[BA]AB^*[A^*B]\\
&=& \lambda^2 B^*A^*[\frac{1}{\lambda} AB]AB^*[\frac{1}{\lambda}BA^*]\\
&=& B^*A^*ABA[B^*B]A^*\\
&=& B^*A^*ABA[BB^*]A^*\\
&=& (AB)^*AB AB (AB)^*
\end{eqnarray*}
We obtain that $AB$ is binormal
\end{pr}
\begin{thm}
Let $A$ be $k$-hyponormal and $B$ isometry such that $AB=\lambda BA\neq 0,\;\; \lambda\in\mathbb{C}$. The following statements are equivalent
\begin{enumerate}
\item $AB$ is $k$-hyponormal
\item $\sigma(AB)\neq \{0\}$
\item $|\lambda|=1$
\end{enumerate}
\end{thm}
\begin{pr}
It is clear $(1)\Rightarrow (2)\Rightarrow (3)$. So we show $(3)\Rightarrow (1)$.
For any unit vector $x\in H$ scince $A$ is $k$-hyponormal and $B$ is isometry then
\begin{eqnarray*}
\|(AB)x\|^k &=& \|A(Bx)\|^k \\
&\leq& \|A^k(Bx)\|\|Bx\|^{k-1} \\
&\leq& \|A^k(Bx)\|
\end{eqnarray*}
On the other hand
By induction we show that $(AB)^k=\lambda^{\frac{k(k-1)}{2}} B^{k-1}A^kB$ for every $k\in\mathbb{N}^*$.\\
For $k=1$, $(AB)^1= \lambda^{\frac{1(1-1)}{2}} B^{1-1}A^1B$.
Assume that it holds for $k\geq2$, we obtain
\begin{eqnarray*}
(AB)^{k+1}=AB(AB)^k&=&(\lambda BA)(\lambda^{\frac{k(k-1)}{2}} B^{k-1}A^kB)\\
&=& \lambda^{\frac{k(k-1)}{2}+1}BA B^{k-1}A^kB\\
&=& \lambda^{\frac{k(k-1)}{2}+1}B(AB)B^{k-2}A^kB\\
&=& \lambda^{\frac{k(k-1)}{2}+1}B(\lambda BA)B^{k-2}A^kB\\
&=& \lambda^{\frac{k(k-1)}{2}+2}B^2 AB^{k-2}A^kB\\
&:&\\
&:&\\
&:&\\
&=& \lambda^{\frac{k(k-1)}{2}+k}B^k AB^{k-k}A^kB\\
&=& \lambda^{\frac{(k+1)k}{2}}B^k A^{k+1}B
\end{eqnarray*}
We conclude that $(AB)^k=\lambda^{\frac{k(k-1)}{2}} B^{k-1}A^kB$ for every $k\in\mathbb{N}^*$.\\
Because $B$ is isometry and $|\lambda|=1$ it follows that
\begin{eqnarray*}
\|(AB)^k\| &=& \|\lambda^{\frac{k(k-1)}{2}} B^{k-1}A^kBx\| \\
&=& |\lambda|^{\frac{k(k-1)}{2}}\|B^{k-1}A^kBx\| \\
&=& \|A^kBx\|
\end{eqnarray*}
Therefore it holds $\|(AB)x\|^k \leq \|(AB)^kx\|$ for any unit vector $x$ and $AB$ is \\ $k$-hyponormal, which completes the proof.
\end{pr}
\begin{thm}
Let $A,B\in \mathcal{B}(\mathcal{H})$ such that $AB=\lambda BA\neq 0, \;\;\lambda\in\mathbb{C}$. Then
\begin{enumerate}
\item if $A^*$ is $M_1$-hyponormal and $B$ is $M_2$-hyponormal, then $|\lambda|\leq (M_1M_2)^\frac{1}{2}$
\item if $A$ is $M_1$-hyponormal and $B^*$ is $M_2$-hyponormal, then $|\lambda|\geq (M_1M_2)^{-\frac{1}{2}}$
\end{enumerate}
\end{thm}
\begin{pr}
\begin{enumerate}
\item We have
\begin{eqnarray*}
|\lambda|\|BA\|&=&\|\lambda BA\|\\
&=& \|AB\|\\
&=& \|B^*A^*AB\|^{\frac{1}{2}} \>\>\>\>(\|T\|=\|TT^*\|^{\frac{1}{2}}\\
&\leq & M_1^{\frac{1}{2}}\|B^*AA^*B\|^{\frac{1}{2}} \>\> (A^* \>\textit{is}\> M_1-hyponormal:\> A^*A\leq M_1AA^*)\\
&\leq& M_1^{\frac{1}{2}} \|A^*B\|\>\>(\|T^*T\|^{\frac{1}{2}}=\|T\|)\\
&\leq& M_1^{\frac{1}{2}} \|A^*BB^*A\|^{\frac{1}{2}}\>\>\>\>(\|T\|=\|TT^*\|^{\frac{1}{2}}) \\
&\leq & (M_1M_2)^{\frac{1}{2}} \|A^*B^*BA\|^{\frac{1}{2}} \>\> (B \>\textit{is}\> M_2-hyponormal:\>BB^*\leq M_2B^*B) \\
&\leq& (M_1M_2)^{\frac{1}{2}} \|BA\|\>\>\>\>(\|T^*T\|^{\frac{1}{2}}=\|T\|).
\end{eqnarray*}
Hence $|\lambda|\|BA\| \leq (M_1M_2)^{\frac{1}{2}} \|BA\|$ and $|\lambda|\leq (M_1M_2)^{\frac{1}{2}}$\\
\item Since $AB=\lambda BA$ and $\lambda\neq 0$ we have $BA=\lambda^{-1}AB$ and apply (1) we obtain $|\lambda^{-1}|\leq (M_2M_1)^{\frac{1}{2}}$,
therefore $|\lambda|\geq (M_2M_1)^{-\frac{1}{2}}$.\\
\end{enumerate}
\end{pr}
\begin{coro}
Let $A,B\in \mathcal{B}(\mathcal{H})$ such that $AB=\lambda BA\neq 0, \;\;\lambda\in\mathbb{C}$. Then
\begin{enumerate}
\item if $A^*$ and $B$ are hyponormal, then $|\lambda|\leq 1$
\item if $A$ and $B^*$ are hyponormal, then $|\lambda|\geq 1$\\
\end{enumerate}
\end{coro}
\begin{thm}
Let $A,B\in \mathcal{B}(\mathcal{H})$ such that $AB=\lambda BA\neq 0,\;\; \lambda\in\mathbb{C}$. Then
if $A^*$ is $M_1$-hyponormal and $B$ is $M_2$-hyponormal then $A^*B$ and $BA^*$ are $M_1M_2|\lambda|^2$-hyponormal
\end{thm}
\begin{pr}
We have
\begin{eqnarray*}
(A^*B)^*A^*B &=& B^*AA^*B \\
&\geq & M_1B^*A^*AB \\
&\geq & M_1 \bar{\lambda}A^*B^*\lambda BA \\
&\geq & M_1 |\lambda|^2 A^*B^*BA\\
&\geq & M_1|\lambda|^2A^*M_2BB^* A\\
&\geq& M_1M_2|\lambda|^2(B^* A)^*B^* A
\end{eqnarray*}
Therefore $A^* B$ is $M_1M_2|\lambda|^2$-hyponormal.\\
Same way we prove that $BA^* $ is $M_1M_2|\lambda|^2$-hyponormal.\\
\end{pr}
|
\subsection{Ramsey's technique of separated oscillating fields}
More than half a century ago Ramsey improved Rabi's resonant frequency technique to measure energy eigenstates of quantum mechanical systems by introducing a free-precession period between two spin-flipping pulses\,\cite{Ramsey1950PR}.
Figure\,\ref{fig:RamseyMethod}a) illustrates this technique, while a typical resonance scan is shown in Fig.\,\ref{fig:RamseyMethod}b). The initial state is a fully polarized, i.e.\ \ensuremath{\Ket{\uparrow}}, ensemble of neutrons exposed to a magnetic field $B_0$. A first rotational oscillating magnetic-field pulse $B_{\rm rf} = B_1\cos\left(\omega_\text{rf}t\right)$, perpendicular to $B_0$, tips the spins into the plane orthogonal to the main magnetic field. The neutron spins then precess freely with their Larmor frequency $\omega_0$ for a duration $T$, accumulating a phase $\phi = \gamma_n B T$, before a second pulse $B_1\cos\left(\omega_\text{rf}t\right)$ in phase with the first is again applied to the neutron ensemble. The essential idea is to compare the phase $\phi$ with $\omega_{\rm rf}T$. If they are identical then $B=\omega_{\rm rf}/\gamma_n$.
The probability to detect a neutron with a final spin state identical to its initial spin state, i.e.\ \ensuremath{\Ket{\downarrow}}, is (see equation~(A.11) in Ref.\,\cite{Piegsa2009NIMA}):
\begin{align}
\mathcal{P}(T,\omega_\text{rf})&=\left|\Bra{\downarrow}U(T,\omega_\text{rf})\Ket{\downarrow}\right|^2 \nonumber \\
&=1-\frac{4\omega_1^2}{\Omega^2}\sin^2\left(\frac{\Omega t_{\pi/2}}{2}\right)
\left[\frac{\Delta}{\Omega}\sin\left(\frac{\Omega t_{\pi/2}}{2}\right)\sin\left(\frac{T\Delta}{2}\right)-\right.
\left.\cos\left(\frac{\Omega t_{\pi/2}}{2}\right)\cos\left(\frac{T\Delta}{2}\right)\right]^2,
\label{eq:RamseyResonance}
\end{align}
\noindent where $U(T,\omega_\text{rf})$ is the time-evolution operator describing the pulse sequence, $\omega_1\!=\!-\gamma_\text{n}B_1$, $\Delta \!=\!\omega_\text{rf}-\omega_0$, and $\Omega\!=\!\sqrt{\Delta^2+\omega_1^2}$. When optimized the spin-flipping pulses have exactly enough power to tip the spins by $\pi/2$, hence, the pulse length and field power fulfill the condition $\gamma_\text{n}B_1t_{\pi/2} =\pi/2$. In this case, and in the central fringe range ($\Delta \ll \omega_1$), equation\,(\ref{eq:RamseyResonance}) simplifies to:
\begin{eqnarray}
\mathcal{P}(T,\omega_\text{rf}) &=&1-4\sin^4\frac{\pi}{4}
\left[\frac{\Delta}{\Omega}\sin\frac{T\Delta}{2}-\cos\frac{T\Delta}{2}\right]^2 \nonumber \\
\mathcal{P}(T,\omega_\text{rf}) &\approx& 1- \cos^2\frac{T\Delta}{2} \nonumber \\
\mathcal{P}(T,\omega_\text{rf}) &=& \frac{1}{2}\left(1 -\cos\left(T\Delta\right)\right).
\label{eq:CosineApproximation}
\end{eqnarray}
\noindent In a real measurement with $N$ neutrons inside a large magnetic field region this becomes
\begin{equation}
N^{\downarrow/\uparrow} = \frac{N}{2}\left\{1 \mp \alpha(T)\cos\left[ \left(\omega_\text{rf} - \gamma_\text{n}B_0\right)\cdot\left(T+\frac{4t_{\pi/2}}{\pi}\right)\right]\right\},
\label{eq:UCNRamseyFormula}
\end{equation}
\noindent where $\alpha(T)$ is the visibility of the central fringe taking into account all depolarization effects\,\cite{Afach2015PRD} and $N^{\downarrow/\uparrow}$ are the neutrons with spin down (high-field seeking) and spin up (low-field seeking). The term $\frac{4t_{\pi/2}}{\pi}$ is necessary to account for field inhomogeneities of $B_1$ and $B_0$ which become relevant when the pulse length $t_{\pi/2}$ is finite\,\cite{Slichter1990book}.
\begin{figure}%
\centering
\subfloat{
\includegraphics[width=0.41\columnwidth]{./art/RamseysMethod.png}}
\subfloat{\includegraphics[width=0.40\columnwidth]{./art/RamseyPlot.png}}
\caption{Ramsey's technique of separated oscillating fields. Note, that neutrons with spin down are high-field seeking neutrons indicated by the yellow arrow as the gyro magnetic ratio is negative. The scheme of the method (left), based on Fig.\,4 of Ref.\,\cite{Harris2007arXiv}, and the data plot (right) are explained in detail in the text. Blue points are UCN counted with spin up $N^{\uparrow}$, while red points are with spin down $N^{\downarrow}$ (data from the PSI-nEDM collaboration\,\cite{Baker2011}). The width at half height $\Delta \nu$ of the central fringe is approximately $1/\left(2T\right)$, and the four vertical lines indicate the working points. }%
\label{fig:RamseyMethod}%
\end{figure}
A Ramsey interference pattern, as shown in Fig.\,\ref{fig:RamseyMethod}b), can be recorded by scanning $\omega_\text{rf}$ while keeping all other conditions constant. The magnetic field $B$ measured by the neutrons then fulfills the resonance condition of the central fringe with $\omega_\text{rf} = \gamma_\text{n}B$.
This procedure is further optimized in searches for an nEDM by only measuring at four points close to highest sensitivity, the working points.
The neutron Larmor frequency is then obtained by fitting expression\,(\ref{eq:UCNRamseyFormula}) to the data.
Separate fits are performed for different electric-field and magnetic-field configurations.
Taking the difference of these Larmor frequencies gives access to the electric dipole moment:
\begin{equation}
\ensuremath{d_\text{n}} = \frac{\hbar (\omega_0^{\parallel}-\omega_0^{\nparallel})}{2(E^{\parallel} - E^{\nparallel})} = \frac{\hbar \Delta\omega}{4E},
\label{eq:nEDM-simple}
\end{equation}
\noindent using equation\,(\ref{eq:DiffConfig}) and assuming no differences in the magnetic field and $E=E^{\parallel} = -E^{\nparallel}$. The statistical sensitivity with which a frequency can be measured in a time $t$ can be deduced from equation\,(\ref{eq:UCNRamseyFormula}):
\begin{equation}
\sigma\left(\omega_0\right) \approx \frac{1}{\alpha T \sqrt{\dot{N}t}},
\label{eq:SensitivityFrequency}
\end{equation}
\noindent where $\dot{N} = \langle N \rangle \cdot f$ is the neutron current with $ \langle N \rangle$ the average number of neutrons detected in one measurement and $f$ the repetition rate. This translates into a statistical sensitivity for the detection of an nEDM of
\begin{equation}
\sigma\left(\ensuremath{d_\text{n}}\right) \approx \frac{\hbar\sqrt{2}}{4\alpha T E \sqrt{\dot{N}t/2}}
= \frac{\hbar}{2\alpha T E \sqrt{\dot{N}t }},
\label{eq:SensitivityEDM}
\end{equation}
\noindent using equation\,(\ref{eq:nEDM-simple}). The additional factor $1/\!\sqrt{2}$ in the denominator accounts for the fact that it needs two measurements to take a difference.
\subsection{Ultracold neutrons}
\subsubsection{Properties of UCN}
From a neutron current point of view it becomes immediately clear that $T\!\sqrt{\dot{N}}$ needs to be maximized for the highest sensitivity. One possible solution, proposed in 1960 by Shapiro\,\cite{Lushchikov1969JETPL,Shapiro1970SPU}, is to increase $T$ by using neutrons that can be stored within vacuum chambers made of adequate materials. Such ultracold neutrons (UCN) are reflected under any incident angle by the neutron optical potential
\begin{equation}
V = \frac{2\pi\hbar^2}{m_\text{n}}\mathcal{N}b,
\label{eq:NeutronOptPot}
\end{equation}
\noindent where $m_\text{n}$ is the neutron mass, $\mathcal{N}$ the nucleon density and $b$ the neutron scattering length of a given material. Well suited materials for storage are, e.g., $^{58}$Ni, diamond-like carbon (DLC), NiMo(85/15), or BeO/Be which all have neutron optical potentials in the range \SIrange{200}{350}{neV}.
These potentials also define the maximum kinetic energy of UCN\@, which is in the same range as the gravitational potential energy of neutrons: $mgh\approx h
\cdot\unit[1.025]{neV/cm}$. Furthermore, strong magnetic fields can be used to polarize or store UCN, as the magnetic potential energy $\boldsymbol{\mu}\cdot\boldsymbol{B} \approx \pm \boldsymbol{B}\cdot\unit[60]{neV/T}$. This means, a magnetic field of \SI{5}{T} creates a potential barrier of \SI{300}{neV} for low-field seekers while high-field seekers are attracted.
For the search of an nEDM with UCN the ultimate goal is to confine UCN as long as possible within a closed material cell. Several imperfections of the material cell will effectively reduce the number of UCN of energy $E$ detected after a time $t$ to $N_0(E)\exp\left(-t/\tau_{\rm s}(E)\right)$. The storage time constant $\tau_{\rm s}(E)$ is the inverse of the sum of all loss rates:
\begin{equation}
\frac{1}{\tau_{\rm s}(E)} = \Gamma(E) = \Gamma_{\beta}+\Gamma_{\eta}(E)+\Gamma_{\rm gap}(E) + \Gamma_{\rm gas}(E).
\label{eq:UCN_lossrates}
\end{equation}
\noindent A free neutron is not stable and decays with a rate of $\Gamma_{\beta}\approx \SI{1/880}{s^{-1}}$. The energy dependent losses due to nuclear absorption and gaps in the surface can be estimated to be
\[
\Gamma_{\eta}(E) = \frac{\mu(E)v(E)}{\lambda(E)} \quad\quad \Gamma_{\eta}(E) = \frac{v(E)S_{\rm gap}}{\lambda(E)S},
\]
\noindent where $\mu(E)$ is the energy dependent loss per bounce probability, typically in the range of some $\pow{1}{-4}$, and $S/S_{\rm gap}$ is the ratio of cell surface to surface of gaps. The last loss rate $\Gamma_{\rm gas}(E)$ from collisions with rest gas molecules can be neglected for a typical vacuum conditions. Hence, the number of UCN $N(t)$ remaining after a time $t$ inside the volume is:
\begin{equation}
N(t) = \frac{\int {\rm d}E~ N_0(E)\exp\left(-\Gamma(E)t\right)}{\int {\rm d}E~ N_0(E)},
\label{eq:UCN_ExpDecay}
\end{equation}
\noindent where $N_0(E)$ is the energy spectrum of UCN directly after closing the door of the storage cell. However, very often the energy spectrum is not very well known and the measured data can be described sufficiently well using a double-exponential decay:
\begin{equation}
N(t) = N_{\rm f}\exp(-\Gamma_{\rm f}t)+N_{\rm s}\exp(-\Gamma_{\rm s}t),
\label{eq:UCN_DoubleExpDecay}
\end{equation}
\noindent where the index `f' describes a fast and the index `s' a slow loss from the storage volume. Good quality storage volumes for searches for an nEDM report values larger than $1/\Gamma_{\rm f} > \SI{60}{s}$ and $1/\Gamma_{\rm f} > \SI{180}{s}$. For a complete review of UCN physics please refer to the books\,\cite{Golub1991,Ignatovich1990}.
\subsubsection{Production of UCN}
Free neutrons are typically produced in research reactors by nuclear fission or in spallation sources where a high-power proton beam is incident on a neutron rich, often stable element (e.g.\ mercury or lead). These neutrons are then cooled by a heavy or normal water moderator to room-temperature. The energy distribution can be very well approximated by a Maxwell-Boltzmann distribution for slightly higher energies. To further shift the energy spectrum a moderator at liquid hydrogen temperature is used as source of cold neutrons.
Any cold neutron source has already a significant number of UCN in the low-energy tail of its distribution. The extraction through aluminum windows and long guides significantly reduces the amount of available UCN outside the biological-protection.
These problems were beautifully circumvented by the conception and design of a phase-space converter, the UCN-turbine (instrument PF2) at the Institut Laue-Langevin, which Doppler-shifts very cold neutrons to the UCN regime\,\cite{Steyerl1975NIM,Steyerl1986PLA}. Until now, this instrument remained the workhorse and benchmark in UCN-physics.
All other current and next-generation UCN sources\,\cite{Trinks2000NIMA,Saunders2013RSI,Lauss2014,Piegsa2014,Serebrov2014TPL,Masuda2015JPS,Zimmer2015PRC} are based on the superthermal concept proposed by Pendlebury and Golub in 1975\,\cite{Golub1975,Golub1977}. The principal idea is to use collective excitations of the conversion medium to down-scatter neutrons from higher energies to the UCN energy regime\,\cite{Kirch2010}. For this process two materials are of main interest: superfluid helium (He-II), using the phonon-roton excitations with a relatively feeble production rate but profiting from a zero absorption cross-section\,\cite{Korobkina2002,Schmidt-Wellenburg2009}, and solid deuterium (sD$_2$) which has a broad range of excitations leading to a high conversion rate while the finite absorption cross-section reduces the effective layer from which a UCN can escape the material\,\cite{Atchison2007PRL}. Both methods can be adapted in such a way that in principle it should be possible to build UCN sources with greatly increased UCN densities, compared to today's standards\,\cite{Kirch2014FPUA}, delivered to experiments at room temperature. An alternative approach are cryogenic nEDM experiments, that intend to avoid transport losses on the path source to room-temperature precession cell altogether, which are discussed in Sec.\,\ref{sec:CryogenicApproach}.
\subsubsection{Detection of UCN}
At the end of the Ramsey cycle the number of neutrons with spin up $N^{\uparrow}$ (low-field seeking) and spin down $N^{\downarrow}$ (high-field seeking) have to be counted, before they can be fitted to expression\,\eqref{eq:UCNRamseyFormula}. For the discrimination between high and low-field seeking neutrons one can use the high magnetic field inside a thin magnetized iron layer $B_{\rm Fe} \approx \SI{2}{T}$ on an aluminum foil. If such a foil is placed inside a sufficiently strong magnetic field from permanent magnets with a flux return yoke, the low-field seeking neutrons will see a potential step of $V_{\rm F} = V_{\rm Fe} + \SI{2}{T}\!\cdot\!\SI{60}{neV/T} = \SI{330}{neV}$, while the potential step for the high-field seeking UCN is reduced to \SI{60}{neV} ($V_\text{Fe} = \SI{210}{neV}$).
Placed approximately \SI{1}{\meter} below the precession cell, such a foil will reflect all low-field seeking (spin up) UCN, while letting high-field seeking (spin down) pass. In combination with an efficient adiabatic spin-flipper\,\cite{Luschikov1984NIMA,Geltenbort2009NIMA} placed in front of such a foil it is possible to invert the spin-state of the neutrons incident on the foil and detect both spin states. The spectrometer from Ref.\,\cite{Baker2014,Pendlebury2015PRD} used one foil and one spin-flipper that was turned on and off during the counting period for spin-analysis. An alternative approach with two detection beam lines each equipped with one foil and one spin-flipper for each precession cell was used in Ref.\,\cite{Serebrov2015PRC} and is currently employed for the single precession cell setup at PSI\,\cite{Afach2015EPJA}. Once the neutron has passed the spin-analysis stage it is detected in a UCN detector.
The neutron, being an uncharged particle, cannot be detected with conventional particle detectors which are based on ionization processes. While high energetic neutrons can be measured by recoil protons from hydrogen-dense materials, low energetic neutrons $E<\unit[25]{meV}$ are best detected exploiting nuclear reactions . Once a neutron is captured by a suitable nuclei a conventional detector may detect the charged particle emitted by the prompt nuclear reaction. Any impulse or energy information of the incident neutron is lost during this reaction, hence a UCN detector is a simple counter. Table\,\ref{tab:DetectionIsotopes} lists the three most widely used isotopes with a large absorption cross section and suitable charged reaction partners.
\begin{table}%
\begin{tabular}{l|crl}
Isotope & $\sigma_{\rm a}$ (barn) & \multicolumn{2}{c}{Nuclear reaction} \\ \hline \hline
\ensuremath{{}^3\mathrm{He}} & 5330 & ${\rm \ensuremath{{}^3\mathrm{He}} +n }~\longrightarrow$ &${\rm ^3H+ p + \SI{764}{keV}}$ \\
$^6$Li & 940 & $\rm ^6Li + n~\longrightarrow$ &${\rm ^3H + \alpha + \SI{4.78}{MeV}}$ \\
{$^{10}$B} & {3835} &$\rm ^{10}B + n~\longrightarrow$ &$ \rm {^7Li} + \alpha + \SI{2.79}{MeV}~(\SI{6}{\%})$\\
& & $\longrightarrow $& ${\rm ^7Li^{\ast} + \alpha + \SI{2.31}{MeV}}~(\SI{94}{\%})$\\
& & &~$\hookrightarrow {\rm ^7Li + \gamma + \SI{0.48}{MeV} }$ \\
\hline
\end{tabular}
\caption{List of typically-used isotopes in UCN detectors with absorption cross section (for $\lambda=\SI{1.8}{\angstrom}$) and nuclear reaction\,\cite{NeutronDataBooklet}.}
\label{tab:DetectionIsotopes}
\end{table}
In the past \ensuremath{{}^3\mathrm{He}}-gas detectors were the most commonly used detectors in UCN physics. As their signal length is rather long (\SI{2}{\micro\second}) these detector face serious dead-time when operated at new and future UCN sources with count rates approaching \SI{1}{\mega\hertz}. Further, the price of \ensuremath{{}^3\mathrm{He}}{} has steadily risen in the last years due to the shortage in supply.
These prospects triggered several studies of alternatives using $^6$Li and $^{10}$B as converter isotope, in the last decade. In particular designs have been investigated using gaseous detection\,\cite{Morris2009NIMA,Klein2011NIMA,Salvat2012NIMA}, silicon state detection\,\cite{Baker2003NIMA,Lasakov2005NIMA,Lauer2011EPJA}, $^6$Li-doped glass scintillators\,\cite{Ban2009NIMA,Goeltl2013EPJA,Jamieson2015NIMA,Ban2016EPJA} or a $^{10}$B/ZnS(Ag) multilayer\,\cite{Wang2015NIMA}.
\subsection{Statistical sensitivity}
The statistical sensitivity also depends on the transverse spin coherence time $T_2^{\ast}$ of the precessing spins and the UCN losses, as described in equation\,\eqref{eq:UCN_lossrates}, inside the precession chamber. Including these into equation\,\eqref{eq:SensitivityEDM} yields for UCN
\begin{equation}
\sigma\left(\ensuremath{d_\text{n}}\right) = \frac{\hbar}{2\alpha_0\exp\left(-\frac{T}{T_2^{\ast}}\right) T E \sqrt{\langle N(f^{-1})\rangle ft}}.
\label{eq:nEDM_SensitivityUCN}
\end{equation}
For a general analysis we simplify equation\,\eqref{eq:UCN_ExpDecay} to a single exponential decay described by $\langle N(f^{-1}) \rangle=N_0\exp(-t/\overline{\tau})$. The minimum of equation\,\eqref{eq:nEDM_SensitivityUCN} is then at
\[
T = \frac{2T_2^{\ast}\overline{\tau}}{T_2^{\ast}+2\overline{\tau}}.
\]
Hence the best attainable sensitivity for a given electric field $E$, initial polarization $\alpha_0$ and a neutron current of $\dot{N}=N_0f$ is
\begin{equation}
\sigma_{\rm min}(\ensuremath{d_\text{n}}) = \frac{\hbar}{4\alpha_0E\sqrt{N_0ft}}\frac{\exp\left(\frac{2\overline{\tau}}{T_2^{\ast}+2\overline{\tau}}\right)(T_2^{\ast}+2\overline{\tau})}{T_2^{\ast}\overline{\tau}\sqrt{\exp\left(-\frac{2T_2^{\ast}}{T_2^{\ast}+2\overline{\tau}}\right)}}.
\label{eq:nEDM_BestSensitivityUCN}
\end{equation}
In the current nEDM-experiment at PSI, on average 14\,000 UCN are counted every $1/f=\SI{300}{s}$, after a free precession period of $T= \unit[180]{s}$. The measured double exponential decay constants are $t_{\rm f} \approx \SI{70}{s}$ and $t_{\rm s}\approx \SI{300}{s}$, which yields for a single exponential approximation $\dot{N}=\SI{140}{s^{-1}}$ and $\overline{\tau}=\SI{163}{s}$. The transverse depolarization time is regularly above $T_2^{\ast}=\SI{1000}{s}$ while the initial polarization is close to 0.86. With an average electric field of $|E|=\SI{11}{kV/cm}$ this gives a sensitivity of
\begin{equation}
\sigma(\ensuremath{d_\text{n}}(t)) = t^{-1/2}\cdot\unit[\pow{3.25}{-23}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}/\sqrt{Hz}},
\label{eq:nEDM_BestSensitivityUCN_Value}
\end{equation}
\noindent with $t$ the effective measurement time. Table\,\ref{tab:Comparison} sets this value in relation to statistical shot-noise limits of other methods to search for an nEDM.
\begin{table}
\begin{tabular}{rccccccc}
& T & N-current & alpha & E-field & $T_2^{\ast}$ & $\overline{\tau}$ & shot-noise limit \\
Method & (s) & ($s^{-1}$) & & (kV/cm) & (s) & (s) & (${\rm \ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}/\sqrt{Hz}}$) \\% & $\sigma(\ensuremath{d_\text{n}})$ in 1 day ($\pow{1.0}{1-26}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$) & $\sigma(\ensuremath{d_\text{n}})$ in 1 year ($\pow{1}{1-26}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$) \\
\hline
last beam & 0.015 & $\pow{1.0}{6}$ & 0.80 & 100 & ~ & ~ & \pow{8670}{-25} \\% & 300 & 22.5 \\
new beam & 0.1 & $\pow{4.0}{8}$ & 0.75 & 100 & ~ & ~ & \pow{22}{-25} \\% & 0.8 & 0.06 \\
RT UCN & 246 & $\pow{1.4}{2}$ & 0.86 & 11 & 1000 & 160 & \pow{325}{-25} \\% & 11.3 & 0.84 \\
future RT UCN & 333 & $\pow{2.8}{3}$ & 0.95 & 11 & 2000 & 200 & \pow{48.5}{-25} \\% & 1.7 & 0.13 \\
cryo-UCN & 570 & $\pow{2.3}{4}$ & 0.99 & 50 & 2000 & 440 & \pow{1}{-25} \\% & 0.1 & 0.007 \\
\hline
\end{tabular}
\caption{Comparison of the shot-noise limit for different methods. Experiments with ultracold neutron at room temperature (RT) will profit from better UCN sources and a better controlled magnetic environment. The shot-noise limit will further decrease when going to a cryogenic setup within superfluid helium. It has been shown that in such a case the electric field can be up to ten times higher and UCN transport losses may be reduced or even become obsolete when producing UCN and measuring the spin precession within the same volume is feasible. The future beam experiment will mainly profit from high counting statistics and a ten time higher electric field which is possible as no insulator will be needed to separate the electrodes. The most important systematic effect, $\boldsymbol{v}\times \boldsymbol{E}$ effect, shall be controlled by exploiting the pulse structure which permits to use a TOF-method to eliminate any $\boldsymbol{v}$-dependency. Please refer to Tab.\,\ref{tab:WWSearches} for a comprehensive overview of current ongoing and proposed searches.}
\label{tab:Comparison}
\end{table}
\subsection{Cryogenic approach}
\label{sec:CryogenicApproach}
In many ways a completely cryogenic experiment is considered as the ultimate drive to highest sensitivity. The essential objective is to perform the entire experiment inside He-II: UCN production and polarization, Ramsey-type measurement, spin-analysis and detection. This would allow optimal exploitation of the high UCN density of a He-II based source in combination with a 5-10 times higher electric field than in vacuum\,\cite{Ito2014arXiv}.
Ultracold neutron sources based on He-II converters at temperatures below \SI{0.5}{K} have a saturation density of $\rho_{\infty} = \tau P$, where the storage lifetime is essentially limited by the neutron beta decay $\tau_{\beta} \approx \SI{880}{s}$ and $P=\diff\phi/\diff\lambda \cdot \SI{5e-9}{\angstrom\per\cubic\ensuremath{\mathrm{cm}}}$ is the production for a differential cold neutron flux at \SI{0.89}{\angstrom}\,\cite{Schmidt-Wellenburg2009}.
Storage times of $1/\Gamma_{\rm f} = \SI{97(13)}{s}$ and $1/\Gamma_{\rm s} = \SI{580(30)}{s}$ extracted with a two-exponential fit were reported from tests of a cyrogenic storage volume of \SI{3}{\liter} coated with deuterated polystyrene\,\cite{Leung2016Mainz}. Combining this with a flux from a CN beam like PF1 at ILL\,\cite{Abele2006} results in a saturation density of $\rho_{\infty}
\approx \SI{4500}{\per\cubic\ensuremath{\mathrm{cm}}}$.
The cryo-EDM collaboration\,\cite{Baker2010JPhCS} followed the classical UCN storage and measurement approach by connecting a cyrogenic precession cell to a He-II UCN source. The UCN were transported from the source in beryllium coated guides immersed in the \SI{0.5}{K} cold liquid helium to the precession cell were the electric and magnetic field were to be applied. After the Ramsey cycle the precession cell was connected to a cryogenic UCN detection system\,\cite{Baker2003NIMA}. The funding for the project was stopped in 2014 after persistent technical difficulties and an accumulated delay of several years.
An even more radical approach is followed by the SNS-EDM collaboration\,\cite{Golub1994PhR} with the intention to avoid UCN transport altogether. In this case the detection of the neutron spin precession is realized using a dilute solution of spin polarized \ensuremath{{}^3\mathrm{He}}{} in He-II. The absorption of a neutron only occurs when the total spin is zero. The reaction products, a proton and a tritium nuclei (see Tab.\,\ref{tab:DetectionIsotopes}), produce ultra violet light in He-II which can be detected by photomultiplier tubes. Therefore it becomes possible to use the intensity of scintillation light, after a classical Ramsey cycle, as measure of spin asymmetry. Alternatively it might be possible to measure the scintillation light from the beating of the two precession frequencies of \ensuremath{{}^3\mathrm{He}}{} and neutron or to employ the `dressed spin' method in the presence of a strong oscillating magnetic field\,\cite{Esler2007PRC}.
Combining these cryogenic prospects in equation\,\ref{eq:nEDM_SensitivityUCN} with $\langle N(T)\rangle f = \SI{2.3e4}{\per\second}$, $t_f=1/\Gamma=\SI{440}{s}$ and $T_2^{\ast} = \SI{2000}{s}$ yield a shot noise limit of \SI{1e-25}{\elementarycharge \ensuremath{\mathrm{cm}} \persqrthz}, which seems by today's technical standard as the ultimate feasible sensitivity. After two years of operation this experiment could reach a sensitivity below \SI{2e-29}{\elementarycharge \ensuremath{\mathrm{cm}}}.
%
\section{INTRODUCTION}
One of the most intriguing questions in cosmology and perhaps even in all of physics is: ``Why is there so much matter in the Universe and so little antimatter?'' This observation is even more puzzling as the near-perfect symmetry of particles and anti-particles is a firmly established fact of collider physics.
Outside the laboratories of particle physics antimatter can be seen only in cosmic rays which are a result of particle collisions in the atmosphere consistent with the same processes observed in particle accelerators or in the decay of $^{40}$K, the only natural positron emitter with a branching ratio of \SI{12}{ppm}\,\cite{Engelkemeir1962PR}.
Until today there is no evidence for any primordial antimatter within our galaxy or even beyond.
There is no indication for any form of co-existence of matter and antimatter in clusters or galaxies within our Universe.
If such regions would exist, one would expect a distinct source of annihilation gammas from any boundary between an antimatter and matter part of the Universe, which has not been observed.
Hence, it is usually conclude that our visible Universe is made entirely of matter and is intrinsic matter non-symmetric.
The natural hypothesis for an initial state prior to the Big Bang is a fully symmetric state indicating the asymmetry of matter and antimatter $\eta = 0$. To describe this asymmetry I use the standard notation:
\begin{equation}
\eta = \frac{n_{\rm b}-n_{\rm \bar{b}}}{n_{\gamma}},
\label{eq:BAU}
\end{equation}
\noindent where $n_{\rm b}$ and $n_{\rm \overline{b}}$ are the number densities of baryons and antibaryons (most matter is made out of baryons which are themselves particles made out of quarks and gluons, e.g.\ protons and neutrons), and $n_{\gamma}$ is the number density of photons emitted during the primordial process of matter/anitmatter creation.
Starting from a symmetric state $\eta = 0$, the Big Bang would have created enormous quantities of baryons and antibaryons, which, while the Universe expanded, would have found each other and annihilated, emitting plenty of photons until the baryon/antibaryon density became low enough for the process to stop.
Assuming such a scenario including the known symmetry violations of the Standard Model of particle physics (SM) results in $\eta \approx 10^{-18}$\cite{Riotto1999ARNPS,Morrissey2012NJP}, while the observed asymmetry is eight orders of magnitude larger when derived from the measurement of the microwave background of the universe $\eta=6.1^{+0.3}_{-0.2}\!\times\!10^{-10}$\,\cite{Dine2012}, or from the abundance of light elements produced in primordial nucleosynthesis with $5.1\!\times\!10^{-10}<\eta<6.7\!\times\!10^{-10}$\,\cite{Cyburt2008JCAP}. The very simple and efficient solution would be to accept an initial condition which is asymmetric. Such a sort of fine tuning is strongly disfavored as it contradicts any naturalness principle. Hence, whatever the answer to this outstanding problem might be, it relates to a fundamental structural asymmetry of physics and might have left a faint footprint on all fundamental particles and their interactions.
With the discovery of the Higgs boson\,\cite{Aad2012PLB,Chatrchyan2012PLB} particle physics at particle accelerators celebrated its latest success, following an exceptional series of discoveries that conformed with the SM, which beautifully describes all particles and interactions known from laboratory experiments.
The possibility that new physics could be at mass scales beyond the reach of collider experiments or coupling very weakly to known particles, in combination with astrophysical observations (e.g.\ dark matter, neutrino oscillations) which cannot be explained by the SM has nurtured vivid interest in high-precision physics at low energies\,\cite{Raidal2008} in recent years.
Some of these experiments look for the remaining faint footprints of structural asymmetries which might have led to the observed baryon asymmetry of the Universe (BAU). One such high-precision search for new physics is the quest to find an electric dipole moment (EDM) of the neutron.
A very intuitive picture of an EDM consists of two point charges, one with charge $+e$, the other with $-e$, separated by a distance $\boldsymbol{r}$ giving rise to $\boldsymbol{d} = e\cdot \boldsymbol{r}$. In a more general classical limit for a continuous charge distribution $\rho(\boldsymbol{r})$ the EDM can be described by
\begin{equation}
\boldsymbol{d}^{\rm cl} = \int{\rm d\boldsymbol{r}\,\boldsymbol{r}\rho(\boldsymbol{r})};
\label{eq:dnClassic}
\end{equation}
\noindent where $\boldsymbol{r}$ is integrated over the extended object. This expression will be redefined in the section on symmetries and EDMs (Sec.\,\ref{sec:Symmetries}) for point-like quantum objects. Thus, the common unit in which EDMs are measured is $\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$: equivalent to a separation between two opposite charged elementary point charges.
Any particle which has a well defined non-degenerate ground state and an EDM ($d\neq0$) violates {\it parity (P)}-symmetry and the symmetry under {\it time}-reversal {\it(T)}. This also indicates a violation of the combined symmetry of {\it charge (C)} and {\it P}, under the assumption of Lorentz and CPT-invariance. Thus, the discovery of an EDM of the neutron would indicate that {\it CP}-symmetry is violated, and this in turn may be connected to the structural asymmetry required to create a matter dominated Universe. {\it CP}-violation is one of three necessary conditions for a baryon asymmetry evolving from an initial symmetric state. These three Sakharov criteria\,\cite{Sakharov1991}, namely i) baryon number violation, ii) {\it C} and {\it CP}-symmetry violation, and iii) departure from thermal equilibrium, were identified as the minimal set of conditions which any theory must fulfill to explain the observed BAU\@. For a more detailed explanation of the role of symmetries in EDMs see Sec.\,\ref{sec:Symmetries}.
While in the recent past remarkable progress has been made in the search for an electron EDM using the enormous electric field inside a ThO molecule\,\cite{Baron2014Science}, and also for the diamagnetic \ensuremath{{}^{199}\text{Hg}}-EDM\,\cite{Graner2016PRL}, that of the neutron has slowed down. The latest result for the neutron EDM, $|\ensuremath{d_\text{n}}|\!<\!\unit[\pow{3}{-26}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$\,\cite{Pendlebury2015PRD}, arises from a re-analysis of the data first published in 2006\,\cite{Baker2006} (then $|\ensuremath{d_\text{n}}|\!<\!\unit[\pow{2.9}{-26}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$) and confirms the original result. Different EDM limits are summarized in Tab.\,\ref{tab:EDMs}, while the interested reader is referred to Ref.\,\cite{Pospelov2005,Engel2013PPNP} or a detailed overview of the interplay between different EDMs and theory\,\cite{Chupp2015PR}.
\begin{table}%
\centering
\begin{tabular}{rclr}
\hline
EDM & limit & C.L.\ & Ref.\\
\hline
ThO & $d_e<\pow{8.7}{-29}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$ &90\% & ~\cite{Baron2014Science}\\
\ensuremath{{}^{199}\text{Hg}} & $d_{\rm Hg}<\pow{7.4}{-30}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$ &95\% &~\cite{Graner2016PRL}\\
neutron & $\ensuremath{d_\text{n}} <\pow{3.0}{-26}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$ &90\% & ~\cite{Pendlebury2015PRD}\\
\hline
\end{tabular}
\caption{Most relevant experimental limits on electric dipole moments.}
\label{tab:EDMs}
\end{table}
The first searches for a neutron EDM, starting in the 1950s\,\cite{Purcell1950PR,Smith1957PR}, were undertaken using thermal, and later cold, neutron beams from reactors. The last beam experiment\,\cite{Dress1977} published a limit of $\ensuremath{d_\text{n}}\!<\!\unit[\pow{3}{-24}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ (C.L.90\%) in 1977 and was limited by the velocity-dependent $\boldsymbol{v\!\times\!E}$ systematic effect. At the beginning of that decade, the first experiments using ultracold neutrons\,\cite{Lushchikov1969JETPL,Shapiro1970SPU} were proposed with first results published at the end of the 1970s\,\cite{Altarev1980NuPhA}.
Today several competing collaborations around the world are pursuing new experiments to improve the sensitivity to the nEDM and by that establish a finite value or push the limit on the nEDM by up to two orders of magnitude in the next decade or so. This review is based on an earlier version\,\cite{Schmidt-Wellenburg2016LASNPA} which was published as part of the proceedings of the XI Latin American Symposium on Nuclear Physics and Applications in Medellin, Columbia. Section\,\ref{sec:ExpTech} presents the state-of-the-art methods and techniques which are used in nEDM searches with UCN, while section\,\ref{sec:Systematics} discusses the most important systematic effects. Finally, in section\,\ref{sec:WWnEDM} an overview of the worldwide activities is given.
\section{The nEDM in the Standard Model and beyond }
\input{SMandBeyond}
\section{Experimental Techniques}
\input{ExperimentalTechniques}
\section{Systematic effects}
\input{Systematics}
\section{WORLD-WIDE EFFORTS}
\input{WWnEDM}
\section{CONCLUSION}
Several groups worldwide are pursuing promising approaches to improve the sensitivity for a permanent electric dipole moment of a fundamental particle. In the next few years a new result will probably improve the current limit on the nEDM slightly, while an order of magnitude improvement seems likely within the next five years or so. Another order of magnitude will require a breakthrough in UCN counting statistics or a dramatic improvement of the electric field which seems to be possible in superfluid helium.
\section{ACKNOWLEDGMENTS}
I would like to thank all my colleagues from the UCN group at PSI and the the nEDM-collaboration for many discussions. In particular I would like to thank
N.~Ayers, K.~Kirch and P.~Harris for reading the manuscript and providing important suggestions and advice.
\input{nEDMReviewAJP_bib}
\end{document}
\subsection{Symmetries and EDMs}
\label{sec:Symmetries}
Symmetries are universally present throughout human culture and can be found throughout many concepts of art and architecture. The eightfold path of Buddhism is visualized in a wheel with an eightfold discrete symmetry. Likewise the Hindu swastika has a fourfold discrete symmetry, and thousands of visitors marvel each day at the symmetric architecture of the Taj Mahal mirrored in the reflecting pool.
At the same time it seems that the breaking of such symmetries, as for instance in the rules of the golden ratio, make a work of art or architecture even more exciting and interesting.
Symmetries and their violation also play an important role in fundamental physics. In physics the concept of symmetry is strictly defined by a transformation of an object which does not alter the object; one says that the object is invariant under this transformation. The more such transformations are allowed, the higher the degree of symmetry. As an example take again the eightfold wheel visualizing the principle virtues of Buddhism. One may mirror it along any of its spokes passing through the origin, or one may rotate it by any multiple of $\pi/4$, and it will not change its appearance as it has an eightfold rotational symmetry. A perfect circle has an even higher symmetry, as one can rotate it by any angle without changing its appearance.
As early as 1918 Emmy Noether\,\cite{Noether1918} showed that the conversation of energy, momentum and angular momentum can be directly derived from the symmetry of the laws of physic under translation in time, and under translation and rotation in space. In this context a physical law is symmetric under transformation if any transformation of the system under this operation does not alter the physical law.
In modern particle physics three discrete symmetries, charge conjugation ({\it C}), parity inversion ({\it P}), and time reversal ({\it T}) , play an outstanding role in our understanding of nature. Their combined conservation ({\it CPT} symmetry) together with Lorentz invariance are cornerstones of today's conception of fundamental physics.
All three of the individual symmetries are violated in the weak sector of the standard model.
\subsubsection*{Charge}
\label{sec:Charge}
The notion of charge in modern particle physics extends beyond the classical notion of electrical charge to all charge quantum number of the particle. In this sense charge conjugation is not only the reversal of the electrical charge of the object, but also of its color charge, its lepton or baryon number, its flavor quantum number and so on. Therefore, charge conjugation transforms a particle into its anti-particle by reversing all charge quantum numbers of the particle, for example the electron into the positron or the neutron ($\mathrm{udd}$) into an anti-neutron ($\mathrm{\bar{u}\bar{d}\bar{d}}$).\\
A system or physical process is considered as invariant under charge conjugation if the observed process is indistinguishable whether anti-particles or particles with the corresponding conjugated fields participate. In this sense the movement of an electron in an electric field is indistinguishable from a positron in an electric field of same magnitude but opposite direction.
\subsubsection*{Parity}
Parity transformation is defined as the sign inversion of all space coordinates in a Cartesian coordinate system. As demonstrated in Fig.\,\ref{fig:IllustrParity}a) this leads to nothing new in two dimension where it can be replaced by a rotation around the origin by $\pi$. Whereas in three dimensions, illustrated in Fig.\,\ref{fig:IllustrParity}b), it becomes obvious that a parity transformation cannot be replaced by any number of discrete rotation. This additional symmetry can also be called the symmetry between right-handiness and left-handedness. There seems to be no obvious reason why in empty space nature would have a preference for right or left-handedness.
\begin{figure}%
\includegraphics[width=0.90\columnwidth]{./art/ParityIn2Dand3D.png}%
\caption{Illustration of parity transformation in comparison to rotational symmetries in two and three dimensions. In two and every higher even dimension the discrete parity transformation can be replaced by a set of discrete rotational transformation. (a) The discrete parity transformation ($\vec{r} \rightarrow -\vec{r}$) of all coordinates in 2 dimensions of the object is equivalent to a rotation by $\pi$ around the origin. Further the object is invariant under parity transformation, as it is indistinguishable from the original object. Note, the little blue square marks one end of the object in order to show that the object is transformed, but is not to be regarded as part of the object itself. (b) The parity transformation in 3 dimension of a screw thread ($\vec{r} \rightarrow -\vec{r}$) cannot be replaced by a series of discrete rotational transformations. Further, the sense of rotation of the screw changes. This is called the handedness. Parity transforms a right handed object into a left handed object.}
\label{fig:IllustrParity}%
\end{figure}
\label{sec:Parity}
\subsubsection*{Time}
Time reversal symmetry indicates that a physical process which is described by a Lagrangian $\mathcal{L}(t)$ or Hamiltonian $\mathcal{H}(t)$ transforms into a process $\mathcal{L}(-t)$,$\mathcal{H}(-t)$ which again is an allowed process by physical laws. This does not mean that the actual ``arrow'' of time has to be inverted, but instead that in a video of the process it becomes indistinguishable whether the video is played forward or backward. The classical example for such processes is the elastic scattering of pool balls on a pool table. In the case of a head-on collision with total momentum transfer the moving ball will come to rest while the ball which was initially at rest will start moving with the same momentum. In this simple case it is obvious that from watching the video is is unclear whether it is played forward or backward. Still our conception of time is mostly define from seemingly irreversible processes: e.g.\ a glass falling off the table and breaking in thousand of pieces or the initial break in snooker just to mention two. In these cases it seems to be easy to distinguish whether the video is played forward or backwards. But amusingly the processes are still invariant under time reversal symmetry. Our everyday experience just tells us that it is extremely unlikely to match the initial conditions so perfectly that an observation of the seemingly backward process is possible in nature.
\label{sec:Time}
\subsubsection*{CP-violation and the nEDM}
In the electroweak sector of the SM {\it C} and {\it P} are maximally violated, recall the (V-A)-structure of the interaction, while the combined symmetry of charge and parity ({\it CP}) is conserved in most cases.
Though high-precision experiments showed that {\it CP} is also violated in rare decays of {\it K}\,\cite{Christenson1964} and \textit{B} mesons\,\cite{Abe2001PRL,Aubert2001}, an essential ingredient for baryogenesis\,\cite{Sakharov1991}. Unfortunately this known source of {\it CP}-violation is too small to explain the observed baryon asymmetry of the universe\,\cite{Riotto1999ARNPS,Morrissey2012NJP} in combination with the other known ingredients from the standard model.
In the classical concept of an EDM described by Eq.\,\eqref{eq:dnClassic} the EDM is a vector changing its direction with parity: $P(\boldsymbol{d}^{\rm cl}) = -\boldsymbol{d}^{\rm cl}$. In a particle with angular momentum $\boldsymbol{j}$ or spin $\sigma = \hbar/2$ all vectors and pseudo-vectors are either aligned or anti-aligned with the pseudo-vector of angular momentum of the particle, as the transverse component is averaged out due to rotation. Therefore one may write
\begin{equation}
\boldsymbol{d}^{\rm cl} = \delta_n \boldsymbol{j} \quad \overset{P}{\longrightarrow}\quad -\boldsymbol{d}^{\rm cl} = \delta_n \boldsymbol{j} .
\label{eq:clHamiltonianP}
\end{equation}
\noindent Both equations can only be fulfilled if either $\boldsymbol{d}^{\rm cl} = 0$, or parity is violated, or the angular momentum is not defining a non-degenerated state.
However, all experiments involving neutrons are obeying to Pauli exclusion principle with a twofold degeneracy in the absence of a magnetic field. This indicates that the spin quantum number is sufficient to describe the neutron ground state. Similarly one can write for a {\it T}-transformation
\begin{equation}
\boldsymbol{d}^{\rm cl} = \delta_n \boldsymbol{j} \quad \overset{T}{\longrightarrow}\quad \boldsymbol{d}^{\rm cl} = -\delta_n \boldsymbol{j} .
\label{eq:clHamiltonianT}
\end{equation}
Which again can only be fulfilled by violating {\it T} symmetry.
In the same way one can show in quantum mechanics that if a neutron were to have an ``electric dipole moment $\ensuremath{d_\text{n}}{}$ then, as
any vector operator in quantum mechanics, it would be connected to
the spin operator as $ \ensuremath{d_\text{n}}= \delta_\text{n} \mathbf{j}/j\hbar$, or, for $j= 1/2$ as $\ensuremath{d_\text{n}}= \delta_\text{n}\boldsymbol{\sigma}$,
where \ensuremath{d_\text{n}}{} gives the size of the EDM''\cite{Dubbers2011}\@. In the non-relativistic limit, the interaction Hamiltonian can be written as:
\begin{equation}
H = -\frac{\hbar}{2}(\delta_\text{n}\boldsymbol{\sigma\!\cdot\!E} + \gamma_\text{n} \boldsymbol{\sigma\!\cdot\!B}),
\label{eq:hamiltonian}
\end{equation}
\noindent where $\delta_\text{n}$ and $\gamma_\text{n}$ can be interpreted as scalar coupling strengths of the neutron spin to the electric and magnetic field. The relative sign of the two dipole coupling strengths is not defined as no electric dipole moment has yet been discovered. The magnetic coupling strength is nothing else than the gyromagnetic ratio of the neutron $\gamma_\text{n}/(2\pi)=\SI{-29.1646943(69)}{MHz\per\tesla}$\,\cite{Greene1978}, which is the ratio of the magnetic moment of the neutron $\boldsymbol{\mu_\text{n}}$ to its angular momentum $\sigma = \hbar/2$. Similarly one can introduce a gyro-electric ratio in combination with the electric dipole moment. Equation\,(\ref{eq:hamiltonian}) and Fig.\,\ref{fig:CPViolCartoon} demonstrate that the eigenstates of the Hamiltonian change when applying either a {\it T} or {\it P}- transformation to the Hamiltonian, indicating again the violation of {\it T} and {\it P}-symmetry.
The CPT-theorem (see standard text books on field theory, e.g.\,\cite{Maggiore2005Book}) is fundamental to any modern quantum field theory and states, that any locally Lorentz-covariant field theory of a point like particle is CPT invariant. As a consequence the observation of a \ensuremath{d_\text{n}}{} would not only indicate the breaking of time reversal symmetry but also CPV and would therefore constrain any model of CPV that attempts to explain the BAU\@.
In any case it would manifest a new, flavor maintaining source of CPV.
\begin{figure}%
\centering
\includegraphics[width=0.3\columnwidth]{./art/CartoonPT-violation.png}%
\caption{Pictogram illustrating the {\it P} and {\it T}-violation of a nEDM in the presence of an electric and magnetic field.}%
\label{fig:CPViolCartoon}%
\end{figure}
\subsection{The standard model prediction}
In the SM two sources of CPV exist: {\it i})~In the weak interaction the weak mass eigenstates of the quarks are not identical to the flavor eigenstates. Both eigenstates are interconnected via the Kobayashi-Maskawa matrix $V_\text{KM}$ which has one single phase $\delta$ which induces the observed CPV in the K and B meson decays.
{\it ii})~The second source is the QCD vacuum polarization term, the only CP-odd term of dimension four in the SM QCD Lagrangian. For a detailed discussion see also the reviews by Pospelov and Ritz\,\cite{Pospelov2005} and by Seng\,\cite{Seng2015PRC}.
The Kobayashi-Maskawa matrix $V_\text{KM}$ can be written as
\begin{equation}
V_\text{KM} = \left(\begin{array}{ccc}
c_{12}c_{13} & s_{12}c_{13} & s_{12}e^{-i\delta} \\
-s_{12}c_{23}-c_{12}s_{23}s_{13}e^{-i\delta} & c_{12}c_{23}-s_{12}s_{23}s_{13}e^{-i\delta} & s_{23}\\
s_{12}s_{23} - c_{13}c_{23}c_{13}e^{-i\delta} & - c_{12}s_{23}s_{13}e^{-i\delta}& c_{23}c_{13}\\
\end{array}\right)
\label{eq:VKM}
\end{equation}
\noindent where $c_{ij} = \cos\theta_{ij}$, $s_{ij} = \sin\theta_{ij}$ and $\delta \approx \unit[1.20]{rad}$ is the CPV phase. It is impossible to write down a tree level diagram generating an electric dipole interaction of one quark of the neutron with the electric field. At the one-loop level, shown in Fig.\,\ref{fig:WeakDiagramloops}a, any phase term of a $V_{ij}$ element at one vertex will be canceled by the complex conjugated phase term at the second vertex $V_{ij}^{\ast}$. Further, Shabalin\,\cite{Shabalin1983} showed that the contributions of all second order processes to an nEDM cancel. The largest SM contribution is at the three-loop level via a strong penguin diagram\,\cite{Khriplovich1982PL} (see Fig.\,\ref{fig:WeakDiagramloops}b) which amounts to an approximate $\ensuremath{d_\text{n}}{}^\text{KM}$ of \unit[\pow{1}{-32}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}\,\cite{Pospelov2005,Seng2015PRC}, well below current and most probable all future experimental sensitivities.
\begin{figure}%
\centering
\includegraphics[width = 0.25\columnwidth]{./art/nEDM_SM-1-loop.png}%
\hspace{3cm}
\includegraphics[width = 0.4\columnwidth]{./art/nEDM3-loop.png}
\caption{Diagrams which involve the CPV phase of the $V_{\rm KM}$-matrix. a)~First loop contribution does not give rise to an EDM, as the phase from the first vertex cancels in the complex conjugate second vertex. b)~This is the largest SM contribution generated by the CPV phase of the $V_{\rm KM}$-matrix. The crossed vertex, shown as insert, is a four-quark vertex known as strong penguin diagram in which not all phase contributions cancel.}%
\label{fig:WeakDiagramloops}%
\end{figure}
An additional source of CPV in the SM is the vacuum term of the QCD-Lagrangian:
\begin{equation}
\mathcal{L}_\text{QCD}^\text{CPV}=\frac{g_s^2}{32\pi^2}\overline{\theta}G^a_{\mu\nu}\widetilde{G}^{\mu\nu,a},
\label{eq:QCDDim4}
\end{equation}
\noindent the only CPV dim-4 operator, where $g_s$ is the coupling constant of the strong interaction, $ \overline{\theta}$ is a phase which also includes the CPV phase of the weak interaction and $G^a_{\mu\nu}$ is the gluon field tensor. The structure of the gluon field tensor times its dual corresponds in electro-magnetism to a scalar product of $\boldsymbol{E}\cdot\boldsymbol{B}$ which is odd under {\it P} and {\it T} reversal. From a dimensional analysis\,\cite{Khriplovich1997Book} one can estimate the size of an nEDM generated by this term:
\begin{equation}
\ensuremath{d_\text{n}}^\text{QCD} \approx \overline{\theta}\cdot\unit[\pow{1}{-16}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}.
\label{eq:QCDThetaApprox}
\end{equation}
\noindent Hence, the current experimental limit on an nEDM is also a limit on $\overline{\theta} \lesssim 10^{-10}$. This is astonishing, as $\overline{\theta}$ is a phase which in principle could acquire any value between $0$ and $2\pi$. It is considered as unnatural that the value is so tiny. Possible solutions to this ``strong {\it CP} problem'' include having at least one mass-less quark, or a mechanism proposed by Peccei and Quinn\,\cite{Peccei1977,Wilczek1978,Weinberg1978}, which gives rise to the axion, a Nambu-Goldstone boson. For searches for the axion and axion like particles see, e.g., Ref.\,\cite{Baer2015}. However, if an nEDM were to be found further measurements of EDMs (proton, electron...) would be necessary to distinguish the CPV source(s) and to explain the role of the tiny $\overline{\theta}$-term.
\subsection{Generic sensitivity of an nEDM to physics beyond the SM}
A comprehensive summery of different BSM scenario which provide viable sources of CPV may be found in the reviews Refs\,\cite{Jaeckel2012JHEP,Engel2013PPNP}. Neglecting for a moment the contribution to an nEDM from the unnatural small $\overline{\theta}$ parameter, it is clear that any observed $\ensuremath{d_\text{n}}>\unit[10^{-30}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ can only be explained by new physics. Generically most BSMs provide several CPV phases and new particles which could already contribute at the one-loop quark level to an observable nEDM\@. A typical order of magnitude analysis from super-symmetric (SUSY) models (e.g.\,\cite{Abel2006JHEP}) gives:
\begin{equation}
\ensuremath{d_\text{n}} \sim \left(\frac{\unit[300]{GeV}}{\Lambda_{\rm SUSY}}\right)^2\sin\phi_{\rm CP}\times10^{-24}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}},
\label{eq:BSMapprox}
\end{equation}
\noindent where $\phi_{\rm CP}$ represents the relevant possible CPV phases of the model and $\Lambda_{\rm SUSY}$ is the SUSY mass scale. The current experimental limit already implies that models either have to be considerably fine tuned to have a small $\phi_{\rm CP}$ or to suppress 1-loop contributions, or that the SUSY-scale is considerably above the weak scale in the range of some TeV\@. Figure\,\ref{fig:SUSYCPproblem} illustrates this conundrum known as SUSY CP problem. The authors of Ref.\,\cite{Engel2013PPNP} generalize this SUSY approach and find similar model-independent constraint for a general BSM scale.
\begin{figure}%
\includegraphics[width=0.6\columnwidth]{SUSY-problem.png}%
\caption{Constraints on two CP violating phases of a generic SUSY-model using limits from the neutron\,\cite{Pendlebury2015PRD}, Hg\,\cite{Graner2016PRL} and ThO\,\cite{Baron2014Science} (electron). The allowed phase space is already stringently reduced for a SUSY scale at \SI{2}{TeV}. Updated figure courtesy of A.~Ritz, original figure appeared in Ref.\,\cite{Pospelov2005}.}
\label{fig:SUSYCPproblem}
\end{figure}
\subsection{Requirements for the stability of the magnetic field}
In order not to degrade an initial statistical sensitivity $\sigma(\ensuremath{d_\text{n}})$ by random magnetic field drifts it is necessary to guarantee that
\begin{equation}
\delta B = \left(\delta B^{\parallel}-\delta B^{\nparallel}\right) \ll \frac{2\left|E\right|\sigma(\ensuremath{d_\text{n}})}{\mu_\text{n}},
\label{eq:MagneticFieldStability}
\end{equation}
\noindent for two measurements with inverted electric field configuration. This results in a required field stability of much better than \SI{25}{fT} over a period of \SI{9}{\hour}, for a field reversal every \SI{4.5}{h} and a neutron statistical sensitivity of $\sigma(\ensuremath{d_\text{n}})=\pow{1}{-24}\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}$ per Ramsey cycle. The stability of the magnetic field may be best characterized by an Allan deviation (AD) of the magnetic field measured using an auxiliary magnetometer. The AD is defined as a function of averaging time $\tau$ by:
\begin{equation}
\sigma_{\rm AD}\left(\tau_{\rm AD})\right) = \sqrt{\frac{1}{2\left(N_{\rm AD}-1\right)}\sum^{N_{\rm AD}-1}_{i=1}\left(B_i(\tau_{\rm AD})-B_{i+1}(\tau_{\rm AD})\right)^2},
\label{eq:AD}
\end{equation}
\noindent where $N_{\rm AD}=T_{\rm AD}/\tau_{\rm AD}$ is the number of non-overlapping samples of magnetic field values $B_i(\tau_{\rm AD})$ each averaged over a time $\tau_{\rm AD}$ in a measurement of total duration $T_{\rm AD}$. Figure\,\ref{fig:AD} shows an AD for several averaging times $\tau_{\rm AD}$ taken from reference\,\cite{Afach2014JAP} describing a system to stabilize the magnetic field for the nEDM search at PSI\@. The best field stability is reached during night times for averaging times of not more than \SI{100}{s}, but even then the typical variation is of the order \SI{100}{pT}, far too large for any competitive nEDM search. For this reason all current and proposed future experiments use a multilayer magnetic shield made of a high-permeability ($\mu$)
NiFe alloy which may have a static shielding factor of better than $100\,000$\,\cite{Sumner1987JPhD,Altarev2014RSI}.
In addition, an active surrounding field compensation system\,\cite{Afach2014JAP} based on several fluxgate magnetometer in close vicinity of the shield may be used to keep the shield in a stable magnetic field. Such a scheme uses a feedback algorithm to actively control a set of large coils around the magnetic shield to correct for magnetic-field changes.
Figure\,\ref{fig:AD_CsM} shows the Allan deviation measured using a three axis vector magnetometer\,\cite{Afach2015OExpress}
inside the PSI magnetic shield.
An improvement of approximate 3 orders of magnitude can be observed when compared to the AD of the outside field. The AD minimum of the field magnitude $\sigma_{\rm AD} \approx \SI{80}{fT}$ is at just below $\tau = \SI{100}{s}$.
Again the field stability would not be nearly as good as required even if the electric field polarity were to be changed after every Ramsey cycle ($1/f=\tau_{\rm AD} \approx \SI{300}{s}$) and the relative long charging times (\SI{200}{s}) of the electrodes were significantly reduced. To circumvent this problem all recent and future searches use one or more monitor magnetometer(s) to correct for magnetic-field drifts.
\begin{figure}%
\includegraphics[width=0.9\columnwidth]{AD_SurroundingField}%
\caption{Allan deviation of the $z$-component of the environmental magnetic field at the location of PSI's nEDM apparatus. The field is measured using 10 fluxgate field sensors placed in the near vicinity of the four-layer magnetic shield. The difference between night time (solid lines) and day time is significant. Figure taken from reference\,\cite{Afach2014JAP}.}%
\label{fig:AD}%
\end{figure}
\begin{figure}%
\includegraphics[width=0.9\columnwidth]{AD_CsM}%
\caption{Allan deviation of the magnitude of the magnetic field inside the four layer magnetic shield at PSI measured by a three-axis atomic vapor (Cs) vector magnetometer, data from Ref.\,\cite{Afach2015OExpress}. Note that the main magnetic field of the experiment is aligned with $z$ and has a magnitude of about \SI{1}{\mbox{\micro T}}.}%
\label{fig:AD_CsM}%
\end{figure}
\subsection{Magnetic field monitors}
As we have seen, fluctuating magnetic fields add noise to the measured nEDM data and limit the sensitivity of the experiment.
Using standard error propagation on equation\,\eqref{eq:DiffConfig2} one finds
\begin{equation}
\delta \ensuremath{d_\text{n}} = \sqrt{\sigma^2(\ensuremath{d_\text{n}}) + 2\left( \frac{\mu_{\rm n}}{2|E|}\delta B\right)^2}.
\label{eq:errorPropagation}
\end{equation}
\noindent Here the first term is the sensitivity due to neutron counting statistics, typically referred to
as the statistical sensitivity. The second term is the systematic effect that worsen
the counting statistics due to its stochastic nature. In general one would like to have the second term at least a factor of four smaller than the first one so that it would have only a
minor effect on the sensitivity. This can be achieved using additional magnetometers to compensate changes of the magnetic field between measurements. This requires a magnetometer with an accuracy of better than
\begin{equation}
\delta B_\text{mag} \leq \frac{\sigma(\ensuremath{d_\text{n}})}{4} \frac{\left|E\right|}{\mu_\text{n}}
\label{eq:SensMagn}
\end{equation}
\noindent per Ramsey cycle. The proportionality constant for a $\unit[12]{kV/cm}$ field is $\mu_\text{n}/\left|E\right| = \unit[6\!\times\!10^{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}/fT}$. Hence, the intrinsic sensitivity of the monitor magnetometer must be better than $\sigma(B_\text{mag})\leq \SI{5}{fT}$ for a statistical sensitivity of $\unit[\pow{1}{-25}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ per cycle, a value typically aimed for by many new searches. Conceptually this might be accomplished by either using several high-sensitivity magnetometers, for example optically pumped atomic vapor magnetometers\,\cite{Budker2007} or SQUIDS\,\cite{Burghoff2007}, in the close vicinity of the neutron precession cell, or by using a magnetically susceptible isotope that can be injected into the same volume where the neutrons precess (cohabiting magnetometer).
\subsubsection{Arrays of high-precision magnetometers for field monitoring}
Point like (actually spheres with $r=\SI{15}{mm}$) magnetometers placed just outside of the precession chamber but still inside the magnetic shield are often proposed or used to monitor the magnetic field. They have the clear advantage not to see the electric field and hence should be free of an electric-field correlated effect.
In the past, several experiments inferred the average magnetic field to which the neutrons were exposed by combining the values of several optically pumped atomic magnetometers (OPM)\,\cite{Smith1990PLB,Altarev1996,Serebrov2015PRC}. Further, the currently operating experiment at the PSI uses in addition to a cohabiting magnetometer an array of OPMs to extract magnetic field information. All of these magnetometers are scalar magnetometers, essentially measuring the magnitude of the magnetic field at a point in space\,\cite{Groeger2006}. By combining the measurement values of these OPM it is possible to extract a mean value and higher-order magnetic-field terms like gradients\,\cite{Afach2014PLB}.
In the case of homogeneous field changes affecting all magnetometers in the same way, the mean value may be used to correct for magnetic field drifts. However, if a field change occurs due to a change in magnetization within or close to the observed volume, for example provoked by a leakage current across the insulator ring (see Sec.\,\ref{sec:correlatedSys} and references therein), the combined value may show no field change even though the neutrons are exposed to a different field. This could effectively lead to the observation of an EDM signal, as pointed out in Ref.\,\cite{Lamoreaux2009}.
In Ref.\,\cite{Afach2014PLB} optical pump cesium magnetometers (Cs-OPM) were successfully used to decompose the magnetic field of the precession chamber into spherical polynomial harmonics to extract the vertical magnetic-field gradient $\partial B_z/\partial z$. New developments focusing on accurate vector information from OPM\,\cite{Afach2015OExpress} will further refine such methods and could prove useful in the future to identify and correct for systematic effects, as from a leakage current or by magnetic-field gradients $\partial B_z/\partial z$ in equations\,(\ref{eq:falseNeutronEDMHg}, \ref{eq:neutronGeomEffect}).
\subsubsection{Cohabiting magnetometers}
The second type of field monitor, using magnetically susceptible isotopes that occupy the same volume as the neutrons, is considered essential to most future experiments and is currently also in use within the spectrometer at PSI\@. Three different isotopes, all of nuclear spin 1/2, are typically considered for this task: \ensuremath{{}^3\mathrm{He}}\,\cite{Golub1983JPL,Ramsey1984}, \ensuremath{{}^{199}\text{Hg}}{} and \ensuremath{{}^{129}\mathrm{Xe}}. While \ensuremath{{}^{199}\text{Hg}}{} is already employed\,\cite{Green1998} the other two are proposed for spectrometers in the future. Although these magnetometers allow for a cycle-to-cycle correction of the neutron precession frequency as shown in Fig.\,\ref{fig:HgCorrectednFrequency}, they pose a certain risk of transferring systematic effects to the neutron measurement. Furthermore, even if all correlated systematic effects (Sec.\,\ref{sec:correlatedSys}) from a co-magnetometer are well under control a certain dilution of the neutron sensitivity may occur.
\begin{figure}%
\includegraphics[width=0.9\columnwidth]{HgCorrectednFrequencyPSI}%
\caption{Neutron resonant frequency before and after correction for magnetic-field drifts using a mercury comagnetometer (data sample from nEDM-PSI collaboration\,\cite{Baker2011}.}%
\label{fig:HgCorrectednFrequency}%
\end{figure}
Ultracold neutrons sample the magnetic field of the precession cell slightly differently then the atoms of the thermal co-magnetometer gas. Stored ensembles of UCN have very low energies which leads to the awkward effect that their center-of-mass $\left\langle z_\text{\tiny UCN}\right\rangle$ is slightly below the geometrical center-of-mass $z_{\rm cm}$ of the precession cell. In the presence of a vertical magnetic-field gradient $g_z=\partial B_z/\partial z$ this leads to an uncompensated contribution which adds to equation\,\eqref{eq:errorPropagation}:
\begin{equation}
\delta \ensuremath{d_\text{n}} = \sqrt{\sigma^2(\ensuremath{d_\text{n}}) + 2\left( \frac{\mu_{\rm n}}{2|E|}\delta B\right)^2 + 2\left(\frac{\mu_{\rm n}}{2|E|}\Delta h_{\rm cm} \delta g_z \right)^2},
\label{eq:errorPropagation2}
\end{equation}
\noindent with the center-of-mass offset $\Delta h_{\rm cm} = (z_{\rm cm}-z_{\rm UCN})\approx \SI{4}{mm}$\,\cite{Afach2015PRD,Afach2015PRL}, for the PSI spectrometer, depending on the UCN energy spectrum. If the random variation of the magnetic-field gradient between electric-field configuration is below
\begin{equation}
\delta g_z \leq \frac{\sigma(\ensuremath{d_\text{n}})}{4\Delta h_{\rm cm}} \frac{\left|E\right|}{\mu_\text{n}},
\label{eq:GradientStabilityRequirement}
\end{equation}
\noindent then no monitor for the gradient may be required: this is typically the case for E-field reversal every \SI{4.5}{\hour} at the PSI experiment. Otherwise, a gradiometer is needed to correct for changes with a sensitivity better than \SI{12}{\femto\tesla\per\centi\meter} for a neutron-only statistical sensitivity of \unit[$\pow{1}{-25}$]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}} per cycle.
An even more subtle effect can further reduce the statistical sensitivity when using slowly oscillating magnetic fields to flip the co-magnetometer spin into the transverse plane. After the neutrons are locked into the cell the polarized co-magnetometer gas is injected and a $\pi/2$-pulse is applied to tip the spin of the co-magnetometer atoms by $\pi/2$. This circular rotating magnetic field is also seen by the neutrons and can lead to a small tilt of the neutron spin.
Using Rabi's equation for a resonant transition from $m = +1/2$ to $m=-1/2$\,\cite{Rabi1954}:
\begin{equation}
P_{-1/2} = \frac{\omega_1}{\Omega}\sin\left(\frac{\Omega t_{\pi/2}}{2}\right),
\label{eq:RabisEquation}
\end{equation}
\noindent with
\[
\omega_1 = \frac{\pi\gamma_n}{2\gamma_i t_{\pi/2}}~,\quad\Omega=\sqrt{\left(\omega_{\rm n} - \omega_{\pi/2}\right)^2-\omega_1^2}\quad\text{and}\quad\omega_{\pi/2}\approx \gamma_i B_0
\]
it is possible to calculate this effect. This leads to tilting of the neutron spins by:
\begin{equation}
\Theta = \frac{R\pi}{\Omega t_{\pi/2}}\sin\left(\frac{\Omega t_{\pi/2}}{2}\right),
\label{eq:XeSpinTilt}
\end{equation}
\noindent where $R = \gamma_{\rm n}/\gamma_i$ is the ratio of gyromagnetic ratios of the neutron and the co-magnetometer isotope.
If the phase relation and timing of the co-magnetometer $\pi/2$~pulse and the neutron $\pi/2$~pulse stays constant over time this only leads to a global shift of the neutron resonant frequency which cancels in the nEDM-analysis. However, if phase and timing is not controlled this can lead to a frequency uncertainty of
\begin{equation}
\delta f_{\rm rms} = \frac{1}{\sqrt{2}}\frac{|\Theta|}{2\pi T},
\label{eq:FrequencyShiftPulse}
\end{equation}
\noindent where $T$ is the free precession time. Figure\,\ref{fig:XeFrequencyShift} shows the systematic uncertainty from this effect on the neutron precession frequency in the case of a \ensuremath{{}^{129}\mathrm{Xe}}\ co-magnetometer. For xenon this effect may be as large as \SI{20}{\micro\hertz} which would be approximately a factor twenty larger than the intrinsic sensitivity of the neutrons in a next generation experiment. Therefore it will be of paramount importance to adjust $t_{\pi/2}$ and $B_0$ such that $\sin\left(\Omega t_{\pi/2}/2\right)=0$ and to control phase and timing of the pulse generation. Furthermore it is possible to reduce this effect by choosing an appropriate envelope of the pulse, such as a triangular or a Hann window.
\begin{figure}%
\includegraphics[width=0.9\columnwidth]{XePuls_InitialAngleNeutron}%
\caption{Systematic frequency error as a function of xenon $\pi/2$-pulse duration $t_{\pi/2}$ and main magnetic field $B_0$.}%
\label{fig:XeFrequencyShift}%
\end{figure}
\subsection{Effects mimicking an EDM signal}
\label{sec:correlatedSys}
Direct effects which mimic an EDM arise from a change of the magnetic field $\delta B(E)$ caused by the electric field. The strength of a false EDM signal can be approximated as in the previous section with $d_{\rm n, f} = \mu_{\rm n}/(2E)\left(\delta B (E^{\parallel}) -\delta B (E^{\nparallel})\right)$
\subsubsection{Leakage currents}
Probably the most obvious direct effect is a leakage current from the charged electrode to ground along a given path. Such a path might result from material imperfections of the insulator ring or from discharges along the wall ( which can lead to visible traces, as reported by several experiments). A current flowing along this path will create a magnetic field which will be seen by neutrons. By reversing the high voltage the current will flow in the opposite direction and create a field with opposite sign, hence $\left(\delta B (E^{\parallel}) -\delta B (E^{\nparallel})\right)= 2\delta B (E^{\parallel})$ if the current is of the same magnitude and follows the same path. Figure\,\ref{fig:LeakageCurrentSketch} shows the idealized geometry of the scenario discussed. Several different current paths were studied in Ref.\,\cite{Zenner2013} using finite elements. The most stringent limit is given by a path which makes a full circle around the insulator ring at the central plane. In this case the neutrons are exposed to a change of the magnetic field of $|\delta B (E^{\parallel})|=\SI{2}{\femto\tesla}$ for a current of \SI{1}{\nano\ampere}, which results in a false nEDM signal of $d_{\rm n,f} \approx \unit[\pow{1}{-26}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$. However, such a current path seems rather unlikely for which reason it became commonly accepted to estimate the effect by a current path as depicted in Fig.\,\ref{fig:LeakageCurrentSketch} with an azimuthal path length of $\diff s = \SI{0.1}{m}$, which means the discharge path is running at up to about 45 degrees. In this case the false nEDM signal would only be $\unit[\pow{5}{-28}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$.
In general most spectrometer designs include a leakage-current monitor which records the integral current flowing across the insulator ring. Additional magnetometers, especially co-magnetometers, will also be sensitive to leakage-current induced field changes and may be used for correction. However, it has been reported that the mean value of scalar magnetometers placed in the close vicinity of the precession chamber might indicate no change in the magnetic field depending on their exact placement, despite the presence of a leakage current large enough to lead to a false nEDM signal. For a detailed discussion of this effect I refer the interested reader to Ref.\,\cite{Lamoreaux2009}.
\begin{figure}%
\includegraphics[width=0.7\columnwidth]{LeakageCurrentSegment}%
\caption{Idealized sketch of the path a leakage current may flow along. The currents flowing along the electrode in the radial direction $j_r$ produce the same magnetic field, in opposite direction, which cancel to first order. Further, the field generated by the fractional current along $z$ parallel to the cylinder axis $j_z$only generates a field in the transverse plane, which is strongly suppressed since the neutrons only see the magnitude of the field, which is dominated by $B_0$.} %
\label{fig:LeakageCurrentSketch}%
\end{figure}
\subsubsection{Systematic effects from electric fields}
Any particle moving through an electric field will see a motional magnetic field seen in its rest frame. According to special relativity the neutron is then exposed to an additional field of:
\begin{equation}
B_{v\times E} = -\gamma\frac{\boldsymbol{v}\times\boldsymbol{E}}{c^2},
\label{eq:vxEEffect}
\end{equation}
\noindent where $\gamma = \left(1-v^2/c^2\right)^{-1/2}$ is approximately 1 for cold and ultracold neutrons. This results in direct systematic effects which lead to a signal mimicking a true nEDM.
For simplicity we consider the case where the electric field and the magnetic field are slightly tilted by an angle $\theta$ with respect to each other in the $x$-$z$-plane ($\boldsymbol{E} = \sin\!\theta\,E, 0 ,\cos\!\theta\,E$) and where a uniform magnetic field gradient of the form $\partial B_z/\partial z = -1/2(\partial B_x/\partial x +\partial B_y/\partial y)$ is present. In this case with $\cos\theta\approx 1$ the magnetic field in the rest frame of the particle moving through the cell at a point in space is
\begin{align}
B_x =& \quad -\frac{x}{2}\frac{\partial B_z}{\partial z} -\frac{v_y E}{c^2} \nonumber \\
B_y =& \quad -\frac{y}{2}\frac{\partial B_z}{\partial z} -\frac{v_x E}{c^2}+\frac{\theta v_zE}{c^2} \\
B_z =& B_0 + z\frac{\partial B_z}{\partial z}+\frac{\theta v_yE}{c^2}. \nonumber
\label{eq:GeomPhase1}
\end{align}
\noindent This leads to a field magnitude of
\begin{align}
|B(\boldsymbol{r})| = B_0 &+ z\frac{\partial B_z}{\partial z}+\frac{x^2+y^2}{8B_0}\left(\frac{\partial B_z}{\partial z}\right)^2 \\
&+\frac{\theta v_yE}{c^2} \label{eq:linearvCrossETerm}\\
&+\frac{\left(xv_y+yv_x+\theta y v_z\right)}{2B_0c^2}\frac{\partial B_z}{\partial z}E \label{eq:GeomTerm}\\
&+\frac{v_y^2+\left(v_x-\theta v_z\right)^2}{2B_0c^4}E^2.
\label{eq:quadraticTermInE}
\end{align}
While the first line (equation 36) is free of any electric-field systematic the following three lines (eqs. 37-39) either have terms linear or quadratic in $E$. The effective magnitude of these terms depends on how the particle samples the precession cell and is discussed in the following paragraphs.
\medskip
\paragraph*{Motional magnetic field in the rest frame of a particle moving through an electric field}
\smallskip
~\\
The latest beam experiment\,\cite{Dress1977} was limited by the motional magnetic field, the linear term in $E$ without gradient in equation\,\eqref{eq:linearvCrossETerm}, seen in the rest frame of a neutron moving through an electric field.
For the configuration of this last beam experiment, with $\boldsymbol{v}= v_y = \unit[150]{m/s}$ and a field strength of \unit[100]{kV/cm}, the term $B_E = v_y E/c^2 =\SI{17}{nT}$.
In the case where the electric field and the magnetic field have a relative angle of $\theta$ the neutrons see a field change of $\Delta B(E) = 2\theta\cdot B_{E}$ when changing polarity of the electric field. The reported error of $\unit[\pow{1.5}{-24}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ is equivalent to a misalignment angle of \SI{0.1}{\milli\radian}, which is extremely demanding. A proposal for a future beam experiment\,\cite{Piegsa2013PRC} intends to circumvent this issue by measuring the nEDM as a function of velocity and then extrapolating to $\boldsymbol{v} = 0$.
Storage experiments using UCN are much less exposed to this effect, as the velocities are a factor 40 lower and in the limit of chaotic trajectories the mean value averages to zero. However, a possible effect which could lead to a relevant signal from $\boldsymbol{v}\times \boldsymbol{E}$ is a remaining ordered motion within the precession chamber.
Such a remaining common rotational motion could result from UCN which enter the precession chamber through an off-axis door or with a predominant tangential velocity component (twist). This could eventually lead to the situation that a fraction of the neutrons move on orbits along the insulator ring resulting in a net rotational flow with velocity $\boldsymbol{v}_{\rm t}$. In this case one can replace $v_y$ with $v_r$ in equation\,\eqref{eq:linearvCrossETerm}, hence, $B_{\rm E}^{\rm rot} = \theta\frac{v_r\cdot E}{c^2}$ would be the result. By increasing the roughness of the precession chamber walls it is possible to reduce the probability of specular reflection which in turn would reduce the probability of ordered motion. Simulations\,\cite{Zenner2013} using Monte Carlo methods for particle trajectories show that an average $v_r < \unit[0.005]{m/s}$ can be obtained for a typical surface roughness. Together with a mechanically feasible alignment of the electric field relative to the magnetic field of better than $\theta \approx
\SI{1}{\milli\radian}$ a remaining systematic effect of $d_{\rm n,f}\leq \unit[\pow{1.7}{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ seems likely.
The detailed analysis of this effect by Pendlebury and coworkers\,\cite{Pendlebury2015PRD} give a limit of
$\unit[\pow{1}{-29}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ arguing that by the time the first $\pi/2$-pulse is applied, tipping the neutrons into the transverse plane, the velocities of the stored UCN are sufficiently randomized.
\medskip
\paragraph*{Motional magnetic field from the quadratic term of $\boldsymbol{v}\times E$}
\smallskip ~\\
Another effect correlated to the motion of particles in an electric field is the quadratic term\,\eqref{eq:quadraticTermInE}, which persists even if $\overline{\boldsymbol{v}} =0$ is fulfilled.
Simply taking the values for the mercury co-magnetometer\,\cite{Baker2014}, $B_0 = \SI{1}{\mbox{\micro T}}$, $v_{\rm Hg}^2 = \left(\SI{150}{m/s}\right)^2$ and $E =\SI{1}{kV/cm}$ results in $ B_{\rm E^2} = \SI{1.4}{pT}$.
This effect is transferred to the neutrons when using the mercury frequency as field monitor. Only a difference in magnitude of the E-field for both polarities will have a remaining contribution to a false nEDM signal. To keep this systematic effect below the current statistical sensitivity the electric-field magnitude would have to be equal to within one part in 10\,000. However, this estimate does not take into account the fact that the direction and velocity of the mercury atoms change with every collision. Hence, the atoms are effectively exposed to a fast and random oscillating motional magnetic field, and the field only has a constant magnitude for $\tau_c$, the average time between two collisions.
The effective magnetic-field shift for this situation is derived as a frequency shift for spin-1/2 particles in Ref.\,\cite{Lamoreaux2009} and can be written as
\begin{equation}
B_{\rm E^2} = \frac{1}{2\gamma}\left\langle \left(B_{\rm v\times E}\right)^2 \frac{\gamma B_0\tau_c - \sin\left( \gamma B_0 \tau_c\right)}{\left(\gamma B_0\right)^2 \tau_c} \right\rangle_{\rm av}.
\label{eq:vCrossECubicShift}
\end{equation}
\noindent The index `av' indicates an average over the statistical ensemble and for times much longer than $\tau_c$. In a nEDM experiment with a co-magnetometer two extreme cases exist. For $\gamma B_0\tau_c \gg 1$ the term $\sin\left( \gamma B_0 \tau_c\right)$ has a zero ensemble average. Further the average over all possible direction for $\boldsymbol{v}$ gives a factor 2/3. Therefore the quadratic motional magnetic field is
\begin{equation}
B_{\rm E^2} = 2/3\frac{v^2 E^2}{2B_0c^4}.
\label{eq:vCrossECubicUCN}
\end{equation}
\noindent In measurements using UCN we are in this regime. For an average velocity of $v = \SI{3}{m/s}$, an electric field of \SI{12}{kV/cm}, $B_0=\SI{1}{\micro\tesla}$ we have a motional field of $B_{\rm E^2}^{\rm UCN} =\SI{0.5}{fT}$. This is completely negligible as the effect only shows up if the magnitude for positive and negative electric fields are different. Modern high voltage power supplies have specifications of better than $|U^+|-|U^-| < 0.001 U$ with $U$ the nominal voltage, which further suppresses this systematic effect by a factor thousand, resulting in a limit of $d_{\rm n,E^2} < \unit[\pow{3}{-30}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$.
In the case when $\gamma B_0\tau_c \ll 1$ the sine-term in equation\,\eqref{eq:vCrossECubicShift} can be expanded:
\begin{equation}
B_{\rm E^2}=2/3\frac{v^2 E^2}{2B_0c^4} \frac{\left(\gamma B_0 \tau_c\right)^2}{6},
\label{eq:vCrossECubicHg}
\end{equation}
\noindent with $\tau_c \approx \unit[\pow{1}{-3}]{s}$ for a precession cell of radius $r=\SI{0.25}{m}$ and height $h=\SI{0.12}{m}$. Mercury atoms in the ballistic limit with $\gamma = \gamma_{\rm Hg}$ match this case very well. For the above parameters this gives a motional magnetic field of $B_{\rm E^2}^{\rm Hg} \leq \SI{5}{\atto\tesla}$ due to the large suppression factor of $\pow{1}{-4}$ of the last quadratic term.
\medskip
\paragraph*{Frequency shifts of particles moving in an inhomogenous magnetic field and an electric field} \smallskip
~\\
Currently the most important correlated systematic effect results from the term in equation\,\eqref{eq:GeomTerm} linear in gradient and electric field. It is of the same order of magnitude as the statistical sensitivity\,\cite{Pignol2012PRA}.
It gives rise to a frequency shift which depends on the way that the particle is sampling the field of the precession chamber. This effect was first discussed in the limit of a trumped shaped field with a vertical gradient $\partial B_z/\partial z = -1/2(\partial B_x/\partial x +\partial B_y/\partial y $ and an electric field\,\cite{Pendlebury2004}. An approach using the general theory of relaxation with correlation functions was then used in Ref.\,\cite{Lamoreaux2005,Barabanov2006} but still applied to a cylindrical uniform field gradient. In the article of Pignol and Roccia\,\cite{Pignol2012PRA} a general form for arbitrary fields and geometries is given:
\begin{equation}
\delta \omega = -\frac{\gamma^2 E}{c}\left\langle xB_x + yB_y\right\rangle,
\label{eq:GenGeomShift}
\end{equation}
\noindent where the brackets refer to a volume average over the cell. For the further discussion we will use the expressions from Ref.\,\cite{Afach2015EPJD} which has confirmed the theoretical predictions in a direct measurement with an array of scalar OPM around the central precession cell of a mercury magnetometer. A UCN based experiment using a thermal co-magnetometer works in the regime of two limiting cases:
\begin{align}
\delta \omega= \frac{\gamma D^2}{16c^2}\frac{\partial B_z}{\partial z}E \quad\quad & \text{non-adiabatic} \\
\delta \omega = \frac{v_{xy}^2E}{2B_0c^2}\frac{\partial B_z}{\partial z} \quad\quad & \text{adiabatic}.
\label{eq:GeomPhase3}
\end{align}
\noindent It is worthwhile to note that the adiabatic case is independent of the gyromagnetic ratio of the particle and is known as geometric phase or Berry phase\,\cite{Berry1984,Commins1991}. Measurements using UCN fall into this regime. Using the relation $\ensuremath{d_\text{n}} = \hbar/2E\cdot\delta\omega$ the shift can be misinterpreted as an EDM of
\begin{equation}
\ensuremath{d_\text{n}}^{\rm false} = \frac{\partial B_z}{\partial z}\cdot \unit[\pow{1.7}{-28}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}} \frac{cm}{pT}},
\label{eq:neutronGeomEffect}
\end{equation}
\noindent for $v_xy = \SI{3}{m/s}$ and $B_0 = \SI{1}{\micro\tesla}$. For current experimental sensitivities this is not a problem, however, in future experiments aiming at sensitivities in the low $\unit[\pow{1}{-28}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ region a perfect control of the magnetic field homogeneity is mandatory. In the case of \ensuremath{{}^{199}\text{Hg}}\ the atoms sense the field in the non-adiabatic limit, and hence the false edm signal is
\begin{equation}
d_{\rm Hg}^{\rm false} = \frac{\partial B_z}{\partial z}\cdot \unit[\pow{1.15}{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}} \frac{cm}{pT}},
\label{eq:falseEDMHg}
\end{equation}
\noindent which is transferred to the neutron measurement:
\begin{equation}
d_{\rm Hg\rightarrow n}^{\rm false} =-\frac{\partial B_z}{\partial z} \cdot \unit[\pow{4.4}{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}} \frac{cm}{pT}}.
\label{eq:falseNeutronEDMHg}
\end{equation}
Already gradients of some \mbox{pT}/cm will be enough to fully dominate an otherwise very sensitive search for an nEDM\@. Ideally one would like to have a magnetometer scheme which is capable of measuring the gradient with a precision of at sub \mbox{pT}/cm-resolution to correct for this effect. The current version of the cesium OPM array in the PSI experiment achieves an accuracy of approximately \unit[5-10]{\mbox{pT}/cm} which is not sufficient for this purpose. An alternative approach is to bring the cohabiting magnetometer also into the adiabatic regime by using a sufficiently high pressure of buffer gas\,\cite{Masuda2012PLA}.
Currently the transferred effect can be mitigated by measuring \ensuremath{d_\text{n}}\ as a function of
\begin{equation}
R = \frac{\omega_{\rm n}}{\omega_{\rm Hg}}=\frac{\gamma_{\rm n}}{\gamma_{\rm Hg}}\left(1+\frac{g_z\!\cdot\!\Delta h_{\rm cm}}{B_0}+ \delta_{\rm T} + \delta_{\rm ER} + \delta_{\rm dip}+ \delta_{\rm LS}\right),
\label{eq:R}
\end{equation}
\noindent where $\Delta h_{\rm cm}$ is the center-of-mass offset between the UCN and the mercury gas. This is one example of the subtle differences in how UCN and thermal atoms sample the precession volume which lead to tiny magnetic-field dependent deviation of the frequencies of the neutron and the co-habiting isotope (for \ensuremath{{}^{199}\text{Hg}}{} see Ref.\,\cite{Afach2014PLB}). Neglecting for the moment potential systematic effects, indicated with $\delta_{\dots}$ in equation\,\eqref{eq:R}, one can redefine a value
\begin{equation}
R'-1 = R\frac{\gamma_{\rm Hg}}{\gamma_{\rm n}} -1 = \frac{g_z\!\cdot\!\Delta h_{\rm cm}}{B_0},
\label{eq:RPrimeMinusOne}
\end{equation}
which can be used as measurement of the vertical magnetic-field gradient $g_z$. The true value of the measured nEDM is the crossing point of two curves $\ensuremath{d_\text{n}}(R'-1)$ vs $(R'-1)$ shown in Fig.\,\ref{fig:GeomSysEffects} found by reversing $B_0$. However, a whole class of systematic effects appear, as $R'-1$ is not only sensitive to the vertical magnetic-field gradient, but also to all other effects which might shift the ratio of precession frequencies $R$.
The systematic shift from transverse fields is a result from the fact that the co-magnetometer operates in the non-adiabatic regime and essentially measures only the vectorial average of the magnetic field $|\langle \boldsymbol{B} \rangle|$. Whereas the neutrons operate in the adiabatic regime and measure the volume averaged field modulus $\langle |\boldsymbol{B}|\rangle$. Hence, in the presence of a residual transverse field like a quadrupole field perpendicular to $B_z$ with $B_x= qy$, $B_y= qx$ a frequency shift of
\begin{equation}
\delta_T = \frac{\langle B_T^2\rangle}{2B_0^2}
\label{eq:transveresFieldSys}
\end{equation}
\noindent may occur, where $\langle B_T^2\rangle$ is the volume average over the transverse magnetic-field modulus squared. The systematic false effect from transverse fields is proportional to $\Delta B_{\rm T}^2 = \langle B_{\rm T}^{\uparrow 2}\rangle -\langle B_{\rm T}^{\uparrow 2}\rangle$. A $\Delta B_{\rm T}^2 = \SI{1}{nT^2}$ shifts the crossing point EDM value by approximately $\unit[\pow{0.3}{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$. It is possible to measure the transverse fields inside the precession chamber using a three-axis fluxgate to a precision of better than \SI{1}{nT}\,\cite{Afach2014PLB} which seems to be sufficient for current searches. In the future similar magnetic field maps recorded with a three-axis OPM\,\cite{Afach2015OExpress} might achieve a much better precision to permit searches in the low $\unit[\pow{1}{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ range.
The systematic effect $\delta_{\rm ER}$ is a shift due to Earth rotation which only appears if the gyromagnetic ratio of the neutron and the co-magnetometer isotope have different relative signs. The effective frequency shift depends on the latitude $\lambda$ of the location of the experiment on earth:
\begin{equation}
\delta_{\rm ER}^{\uparrow/\downarrow} = -{\rm sgn}(B_0)\left(\frac{f_{\rm E}}{f_{\rm n}}+\frac{f_{\rm E}}{f_{\rm Hg}}\right)\sin\left(\lambda\right).
\label{eq:EarthRotation}
\end{equation}
Where the Earth's rotation frequency $f_{\rm E}=\SI{11.6}{\micro\hertz}$.
For the experiment at the Paul Scherrer Institute ($\lambda \approx \SI{46.6}{\degree}$ north) this results in a shift of $\delta_{\rm ER}^{\uparrow/\downarrow} =\mp \unit[\pow{1.4}{-6}]{Hz}$.
Which, when forgotten, changes the crossing point to $d_{\times} = \unit[\pow{1.76}{-25}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$.
The next possible contribution comes from small magnetic permanent dipole fields close to the cell's surface. This will locally cause an inhomogeneous $\partial B_z /\partial z$ gradient which adds to $d_{\rm Hg\rightarrow n}^{\rm false}$ of equation\,\eqref{eq:falseNeutronEDMHg}. Further it will induce a frequency shift $\delta_{\rm dip}$ moving the lines horizontally by an equal amount but in opposite directions. Both the exact magnitude and orientation as well as position of a permanent dipole in the cell container influence the strength of the systematic effect. A possible method to avoid such an effect is to thoroughly scan all parts which constitute the precession cell assembly. Pendlebury and coworkers\,\cite{Pendlebury2015PRD} allowed for a systematic effect in their analysis of $\unit[\pow{\pm 6}{-27}]{\ensuremath{\si{\elementarycharge}\!\cdot\!\ensuremath{\mathrm{cm}}}}$ for an undetected permanent dipole moment in the mercury door.
\begin{figure}%
\includegraphics[width=0.95\columnwidth]{GeomSysEffects.png}%
\caption{Neutron EDM as function of $R'$ for mercury as co-magnetometer, indicating systematic effects which may occur when using the crossing point analysis as in Ref.\,\cite{Pendlebury2015PRD}. The solid orange ($B_0^{\uparrow}$) and blue ($B_0^{\downarrow}$) line indicate the perfect systematic free measurement. Some possible shifts are indicated individually. Earth rotation shift $\delta_{\rm ER}^{\uparrow/\downarrow}$ for main magnetic field up/down; shift from a permanent dipole close to the precession chamber $\delta_{\rm dip}$; shifts from transverse magnetic fields of different magnitude $\delta_{\rm T}$. The black arrows pointing to the $y$-scale indicate the magnitude of the systematic effect. Note that, if the transverse fields for both magnetic field configurations is identical, $B_{\rm T}^{\uparrow} = B_{\rm T}^{\downarrow}$, the systematic effect vanishes.}%
\label{fig:GeomSysEffects}%
\end{figure}
\subsection{LANL - Los Alamos, United States of America}
|
\section{Introduction}
\label{sec:introduction}
Supersymmetry (SUSY) provides a most promising solution to the hierarchy problem to the standard model (SM). The quadratically divergent contributions to the Higgs mass-squared parameter from couplings to the SM fields are canceled by those of their superpartners. To keep the scale of the electroweak symmetry breaking natural, the superpartners are expected to have masses around or beneath the TeV scale. In particular, since the largest coupling to the Higgs in the SM is from the top quark, the superpartners of the top quark (top squarks or stops) play the most important role in canceling the quadratic divergence and are expected to be close to the weak scale in a natural theory.
On the other hand, the Higgs boson mass of 125 GeV also has important implications for the stop masses. In the minimal supersymmetric standard model (MSSM), the tree-level Higgs boson mass has an upper bound of $M_Z$. To get to 125 GeV, it requires large radiative contributions from the stop loops~\cite{Haber:1990aw,Okada:1990vk,Ellis:1990nz}. This could happen if the stops are heavy and/or the trilinear $A_t$ term of the stop sector is large~\cite{Hall:2011aa,Arbey:2011ab,Draper:2011aa,Carena:2011aa,Akula:2011aa}. To keep the fine-tuning minimal, it is preferable to have a large $A_t$ term so that the stops masses can be lowered to $\sim 1$ TeV or below. A large $A_t$ term implies large off-diagonal masses of the stop mass matrix so there will be a substantial mixing between the left-handed and the right-handed stops in the mass eigenstates. As a consequence, there will also be a sizable mass difference between the two stop mass eigenstates.
As a key to the naturalness problem, the stops have been extensively searched for at the LHC. Assuming that the lightest neutralino $\tilde{\chi}^0$ is the lightest supersymmetric particle (LSP) and is stable, the search limit for the $\tilde{t} \to t \tilde{\chi}^0$ decay (assuming 100\% branch fraction) has reached $\sim 860$ GeV for $m_{\tilde{\chi}^0} \lesssim 250$~GeV at the current Run 2 of LHC~\cite{ATLAS:2016jaa,ATLAS:2016ljb,CMS:2016xva,CMS:2016vew,CMS:2016hxa,CMS:2016inz}.
From the naturalness point view, some fine-tuning is already required if the lightest stop is heavier than 860 GeV. However, the search limits are significantly weakened in the compressed region, where $m_{\tilde{t}_1}- m_{\tilde{\chi}^0} \lesssim m_t$. In particular, there is a gap along $m_{\tilde{t}_1}- m_{\tilde{\chi}^0} \approx m_t$ in the $m_{\tilde{t}_1}- m_{\tilde{\chi}^0}$ plane. In this case the top quark and the neutralino from the stop decay are roughly static in the stop rest frame. For the stop pair production, the neutralinos travel along with the same velocities as the original stops and their momenta tend to cancel each other out, leaving little missing transverse energy (MET) in the signal. Consequently, it is difficult to be distinguished from the SM top pair production background and it is still possible to have a relatively light $\tilde{t}_1$. There have been studies trying to identify useful variables to probe this compressed region but the reach is limited~\cite{Alves:2012ft,Han:2012fw,Kilic:2012kw,Dutta:2013gga}. A more promising strategy is to consider the stop pair production with a hard initial state radiation (ISR) jet, then the neutralinos are boosted in the opposite direction to the ISR jet, giving rise to some MET. It may have a discovery reach up to $m_{\tilde{t}_1}\sim 400-500$~GeV at LHC 13 TeV with 300 fb$^{-1}$~\cite{Hagiwara:2013tva,An:2015uwa,Macaluso:2015wja,Cheng:2016mcw}.
Since naturalness needs both stops to be not too heavy, if $\tilde{t}_1$ happens to lie in the compressed region, it may be easier to search for $\tilde{t}_2$ even though it is heavier. Indeed,
both ATLAS and CMS have performed searches for the heavier stop for $m_{\tilde{t}_1} \approx m_t + m_{\tilde{\chi}_1^0}$ so that $\tilde{t}_1$ escapes the detection~\cite{Aad:2015pfx,Aad:2014mha,Khachatryan:2014doa,ATLAS:2016tpc}. These searches adopted the simplified model approach, assuming that the heavier stop decays to the lighter stop plus a $Z$ or a Higgs boson ($ \tilde{t}_2 \to \tilde{t}_1 + Z/h$) with a 100\% branching fraction. The exclusion limit for the $\tilde{t}_2$ mass goes up to $\sim 730$ GeV for the $\tilde{t}_2 \to Z \tilde{t}_1$ decay mode with 13 TeV Run 2~\cite{ATLAS:2016tpc} and about 600 GeV for the $\tilde{t}_2 \to h \tilde{t}_1$ decay mode with the 8 TeV Run 1 data~\cite{Aad:2015pfx} (the corresponding Run 2 analysis has not appeared yet).
The simplified model approach is a good strategy if there is a dominant decay channel or the search limit is dictated by a certain decay channel. One can easily recast the search result to a wide range of models which have similar decay processes and final states, by rescaling the cross sections and branching ratios.
However, it is seldom a good approximation for the system of the two stops. Because the left-handed stop and sbottom belong to an $SU(2)_W$ doublet, there is always a sbottom with mass within the vicinity of the two stops. The presence of the sbottom will give additional decay modes of $\tilde{t}_2$. In addition, there could be other charginos and neutralinos lighter than $\tilde{t}_2$ besides $\tilde{\chi}^0$. If this is the case, there will also be decay chains going through them as intermediate states. As a result, the $\tilde{t}_2$ decays often have many decay channels without a dominant one~\cite{Eckel:2014wta,Guo:2013iij}. Different decay channels have different final states and hence require different signal selection criteria. It makes the $\tilde{t}_2$ search in the compressed region in a realistic scenario more complicated than simply rescaling the simplified model analysis.
An alternative approach to the simplified model is the pMSSM~\cite{Djouadi:2002ze,Berger:2008cq,CahillRowley:2012cb}, which parametrizes MSSM with some modest assumptions. The assumptions include $R$ parity conservation with the lightest neutralino being the LSP, minimal flavor violation at the TeV scale with no CP violation in the SUSY sector, and degenerate sfermion masses for the first two generations. It contains 19 phenomenological parameters and a scan over these parameters generates a large model samples for phenomenological studies.
However, if one is interested in the stop system, scanning over the full 19-parameter space may be an overkill.\footnote{A light stop study in the pMSSM approach can be found in Ref.~\cite{Belanger:2015vwa}.} To study the possible interesting decay patterns of the stop and sbottom system and their experimental signals one should focus on the most relevant parameters. This is the approach taken in this paper. We divide the models into two scenarios. In the first scenario all neutralinos and charginos except the LSP are heavier than $\tilde{t}_2$ so they decouple. The only relevant particles are $\tilde{t}_1$, $\tilde{t}_2$, $\tilde{b}_1$, and $\tilde{\chi}_1^0$, whose masses and interactions are only governed by a few parameters. We scan through them and find model points with different characteristic decay patterns. We identify categories of signal channels which may be sensitive to various final states of the decay chains and study the signal significance over the backgrounds. The real search reach may require division and combination of many different channels. In the second scenario we consider additional charginos and neutralinos below the mass of $\tilde{t}_2$, which can introduce even more complicated decay patterns. We perform the similar study as in the first scenario for the model points where the additional charginos and neutralinos play important roles in the decay chains.
This paper is organized as follows. In Sec.~\ref{sec:MSSM}, we discuss the spectrum of the third generation squarks in MSSM based on the naturalness and the Higgs boson mass. We also summarize the current experimental constraints. In Sec.~\ref{sec:benchmark}, we consider some benchmark points for the stop and sbottom spectrum where $\tilde{t}_1$ is hidden in the compressed region. The benchmark models are divided into two types, depending on whether there are additional neutralinos and charginos which can appear in the decay chains of $\tilde{t}_2$ and $\tilde{b}_1$, We present the branching ratios of various decay chains and the fractions of final states of these benchmark points. In Sec.~\ref{sec:collider}, we discuss categories of signals for the $\tilde{t}_2$ and $\tilde{b}_1$ searches when $\tilde{t}_1$ is hidden. We perform simplified collider studies for the benchmark models to explore the future sensitivities at the 14 TeV LHC. The conclusions are drawn in Sec.~\ref{sec:conclusions}.
The compatibilities of our benchmark points with the current experimental constraints are examined in the Appendix.
\section{Stop Masses in MSSM}
\label{sec:MSSM}
\subsection{Argument of Naturalness}
In MSSM, after minimizing the Higgs potential, the $Z$ boson mass is given by~\cite{Martin:1997ns}
\begin{equation}
\frac{m_z^2}{2}=\frac{m_{H_d}^2+\delta_{H_d}^2-\tan\beta^2(m_{H_u}^2+\delta_{H_u}^2)}{\tan\beta^2-1}-\mu^2,
\label{eq:mz}
\end{equation}
where $m_{H_u}^2$ and $m_{H_d}^2$ are the soft SUSY breaking masses of the $H_u$ and $H_d$ doublets at an UV cutoff scale $\Lambda$, $\delta_{H_u}$ and $\delta_{H_d}$ represent the radiative contributions to the soft SUSY breaking masses below the cutoff, and $\mu$ is the SUSY-preserving Higgs mass parameter which is also the approximate Higgsino mass. To avoid fine-tuning on the $Z$ mass, there should not be a large cancellation among various terms in the above equation. In particular, the radiative correction $\delta_{H_u}$ receives the largest contribution from the stop loops:
\begin{equation}
\delta_{H_u}^2 = - \frac{3y_t^2}{8\pi^2}(m_{Q_3}^2+m_{u_3}^2+|A_t|^2) \ln \left(\frac{\Lambda}{\mu_{\text{IR}}}\right)
\end{equation}
where $m_{Q_3}, m_{u_3}$ are soft-breaking mass terms for the left-handed top-bottom doublet and the right-handed top squarks, $A_t$ is the trilinear soft-SUSY breaking of the corresponding Yukawa interaction, and $\mu_{\text{IR}}$ is taken to be the geometric average of the stop masses $m_{\tilde{t}} = \sqrt{m_{\tilde{t}_1} m_{\tilde{t}_2}}$. The tuning of $m_Z^2$ due to the stop mass contribution
$
\frac{m_Z^2/2}{|\delta_{H_u}^2|}
$
is already at the level of $\lesssim 1\%$ for $m_{Q_3}, m_{u_3} \sim 1$~TeV and $\Lambda \sim 100$ TeV~\cite{Harnik:2003rs}. Therefore, naturalness argument would prefer both stops to have masses below or around 1 TeV.
\subsection{Higgs boson mass}
At the tree level, MSSM also predicts the light Higgs boson mass to be less than the $Z$ mass:
\begin{equation}
m_{h,\text{tree}}^2 =m_z^2\cos^22\beta .
\end{equation}
This contradicts the observed Higgs mass at 125 GeV. The loop corrections can raise the Higgs boson mass to evade the tree-level upper bound of $m_Z$. However, to reach 125 GeV the loop contributions must be significant. The dominant contribution comes from the stop loop, which implies constraints on the masses of the stop sector. The Higgs boson mass including the leading one-loop stop contribution is given by~\cite{Okada:1990gg,Casas:1994us,Carena:1995bx}
\begin{equation}
m_h^2=m_z^2\cos^22\beta+\frac{3m_t^4}{4\pi^2v^2} \left[ \log\frac{m_{\tilde{t}}^2}{m_t^2} +\frac{X_t^2}{m_{\tilde{t}}^2}
\left(1-\frac{X_t^2}{m_{\tilde{t}}^2}\right)
\right] ,
\end{equation}
where $m_{\tilde{t}} = \sqrt{m_{\tilde{t}_1} m_{\tilde{t}_2}}$ as defined previously and $X_t = A_t - \mu \cot \beta$ is the stop mass mixing parameter. From the formula one can see that without the $X_t$ term, the stop masses need to be raised to very high values in order to generate a Higgs massif 125 GeV. This would be in severe conflict with naturalness. To minimize the fine-tuning, the second term in the bracket should be large and the one-loop correction is maximized when $X_t = \sqrt{6}m_{\tilde{t}}$. Such a large $X_t$ implies a large mixing between the left-handed and right-handed stops, which has interesting phenomenological implications.\footnote{If there are additional contributions to the Higgs quartic coupling, such as in the Next-to-Minimal Supersymmetric Standard Model (NMSSM)~\cite{Ellwanger:2009dp}, it is easier to have a 125 GeV Higgs boson with light stops compatible with naturalness~\cite{Hall:2011aa,Beuria:2015mta}. The stop mixing does not need to be large in that case and the stop spectrum could be more compact. Nevertheless, the results in this paper also applies to a large region of parameter space in NMSSM.}
\subsection{Stop and sbottom masses}
The mass matrices for the stops, $\tilde{t}_L, \tilde{t}_R$, and sbottoms, $\tilde{b}_L\,$ and $\tilde{b}_R$ are given by
\begin{eqnarray}
M_{\tilde{t}}^2 =\left(\begin{array}{cc}
m_{Q_3}^2 + m_t^2 +\Delta_{\tilde{u}L} & m_t X_t \\
m_t X_t & m_{u_3}^2 + m_t^2 +\Delta_{\tilde{u}R}
\end{array} \right),
\end{eqnarray}
\begin{eqnarray}
M_{\tilde{b}}^2 =\left(\begin{array}{cc}
m_{Q_3}^2 + m_b^2 +\Delta_{\tilde{d}L} & m_b X_b \\
m_b X_b & m_{d_3}^2 + m_b^2 +\Delta_{\tilde{d}R}
\end{array} \right)
\end{eqnarray}
where $X_b = A_b - \mu \tan\beta$ is the term related to $\tilde{b}_L$ and $\tilde{b}_R$ mixing from trilinear couplings, and
\begin{equation}
\Delta_{\tilde{u}L}= (\frac{1}{2}-\frac{2}{3} \sin^2\theta_W) \cos 2\beta\, m_Z^2, \quad \Delta_{\tilde{u}R}=-\frac{2}{3}\sin^2\theta_W \cos 2\beta\ m_Z^2,
\end{equation}
\begin{equation}
\Delta_{\tilde{d}L}= (-\frac{1}{2}+\frac{1}{3} \sin^2\theta_W) \cos 2\beta m_Z^2, \quad \Delta_{\tilde{d}R}=\frac{1}{3}\sin^2\theta_W \cos 2\beta\ m_Z^2
\end{equation}
represent the $D$-term contributions.
Notice that in the limit where there is no mixing, both $\tilde{t}_L$ and $\tilde{b}_L$ masses are controlled by the soft breaking mass $m_{Q_3}$ and they are expected to be nearly degenerate if $m_{Q_3} \gg m_t,\, m_Z$, with $m_{\tilde{t}_L} \gtrsim m_{\tilde{b}_L}$. However, with a large $X_t$, there is a significant mixing between $\tilde{t}_L$ and $\tilde{t}_R$ and the mass spectrum of the stops will be modified. The two mass eigenstates are repelled from one another by the mixing term and the mass gap between them further increases. Consequently, there is at least one sbottom lighter than the heavier stop. The spectrum of the stop and sbottom sector has important implications for their decay patterns and collider searches as we will see.
To scan the MSSM parameter space we specify the parameters at the cutoff scale $\Lambda$ which is taken to be 100 TeV. Since we are focusing on the spectrum of the third generation squark (and neutralino/chargino for the decay patterns), we decouple the first two generation of sfermions and third generation sleptons by setting their soft SUSY-breaking masses to 3 TeV. We also set the gluino mass to 2.5 TeV, beyond the current and near future reaches.\footnote{A lighter gluino does not affect the direct stop and sbottom productions. However, it may give the first SUSY signals at the LHC. Its decays through stops and sbottoms will mix into the signals for direct stop and sbottom productions, so we choose a heavier gluino to avoid this complication.}
In order to generate stop masses which are potentially within the LHC Run 2 reach, the diagonal stop soft breaking masses $m_{Q_3}$ and $m_{u_3}$ are varied from 250 GeV to 1.4 TeV. The $X_t$ term is scanned from $-3m_{\tilde{t}}$ to $3m_{\tilde{t}}$, where $m_{\tilde{t}}=\sqrt{m_{Q_3} m_{u_3}} $ as mentioned before. On the other hand, the right-handed sbottom soft-breaking mass $m_{d_3}$ is varied from 100 to 3000 GeV. For the Higgs sector, $\tan\beta$ varies from 2 to 50 and the Higgsino mass parameter $\mu$ varies from 100 GeV to 3 TeV. Although the $SU(2)$ and $U(1)$ gaugino masses $M_2$ and $M_1$ have little effects in the stop/sbottom spectrum or the Higgs boson mass, the presence of the neutralinos and charginos can affect the decay chains of the stop/sbottom. Thus, we let $M_1$ and $M_2$ vary in the ranges of 50--1500 GeV and 250-1500 GeV respectively.
We use FeynHiggs~\cite{Feynhiggs} to generate the SUSY spectrum and to calculate the Higgs boson mass. Given the uncertainties in different approaches in the Higgs mass calculation and higher order corrections, we require the resulting Higgs boson mass to be bigger than 122 GeV as our selection criterion.\footnote{The effective field theory approach generally gives a lower Higgs boson mass~\cite{Vega:2015fna} so we do not impose an upper limit on the Higgs mass.} The masses of the two stop mass eigenstates and the corresponding mixing term $X_t$ which can satisfy the Higgs boson mass requirement are shown in Fig.~\ref{fig:stop_vs_Xt}
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{subfigure}[b]{0.8\textwidth}
\includegraphics[width=\textwidth]{M_Stop_X.pdf}
\end{subfigure}
\caption{The model points that satisfies $m_h>122$GeV presented in $X_t $ vs. $m_{\tilde{t}} $ plane. Each red dot represents the second stop mass in GeV, and the blue one represents the lightest stop. All sample points presented have a proper LSP with a mass $|m_{\tilde{t}_1} - m_{\tilde{\chi}_1} - m_t| \leqslant30$ GeV.\label{fig:stop_vs_Xt}
}
\end{center}
\end{figure}
We can see that the lightest stop mass can be as low as 250 GeV in extreme cases, while the second stop can be as low as around 600 GeV.
When it comes to how the sbottom and the second stop decay, an interesting question is the mass differences between the stops and the sbottom. In Fig.~\ref{fig:stop1_stop2} we show $m_{\tilde{t}_1}$ vs. $m_{\tilde{t}_2}$ for allowed points. There is always a significant gap between the two stop masses due to the large mass mixing term. Most points have a mass difference greater than 300 GeV, which means that $\tilde{t}_2 \to \tilde{t}_1 + Z(h)$ decays are always open. Fig.~\ref{fig:stop_sbottom} shows the mass difference between $\tilde{b}_1$ and $\tilde{t}_1$ vs. the mass difference between $\tilde{t}_2$ and $\tilde{b}_1$. We see that $\tilde{b}_1$ is always lighter than $\tilde{t}_2$. There are also points with $\tilde{b}_1$ lighter than $\tilde{t}_1$ but for these points the $\tilde{b}_1$ search will provide the strongest constraint~\cite{AdeelAjaib:2011ec,Alvarez:2012wf,Lee:2012sy,Bi:2012jv,Dutta:2015hra,Han:2015tua,Beuria:2016mur,CMS:2016xva}. For all points we have either $m_{\tilde{b}_1} -m_{\tilde{t}_1} > m_W$ or $m_{\tilde{t}_2} -m_{\tilde{b}_1} > m_W$, and there is also a significant fraction of points where both inequalities hold. For $m_{\tilde{t}_2} -m_{\tilde{b}_1} > m_W$, the decay channel $\tilde{t}_2 \to \tilde{b}_1 +W$ is open, which has not been considered in current $\tilde{t}_2$ searches. Similarly $\tilde{b}_1 \to \tilde{t}_1 +W$ will be open if $m_{\tilde{b}_1} -m_{\tilde{t}_1} > m_W$~\cite{Li:2010zv,Datta:2011ef}. These decay channels should be included in searches for $\tilde{t}_2$ and $\tilde{b}_1$ since they occur naturally in MSSM. There will be even more possible decay channels if some charginos and neutralinos have masses between these stop and sbottom states. We will perform some benchmark point studies in the rest of the paper to point out the final states and channels that are relevant for the second stop and sbottom searches.
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{subfigure}[b]{0.6\textwidth}
\includegraphics[width=\textwidth]{Mt12.pdf}
\end{subfigure}
\caption{The mass and mass difference between the second and first stops. Two diagonal dashed lines represent $m_{\tilde{t}2}-m_{\tilde{t}1}=0,300$ GeV. Colored points are the benchmark models.
\label{fig:stop1_stop2}}
\end{center}
\end{figure}
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\includegraphics[scale=0.6]{Mt12_b1_detail.pdf}
\caption{A detailed look at $m_{\tilde{t}_2}-m_{\tilde{b}_1}$ vs. $m_{\tilde{b}_1}-m_{\tilde{t}_1}$ plane, including the right hierarchy only. The vertical dashed line represents the mass relation $m_{\tilde{t}_2}-m_{\tilde{b}_1} = m_W$. The on-shell charged current decay is kinematically forbidden to the left of this line. The horizontal lines are $m_{\tilde{b}_1}-m_{\tilde{t}_1}= m_W$. Colored points are the benchmark models.\label{fig:stop_sbottom}}
\end{figure}
\subsection{Current experimental bounds}
The third generation squarks are extensively searched at the LHC experiments. Here we give a brief summary of the experimental constraints most relevant to our discussion.
For $\tilde{t}_1$ decaying 100\% to $t \tilde{\chi}_1^0$, 13 TeV Run 2 results based on the integrated luminosity $\sim 13$ fb$^{-1}$ exclude $\tilde{t}_1$ mass up to 860 GeV for a light ($\lesssim 250$ GeV) $\tilde{\chi}_1^0$~\cite{CMS:2016hxa,CMS:2016vew,ATLAS:2016jaa,ATLAS:2016ljb}.
However, this does not apply to the compressed region when the mass difference between $\tilde{t}_1$ and $\tilde{\chi}_1^0$ is small. For $m_{\tilde{t}_1} -m_{\tilde{\chi}_1^0} \approx m_t$, while there was no bound from Run 1, with Run 2 data ATLAS has obtained a limit on $m_{\tilde{t}_1}$ of 380 GeV~\cite{ATLAS:2016jaa} using the variable proposed in Ref.~\cite{An:2015uwa}. For even smaller mass difference, there are some constraints from several search modes $\tilde{t}_1 \to Wb \tilde{\chi}_1^0$, $c\tilde{\chi}_1^0$, or $b f f' \tilde{\chi}_1^0$ depending on the mass difference~\cite{Hikasa:1987db,Muhlleitner:2011ww,Boehm:1999tr,Das:2001kd}. The bounds on $m_{\tilde{t}_1}$ in these cases are around 300 GeV from the Run 1 data~\cite{Aad:2015pfx,Khachatryan:2016pup,Aad:2014kra,Aad:2014nra,Chatrchyan:2013xna,Khachatryan:2015pot,Khachatryan:2015wza,Khachatryan:2016pxa}. The most recent Run 2 analyses exclude a stop mass up to 365 GeV for $m_{\tilde{t}_1} -m_{\tilde{\chi}_1^0}= 90$~GeV on the 3-body decay mode $\tilde{t}_1 \to Wb \tilde{\chi}_1^0$~\cite{ATLAS:2016xcm}, and $\sim 450 (360)$ GeV for the 4-body decays $b f f' \tilde{\chi}_1^0$ in the fully hadronic final state (opposite-sign leptons) search~\cite{CMS:2016hxa,CMS:2016zvj}. For $\tilde{b}_1$ decaying 100\% to $b+\tilde{\chi}_1^0$, The 13 fb$^{-1}$ Run 2 results exclude $m_{\tilde{b}_1} $ up to $\sim 1$ TeV for $m_{\tilde{\chi}_1^0}$ up to $\sim 400$ GeV~\cite{CMS:2016xva}.
For $\tilde{b}_1$ decaying to $t+\tilde{\chi}_1^-$ and then $\tilde{\chi}_1^- \to W^- + \tilde{\chi}_1^0$, the 13 fb$^{-1}$ Run 2 analysis reached a mass limit of 690 GeV for a light $\tilde{\chi}_1^0$, assuming $m_{\tilde{\chi}_1^\pm} = m_{\tilde{\chi}_1^0} +100$ GeV, while $m_{\tilde{\chi}_1^0} \lesssim 260$ GeV are also excluded for $m_{\tilde{b}_1} \approx 540$ GeV~\cite{ATLAS:2016kjm}.
If the LSP is Wino-like or Higgsino-like, then one expects that there is a chargino state with its mass close to that of the LSP. The decay $\tilde{t}_1 \to b \tilde{\chi}_1^\pm$ will be open as long as $\tilde{t}_1$ is heavier. The decay products of the $\tilde{\chi}_1^\pm$ to $\tilde{\chi}_1^0$ are likely to be too soft to be detected. The signal is similar to the $\tilde{b}_1$ search discussed above with the similar limit.
For $m_{\tilde{t}_1} \approx m_t + m_{\tilde{\chi}_1^0}$, the current bound on the $\tilde{t}_1$ mass is rather weak. Both ATLAS and CMS performed searches for the $\tilde{t}_2$ states with the decays to $Z \tilde{t}_1 $, $h \tilde{t}_1 $. The bound on $m_{\tilde{t}_2}$ from the most recent ATLAS Run 2 analysis is up to 730 GeV for 100\% decay to $Z \tilde{t}_1$~\cite{ATLAS:2016tpc}. The analysis for the $h \tilde{t}_1$ mode with Run 2 data has not come out yet. The limit from Run 1 data is about 600 GeV~\cite{Aad:2015pfx,Khachatryan:2014doa} and similar limits were obtained if $\tilde{t}_2$ decays to a mixture of the $Z \tilde{t}_1$, $h \tilde{t}_1 $, and $t\tilde{\chi}_1^0$ final states.
\section{Benchmark Points for Case Studies}
\label{sec:benchmark}
We are interested in searches for the heavier stop and sbottom in the case when the lightest stop is hidden. Therefore we choose the benchmark points of our study to satisfy $0\leqslantm_{\tilde{t}_1}-m_t-m_{\tilde{\chi}_1} \leqslant20$ GeV where the bound is weakest. Of course, for a heavier $\tilde{t}_1$ the difference can be bigger from current bounds. For $m_{\tilde{t}_1}$ being much closer to $m_{\tilde{\chi}_1}^0$, there have been several studies focusing on these scenarios~\cite{Beuria:2015mta,Kaufman:2015nda,Ghosh:2013qga}. We also assume that $\tilde{\chi}_1^0$ is Bino-like so that there is no nearby chargino state. Otherwise the $\tilde{t}_1 \to b \tilde{\chi}_1^\pm$ search would provide a strong constraint. Moreover, if $\tilde{b}_1$ is close to or even lighter than $\tilde{t}_1$, then the sbottom search in $\tilde{b}_1 \to b\tilde{\chi}_1^0$ decay will provide the strongest constraint. As we see from the summary in the last section, the constraint on $b$-jets + MET is very strong and the exclusion limit has reached $\sim 1$ TeV for these decay modes. Thus, we will focus on points where $\tilde{b}_1$ is somewhat heavier where other decay modes such as $\tilde{b}_1 \to W \tilde{t}_1$ are open, so that the traditional search based on $\tilde{b}_1 \to b\tilde{\chi}_1^0$ is not effective. We assume that the gluino is heavy enough so that the stops and sbottoms are dominated by the direct pair production. Otherwise the gluino cascade decays would be the strongest probe.
Given the spectrum $m_{\tilde{\chi}_1}^0 < m_{\tilde{t}_1} < m_{\tilde{b}_1} < m_{\tilde{t}_2}$, we can find two main classes of model points depending on whether there are other neutralinos and charginos lying between them. We denote Type A models to have the lightest charginos and the second lightest neutralinos heavier than our second stop, so that they decouple from our discussion, given that their direct pair production rate is much smaller than that of squarks of the same mass.
In Type B models, the second neutralino and accompanying charginos are lighter than $\tilde{t}_2$ so that they may appear in the cascade decays of $\tilde{t}_2$. These neutralinos and charginos may be either Wino-like or Higgsino-like, or both. From the naturalness point of view, it is preferable to have the Higgsinos not too heavy. In addition, Winos couple to the squarks via the $SU(2)$ gauge coupling instead of the large top Yukawa coupling. The branching fractions of stop and sbootom decays through the Winos are often small. Then the decay patterns are mostly similar to the Type A models. Therefore for Type B models we focus on the cases where the relevant charginos and neutralinos are mostly Higgsino-like. The benchmark points are selected to be compatible with the current experimental constraints as examined in the Appendix.
\subsection{Type A}
From the parameter space scan described earlier, we select several benchmark points of Type A spectrum, listed in Table~\ref{table:A_spectrum}.\footnote{In the first version of the paper, the A1 point has been ruled out after the release of new data at ICHEP 2016, so it is removed and the original A2 is moved to A1. We add two new benchmark points A2 and A3 which have similar decay patterns as the original A1, but with heavier spectra. The original A3-A5 are shifted to A4-A6.} The range of $\tilde{t}_2$ and $\tilde{b}_1$ masses is chosen from $\sim 700$ GeV up to $\sim 1.2$ TeV. A1 and A2 have relatively light spectra which are not far from the current bounds. They may soon be discovered or excluded at LHC Run 2 with more luminosities. Benchmarks A4--A6 have heavy $\tilde{t}_2$ and $\tilde{b}_1$ which are close to the reach limits of the 14 TeV LHC.
As expected, to obtain a Higgs boson mass close to 125 GeV, these benchmark points all have large mixing $|X_t/m_{\tilde{t}}|$ between the left- and right-handed stops.
\begin{table}[ht]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Spectrum & A1 & A2 & A3 & A4 & A5 & A6 \\
\hline \hline
$m_{\tilde{t}_2}$(GeV) & 815.4 & 887.1 & 1077.3 & 1230.6 & 1253.2 & 1262 \\
\hline
$m_{\tilde{b}_1}$(GeV) & 693.0 & 704.5 & 812.8 & 1029.5 & 1143.8 & 1229\\
\hline
$m_{\tilde{t}_1}$(GeV) & 491.0 & 605.5 & 687.6 & 904.0 & 916.5 & 640.1\\
\hline
$m_{\tilde{\chi}_1^0}$(GeV) & 304.9 & 414.2 & 498.0 & 710.8 & 724.2 & 459.4\\
\hline
$X_t/m_{\tilde{t}}$ & -1.81 & 1.58 & -2.17 & -1.84 & -1.82 & 1.51\\
\hline
$m_h$(GeV) & 122.8 & 122.7 & 123.4 & 124.9 & 124.6 & 122.3\\
\hline
\end{tabular}
\end{center}
\caption{Spectra of Type A Benchmark points.\label{table:A_spectrum}}
\end{table}
Due to the large mixing term, it is typical to have a large mass gap between $\tilde{t}_1$ and $\tilde{t}_2$, leaving a relatively large phase space for $\tilde{t}_2 \to \tilde{t}_1 + Z/h$ decays. In the scanned parameter space it is also common to have the mass difference $m_{\tilde{t}_2} - m_{\tilde{b}_1} \geq m_W$, which opens up the decay mode $\tilde{t}_2 \to \tilde{b}_1 + W$. This decay mode has not been considered by the experimental analysis of $\tilde{t}_2$ searches.
In Table~\ref{table:A_decay} we list the branching ratios of various decay modes of $\tilde{t}_2$ and $\tilde{b}_1$ for the benchmark points. The branching ratios are calculated by SDECAY~\cite{Muhlleitner:2003vg}.
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
\hline
Channel & A1 & A2 & A3 & A4 & A5 & A6\\
\hline \hline
$\tilde{t}_2 \to \tilde{b}_1 + W^+$ & 16.5 & 42.0 & 48.2 & 42.1 & 8.1 & 0 \\
\hline
$\tilde{t}_2 \to \tilde{t}_1 + Z$ & 74.5 & 47.6 & 44.6 & 52.9 & 79.2 & 53.1 \\
\hline
$\tilde{t}_2 \to \tilde{t}_1 + h$ & 5.9 & 3.9 & 5.0 & 2.5 & 10.7 & 45.6 \\
\hline
$\tilde{t}_2 \to t+\tilde{\chi}_1^0$ & 3.1 & 6.5 & 2.2 & 2.5 & 2.0 & 1.2\\
\hline \hline
$\tilde{b}_1 \to \tilde{t}_1 + W^-$ & 99 & 90.1 & 98.0 & 98.3 & 99.5 & 99.3 \\
\hline
$\tilde{b}_1 \to b+\tilde{\chi}_1^0$ & 1.0 & 9.9 & 2.0 & 1.7 & 0.5 & 0.7\\
\hline
\end{tabular}
\end{center}
\caption{Branching ratios of the major decays of $\tilde{t}_2$ and $\tilde{b}_1$ for the Type A benchmark models. The rest of the decay branching ratio is $\tilde{t}_2$ and $\tilde{b}_1$ decaying to the corresponding quarks and an LSP.\label{table:A_decay}}
\end{table}
As it can be seen, in Type A models, the $\tilde{t}_2 \to \tilde{t}_1 Z$ decay branching fraction is always significant, and it is dominant for A1 and A5. The $\tilde{t}_2 \to \tilde{t}_1 h$ branching ratio is smaller than the $\tilde{t}_1 Z$ branching ratio. It is controlled by the difference between two diagonal soft breaking masses. More specifically, the $h\tilde{t}_2 \tilde{t}_1$ coupling is proportional to $\cos 2 \theta_t$ where $\theta_t$ is the mixing angle diagonalizing the stop mass matrix. If the two diagonal soft breaking masses are exactly equal, $m_{u_3} = m_{Q_3}$, then $\theta_t=\frac{\pi}{4}$ and the mass eigenstates $\tilde{t}_1,\tilde{t}_2$ will be equal mixtures between $\tilde{t}_L$ and $\tilde{t}_R$. In this case the $h\tilde{t}_2 \tilde{t}_1$ coupling cancels between the contributions coming from the left-handed and the right-handed stops, and the $\tilde{t}_2 \to \tilde{t}_1 + h$ decay will be highly suppressed.
In models where $\tilde{t}_2 \to \tilde{b}_1 + W$ decay is kinematically allowed, the branching ratio of this channel increases rapidly as the allowed phase space expands, and easily becomes comparable to the $\tilde{t}_1 + Z$ decay channel, as in models A2--A4. The model A6 is chosen such that the $\tilde{t}_2 \to \tilde{b}_1 + W$ decay is closed while the $\tilde{t}_2 \to \tilde{t}_1 h$ branching ratio is significant.
On the other hand, in Type A spectrum, $\tilde{b}_1$ decays predominately
to $\tilde{t}_1 W$ if the phase space allows. This is not covered by the current experimental searches for direct pair-production of $\tilde{b}_1$~\cite{Aad:2015pfx,Khachatryan:2015wza,Khachatryan:2016xvy,CMS:2014dpa,Chatrchyan:2013fea,Aad:2014pda,Khachatryan:2016kod}. We would advocate that this decay mode should be included in the future direct $\tilde{b}_1$ pair production search analysis. In these benchmark models, the branching ratios of the direct decays of $\tilde{t}_2$ and $\tilde{b}_1$ to an LSP plus $t$ or $b$ are small, so the searches using these decays will not be effective.
In Table~\ref{table:A_final} we list the fractions of the final states in terms of $t, b, W, Z, h$ (without their subsequent decays) of the $\tilde{t}_2 \tilde{t}_2$ and $\tilde{b}_1 \tilde{b}_1$ pair productions, aside from a pair of $\tilde{\chi}_1^0$'s which are implicitly understood. The recent advances in jet substructure techniques to tag hadronically decayed top quarks may help to identify final states which contain them~\cite{Kaplan:2008ie,CMS:2009lxa}, so in the list we keep the top quark instead of its decay product in the final states. As a result, all $W$ bosons and $b$ jets listed in Table~\ref{table:A_final} are coming from SUSY particle decays instead of top decays.
\begin{table}
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
& A1 & A2 & A3 & A4 & A5 & A6\\
\hline \hline
$\sigma(\tilde{t}_2\tilde{t}_2) (fb)$ & 33.8 & 19.4 & 5.1 & 1.9 & 1.7 & 1.6\\
\hline \hline
$ttZZ$ & 55.5 & 22.6 & 19.9 & 28 & 62.7 & 28.2\\
\hline
$ttZWW$ & 24.6 & 40.0 & 42.1 & 44.5 & 12.8 & 0\\
\hline
$ttZh$ & 8.8 & 3.7 & 4.5 & 2.6 & 16.9 & 48.4\\
\hline
$tt4W$ & 2.7 & 17.6 & 22.3 & 17.7 & 0.7 & 0\\
\hline
$tthWW$ & 1.9 & 3.3 & 4.7 & 2.1 & 1.7 & 0\\
\hline
$tthh$ & 0.4 & 0.1 & 0.3 & $\sim$0 & 1.1 & 20.8\\
\hline \hline
$\sigma(\tilde{b}_1\tilde{b}_1) (fb)$ & 94.5 & 85.2 & 34.5 & 7 & 3.3 & 1.9\\
\hline \hline
$ttWW$ & 98 & 81 & 96 & 96.6 & 99 & 98.6\\
\hline
$tbW$ & 2 & 19.6 & 4 & 3.3 & 1.0 & 1.3\\
\hline
\end{tabular}
\caption{14 TeV production cross sections and fractions of the final states (in terms of $t,\, b,\, Z,\, W,\, h$) of $\tilde{t}_2$ and $\tilde{b}_1$ for Type A models. All final states also contain an additional $\tilde{\chi}^0_1\tilde{\chi}^0_1$ pair which becomes MET. In the list, the $W$ and $b$ jets in the final states are produced from squark decays rather than top decay.\label{table:A_final}}
\end{center}
\end{table}
It is pretty common for Type A models the heavy stop pairs produce an excess in $W$ bosons. Since each top quark itself gives another $W$ boson in its decay, the chances for the $\tilde{t}_2$ pairs in our model points to give a final state with 4 or more $W$'s varies up to $\sim 70 \%$. The high multiplicity of $W$ bosons can lead to excesses in same-sign dilepton and multiple lepton events. As a result, these types of signals are also useful for $\tilde{t}_2$ searches in our benchmark models. Also, top pair associated with additional $ZZ$, $Zh$ or $hh$ constitutes considerable fractions of final states for some model points. They are the basis of the existing experimental searches for $\tilde{t}_2\ttwo$ direction production.
Sbottom pairs in our benchmark points predominately decay into $4W + 2b$ final state, which also leads to an excess in same-sign dilepton and multiple lepton signals.
\subsection{Type B}
In Type B models there are additional neutralinos and charginos lighter than $\tilde{t}_2$. To be distinct from the phenomenology of Type A models, we select benchmark points where these neutralinos and charginos have significant appearance in the decay chains from $\tilde{t}_2$. Because Higgsinos have larger couplings to the stops than the Winos, for the benchmark points we choose these neutralinos and charginos to be Higgsino-like.
\begin{table}[ht]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Spectrum & B1 & B2 & B3 & B4 \\
\hline \hline
$m_{\tilde{t}2}$(GeV) & 952.6 & 1067.1 & 1131.9 & 1265.4 \\
\hline
$m_{\tilde{b}1}$(GeV) & 832.5 & 749.1 & 870.4 & 1232.6 \\
\hline
$m_{\tilde{t}1}$(GeV) & 654.4 & 677.8 & 785.6 & 585.3 \\
\hline
$m_{\tilde{\chi}_1^0}$(GeV) & 478.6 & 499.5 & 594.8 & 407.7\\
\hline
$m_{\tilde{\chi}_2^0}$(GeV) & 774.4 & 702.8 & 824.1 & 823.3 \\
\hline
$m_{\tilde{\chi}_3^0}$(GeV) & 775.7 & 703.2 & 824.3 & 825.7 \\
\hline
$m_{\tilde{\chi}^\pm}$(GeV) & 772.3 & 699.5 & 820.7 & 822.1 \\
\hline
$X_t/m_{\tilde{t}}$ & 1.65 & -1.78 & -1.74 & -1.69 \\
\hline
$m_h$(GeV) & 124.0 & 124.7 & 124.9 & 123.5 \\
\hline
\end{tabular}
\end{center}
\caption{Spectra of Type B benchmark points. $\tilde{\chi}_2^0, \tilde{\chi}_3^0, \tilde{\chi}^\pm$ are Higgsino-like.\label{table:B_spectrum}}
\end{table}
The spectra of Type B benchmark points are listed in Table~\ref{table:B_spectrum}.
The mass gap between $\tilde{t}_2,\tilde{b}_1$ and $\tilde{\chi}^\pm,\tilde{\chi}_2^0$ are chosen to be big enough to allow sufficient phase space for decays through the additional neutralinos or charginos, such as $\tilde{t}_2 \to t \tilde{\chi}_2^0$ or $b\tilde{\chi}^\pm$. Therefore these benchmark points have relatively heavy $\tilde{t}_2$, and $m_{\tilde{t}_2}$ ranges from $\sim$ 900 GeV to 1.3 TeV. For a more compressed spectrum, the decays of $\tilde{t}_2$ and $\tilde{b}_1$ into second neutralino and charginos are suppressed by the phase space, then the decay patterns will be similar to those of Type A benchmark points. If $\tilde{\chi}^0_{2,3}, \tilde{\chi}^\pm$ are lighter than $\tilde{b}_1$, $\tilde{b}_1$ can also decay to $b\tilde{\chi}_{2,3}$ or $t\tilde{\chi}^\pm$ (if kinematically open), in addition the $\tilde{t}_2 \to b\tilde{\chi}^\pm, \, t\tilde{\chi}^0_{2,3}$ decays. The branching ratios of SUSY particle decays for the benchmark points are listed in Table~\ref{table:B_decay}.
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Channel & B1 & B2 & B3 & B4\\
\hline \hline
$\tilde{t}_2 \to \tilde{t}_1 + Z$ & 58.2 & 24.4 & 30.2 & 30.2 \\
\hline
$\tilde{t}_2 \to \tilde{b}_1+W^+$ & 12.9 & 36.6 & 38.0 & 0 \\
\hline
$\tilde{t}_2 \to \tilde{t}_1 + h$ & 3.3 & 9.6 & 7.8 & 25.8 \\
\hline
$\tilde{t}_2 \to t+\tilde{\chi}_2^0$ & 2.7 & 3.9 & 7.7 & 9.6 \\
\hline
$\tilde{t}_2 \to t+\tilde{\chi}_3^0$ & 0.1 & 7.8 & 2.9 & 7.6 \\
\hline
$\tilde{t}_2 \to b+\tilde{\chi}_1^\pm$ & 20.9 & 16.0 & 11.7 & 26.5 \\
\hline \hline
$\tilde{b}_1 \to \tilde{t}_1 + W^-$ & 90.0 & 0 & 31.3 & 58.6 \\
\hline
$\tilde{b}_1 \to b + \tilde{\chi}_2^0$ & 3.7 & 35.1 & 19.7 & 12.6 \\
\hline
$\tilde{b}_1 \to b + \tilde{\chi}_3^0$ & 3.6 & 33.4 & 20.0 & 12.5\\
\hline
$\tilde{b}_1 \to t + \tilde{\chi}^\pm$ & 0 & 0 & 0 & 15.6 \\
\hline \hline
$\tilde{\chi}_2^0 \to t+ \tilde{t}_1$ & 0 & 0 & 0 & 92.1 \\
\hline
$\tilde{\chi}_2^0 \to h+ \tilde{\chi}_1^0$ & 94.2 & 1.8 & 96.9 & 7.1 \\
\hline
$\tilde{\chi}_2^0 \to Z+ \tilde{\chi}_1^0$ & 5.8 & 98.2 & 3.1 & 0.8 \\
\hline \hline
$\tilde{\chi}_3^0 \to t+ \tilde{t}_1$ & 0 & 0 & 0 & 86.9 \\
\hline
$\tilde{\chi}_3^0 \to h+ \tilde{\chi}_1^0$ & 4.4 & 96.1 & 1.7 & 1.2 \\
\hline
$\tilde{\chi}_3^0 \to Z+ \tilde{\chi}_1^0$ & 95.6 & 3.9 & 98.3 & 11.9 \\
\hline \hline
$\tilde{\chi}^\pm \to b+ \tilde{t}_1$ & 86.8 & 37.2 & 29.8 & 93.0 \\
\hline
$\tilde{\chi}^\pm \to W+ \tilde{\chi}_1^0$ & 13.2 & 62.8 & 70.2 & 7.0 \\
\hline
\end{tabular}
\end{center}
\caption{Branching ratios of the major decays of $\tilde{t}_2$ and $\tilde{b}_1$ for the Type B benchmark models. The rest of the decay branching ratio is $\tilde{t}_2$ and $\tilde{b}_1$ decaying to the corresponding quarks and an LSP.\label{table:B_decay}}
\end{table}
From the Table~\ref{table:B_decay} we see that $\tilde{t}_2 \to t \tilde{\chi}^0_{2,3}$ decay branching ratios are usually small, due to the phase space suppression. The $\tilde{t}_2 \to b \tilde{\chi}^\pm$ branching ratio, on the other hand, can get up to about 1/4 for model B4. Similarly, the $\tilde{b}_1 \to t \tilde{\chi}^\pm$ branching ratio is typically small but $\tilde{b}_1 \to b \tilde{\chi}^0_{2,3}$ branching ratios can be quite big. In model B2, the sum of the branching ratios of $\tilde{b}_1 \to b\tilde{\chi}^0_2$ and $b\tilde{\chi}^0_3$ is close to 70\%. The heavy neutralinos tends to decay to $\tilde{\chi}_1^0 + h/Z$, unless the decay channel to $t \tilde{t}_1$ is open (as in model B4), in which case it becomes the dominant decay channel. The chargino decays to $b\tilde{t}_1$ or $W\tilde{\chi}_1^0$. Which branching ratio is larger depends on the model point.
Similar to the Type A case, we can list the fractions of the final states of the $\tilde{t}_2\ttwo$ and $\tilde{b}_1\bone$ pair productions, in terms of $t, b, W, Z, h$ aside from the $\tilde{\chi}_1^0$. Because there are many more possible decay channels for Type B models, we only list the final states of a single decay chain in Table~\ref{table:B_final}. The complete final states can be obtained by simply squaring the Table. In addition to the final states which have been present in Type A models, Type B models can produce final states with a large number (up to 6) of $t$ or $b$ quarks. Therefore, search channels for multiple tops or bottoms could be interesting and important for some Type B spectra (e.g., B4).
\begin{table}
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
& B1 & B2 & B3 & B4\\
\hline \hline
$\sigma(\tilde{t}_2\tilde{t}_2) (fb)$ & 12.1 & 5.5 & 3.6 & 1.6 \\
\hline \hline
$tZ/bWZ$ & 58.5/0.5 & 28.5/13.1 & 33.3/7.7 & 31.2/0 \\
\hline
$th/bWh$ & 5.8/0.5 & 17.2/12.0 & 15.3/7.4 & 26.6/0 \\
\hline
$tbb$ & 18.1 & 6.0 & 3.5 & 24.6 \\
\hline
$Wb$ & 2.8 & 21.5 & 19.2 & 1.9 \\
\hline
$3t$ & 0 & 0 & 0 & 15.4 \\
\hline
$tWW$ & 11.6 & 0 & 11.9 & 0 \\
\hline \hline
$\sigma(\tilde{b}_1\tilde{b}_1) (fb)$ & 29.6 & 58.1 & 22.1 & 1.9 \\
\hline \hline
$tW$ & 90 & 0 & 31.3 & 59.7 \\
\hline
$ttb$ & 0 & 0 & 0 & 37.0 \\
\hline
$hb$ & 3.7 & 32.7 & 19.4 & 1.0 \\
\hline
$Zb$ & 3.6 & 35.8 & 20.3 & 1.6 \\
\hline
\end{tabular}
\caption{14 TeV production cross sections and fractions of the final states of $\tilde{t}_2$ and $\tilde{b}_1$ for Type B models. Only final states of a single $\tilde{t}_2$ or $\tilde{b}_1$ decays are listed. An additional LSP $\tilde{\chi}_1^0$ is implicitly understood. The fractions of the finals states of the squark pair can be easily obtained by squaring this table.\label{table:B_final}}
\end{center}
\end{table}
\section{Collider Studies for LHC 14 TeV}
\label{sec:collider}
Given the complex decay chains and many possible final state combinations of the second stop and the sbottom, we expect that there exist many possible signal channels. The best experimental reach may come from a combination of different signals. In this section we perform a rudimentary collider study for the benchmark points discussed in the previous section for the 14 TeV LHC and point out the interesting channels. Even though some of the channels were not considered in the $\tilde{t}_2$ simplified model analyses by ATLAS and CMS, most of them have been used in other SUSY or new physics searches. A purpose of this study is to point out their relevance for the $\tilde{t}_2$ and $\tilde{b}_1$ searches. These existing analyses could readily be adapted to the current case, maybe with some optimizations of cuts for the current scenario.
For SUSY signals, we use MadGraph 5~\cite{Alwall:2014hca} to generate $\tilde{t}_2$ and $\tilde{b}_1$ pair events, with PYTHIA 6~\cite{Sjostrand:2006za} for parton showering and hadronization simulations. Detector simulations were done by Delphes 3~\citep{deFavereau:2013fsa}, with the anti-$k_t$ jet algorithm~\cite{Cacciari:2008gp} and $\Delta R=0.5$. We use the CTEQ6L~\cite{Pumplin:2002vw} PDF in order to match the Snowmass background simulations. All signal cross sections are normalized to NLO+NLL results with gluino decoupled~\cite{Borschensky:2014cia}.
We adopt the backgrounds generated by the Snowmass 2013 Energy Frontier Simulation group~\cite{Snowmass1,Snowmass2}. No-pileup effects are included for either the signals or the backgrounds. For this analysis the dominant background is the top-pair production, and top pair plus an extra boson ($W$, $Z$ or $h$). We also include other backgrounds such as the single/multiple boson(s) with jets, and the single top production.
\subsection{Basic cuts and tagging for backgrounds and signals}
\label{subsect:tagging}
The lepton and $b$ jet tagging efficiencies are taken to be the same as those of the Snowmass 2013 Energy Frontier Simulations. Additionally, we drop all leptons having a $\Delta R \leqslant 0.4$ with any jets. For both $e$ and $\mu$, we require them to have $p_T>15$~GeV and $\eta<2.5$. We will refer to these isolated leptons simply as leptons for the rest of our discussion.
Each jet candidate is required to have $\eta<2.5$ and $p_T>25$~GeV, which is rather loose since our analysis is not very sensitive to non-$b$-tagged jets. The maximum tagging efficiency for $b$ jets is $\approx 70\%$. The $b$-jet mis-tagging rate from light flavors is $0.1\%$. A $b$ jet candidate must have $p_T>30$~GeV. Further selection rules for leptons and jets will be described in detail later.
Since all SUSY signals we discussed have two LSP which could lead to significant MET, an event with a higher MET is preferred, especially for those with no or only one lepton in the final state. On the other hand, SM background events with more leptons could get a large MET from the associated neutrinos. Therefore, for a higher lepton multiplicity we would have a lower the MET cut in order to include more signal events. As a result, we set up a preliminary criteria based on MET: for all-hadronic or one-leptonic channels, all signal events are required to have a MET greater than 200 GeV. On the other hand, events with two/more than two leptons would be vetoed if their MET is less than 150/100~GeV.
To further suppress the backgrounds, scalar $H_T$ could provide a good discrimination between the signal and backgrounds, since we are interested in the pair production of heavy particles. We require that each event should have scalar $H_T \geq 800$ GeV as a preliminary selection rule.
A pair of opposite-sign leptons of the same flavor can come from $Z$ decays. The $Z$ boson produced in $\tilde{t}_2$ decay could be somewhat boosted due to the large $m_{\tilde{t}_2}-m_{\tilde{t}_1}$. To define the $Z$ candidate, we require that the opposite-sign same-flavor lepton pair lepton pair to have $|\sqrt{(\Sigma p_\ell)^2} - m_Z | \leqslant10$~GeV and $\Delta R < 1.5$.
Opposite-sign lepton pairs also often arise from SM background such as $t\bar{t}$. For our signals, $\tilde{t}_2$ pair decays usually produce extra $W/Z$ bosons. If they are boosted and decay hadronically, they may be tagged to help discriminating signals and backgrounds. We follow the method presented in Ref.~\cite{Han:2011ab}: A Cambridge-Archen jet algorithm~\cite{Dokshitzer:1997in} is adopted to identify fat jets from vector boson decays, with $\Delta R=0.8$ and $p_T>200$ GeV. Any fat jet constructed this way with a invariant mass between 60 and 110 GeV would be our vector jet candidate. Furthermore, we require $N$-subjettiness~\cite{Thaler:2010tr} $\tau_{21}=\frac{\tau_2}{\tau_1}< 0.5$, which means that its substructure is more likely to have two subjets rather than coming from QCD backgrounds.
\subsection{Signal channels and benchmark results}
In brief, the SUSY signals we are looking for in this work can be understood as a bottom quark pair accompanied by multiple bosons ($W,Z$ or $h$) and a pair of missing neutralinos, with possible extra $t,b$ pairs. Consequently, events with multiple $b$ jets are favored. Compared with the SM background such as $t\bar{t}$ events which can give two $b$ jets at parton level, our signals can have more $b$ jets in the final states. $\tilde{t}_1\tone hZ/\tilde{t}_1\tone ZZ/\tilde{t}_1\tone hh$ decays of $\tilde{t}_2\ttwo$ with an $h$ or $Z$ decay to a pair of $b$ quarks and $\tilde{t}_1\tone$ cascade to a pair of tops can give us 4 or more $b$ jets. Thus, requiring more than 3 $b$ jets in each event would be a good way to handle SM backgrounds~\cite{Aad:2014lra}. On the other hand, signal events with two or less $b$ jets would be overwhelmed by SM backgrounds such as $t\bar{t}$ production. Therefore, for events with $N_b \leqslant 2$ would need features such as a large multiplicity of leptons or same-sign dileptons to increase their signal sensitivity. The distribution of $N_b$ with zero and non-zero lepton(s) in final states are shown in Fig.~\ref{fig:NUMB}:
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{NUMB_0.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{NUMB_L.pdf}
\end{subfigure}
\caption{Event distribution binned by $N_b$ for both signals and SM background with $\slashed{E}_T \geq 200$~GeV. \textbf{Left:} $N_b$ distribution for events without leptons. The red shade is the SM background. The blue (green) lines are $\tilde{t}_2$ ($\tilde{b}_1$) contributions, and the solid (dashed) lines are for the A1 (B1) benchmark model.
\textbf{Right:} the same distributions for events with at least one lepton.}
\label{fig:NUMB}
\end{center}
\end{figure}
Based on these considerations we have tried various search channels by $N_\ell$ and $N_b$. The set of useful signal channels that we found are classified as:
\begin{enumerate}
\item No lepton and no less than 3 $b$ jets with a large MET and no less than 6 jets in total ($0\ell3b$).
\item 1 lepton and no less than 3 $b$ jets and no less than 6 jets in total ($1\ell3b$).
\item 2 opposite-sign (OS) leptons forming a $Z$ candidate with no less than 2 $b$ jets, and 5 or more jets in total ($Z2b$).
\item 2 same-sign (SS) leptons with at least one $b$ jet, also no less than 5 jets in total ($SS+nb$).
\item 3 or more leptons and at least one $b$ jets, the number of jets in total $\geqslant$ 2 (Multi-$\ell$).
\end{enumerate}
Different channels are classified as such that for zero or one lepton channels, more $b$ jets and total number of jets are required to optimize the sensitivity. As the lepton multiplicity increases, we loosen up the requirements on $N_b$, $N_j$ and MET in order to keep more signal events. Some of the channels are essentially the same as those which are already used in experimental $\tilde{t}_2$ searches. Others are also close to some other SUSY or new physics searches but have not been applied to $\tilde{t}_2$ searches. We also identified some additional requirements which may enhance the signal and background discrimination in some channels. For each signal channel, it would be beneficial to further divide into signal regions based on various energy distributions to utilize the possible different distributions between signals and backgrounds. This requires more sophisticated event simulations to produce accurate event distributions. For simplicity, here we will treat each signal channel as a whole and leave more detailed analyses for the experimental collaborations.
\subsubsection{No lepton, large MET, with three or more $b$ jets ($0\ell3b$)}
In this channel, we require no isolated lepton in the final states, $N_b \geqslant 3$, and the total number of jets $\geqslant6$. For such fully-hadronic events to be triggered, we require that the leading jet has $p_T \geqslant 250$~GeV and two subleading jets have $p_T \geqslant 90$~GeV. For $b$ jets we also require at least two $b$ jets to have $p_T \geqslant 90$~GeV and the rest of $b$ jets to have $p_T>30$~GeV.
A large $\slashed{E}_T$ cut is imposed in this channel to greatly suppress the contribution from QCD background, allowing us to utilize the Snowmass 2013 Energy Frontier Simulation results. The $\slashed{E}_T$ distributions for some benchmark signals and the background are shown in Fig.~\ref{fig:0l3b}. We can see that both $\tilde{b}_1\bone$ and $\tilde{t}_2\ttwo$ events on average have higher $\slashed{E}_T$ than background events. We impose $\slashed{E}_T >280$~GeV for event selections in this channel.
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\includegraphics[scale=0.6]{Res_0l3bA_1.pdf}
\caption{The $\slashed{E}_T$ distributions of the $0\ell3b$ channel for benchmarks A1, B1 and the SM background.
The red shade is the SM background. The blue lines are $\tilde{t}_2$ contributions and the green lines are $\tilde{b}_1$ contributions. The solid lines are for A1 and the dashed lines are for B1.\label{fig:0l3b}
}
\end{center}
\end{figure}
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
Bkg Total &Model & A1& A2&A3&A4&A5&A6&B1&B2&B3&B4\\\hline
\multirow{2}{*}{191.4}&$\tilde{t}_2\ttwo$&119.9&51.2&24.9&7.6&11.8 &33.2 &61.5 &57.6 &26.1 &51.6 \\
&$\tilde{b}_1\bone$&67.0&56.4&20.5&5.0&5.2 &7.6 &25.7 &86.2 &24.6 &31.1 \\\hline
\end{tabular}
\end{center}
\caption{Background and signal events for $0\ell3b$ channel normalized to 300 fb$^{-1}$ integrated luminosity at 14 TeV LHC for each benchmark model point.\label{table:0l3b}}
\end{table}
The number of signal events of each benchmark model and the background events passing the selection in this channel for a 300 fb$^{-1}$ integrated luminosity is listed in Table~\ref{table:0l3b}. We can see that the $0\ell3b$ channel can be useful for both $\tilde{t}_2$ and $\tilde{b}_1$ searches. Na\"ively one might expect that this channel is more useful for the $\tilde{t}_2$ search than the $\tilde{b}_1$ search because the $Z$ or $h$ bosons from $\tilde{t}_2$ decays give the extra $b$ jets. This is true for some (e.g., $\tilde{t}_2$ has a large branching ratio in decaying to $h/Z$ in A1, A5, A6, which produces more $b$ jets on average) but not all of the benchmark models. For example, in some Type B models (in particular B2), $Z$ or $h$ bosons can also arise from the cascade decays of $\tilde{b}_1$ through heavier neutralinos, so $\tilde{b}_1$ can provide comparable number of events as $\tilde{t}_2$. Even though in Type A models $\tilde{b}_1\bone$ pair production can only give two $b$ jets at the parton level, additional $b$-tagged jets can arise from QCD radiations and mis-tagged light jets. In addition, $\tilde{b}_1$ has a larger production cross section than $\tilde{t}_2$ because it is lighter. As a result, the number of events from $\tilde{b}_1$ and from $\tilde{t}_2$ can be comparable in many Type A models, and this channel can also be sensitive to $\tilde{b}_1$ pair production. Among the heavier benchmarks A4-A6, this channel is most sensitive to A6 and its contribution mainly comes from $\tilde{t}_2$ because of the large $\tilde{t}_2 \to h \tilde{t}_1$ branching ratio.
For Type B models, B3 has fewer signal events due to the more compressed spectrum which results in softer final state particles and hence a lower signal efficiency. This is also true for other signal channels discussed later. For B4, even though the production cross sections of $\tilde{t}_2$ and $\tilde{b}_1$ pairs are small, the final state particles are harder due to the large mass splittings. The signal efficiency is better. In addition, there are significant branching ratios of $\tilde{t}_2$ and $\tilde{b}_1$ decaying to $tbb$ and $ttb$ which also give a higher number of $b$ jets.
\subsubsection{One lepton with three or more $b$ jets ($1\ell3b$)}
In this channel, the isolated lepton is required to have $p_{T\ell}\geqslant25$~GeV. We also require more than 5 jets with $p_{T}\geqslant25$~GeV for each jet. The leading $b$ jet should have $p_{Tb}\geqslant 90$~GeV and other $b$ jets have $p_T>30$~GeV. The $\slashed{E}_T$ is required to be $\geqslant 200$~GeV. For the SM backgrounds, a significant $\slashed{E}_T$ can arise due to the neutrino produced by $W\to \ell\nu$ decay. A cut on the transverse mass $M_T=\sqrt{2(p_{T\ell}\slashed{E}_T)(1-\cos\Delta\phi)}$ can be an effective way to eliminate most of the SM backgrounds, since $M_T$ from single $W$ leptonic decay would have a drop off around $m_W$. The $M_T$ distributions for signals and the background for the single lepton channel are shown in Fig.~\ref{fig:1l3b}. We impose a $M_T>160$~GeV cut to enhance the signal significance.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.6]{Res_1l3b.pdf}
\caption{$M_T$ distributions for of $1\ell3b$ channel.}
\label{fig:1l3b}
\end{center}
\end{figure}
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
Bkg Total &Model & A1& A2&A3&A4&A5&A6&B1&B2&B3&B4\\\hline
\multirow{2}{*}{42.5}&$\tilde{t}_2\ttwo$&39.9&23.8 &10.8 &3.8 &3.2 &6.2 &20.2 &14.6 &7.8 &13.1 \\
&$\tilde{b}_1\bone$&27.1&13.5 &8.9 &2.6 &2.4 &3.1 &9.3 &1.3 &3.0 &12.5 \\\hline
\end{tabular}
\end{center}
\caption{Background and signal events for the $1\ell3b$ channel with a 300 fb$^{-1}$ integrated luminosity for each model point.\label{table:1l3b}}
\end{table}
This is one of the main channels for current experimental $\tilde{t}_2$ searches~~\cite{Aad:2015pfx,Khachatryan:2014doa}. From Table~\ref{table:1l3b} we see that $\tilde{b}_1$ could also give nearly comparable contributions for many benchmark points (except B2 where $W$'s are not produced in $\tilde{b}_1$ decays), due to the larger cross section for being lighter.
\subsubsection{Opposite-sign dilepton (Z) channels}
Events with an OS lepton pair and $b$ jets would be dominated by SM $t\bar{t}$ production. Here we define the OS lepton channels by having two OS leptons with $p_T>15$~GeV, with $\slashed{E}_T>150$~GeV.
Also each event is required to have $\geqslant$ 2 $b$-jets with $p_T>30$~GeV and $\geqslant$ 5 jets of $p_T>25$~GeV in total. A rough estimation for this channel predicts $\sim 800$ background events and $\lesssim 10$ signal events for those heavy benchmark points such as A4-A6 or B2-B4, yielding a significance too low to be useful. Therefore, additional requirement is needed to suppress the background and we focus on the case where the lepton pair come from the $Z$ decay, since there is always a significant branching ratio of producing $Z$'s in $\tilde{t}_2$ decays. In addition, $Z$ can also arise from heavier neutralino decays in Type B models. To be identified as a $Z$, the OS leptons of the same flavor is required to have $\Delta R <1.5$ and an invariant mass $|m_{\ell\ell}-m_Z|<10$~GeV.
The $Z2b$ is also a standard search channel for the $\tilde{t}_2$ pair production~\cite{Aad:2015pfx,Aad:2014mha,Khachatryan:2014doa}. There is also an attempt to explain the ATLAS $Z$+jets+MET excess~\cite{Aad:2015wqa} by the two mixed stop system~\cite{Collins:2015boa}. Here we notice that most signal events contain additional $W$ or $Z$ bosons. Some of them may be boosted if they come from the decay of a heavy particle. We therefore explore the possibility that additional vector tags may be helpful with the signal and background discrimination. We divide the events passing the above requirements into exclusive channels with no vector tag ($Z2b$) and with at least one vector tag ($Z2bV$). They are listed in Table~\ref{table:Z2bV}. We see that the channel with a vector tag in general does better than without the vector tag,
especially for model points with larger $m_{\tilde{t}_2}-m_{\tilde{t}_1}$ and higher BR$(\tilde{t}_2 \to Z\tilde{t}_1)$ (e.g., A1, A6 and B4). Since the vector-tagging technique we used here is rather crude, further improvements with advanced vector-tagging techniques are possible.
These channels are in general more useful for the $\tilde{t}_2$ search, because in most benchmark models, $Z$ is rarely produced in $\tilde{b}_1$ decays. The exceptions are B2 and B3, where $Z$ can be produced from the neutralino decays. However, the signal efficiency for B3 is small due to its compressed spectrum. Only for B2 $\tilde{b}_1$ can give a comparable contribution to $\tilde{t}_2$.
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
Channel&Bkg Total &Model & A1& A2&A3&A4&A5&A6&B1&B2&B3&B4\\\hline
\multirow{2}{*}{$Z2bV$}&\multirow{2}{*}{16.7}&$\tilde{t}_2\ttwo$&27.1&6.5&3.0&1.2&1.5&2.8&9.2&3.2&2.0&2.4 \\
& &$\tilde{b}_1\bone$&1.5&0.0&0.2&0.0&0.0&0.0&0.5&3.1&0.3&0.1 \\\hline
\multirow{2}{*}{$Z2b$}&\multirow{2}{*}{29.2}&$\tilde{t}_2\ttwo$&20.5&7.0&2.7&0.9&1.4&1.2&9.0&2.7&1.6&1.0 \\
& &$\tilde{b}_1\bone$&1.7&0.0&0.5&0.1&0.0&0.0&0.5&3.5&0.3&0.1 \\\hline
\end{tabular}
\end{center}
\caption{Background and signal events for $Z2bV$ and $Z2b$ channel for each model point with a 300 fb$^{-1}$ integrated luminosity.}
\label{table:Z2bV}
\end{table}
The significance of this channel may be further improved if one can suppress the fake $Z$ bosons made of two leptons coming from opposite sign $W$'s from the SM $t\bar{t}$ background. This contamination may be estimated by the opposite-sign, opposite-flavor dilepton events which satisfy the invariant mass and $\Delta R$ requirements, after taking into account the different efficiencies of the electron and the muon. Here instead we introduce a simple kinematic variable dubbed ``leverage.'' It is inspired by $M_T$ and can be considered as a generalization applying to more than one leptons (or even other final state particles) together with MET. With multiple leptons it is defined as
\begin{equation}
L_\ell=\left(p_T^{\rm miss
{\sum} (1-\cos \Delta \phi_i)\right)/N_\ell
\end{equation}
In the SM fake $Z$ events, since we require $\Delta R\leqslant1.5$, the MET given by two neutrinos from $W$ decay would tend to be in the same direction as the fake $Z$ direction. On the contrary, there is much less such correlation for SUSY events, since $Z$ produced by $\tilde{t}_2$ decay can have a different direction from the MET. As it can be seen in both plots in Fig.~\ref{fig:Zbb}, most of the background events have a small $L_{\ell}$ and a suitable cut on $L_{\ell}$ can increase both the significance and $S/B$ ratio effectively. To enhance the signal significance, we put an additional $L_{\ell}>40$~GeV cut for both $Z2bV$ and $Z2b$ channels besides the cuts aforementioned. The numbers of events after the cut are listed in Table~\ref{table:L_cut}. We can see that it significantly reduces the remaining background events while retaining most of the signal events.
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\includegraphics[scale=0.35]{Res_Z2bV_2.pdf}
\includegraphics[scale=0.35]{Res_Z2b_2.pdf}
\caption{Histogram of leptonic $Z$ with $\geqslant 2\, b$-jets and $
\geqslant 5$ jets binned by $ L_{\ell}$. \textbf{Left:} With at least one extra vector-tagged jet. \textbf{Right:} without vector-tagged jets.}
\label{fig:Zbb}
\end{center}
\end{figure}
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
Channel&Bkg Total &Model & A1& A2&A3&A4&A5&A6&B1&B2&B3&B4\\\hline
\multirow{2}{*}{$Z2bV$}&\multirow{2}{*}{4.0}&$\tilde{t}_2\ttwo$&23.1&5.2 &2.7 &0.9 &1.3 &2.3 &7.6 &2.7 &1.9 &2.1 \\
& &$\tilde{b}_1\bone$&0.4 &0.0&0.2 &0.0 &0.0 &0.0 &0.4 &1.3 &0 &0.1 \\\hline
\multirow{2}{*}{$Z2b$}&\multirow{2}{*}{4.9}&$\tilde{t}_2\ttwo$&17.8 &5.9&1.9 &0.9 &1.2 &1.0 &7.9 &2.2 &1.1 &0.9 \\
& &$\tilde{b}_1\bone$&0.4 &0.0&0.0 &0.1 &0.03 &0.03 &0.4 &3.1 &0.2 &0.0 \\\hline
\end{tabular}
\end{center}
\caption{Background and signal events for $Z2bV$ and $Z2b$ channel for each model point after the $L_{\ell}>40$~GeV cut with a 300 fb$^{-1}$ integrated luminosity.}
\label{table:L_cut}
\end{table}
\subsubsection{Same-sign dilepton channel}
The previous channels are sensitive to the $\tilde{t}_2 \to \tilde{t}_1+ Z/h$ decays which are the focus of existing experimental analyses based on the simplified model approach. However, as we see in the benchmark models, it is common to have large fractions of final states with high multiplicities of $W$ bosons from both $\tilde{t}_2$ and $\tilde{b}_1$ decays. It is therefore important to study signals from multiple $W$'s in $\tilde{t}_2$ and $\tilde{b}_1$ searches. A very useful signal for multiple $W$'s is the same-sign dilepton which is relatively rare in the SM.
For this study we define the same-sign dilepton channel to have two leptons of the same charge and $p_T>15$~GeV, with $\slashed{E}_T >190$~GeV for each event. Also, we require at least one $b$-jet with $p_T>30$~GeV and the total number of jets of $p_T>25$~GeV is required to be more than 4. We further divide these events into two channels: $SS1b$ for events with exactly one $b$-tagged jet and $SS2b$ for events with two or more $b$-tagged jets.
Ignoring the contribution of misidentified leptons, only a few SM processes can generate a pair of same-sign dilepton, such as multiple vector boson or $t\bar{t}+$ boson production, etc. In our case, the dominant background is $t\bar{t}W/t\bar{t}Z/t\bar{t}h$ production. However, these processes' contribution suffer from small production cross sections ($\sim 3$~pb in total). The distributions of the signal and background events for the $SS1b$ and $SS2b$ channels in $\slashed{E}_T$ are shown in Fig.~\ref{fig:SS}. We see that our $\slashed{E}_T$ cut can further reduce the SM backgrounds.
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\includegraphics[scale=0.4]{Res_SS1b.pdf}
\includegraphics[scale=0.4]{Res_SS2b.pdf}
\caption{The $\slashed{E}_T$ distributions of the same-sign lepton pair with \textbf{Left:} 1 $b$ jet and \textbf{Right:} more than 1 $b$ jets.}
\label{fig:SS}
\end{center}
\end{figure}
\begin{table}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
Channel&Bkg Total &Model & A1& A2&A3&A4&A5&A6&B1&B2&B3&B4\\\hline
\multirow{2}{*}{$SS1b$}&\multirow{2}{*}{76.6}&$\tilde{t}_2\ttwo$&26.2&20.9 &10.1 &3.0 &1.1 &1.2 &7.1 &2.0 &2.6 &1.4 \\
& &$\tilde{b}_1\bone$&62.0&29.8 &16.2 &3.4 &3.9 &5.8 &19.5 &0.0 &1.9 &4.5 \\\hline
\multirow{2}{*}{$SS2b$}&\multirow{2}{*}{10.8}&$\tilde{t}_2\ttwo$&16.8&15.5 &7.1 &2.1 &1.0 &1.1 &7.4 &2.0 &3.4 &2.9 \\
& &$\tilde{b}_1\bone$&34.9&15.1 &8.2 &2.1 &1.7 &3.5 &11.5 &0.0 &1.0 &5.3 \\\hline
\end{tabular}
\end{center}
\caption{Background and signal events for $SS1b$ and $SS2b$ channel for each model point with a 300 fb$^{-1}$ integrated luminosity.\label{table:ss2l}}
\end{table}
One can see from Table~\ref{table:ss2l} that the numbers of signal events are comparable for $SS1b$ and $SS2b$ channels. However, $SS1b$ has much more background events due to mis-tagged $b$ jets, so $SS2b$ is expected to have better reaches. In general, for benchmark points where $\tilde{b}_1$ has a large branching ratio of decaying to $\tilde{t}_1 W$ (i.e., other than B2 and B3), more signal events come from $\tilde{b}_1$ due to the larger production cross section. Nevertheless, $\tilde{t}_2$ can also give a significant contribution. As $\tilde{t}_2$ pair decays may produce more than 4 $W$'s in the final states, the probability of getting same-sign dileptons could be helped by the combinatorial factor. Also, in models where the mass difference between $\tilde{b}_1$ and $\tilde{t}_1$ is small, the leptons coming from $\tilde{b}_1$ are softer and hence have a lower signal efficiency. For example, A1 and A2 have similar production rate for $\tilde{b}_1$ pairs, but the signal efficiency of the latter is almost halved because of its small $\tilde{b}_1$--$\tilde{t}_1$ mass difference. In this case, the contributions from $\tilde{t}_2$ and $\tilde{b}_1$ can even be comparable.
The same-sign dilepton signals, although not used in $\tilde{t}_2$ search yet, have been applied to many other new physics searches. Interestingly, excesses in the same-sign dilepton channel with $b$-jets and MET are found in both Run 1 and Run 2 of LHC in many separate analyses by the ATLAS and CMS collaborations. These include CMS SUSY Search~\cite{Chatrchyan:2013fea}, ATLAS SUSY Search~\cite{Aad:2014pda}, CMS $tth$ Search~\cite{CMS:2013tfa}, ATLAS Exotica Search~\cite{Aad:2015gdg}, and ATLAS $tth$ Search~\cite{Aad:2015iha} in Run 1 and ATLAS $tth$ Search~\cite{ATLAS:2016ldo}, CMS $tth$ Search~\cite{CMS:2016vqb} in Run 2. On the other hand, there are no significant excesses in Run 2 SUSY searches~\cite{ATLAS:2016kjm,CMS:2016vfu}. The SUSY analyses were based on the simplified model of sbottom decay $\tilde{b}_1 \to t +(\tilde{\chi}_1^\pm \to W^\pm +\tilde{\chi}_1^0)$. However, the branching ratio of such a decay chain is never close to 100\% due to the presence of other decays ($\tilde{b}_1 \to b +\tilde{\chi}_2^0 \mbox{ or } \tilde{\chi}_1^0$), and these other decays generally give stronger constraints as we saw in the previous section. Ref.~\cite{Huang:2015fba} proposed to explain the excesses from the right-handed stop production with the decay $\tilde{t}_R \to t+ (\tilde{B} \to W^\pm + \tilde{W}^\mp)$ where $\tilde{W}^\mp$ is closely degenerate with the neutral Wino which is assumed to be the LSP, and hence its decay products are too soft to be seen. With a suitable arrangement of the spectrum, the branching ratio of this decay chain can be close to 100\%, and the excesses could be explained by a $\tilde{t}_R$ of $\sim 550$~GeV.
The spectrum in our study provides an alternative way to explain the excesses. From Table~\ref{table:A_final}, we can see that the sbottom pair decays to the $2t+2W+\slashed{E}_T$ final state which gives the desired signals are close to 100\% in Type A models. In addition, a substantial fraction of $\tilde{t}_2$ pairs also give $ttWW+\slashed{E}_T+X$ final states which contribute to the signal. We expect that our benchmark models with light spectra may produce same-sign dilepton events compatible with the excesses observed in experiments. To minimize the systematic uncertainties in comparing with experimental excesses, we follow Ref.~\cite{Huang:2015fba} to normalize the signal strength of our benchmark points to the SM $t\bar{t}h$ signal strength, then compare the simulation results at 13 TeV to the best fit signal strengths of the new Run 2 results: $\mu=4.0^{+2.1}_{-1.7}$ of the ATLAS $2l0\tau_{had}$ signal region~\cite{ATLAS:2016ldo} and $\mu=2.7^{+1.1}_{-1.0}$ of the CMS 2LSS category for $t\bar{t}h$ searches~\cite{CMS:2016vqb}. For our benchmark model A1 (A2), we get a total signal strength of $\mu=2.4 (1.8)$, with $\mu_{\tilde{b}_1}=1.1 (0.6)$ from $\tilde{b}_1$ pairs and $\mu_{\tilde{t}_2}=0.3 (0.2)$ from $\tilde{t}_2$ pairs. They are in the ballpark of the observed excesses in Run 2. Furthermore, we also check our benchmark points with the 95\% CL upper limits on the number of SS2L events in the ATLAS SUSY search~\cite{ATLAS:2016kjm}. For our A1 (A2), we get 8.0(4.8)/0.9(0.4) events in the corresponding SR1b/SR3b signal regions where the observed 95\% CL upper limits are 10.3/4.9 .
\subsubsection{Channels with multiple leptons}
Multiple $W$'s and $Z$'s also give rise to multilepton signals.
For multilepton channels, we require 3 or more leptons with $p_T>15$~GeV and $\slashed{E}_T>100$~GeV in the final state. Since the chance of finding more than two leptons in an event is rare, we loosen the $N_b$ cut to $\geqslant 1$ and $N_j\geqslant 2$ in these cases. Each $b$ jet is required to have $p_T>30$~GeV while for other jets $p_T$ should be greater than 25~GeV. The multilepton events are then split into two different channels based on the number of $Z$-like lepton pairs. We ascribe those events with at least one $Z$-like lepton pair to the $\ell Zb$ channel, and the rest are recorded by the $3\ell b$ channel.
\begin{figure}[t]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\includegraphics[scale=0.35]{Res_Zlb_1.pdf}
\includegraphics[scale=0.35]{Res_3lb_1.pdf}
\caption{Histograms in $\slashed{E}_T$ of multiple lepton channels. We require $\geqslant1\, b$ jet and $\geqslant2$ jets in total. \textbf{Left:} Channel with at least 1 $Z$ like lepton pairs and extra lepton(s). \textbf{Right:} Channel without any $Z$ like lepton pair.}
\label{fig:multilep}
\end{center}
\end{figure}
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline\hline
Channel&Bkg Total &Model & A1& A2&A3&A4&A5&A6&B1&B2&B3&B4\\\hline
\multirow{2}{*}{$\ell Z b$}&\multirow{2}{*}{279.3}&$\tilde{t}_2\ttwo$&57.1&20.1&9.2&3.0&3.4&4.5&18.2&6.7&5.0&3.2\\
& &$\tilde{b}_1\bone$&3.1&1.6&0.5&0.2&0.2&0.2&0.9&0.9&0.5&0.2\\\hline
\multirow{2}{*}{$3 \ell b$}&\multirow{2}{*}{250.8}&$\tilde{t}_2\ttwo$&39.1&27.2&12.8&3.5&1.7&1.7&12.2&2.7&4.0&2.6\\
& &$\tilde{b}_1\bone$&66.6&23.9&13.0&3.2&2.9&4.8&18.6&0.9&1.9&5.0\\\hline
\end{tabular}
\end{center}
\caption{Background and signal numbers for each model point in $\ell Zb$ and $3\ell b$ channels with a 300 fb$^{-1}$ integrated luminosity.}
\label{table:multilep}
\end{table}
The SM backgrounds for multileptons with $b$ jets coming from $t\bar{t}+W/Z/h$ and vector boson pair produced with extra jets. The $\slashed{E}_T$ distributions of signals and backgrounds for both channels are shown in Fig.~\ref{fig:multilep}. We see that there are still substantial backgrounds for both channels. The numbers of background and signal events for the benchmark points with a 300 fb$^{-1}$ integrated luminosity are listed in Table~\ref{table:multilep}. We note that $Z$ bosons are mostly produced in $\tilde{t}_2$ decay so $\ell Zb$ signal is $\tilde{t}_2$ specific. As a result the $\ell Z b$ channel has less total signal events than the $3\ell b$ channel. Anyway, the large number of background events make either channel less effective for our benchmark models than the previous channels.
\subsubsection{Other potential channels}
As we discussed, the decay chain $\tilde{t}_2 \to W\tilde{b}_1 \to WW \tilde{t}_1 \to WWt \tilde{\chi}_1^0$ is common for a typical spectrum and has a significant branching ratio in several benchmark models (e.g., A2--A4). In this case there can be up to 6 $W$'s in the final state, which could give rise to same-sign trilepton events. The SM background would be extremely low. The signal event rate is low so one does not expect it to be the discovery channel. However, it may provide useful information about the spectrum after the discovery.
For type B models, the long decay chains may even produce multiple $t$'s in the final states. This occurs in the benchmark B4, where the $\tilde{t}_2$ decay can give 3 $t$'s and the $\tilde{b}_1$ decay can produce $ttb$. With 6 tops it can also produce same-sign trilepton events. In addition, specialized searches for multiple top final states such as Ref.~\cite{multitop} may be sensitive to this type of spectrum, although it is not expected to be the first discovery channel either.
\subsection{Significance}
From the tables of signal and background events, we can calculate the signal significances of each channel for the benchmark models. However, there are other effects which can affect the real significances of each channel. First, we need to consider the effect of systematic errors, especially when $S/B$ ratio is small, the uncertainties in the background normalization can overwhelm the signal. There are many factors which could introduce systematic errors, such as ($b$) jet tagging efficiency, detector resolutions, or PDF uncertainties, etc.. Second, if the distributions of the signal and the background on some kinematic variables are very distinct and accurately known, the significance could be enhanced by dividing each channel into several signal regions according to the values of the variables. Here we just use the likelihood method to evaluate the effects of the systematic errors from the background normalizations. We assume that the overall number of background events respects the normal distribution with a fractional uncertainty $\sigma_{B}\propto B$. The likelihood is defined to be
\begin{equation}
Q=\frac{\int \mathcal{L}(S+B,S+B')P(B')dB'}{\int \mathcal{L}(S+B,B')P(B')dB'}
\end{equation}
where $S$ and $B$ are corresponding numbers of signal and background events, $\mathcal{L}(x,\mu)=\frac{\mu^{x}e^{-\mu}}{x!}$, and $P(B)$ is the normalized normal distribution with the mean $B$ and a standard deviation $\sigma_B$. The final significance from this method is simply given by $\sqrt{2\log(Q)}$. For the case with no systematic error, $\sigma_B=0$, this equation simply reduces to the standard formula~\cite{Cowan:2010js}:
\begin{equation}
\sigma = \sqrt{\left[2(S+B)\log\left( \frac{S+B}{B} \right)-S\right]}.
\label{eq:significance}
\end{equation}
The systematic uncertainties in general depend on signal channels. Without knowing the exact numbers we will calculate the significances with the assumption of a 10\% uncertainty in background normalization for each channel and compare them with the significances obtained without systematic errors to see their effects.
The result for Type A and Type B models by combining signal events from both $\tilde{t}_2$ and $\tilde{b}_1$ are listed in Table~\ref{table:A_significance} and Table~\ref{table:B_significance} respectively. (Individual significances from $\tilde{t}_2$ or $\tilde{b}_1$ can also be easily obtained from the Tables of event numbers in the previous subsection.) The significances for an integrated luminosity different from 300 fb$^{-1}$ can be obtained by a simple rescaling.
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Significance & A1 & A2 & A3 & A4 & A5 & A6\\
\hline
$0\ell+3b$ & 11.9 (7.3)& 7.2 (4.3) & 3.2 (1.9)& 0.9 (0.6) & 1.2 (0.7) & 2.9 (1.7)\\
\hline
$1\ell+3b$ & 8.6 (7.1)& 5.1 (4.2) & 2.8 (2.4) & 1.0 (0.8)& 0.9 (0.8) & 1.4 (1.2)\\
\hline
$Z2b+V$ & 7.7 (7.4)& 2.2 (2.2) & 1.2 (1.2)& 0.4 (0.4)& 0.6 (0.6)& 1.1 (1.1) \\
\hline
$Z2b$ & 6.0 (5.7) & 2.3 (2.2) &0.8 (0.8) & 0.4 (0.4)& 0.5 (0.5) & 0.5 (0.5) \\
\hline
$SS1b$ & 8.7 (6.6) & 5.3 (4.0) &2.9 (2.2)& 0.7 (0.6) & 0.6 (0.5)& 0.8 (0.6)\\
\hline
$SS2b$ & 10.8 (9.9)& 7.1 (6.6) & 4.1 (3.8) & 1.2 (1.2)& 0.8 (0.8)& 1.3 (1.3)\\
\hline
$\ell Zb$ & 3.5 (1.8)& 1.3 (0.7) & 0.6 (0.3) & 0.2 (0.1)& 0.2 (0.1)& 0.3 (0.2) \\
\hline
$3\ell b$ & 6.3 (3.4)& 3.1 (1.7) & 1.6 (0.9) & 0.4 (0.2)& 0.3 (0.2)& 0.4 (0.2)\\
\hline
Total & 23.5 (18.7)& 13.3 (10.4) &6.9 (5.6) & 2.1 (1.8) & 2.0 (1.6)& 3.7 (2.8)\\
\hline
\end{tabular}
\end{center}
\caption{The significances for a 300 fb$^{-1}$ integrated luminosity of Type A Benchmark points from various channels. The numbers in the brackets are significances with a $10\%$ background systematic uncertainty.}
\label{table:A_significance}
\end{table}
\begin{table}[ht]
\captionsetup{justification=raggedright,
singlelinecheck=false}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
Significance & B1 & B2 & B3 & B4 \\
\hline
$0\ell+3b$ & 5.9 (3.5) & 9.4 (5.6) & 3.5 (2.1) & 5.6 (3.3)\\
\hline
$1\ell+3b$ & 4.1 (3.4) & 2.3 (2.0) & 1.6 (1.4) & 3.6 (3.0)\\
\hline
$Z2b+V$ & 3.3 (3.2)& 1.8 (1.7)& 0.9 (0.9)& 1.0 (1.0)\\
\hline
$Z2b$ & 3.1 (3.0) & 2.1 (2.0)& 0.6 (0.6)& 0.4 (0.4) \\
\hline
$SS1b$ & 2.9 (2.2) & 0.2 (0.2)& 0.5 (0.4)& 0.7 (0.5)\\
\hline
$SS2b$ & 4.7 (4.4) & 0.6 (0.6)& 1.3 (1.2)& 2.3 (2.1)\\
\hline
$\ell Zb$ & 1.1 (0.6)& 0.5 (0.3)& 0.3 (0.2) & 0.2 (0.1)\\
\hline
$3\ell b$ & 1.9 (1.0)& 0.2 (0.1)& 0.4 (0.2)& 0.5 (0.3)\\
\hline
Total & 10.4 (8.3)& 10.1 (6.6)& 4.3 (3.0)& 7.2 (5.2) \\
\hline
\end{tabular}
\end{center}
\caption{The significances for a 300 fb$^{-1}$ integrated luminosity of Type B Benchmark points from various channels. The numbers in the brackets are significances with a $10\%$ background systematic uncertainty.}
\label{table:B_significance}
\end{table}
Due to the unknown systematic uncertainties in different channels and possible improvements from dividing events into different signal regions, the numbers in Tables~\ref{table:A_significance} and \ref{table:B_significance} should not be taken literally. However, they provide a good guidance on the effectiveness of various search channels for different benchmark models. In Type A models, the same-sign dilepton signals are often the most effective channels due to low backgrounds and large fractions of final states containing multiple $W$'s. However, for A6, the channels with multi-$b$-jets have the best reach because the large decay branching fraction for $\tilde{t}_2 \to h \tilde{t}_1$ and a small cross section for the heavy $\tilde{b}_1$. For Type B models, the longer decay chains through charginos and neutralinos often produce more $Z$, $h$, $t$ and $b$ in the final states, which results in more $b$'s, as can be seen from Table~\ref{table:B_final}. Therefore, the $0l3b$ may have the best reach. Most channels receive contributions from both $\tilde{t}_2$ and $\tilde{b}_1$. On the other hand, the $Z2b(V)$ and $\ell Zb$ channels are almost stop specific and receive little contribution from the sbottom. Due to the complicated decay patterns of $\tilde{t}_2$ and $\tilde{b}_1$, a combination of different search channels not only enhances the search reach, but also provides important information about the spectrum and the decay patterns upon the discovery by comparing signals in different channels.
\section{Conclusions}
\label{sec:conclusions}
If the hierarchy problem is solved by SUSY, the stops are likely to be light enough to be accessible at the LHC. On the other hand, to obtain a Higgs boson mass of 125 GeV, a large $A_t$ term in the stop sector is needed in MSSM to increase the radiative contribution if both stops are light. The large mixing between the left-handed and right-handed stops implies a large splitting ($\gtrsim 300$ GeV most of the time for stops lighter than $\sim 1$ TeV) between the two mass eigenstate and at least a sbottom lighter than the second stop. Such a stop and sbottom spectrum often embroil complex decay chains for the stop and sbottom sector.
Even though intensive LHC searches for the lightest stop have put strong constraints on its mass, it can be hidden in the compressed region where it is difficult to have an effective search. In that case, the second stop and the sbottom may be easier to discover. The existing experimental searches are based on the simplified model approach. In particular for the second stop, only $\tilde{t}_2 \to \tilde{t}_1 Z$ and and $\tilde{t}_2 \to \tilde{t}_1 h$ decays are assumed. We consider many benchmark models for a more complete stop and sbottom spectrum and find that the simplified models are seldom good approximations to the realistic models. The decay patterns are complex and often without a dominant mode. The branching ratios of $\tilde{t}_2 \to \tilde{b}_1 W$ and $\tilde{b}_1 \to \tilde{t}_1 W$ decays can be quite substantial and they were not considered in the existing experimental searches. If there are additional charginos and neutralinos lighter than $\tilde{t}_2$ or $\tilde{b}_1$, they can appear in the decay chains and make the decay pattern even more complex.
In this paper, we perform a study of collider searches of $\tilde{t}_2$ and $\tilde{b}_1$ at 14 TeV LHC assuming that $\tilde{t}_1$ is hidden in the compressed region. The study is based on general MSSM but focuses on the stop and sbottom sector. The spectra are divided into two types, depending whether there are additional charginos and neutralinos besides the LSP below the $\tilde{t}_2$ mass. We derive the branching ratios of various decay modes and obtain the fractions of possible final states. From there we can identify potentially useful signal channels for the $\tilde{t}_2$ and $\tilde{b}_1$ searches. In additional to the standard signals based on multi-$b$-jets and leptonic decaying $Z$'s of current experimental searches, we find that same-sign dilepton and multi-lepton signals are also important for $\tilde{t}_2$ and $\tilde{b}_1$ searches, because many benchmark models produce large fractions of multi-$W$ final states. The same-sign dilepton excesses observed in the LHC Run 1 and Run 2 data may be explained by some of our benchmark models if they turn out to be real. For the standard leptonic decaying $Z$, we find that additional vector-tags and a new kinematic variable which we called ``leverage'' can further help to increase signal significance.
Due to the complex decay patterns, which signal channels are most useful for $\tilde{t}_2$ and $\tilde{b}_1$ searches depend on the models and spectra. More often than not there is no dominant decay mode and a combination of many different signal channels is needed to obtain the best reach. Also, most signal channels receive contributions from both $\tilde{t}_2$ and $\tilde{b}_1$. Some signals with a reconstructed $Z$ may be thought as more $\tilde{t}_2$ specific. However, in Type B models, $Z$ can also appear in the $\tilde{b}_1$ decay chain from $\tilde{\chi}_2^0$ or $\tilde{\chi}_3^0$ decay if they are lighter than $\tilde{b}_1$. It is therefore difficult to perform independent searches for $\tilde{t}_2$ and $\tilde{b}_1$ in a realistic scenario. Observation of signals in multiple channels and their kinematic distributions will be needed to help disentangling the ultimate underlying theory and its spectrum.
\section*{Acknowledgments}
We would like to thank Tim Cohen, Dmitri Denisov, and John Stupak for help with the Snowmass background simulations, also Thomas Hahn for FeynHiggs coding and Margarete M\"uhlleitner for the SUSYhit support. We also thank Ian Low and Angelo Monteux for discussion and email correspondences. H.-C. C. thanks Academia Sinica in Taiwan and Kavli Institute for Theoretical Physics China for hospitality where this manuscript is finished. This work is supported in part by the US Department of Energy grant DE-SC-000999.
\begin{appendix}
\section{Compatibilities of the benchmark models with current constraints from 13 TeV LHC data}
In this Appendix, we check our benchmark models against the experimental constraints from the most recent LHC 13 TeV Run 2 results. For our analyses, the most relevant search channels from current LHC public results are multiple-$b$ jets with 0 or 1 lepton and same-sign dilepton with one or more $b$-jets. Other channels are either less significant such as multi-lepton channels, or not available yet such as the $Z2b$ channel.
\subsection{Multiple $b$ + 0/1$\ell$}
For multiple $b$ + 0/1$\ell$ channels, the strongest constraints come from the searches of gluinos decaying via stop or sbottom~\cite{ATLAS:2016uzr}. In order to test the viability of our benchmark points, we adopt a cut similar to the ATLAS study:
All candidate jets should have $p_T>30$~GeV with $|\eta|<2.5$. All candidate leptons should have $p_T>20$~GeV with $|\eta|<2.5$. The 4 leading hardest jets are required to have a $\Delta\phi>0.4$ from the MET. At least 3 candidate jets should be $b$-tagged.
We define the inclusive effective invariant mass $M_{\text{eff}}$ to be the scalar sum of $p_T$ of all candidate jets and leptons and MET. The transverse mass of $b$ jets $M^b_{T,min}$ is defined to be $\text{Min}(\sqrt{2 p_{Tb} (1-\cos\Delta\phi_{i})}), i=1,2,3$, the minimum is taking among three leading $b$ jets. Finally, we define the total jet mass variable $M^{\Sigma}_J$ to be the mass sum of the 4 leading fat jets with $p_T>100$~GeV. Five signal regions used to search for $\tilde{g}\to \tilde{t}t$ are given by:
\begin{table}[h]
\begin{tabular}{|c|c|c|}
\hline
Name & Definition & ATLAS 95$\%$ CL limit\\
\hline
Gtt0LA & $N_j \geqslant$ 8, $M^b_{T,min}>80$, MET$>$400, $M_J^\Sigma>200$, $M_{\text{eff}}>2000$ &3.8\\
\hline
Gtt0LB & $N_j \geqslant$ 8, $M^b_{T,min}>80$, MET$>$400, $M_{\text{eff}}>1250$ &13.3\\
\hline
Gtt1LA & $N_j \geqslant$ 6, $M^b_{T,min}>120$, $M_T>200$, MET$>$200, $M_J^\Sigma>200$, $M_{\text{eff}}>2000$ &3.8\\
\hline
Gtt1LB & $N_j \geqslant$ 6, $M^b_{T,min}>120$, $M_T>200$, MET$>$350, $M_J^\Sigma>150$, $M_{\text{eff}}>1500$ &4.9\\
\hline
Gtt1LA & $N_j \geqslant$ 6, $N_{jb} \geqslant$ 4, $M^b_{T,min}>80$, $M_T>150$, MET$>$200, $M_{\text{eff}}>500$ &5.7\\
\hline
\end{tabular}
\end{table}
These are then applied upon our benchmark points. For A1, we expect 0.4 and 0.9 events for two 0L signal regions. For 1L + multi-$b$ channels Gtt1L-A/B and C, A1 gives 0.1, 0.3 and 0.5 signal events with 14.8 $\text{fb}^{-1}$. All of them are an order of magnitude less than the 95\% CL exclusion limit. The next lightest benchmark point A2 contributes 0.2, 0.7, 0.03, 0.08 and 0.2 to these 5 signal regions. Other Type A benchmark points yield smaller numbers. For Type B models, B1 would contribute 0.2 and 0.5 events to two 0L signal regions, and 0.06, 0.2, 0.3 events for the 1L regions. The contributions from other type B benchmark points are even less. The smallness of the number of events is partly due to the hard cuts imposed in the analyses which are designed for the gluino search. Therefore, none of our benchmark points are excluded by the multi-$b$ +0/1 lepton searches.
\subsection{Same-sign dilepton +$b$ jets}
For this signal we compare to the ATLAS same-sign dilepton SUSY search~\cite{ATLAS:2016kjm}. The most relevant signal region is SR1b and SR3b. We require each candidate jet must have $p_T>25$~GeV and $|\eta|<2.5$. Each candidate lepton need to have $p_T>10$~GeV and $|\eta|<2.5$. The signal requires a pair of same-sign lepton or more than 3 leptons.
\begin{table}[h]
\begin{tabular}{|c|c|c|}
\hline
Name & Definition & ATLAS 95$\%$ CL limit\\
\hline
SR1b & $N_j \geqslant$ 6, $N_b\geqslant1$, MET$>$200, $M_{\text{eff}}>650$ &10.3\\
\hline
SR3b & $N_j \geqslant$ 6, $N_b\geqslant3$, MET$>$160, $M_{\text{eff}}>600$ &4.9\\
\hline
\end{tabular}
\end{table}
We find that at 13~TeV and the same luminosity, the lightest benchmark A1 gives 8.0 events for SR1b and 0.9 events for SR3b, thus is not excluded. Other benchmark points produce even less events. For instance, A2 gives 4.9 and 0.4 for these two signal regions while B1 gives 2.5 and 0.2. Note that with more data, A1 might be excluded soon, while testing other benchmark points may take longer time because their contributions are well below the upper limits.
\subsection{CMS jets+MET data}
The models points can also be constrained by SUSY searches with multi-jets and MET. In particular, CMS showed a deficit in the aggregate search region SR11 in ``Additional Table 5'' of Ref.~\cite{CMS:2016mwj}. This aggregate region is targeted for a light SUSY spectrum. It requires to 7+ jets with 1+ b jets, each of them should have $p_T>30$~GeV, $H_T \geq 300$~GeV and $H_T^{\rm miss}\geq 300$~GeV. Also no isolated leptons or isolated-charged particles are found. For the two leading jets, their $\Delta\phi$ with the MET are required to be greater than 0.5, while for the third and the fourth leading jets, $\Delta\phi>0.3$ is required. The predicted number of background events is $385^{+19+27}_{-17-27}$, while the observed number is 316, showing a $\sim 2\sigma$ deficit. This can put any new physics models that contribute a significant number of events in this region in tension with the data.
As a test of our benchmark models, the lightest A1 point gives 35 events for SR11, which roughly equals the $1\sigma$ uncertainty of the predicted background events, and A2 contributes about 22 events. They are in some tension with the observed data, but not much more than the SM itself. We note that among the 12 combined signal regions, SR11 is the only one showing a significant deficit. We hope that with more statistics and a cross-check with ATLAS will clarify this situation soon.
\end{appendix}
|
\section{Introduction}
The electroweak interactions have been tested and confirmed mainly by
parity-violating lepton scattering, decays of hadrons, and
$\beta$ decays of nuclei. Recently, Parity-violating (PV) hadronic
processes play yet another important role of a touchstone to
examine the standard model (SM) and physics beyond the standard model
(BSM)~(See for example recent reviews~\cite{RamseyMusolf:2006dz,
Holstein:2009zzc, Cirigliano:2013xha, Cirigliano:2013lpa,
Haxton:2013aca, Schindler:2013yua}).
There are mainly two different ways of describing PV hadronic
reactions: One is to consider one-boson exchanges such as $\pi$,
$\rho$, and $\omega$ mesons \`{a} la the strong nucleon–nucleon
($NN$) potential~\cite{Tadic:1969xx, Donoghue:1975xk,
Desplanques:1979hn}. The other is to employ effective field
theory~\cite{Zhu:2004vw, Schindler:2013yua}. In both
methods, the PV pion-nucleon coupling constant is the most essential
quantity, since it governs the PV hadronic processes in long range .
Desplanques, Donoghue and Holstein (DDH)~\cite{Desplanques:1979hn}
estimated the value of the PV $\pi NN$ coupling constant, also known
as the so-called ``DDH best value'': $h^1_{\pi NN}=4.6 \times
10^{-7}$. A great deal of experimental and theoretical efforts has
been devoted to extract the precise value of the $\pi NN$ coupling
constant (for recent reviews, see \cite{haxton2013, schin2013}). For
example, its contribution is exclusively dominant in the PV asymmetry
in $\vec{n} p \to d\gamma$~\cite{hyun2001, hyun2005, hyun2007},
and $\vec{n} d \to t \gamma$~\cite{des1986}. The PV $\pi NN$ coupling
constant has been studied in various different theoretical approaches
such as the Skyrme models~\cite{kaiser1988, weigel1999, ks1993},
quark models~\cite{dubovik1986}, the chiral-quark soliton
model~\cite{Lee:2004tr, Lee:2012fx}, QCD sum rule~\cite{henley1998},
and so on. However, all these values of $h^1_{\pi NN}$ are far from
consensus and are given in the wide range between $10^{-8}$
\cite{kaiser1988} and $\sim 5\times 10^{-7}$ \cite{ks1993}. A recent
analysis of lattice QCD yields $h_{\pi NN}^{1,\mathrm{con}} =
(1.099\pm0.505_{-0.064}^{+0.058})\times 10^{-7}$ for which only the
contribution of the connected diagrams to $h_{\pi NN}^1$ has been
considered~\cite{wasem2012}. On the experimental side, though the
accuracy of the measurements has been much improved, an upper bound on
the value of $h^1_{\pi NN}$~\cite{Gericke:2011zz} is only known. Thus,
more systematic and quantitative studies are required in order to
obtain the value of the PV $\pi NN$ coupling constant.
The main dynamical origin of hadronic parity violation comes from the
flavor-conserving effective weak Hamiltonian, which was already
investigated~\cite{Donoghue:1975xk, Korner:1978sz,
Guberina:1978wg, Desplanques:1979hn, Karino:1980vn, Karino:1981yj,
Dai:1991bx}. In particular, the PV $\pi NN$ coupling constant can be
obtained from the isovector ($\Delta I=1$) effective weak Hamiltonian,
which was first derived in Ref.~\cite{Dai:1991bx} at the one-loop
level with the effects of heavy quarks taken into account. Very
recently, Tiburzi~\cite{Tiburzi:2012hx} investigated systematically
the $\Delta S=0$ effective weak Hamiltonian with QCD corrections at
next-to-leading order (NLO). The effects of the NLO corrections have
changed the Wilson coefficients about $(10-20)\,\%$ at the typical scale
of light hadrons ($\mu=1\,\mathrm{GeV}$). Considering the fact that
the PV $\pi NN$ coupling constant is very tiny, we expect that
the corrections from QCD at NLO may come into play. Thus, it is of
great interest to examine the NLO corrections to the PV $\pi NN$
coupling constant.
In the present work, we investigate the PV $\pi NN$ coupling constant,
$h^1_{\pi NN}$, within the framework of the chiral quark-soliton model
($\chi$QSM) together with the effective weak Hamiltonian at
NLO~\cite{Tiburzi:2012hx}. Recently, the present authors computed the
PV $\pi NN$ coupling constant~\cite{Lee:2012fx}
in the same framework, employing the effective weak Hamiltonian from
Ref.~\cite{Desplanques:1979hn}. We first derived the effective weak
chiral Lagrangian, based on the nonlocal chiral-quark model
(N$\chi$QM) from the instanton vacuum associating with the effective
weak Hamiltonian~\cite{Lee:2004tr}. If one performs the gradient
expansion for the effective chiral action of the $\chi$QSM with the
effective weak Hamiltonian, we would obtain exactly the same
expressions starting directly from the effective weak chiral
Lagrangian. Using this gradient expansion, we were able to obtain
the PV $\pi NN$ coupling constant to be about
$1\times 10^{-8}$ at $\mu=1$ GeV. We also found that the $h_{\pi
NN}^1$ is rather sensitive to the Wilson coefficients. In this
respect, it is of great importance to reexamine the PV $\pi NN$
coupling constant, the $\Delta I=1$ effective weak Hamiltonian being
employed with the NLO QCD effects. As we will show in this work, the
value of $h_{\pi NN}^1$ indeed turns out to be different from the
previous result. Moreover, the effects from the next-to-leading-order
corrections reduce the reading-order result by about $20~\%$.
The paper is organized in the following order: In Section
\ref{sec:weaklag}, we present briefly the general procedure to obtain
the PV $\pi NN$ coupling constants within the $\chi$QSM. We
first derive the flavor-conserving effective weak chiral Lagrangian,
starting from the nonlocal chiral quark model from the instanton
vacuum. In Section III we compute the correlation function corresponding to the PV
$\pi NN$ coupling constant. In Section \ref{sec:con}, we discuss the
result, and conclude the work.
\section{$\Delta I=1$ Effective weak chiral
Lagrangian}\label{sec:weaklag}
We start with the $\Delta I=1$ flavor-conserving effective weak
Hamiltonian including the NLO corrections~\cite{Tiburzi:2012hx}, which
is expressed as
\begin{align}
\mathcal{H}_{W}^{\Delta I=1} = \frac{G_F}{\sqrt{2}}
\frac{\sin^2\theta_W}{3}\sum_{i=1}^{8}c_i(\mu) O_i(\mu)\ ,
\end{align}
where $G_F$ and $\sin\theta_W$ denote the Fermi constant and the
Weinberg angle, respectively. The eight different operators $O_i$ are
defined generically as two-body operators:
$O_i=(\bar{\psi}\gamma_\mu\gamma_5 M \psi) (\bar{\psi} \gamma_\mu N
\psi)$. The $c_i$ stands for the Wilson coefficient corresponding to
the $O_i$, which depends on the renormalization scale
$\mu$. Introducing the Gell-Mann matrices in SU(3) flavor space, we
can write $O_i$ as
\begin{align}
O_1&=\sqrt{\frac{1}{3}} ({\bar \psi}\gamma_\mu\gamma_5\lambda_3\psi)
\bigg({\bar \psi} \gamma_\mu \bigg[\sqrt{2} \lambda_0 +
\lambda_8\bigg]\psi\bigg)\ , \cr
O_2&= \sqrt{\frac{1}{3}}{\bar \psi}^a \gamma_\mu\gamma_5 \lambda_3\psi^b
{\bar \psi}^b\gamma_\mu \bigg[\sqrt{2} \lambda_0+\lambda_8
\bigg]\psi^a\ , \cr
O_3& = \sqrt{\frac{1}{3}} ({\bar \psi}\gamma_\mu\lambda_3\psi)
\bigg({\bar \psi}\gamma_\mu \gamma_5 \bigg[\sqrt{2}\lambda_0 +
\lambda_8\bigg]\psi\bigg)\ , \cr
O_4&=\sqrt{\frac{1}{3}}{\bar \psi}^a\gamma_\mu\lambda_3\psi^b
{\bar \psi}^b\gamma_\mu\gamma_5 \bigg[\sqrt{2} \lambda_0 +
\lambda_8\bigg]\psi^a\ , \cr
O_5&=\sqrt{\frac{1}{6}}({\bar \psi}\gamma_\mu\gamma_5\lambda_3\psi)
\bigg({\bar \psi}\gamma_\mu \bigg[\lambda_0- \sqrt{2}\lambda_8
\bigg]\psi\bigg)\ , \cr
O_6&=\sqrt{\frac{1}{6}}{\bar \psi}^a \gamma_\mu \gamma_5 \lambda_3\psi^b
{\bar \psi}^b\gamma_\mu \bigg[\lambda_0-\sqrt{2}
\lambda_8\bigg]\psi^a\ , \cr
O_7&=\sqrt{\frac{1}{6}}({\bar \psi}\gamma_\mu\lambda_3\psi)
\bigg({\bar \psi}\gamma_\mu \gamma_5 \bigg[\lambda_0 -
\sqrt{2}\lambda_8\bigg]\psi\bigg)\ ,
\cr
O_8&=\sqrt{\frac{1}{6}}{\bar \psi}^a \gamma_\mu \lambda_3\psi^b
{\bar \psi}^b\gamma_\mu \gamma_5
\bigg[\lambda_0-\sqrt{2}\lambda_8\bigg]\psi^a\ ,
\label{eq:twobody}
\end{align}
where $\lambda_0$, $\lambda_3$, and $\lambda_8$ are the Gell-Mann
matrices represented in flavor SU(3) space as
$\lambda_0=\sqrt{2/3}\,\mathrm{diag}(1,\,1,\,1)$,
$\lambda_3=\mathrm{diag}(1,\,-1,\,0)$, and
$\lambda_8=\sqrt{1/3}\,\mathrm{diag}(1,\,1,\,-2)$, respectively.
The quark field is given as a triplet in flavor SU(3)
\begin{align}
\psi = \left(\begin{array}{c} u \\ d \\ s \end{array} \right),
\nonumber
\end{align}
where $u$, $d$ and $s$ represent the up, down and strange quark
fields, respectively. The repeated indices $a$ and $b$ designate the
color-singlet contraction and the parentheses $(\bar{\psi}\Gamma \psi)$
without showing the color indices are already color-singlet contracted.
Applying the following Fiertz identity to $O_2, O_4, O_6,$ and $O_8$,
\begin{align}
\delta_{bc}\delta_{ad}=\frac{1}{2}(t^A)_{ab}(t^A)_{cd}
\end{align}
where $t^A$ denote the Gell-Mann matrices in color space,
we are able to express the effective weak Hamiltonian in the following
form
\begin{align}
\mathcal{H}_{W}^{\Delta I=1} &=\frac{G_F}{\sqrt{6}}
\frac{\sin^2\theta_W}{3} \bigg\{({\bar
\psi}\gamma_\mu \gamma_5\lambda_3\psi)
\bigg({\bar \psi}\gamma_\mu\bigg[\lambda_0 \bigg(\sqrt{2}c_1 +
\frac{c_5}{\sqrt{2}} \bigg) + \lambda_8(c_1-c_5)
\bigg]\psi\bigg)\cr
&+\frac{1}{2}({\bar \psi}\gamma_\mu\gamma_5 \lambda_3t^A\psi)
\bigg({\bar \psi}\gamma_\mu\bigg[\lambda_0
\bigg(\sqrt{2}c_2+\frac{c_6}{\sqrt{2}}\bigg)
+\lambda_8(c_2-c_6)\bigg]t^A\psi\bigg) \cr
&+({\bar \psi}\gamma_\mu\lambda_3\psi)
\bigg({\bar \psi}\gamma_\mu\gamma_5 \bigg[\lambda_0
\bigg(\sqrt{2}c_3+\frac{c_7}{\sqrt{2}}\bigg)
+\lambda_8(c_3-c_7)\bigg]\psi\bigg) \cr
&+\frac{1}{2}({\bar \psi}\gamma_\mu\lambda_3t^A\psi)
\bigg({\bar \psi}\gamma_\mu\gamma_5\bigg[\lambda_0 \bigg(\sqrt{2}c_4 +
\frac{c_8}{\sqrt{2}}\bigg) +\lambda_8(c_4-c_8) \bigg] t^A\psi \bigg)
\bigg\}.
\end{align}
We rewrite the Hamiltonian in terms of the \textit{effective}
four-quark operators that contain already the Wilson coefficients
$\mathcal{Q}_i(z;\mu)$
\begin{align}
\mathcal{H}_{W}^{\Delta I=1} =
\frac{G_F}{\sqrt{6}}\frac{\sin^2\theta_W}{3}\left(
\mathcal{Q}_1+\mathcal{Q}_2+\mathcal{Q}_3+\mathcal{Q}_4\right),
\end{align}
where the four-quark operators $\mathcal{Q}_i(z;\mu)$ are defined as
\begin{align}
\mathcal{Q}_i(z;\mu) = \alpha_i \left( \bar{\psi} \Gamma^{(i)}_1 \psi \right)
\left( \bar{\psi} \Gamma^{(i)}_2 \psi \right),
\end{align}
where $\alpha_i=1$ for $i=1,\, 3$ and $\alpha_i = 1/2$ for $i=2,\, 4$.
The $\Gamma^{(i)}_j$ are defined as
\begin{align}
\Gamma_1^{(1)}&=\gamma_\mu\gamma_5\lambda_3, \;\;
\Gamma_2^{(1)}=\gamma_\mu\Lambda^{(1)}, \;\;\;\;\;\;\;
\Gamma_1^{(2)}=\gamma_\mu\gamma_5\lambda_3 t^A, \;\;
\Gamma_2^{(2)}=\gamma_\mu\Lambda^{(2)}t^A, \cr
\Gamma_1^{(3)}&=\gamma_\mu\lambda_3, \;\;\;\;\;
\Gamma_2^{(3)}=\gamma_\mu\gamma_5\Lambda^{(3)},
\;\;\;\; \Gamma_1^{(4)}=\gamma_\mu\lambda_3t^A,
\;\;\;\;\;\,
\Gamma_2^{(4)}=\gamma_\mu\gamma_5\Lambda^{(4)}t^A
\label{eq:Gamma}
\end{align}
with flavor matrices defined as
\begin{align}
\Lambda^{(1)} &= \lambda_0\bigg(\sqrt{2}c_1+\frac{c_5}{\sqrt{2}}\bigg)
+\lambda_8(c_1-c_5), \;\;\;\;
\Lambda^{(2)}=\lambda_0
\bigg(\sqrt{2}c_2+\frac{c_6}{\sqrt{2}}\bigg)+\lambda_8(c_2-c_6), \cr
\Lambda^{(3)}& = \lambda_0
\bigg(\sqrt{2}c_3+\frac{c_7}{\sqrt{2}}\bigg) +
\lambda_8(c_3-c_7),\;\;\;\;
\Lambda^{(4)} = \lambda_0 \bigg(\sqrt{2}c_4 +
\frac{c_8}{\sqrt{2}}\bigg) + \lambda_8(c_4-c_8)\ .
\end{align}
In order to compute the $\Delta I=1$ flavor-conserving effective weak
chiral Lagrangian, we employ the N$\chi$QM from the instanton
vacuum. The effective weak chiral Lagrangian is defined as a
vacuum expectation value (VEV) of the effective weak
Hamiltonian~\cite{Franz:1999wr, Franz:1999ik}
\begin{align}
\mathcal{L}_W^{\Delta I=1} = \int D \psi D
\psi^\dagger \mathcal{H}_W^{\Delta I=1} \exp \left[
\int d^4 z\,\psi^\dagger(z) \mathcal{D} \psi(z) \right],
\end{align}
where $\mathcal{D}$ represents the nonlocal covariant Dirac operator
defined as
\begin{align}
\mathcal{D}(-i\partial) \equiv i \gamma_\mu \partial_\mu +
i\sqrt{M(-i\partial)} U^{\gamma_5}(x) \sqrt{M(-i\partial)},
\end{align}
where $U^{\gamma_5}$ represents the chiral field defined as
\begin{equation}
U^{\gamma_5} \;=\; \frac{1+\gamma_5}{2} U + \frac{1-\gamma_5}{2}
U^\dagger
\end{equation}
with the Goldstone boson field $U=\exp(i\lambda^a \pi^a/f_\pi)$.
Then, the flavor-conserving effective weak chiral Lagrangian can be
expressed in terms of the VEV of the four-quark operator
\begin{align}
{\cal L}_{\mathrm{eff}} = \frac{G_F}{\sqrt{6}} \;
\frac{\sin^2\theta_W}{3} \sum_{i=1}^4\langle \mathcal{Q}_i \rangle.
\end{align}
We refer to Refs.~\cite{Franz:1999wr, Franz:1999ik, Lee:2004tr} for
details of how to compute the VEV of $\mathcal{Q}_i$.
The flavor-conserving effective weak chiral Lagrangian in the $\Delta
I=1$ channel is obtained in terms of the low-energy constants
$\mathcal{N}_i$ and $\mathcal{M}_i$
\begin{align}
\mathcal{L}_{\mathrm{eff}}^{\Delta I=1} & =
\mathcal{N}_1\langle(R_\mu-L_\mu)\lambda_3 \rangle
\langle(R_\mu+L_\mu)\lambda_0\rangle
+ \mathcal{N}_2\langle(R_\mu-L_\mu)\lambda_3\rangle
\langle(R_\mu+L_\mu)\lambda_8\rangle
\cr
&+\mathcal{N}_3\langle(R_\mu-L_\mu)\lambda_0\rangle
\langle(R_\mu+L_\mu)\lambda_3\rangle
+\mathcal{N}_4 \langle(R_\mu-L_\mu)\lambda_8\rangle
\langle(R_\mu+L_\mu)\lambda_3\rangle
\cr
& + \mathcal{N}_{5} \langle \lambda_3 U\lambda_0U^\dagger - \lambda_3
U^\dagger\lambda_0U \rangle
+ \mathcal{N}_{6} \langle \lambda_3
U\lambda_8U^\dagger-\lambda_3 U^\dagger\lambda_8U \rangle
\cr
& + \mathcal{N}_{7} \langle L_\mu\lambda_3 L_\mu U^\dagger\lambda_0U
-R_\mu\lambda_3 R_\mu U\lambda_0 U^\dagger \rangle
\cr
& + \mathcal{N}_{8} \langle L_\mu\lambda_3 L_\mu U^\dagger\lambda_8U
-R_\mu\lambda_3 R_\mu U\lambda_8U^\dagger \rangle
\cr
& + \mathcal{N}_{9} \langle \left(\lambda_3 R_\mu R_\mu
+ R_\mu R_\mu\lambda_3\right) U\lambda_0U^\dagger
-\left(\lambda_3 L_\mu L_\mu
+L_\mu L_\mu \lambda_3\right) U^\dagger\lambda_0U \rangle
\cr
& + \mathcal{N}_{10} \langle \left(\lambda_3 R_\mu R_\mu
+ R_\mu R_\mu\lambda_3 \right) U\lambda_8U^\dagger
-\left(\lambda_3 L_\mu L_\mu
+L_\mu L_\mu \lambda_3 \right) U^\dagger\lambda_8U \rangle
\cr
& + \mathcal{N}_{11} \langle (R_\mu\lambda_3 R_\mu-L_\mu\lambda_3
L_\mu) \lambda_0\rangle
+ \mathcal{N}_{12} \langle (R_\mu\lambda_3 R_\mu-L_\mu\lambda_3 L_\mu)
\lambda_8\rangle,
\cr
& + \mathcal{M}_1 \langle \lambda_3 U \lambda_0U^\dagger - \lambda_3
U^\dagger \lambda_0U \rangle
+ \mathcal{M}_2 \langle \lambda_3 U\lambda_8U^\dagger - \lambda_3
U^\dagger \lambda_8U \rangle
\cr
&+ \mathcal{M}_3 \langle L_\mu\lambda_3 L_\mu U^\dagger\lambda_0U
-R_\mu\lambda_3 R_\mu U\lambda_0U^\dagger\rangle
+ \mathcal{M}_4 \langle L_\mu\lambda_3 L_\mu U^\dagger\lambda_8U
-R_\mu\lambda_3 R_\mu U\lambda_8U^\dagger\rangle
\cr
& + \mathcal{M}_5 \langle \left(\lambda_3 R_\mu R_\mu
+ R_\mu R_\mu\lambda_3 \right) U\lambda_0U^\dagger
-\left( \lambda_3 L_\mu L_\mu
+L_\mu L_\mu \lambda_3 \right) U^\dagger\lambda_0U \rangle
\cr
& + \mathcal{M}_{6} \langle \left(\lambda_3 R_\mu R_\mu
+ R_\mu R_\mu\lambda_3 \right) U\lambda_8U^\dagger
-\left(\lambda_3 L_\mu L_\mu
+L_\mu L_\mu \lambda_3 \right) U^\dagger\lambda_8U \rangle
\cr
& + \mathcal{M}_{7} \langle (R_\mu\lambda_3 R_\mu-L_\mu\lambda_3
L_\mu) \lambda_0 \rangle
+ \mathcal{M}_{8} \langle (R_\mu\lambda_3 R_\mu-L_\mu\lambda_3 L_\mu)
\lambda_8 \rangle
\label{eq:ExL1}
\end{align}
where $\langle \cdots \rangle$ means the trace over the flavor. The
right and left currents $R_\mu$ and $L_\mu$ are defined respectively
as
\begin{align}
R_\mu = i U \partial_\mu U^\dagger,\,\,
L_\mu = i U^\dagger \partial_\mu U\ .
\end{align}
The weak low-energy constants (WLECs) $\mathcal{N}_i$ are the leading
order in the large $N_c$ limit whereas $\mathcal{M}_i$ are of the
subleading order. They are expressed as
\begin{align}
\mathcal{N}_1 &= 4 N_c^2 J_2^2 \mathcal{C} \left(
\sqrt{2}c_1+\frac{c_5}{\sqrt{2}}\right),\;\;\;\;
\mathcal{N}_2 = 4 N_c^2 J_2^2 \mathcal{C}\left(c_1-c_5 \right),\cr
\mathcal{N}_3 & = 4 N_c^2 J_2^2 \mathcal{C} \left(\sqrt{2}c_3 +
\frac{c_7}{\sqrt{2}}\right), \;\;\;\;
\mathcal{N}_4 = 4 N_c^2 J_2^2 \mathcal{C} \left(c_3-c_7\right), \
\cr
\mathcal{N}_5 & = 8 N_c^2 J_1^2 \mathcal{C} \left(\sqrt{2}c_2 - \sqrt{2}c_4 +
\frac{c_6}{\sqrt{2}}-\frac{c_8}{\sqrt{2}}\right),\;\;\;\;
\mathcal{N}_6 = 8 N_c^2 J_1^2 \mathcal{C}\left(c_2-c_4-c_6+c_8\right),
\cr
\mathcal{N}_7 & = 16 N_c^2 J_1 J_3 \mathcal{C} \left(\sqrt{2}c_2 -
\sqrt{2}c_4 + \frac{c_6}{\sqrt{2}}-\frac{c_8}{\sqrt{2}}\right),\;\;\;\;
\mathcal{N}_8 = 16 N_c^2 J_1 J_3
\mathcal{C}\left(c_2-c_4-c_6+c_8\right),
\cr
\mathcal{N}_9 & = 8 N_c^2 J_1J_4 \mathcal{C} \left(\sqrt{2}c_2 - \sqrt{2}c_4 +
\frac{c_6}{\sqrt{2}}-\frac{c_8}{\sqrt{2}}\right),\;\;\;\;
\mathcal{N}_{10} = 8 N_c^2 J_1 J_4
\mathcal{C}\left(c_2-c_4-c_6+c_8\right),
\cr
\mathcal{N}_{11} & = 4 N_c^2 J_2^2 \mathcal{C}\left(\sqrt{2}c_2 +
\sqrt{2}c_4 +
\frac{c_6}{\sqrt{2}}+\frac{c_8}{\sqrt{2}}\right),
\;\;\;\;
\mathcal{N}_{12} = 4 N_c^2 J_2^2 \mathcal{C}
\left(c_2+c_4-c_6-c_8\right), \cr
\mathcal{M}_1 & = 8 N_c J_1^2 \mathcal{C} \left(
\sqrt{2}c_1 - \sqrt{2}c_3 +
\frac{c_5}{\sqrt{2}}-\frac{c_7}{\sqrt{2}}\right), \;\;\;\;
\mathcal{M}_2 = 8 N_c J_1^2 \mathcal{C}\left(c_1-c_3-c_5+c_7\right),
\cr
\mathcal{M}_3 & = 16 N_c J_1 J_3 \mathcal{C} \left(
\sqrt{2}c_1 - \sqrt{2}c_3 +
\frac{c_5}{\sqrt{2}}-\frac{c_7}{\sqrt{2}}\right), \;\;\;\;
\mathcal{M}_4 = 16 N_c J_1 J_3 \mathcal{C}\left(c_1-c_3-c_5+c_7\right),
\cr
\mathcal{M}_5 & = 8 N_c J_1 J_4 \mathcal{C} \left(
\sqrt{2}c_1 - \sqrt{2}c_3 +
\frac{c_5}{\sqrt{2}}-\frac{c_7}{\sqrt{2}}\right), \;\;\;\;
\mathcal{M}_6 = 8 N_c J_1 J_4 \mathcal{C}\left(c_1-c_3-c_5+c_7\right),
\cr
\mathcal{M}_7 & = 4 N_c J_2^2 \mathcal{C} \left( \sqrt{2}c_1 +
\sqrt{2}c_3 + \frac{c_5}{\sqrt{2}}+\frac{c_7}{\sqrt{2}}\right),\;\;\;\;
\mathcal{M}_8 = 4 N_c J_2^2 \mathcal{C} \left(c_1+c_3-c_5-c_7\right),
\label{eq:LECs}
\end{align}
where the integrals $J_1$, $J_2$, $J_3$, and $J_4$ are defined
respectively as
\begin{align}
J_{1}&=-\int \frac{d^4 k}{(2\pi)^4}\ \frac{M(k)}{k^2+M^2(k)}
=\frac{\left\langle\overline{\psi}{\psi}\right\rangle_{M}}{4N_c},
\cr
J_{2}&=\int \frac{d^4 k}{(2\pi)^4}\ \frac{M^2(k)
-k^2M(k) \tilde{M}^\prime + k^4 M'^2(k)}{(k^2+M^2(k))^2} =
\frac{f_\pi^2}{4N_c},
\cr
J_{3}&=\int \frac{d^4 k}{(2\pi)^4}\
\left[\frac{\frac{1}{4}\tilde{M}^{\prime\prime}k^2
+\frac{1}{2}\tilde{M}^\prime-\frac{\tilde{M}^{\prime2}}{8M}k^2}{k^2+M^2(k)}
\right. \cr
&-\left. \frac{M+M^2\tilde{M}^\prime +
\frac{k^2}{2}M^2\tilde{M}^{\prime\prime} +
\frac{1}{2}k^2M\tilde{M}^{\prime2}
+\frac{k^2}{4}\tilde{M}^\prime}{(k^2+M^2(k))^2}
+k^2\frac{\frac{1}{2}M+2M^2\tilde{M}^\prime
+M^3\tilde{M}^{\prime2}}{(k^2+M^2(k))^3}\right],\cr
J_4&=\int \frac{d^4 k}{(2\pi)^4}\
\frac{-M^3+k^2M^2\tilde{M}^\prime}{(k^2+M^2(k))^3}.
\label{eq:coefficient4}
\end{align}
$M(k)$ represents the momentum-dependent quark mass, and $\tilde{M}'$
and $\tilde{M}''$ are defined as
\begin{align}
\tilde{M}'\equiv \frac{M'}{2|k|}\ ,\ \
\tilde{M}''\equiv \frac{1}{4|k|^3}(M''|k|-M') .
\end{align}
The $\mathcal{C}$ contains the Fermi constant and the Weinberg angle
\begin{align}
\mathcal{C} = \frac{G_F}{\sqrt{6}} \frac{\sin^2 \theta_W}{3}.
\end{align}
To compute the WLECs in Eq.(\ref{eq:LECs}), we use the
momentum-dependent quark mass derived from the instanton
vacuum~\cite{Diakonov:1985eg}
\begin{align}
M(k) = M_0 F^2(k\rho)
\end{align}
with
\begin{align}
F(k\rho) = 2 z \left( I_0(z)K_1(z) - I_1(z) K_0(z) - \frac{1}{z}
I_1(z) K_1(z) \right),
\end{align}
where $I_i$ and $K_i$ are the modified Bessel functions, and
$z =k/2\Lambda$.
The value of the
dynamical quark mass at the zero virtuality of the quark is also
obtained from the instanton vacuum, i.e. $M_0=350\,\mathrm{MeV}$,
given the average size of the instanton and the interdistance between
instantons $R\approx 1\,\mathrm{fm}$~\cite{Diakonov:1985eg}.
The parameter $\Lambda$ is determined to reproduce the physical
value of $f_\pi$ through Eq.~(\ref{eq:coefficient4}).
As discussed already in Ref.~\cite{Franz:1999ik}, the vector and
axial-vector currents are not conserved in the presence of the
nonlocal interaction arising from the momentum-dependent quark
mass, that is, the corresponding gauge symmetries are broken. In order
to keep the currents conserved, we need to make the effective chiral
action gauge-invariant. In Ref.~\cite{Musakhanov:2002xa, Kim:2004hd},
the gauged effective chiral action was derived, based on the instanton
vacuum. Had we naively computed $J_2=f_\pi^2/4N_c$ without the current
conservation being considered, then we would have ended up with
the Pagels-Stokar formula for $f_\pi^2$~\cite{Pagels:1979hd}, which
does not satisfy the gauge invariance. The numerical
results for the integrals given in Eq.(\ref{eq:coefficient4}) are
obtained as
\begin{align}
J_1=(-112.31)^3\,\mathrm{MeV}^3, \;\;\;\;
J_2=(26.673)^2\,\mathrm{MeV}^2,\;\;\;\;
J_3=-1.7403\,\mathrm{MeV},\;\;\;\;
J_4=-0.601\,\mathrm{MeV}.
\label{eq:integrals}
\end{align}
Note that the value of $J_1$ is related to that of the quark
condensate $\langle \overline{\psi}\psi\rangle_M
=(-257.13\,\mathrm{MeV})^3$ and that of $J_2$ corresponds to the value
of $f_\pi=92.4$ MeV.
\begin{table}[htp]
\centering
\begin{tabular}{ccccc} \hline\hline
& LO & NLO ($Z$) & NLO ($Z+W$) &KS \\ \hline
$c_1$ & $\phantom{-}0.264$ & $-0.054$ & $-0.055$ & $\phantom{-}0.403$
\\
$c_2$ & $\phantom{-}0.981$ & $\phantom{-}0.803$ & $\phantom{-}0.810$
& $\phantom{-}0.765$ \\
$c_3$ & $-0.592$ & $-0.629$ & $-0.627$ & $-0.463$ \\
$c_4$ & $0$ & $0$ & $0$ & $0$ \\ \hline
$c_5$ & $\phantom{-}5.97$ & $\phantom{-}4.85$ & $\phantom{-}5.09$ &
$\phantom{-}5.61$ \\
$c_6$ & $-2.30$ & $-2.14$ & $-2.55$ & $-1.90$ \\
$c_7$ & $\phantom{-}5.12$ & $\phantom{-}4.27$ & $\phantom{-}4.51$ &
$\phantom{-}4.74$ \\
$c_8$ & $-3.29$ & $-2.94$ & $-3.36$ & $-2.67$ \\
\hline \hline
\end{tabular}
\caption{The Wilson coefficients $c_i$ derived from
Ref.~\cite{Tiburzi:2012hx}. The last column denoted by KS lists the
values derived in Ref.~\cite{ks1993} at the one-loop level.}
\label{tab:cs}
\end{table}
In Table~\ref{tab:cs}, the values of the Wilson coefficients are
listed. Those in the first three columns are taken from
Ref.~\cite{Tiburzi:2012hx}. The first column lists the results for the
Wilson coefficients in the LO, whereas the second and third ones
correspond to those from the NLO contributions together with the
LO terms. The $Z$ and $Z+W$ in the second and
third columns stand respectively for the considerations of $Z$ and
$Z+W$ boson exchanges. The last column presents those at the one-loop
level from Ref.~\cite{ks1993}. As already discussed in
Ref.~\cite{Tiburzi:2012hx}, there are certain effects from the NLO
contributions.
\begin{table}[htp]
\centering
\begin{tabular}{ccccc} \hline\hline
WLEC & LO & NLO($Z$) & NLO($Z+W$) & KS \\ \hline
${\cal N}_1$ & $\phantom{-}0.307$ & $\phantom{-}0.224$ &
$\phantom{-}0.235$ & $\phantom{-}0.303$ \\
${\cal N}_2$ & $-0.381$ & $-0.328$ & $-0.344$ & $-0.348$ \\
${\cal N}_3$ & $\phantom{-}0.186$ & $\phantom{-}0.142$ &
$\phantom{-}0.154$ & $\phantom{-}0.180$ \\
${\cal N}_4$ & $-0.382$ & $-0.327$ & $-0.343$ & $-0.348$ \\
${\cal N}_5$ & $\phantom{-}1.106$ & $\phantom{-}0.901$ &
$\phantom{-} 0.910$ & $\phantom{-}0.861$ \\
${\cal N}_6$ & $-0.005$ & $\phantom{-} 0.002$ & $0$ & $-0.003$ \\
${\cal N}_7$ & $\phantom{-} 2.716$ & $\phantom{-} 2.214$ &
$\phantom{-} 2.236$ & $\phantom{-} 2.117$ \\
${\cal N}_8$ & $-0.012$ & $\phantom{-}0.004$ & $0$ & $-0.007$ \\
${\cal N}_9$ & $\phantom{-} 0.469$ & $\phantom{-} 0.382$ &
$\phantom{-} 0.386$ & $\phantom{-} 0.365$ \\
${\cal N}_{10}$ & $-0.002$ & $\phantom{-}0.001$ & $0$ & $-0.001$ \\
${\cal N}_{11}$ & $-0.171$ & $-0.164$ & $-0.203$ & $-0.144$ \\
${\cal N}_{12}$ & $\phantom{-} 0.439$ & $\phantom{-} 0.393$ &
$\phantom{-} 0.449$ & $\phantom{-} 0.356$
\\ \hline
${\cal M}_1$ & $\phantom{-}0.320$ & $\phantom{-}0.216$ &
$\phantom{-}0.215$ & $\phantom{-} 0.325$ \\
${\cal M}_2$ & $\phantom{-}0.001$ & $-0.001$ & $-0.001$ & $-0.001$ \\
${\cal M}_3$ & $\phantom{-}0.786$ & $\phantom{-}0.531$ &
$\phantom{-} 0.529$ & $\phantom{-} 0.798$ \\
${\cal M}_4$ & $\phantom{-}0.003$ & $-0.002$ & $-0.004$ & $-0.002$ \\
${\cal M}_5$ & $\phantom{-}0.136$ & $\phantom{-}0.092$ &
$\phantom{-}0.091$ & $\phantom{-}0.138$ \\
${\cal M}_6$ & $\phantom{-}0.0004$ & $-0.0004$ & $-0.001$ & $-0.0003$ \\
${\cal M}_7$ & $\phantom{-}0.164$ & $\phantom{-}0.122$ &
$\phantom{-}0.130$ & $\phantom{-}0.161$ \\
${\cal M}_8$ & $-0.254$ & $-0.218$ & $-0.229$ & $-0.232$ \\
\hline\hline
\end{tabular}
\caption{The results for the low-energy constants in units of in
$10^{-4}~{\rm MeV}^2$. Notations are the same as in
Table~\ref{tab:cs}.}
\label{tab:lecs}
\end{table}
Based on these values of the Wilson coefficients, we
list in Table~\ref{tab:lecs} the results for the WLECs given in
Eq.(\ref{eq:LECs}). Note that the WLECS ${\cal N}_6$, ${\cal N}_8$, and ${\cal N}_{10}$
are null. This is due to the fact that they correspond to the
operators
\begin{align}
O_2'= -O_2 + O_4 + O_6 - O_8,
\end{align}
for which the corresponding Wilson coefficient vanishes because
$O_2'$ is not generated by QCD radiative
corrections~\cite{Tiburzi:2012hx}.
Though there are arguments that sizable contributions in $\Delta I=1$
channel come from the operators with strangeness~\cite{ks1993}, we
will restrict ourselves to the case of SU(2). The calculation in SU(2)
has several merits in particular in the present work. Firstly, the
chiral solitonic approach in SU(2) is much simpler and physically
clearer than that in SU(3). Secondly, the SU(2) approach allows one to
understand better the PV $\pi NN$ constant based on the effective weak
Hamiltonian. A more quantitative work within SU(3) will appear
elsewhere. In the case of SU(2), we reduce $\lambda_0$, $\lambda_3$, and
$\lambda_8$ to
\begin{align}
\lambda_0~\rightarrow~\sqrt{\frac{2}{3}}\bm{1},\;\;\;\;
\lambda_3~\rightarrow~\tau_3,\;\;\;\;
\lambda_8~\rightarrow~\frac{1}{\sqrt{3}}\bm{1}.
\end{align}
As a result, the $\Delta I=1$ effective weak Lagrangian is simplified
as
\begin{align}
\mathcal{L}_{\mathrm{eff}}^{\mathrm{SU(2)}}
&= \beta_1 \langle(R_\mu-L_\mu)\tau_3\rangle \langle
R_\mu+L_\mu\rangle
+ \beta_2
\langle R_\mu-L_\mu\rangle \langle(R_\mu+L_\mu)\tau_3\rangle
\cr
& + \beta_3
\langle (R_\mu\ R_\mu-L_\mu L_\mu)\tau_3\rangle,
+ \beta_4 \langle (R_\mu\ R_\mu-L_\mu L_\mu)\tau_3\rangle,
\label{eq:su2Lag}
\end{align}
where $\beta_i$ are defined in terms of the WLECs
\begin{align}
\beta_1 &= \frac1{\sqrt{3}}\left(\sqrt{2} \mathcal{N}_1 +
\mathcal{N}_2\right), \;\;\;\;
\beta_2 = \frac1{\sqrt{3}}\left(\sqrt{2} \mathcal{N}_3 +
\mathcal{N}_4\right),\cr
\beta_3 &= \frac1{\sqrt{3}}\left[\sqrt{2} \mathcal{N}_{11} + \mathcal{N}_{12}
+ \sqrt{2}(2\mathcal{N}_9 -\mathcal{N}_7) +
2 \mathcal{N}_{10} - \mathcal{N}_8 \right],\cr
\beta_4 &= \frac{1}{\sqrt{3}}\left[\sqrt{2}\mathcal{M}_7 + \mathcal{M}_8
+ \sqrt{2}(2\mathcal{M}_5 -\mathcal{M}_3) +
2\mathcal{M}_{6} - \mathcal{M}_4 \right].
\label{eq:betas}
\end{align}
As will be shown soon, $\beta_1$ and $\beta_2$ do not
contribute at all to the PV $\pi NN$ coupling constant. On the other
hand, $\beta_3$ and $\beta_4$ do come into play, so that we need to
examine them in detail. We can explicitly express $\beta_3$ and
$\beta_4$ in terms of the Wilson coefficients such that we can see
which terms contribute dominantly to the PV $\pi NN$ coupling
constant. $\beta_3$ and $\beta_4$ are rewritten as
\begin{align}
\beta_3 &= \frac{G_F\sin^2 \theta_W }{12\sqrt{2}} \left[
(c_2+c_4) f_\pi^4 - 16 N_c (c_2-c_4) \langle \overline{\psi}\psi\rangle_M
(J_3-J_4) \right],\cr
\beta_4 &=\frac{G_F\sin^2 \theta_W }{12\sqrt{2}N_c} \left[
(c_1+c_3) f_\pi^4 - 16 N_c (c_1-c_3) \langle \overline{\psi}\psi\rangle_M
(J_3-J_4) \right],
\label{eq:beta34}
\end{align}
which clearly shows that $\beta_4$ is the subleading order in the large
$N_c$ limit with respect to $\beta_3$. Note that the structure of the
$\beta_4$ is the same as that of $\beta_3$ except for the Wilson
coefficients and the $1/N_c$ factor. The magnitudes of the second
terms in Eq.(\ref{eq:beta34}) are much larger than those of the first
ones, we can ignore approximately the first terms. That is, $\beta_3$
and $\beta_4$ can be expressed as
\begin{align}
\beta_3 \approx \frac{G_F\sin^2 \theta_W }{12\sqrt{2}} c_2 \langle
\overline{\psi}\psi \rangle_M (J_4-J_3),\;\;\;\;
\beta_4 \approx -\frac{c_3}{c_2 N_c} \beta_3,
\label{eq:betaapprx}
\end{align}
which indicates that $\beta_3$ is larger than $\beta_4$ approximately
by $(70-75)\,\%$.
\begin{table}[htp]
\centering
\begin{tabular}{ccccc} \hline \hline
WLEC & LO & NLO($Z$) & NLO($Z+W$) & KS \\ \hline
$\beta_1$ & $\phantom{-}0.031$ & $-0.006$ & $-0.006$ &
$\phantom{-}0.047$ \\
$\beta_2$ & $-0.069$ & $-0.073$ & $-0.073$ & $-0.054$ \\
$\beta_3$ & $-1.334$ & $-1.092$ & $-1.102$ & $-1.041$ \\
$\beta_4$ & $-0.434$ & $-0.309$ & $-0.308$ & $-0.428$ \\
\hline\hline
\end{tabular}
\label{tab:betas}
\caption{The values of $\beta_i$ in units of $10^{-4}~{\rm
MeV}^2$. Notations are the same as in Table~\ref{tab:cs}.}
\end{table}
In Table~\ref{tab:betas}, we list the results for the $\beta_i$. Note
that $\beta_3$ and $\beta_4$ have the same sign because $c_2$ and
$c_3$ have different relative signs as shown in Table~\ref{tab:cs}.
The magnitude of $\beta_3$ indeed turns out to be about 75 \% larger
than the $\beta_4$, as expected from Eq.(\ref{eq:betaapprx}).
\section{Parity-violating $\pi NN$ coupling
constant}\label{sec:result}
We are now in a position to determine the PV $\pi NN$ coupling
constant. Starting from Eq.(\ref{eq:su2Lag}), we are able to derive
the PV $\pi NN$ coupling constant. We already have shown explicitly
how one can obtain the PV $\pi NN$ coupling constant, based on the
$\chi$QSM~\cite{Lee:2012fx}. Thus, we want to briefly explain the
procedure of computing the $h_{\pi NN}^1$ within the model. The PV
$\pi NN$ coupling constant can be derived by solving the following
matrix element:
\begin{align}
\langle N| \mathcal{H}_W^{\Delta I=1} |\pi^a N\rangle &=
\frac{G_F}{\sqrt{6}} \frac{\sin^2 \theta_W}{3} \sum_{i=1}^4
\langle N | \mathcal{Q}_i(z;\mu) |\pi^a N\rangle \cr
&= \frac{G_F}{\sqrt{6}} \frac{\sin^2 \theta_W}{3} \sum_{i=1}^4 \int
d^4 \xi (k^2 + m_\pi^2) e^{ik\cdot\xi} \langle N |
\mathcal{T} [\mathcal{Q}_i(z;\mu) \pi^a(\xi)]|N\rangle,
\label{eq:mat1}
\end{align}
where the nucleon states can be constructed by using the Ioffe-type
current in Euclidean space ($x_{0}=-ix_{4}$)~\cite{Diakonov:1987ty,
Christov:1995vm}:
\begin{align}
|N(p_{1})\rangle &= \lim _{y_{4}\rightarrow -\infty }
e^{p_{4}y_{4}}{\mathcal{N}}^{*}(p_{1})\int d^{3}y
e^{i{\bm p}_{1}\cdot {\bm y}}J^{\dagger }_{N}(y)|0\rangle,\cr
\langle N(p_{2})| &= \lim _{x_{4}\rightarrow +\infty }
e^{-p_{0}x_{4}}{\mathcal{N}}(p_{2})\int d^{3}x
e^{-i{\bm p}_{2}\cdot {\bm x}}\langle 0| J_{N}(x).
\end{align}
The $J^{\dagger }_{N}$ ($J_{N}$) constitutes $N_{c}$ quarks
\begin{align}
J_{N}(x)\; =\; \frac{1}{N_{c}!}\epsilon ^{c_{1}c_{2}\cdots c_{N_{c}}}
\Gamma ^{s_{1} s_{2}\cdots
s_{N_{c}}}_{(TT_{3}Y)(JJ_{3}Y_{R})}\psi _{s_{1}c_{1}}(x)
\cdots \psi _{s_{N_{c}}c_{N_{c}}}(x)\,,
\end{align}
where $s _{1}\cdots s _{N_{c}}$ and $c_{1}\cdots c_{N_{c}}$
stand for spin-isospin and color indices, respectively.
The $\Gamma ^{\{s \}}_{(TT_{3})(JJ_{3})}$ provides
the quantum numbers $(TT_{3})(JJ_{3})$ for the nucleon: $T=1/2$, $Y=1$
and $J=1/2$. The nucleon creation operator $J_N^\dagger$ can be
obtained by taking the Hermitian conjugate of $J_N$.
The matrix elements in Eq.(\ref{eq:mat1}) is just the four-point
correlation function given as
\begin{align}
\label{Eq:correl}
\lim_{y_{0}\rightarrow -\infty \atop x_{0}\rightarrow +\infty }
\sum_{i=1}^{4} \langle 0 | {\mathcal
T}[J_{N}(x)\mathcal{Q}^i(z;\mu) \partial_\mu A_\mu^a (\xi) J^{\dagger
}_{N}(y) ]|0\rangle\; =\;
\lim _{y_{0}\rightarrow -\infty
\atop x_{0}\rightarrow +\infty }{\mathcal{K}},
\end{align}
where $A_\mu^a$ stands for the axial-vector current. Note that we have
used the partial conservation of the axial-vector current
(PCAC). In principle, the four-point correlation function
$\mathcal{K}$ can be computed by solving the following functional
integral
\begin{equation}
\mathcal{K} \;=\; \frac1{\mathcal{Z}} \int D\psi D\psi^\dagger DU
J_{N}(x){\cal Q}^i(z;\mu) \partial_\mu A_\mu^a (\xi) J^{\dagger }_{N}(y)
\exp\left[\int d^4x \psi^\dagger \left(i\rlap{/}{\partial} + i
\sqrt{M(-\partial^2)}U^{\gamma_5}\sqrt{M(-\partial)^2}\right)
\psi\right].
\label{eq:corr2}
\end{equation}
As was already mentioned in the previous work~\cite{Lee:2012fx},
it is extremely complicated to deal with Eq.~(\ref{eq:corr2})
technically, since the PV $\pi NN$ coupling constant arises from both
the two-body quark operators $\mathcal{Q}^i$ and the axial-vector
one, which causes laborious triple sums over quark levels already at
the leading order in the large $N_c$ expansion. Thus, we employ the
gradient expansion method as in Ref.~\cite{Lee:2012fx}. In the
gradient expansion, $(\rlap{/}{\partial}U/M)\ll 1$ is used as an
expansion method~\cite{Diakonov:1987ty} to expand the quark
propagator in the pion background field, with the pion momentum
assumed to be small. Equivalently, we can directly start from the
effective weak chiral Lagrangian in
Eqs.(\ref{eq:ExL1},\ref{eq:su2Lag}) already derived in the previous
Section.
The classical soliton is assumed to have a hedgehog symmetry, so that
it can be parametrized in terms of the soliton profile function $P(r)$
\begin{align}
U_0 = \exp \left( i \bm{\tau} \cdot \hat{\bm{r}} \, P(r) \right).
\end{align}
In principle, $P(r)$ can be found by solving the equations of motion
self-consistently~\cite{Christov:1995vm}. However, we will employ a
parametrized form of $P(r)$ which is very close to the self-consistent
one. The classical soliton field can be fluctuated such that the pion
field can be coupled to a $\Delta I=1$ two-body quark operator
\begin{align}
U=\exp\left(i\frac{\bm{\tau}\cdot \bm{\pi}}{2f_\pi}\right)U_0
\exp\left(i\frac{\bm{\tau}\cdot \bm{\pi}}{2f_\pi}\right).
\end{align}
Since the trace of the left and right currents over flavor space
vanish, i.e.
\begin{align}
\langle R_\mu+L_\mu\rangle=0,\;\;\;\;
\langle R_\mu-L_\mu\rangle=0,
\end{align}
the terms with $\mathcal{N}_1$, $\mathcal{N}_2$, $\mathcal{N}_3$, and
$\mathcal{N}_4$ do not contribute to $h^1_{\pi NN}$ as shown in our
previous analysis~\cite{Lee:2012fx} with the DDH effective
Hamiltonian~\cite{Desplanques:1979hn}. Considering the fact that
$\mathcal{N}_1$ and $\mathcal{N}_2$ contain the Wilson coefficient
$c_5$, which is the most dominant one, and $\mathcal{N}_3$ and
$\mathcal{N}_4$ have $c_7$ that is the second largest one, one can
explain a part of the reason why $h_{\pi NN}^1$ turns out to be rather
small in the present approach.
When it comes to all other terms, we can approximately rewrite
$L_\mu^2$ and $R_\mu^2$ as
\begin{align}
L_\mu L_\mu \simeq
\frac{i}{2f_\pi}\left(L_\mu^0L_\mu^0\bm{\tau}\cdot\bm{\pi}
-\bm{\tau}\cdot\bm{\pi}L_\mu^0L_\mu^0\right),\;\;\;\;
R_\mu R_\mu \simeq \frac{i}{2f_\pi} \left(\bm{\tau}\cdot \bm{\pi}
R_\mu^0R_\mu^0 -R_\mu^0R_\mu^0
\bm{\tau}\cdot\bm{\pi}\right)
\end{align}
with $L^0_\mu=iU_0^\dagger\partial_\mu U_0$ and $R^0_\mu=iU_0\partial_\mu U_0^\dagger$,
so that we get
\begin{align}
\langle(R_\mu R_\mu -L_\mu L_\mu)\tau_3 \rangle
=i \frac{\sqrt{2}}{f_\pi}\langle(R_\mu^0 R_\mu^0+L_\mu^0L_\mu^0)
(\tau^-\pi^+ -\tau^+\pi^-)\rangle\ ,
\end{align}
where $\tau^\pm$ and $\pi^\pm$ are defined in the spherical basis as
\begin{align}
\tau^{\pm}=\mp \frac{1}{2}(\tau^1\pm i\tau^2)\ ,\
\pi^{\pm}=\mp \frac{1}{\sqrt{2}}(\pi^1\pm i\pi^2)\ .
\end{align}
The relevant effective Lagrangian is then expressed as
\begin{align}
\mathcal{L}_{\mathrm{eff}}^{\mathrm{SU(2)}}
=(\beta_3+\beta_4) \frac{i\sqrt{2}}{f_\pi}\langle(R_\mu^0
R_\mu^0+L_\mu^0L_\mu^0) (\tau^-\pi^+ -\tau^+\pi^-)\rangle,
\end{align}
where $\beta_3$ and $\beta_4$ are defined already in
Eq.(\ref{eq:betas}).
Since we have already explained how the quantization of the soliton is
performed in the context of the PV $\pi NN$ coupling constant in
Ref.~\cite{Lee:2012fx}, we proceed to compute the $h_{\pi
NN}^1$ within this framework. For simplicity, let us consider the PV
process $n+\pi^+\to p$. Then, we need to compute the following trace
\begin{align}
\langle(R_\mu^0 R_\mu^0+L_\mu^0L_\mu^0)\tau^+\rangle .
\end{align}
Defining isovector fields $r^i_\mu$ and $l_\mu^i$ as
\begin{align}
R_\mu^0=-r_\mu^i\tau^i\ , \ L_\mu^0=-l_\mu^i\tau^i
\end{align}
and using an identity $\langle \tau^i\tau^j\tau^k\rangle =
2i\epsilon^{ijk}$, we obtain
\begin{align}
\langle R_\mu^0 R_\mu^0\tau^+\rangle =
(r_\mu^1+ir_\mu^2)r_\mu^3-r_\mu^3(r_\mu^1+ir_\mu^2),
\;\;\;\;
\langle L_\mu^0 L_\mu^0\tau^+\rangle
= (l_\mu^1+il_\mu^2)l_\mu^3-l_\mu^3(l_\mu^1+il_\mu^2).
\end{align}
Thus, we arrive at the final form of the effective Lagrangian
\begin{align}
{\cal L}^{\rm SU(2)}_{\rm eff}
=(\beta_3+\beta_4) \frac{i\sqrt{2}}{f_\pi} \left[(r_\mu^1+ir_\mu^2)r_\mu^3
- r_\mu^3(r_\mu^1+ir_\mu^2)+(r\rightarrow l)\right]\pi^+,
\end{align}
from which we can derive the PV $\pi NN$ coupling constant.
Using the collective quantization discussed in Ref.~\cite{Lee:2012fx},
we get
\begin{align}
&\int d^3x\langle p\uparrow|r_\mu^3(r_\mu^1+ir_\mu^2)|n\uparrow\rangle
=-\int d^3x\langle p\uparrow|(r_\mu^1+ir_\mu^2)r_\mu^3|n\uparrow\rangle
\cr
&=\frac{2\pi}{3I^2}\int dr\,r^2 \sin^2P(r)
\left[\sin^2P(r)-3\cos^2P(r)\right],
\end{align}
where $I$ denotes the moment of inertia~\cite{Diakonov:1987ty}
expressed as
\begin{align}
I = \frac{N_c}{12} \int_{-\infty}^\infty
\frac{d\omega}{2\pi}\mathrm{Tr}\left(\tau^i\frac1{\omega + i
H}\tau^i\frac1{\omega + i H} \right) \approx \frac83 \pi f_\pi^2
\int_0^\infty dr\, r^2 \sin^2 P(r).
\end{align}
Here, $\omega$ is the energy frequencies of the quark levels and
$\mathrm{Tr}$ stands for the functional trace over coordinate space,
isospin and Dirac spin space. The second term was derived approximately
by the gradient expansion. Similarly, we obtain the same
result for $l_\mu$. Having carried out the calculation of the matrix
element for the collective operators, we finally derive the PV $\pi
NN$ coupling constant $h_{\pi NN}^1$ as
\begin{align}
h_{\pi NN}^1=i\langle p\uparrow|{\cal L}^{\rm SU(2)}_{\rm eff}|n\uparrow,
\pi^+\rangle = \frac{8\sqrt{2}\pi}{3 f_\pi I^2}\, (\beta_3+\beta_4)\, \int dr\, r^2
\sin^2P(r) \left[\sin^2P(r) - 3 \cos^2P(r)\right].
\label{eq:final}
\end{align}
It is interesting to see that Eq.(\ref{eq:final}) is exactly the same
as the expression obtained in Ref.~\cite{Lee:2012fx} except for the
coefficient $\beta_3+\beta_4$.
In order to compute the PV $\pi NN$ coupling constant, we
employ the following numerical values of the constants involved in the
present work: the Fermi constant $G_F=1.16637\times
10^{-5}\,\mathrm{GeV}^{-2}$, the Weinberg angle
$\sin^2\theta_W=0.23116$, and the pion decay constant is obtained to
be $f_\pi=0.0924\,\mathrm{GeV}$ given in
Eq.(\ref{eq:coefficient4}). Concerning the profile function,
we have already examined the dependence of $h_{\pi NN}^1$ on types of
the profile functions~\cite{Lee:2012fx}. The physical profile function
produces the largest value, compared to the linear and arctangent
profile functions. In the present work, we employ the physical profile
function expressed as
\begin{align}
P(r)=\left\{\begin{array}{ll}
2\arctan\left(\frac{r_0}{r}\right)^2, \hspace{2.0cm} \mbox{($r\leq r_x$)},\\
P_0 \, e^{-m_\pi r}(1+m_\pi r)/r^2, \hspace{0.7cm} \mbox{($r>r_x$)},\\
\end{array}\right.
\label{eq:profile}
\end{align}
where $r_0$ is defined as $r_0=\sqrt\frac{3g_A}{16\pi f_\pi^2}$ with
the axial-vector constant $g_A=1.26$. $P_0$ and $r_x$ are given as
$P_0 =2r_0^2$, and $r_x=0.752\, \mathrm{fm}$, respectively. The
profile function in Eq.(\ref{eq:profile}) satisfies a correct behavior
of the Yukawa tail. Then, the moment of inertia is obtained to be
$I=3.32{\rm GeV}^{-1}$.
\begin{table}[htp]
\centering
\begin{tabular}{ccccc} \hline\hline
& LO & NLO($Z$) & NLO($Z+W$) & KS \\ \hline
$h_{\pi NN}^1$ & $10.96$ & $8.69$ & $8.74$ & $9.11$ \\
\hline\hline
\end{tabular}
\caption{$h_{\pi NN}^1$ in units of $10^{-8}$. Notations are the same
as in Table~\ref{tab:cs}.}
\label{tab:h1pi}
\end{table}
Numerical results for $h_{\pi NN}^1$ are summarized in Table~\ref{tab:h1pi}.
In Ref.~\cite{Tiburzi:2012hx}, it was shown that NLO contributions
alter the values of the Wilson coefficients at $\mu=1$ GeV by about
$(10-20)\,\%$, which actually lessens the value of the $h_{\pi
NN}^1$. by about $25\,\%$ as shown in Table~\ref{tab:h1pi}.
As already examined in Eqs.(\ref{eq:beta34}, \ref{eq:betaapprx}),
$\beta_3$ plays a dominant role in determining $h_{\pi NN}^1$. Thus,
the most important operator in the $\Delta I=1$ effective weak
Hamiltonian is $O_2$ in Eq.(\ref{eq:twobody}), which contains the
Wilson coefficient $c_2$. As clearly shown in Table~\ref{tab:h1pi},
the NLO QCD radiative corrections suppress the PV $\pi NN$ coupling
constant. In fact, we have shown already in the previous
work~\cite{Lee:2012fx}, the QCD radiative corrections strongly
diminish the value of $h_{\pi NN}^1$. This behavior contrasts
with the case of $K$ nonleptonic decays, where the penguin diagrams
enhance the contribution to the $\Delta I=1/2$ channel.
\begin{table}[htp]
\centering
\begin{tabular}{ccccccc} \hline\hline
DDH~\cite{Desplanques:1979hn} & DZ~\cite{dubovik1986} &
KS~\cite{ks1993} & QCD sum rules~\cite{henley1998} & Skyrme
Model~\cite{weigel1999} & Lattice QCD~\cite{wasem2012} & Present work
\\ \hline
$45.6$ & $11.4$ & $60$ & $2$ & $8.0-1.3$ &
$10.99\pm 5.05_{-0.64}^{+0.58}$ & $8.74$ \\
\hline\hline
\end{tabular}
\caption{Comparion of $h_{\pi NN}^1$ in units of $10^{-8}$ with
various threoretical works. }
\label{tab:5}
\end{table}
In Table~\ref{tab:5}, we compare the present result with those of
various theoretical works. The present result turns out to be about 5
times smaller than the DDH ``best value''. We find that the result
from the QCD sum rules predicts the smallest value of $h^1_{\pi NN}$
whereas Ref.~\cite{ks1993} yields the largest result, in which the
importance of the strangeness contribution was emphasized. Compared
to the value of $h_{\pi NN}^1$ from lattice QCD with connected
diagrams considered only, the present result is in good agreement with
it.
\section{Summary and Outlook}\label{sec:con}
In the present work, we investigated the parity-violating
pion-nucleon coupling constant. Starting from the $\Delta I=1$
effective weak Hamiltonian~\cite{Tiburzi:2012hx} that considered
the next-to-leading order QCD radiative corrections, we derived the
effective weak chiral Lagrangian with the weak low-energy constants
determined in the $\Delta I=1$ and $\Delta S=0$ channel.
In order to calculate the parity-violating pion-nucleon coupling
constant $h_{\pi NN}^1$, we employed the chiral quark-soliton
model. Using the gradient expansion, which is equivalent to using the
effective weak chiral Lagrangian directly, we were able to compute
the values of $h_{\pi NN}^1$. We found that the first four terms of
the Lagrangian did not contribute at all to $h_{\pi NN}^1$, which
partially explains why the value of $h_{\pi NN}^1$ should be
small. It was also found that the main contribution to $h_{\pi NN}^1$
arose from the operator $O_2$ in the effective weak Hamiltonian.
We also noted that the next-to-leading-order QCD radiative corrections
further suppress the value of $h_{\pi NN}^1$ and as a result we
obtained $h_{\pi NN}^1=8.74\times 10^{-8}$. We compared this result with those from
various theoretical models including the recent result from lattice
QCD. The present result was shown to be in agreement with that
from lattice QCD.
The present work can be extended to the SU(3) case in which the
strange quark comes into play. Another merit of the chiral
quark-soliton model is that the explicit breaking of flavor
SU(3) symmetry can be treated systematically, the strange
quark mass being considered as a perturbation. Thus, it is interesting
to examine the contribution of the strange quark and its current quark
mass to the parity-violating pion-nucleon coupling constant. Other coupling
constants such as $h_{\rho NN}$ and $h_{\omega NN}$ can be studied
within the same framework. The related works are under way.
\section*{Acknowledments}
The work of H.-Ch.K. was supported by Basic Science
Research Program through the National Research Foundation of Korea
funded by the Ministry of Education, Science and Technology (Grant
Number: NRF-2015R1D1A1A01060707).
The work of H.J.L. was supported by Basic Science
Research Program through the National Research Foundation of Korea (NRF)
funded by the Ministry of Education (Grant
Number: NRF-2013R1A1A2009695).
|
\section{Introduction}\label{transfo}
The performance measure of an algorithm involves the evaluation of the quality of the approximated optimum with regards to the real optimum. This can be done by using the concept of \emph{Regret}, well studied in the machine learning literature and used in the noisy optimization literature, sometimes under other names. Basically, in the optimization framework, the regret acounts for the ``loss'' of choosing the point used in the algorithm over the best possible choice: the optimum. Therefore, we measure the difference between the point used/recommended by the algorithm and the optimum in terms of the objective function.
In general, an optimization algorithm searches for the optimum, and to do so, it produces iteratively \emph{search points} which will be evaluated through the objective function. And at regular steps, the algorithm must return a \emph{recommendation point} that will be the best approximation to the optimum so far. Note that the recommendation point can be equal to a search point, but not necessarily.
The most usual form of regret is termed \emph{Simple Regret}. The \emph{Simple Regret} measures the distance, in terms of fitness values, between the optimum and the recommendation point output by the algorithm. It is widely used (possibly without this name) in noisy optimization \cite{fabian,chen1988,BubeckE09}. However, some test beds, notably the Bbob/Coco framework in the first version, did not allow the distinction between search points (at which the fitness function is evaluated) and recommendations (which are output by the algorithm as an approximation of the optimum), so that the Simple Regret can not be checked. This leads to the use of an \emph{Approximate Simple Regret} (name by us), which evaluates the fitness difference between the search points (and not the recommendations) and the optimum. Later, another form of regret, that we will term here \emph{Robust Simple Regret}, was also proposed, using recommendation points. We analyze in this paper the use of different regrets that aim to estimate the quality of the approximated optimum in a similar way. In particular, we show to which extent they lead to incompatible performance evaluations of the same algorithm over the same class of noisy optimization problems, i.e. the convergence rate for the Approximate or the Robust Simple Regret overestimates or underestimates the convergence rate of the more natural simple regret. We also prove some new results in terms of Simple Regret itself.
\section{Framework and Regrets}\label{sec:framework}
This section is devoted to the formalization of the noisy optimization problem considered and the analysed regrets. We will focus on the \emph{Simple Regret} and the alternative definitions that aim to approximate it (denoted here \emph{Approximate Simple Regret} and \emph{Robust Simple Regret}). At the end of the section we will highlight some general relationships between the presented regrets.
\subsection{Continuous Noisy Optimization}\label{sec:noisyOpt}
Given a fitness function $F:D\subset {\mathbb R}^d\to {\mathbb R}$, also known as objective function, optimization (minimization) is the search for the optimum point $x^*$ such that $\forall x\in D, F(x^*)\leq F(x)$. The fitness function is corrupted by additive noise. In other words, given a search point $x\in D$, evaluating $F$ in $x$ results in an altered fitness value $f(x,\bm{w})$ as follows:
\begin{equation}\label{additnoise}
f(x,\bm{w})=F(x)+\bm{w},
proof}{\fbox\\{equation}
where $\bm{w}$ is an independent random variable of mean zero and variance $\sigma$. In the present paper, we will consider a simple case, namely a standard Gaussian additive noise
\footnote{More general cases such that ${\mathbb E} f(x,\bm{w})=F(x)$ can be considered, as most algorithms do not request the noise to be additive and independent; the key point is the absence of bias. \mlc{maybe some cites here? To me, this comment is not clear enough.}\sam{Maybe the comment is not necessary?}}. In addition,
we assume that $F(x)=\|x-x^*\|^2$,
where $x^*$ is randomly uniformly drawn in the domain $D$.
\emph{Noisy optimization} is then the search for the optimum $x^*$ such that ${\mathbb E}_\bm{w} f(x^*,\bm{w})$ is approximately minimum, where ${\mathbb E}_\bm{w}$ denotes the expectation over the noise $\bm{w}$.
Consider a noisy black-box optimization scenario: for a point $x$, the only available information is the noisy value of $F$ in $x$ as given by $f(x,\bm{w})$ for some independent $\bm{w}$. An optimization algorithm generates $x_1,x_2,\dots,x_n,\dots$, successive \emph{search points} at which the objective function is evaluated in a noisy manner. It also generates
$\tilde x_1,\tilde x_2,\dots,\tilde x_n,\dots$ which are \emph{recommendations} or \emph{approximations of the optimum} after $n$ fitness evaluations are performed.
\subsection{Simple Regret and variants}\label{sec:criteria}
The \emph{Simple Regret} ($SR$) focuses only on approximating the optimum in terms of fitness values. Its definition is:
\begin{equation*}
SR_n={\mathbb E}_\bm{w} \left(f(\tilde x_n,\bm{w})-f(x^*,\bm{w})\right) = F(\tilde x_n)-F(x^*)
proof}{\fbox\\{equation*} Notice that the expectation operates only on $\bm{w}$ in $f(\tilde x_n,\bm{w})$, and not on $\tilde x_n$. As a consequence $SR_n$ is a random variable due to the stochasticity of the noisy evaluations of the search points or the (possible) internal randomization of the optimization algorithm.
The $SR$ can be a part of the performance evaluation of an algorithm. In the noise-free case it can be used to determine the \emph{precision} of a method, by ensuring that the algorithm outputs a recommendation $\tilde{x}_m$ satisfying $SR_m \leq \epsilon $.
Even more, when testing algorithms, it is common to use the ``first hitting time'' (FHT). FHT in fact refers to the first ``stable'' hitting time, i.e. the next recommendation is at least as good as the previous one.
This is a reasonable assumption for algorithms solving noise-free problems. In this case the FHT is the minimum $n$ such that $SR_n\leq \epsilon $, provided that the recommendation is defined as $\tilde x_n=x_{i(n)}$ with $1\leq i(n)\leq n$ minimizing $SR_{i(n)}$. However, there is no exact equivalence or natural extension for the FHT with precision $\epsilon $ on noisy optimization. The algorithm only has access to noisy evaluations hence it cannot compute with ``certainty'' $SR_n$, which corresponds to the precision of the algorithm.
An alternative definition, that aims to measure the precision in a similar way to $SR$, is the \textit{Approximate Simple Regret}\footnote{The name is proposed by us.} ($ASR$), defined by:
\begin{equation*}
ASR_n=\min_{m\leq n} F(x_m)-F(x^*).
proof}{\fbox\\{equation*}
$ASR$ takes in account the ``best'' evaluations among all the search points.
It is used in the Bbob/Coco framework \cite{nbbob1,nbbob2,nbbob3,nbbob4,nbbob5,nbbob6,nbbob7,nbbob8,nbbob9}, and in some theoretical papers \cite{fogaasr}.
Notice that since $ASR$ is non-increasing, the FHT can be computed.
In this paper we will also discuss another variant of regret, the \emph{Robust Simple Regret}\footnote{Discussed on Bbob-discuss mailing list (\url{http://lists.lri.fr/pipermail/bbob-discuss/2014-October.txt}. The name is proposed by us.} ($RSR$), defined by:
\begin{equation*}
RSR_n=\min_{k\leq n}\max_{(k-\lfloor g(k)\rfloor)< m\leq k}\left( F(\tilde x_m)-F(x^*)\right),
proof}{\fbox\\{equation*}
where $g(n)$ is a polylogarithmic function of $n$ and $\lfloor \cdot \rfloor$ is the floor function.
The $RSR$ is the ``best'' SR since the beginning of the run, sustained during $\lfloor g(k)\rfloor$ consecutive recommendations \footnote{If $(k-\lfloor g(k)\rfloor)<0$, then the $\max$ on the definition considers indexes between $1$ and $k$}. The polylogarithmic nature of $g(\cdot)$ is explained by the following argument: $g(k)$ be large enough, so that we have a
correct recommendation confirmed over $g(k)$ iterations, but small enough, so that we do not have to wait many evaluations before acknowledging that a correct recommendation has been found.
The $RSR$ uses the recommendation $\tilde x_m$ instead of the search point $x_m$ used in $ASR$. But it uses the worst of a sequence of recommendations.
As a side note about the definition of $RSR$, it was originally proposed to use a quantile instead of the maximum. The ``quantile version'' (without this name), was proposed to become part of the performance measure in Bbob/Coco. However, we will show that it is possible to get a $RSR$ decreasing quicker than the $SR$, so that $RSR$ is a poor approximation of $SR$. The result is valid even with the quantile $100\%$, i.e. the maximum. The same is possible with any other quantile.
The introduction of $RSR$ apparently outplays $ASR$ as an approximation of $SR$ by two means. First, by using recommendations rather than search points. Second, by checking on multiple recommendations that the optimum is correctly found with a given precision. In addition, as well as $ASR$, it is non-increasing, therefore it can be used for fastening experiments on testbed. Please note however that this advantage makes sense only when the target fitness value is known, which is rarely the case except in an artificial testbed
To investigate the convergence rate of the regrets, we will use a slightly different notation than classical works on noisy optimization. Usually the rates are given in terms of $O(h(n))$ where $h(n)$ is some function depending on the number of evaluations $n$. The state of the art shows that in many cases (\cite{dupac,fabian,shamir,fogaasr}) $SR_n = O(n^\psi)$ where $\psi<0$ implies that the algorithm converges. Therefore, there is a linear relationship between $\log(SR_n)$ and $\log(n)$ with a \emph{slope} $\psi$, where $\log(\cdot)$ is the natural logarithm. We will then refer to the \emph{slope of the regret} when speaking about the convergence rate of the regret. The definition of the \emph{slope of the SR} is:
\begin{equation*}
s(SR) = \underset{n\to\infty}{\lim\sup}~ \frac{\log(SR_n)}{\log(n)}
proof}{\fbox\\{equation*}
We have the corresponding definition for the slope of $ASR$ and $RSR$. Notice that if the slope is close to $0$, then the algorithm (at best) converges \emph{slowly}. On the contrary, if the slope is negative and far away from $0$, then the algorithm is \emph{fast}.
\subsection{General results for $SR$, $ASR$ and $RSR$: $RSR$ is an optimistic approximation of $SR$}
The problems analysed in this paper arise from the gap between $s(SR)$ and $s(ASR)$ or $s(RSR)$. Ideally we would like to have a regret that can be used easily and that \emph{approximates} the Simple Regret. In the following sections (\ref{sec:EA} and \ref{sec:stoc}) we will see with specific examples there is indeed a gap between $s(SR)$ and $s(ASR)$. In some cases using $ASR$ overestimates the performance of algorithms and in others it underestimates their performance. An extreme case is detailed in section \ref{sec:stoc}, where we prove that Alg.~\ref{alg:shamir} has optimal convergence rate in term of $SR$ whilst for $ASR$ it does not converge at all.
In general, by definition, we have that for any algorithm,
$s(RSR)\leq s(SR)$. From this point of view, $RSR$ is a correct lower bound for $SR$. In other words, if an algorithm is fast in terms of $SR$, its performance measured by $RSR$ will be at least as good.
Unfortunately, this bound is not nearly tight. We will prove that a small modification on the algorithm induces $s(RSR)\leq s(ASR)$ whereas $s(SR)$ is the same - so that, for cases in which $s(ASR)<s(SR)$ (sometimes by far, as explained in later sections), we get $s(RSR) < s(SR)$ (sometimes by far)
Let $A$ be an optimization algorithm and its search points $(x_i)_{i \geq 1}$. Consider another algorithm denoted $A_g$ and its successive search points $(X_i)_{i \geq 1}$. The search points of $A_g$ are obtained by repeating $\lfloor g(n)\rfloor$ times, for any $n$, the search points $x_n$ of $A$. Hence, we get the assignment:
\begin{tabular}{rcrrl}
$X_{1}=$ & $\dots$ & $=$ & $ X_{1+\lfloor g(1)\rfloor}$ & $\leftarrow x_1$\\
$X_{2+\lfloor g(1) \rfloor}=$ & $\dots$& $=$ & $X_{2+\lfloor g(1)\rfloor+\lfloor g(2)\rfloor} $ & $\leftarrow x_2$\\
& $\vdots$ & \\
$X_{n+\sum_{j=1}^{n-1} \lfloor g(j)\rfloor} =$& $\dots$ & $=$ & $X_{n+\sum_{j=1}^{n} \lfloor g(j)\rfloor} $& $\leftarrow x_{n}$
proof}{\fbox\\{tabular}
Let the recommendation points of $A_g$ be defined by $\tilde{X}_n = X_n$ for any $n$. Therefore $A_g$ is a slightly modified version of $A$ since there is an additional polylogarithmic number of evaluation in $A_g$, assuming $g$ is polylogarithmic. The $RSR$ of $A_g$ (say $RSR_{A_g}$) converges approximately as fast as the $ASR$ of the original algorithm (say $ASR_A$). The extra polylogarithmic number of evaluations does not affect the linear convergence in log/log scale. Hence, for any algorithm $A$, $s(RSR_{A_g}) \leq s(ASR_A)$.
The general relationships between the slopes of the $SR$ and its approximations are not conclusive since the bounds are not tight. We will show some gaps between the different approximations of $SR$ and the $SR$ itself.
In the following sections we present five algorithms that will serve as clear examples to see the differences of using one or another regret as performance measures.
We will focus in two types of algorithms: the first group, in Section~\ref{sec:EA}, consists of Evolutionary Algorithms and Random Search and the second, in Section~\ref{sec:stoc}, of algorithms using approximations of the gradient of the objective function.
For each class of algorithms, we exhibit convergence rate bounds on $s(SR)$, $s(ASR)$ and $s(RSR)$. Section~\ref{sec:experiments} displays some experimental works in order to confront theory, conjecture and practice.
\section{Evolutionary Algorithms}\label{sec:EA}
On the group of Evolutionary Algorithms (EAs), we present Random Search (RS), Evolution Strategy (ES) and Evolution Strategy with resampling (ES+r). They all use comparisons between fitness values to optimize the function.
\vfill\break
\subsection{Random Search}
Random Search (Alg. \ref{alg:rs}) is the most basic of stochastic algorithms \cite{rastrigin1964convergence}.
The search points $x_1,\dots,x_n,\dots$ are independently identically drawn according to some probability distribution. $\tilde x_n$ is usually the search point with the best fitness so far, i.e. with $y_i$ the fitness value obtained for $x_i$, we have $\tilde x_n=x_i$ with $i\in\{1,\dots,n\}$ minimizing $y_i$.
\begin{algorithm}[H]
\caption{\label{alg:rs} {\scriptsize Random Search.}}
\begin{algorithmic}[1]\scriptsize
\State{{\bf Initialize:} Candidate solution $\tilde x$ randomly drawn in $[0,1]^d$}
\State{$bestfitness \leftarrow f(\tilde x)$}
\State{{\bf Initialize:} $n\leftarrow1$}
\While{not terminated}
\State{Randomly draw $y$ in $[0,1]^d$.
\State{$fitness\leftarrow f(y)$}
\If{$fitness<bestfitness$ }\label{line:sel1}
\State{ $\tilde x \leftarrow y$}
\State{ $bestfitness\leftarrow fitness$}\label{line:sel2}
\EndIf
\State{$n\leftarrow n+1$}
\EndWhile
\Return{$\tilde{x}$}
proof}{\fbox\\{algorithmic}
proof}{\fbox\\{algorithm}
We consider in this paper a simple variant of RS to show clearly the contrast between $SR$ and $ASR$.
{\bf Framework for RS:} each search point is randomly drawn independently and uniformly, sampled once and only once, with the uniform probability distribution over $[0,1]^d$. The objective function is the noisy sphere function $f$:
\begin{equation}
f(x)=\|x-x^*\|^2+\mathcal{N}.\label{noisysphere}
proof}{\fbox\\{equation}
where $\mathcal{N}$ is a standard Gaussian variable.
In this setting, existing results in the literature imply a bound on $s(ASR)$ as explained in Property \ref{prop:RS}. We will prove then that the slope of the Simple Regret is not negative, as formalized in Theorem \ref{rswn}
\subsubsection{Approximate Simple Regret: $s(ASR)=-2/d$}
\begin{property}\label{prop:RS}
Consider RS described in Alg.~\ref{alg:rs}, with the framework above.
Then almost surely $ASR_n=O\left(\frac{1}{n^{2/d}}\right)$.
proof}{\fbox\\{property}
\begin{proof}
\cite{deheuvels} has shown that among $n$ points generated independently and uniformly over $[0,1]^{d}$ the closest
search point to the optimum is almost surely at distance $O\left(\frac{1}{n^{1/d}}\right)$ from the optimal point $x^*$ within a logarithmic factor. Hence for the sphere function\footnote{The result also holds for a function locally quadratic around a unique global optimum.}, the Approximate Simple Regret $ASR_n$ almost surely satisfies: $ASR_n = O\left(\frac{1}{n^{2/d}}\right)$ up to logarithmic factors.
proof}{\fbox\\{proof}
\subsubsection{Simple Regret: $s(SR)$ is not negative}
\begin{theorem}\label{rswn}
With the framework above, for all $\beta>0$, the expected simple regret ${\mathbb E}(SR_n)$ is not $O(n^{-\beta})$.
proof}{\fbox\\{theorem}
\begin{proof} See supplementary material. proof}{\fbox\\{proof}
{\bf Remark:} Roughly speaking, the proof of the theorem is based on the fact that with a non-zero probability a search point which does not have the best fitness value, is selected as the best point in Lines~\ref{line:sel1}-\ref{line:sel2} of Alg. \ref{alg:rs}.
\def\movedintosup{
The following sections prove this theorem.\mlc{if we are too long, we can put this proof in supplementary material.}
\begin{lemma}[the quantiles of the standard
Gaussian random variable.]
Let $Q(q)$ be the quantile of the standard centered Gaussian $\mathcal{N}$,
i.e. $\forall q\in (0,1), P(\mathcal{N} \leq Q(q) ) =q$.
Then, there exist $X(q)=-\sqrt{-\log(q S_X(q) )}$ and $Y(q)=-\sqrt{-\log(q S_Y(q))}$ such that
$$X(q) \leq Q(q) \leq Y(q),$$
for some $S_X(q)$ and $S_Y(q)$ polylogarithmic as functions of $q$.
proof}{\fbox\\{lemma}
Proof: See e.g. http://www.johndcook.com/normalbounds.pdf TODO we should put this into a clean reference.QED
Let us consider $N$ the number of search points,
$x_1,\dots,x_N$ the $N$ i.i.d search points,
$\epsilon(N)>0$ and $C(N)>0$
with $\epsilon(N) =o( C(N) )$,
and $\epsilon(N)$ and $C(N)$ both $\Omega(1/N)$ and $o(1)$.
Define $N_g(N)=\{ i\in \{1,\dots,N\}; || x_i - x* || \leq \sqrt{\epsilon(N)} \} $ the number of search points with norm $\leq \sqrt{\epsilon(N)}$ (i.e. ``good'' search points, with simple regret better than $\epsilon(N)$ ).
Define $N_b(N)$ the number of search points with norm $> \sqrt{\epsilon(N)}$ but norm $\leq \sqrt{C(N)}$ (i.e. bad search points, but not very bad...), i.e. $N_b(N)\leq N-N_g(N)$. \sam{precise here if we abuse of notation or change the notation ($N_{b}(N)$ is a set) }
\begin{lemma}[Linear numbers of good and bad points]\label{lemmaCheb}
There exist constants $K_g>0$ and $K_b>0$, such that with probability $\geq \frac12$,
\begin{equation}
N_g(N) \leq K_g N\sqrt{\epsilon(N)}\label{calvin}
proof}{\fbox\\{equation}
and
\begin{equation}
N_b(N) \geq K_b N \sqrt{C(N)}.\label{hobbes}
proof}{\fbox\\{equation}
proof}{\fbox\\{lemma}
\begin{proof}
Consider a search point $x_i$. It is ``good'', in the sense above, with probability $\Omega(\sqrt{\epsilon_N})$. This holds for each of the search points; therefore the number of good points is the sum of $N$ Bernoulli random variables with parameter $\Omega(\sqrt{\epsilon_N})$. The expectation and the variance are therefore $\Omega(N\sqrt{\epsilon_N})$.
By Chebyshev inequality, there is $\alpha>0$ such that a random variable $X$ is $O({\mathbb E} X+\alpha \sqrt{Var\ X})$ with probability at least $\frac12$. This implies that the number $N_g$ of ``good'' points is
\begin{equation}
N_g=O(N\sqrt{\epsilon _N}+\alpha\sqrt{N\sqrt{\epsilon _N}}).\label{bolalapouf}
proof}{\fbox\\{equation}
By the assumption $\epsilon _N=\Omega(1/N)$, $N\sqrt{\epsilon _N}\to\infty$; therefore
Eq. \ref{bolalapouf} implies $N_g=O(N\sqrt{\epsilon _N})$.
The proof is similar for the number of ``bad'' points.
proof}{\fbox\\{proof}
Consider $G=\{i\in \{1,\dots,N\}; ||x_i-x^*||\leq \sqrt{\epsilon(N)}\}$ the indices of good search points, and $B=\{i\in\{1,\dots,N\}; \sqrt{C(N)} \geq ||x_i-x^*||>\sqrt{\epsilon(N)}\}$ the indices of bad search points.
\begin{proposition}\label{zeprop}
For some $c>0$, for all $N$, with probability at least $c$,
the minimum of the noisy fitness values for the $N_g$ good points
verifies $\inf_{i \in G} y_i \geq X(1/N_g)$, and the best noisy fitness for the $N_b$
bad points verifies $\inf_{i\in B} y_i \leq C+Y(1/N_b)$.
proof}{\fbox\\{proposition}
\begin{proof}
Consider the case in which $N_g$
and $N_b$ are upper and lower bounded (respectively) as explained
in Lemma \ref{lemmaCheb}.
This happens with probability $\geq \frac12$, by that lemma.
Consider some $n_i$ (for $i\in \{1,2,\dots,N\}$), which are independently identically distributed according to some absolutely continuous density.
Define $q_N$ the $\frac1N$ quantile of their common probability distribution.
Then the probability that $\inf \{n_1,n_2,\dots,n_N\}$ is less or equal to $q_N$ is, by definition,
$1-(1-1/N)^N$. This converges to the constant $1-\exp(-1)$ as $N\to \infty$.
Therefore, the probability that one of the $N_g$
good points has a noise $\leq Q(1/N_g)$ is
upper and lower bounded by a constant;
the probability that one of the $N_b$ bad
points
points has a noise $\leq Q(1/N_b)$ is
upper and lower bounded by a constant;
these events are independent, so the
probability that both happen simultaneously
is a constant; these events are also
independent of the 0.5 probability from Lemma \ref{lemmaCheb},
so with lower bounded probability all
these events happen simultaneously.proof}{\fbox\\{proof}
{\bf{Proof of Theorem \ref{rswn}:}}
\begin{proof}
Let us assume that the expected
simple regret has slope $<-\beta$ for some
$1>\beta>0$; then define
$\epsilon(N)=N^{-\beta}$
and $\alpha=\beta/k$
for some $0<k<1$, and $C(N)=N^{-\alpha}$;
then Lemma \ref{lemmaCheb} and Proposition \ref{zeprop} implies that there is a $c>0$ such that with probability $c$, for $N$ sufficiently large,
\begin{itemize}
\item there are much more good points than bad points, i.e.
\begin{equation}
N_b=o(N_g)\label{proutproutprout}
proof}{\fbox\\{equation}
thanks to Eq. \ref{calvin} and \ref{hobbes}.
\item all good points have noisy fitness at least $X(1/N_g)=\tilde \Theta(-\sqrt{-\log(N_g)})$;
\item at least one bad point has fitness at most $C(N)+Y(1/N_b)=\tilde\Theta(N^{-\alpha}-\sqrt{-\log(N_b)})=\tilde\Theta(-\sqrt{-\log(N_b)})$;
\item therefore (by the two points above, and using Eq. \ref{proutproutprout}, one bad point is selected; \sam{when is it selected? we always keep the point with the best fitness}
proof}{\fbox\\{itemize}
so that the simple regret is larger than $\epsilon(N)$
with probability $c$ for all $N$ sufficiently
large.
This implies that the expected simple
regret is at least $c \epsilon(N)$, which contradicts the slope $<-\beta$. proof}{\fbox\\{proof}
}
\subsection{Evolution Strategies ($ES$)}
Evolution Strategies~\cite{Rechenberg73,Schw74b} are algorithms included in the category of Evolutionary Algorithms (EAs).
In general, EAs evolve a population until they find an optimum for the objective or fitness function. The process starts by a population randomly generated. Then the algorithm iterates creating new individuals using crossover and mutation and then evaluating this new population of offspring and selecting the ones - regarding to their fitness values - that will become the parents of the next generation.
ES have some more specific selection and mutation processes. Usually the mutation is performed by creating new individuals starting from the parent and adding a random value to it (usually normally distributed around the parent). There are various rules for choosing the step-size
The selection in ES is usually deterministic and rank-based. This is, the individuals chosen to be the parents of the next generations are the ones that have the best fitness values.
When dealing with noisy function, the sorting step of the ES is disturbed by the noise and misranking might occur. To tackle this problem, Arnold and Beyer, in \cite{abinvestigation,abnoise} propose to increase the population size. An alternative is to evaluate multiple times the same search point and average the resulting fitness values. For a given search point $x\in D$, $r$ evaluations are performed: $\left(f(x,w_1), \dots, f(x,w_r)\right)$ and the fitness value used in the comparisons is the average of these evaluations $\frac{1}{r} \sum_{i=1}^{r} f(x,w_i)$. In particular, the variance of the noise is divided by $r$. Several rules have been studied: constant \cite{bignoise3}, adaptive (polynomial in the inverse of the step-size), polynomial and exponential \cite{bignoise2} number of resamplings. A general $(\mu,\lambda)$-ES is presented in Algorithm \ref{alg:es}.
\begin{algorithm}[H]
\caption{\label{alg:es} {\scriptsize $(\mu,\lambda)$-Evolution Strategy. The resampling function $r$ may be constant or depend on the number of iterations and possibly on the step-size. When $r=1$, the algorithm reduces to an ES without resampling. ${\cal{N}}$ is a standard Gaussian of dimension $d$. Here, the index $n$ is the number of \emph{iterations}}}
\begin{algorithmic}[1]\scriptsize
\State{{\bf Input:} Parameters $\mu$, $\lambda$ and resampling function $r$ }
\State{{\bf Initialize:} Parent $\tilde x$ and Step-size $\sigma$}
\State{{\bf Initialize:} $n\leftarrow1$}
\While{not terminated}
\State{{\bf Mutation step:}} $\forall i\in \{1,\dots,\lambda\}$,
$x^{(i)}\leftarrow \tilde{x}+ \sigma\mathcal{N}$\label{line:mutate}
\State{{\bf Evaluation step:}}
$\forall i\in \{1,\dots,\lambda\}$
${y^{(i)} \leftarrow\frac{1}{r(n)}\sum_{j=1}^{r(n)}f(x^{(i)}, w_{j})}$\label{line:eval}
\State{{\bf Selection step:} Sort the population according to their fitness and select the $\mu$ best: $(x^{(i)})^\mu$}
\State{{\bf Update Parent:} $\tilde{x}$ from $\sigma$, $(y^{(i)})^\mu$ and $(x^{(i)})^\mu$}
\State{{\bf Update Step-size:} $\sigma$ from $\sigma$, $(y^{(i)})^\mu$ and $(x^{(i)})^\mu$}
\State{$n\leftarrow n+1$}
\EndWhile
\Return{$\tilde{x}$}
proof}{\fbox\\{algorithmic}
proof}{\fbox\\{algorithm}
\subsubsection{Regrets for ES without resampling}\label{subsubsec:ES}
It is known \cite{stocopti5} that when the noise strength is too big, classical evolution strategies (without reevaluations or other noise adaptation scheme) do not converge, they stagnate. \cite{BeyerMutate} experimentally shows that an ES without any adaptation to the noisy setting stagnates around some step-size and at some distance of the optimum. The divergence results suggest that the ES in this case is only as a more sophisticated version of RS. The steps of the ES are more complicated, but not sufficiently adapted to handle the noise of the function. We propose then a Conjecture on the convergence rates for ES.
\begin{conjecture}[Convergence rates for ES]\label{conj:es}
Evolution Strategies without a noise handling procedure have the same convergence rates as Random Search for all regrets.
proof}{\fbox\\{conjecture}
\subsubsection{Simple regret for ES with resamplings}\label{SR_ES}
We will see that the results are more encouraging than in Section \ref{subsubsec:ES} when we consider an ES with some adaptation to mitigate the effect of the noise. We will assume that the function $r$ (number of revaluations per point) in Alg.~\ref{alg:es} grows polynomially or exponentially with the number of iterations.
The work in \cite{bignoise2} shows that ES that include an exponential number of revaluations converges with high probability to the optimum. The convergence rate is $s(SR) = K$ for some $K<0$ under assumptions about the convergence in ES in the noise-free case. Moreover \cite{esareslow} shows that ES, under general conditions, must exhibit $K> -\frac12$.
There is no formal proof of an upper bound that can theoretically ensure a value or a range for $s(SR)$.
However, the experiments on \cite{bignoise2} suggest that the $K=-\frac12$ is reached for functions with a quadratic Taylor expansion and additive noise (as in Eq.~\ref{additnoise}).
Hence we propose Conjecture~\ref{conj:esSReval}:
\begin{conjecture}[$SR$ for ES + $r$]\label{conj:esSReval}
Consider $0<\delta<1$. For some resampling parameters {{(i.e. for some revaluation function $r$)}}, Evolution Strategies with resampling (Alg.~\ref{alg:es}) satisfy $s(SR)=-1/2$ with probability $1-\delta$.
proof}{\fbox\\{conjecture}
This conjecture applies to some ES with step-size scaling as the distance to the optimum, i.e. $\sigma_n$ used for generating the $n^{th}$ search point has the same magnitude as $\|\tilde{x}_n - x^{*}\|$ (\cite{Rechenberg73,Beyer:bookES}). \cite{BeyerMutate} has proposed variants of ES for quickening the convergence thanks to large mutations and small inheritance. Such an approach is not covered by the bound in \cite{esareslow} and it is for sure an interesting research direction - maybe it might reach slope $s(SR)=-1$.
\subsubsection{Approximate simple regret for ES with resamplings}\label{asrreval}
We have seen that an ES can reach a slope of $SR$ approximately $-\frac12$, when using resamplings.
However, $ASR$ can be better
by slightly modifying the original ES, and therefore achieving a faster convergence rate than the real one represented by $s(SR)$
We consider an ES - called $MES+R$ for Modified ES with Resampling. Let $r_n$ be exponential in the number of iterations: $r_n=R\cdot \zeta^n$, $R>0$, $\zeta>1$.
$MES+R$ is as in Alg. \ref{alg:es} with the following modifications. At iteration $n$:
{\bf Generation:} (Alg.~\ref{alg:es}, Line \ref{line:mutate})
Generate additional $r_n$ ``fake'' offspring $\{ x^{(i)f}: 1\leq i\leq r_n \}$, with the same probability distribution as the $\lambda$ offspring. They will be evaluated one time each, but they will {\emph{not}} be taken into account for the selection. Note that this means that they are part of the sequence of points considered by $ASR$, but not by $SR$.
{\bf Evaluation:} (Alg. \ref{alg:es}, Line \ref{line:eval}) Evaluate $r_n$ times each ``true'' offspring $\{ x^{(i)} : 1\leq i\leq\lambda \}$ to obtain their corresponding fitness value $y^{(i)}$. Evaluate one time each ``fake'' offspring.
Therefore, performing $(\lambda+1)r_n$ function evaluations in each iteration.
Then, under some reasonable convergence assumptions which are detailed in theorem \ref{thm:asr} below, the $ASR$ reaches a faster rate: $s(ASR)=-1/2-2/d$ with high probability.
\begin{theorem}\label{thm:asr}
Consider $0<\delta<1$.
Consider an objective function $F(x)=\|x\|^2$, where $x\in {\mathbb R}^d$.
Consider a $MES+R$ as described previously. Assume that:
\begin{enumerate}[label=(\roman*)]
\item $\sigma_n$ and $\|\tilde{x}_n\|$ have the same order of magnitude:
\begin{equation}\label{xsigma}
\|\tilde{x}_n\|=\Theta(\sigma_n).
proof}{\fbox\\{equation}
\item\label{logalso} $\log-\log$ convergence occurs for the $SR$:
\begin{equation}\label{loglog}
\frac{\log(\|\tilde{x}_n\|)}{\log(n)}\underset{n\rightarrow+\infty}{\longrightarrow }-\frac12 ~\mbox{ with probability $1-\delta$,}
proof}{\fbox\\{equation}
proof}{\fbox\\{enumerate}
Then, with probability at least $1-\delta$, $s(ASR)=-1/2-2/d$.
proof}{\fbox\\{theorem}
\begin{proof} See supplementary material. proof}{\fbox\\{proof}
\begin{remark}\label{rmq:sr}
The assumption of $s(SR)=-1/2$ is based on the convergence of ES in the noise-free case and it is essential to prove Theorem~\ref{thm:asr}. This rate of convergence can be proved when the ES converges in the noise-free case (details on \cite{bignoise2}).
But the convergence of ES has not been formally proved, not even for the noise-free case. There is an important element given in \cite{TCSAnne04-corr}, showing that $\frac1n \log ||x_n-x^*||$ converges to some constant, but this constant is not proved negative. Furthermore, parameters ensuring convergence in the noisy case are unspecified in \cite{bignoise2}.
proof}{\fbox\\{remark}
\section{Stochastic Gradient Descent}\label{sec:stoc}
For the group of Stochastic Gradient Descent Algorithms, we consider the ones presented by Shamir \cite{shamir} and Fabian \cite{fabian} which approximate the gradient of the objective function. We will denote them Shamir algorithm and Fabian algorithm respectively. Both of them approximate the gradient of the function using function evaluations by different methods, therefore they remain in the black-box category. Fabian algorithm uses the average of redundant finite differences and Shamir algorithm a one point estimate gradient technique.
The convergence rates in terms of $SR$ are proved in~\cite{shamir} and ~\cite{fabian}. For Shamir it is shown that $s(SR)=-1$ in expectation for quadratic functions.
Fabian ensures a rate $s(SR)=-1$ approximately and asymptotically only for limit values of parameters. However, it requires only smooth enough functions, so the class of functions is wider than the one considered in~ \cite{shamir}. This rate $s(SR)=-1$ has been proved tight in ~\cite{chen1988}. Hence, Shamir and Fabian algorithm are faster than ES's, which cannot do better than $s(SR)=-\frac12$, at least under their usual form~\cite{esareslow}.
\subsection{Shamir's quadratic algorithm}
Shamir algorithm presented in~\cite{shamir} for quadratic functions is Algorithm \ref{alg:shamir}.
\begin{algorithm
\caption{ {\scriptsize \label{alg:shamir} Shamir Algorithm for Quadratic functions. $\Pi_{W} $ represents the projection over the space $W$} }
\begin{algorithmic}[1]\scriptsize
\State{{\bf Input:} Parameters $\lambda$ and $\epsilon $}
\State{{\bf Initialization:} $\hat{x}_1\leftarrow 0$, $n\leftarrow 1$}
\While{not terminated}
\State{Pick $r\in \{-1,1\}^{d}$ uniformly at random}
\State{$x_{n}\leftarrow\hat{x}_{n}+\frac{\epsilon }{\sqrt{d}}r$}
\State{{\bf Evaluate:} $v\leftarrow f(x_n,w)$}
\State{$\hat{g}\leftarrow\frac{\sqrt{d}v}{\epsilon }r$}
\State{$\hat{x}_{n+1}\leftarrow\Pi_{W}\left(\hat{x}_{n}+\frac{1}{\lambda n}\hat{g} \right)$}
\State{{\bf Recommend:} $\tilde{x}_n\leftarrow\lceil\frac{2}{n}\rceil\sum_{k=\lceil n/2\rceil}^{n} \hat{x}_{k}$}
\State{$n\leftarrow n+1$}
\EndWhile
\Return{$\tilde{x}_{n}$}
proof}{\fbox\\{algorithmic}
proof}{\fbox\\{algorithm}
One of the key points in Alg.~\ref{alg:shamir} are that there is only one evaluation per iteration (somehow in the spirit of Simultaneous Perturbation Stochastic Approximation SPSA~\cite{spall00adaptive,beynoise}). The second important point is that the expectation of the distance between search points and recommendations is constant, which implies that the search points do not converge towards the optimum! This is not a problem for the convergence in terms of $SR$, when search points $x_n$ and recommendations $\tilde{x}_n$ are distinguished, but it makes a difference for $ASR$.
Shamir algorithm has an optimal convergence rate in expectation ($s(SR)=-1$) for quadratic functions. This fact should be acknowledge by any other regret used to evaluate its performance which aims to aproximate the $SR$.
But intuitively in the framework of Shamir algorithm, the $s(ASR)$ is presumably a bad approximation of $s(SR)$
due to the queries at a constant distance of the current recommendation. This convergence rate in terms of $s(ASR)$ could not be obtained from the results in ~\cite{shamir}. Nonetheless, we prove in a general way that as long as the results for Shamir algorithm are satisfied \emph{almost surely}, then $s(ASR)=0$ a.s. Therefore we present the latter result in Theorem ~\ref{thm:aSRgradient} and a conjecture on the convergence rate of $s(ASR)$ in expectation for Shamir algorithm.
\begin{theorem
\label{thm:aSRgradient}
Assume that the optimum $x^*$ is unique and that $(\tilde x_{n})$ is a sequence of recommendation points converging a.s. to $x^*$. Assume that the sequence of evaluation points $(x_n)$ is such that $\forall n, x_n\neq x^*$ a.s. and that $ \|x_n-\tilde{x}_n\| $ is constant.
Then, a.s.
\begin{equation*}
s(ASR)=0.
proof}{\fbox\\{equation*}
proof}{\fbox\\{theorem}
\begin{proof}
$\tilde x_n$ converges almost surely to the optimum and $x_n$ is at a constant distance from $\tilde x_n$.
Therefore the distance between $x_n$ and the optimum converges to a constant.
This implies that the minimum $\min_{i=1}^n \|x_i-x^*\|^2$ is lower bounded by some constant.
Therefore $s(ASR)=0$.
proof}{\fbox\\{proof}
\begin{conjecture}[$ASR$ for the Shamir algorithm]
Shamir algorithm also verifies $s(ASR)=0$ a.s. on quadratic functions.
proof}{\fbox\\{conjecture}
\subsection{Fabian Algorithm}
Algorithm \ref{alg:fabian} presents the algorithm studied in \cite{fabian}. Unlike Algorithm~\ref{alg:shamir}, Fabian algorithm performs several evaluations per iteration, and the distance between search point and recommandation decreases.
\begin{algorithm
\caption{ {\scriptsize \label{alg:fabian} Fabian Algorithm. $e_i$ is the $i^{th}$ vector of the standard orthogonal basis of ${\mathbb R}^d$ and $e_{1,s/2}$ is the $1^{st}$ vector of the standard orthogonal basis of ${\mathbb R}^{s/2}$. $v_i$ is the $i^{th}$ coordinate of vector $v$. ($\hat{x}_{i}$) is the $i^{th}$ coordinate of intermediate points ($\hat{x}$). ($x^{(i,j)+}$) and ($x^{(i,j)-}$) are the search points and $\tilde{x}$ is the recommendation. Here, the index $n$ is the number of \emph{iterations}.} }
\begin{algorithmic}[1]\scriptsize
\State{{\bf{ Input:}} An even integer $s>0$. Parameters $a$, $\alpha$, $c$, $\gamma$.}
\State{{\bf Initialization:}\\
$u_{i} \leftarrow \frac{1}{i},\ \forall\ i\in\{1,\dots,s/2\}$\\
Matrix $U\leftarrow\left( \| u_{j}^{2i-1} \| \right)_{1\leq i,j \leq s/2}$\\
Vector $v\leftarrow\frac12 U^{-1}e_{1,s/2}$\\
$\tilde{x}\leftarrow x\in [0,1]^{d}$ uniformly at random\\
$n \leftarrow 1$}
\While{not terminated}
\State{$a_n \leftarrow\frac{a}{n^{\alpha}}$, $c_n\leftarrow\frac{c}{n^{\gamma}}$}
\State{$\forall j\in\{1,\dots,s/2\}$,\ $\forall i\in\{1,\dots,d\}$\\
{~~~~\bf Evaluate:}
\begin{eqnarray*}
x^{(i,j)+}\leftarrow \tilde{x} + c_n u_{j}e_{i} \qquad
x^{(i,j)-}\leftarrow \tilde{x} - c_n u_{j}e_{i}
proof}{\fbox\\{eqnarray*}}
\State{$\hat{x}_{i}\leftarrow\frac{1}{c_n} \sum_{j=1}^{s/2} v_j \left(f(x^{(i,j)+})-f(x^{(i,j)-})\right)$}
\State{{\bf Recommend:} $\tilde{x}\leftarrow\tilde{x}-a_{n}\hat{x}$}
\State{$n\leftarrow n+1$}
\EndWhile
\Return{$\tilde{x}$}
proof}{\fbox\\{algorithmic}
proof}{\fbox\\{algorithm}
The work in \cite{fabian} gives the convergence rate in terms of $SR$. The result is presented here as Theorem \ref{thm:fabSR}. The value of the $s(SR)$ depends on the parameters of the algorithm and it is ensured a.s.
\begin{theorem}[Simple Regret of Fabian algorithm]\label{thm:fabSR}
Let $s$ be an even positive integer and $F$ be a function {${(s+1)}$-times differentiable} in the neighborhood of its optimum $x^*$. Assume that its Hessian and its $(s+1)^{th}$ derivative are bounded in norm. Assume that the parameters given in input of Algorithm \ref{alg:fabian} satisfy: $a>0$, $c>0$, $\alpha= 1$, $0<\gamma<1/2$ and $2\lambda_0 a >\beta_0$ where $\lambda_0$ is the smallest eigenvalue of the Hessian. Let ${\beta_0=\min\left(2 s \gamma, 1 - 2 \gamma \right)}$. Then, a.s.:
\begin{equation}\label{fabiantropfort}
n^{\beta}(\tilde{x}_n-x^*) \rightarrow 0\ \forall\ \beta <\beta_0/2
proof}{\fbox\\{equation}
In particular, when $F$ is smooth enough, we get ${s(SR)=-2\beta}$.
proof}{\fbox\\{theorem}
\begin{remark}
Note that $s(SR)$ optimal when $\gamma = \frac{1}{2}(s+1)^{-1}$. In this case, $\beta_0=\frac{s}{s+1} \underset{s\rightarrow \infty}{\rightarrow} 1$. $\beta_0$ can be made arbitrarily close to $1$, so $2\beta$ also, but then $\gamma$ goes to $0$. Hence we get the values of Table \ref{maintable}, with $2\beta=1-e$, $e>0$ and close to $0$.
proof}{\fbox\\{remark}
This shows that the Fabian algorithm can have $s(SR)$ arbitrarily close to $-1$, which is optimal. As in the case of Shamir, this optimal performance should be captured by the regret used to evaluate the algorithm.
Unfortunately, this is not the case, as detailed in Theorem \ref{thm:fabASR}.
\begin{theorem}[$ASR$ of Fabian algorithm]\label{thm:fabASR}
Let $F$ be a $\lambda$-convex and $\mu$-smooth function corrupted by an additive noise with upper bounded density and with optimum randomly drawn according to a distribution with upper bounded density. Then, a.s.,
\begin{equation*}
s(ASR)= -\min(2\beta,2\gamma).
proof}{\fbox\\{equation*}
proof}{\fbox\\{theorem}
\begin{proof} See supplementary material. proof}{\fbox\\{proof}
\begin{remark}
Theorem \ref{thm:fabASR} shows that $s(ASR)=-\min(2\beta,2\gamma)$, i.e., when then Fabian algorithm is optimized for $SR$, $s(ASR)$ is close to $-2\gamma$, close to $0$.
proof}{\fbox\\{remark}
\subsection{Shamir and Fabian adapted for $ASR$}
Both algorithms presented in this section have a clear difference between the search and recommendation points. This fact is not automatically distinguished when we are evaluating their performance using for example a test bed. If we modify the algorithms we can achieve $ASR$ approximating well the optimal behavior reported by $SR$. A modification such as sampling one point out of two at the current recommendation, without using it in the algorithm can imply $s(ASR)=s(SR)$ arbitrarily close to $-1$.
\begin{sidewaystable}
\centering
\ra{1.3}
\begin{tabular}{@{}rrrcrrcrr@{}}\toprule
& \multicolumn{2}{c}{SR} & \phantom{abc}& \multicolumn{2}{c}{ASR} &
\phantom{abc} & \multicolumn{2}{c}{RSR}\\ \cmidrule{2-3} \cmidrule{5-6} \cmidrule{8-9}
& conv. & type && conv. & type && conv. & type \\ \midrule
\bf{Evolutionary Algorithms}\\
RS & $\mathbf{0}$ & \textbf{expect.} && $\mathbf{ -\frac2d }$ & \textbf{a.s.} && $\mathbf{ -\frac2d }$ & a.s. \\
ES & $0$ & expect. && $-\frac2d$ & a.s. && $-\frac2d$ & a.s. \\
ES + $r$ & $- \frac12$ & high prob. && $- \frac12$ & high prob. && $- \frac12$ & high prob. \\
MES+ $r$ & $- \frac12$ & high prob. && $\mathbf{ - \frac12 - \frac2d }$ & \textbf{high prob.} && $\mathbf{ - \frac12 - \frac2d } $ & \textbf{high prob.} \\
\bf{Stochastic Gradient}\\
Shamir & $\mathbf{-1}$& \textbf{expect.} && $\mathbf{0}$ & \textbf{expect.} && $\mathbf{-1}$ & \textbf{expect.} \\
Shamir for ASR & $\mathbf{ -1 }$& \textbf{expect. }&& $\mathbf{-1}$ & \textbf{expect.} && $\mathbf{-1}$ & \textbf{expect.}\\
Fabian & $\mathbf{-(1-e)}$ & \textbf{a.s.} && $\mathbf{-e'}$ & \textbf{a.s.} && $\mathbf{-(1-e)}$& \textbf{a.s.}\\
Fabian for ASR & $\mathbf{-(1-e)}$ & \textbf{a.s.} && $\mathbf{-(1-e)}$ & \textbf{a.s.} && $\mathbf{-(1-e)}$ & \textbf{a.s.} \\
\bottomrule
proof}{\fbox\\{tabular}
\caption{\label{maintable}Convergence rates for the regrets analysed on this paper. The ``convergence'' column refers to the convergence rate and the ``type'' column specifies the type of convergence: with high probability, in expectation, almost surely. The results in bold are proved and the others are conjectures, all of them presented in this paper.
}
proof}{\fbox\\{sidewaystable}
\vfill\break
\section{Experiments}\label{sec:experiments}
We present experimental results for part of the algorithms theoretically analysed \footnote{In addition, the experimental results we include the algorithm $UHCMAES$, as another example of an $ES$. For more information, see ~\cite{hansen2009tec}.}. We will analyse the convergence rate of this algorithms in terms of \emph{slope} (see section \ref{sec:criteria} for the definition). As in the theoretical part,
the function to optimize is the noisy sphere: $f(x)=\|x-x^*\|^2+\vartheta\mathcal{N}$ where $\vartheta=0.3$ and $\mathcal{N}$ is a standard gaussian distribution\footnote{ The choice $\vartheta=0.3$ is made only to illustrate in the experiments the effect of the regret choice in a reasonable time bugdet. The noise is weaker than in the case of a standard gaussian and the algorithms can deal with it faster. The optimum $x^*$ for the experiments of each algorithm is different, which does not affect the result since the regret compares the function value on the search/recommended points and on the optimum.}. The dimension of the problem is $2$. The results in Fig. \ref{zola} correspond to the mean of 10 runs for each of the algorithms.
\begin{table}[h]
\scriptsize
\begin{center}
\begin{tabular}{rl}
\toprule
Algorithms & Set of parameters \\
\midrule
UHCMAES ~\cite{hansen2009tec} & $x_{\text{initial}}=1$, $\sigma_{\text{in}}=1$ \\
Shamir &$\epsilon=0.3$, $\lambda=0.1$, $B=3$\\
$(1+1)$-ES & \\
$(1+1)$-ES resamp & resamp$=2^n$ \\
Fabian & s=4 $\alpha=1$, $\gamma=0.01$ \\
\bottomrule
proof}{\fbox\\{tabular}
proof}{\fbox\\{center}
\label{table:exps}
proof}{\fbox\\{table}
\begin{figure*}[ht]
\begin{minipage}{\textwidth}
\centering
\subfloat[Simple Regret]{\label{fig:SR}\includegraphics[width=0.45\textwidth]{SR.eps}}
\subfloat[Approximate Simple Regret]{\label{fig:ASR}\includegraphics[width=0.45\textwidth]{ASR.eps}}
\caption{\label{zola} Figure \ref{fig:SR} presents the Slope of Simple Regret for each algorithm on the first $(1\cdot10^6)$ function evaluations. Stochastic Gradient algorithms reach $s(SR)=-1$ while the evolutionary algorithms present $s(SR)=-0.2$. Figure \ref{fig:ASR} presents the Slope of Approximate Simple Regret. Observe that the performance of the algorithms is inverted with regards to the figure \ref{fig:SR} : now the Stochastic Gradient algorithms have the worse performance.}
proof}{\fbox\\{minipage}
proof}{\fbox\\{figure*}
The results in figure (\ref{fig:SR}) show the comparison between algorithms with regards to the $SR$ criterion. The budget is limited to $(1\cdot10^6)$ function evaluations. We can see clearly the difference between the algorithm that are gradient-based and the $ES$. The algorithms Fabian and Shamir achieve $s(SR)=-1$ whereas the $ES$ presented cannot do better than $s(SR)=-0.25$. The figure
(\ref{fig:ASR}) shows that the use of $ASR$ changes completely the performance of the algorithms. In this case, the gradient-based algorithms are the ones with the worst performance. The results confirm the theoretical work (and the conjectures) presented in sections \ref{sec:EA} and \ref{sec:stoc}.
\section{Conclusion}\label{sec:conc}
In this paper we analyse the use of regrets to measure performance of algorithms in noisy optimization problems. We take into account the \emph{Simple Regret} ($SR$) and other two forms of regret used in practice to approximate the $SR$. We show that the convergence rates of the same algorithm over the same class of functions depend on the considered regret. Ultimately this leads to inconsistent results: algorithms are efficient for $SR$ and the opposite for an approximation of $SR$.
Table \ref{maintable} summarizes our results, detailing if they are proved or conjectured and what type of convergence is found.
{\bf{Approximations of $SR$.}} The $ASR$ appears to be a poor approximation of $SR$. Two situations are exposed in this paper.
First, algorithms with fast convergence for $SR$ can have a slow convergence for $ASR$ (Fabian and Shamir algorithm). However, this can be solved by modifying the algorithm in order to sample, sometimes, a recommended point.
Second, an algorithm with slow convergence for $SR$ can have a fast convergence for $ASR$, and this is a serious issue. There is no simple ``patch'' to deal with this problem. Test beds using $ASR$ will underestimate algorithms which include random exploration. This is partially, but not totally, solved by $RSR$. We did not come up with a satisfactory regret definition, which would be consistent with $SR$ (at least showing similar convergence rates) and without drawbacks as those presented above.
However, the difference (between $SR$ on the one hand, and $ASR$/$RSR$ on the other hand) decreases with the dimension for most algorithms.
{\bf{Simple Regret.}}
$SR$ is clearly a natural way to measure the quality of an approximated optimum output by an algorithm. The drawback of $SR$ is that it is not necessarily non-increasing, which is an issue for the concept of ``stable first hitting time''.
Note that a particularly interesting result is the one of Evolution Strategies correctly tuned in terms of its resampling scheme. It appears, from theoretical and experimental results, that it reaches half the speed of classical noisy optimization algorithms in terms of $SR$, on the log-log scale. This corresponds to a squared computation time. \footnote{This is not the case, however, for some ES with fast rates in noisy cases, such as mutate-large-inherit-small\cite{BeyerMutate}.}\\
{\bf{Further work.}}
We have compared convergence rates through their \emph{slopes}, but in some cases we have slopes for almost sure convergence, in other cases slope with high probability, and in others slope in expectation. There is room for refinement of the results. As for the approximations of $SR$, there might be other definitions of regret that approximate it in a better way. Evidently, the use of an approximation makes sense in the case where there are technical issues that do not allow the direct computation of $SR$.
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{sec:intro}
The discovery of the Higgs boson, at the Large Hadron
Collider~(LHC)~\cite{Aad:2012tfa,Chatrchyan:2012xdj}, opened a new era in the
exploration of the electroweak sector of the Standard Model~(SM).
The measured value of the Higgs mass uniquely specifies all
of the couplings and turns the SM into a fully predictive theory.
Hence, we are at a position to perform stringent tests of our current
modelling of these fundamental interactions.
This is only possible if we are in possession of precise theoretical
predictions for the Higgs production cross sections.
Most of the Higgs boson particles observed at hadron colliders are produced
through the process of gluon fusion, a channel that is known to exhibit very
slow convergence in perturbative Quantum Chromodynamics~(QCD).
At LHC energies, the next-to-leading order~(NLO) corrections to the total
cross section for the inclusive production of the Higgs boson through gluon fusion
turn out to be as large as 70\%, and the next-to-next-to-leading order~(NNLO)
corrections increase the cross section by another~30\%~\cite{Harlander:2002wh,
Anastasiou:2002yz, Catani:2007vq}.
The theoretical uncertainty of the NNLO result, arising from the missing higher
orders and obtained by the standard renormalization and factorization scale
variations, is estimated at around 10\%, and is hence at the level of the experimental
accuracy of the Run 1 LHC measurements.
This large uncertainty at NNLO has motivated the efforts to further improve the
precision by calculating the full
next-to-next-to-next-to-leading order~(\text{N$^3$LO}\xspace) result for inclusive Higgs
boson production in gluon fusion~\cite{Anastasiou:2015ema}.
Adding these contributions to the predictions for the cross section reduces
their scale uncertainties down to the level of~3\%.
Apart from the inclusive Higgs cross section, which is the most fundamental
quantity, as it enables one to predict the total number of Higgs particles produced
at a given energy and luminosity, one is also equally interested in more differential
observables.
Therefore,
a significant amount of work has also gone into obtaining predictions for
differential cross sections for Higgs production in gluon fusion beyond NLO.
In particular,
differential observables have been predicted within frameworks of
analytic resummation, like for example small-$q_T$ resummation performed
in QCD in coordinate space up to the NNLL+NLO accuracy~\cite{deFlorian:2011xf}
and directly in momentum space up to NNLL+NNLO~\cite{Monni:2016ktx} as well as
within SCET~\cite{Becher:2012yn} up to NNLL+NLO.
Differential cross sections for Higgs production in gluon fusion have been also
widely studied with approaches in which fixed-order NLO results are matched to parton
shower (NLO+PS) such as the MiNLO
method~\cite{Hamilton:2012np,Hamilton:2012rf}. Recently, NNLO+PS matched
results were computed with the UN$^2$LOPS technique~\cite{Hoche:2014dla} as well
as with an extended version of
MiNLO~\cite{Hamilton:2013fea,Frederix:2015fyz,Hamilton:2016bfu} combined with
the HNNLO program~\cite{Catani:2007vq, Grazzini:2008tf}.
The current methods of performing
NNLO+PS~\cite{Hamilton:2012rf,Hoeche:2014aia,Alioli:2013hqa,Hoche:2015sya,Alioli:2015toa,Hamilton:2015nsa,Frederix:2015fyz,Hamilton:2016bfu}
represent
clear progress in the matching of fixed-order NNLO QCD calculations with
parton shower Monte Carlos (PSMCs).
The next challenge towards even higher-precision perturbative QCD
calculations, needed until the end of the LHC era two decades from now, is the
combination of the fully exclusive NNLO corrections for the hard process with
NLO parton shower (NNLO+NLOPS).
In this article, we present NLO+PS predictions for various differential
distributions computed with the {\textsf{KrkNLO}}{} approach~\cite{Jadach:2011cr, Jadach:2015zsq}.
The main advantage of the {\textsf{KrkNLO}}{} method with respect to other methods of
matching the fixed-order NLO calculations with PSMCs (\mcatnlo{} and \powheg{})
is its simplicity, which stems from the fact that the entire NLO corrections are
implemented using a simple, positive, multiplicative MC weight in combination with
pre-calculated MC-scheme PDFs. The present work is relevant for the above future
developments in the sense that it presents a simplified method of correcting the
hard process to NLO level in combination with a leading order (LO) parton shower (PS)
that will hopefully pave the way to NNLO hard processes combined with a NLO PSMC;
NLOPS is a parton shower MC implementing the NLO evolution kernels in the fully
exclusive form, thus providing the full set of the soft-collinear counter-terms
for the hard process. Ref.~\cite{Jadach:2013dfd} reviews several feasibility
studies which show that constructing such a NLOPS is, in principle, plausible.
In our opinion, any simplifications of the NLO+PS matching, as in the {\textsf{KrkNLO}}{}
method, will be instrumental in the progression towards more ambitious fully
exclusive NNLO+NLOPS projects.
The {\textsf{KrkNLO}}\ method was first introduced Ref.~\cite{Jadach:2015mza}
for $Z/\gamma^*$ production in hadron collisions (the Drell--Yan process, DY)
and was also presented in Ref.~\cite{Jadach:2015zsq}.
These developments followed the initial study in Refs.~\cite{Jadach:2011cr,Jadach:2012vs}
on the inclusion of fixed-order NLO QCD corrections to the hard process
in LO PSMC through an appropriate Monte Carlo weight.
This study was performed
with the use of some dedicated toy-model parton-shower generator
and for gluonstrahlung from quarks only, albeit
for two processes: DY production and deep-inelastic electron--proton scattering (DIS).
The first realistic numerical results (total cross sections
and distributions of $Z/\gamma^*$ transverse momentum
and rapidity) of the {\textsf{KrkNLO}}{} method,
based on its implementation in the \sherpa{}~\cite{Gleisberg:2008ta} PSMC,
were presented in Ref.~\cite{Jadach:2015mza} for the DY process.
The {\textsf{KrkNLO}}{} results were compared with the fixed-order
NLO predictions of \mcfm{}~\cite{MCFM}
and those of other NLO-PSMC matching methods,
namely \mcatnlo{}~\cite{Frixione:2002ik,Frixione:2006he} and
\powheg{}~\cite{Nason:2004rx,Frixione:2007vw},
as well as with the fixed-order NNLO calculations of \dynnlo{}~\cite{Catani:2009sm}.
A satisfactory agreement with other NLO calculations was found. Moreover,
for the boson transverse momentum
the agreement with the NNLO predictions was better than for \mcatnlo{} and \powheg{},
which may be explained by effects beyond NLO accounted for in {\textsf{KrkNLO}}{} as a result of
using the MC factorization scheme and multiplicative virtual+soft-real corrections,
see Ref.~\cite{Jadach:2015mza} for more details.
In that paper the advantages of the {\textsf{KrkNLO}}{} method
over the \mcatnlo{} and \powheg{} techniques were also discussed.
In Ref.~\cite{Jadach:2015mza} the concept of the Monte Carlo (MC)
factorization scheme was introduced as a necessary
ingredient of the {\textsf{KrkNLO}}{} method; this was further developed in
Ref.~\cite{Jadach:2016acv}, see also Ref.~\cite{Jadach:2016viv}.
Appropriate MC-scheme parton distribution functions (PDFs) were defined and
constructed from PDFs in the standard $\overline{\text{MS}}$ scheme.
However, those MC-scheme PDFs included only contributions from
parton--parton transitions that were sufficient
for the DY process in the NLO approximation -- in the following we denote them
with $\rm MC_{DY}$.
The complete PDFs in the MC factorization scheme,
which include all of the LO parton--parton transitions,
were defined, computed and compared with the $\overline{\text{MS}}$ PDFs in Ref.~\cite{Jadach:2016acv},
where Higgs-boson production from initial-state gluons was considered.
In that paper only the values of the total cross section
for the Higgs production at the LHC were shown;
further results for this process will be presented in this work.
However, before presenting the Higgs results,
we first validate the implementation of the {\textsf{KrkNLO}}{} method in
the \herwig{} program with respect to its previous implementation
in \sherpa{} using the DY process.
The \herwig{} implementation of the {\textsf{KrkNLO}}{} method will be our basic platform
for its future developments and applications to other processes.
Then, we will compare the {\textsf{KrkNLO}}{} results for the DY process
based on the complete MC-scheme PDFs with
those where the $\rm MC_{DY}$ PDFs,
as defined in Ref.~\cite{Jadach:2015mza}, are used.
For the process of Higgs-production in gluon--gluon fusion,
we present the results for the total cross section
and the distributions of the Higgs transverse momentum and rapidity at the LHC.
The predictions of the {\textsf{KrkNLO}}{} method are compared
with those of the fixed-order NLO and NNLO calculations
of \hnnlo{}~\cite{Catani:2007vq}, the results
of the NLO-PSMC matching approaches of
\mcatnlo{}~\cite{Frixione:2006he} and \powheg{}~\cite{Nason:2004rx, Frixione:2007vw},
as well as resummed calculations from \hqt{}~\cite{Bozzi:2005wk,deFlorian:2011xf}.
We also confront the predictions of all the above matching methods with the LHC data
of the ATLAS experiment~\cite{Aad:2015lha}.
The outline of this paper is the following: In Section~2,
after describing the set-up for our numerical computations,
we present and discuss the results of the {\textsf{KrkNLO}}{} method;
we do this first for the DY process and then for Higgs production
in gluon--gluon fusion at the LHC.
Section~3 concludes our work and provides some outlook for future developments.
In Appendix~A we present comparisons of various PDF parametrizations
in the $\overline{\text{MS}}$ and MC factorization schemes.
\section{Numerical results}
\label{sec:numres}
\subsection{Set-up}
\label{ssec:setup}
For the numerical evaluation of the cross sections%
\footnote{Unless stated otherwise in the text.}
at the LHC for the proton--proton collision energy of $\sqrt{s}=8\,$TeV
we have chosen the following set of the Standard Model input parameters:
\begin{eqnarray}\label{eq:pars}
M_Z = 91.1876 \; {\rm GeV}, & \quad & \Gamma_Z = 2.4952 \; {\rm GeV},
\nonumber \\
M_W = 80.4030 \; {\rm GeV}, & \quad & \Gamma_W = 2.1240 \; {\rm GeV},
\\
M_H = 126 \; {\rm GeV},
& \quad & m_t = 173.2 \; {\rm GeV},
\nonumber \\
& \quad & G_{\mu} = 1.16637\times 10^{-5} \; {\rm GeV}^{-2},
\nonumber
\end{eqnarray}
and use the $G_{\mu}$-scheme~\cite{LHC-YR} for the electroweak sector of the Standard Model.
We take the renormalization and
factorization scales to be $\mu_R^2=\mu_{F}^2=M_B^2$, where $M_B$ is the mass
of the $Z$ or Higgs boson for the respective processes.
In the case of the Drell--Yan process, detector acceptance cuts are imposed
only on the invariant mass of the final-state lepton pair ($Z/\gamma^*$-boson):
\begin{equation}
50~{\rm GeV}<M_{l\bar{l}}<150~{\rm GeV}.
\label{eq:Mllcut}
\end{equation}
For the Higgs-production process we do not apply any cuts, and for simplicity we
set the Higgs boson to be stable.
The LO, NLO and NNLO Higgs-production matrix elements
are calculated in the $m_t\rightarrow \infty$ and $m_{q\neq t} = 0$ approximation.
To compute the hadronic cross-section, we
employ $\overline{\text{MS}}$ PDFs from the \LHAPDF{} library \cite{Whalley:2005nh} and their MC-scheme
counterparts when using {\textsf{KrkNLO}}{}. The value of ${\alpha_s}$ is chosen to match the
value used in the PDFs. The PDF set used is detailed in the relevant subsection.
\subsection{Drell--Yan process}
\label{ssec:DY}
The results of the {\textsf{KrkNLO}}{} method implemented on top of the \sherpa{}
PSMC~\cite{Gleisberg:2008ta}
for the Drell--Yan process were already presented in Ref.~\cite{Jadach:2015mza}.
As with these results, for this process we use the {\tt MSTW2008} LO set of parton
distribution functions~\cite{Martin:2009iq}, which has ${\alpha_s}(M_Z^2)=0.13938690$.
Here we use these \sherpa{} results to validate a new implementation of {\textsf{KrkNLO}}{}
in the \herwig{} program \cite{Bahr:2008pv,Bellm:2015jjp}.
This version of the \herwig{} PSMC features a new parton-shower
algorithm based on the Catani--Seymour (CS) dipole
\cite{Catani:1996vz} formalism, and is therefore well suited
to the implementation of the {\textsf{KrkNLO}}{} method.
Details on how to implement the {\textsf{KrkNLO}}{} method in the CS-dipole PSMC
are given in Ref.~\cite{Jadach:2015mza} for
the DY process and in Ref.~\cite{Jadach:2016acv} for the Higgs-boson production.
In short, it amounts to replacing the $\overline{\text{MS}}$ PDFs with
the MC-scheme PDFs and applying to each event generated
by the PSMC an appropriate Monte Carlo weight that introduces
the NLO QCD correction to a given hard process.
This weight is positive-definite and can be computed simply
from information provided in an event record.
\subsubsection{LO results}
\label{sssec:dylo}
As a basis for this validation, we first compare the LO-level predictions
from \mcfm{} and \sherpa{} with those of \herwig{}
using an identical choice of parameters (see Section. \ref{ssec:setup}).
The results for the total cross section presented in Table~\ref{tab:Z-lo-settings-val} show
a very good agreement (within statistical errors) between different programs.
\begin{table}[t]
\centering
\begin{tabular}{|c||c|c|c|}
\hline
& \mcfm{} & \sherpa{} & \herwig{} \\
\hline\hline
$\sigma_\text{tot}$ [pb] & $936.9\,(1)$
& $937.2\,(2)$ & $ 936.6\,(2)$ \\
\hline
\end{tabular}
\caption{\sf
Values of the total cross section with statistical errors at the Born level for
the Drell--Yan process with PDFs in the $\overline{\text{MS}}$ factorization scheme.
}
\label{tab:Z-lo-settings-val}
\end{table}
\begin{figure}[!h]
\centering
\includegraphics[width=0.46\textwidth]{Z_LO_PteeFullPS.pdf}
\includegraphics[width=0.46\textwidth]{Z_LO_PteeFullPSlog_mcfm.pdf}\\
\includegraphics[width=0.46\textwidth]{Z_LO_Z_mass.pdf}
\includegraphics[width=0.46\textwidth]{Z_LO_yZmcfm.pdf}\\
\caption{\sf
Comparisons of the $Z/\gamma^*$ distributions
from \sherpa{} and \herwig{} for the LO-level Drell--Yan process
with $e^+e^-$ pairs in the final state.
}
\label{fig:LO-Z-pt}
\end{figure}
Since the {\textsf{KrkNLO}}{} method depends on details of the parton-shower algorith
\footnote{For example, in \sherpa{} the initial-state parton shower has a prefactor of $1/2$
in the scale of the running ${\alpha_s}$ in the calculation
of the Sudakov form factor.} both dipole
showers~\cite{Schumann:2007mg,Platzer:2011bc} have to be set as similar as possible.
The level of agreement of the two PSMCs is presented in Fig.~\ref{fig:LO-Z-pt},
where we show distributions of the final-state $e^+e^-$-pair
(the decay product of $Z/\gamma^*$-boson)
transverse momentum $p_{Tee}$, invariant mass $m_{ee}$ and rapidity $y_{ee}$.
We can see that all of the distributions are
in a good agreement. In the case of the transverse momentum distribution there are
some differences at small $p_T$ that result from different treatment of intrinsic-$k_T$
and the soft-parton limit in the two programs, where differences in the
latter emerge from the different ordering variables employed in the two programs.
The differences at high $p_T$ are due to limited statistics in this region.
\subsubsection{KrkNLO results: H7 vs. Sherpa implementations}
\label{sssec:dynlo}
\begin{table}[!h]
\centering
\begin{tabular}{|c||c|c|c|}
\hline
& \mcfm{} & {\textsf{KrkNLO}}{} \sherpa{} & {\textsf{KrkNLO}}{} \herwig{} \\
\hline\hline
$\sigma_\text{tot}$ [pb] & $1086.5\,(1)$ & $1045.2\,(3)$ & $ 1046.5\,(2)$ \\
\hline
\end{tabular}
\caption{\sf
Values of the total cross section with statistical errors
for the Drell--Yan process
from the {\textsf{KrkNLO}}{} method implemented in \sherpa{} and \herwig{} compared to
the fixed-order result of \mcfm{}.
}
\label{tab:Z-nlo-settings-val}
\end{table}
With the consistent predictions obtained at the LO level,
we are now ready to examine the consistency between
the \sherpa{} and \herwig{} implementations of {\textsf{KrkNLO}}{}. For the comparisons, we
consider both the $q{\bar q}$ and $qg$ NLO channels of the DY process,
with the backward evolution of the parton shower
running to the end, as in the standard PSMC set-up.
In this set-up the argument of ${\alpha_s}$ in the hard-real NLO corrections is
the evolution variable $q^2$, {\it i.e. } ${\alpha_s}(q^2)$,
and in the virtual+soft-real correction it is set to $M_Z$, {\it i.e. } ${\alpha_s}(M_Z^2)$.
Here both programs use the $\rm MC_{DY}$ version of PDFs,
those in which only the parton--parton transitions
relevant to the DY process up to NLO are considered, see Ref.~\cite{Jadach:2015mza}.
Once again, we start from the comparison of the total cross section
-- the results are collected in
Table~\ref{tab:Z-nlo-settings-val}.
We see that both implementations of the {\textsf{KrkNLO}}{} method give cross sections
that agree at the per-mille level -- the tiny residual discrepancy is due to the
aforementioned differences in the low-$p_T$ region between \herwig{} and \sherpa{}
which affect the {\textsf{KrkNLO}}{} correcting weights.
These values also agree with our previous results in Ref.~\cite{Jadach:2015mza}
(see Table~5 there).
\begin{figure}[!h]
\centering
\includegraphics[width=0.46\textwidth]{Z_NLO_PteeFullPS.pdf}
\includegraphics[width=0.46\textwidth]{Z_NLO_PteeFullPSlog_mcfm.pdf}\\
\includegraphics[width=0.46\textwidth]{Z_NLO_Z_mass.pdf}
\includegraphics[width=0.46\textwidth]{Z_NLO_yZmcfm.pdf}\\
\caption{\sf
Comparisons of the $Z/\gamma^*$ distributions from the
{\textsf{KrkNLO}}{} method, as implemented in \sherpa{} and \herwig{},
for the Drell--Yan process with $e^+e^-$ pairs in the final state,
see text for details.
}
\label{fig:NLO-2ch-pt}
\end{figure}
In Fig.~\ref{fig:NLO-2ch-pt}, we show similar distributions
as in Fig.~\ref{fig:LO-Z-pt}, but this time include
the NLO QCD corrections according to the {\textsf{KrkNLO}}{} method.
Again, a good agreement between the two programs is found.
Only in the low $p_T$ region of the $p_{Tee}$ distributions are some differences visible,
but they reflect effects already seen in Fig.~\ref{fig:LO-Z-pt}.
Given this agreement we are able to validate our implementation
of the {\textsf{KrkNLO}}{} method in \herwig{}.
\subsubsection{KrkNLO results: MC vs. $\bf MC_{DY}$ prescriptions}
\label{sssec:dymc}
\begin{table}[!h]
\centering
\begin{tabular}{|c||c|c|c|}
\hline
& \mcfm{}: $\overline{\text{MS}}$ PDFs & {\textsf{KrkNLO}}{}: $\rm MC_{DY}$ PDFs & {\textsf{KrkNLO}}{}: MC PDFs\\
\hline \hline
$\sigma_\text{tot}$ [pb] & $1086.5\,(1)$ & $1046.5\,(2)$
& $1022.3\,(2) $ \\
\hline
\end{tabular}
\caption{\sf
Values of the total cross section with statistical errors
for the Drell--Yan process from {\textsf{KrkNLO}}{}
implemented in \herwig{} for two variants of MC-scheme PDFs
compared with the fixed-order NLO result from \mcfm{},
see text for details.
}
\label{tab:Z-nlo-PDF-MC-MCDY}
\end{table}
\begin{figure}[!h]
\centering
\includegraphics[width=0.46\textwidth]{Z_NLO_PDF_MC_DY_PteeFullPS.pdf}
\includegraphics[width=0.46\textwidth]{Z_NLO_PDF_MC_DY_PteeFullPSlog_mcfm.pdf}\\
\includegraphics[width=0.46\textwidth]{Z_NLO_PDF_MC_DY_Z_mass.pdf}
\includegraphics[width=0.46\textwidth]{Z_NLO_PDF_MC_DY_yZmcfm.pdf}\\
\caption{\sf
Comparisons of the $Z/\gamma^*$ distributions, as in the previous plots,
from {\textsf{KrkNLO}}{} implemented in \herwig{}
for both the $\rm MC_{DY}$ PDFs and the complete MC PDFs.
In the {\textsf{KrkNLO}}{}, the NLO weights use ${\alpha_s}(q^2)$
in the hard-real corrections and ${\alpha_s}(M_Z^2)$ in the virtual+soft-real ones.
}
\label{fig:NLO-2ch-pt-PDF-MC-MCDY}
\end{figure}
Having validated the implementation of the {\textsf{KrkNLO}}{} method in \herwig{} we are
ready to present the first new results. We start from the comparison of the
{\textsf{KrkNLO}}{} predictions based on the $\rm MC_{DY}$ PDFs, defined in
Ref.~\cite{Jadach:2015mza}, with those in which the complete MC-scheme PDFs,
first introduced in Ref.~\cite{Jadach:2016acv}, are used. The main difference
between the MC$_\text{DY}$ and MC PDFs is that in the former
only the quark PDFs are transformed from
the $\overline{\text{MS}}$ to MC scheme and the gluon PDF is unchanged, whereas in the latter
the gluon PDF is also transformed to the MC scheme;
this is described in Ref.~\cite{Jadach:2016acv}.
We note that for the DY process the transformation of
the gluon PDF comes as an effect beyond NLO, so formally, for any predictions at
the NLO level, the $\rm MC_{DY}$ PDFs are sufficient.
However, this is not the case for processes
in which initial-state gluons are present at the LO level, as is the case for
Higgs boson production in gluon--gluon fusion. Therefore, for future
applications of the {\textsf{KrkNLO}}{} method to a generic process we shall use the
complete MC PDFs of Ref.~\cite{Jadach:2016acv}.
From Table~\ref{tab:Z-nlo-PDF-MC-MCDY} we see that the differences
between the values of the total cross section
corresponding to these two variants of the MC-scheme PDFs are rather small,
$\sim 2\%$, well within uncertainties of the NLO predictions. For comparison,
we also show the fixed-order NLO result obtained from \mcfm{} using the $\overline{\text{MS}}$ PDFs.
The differences with respect to the {\textsf{KrkNLO}}{} results are at the level of
$4$--$6\%$, also within the NLO accuracy.
For the distributions of the quantities as shown in the previous plots,
presented in Fig.~\ref{fig:NLO-2ch-pt-PDF-MC-MCDY},
the differences are at the level of a few per-cent,
except for in the low $p_T$ region where they can grow up to $\sim 50\%$,
but this region is very sensitive to soft gluon effects
(and thus to the gluon PDF) that are formally beyond the NLO approximation.
\subsection{Higgs-boson production}
\label{ssec:higgs}
In this section we present results for Higgs-boson production in
gluon--gluon fusion at the LHC obtained with the {\textsf{KrkNLO}}{} method implemented in
\herwig{} and compare them with predictions of other NLO-PSMC matching
approaches, namely \mcatnlo{}~\cite{Frixione:2006he} and
\powheg{}~\cite{Frixione:2007vw} as implemented in \herwig{},
as well as with fixed-order calculations at
NLO and NNLO from \hnnlo{}~\cite{Catani:2007vq}
and an NNLL+NNLO calculation from \hqt{}~\cite{deFlorian:2011xf,Bozzi:2005wk}.
We opt to use the {\tt CT10nlo} PDF
set \cite{Lai:2010vv} which has ${\alpha_s}(M_Z^2)=0.118$, and set \herwig{}
equivalently, such that we have a consistent ${\alpha_s}$ setting across all
predictions\footnote{Aside from small differences in the running ${\alpha_s}$
between \hnnlo{} and \herwig{}.}.
Finally, we confront the theoretical predictions
of all the above matching methods with experimental
measurements performed at the LHC during Run~1
by the ATLAS collaboration~\cite{Aad:2015lha}.
\subsubsection{LO results}
\label{sssec:hlo}
\begin{table}[!h]
\centering
\begin{tabular}{|c||c|c|c|}
\hline
& \hnnlo{@LO} & \herwig{} \\
\hline\hline
$\sigma_\text{tot}$ [pb] & $5.565\,(1)$ & $ 5.564\,(1)$ \\
\hline
\end{tabular}
\caption{\sf
Values of the total cross section with statistical errors (in brackets) at the LO level for
Higgs production in gluon--gluon fusion at the LHC
for the $\overline{\text{MS}}$ CT10nlo (${\alpha_s}=0.118$) PDFs from \hnnlo{} and \herwig{} (fixed order),
see text for details.
}
\label{tab:H-lo-settings-val}
\end{table}
We start from the numerical cross-check
at the LO level of different programs used in our study.
In Table~\ref{tab:H-lo-settings-val} we show the results for the total cross section obtained
from \hnnlo{} and \herwig{} (fixed order). These values are in very good agreement,
despite small differences in the running of ${\alpha_s}$.%
We are therefore assured that all of the parameters
as well as PDFs used in computation of Higgs-boson production
in gluon--gluon fusion are consistently set in these programs.
\subsubsection{NLO results}
\label{sssec:hnlo}
\begin{table}[!h]
\centering
\begin{tabular}{|c||c|c|c|c|c|c|}
\hline
& & \multicolumn{2}{c|}{\powheg} & & \multicolumn{2}{c|}{\hnnlo{}} \\
\cline{3-4} \cline{6-7}
& \mcatnlo{} & Default & Original &{\textsf{KrkNLO}}{} & NLO & NNLO \\
\hline \hline
$\sigma_\text{tot}$ [pb] & $ 12.700\,(2)$ & $ 12.699\,(2)$
& $12.697\,(2) $ & $12.939\,(2) $ & $12.640\,(1)$ & $17.063\,(15)$\\
\hline
\end{tabular}
\caption{\sf
Values of the NLO total cross section with statistical errors (in brackets) for Higgs-boson production in gluon--gluon fusion at the LHC for {\textsf{KrkNLO}}{}, \mcatnlo{} and \powheg{} as calculated by \herwig{},
as well as the NLO and NNLO result from \hnnlo{}, see text for details.
}
\label{tab:H-nlo-settings-val-cteq}
\end{table}
Here we present the results from {\textsf{KrkNLO}}{} alongside those of the \mcatnlo{} and
\powheg{} methods implemented in \herwig{} as well as the NLO and NNLO results from
\hnnlo{} as well as a results from \hqt{}.
The {\textsf{KrkNLO}}{} setup uses ${\alpha_s}(q^2)$ for the hard-real corrections and ${\alpha_s}(M_H^2)$ for the virtual+soft-real corrections.
We show two variants of the \powheg{}
method: The first one, \powheg{~(Default)}, is the default set-up in \herwig{}
and it restricts the transverse momentum of parton-shower emissions to be less than
the factorization scale, as is done in the \mcatnlo{} setup, which follows the
initial work of Ref.~\cite{Frixione:2002ik}; the second one, \powheg{~(Original)},
is closer to its original implementation~\cite{Frixione:2007vw} which has no such
restriction.
This amounts to \powheg{~(Default)} generating both S and H-events, with
\powheg{~(Original)} only generating S-events.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\textwidth]{H_NNLO_CT10_XS.pdf}
\caption{\sf
Comparison of the total cross-section for Higgs-boson production in gluon-gluon
fusion at NLO, from \hnnlo{} and {\textsf{KrkNLO}}{}, as well as at NNLO from \hnnlo{}. The
error bars, shown for \hnnlo{}, are obtained from the independent variations of
the renormalization and factorization scales by factors of $1/2$ and $2$ from $M_H$.
}
\label{fig:NLO-H-XS}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.46\textwidth]{H_NLO_CT10_X_pT.pdf}
\includegraphics[width=0.46\textwidth]{H_NLO_CT10_X_pT_peak.pdf}\\
\includegraphics[width=0.46\textwidth]{H_NLO_CT10_X_y.pdf}
\includegraphics[width=0.46\textwidth]{H_NLO_CT10_jet1_pT.pdf}
\caption{\sf
Comparisons of the Higgs-boson transverse momentum and rapidity distributions from the
{\textsf{KrkNLO}}{}, \mcatnlo{} and \powheg{} methods implemented in \herwig{}
for Higgs-boson production in gluon--gluon fusion at the LHC, see text for details.
}
\label{fig:NLO-H}
\end{figure}
The values of the total cross section, with statistical errors, for the Higgs-boson
production process are shown in Table~\ref{tab:H-nlo-settings-val-cteq}.
We can see that, as expected, both the \mcatnlo{} and \powheg{} results
give the same total cross sections.
The {\textsf{KrkNLO}}{} method gives a slightly higher value of the cross section
than the other methods.
This can be explained by the beyond-NLO contributions that are partially
accounted for in the {\textsf{KrkNLO}}{} result.
Additionally, in Fig.~\ref{fig:NLO-H-XS} we show the total
cross-sections from \hnnlo{}, along with error bands generated by the
variation of the renormalization and factorization scales by factors of $1/2$ and $2$
around $M_H$.
We see that the prediction of {\textsf{KrkNLO}}{} is within the NLO
uncertainty estimate. Furthermore, we see that the uncertainty estimate for NNLO
is still rather large, at around $10\%$, and does not overlap with the NLO range.
\begin{figure}[!h]
\centering
\includegraphics[width=0.79\textwidth]{H_NNLO_CT10_X_pT_peak_HNNLO.pdf}
\\
\includegraphics[width=0.79\textwidth]{H_HqT_CT10_X_pT_peak_HNNLO.pdf}
\caption{\sf
Higgs-boson transverse-momentum distributions from
{\textsf{KrkNLO}}{}, \powheg{} and \mcatnlo{}. The upper plot compares our results with the
fixed-order NNLO result from the \hnnlo{} program and are shown relative to the
NLO results from \hnnlo{}. The lower plot shows our results in comparison to \hqt{},
these are shown relative to the \hqt{} NNLO prediction. The content of the error
bands is described in the main text.
}
\label{fig:NNLO-H}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.46\textwidth]{H_NLO_KrkNLO_pdf_X_pT.pdf}
\includegraphics[width=0.46\textwidth]{H_NLO_KrkNLO_pdf_X_pT_peak.pdf}\\
\includegraphics[width=0.46\textwidth]{H_NLO_KrkNLO_pdf_X_y.pdf}
\caption{\sf
Comparisons of the Higgs-boson transverse momentum and rapidity distributions from the
{\textsf{KrkNLO}}{} method using different PDF sets in the MC factorization scheme
for the Higgs-boson production in gluon--gluon fusion at the LHC, see text for details.
}
\label{fig:NLO-H-pdf}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{H_NLO_CT10_d01-x01-y01.pdf}
\includegraphics[width=0.49\textwidth]{H_NLO_CT10_d02-x01-y01.pdf}\\
\includegraphics[width=0.49\textwidth]{H_NLO_CT10_d03-x01-y01.pdf}
\includegraphics[width=0.49\textwidth]{H_NLO_CT10_d04-x01-y01.pdf}\\
\caption{\sf
Comparisons of the predictions of the {\textsf{KrkNLO}}{}, \mcatnlo{} and \powheg{}
methods implemented in \herwig{} for Higgs-boson production
with the ATLAS data from LHC Run~1.
The gluon--gluon fusion results from \herwig{} are plotted on top of the XH
results from Ref.~\cite{Aad:2015lha}.
}
\label{fig:NLO-H-ATLAS}
\end{figure}
In Fig.~\ref{fig:NLO-H} we present the distributions
of the Higgs-boson transverse momentum $p_T^\text{H}$
and rapidity $y_\text{H}$, comparing the results from {\textsf{KrkNLO}}{}
with the ones from \mcatnlo{} and \powheg{}.
All of the predictions agree within a few per-cent for the $p_T^\text{H}$ range from
$\sim 5\,$GeV to $\sim 100\,$GeV and for the rapidity range $[-3,3]$.
In the region $p_T^\text{H} > M_H$ larger differences between {\textsf{KrkNLO}}{} and
\mcatnlo{}/\powheg{~(Default)} are visible, whereas {\textsf{KrkNLO}}{} produces a similar shape to \powheg{~(Original)}.
The main reason for this is that both \mcatnlo{} and \powheg{~(Default)} restrict
the value of the transverse momentum of parton-shower emissions
to be below the value of $M_H$,
whereas for {\textsf{KrkNLO}}{}, and also \powheg{~(Original)}, there is no such a restriction;
this can be seen in the upper-left plot, where a sharp drop of the $p_T$
spectrum for $p_T \gtrsim M_H$
is visible in the case where the emissions are restricted
by this hard-cutoff. However, this spectrum can be smoothed by relaxing this
condition, as shown in Refs.~\cite{Alwall:2014hca, Bellm:2016rhh}.
Of course, such differences are acceptable within the NLO approximation.
Next, in Fig.~\ref{fig:NNLO-H} we compare {\textsf{KrkNLO}}{},
\mcatnlo{} and \powheg{} predictions from \herwig{} for the
Higgs-boson transverse momentum distribution
with the corresponding result obtained from
\hnnlo{}~\cite{Catani:2007vq, Grazzini:2008tf} and \hqt{}~\cite{deFlorian:2011xf,Bozzi:2005wk}.
The error bands \hnnlo{} NNLO distributions were obtained by varying the
renormalization and factorization scales by the typical factors of $1/2$ and $2$
around $M_H$ as an estimate of the uncertainty from neglected higher orders. The
\hqt{} distributions were obtained similarly, but also include variations of the
resummation scale of $1/2$ and $2$ at the central value of $M_H/2$ as
recommended in Ref.~\cite{Bozzi:2005wk}.
The \hnnlo{} comparison, upper plot, is shown relative to the NLO distribution
from \hnnlo{} and the \hqt{} comparison, lower plot, is shown relative to the NNLO
distribution from \hqt{}.
As we see in the upper plot of Fig.~\ref{fig:NNLO-H}, both the {\textsf{KrkNLO}}{} and the
NNLO results show the same trends, quickly raising above the NLO result at low and moderate
$p_T^\text{H}$ and remaining above it at high $p_T^\text{H}$.
The fact that our method gives a result that is higher than the NLO one
at high $p_T^\text{H}$
is a consequence of the mixed real-virtual $\order{{\alpha_s}^2}$ terms, which
constitute part of the NNLO correction and arise because of the multiplicative
nature of the {\textsf{KrkNLO}}{} approach.
In the upper plot of Fig.~\ref{fig:NNLO-H} we also show similar comparisons with NNLO
for \mcatnlo{} and two versions of \powheg{}.
The behaviour at low $p_T^\text{H}$ is close to that
from {\textsf{KrkNLO}}{}. At high $p_T^\text{H}$, however,
\mcatnlo{} and \powheg{~(Default)}, by
construction, converge to the NLO results,
departing from the NNLO predictions. On the other hand,
{\textsf{KrkNLO}}{} and \powheg{~(Original)} are closer to the NNLO predictions but for larger $p_T^\text{H}$ values
they are marginally harder.
The lower plot of Fig.~\ref{fig:NNLO-H} shows results from \mcatnlo{}, \powheg{} and {\textsf{KrkNLO}}{} compared to the resummed calculation from \hqt{} (for the ``switched'' option).
All of the NLO+PS give similar results up to roughly 80~GeV. The \hqt{} result peaks towards lower values of $p_T^\text{H}$ compared to the other predictions.
In Fig.~\ref{fig:NLO-H-pdf} we show the results of the {\textsf{KrkNLO}}{} method
obtained for different modern $\overline{\text{MS}}$
PDF sets: {\tt CT10nlo} (as used in this section), {\tt CT14lo}, {\tt HERAPDF20},
{\tt MMHT2014lo} and {\tt NNPDF30lo}.
Except for {\tt CT10}, we have used the LO versions of the corresponding PDF parametrizations,
since at NLO they become negative at some regions of the $x$ variable and this poses a problem for
the \herwig{} PSMC generator.
We can see that the distributions can vary even by up to 40\%. The biggest
differences are observed at low transverse momenta and large rapidities.
In Appendix~\ref{app:pdfs} we compare all the different PDFs in the $\overline{\text{MS}}$ and
MC schemes and show that the differences at the level of parton distribution
functions, see Fig.~\ref{fig:pdfs-comp}, are commensurate to those observed in
Fig.~\ref{fig:NLO-H-pdf} for the differential cross sections. Further studies of
systematic effects due to PDFs are left for the future.
Finally, in Fig.~\ref{fig:NLO-H-ATLAS} we compare the predictions
for the distributions of the Higgs-boson transverse
momentum and rapidity, the number of jets and the transverse momentum
of the hardest jet from {\textsf{KrkNLO}}{}, \mcatnlo{}
and \powheg{} with the ATLAS data collected in LHC Run~1
with a collision energy of $\sqrt{s} = 8\,$TeV~\cite{Aad:2015lha}.
To our generated results for the gluon--gluon fusion we have added the contributions
from other Higgs-production channels, denoted XH, taken from Ref.~\cite{Aad:2015lha}
-- they account for $\sim 12\%$ of the cross section.
All of the data points lie above the theoretical predictions,
although the experimental errors are large,
The NLO-PSMC matching methods offer largely equivalent predictions
with {\textsf{KrkNLO}}{} and \powheg{~(Original)}
predicting slightly harder spectrum for high $p_\perp$ and higher rates for larger jet multiplicities
(similar trends are also seen in Fig.~\ref{fig:NLO-H}).
Further differences were previously
discussed in the context of Fig.~\ref{fig:NLO-H}.
In our simulations we have used the {\tt CT10nlo} PDF parametrization, the same that was used
in Ref.~\cite{Aad:2015lha} for theoretical predictions. However, we have checked that
changes of our results due to the use of different PDF sets, discussed in Appendix A,
are much smaller than the experimental errors of the ATLAS data and negligible
compared to the differences between these results and the data points.
\section{Conclusions and outlook}
In this paper we have presented the numerical results
of the {\textsf{KrkNLO}}{} method for the Drell--Yan (DY) process and Higgs-boson
production in gluon--gluon fusion at the LHC
for the collision energy $\sqrt{s} = 8\,$TeV.
Firstly, we have validated the implementation of the {\textsf{KrkNLO}}{} method
in the \herwig{} PSMC by comparing its results
for the DY process with previous results obtained with the \sherpa{}-based implementation.
Then, we have presented new results for the DY process
with the complete MC-scheme PDFs that were recently defined
in Ref.~\cite{Jadach:2016acv}.
These results have been compared with those for the older variant of the MC-scheme PDFs,
called here the $\rm MC_{DY}$ PDFs,
that were introduced in Ref.~\cite{Jadach:2015mza} exclusively for the DY process.
The agreement between the results for these two variants of the MC-scheme PDFs
has been found at the level of $\sim 2\%$, which is well within the NLO accuracy.
Our main numerical results in this paper concern the Higgs-boson production process.
We have presented the results for the total cross section,
as well as distributions of the Higgs-boson transverse momentum
and rapidity.
The {\textsf{KrkNLO}}{} predictions have been compared with those
of the \mcatnlo{} and \powheg{} methods.
A good agreement, within the NLO accuracy,
has been found between the default options of these methods.
For $p_T^\text{H} > M_H$ the {\textsf{KrkNLO}}{} result shows better agreement with the
\powheg{~(Original)} option, a result of the restriction on the transverse momentum
of parton-shower emissions to below the factorization scale present in the other setups.
Finally, the theoretical predictions of the above NLO-PSMC matching
methods have been compared with the ATLAS data
from LHC Run~1 for several observables for the Higgs-production process.
All of the matching methods
underestimate the ATLAS measurements, however the experimental errors are large.
The {\textsf{KrkNLO}}{} results offer comparable predictions to other matching methods in
all distributions and, along with \powheg{ (Original)},
predict marginally harder spectrum for high $p_T$ and larger jet multiplicities.
It is worth mentioning that all the calculations are performed in the infinite top-quark mass
approximation, therefore including finite quark mass effects, which are important
for large transverse momenta, would most likely bring the predictions
closer to the experimental data.
As a next step in our numerical predictions with the {\textsf{KrkNLO}}{} method
we plan to perform a more detailed study
of the Drell--Yan processes,
involving both the neutral ($Z/\gamma^*$) and charged ($W^{\pm}$) modes,
in presence of experimental cuts and a focus on leptonic observables.
In order to do this, we need to supplement the NLO-correcting weights with
appropriate spin correlations for vector-boson decay products
(which is not needed in the case of the scalar Higgs boson).
This can be done rather easily within the \herwig{} PSMC algorithm
using the method proposed in Ref.~\cite{Seymour:1994we}.
Future work will seek to apply the {\textsf{KrkNLO}}{} method to other processes investigated
at the LHC, first of all looking at electroweak-boson pair-production
($VV$, where $V=\gamma^*,Z,W^+,W^-$) and $V+$jet production.
This would be an important test of feasibility and universality of the method.
Future work will also comprise an appraisal of uncertainties of the
{\textsf{KrkNLO}}{} approach, similar to that of Ref.~\cite{Bellm:2016voq}. Beyond this, the
natural extension for {\textsf{KrkNLO}}{} is NNLO + NLOPS, where NLOPS is PSMC that
implements the NLO evolution kernels in the fully exclusive form, and thus
provides the full set of the soft-collinear counter-terms for the hard process.
Ref.~\cite{Jadach:2013dfd} reviews several feasibility studies showing that
construction of such a NLOPS is, in principle, plausible. Moreover,
the simplifications due to the {\textsf{KrkNLO}}{} method in achieving NLO+PS will, in our opinion,
be instrumental towards these more ambitions research directions.
The {\textsf{KrkNLO}}{} project will be available on {\tt hepforge} at
\url{https://krknlo.hepforge.org}. This will become the home site of the {\textsf{KrkNLO}}{}
development, containing relevant codes and the MC-scheme PDFs as well as set-up
instructions to facilitate its use within \herwig{}.
|
\section{The BCPI Problem}\label{sec:BCPI}
\section{Balancing the Supply/Demand Only}\label{sec:BCPI}
In this section, we study the single objective problem of
finding a partition of the graph into connected subgraphs that balances (approximately)
the supply and demand in each part of the partition, without any regard
to the sizes of the parts.
We can state the optimization problem as follows, and will refer to it as the Balanced Connected Partition with Integer weights (BCPI) problem.
\begin{definition}
Given a graph $G=(V,E)$ with a weight (supply/demand) function $p~:~V~\rightarrow~\mathbb{Z}$ satisfying $\sum_{j\in V} p(j)=0$. The BCPI problem is the problem of partitioning $V$ into $(V_1,V_2)$ such that
\begin{enumerate}
\item $V_1\cap V_2=\emptyset$ and $V_1\cup V_2=V$,
\item $G[V_1]$ and $G[V_2]$ are connected,
\item $|p(V_1)|+|p(V_2)|$ is minimized, where $p(V_i)=\sum_{j\in V_i} p(j).$
\end{enumerate}
\end{definition}
Clearly, the minimum possible value for $|p(V_1)|+|p(V_2)|$ that we can hope for is 0,
which occurs iff $p(V_1)=p(V_2) =0$.
It is easy to show that the problem of determining whether there exists such a `perfect' partition
(and hence the BCPI problem) is strongly NP-hard.
The proof is very similar to analogous results concerning the partition of a graph
into two connected subgraphs with equal sizes (or weights, when nodes have positive weights)
\cite{chataigner2007approximation,dyer}
\begin{proposition}\label{lem:BCPI_hard}
(1) It is strongly NP-hard to determine whether there is a solution to the BCPI problem with value 0, even when $G$ is 2-connected. \\
(2) If $G$ is not 2-connected, then this problem is NP-hard even when $\forall i, p(i)=\pm1$.
\end{proposition}
Although it is NP-hard to tell whether there is a solution satisfying $p(V_1)=p(V_2) =0$,
even when $\forall i, p(i)=\pm1$, in this case, if the graph $G$ is 2-connected there is always such a solution.
For general weights $p$, there is a solution such that $|p(V_1)|,|p(V_2)|\leq \max_{j\in V}|p(j)|/2$
and it can be found easily in linear time using the $st$-numbering between two nodes (see the Appendix \ref{sec:proof1}).
\begin{proposition}\label{lem:2_poly}
Let $G$ be a 2-connected graph and $u,v$ any two nodes in $V$ such that $p(u) p(v)>0$.\\
(1) There is a solution such that $u\in V_1$, $v\in V_2$, and $|p(V_1)|=|p(V_2)|\leq \max_{j\in V}|p(j)|/2$.\\
(2) If $\forall i, p(i)=\pm1$, we can find a solution such that $u\in V_1$, $v\in V_2$, and $p(V_1)=p(V_2)=0$.\\
In both cases, the solution can be found in $O(|E|)$ time.
\end{proposition}
\begin{remark}
The bound in Proposition~\ref{lem:2_poly} (1) is tight. A simple example is a cycle of length 4 like $(v_1,v_2,v_3,v_4)$ with $p(v_1)=-p$, $p(v_2)=-p/2$, $p(v_3)=p$, and $p(v_4)=p/2$. It is easy to see that in this example $|p(V_1)|+|p(V_2)|= \max_{j\in V}|p(j)|=p$ is the best that one can do.
\end{remark}
\subsection{Connected Partitioning into Many Parts}
The BCPI problem can be extended to partitioning a graph into $k=3$ or more parts.
Let $G=(V,E)$ be a graph with a weight function $p~:~V~\rightarrow~\mathbb{Z}$ satisfying $\sum_{j\in V} p(j)=0$. The $\text{BCPI}_k$ problem is the problem of partitioning $G$ into $(V_1,V_2,\dots,V_k)$ such that for any $1\leq i\leq k$, $G[V_i]$ is connected and $\sum_{i=1}^k |p(V_i)|$ is minimized.
In the following proposition, we show that for $k=3$, if $ p(i)=\pm1, \forall i$, then
there is always a perfect partition (i.e., with $p(V_1)=p(V_2)=p(V_3)=0$) and it can be found efficiently.
For general $p$, we can find a partition such that $|p(V_1)|+|p(V_2)|+|p(V_3)|\leq 2 \max_{j\in V}|p(j)|$.
The proof and the algorithm use a similar approach as the algorithm in~\cite{wada1994efficient}
for partitioning a 3-connected graph to three connected parts with prescribed sizes, using the nonseparating ear decomposition of 3-connected graphs as described in Subsection~\ref{subsec:ear-decomposition}.
The proof is given in the Appendix \ref{sec:proof1}.
\begin{proposition}\label{lem:3-connected-zero-p}
Let $G$ be a 3-connected graph and $u,v,w$ three nodes in $V$ such that $p(u), p(v),p(w)>0$ or $p(u), p(v),p(w)<0$.\\
(1) There is a solution such that $u\in V_1$, $v\in V_2$, $w\in V_3$, and $|p(V_1)|+|p(V_2)|+|p(V_3)|\leq 2 \max_{j\in V}|p(j)|$. \\
(2) If $\forall i, p(i)=\pm1$, then there is a solution such that $u\in V_1$, $v\in V_2$, $w\in V_3$, and
$|p(V_1)|=|p(V_2)|=|p(V_3)|=0$.\\
In both cases, the solution can be found in $O(|E|)$ time.
\end{proposition}
\section{Balancing Both Objectives}\label{sec:BPGI}
In this section, we formally define and study the Doubly Balanced Connected graph Partitioning (DBCP) problem.
\begin{definition}
Given a graph $G=(V,E)$ with a weight (supply/demand) function $p~:~V~\rightarrow~\mathbb{Z}$ satisfying $p(V)=\sum_{j\in V} p(j)=0$ and constants $c_p\geq 0$, $c_s\geq 1$. The \emph{DBCP problem} is the problem of partitioning $V$ into $(V_1,V_2)$ such that
\begin{enumerate}
\item $V_1\cap V_2=\emptyset$ and $V_1\cup V_2=V$,
\item $G[V_1]$ and $G[V_2]$ are connected,
\item $|p(V_1)|,|p(V_2)|\leq c_p$ and $\max\{\frac{|V_1|}{|V_2|},\frac{|V_2|}{|V_1|}\}\leq c_s$, where $p(V_i)=\sum_{j\in V_i} p(j).$
\end{enumerate}
\end{definition}
\begin{remark}
Our techniques apply also to the case that $p(V)\neq0$.
In this case, the requirement 3 on $p(V_1)$ and $p(V_2)$ is
$|p(V_1)-p(V)/2|,|p(V_2)-p(V)/2|\leq c_p$, i.e., the excess supply/demand
is split approximately evenly between the two parts.
\end{remark}
We will concentrate on 2-connected and 3-connected graphs and
show that we can construct efficiently good partitions.
For most of the section we will focus on the case that $p(i)=\pm1, \forall i \in V$.
This case contains all the essential ideas.
All the techniques generalize to the case of arbitrary $p$,
and we will state the corresponding theorems.
We observed in Section 2 that if the graph is 2-connected and $p(i)=\pm1, \forall i \in V$ then there is always a connected partition
that is perfect with respect to the weight objective,
$p(V_1)=p(V_2)=0$, i.e., (3) is satisfied with $c_p=0$.
We know also from \cite{lovasz1977homology,gyori1976division}
that there is always a connected partition
that is perfect with respect to the size objective,
$|V_1|=|V_2|$, i.e., condition 3 is satisfied with $c_s=1$.
The following observations show that combining the two objectives makes
the problem more challenging.
If we insist on $c_p=0$, then $c_s$ cannot be bounded in general,
(it will be $\Omega(|V|)$),
and if we insist on $c_s=1$, then $c_p$ cannot be bounded.
The series-parallel graphs of Figure~\ref{fig:example_bound} provide simple counterexamples.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{Figures/BoundBalance.pdf}
\caption{Series-parallel graphs with $2s+1$ paths of length $4t+2$ used in Observations~\ref{obs:p=0} and \ref{obs:s=1}.}
\label{fig:example_bound}
\end{figure}
\begin{observation}\label{obs:p=0}
If $c_p=0$, then for any $c_s<|V|/2-1$, there exist a 2-connected graph $G$ such that the DBCP problem does not have a solution even when $\forall i, p(i)=\pm1$.
\end{observation}
\begin{proof}
In the graph depicted in Figure~\ref{fig:example_bound}, set $t=0$.
\end{proof}
\begin{observation}\label{obs:s=1}
If $c_s=1$, then for any $c_p<|V|/6$, there exist a 2-connected graph $G$ such that the DBCP problem does not have a solution even when $\forall i, p(i)=\pm1$.
\end{observation}
\begin{proof}
In the graph depicted in Figure~\ref{fig:example_bound}, set $s=1$.
\end{proof}
Thus, $c_p$ has to be at least 1 to have any hope for a bounded $c_s$.
We show in this section that $c_p=1$ suffices for all 2-connected graphs.
We first treat 3-connected graphs.
\subsection{3-Connected Graphs}\label{subsec:3-connected}
Let $G=(V,E)$ be a 3-connected graph.
Assume for the most of this section that $\forall i, p(i)=\pm1$ and $p(V)=0$
(we will state the results for general $p$ at the end).
We show that $G$ has a partition that is essentially perfect with
respect to both objectives, i.e., with $c_p=0$ and $c_s=1$.
We say ``essentially", because $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|$
imply that $|V_1|=|V_2|$ are even, and hence $V$ must be a multiple of 4.
If this is the case, then indeed we can find such a perfect partition.
If $|V|\equiv 2 (\mathrm{mod}~4)$ ($|V|$ has to be even since $p(V)=0$), then
we can find an `almost perfect' partition, one in which
$|p(V_1)|=|p(V_2)|=1$ and $|V_1|=|V_2|$ (or one in which
$p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|+2$).
We first treat the case that $G$ contains a triangle (i.e., cycle of length 3).
In the following Lemma, we use the embedding for $k$-connected graphs introduced in~\cite{linial1988rubber} and as described in Subsection~\ref{subsec:embedding}, to show that if $G$ is 3-connected with a triangle and all weights are $\pm 1$, then the DBCP problem has a perfect solution.
\begin{lemma}\label{lem:3-connected-tri}
If $G$ is 3-connected with a triangle, $\forall i, p(i)=\pm1$, and $|V|\equiv0(\mathrm{mod}~4)$, then there exists a solution to the DBCP problem
with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|$.
If $|V|\equiv2 (\mathrm{mod}~4)$, then there is a solution with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|+2$.
Moreover, this partition can be found in polynomial time.
\end{lemma}
\begin{proof}
Assume that $|V|\equiv0(\mathrm{mod}~4)$; the proof for the case $|V|\equiv2(\mathrm{mod}~4)$ is similar.
In~\cite{linial1988rubber} as described in Subsection~\ref{subsec:embedding}, it is proved that if $G$ is a $k$-connected graph, then for every $X\subset V$ with $|X|=k$, $G$ has a convex $X$-embedding in general position. Moreover, this embedding can be found by solving a set of linear equations of size $|V|$.
Now, assume $v,u,w\in V$ form a triangle in $G$. Set $X=\{v,u,w\}$. Using the theorem, $G$ has a convex $X$-embedding $f:V\rightarrow\mathbb{R}^2$ in general position. Consider a circle $\mathcal{C}$ around the triangle $f(u),f(v),f(w)$ in $\mathbb{R}^2$ as shown in an example in Fig.~\ref{fig:embedding_1}. Also consider a directed line $\mathcal{L}$ tangent to the circle $C$ at point $A$. If we project the nodes of $G$ onto the line $\mathcal{L}$, since the embedding is convex and also $\{u,v\},\{u,w\},\{w,v\}\in E$, the order of the nodes' projection gives an $st$-numbering between the first and the last node (notice that the first and last nodes are always from the set $X$). For instance in Fig.~\ref{fig:embedding_1}, the order of projections give an $st$-numbering between the nodes $u$ and $v$ in $G$. Hence, if we set $V_1$ to be the $|V|/2$ nodes whose projections come first and $V_2$ are the $|V|/2$ nodes whose projections come last, then $G[V_1]$ and $G[V_2]$ are both connected and $|V_1|=|V_2|=|V|/2$. The only thing that may not match is $p(V_1)$ and $p(V_2)$. Notice that for each directed line tangent to the circle $\mathcal{C}$, we can similarly get a partition such that $|V_1|=|V_2|=|V|/2$. So all we need is a point $D$ on the circle $\mathcal{C}$ such that if we partition based on the directed line tangent to $C$ at point $D$, then $p(V_1)=p(V_2)=0$. To find such a point, we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ ($AB$ is a diameter of the circle $\mathcal{C}$) and consider the resulting partition. Notice that if at point $A$, $p(V_1)>0$, then at point $B$ since $V_1$ and $V_2$ completely switch places compared to the partition at point $A$, $p(V_1)<0$. Hence, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ and keep it tangent to the circle, in the resulting partitions, $p(V_1)$ goes from some positive value to a non-positive value. Notice that the partition $(V_1,V_2)$ changes only if $\mathcal{L}$ passes a point $D$ on the circle such that at $D$, $\mathcal{L}$ is perpendicular to a line that connects $f(i)$ to $f(j)$ for some $i,j\in V$. Now, since the embedding is in general position, there are exactly two points on every line that connects two points $f(i)$ and $f(j)$, so $V_1$ changes at most by one node leaving $V_1$ and one node entering $V_1$ at each step as we move $\mathcal{L}$. Hence, $p(V_1)$ changes by either $\pm 2$ or $0$ value at each change. Now, since $|V|\equiv 0 (\mathrm{mod}~ 4)$, $p(V_1)$ has an even value in all the resulting partitions. Therefore, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$, there must be a point $D$ such that in the resulted partition $p(V_1)=p(V_2)=0$.
It is also easy to see that since $V_1$ may change only when a line that passes through 2 nodes of graph $G$ is perpendicular to $\mathcal{L}$, we can find $D$ in at most $O(|V|^2)$ steps.
\end{proof}
\begin{figure}[t]
\centering
\includegraphics[scale=0.75]{Figures/Embedding_1.pdf}
\caption{Proof of Lemma~\ref{lem:3-connected-tri}.}
\label{fig:embedding_1}
\end{figure}
When $G$ is a triangle-free 3-connected graph, however, the proof in Lemma~\ref{lem:3-connected-tri} cannot be directly used anymore. The reason is if for example $\{u,v\}\notin E$ and we project the nodes of $G$ onto the line $\mathcal{L}$, this time the order of the nodes projection does not give an $st$-numbering between the first and the last node if for example $u$ and $w$ are the first and last node, since some of the middle nodes may only be connected to $v$. To prove a similar result for triangle-free 3-connected case, we first provide the following two Lemmas. The main purpose of the following two Lemmas are to compensate for the triangle-freeness of $G$ in the proof of Lemma~\ref{lem:3-connected-tri}. The idea is to show that in every 3-connected graph, there is a triple $\{u,w,v\}\in V$, such that $\{u,w\},\{w,v\}\in E$ and in every partition that we get by the approach used in the proof of Lemma~\ref{lem:3-connected-tri}, if $u$ and $v$ are in $V_i$, so is a path between $u$ and $v$.
\begin{lemma}\label{lem:cut}
If $G$ is 3-connected, then there exists a set $\{u,v,w\}\in V$ and a partition of $V$ into $(V_1',V_2')$ such that:
\begin{enumerate}
\item $V_1'\cap V_2' =\emptyset$ and $V_1'\cup V_2'=V$,
\item $G[V_1']$ and $G[V_2']$ are connected,
\item $\{u,w\},\{v,w\}\in E$,
\item $w\in V_1'$, $u,v\in V_2'$,
\item $|V_2'|\leq |V|/2$.
\end{enumerate}
Moreover, such a partition and $\{u,v,w\}$ can be found in $O(|E|)$ time.
\end{lemma}
\begin{proof}
Using the algorithm presented in~\cite{cheriyan1988finding}, we can find a non-separating cycle $C_0$ in $G$ such that every node in $C_0$ has a neighbor in $G\backslash C_0$. Now, we consider two cases:
\begin{itemize}
\item[(i)] If $|C_0|\leq |V|/2+1$, then select any three consecutive nodes $(u,w,v)$ of $C_0$ and set $V_2'=C_0\backslash \{w\}$ and $V_1'=V\backslash V_2'$.
\item[(ii)] If $|C_0|> |V|/2+1$, since every node in $C_0$ has a neighbor in $G\backslash C_0$, there exists a node $w\in V\backslash C_0$ such that $|N(w)\cap C_0|\geq 2$. Select two nodes $u,v\in N(w)\cap C_0$. There exists a path $P$ in $C_0$ between $u$ and $v$ such that $|P|<|V|/2-1$. Set $V_2'=\{u,v\}\cup P$ and $V_1'=V\backslash V_2'$.\qedhere
\end{itemize}
\end{proof}
\begin{figure}[t]
\centering
\includegraphics[scale=0.9]{Figures/3-connected.pdf}
\caption{Proof of Lemma~\ref{lem:embedding} and Theorem~\ref{th:3-connected}.}
\label{fig:3-connected}
\end{figure}
\begin{lemma}\label{lem:embedding}
Given a partition $(V_1',V_2')$ of a 3-connected graph $G$ with properties described in Lemma~\ref{lem:cut}, $G$ has a convex $X$-embedding in general position with mapping $f:V\rightarrow \mathbb{R}^2$ such that:
\begin{enumerate}
\item $X=\{u,w,v\}$, $f(u)=(0,0)$, $f(v)=(1,0)$, and $f(w)=(0,1)$,
\item Every node $i$ in $V_1'$ is mapped to a point $(f_1(i),f_2(i))$ with $f_2(i)\geq 1/2$,
\item Every node $i$ in $V_2'$ is mapped to a point $(f_1(i),f_2(i))$ with $f_1(i)\geq f_2(i)$ and $f_1(i)+2f_2(i)\leq 1$.
\end{enumerate}
Moreover, such an embedding can be found in in polynomial time.
\end{lemma}
\begin{proof}[Sketch of the proof]
Set $X=\{v,u,w\}$. Using~\cite{linial1988rubber}, $G$ has a convex $X$-embedding in $\mathbb{R}^2$ in general position with mapping $f:V\rightarrow \mathbb{R}^2$ such that $f(u)=(0,0)$, $f(v)=(1,0)$, and $f(w)=(0,1)$. In the $X$-embedding of the nodes, we have a freedom to set the elasticity coefficient vector $\vec{c}$ to anything that we want (except a measure zero set of vectors). So for any edge $\{i,j\}\in G[V_1]\cup G[V_2]$, set $c_{ij}=g$; and for any $\{i,j\} \in E[V_1',V_2']$, set $c_{ij}=1$. Since both $G[V_1']$ and $G[V_2']$ are connected, as we increase $g$, nodes in $V_1'$ get closer to $w$ and nodes in $V_2'$ get closer to the line $uv$ (as $g\to \infty$, nodes in $V_1'$ get in the same position as $w$ and node in $V_2'$ get on the line $uv$). Hence, intuitively there exists a value $g$, for which all the nodes in $V_1$ are above line $\mathcal{L}_1$ and all the nodes in $V_2'$ are below the lines $\mathcal{L}_2$ and $\mathcal{L}_3$ as depicted in Fig.~\ref{fig:3-connected}. In the appendix we give the detailed proof and show that a $g$ with polynomially many bits suffices.
\end{proof}
Using Lemmas~\ref{lem:cut} and \ref{lem:embedding}, we are now able to prove that for any 3-connected graph $G$ such that all the weights are $\pm 1$, the DBCP problem has a solution for $c_p=0$ and $c_s=1$. The idea of the proof is similar to the proof of Lemma~\ref{lem:3-connected-tri}, however, we need to use Lemma~\ref{lem:cut} to find a desirable partition $(V_1',V_2')$ and then use this partition to find an embedding with properties as described in Lemma~\ref{lem:embedding}. By using this embedding, we
can show that in every partition that we obtain by the approach in the proof of Lemma~\ref{lem:3-connected-tri}, if $u$ and $v$ are in $V_i$, so is a path between $u$ and $v$. This implies then that $G[V_1]$ and $G[V_2]$ are connected. So we can use similar arguments as in the proof of Lemma~\ref{lem:3-connected-tri} to prove the following theorem (see the Appendix~\ref{sec:proof2} for the proof details).
\begin{theorem}\label{th:3-connected}
If $G$ is 3-connected, $\forall i, p(i)=\pm1$, and $|V|\equiv0(\mathrm{mod}~4)$,
then there exists a solution to the DBCP problem
with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|$.
If $|V|\equiv2(\mathrm{mod}~4)$, then there is a solution with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|+2$.
Moreover, this partition can be found in polynomial time.
\end{theorem}
It is easy to check for a 3-connected graph $G$, by using the same approach as in the proof of Lemma~\ref{lem:3-connected-tri} and Theorem~\ref{th:3-connected}, that even when the weights are arbitrary (not necessarily $\pm 1$) and also $p(V)\neq 0$, we can still find a connected partition $(V_1,V_2)$ for $G$ such that
$|p(V_1)-p(V)/2|,|p(V_1)-p(V)/2| \leq \max_{i\in V}|p(i)|$ and $|V_1|=|V_2|$.
\begin{corollary}\label{cor:3-general}
If $G$ is 3-connected, then the DBCP problem (with arbitrary $p$, and not necessarily satisfying $p(V)= 0$) has a
solution for $c_p=\max_{i\in V}|p(i)|$ and $c_s=1$. Moreover, this solution can be found in polynomial time.
\end{corollary}
\subsection{2-Connected Graphs}\label{subsec:2-connected}
We first define a \emph{pseudo-path} between two nodes in a graph as below. The definition is inspired by the definition of the $st$-numbering.
\begin{definition}
A \emph{pseudo-path} between nodes $u$ and $v$ in $G=(V,E)$, is a sequence of nodes $v_1,\dots,v_t$ such that if $v_0= u$ and $v_{t+1}=v$, then
for any $1\leq i\leq t$, $v_i$ has neighbors $v_j$ and $v_k$ such that $j<i<k$.
\end{definition}
Using the pseudo-path notion, in the following lemma we show that if $G$ is 2-connected and has a separation pair such that none of the resulting components are too large, then the DBCP problem always has a solution for some $c_p=c_s=O(1)$. The idea used in the proof of this lemma is one of the building blocks of the proof for the general 2-connected graph case.
\begin{lemma}\label{lem:parallel}
Given a 2-connected graph $G$ and an integer $q>1$, if $\forall i: p(i)=\pm1$ and $G$ has a separation pair $\{u,v\}\subset V$ such that for every connected component $G_i=(V_i,E_i)$ of $G[V\backslash \{u,v\}]$, $|V_i|< (q-1)|V|/q$, then the DBCP problem has a solution for $c_p=1$, $c_s=q-1$, and it can be found in $O(|E|)$ time.
\end{lemma}
\begin{proof}
Assume for simplicity that $|V|$ is divisible by $q$. There is a separation pair $\{u,v\}\in V$ such that if $G_1,\dots,G_k$ are the connected components of $G\backslash \{u,v\}$, for any $i$, $|V_i|< (q-1)|V|/q$. Since $G$ is 2-connected, $G_1,\dots,G_k$ can be presented by pseudo-paths $P_1,\dots,P_k$ between $u$ and $v$. Assume $P_1,\dots,P_k$ are in increasing order based on their lengths. There exists two subsets of the pseudo-paths $S_1$ and $S_2$ such that $S_1\cap S_2=\emptyset$, $S_1\cup S_2=\{P_1,\dots,P_{k}\}$ and $\sum_{P_j\in S_i} |P_j| \geq |V|/q-1$ for $i=1,2$. The proof is very simple. If $|P_k|< |V|/q$, the greedy algorithm for the partition problem gives the desired partition of $\{P_1,\dots,P_{k}\}$. If $|P_k|\geq |V|/q$, since $|P_k|< (q-1)|V|/q$, $S_1=\{P_1,\dots,P_{k-1}\}$ and $S_2=\{P_k\}$ is the desired partition.
Now, if we put all the pseudo-paths in $S_1$ back to back, they will form a longer pseudo-path $Q_1$ between $u$ and $v$. Similarly, we can form another pseudo-path $Q_2$ from the paths in $S_2$. Without loss of generality we can assume $|Q_1|\geq |Q_2|$. From $u$, including $u$ itself, we count $|V|/q$ of the nodes in $Q_1$ towards $v$ and put them in a set $V'$. Without loss of generality, we can assume $p(V')\geq 0$. If $p(V')=0$, then $(V',V\backslash V')$ is a good partition and we are done. Hence, assume $p(V')>0$. We keep $V'$ fixed and make a new set $V''$ by continuing to add nodes from $Q_1$ to $V'$ one by one until we get to $v$.
If $p(V'')$ hits 0 as we add nodes one by one, we stop and $(V'',V\backslash V'')$ is a good partition and we are done. So, assume $V''=Q_1\cup \{u\}$ and $p(V'')>0$. Since $|Q_2\cup\{v\}|\geq |V|/q$, $|V''|\leq (q-1)|V|/q$. If $|V''| < (q-1)|V|/q$, we add nodes from $Q_2$ one by one toward $u$ until either $|V''|=0$ or $|V''| = (q-1)|V|/q$. If we hit 0 first (i.e., $p(V'')=0$) and $|V''|\leq (q-1)|V|/q$, define $V_1=V''\backslash\{u\}$, then $(V_1,V\backslash V_1)$ is a good partition. So assume $|V''|=(q-1)|V|/q$ and $p(V'')>0$. Define $V'''=V\backslash V''$. Since $p(V'')>0$ and $|V''|=(q-1)|V|/q$, then $p(V''')<0$ and $|V'''|=|V|/q$. Also notice that $V'''\subseteq Q_2$. We consider two cases. Either $|p(V')|\geq|p(V''')|$ or $|p(V')|<|p(V''')|$.
If $|p(V')|\geq|p(V''')|$, then if we start from $u$ and pick nodes one by one from $Q_1$ in order, we can get a subset $V'_1$ of $V'$ such that $|p(V'_1)|=|p(V''')|$. Hence, if we define $V_1=V_1'\cup V'''$, then $(V_1,V\backslash V_1)$ is a good partition.
If $|p(V')|<|p(V''')|$, then we can build a new set $V_1$ by adding nodes one by one from $Q_2$ to $V'$ until $P(V_1)=0$. It is easy to see that since $|p(V')|<|p(V''')|$, then $V_1\backslash V'\subseteq V'''$. Hence, $(V_1, V\backslash V_1)$ is a good partition.
\end{proof}
\begin{corollary}\label{cor:serpar}
If $G$ is a 2-connected series-parallel graph and $\forall i: p(i)=\pm1$, then the DBCP problem has a solution for $c_p=1,~c_s=2$, and the solution can be found in $O(|E|)$ time.
\end{corollary}
The graph in Figure~\ref{fig:example_bound} with $s=1$ shows that these parameters are
the best possible for series parallel graphs: if $c_p=O(1)$ then $c_s$ must be at least 2.
To generalize Lemma~\ref{lem:parallel} to all 2-connected graphs, we need to define the \emph{contractible} subgraph and the \emph{contraction} of a given graph as below.
\begin{definition}\label{def:contraction}
We say an induced subgraph $G_1$ of a 2-connected graph $G$ is \emph{contractible}, if there is a separating pair $\{u,v\}\subset V$ such that $G_1=(V_1,E_1)$ is a connected component of $G[V\backslash\{u,v\}]$. Moreover, if we replace $G_1$ by a weighted edge $e'$ with weight $w(e')=|V_1|$ between the nodes $u$ and $v$ in $G$ to obtain a smaller graph $G'$, we say $G$ is \emph{contracted} to $G'$.
\end{definition}
\begin{remark}
Notice that every contractible subgraph of a 2-connected graph $G$ can also be represented by a pseudo-path between its associated separating pair. We use this property in the proof of Theorem~\ref{th:2-connected}.
\end{remark}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.6\textwidth}
\includegraphics[scale=0.5]{Figures/Transformation_1.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\includegraphics[scale=0.5]{Figures/Transformation_2.pdf}
\caption{}
\end{subfigure}
\caption{Lemma~\ref{lem:2-connected-pair}.}
\label{fig:lem_2-connected-pair}
\end{figure}
Using the notion of the graph contraction, in the following lemma, we show that to partition a 2-connected graph, we can reduce it to one of two cases: either $G$ can be considered as a graph with a set of short pseudo-paths between two nodes, or it can be contracted into a 3-connected graph as illustrated in Fig.~\ref{fig:lem_2-connected-pair}.
\begin{lemma}\label{lem:2-connected-pair}
In every 2-connected graph $G=(V,E)$, given an integer $q\geq 3$, one of the following cases holds, and we can determine which in $O(|E|)$ time:
\begin{enumerate}
\item There is a separation pair $\{u,v\}\subset V$ such that if $G_1,\dots,G_k$ are the connected components of $G[V\backslash \{u,v\}]$, for all $i$, $|V_i|< (q-1)|V|/q$.
\item After a set of contractions, $G$ can be transformed into a 3-connected graph $G^*=(V^*,E^*)$ with weighted edges representing contracted subgraphs such that for every $e^*\in E^*$, $w(e^*)<|V|/q$.
\end{enumerate}
\end{lemma}
\begin{proof}
If there is no separation pairs in $G$, then $G$ is 3-connected and there is nothing left to prove. So assume $\{u,v\}\subset V$ is a separation pair and $G_1,\dots,G_k$ are the connected components of $G[V\backslash \{u,v\}]$. If $\forall i, |V_i|< (q-1)|V|/q$, we are done. So let's assume there is a connected component $G_j$ such that $|V_j|\geq (q-1)|V|/q$. Then for every $i\neq j$, $G_i$ can be contracted and represented by an edge of weight less than $|V|/q$ between $u$ and $v$. Now, we repeat the process by considering the weight of the edges in the size of each connected component. An example for each case is shown in Fig.~\ref{fig:lem_2-connected-pair} for $q=3$. We can find either a suitable separation pair as in case 1 or a
suitable contracted graph $G^*$ as in case 2 in linear time using the Hopcroft-Tarjan
algorithm for finding the triconnected components \cite{hopcroft1973dividing}.
\end{proof}
Using Lemma~\ref{lem:2-connected-pair} for $q=4$, then Lemma~\ref{lem:parallel}, and the idea of the proof for Theorem~\ref{th:3-connected}, we can prove that when $G$ is 2-connected and all $p(i)=\pm 1$, the DBCP problem has a solution for $c_p=1$ and $c_s=3$. There are some subtleties in adapting Lemma~\ref{lem:cut} for this case to account for the fact that the edges of $G^*$ are now weighted, and the partition $(V_1', V_2')$ has to take into account the edge weights.
We find a suitable convex embedding of the 3-connected graph $G^*$
and then embed the nodes of the contracted pseudopaths appropriately
along the segments corresponding to the weighted edges. Some care is needed
to carry out the argument of the 3-connected case, since as the line
tangent to the circle rotates, the order of the projections of many nodes may change at once, namely the nodes on an edge perpendicular to the rotating line.
The details of the proof are given in Appendix~\ref{sec:proof2}.
We have:
\begin{theorem}\label{th:2-connected}
If $G$ is 2-connected, $\forall i, p(i)=\pm1$, then the DBCP problem has a solution for $c_p=1$ and $c_s=3$. Moreover, this solution can be found in polynomial time.
\end{theorem}
Similar to Corollary~\ref{cor:3-general}, the approach used in the proof of Theorem~\ref{th:2-connected}, can also be used for the case when the weights are arbitrary (not necessarily $\pm 1$) and $p(V)\neq 0$. It is easy to verify that in this case, if $G$ is 2-connected, the DBCP problem has a
connected partition $(V_1,V_2)$ such that
$|p(V_1)-p(V)/2|, |p(V_2)-p(V)/2| \leq \max_{j\in V} |p(j)|$ and
$|V_1|, |V_2| \geq |V|/4$.
\begin{corollary}\label{cor:2-general}
If $G$ is 2-connected, then the DBCP problem (with general $p$ and not necessarily satisfying $p(V)= 0$) has a solution for $c_p=\max_{j\in V} |p(j)|$ and $c_s=3$. Moreover, this solution can be found in polynomial time.
\end{corollary}
\section{Graphs with Two Types of Nodes}\label{sec:blue_red}
Assume $G$ is a connected graph with nodes colored either red ($R\subseteq V$) or blue ($B\subseteq V$). Let $|V|=n$, $|R|=n_r$, and $|B|=n_b$.
If $G$ is 3-connected, set $p(i)=1$ if $i \in R$ and $p(i) =-1$ if $i \in B$.
Corollary~\ref{cor:3-general} implies then that there is always a connected partition $(V_1,V_2)$ of $V$ that splits both the blue and the red nodes evenly (assuming $n_r$ and $n_b$ are both even), i.e., such that $|V_1|=|V_2|$, $|R\cap V_1|=|R\cap V_2|$, and $|B\cap V_1|=|B\cap V_2|$. (If $n_r$ and/or $n_b$ are not even, then one side will
contain one more red or blue node.)
\begin{corollary}\label{cor:blue_red3}
Given a 3-connected graph $G$ with nodes colored either red ($R\subseteq V$) or blue ($B\subseteq V$). There is always a partition $(V_1,V_2)$ of $V$ such that $G[V_1]$ and $G[V_2]$ are connected, $|V_1|=|V_2|$, $|R\cap V_1|=|R\cap V_2|$, and $|B\cap V_1|=|B\cap V_2|$ (assuming $|R|$ and $|B|$ are both even). Such a partition can be
computed in polynomial time.
\end{corollary}
If $G$ is only 2-connected, we may not always get a perfect partition.
Assume wlog that $n_r \leq n_b$. If for every $v\in R$ and $u\in B$, we set $p(v)=1$ and $p(u)=-n_r/n_b$, Corollary~\ref{cor:2-general} implies that there is always a connected partition $(V_1,V_2)$ of $V$ such that both $|(|R\cap V_1|-n_r/n_b |B\cap V_1|)|\leq 1$ and $|(|R\cap V_2|-n_r/n_b |B\cap V_2|)|\leq 1$, and also $\max\{\frac{|V_1|}{|V_2|},\frac{|V_2|}{|V_1|}\}\leq 3$.
Thus, the ratio of red to blue nodes in each side $V_i$ differs from
the ratio $n_r/n_b$ in the whole graph by $O(1/n)$. Hence if the
numbers of red and blue nodes are $\omega(1)$, then the two types are
presented in both sides of the partition in approximately the same
proportion as in the whole graph.
\begin{corollary}
Given a 2-connected graph $G$ with nodes colored either red ($R\subseteq V$) or blue ($B\subseteq V$), and assume wlog $|R| \leq |B|$. We can always find
in polynomial time a partition $(V_1,V_2)$ of $V$
such that $G[V_1]$ and $G[V_2]$ are connected, $|V_1|,|V_2| \geq |V|/4$,
and the ratio of red to blue nodes in each side $V_i$ differs from the ratio $|R|/|B|$
in the whole graph by $O(1/n)$.
\end{corollary}
\section{Conclusion}\label{sec:conclusion}
In this paper, we introduced and studied the problem of partitioning a graph
into two connected subgraphs that satisfy simultaneouly two objectives:
(1) they balance the supply and demand within each side of the partition
(or more generally, for the case of $p(V) \neq 0$, they split approximately equally
the excess supply/demand between the two sides), and (2) the two sides are large and
have roughly comparable size (they are both $\Omega(|V|)$).
We showed that for 2-connected graphs it is always possible to achieve both objectives at the same time, and for 3-connected graphs there is a partition
that is essentially perfectly balanced in both objectives. Furthermore,
these partitions can be computed in polynomial time.
This is a paradigmatic bi-objective balancing problem.
We observed how it can be easily used to find a connected partition of a graph with two types of nodes that is balanced with respect to the sizes of both types.
Overall, we believe that the novel techniques used in this paper can be applied to partitioning heterogeneous networks in various contexts.
There are several interesting further directions that suggest themselves.
First, extend the theory and algorithms to find doubly balanced connected partitions to more than two parts.
Second, even considering only the supply/demand objective, does the analogue of the results
of Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division} for the connected $k$-way partitioning
of $k$-connected graphs with respect to size (which corresponds to $p(i)=1$) extend to the
supply/demand case ($p(i)=\pm1$) for $k>3$? And is there a polynomial algorithm
that constructs such a partition?
Finally, extend the results of Section \ref{sec:blue_red} to graphs with more than two types of nodes,
that is, can we partition (under suitable conditions) a graph with several types of nodes to two (or more)
large connected subgraphs that preserve approximately the diversity (the proportions of the types) of the whole
population?
\newpage
\section{The BCPI Problem}\label{sec:BCPI}
\section{Balancing the Supply/Demand Only}\label{sec:BCPI}
In this section, we study the single objective problem of
finding a partition of the graph into connected subgraphs that balances (approximately)
the supply and demand in each part of the partition, without any regard
to the sizes of the parts.
We can state the optimization problem as follows, and will refer to it as the Balanced Connected Partition with Integer weights (BCPI) problem.
\begin{definition}
Given a graph $G=(V,E)$ with a weight (supply/demand) function $p~:~V~\rightarrow~\mathbb{Z}$ satisfying $\sum_{j\in V} p(j)=0$. The BCPI problem is the problem of partitioning $V$ into $(V_1,V_2)$ such that
\begin{enumerate}
\item $V_1\cap V_2=\emptyset$ and $V_1\cup V_2=V$,
\item $G[V_1]$ and $G[V_2]$ are connected,
\item $|p(V_1)|+|p(V_2)|$ is minimized, where $p(V_i)=\sum_{j\in V_i} p(j).$
\end{enumerate}
\end{definition}
Clearly, the minimum possible value for $|p(V_1)|+|p(V_2)|$ that we can hope for is 0,
which occurs iff $p(V_1)=p(V_2) =0$.
It is easy to show that the problem of determining whether there exists such a `perfect' partition
(and hence the BCPI problem) is strongly NP-hard.
The proof is very similar to analogous results concerning the partition of a graph
into two connected subgraphs with equal sizes (or weights, when nodes have positive weights)
\cite{chataigner2007approximation,dyer}
\begin{proposition}\label{lem:BCPI_hard}
(1) It is strongly NP-hard to determine whether there is a solution to the BCPI problem with value 0, even when $G$ is 2-connected. \\
(2) If $G$ is not 2-connected, then this problem is NP-hard even when $\forall i, p(i)=\pm1$.
\end{proposition}
Although it is NP-hard to tell whether there is a solution satisfying $p(V_1)=p(V_2) =0$,
even when $\forall i, p(i)=\pm1$, in this case, if the graph $G$ is 2-connected there is always such a solution.
For general weights $p$, there is a solution such that $|p(V_1)|,|p(V_2)|\leq \max_{j\in V}|p(j)|/2$
and it can be found easily in linear time using the $st$-numbering between two nodes (see the Appendix \ref{sec:proof1}).
\begin{proposition}\label{lem:2_poly}
Let $G$ be a 2-connected graph and $u,v$ any two nodes in $V$ such that $p(u) p(v)>0$.\\
(1) There is a solution such that $u\in V_1$, $v\in V_2$, and $|p(V_1)|=|p(V_2)|\leq \max_{j\in V}|p(j)|/2$.\\
(2) If $\forall i, p(i)=\pm1$, we can find a solution such that $u\in V_1$, $v\in V_2$, and $p(V_1)=p(V_2)=0$.\\
In both cases, the solution can be found in $O(|E|)$ time.
\end{proposition}
\begin{remark}
The bound in Proposition~\ref{lem:2_poly} (1) is tight. A simple example is a cycle of length 4 like $(v_1,v_2,v_3,v_4)$ with $p(v_1)=-p$, $p(v_2)=-p/2$, $p(v_3)=p$, and $p(v_4)=p/2$. It is easy to see that in this example $|p(V_1)|+|p(V_2)|= \max_{j\in V}|p(j)|=p$ is the best that one can do.
\end{remark}
\subsection{Connected Partitioning into Many Parts}
The BCPI problem can be extended to partitioning a graph into $k=3$ or more parts.
Let $G=(V,E)$ be a graph with a weight function $p~:~V~\rightarrow~\mathbb{Z}$ satisfying $\sum_{j\in V} p(j)=0$. The $\text{BCPI}_k$ problem is the problem of partitioning $G$ into $(V_1,V_2,\dots,V_k)$ such that for any $1\leq i\leq k$, $G[V_i]$ is connected and $\sum_{i=1}^k |p(V_i)|$ is minimized.
In the following proposition, we show that for $k=3$, if $ p(i)=\pm1, \forall i$, then
there is always a perfect partition (i.e., with $p(V_1)=p(V_2)=p(V_3)=0$) and it can be found efficiently.
For general $p$, we can find a partition such that $|p(V_1)|+|p(V_2)|+|p(V_3)|\leq 2 \max_{j\in V}|p(j)|$.
The proof and the algorithm use a similar approach as the algorithm in~\cite{wada1994efficient}
for partitioning a 3-connected graph to three connected parts with prescribed sizes, using the nonseparating ear decomposition of 3-connected graphs as described in Subsection~\ref{subsec:ear-decomposition}.
The proof is given in the Appendix \ref{sec:proof1}.
\begin{proposition}\label{lem:3-connected-zero-p}
Let $G$ be a 3-connected graph and $u,v,w$ three nodes in $V$ such that $p(u), p(v),p(w)>0$ or $p(u), p(v),p(w)<0$.\\
(1) There is a solution such that $u\in V_1$, $v\in V_2$, $w\in V_3$, and $|p(V_1)|+|p(V_2)|+|p(V_3)|\leq 2 \max_{j\in V}|p(j)|$. \\
(2) If $\forall i, p(i)=\pm1$, then there is a solution such that $u\in V_1$, $v\in V_2$, $w\in V_3$, and
$|p(V_1)|=|p(V_2)|=|p(V_3)|=0$.\\
In both cases, the solution can be found in $O(|E|)$ time.
\end{proposition}
\section{Balancing Both Objectives}\label{sec:BPGI}
In this section, we formally define and study the Doubly Balanced Connected graph Partitioning (DBCP) problem.
\begin{definition}
Given a graph $G=(V,E)$ with a weight (supply/demand) function $p~:~V~\rightarrow~\mathbb{Z}$ satisfying $p(V)=\sum_{j\in V} p(j)=0$ and constants $c_p\geq 0$, $c_s\geq 1$. The \emph{DBCP problem} is the problem of partitioning $V$ into $(V_1,V_2)$ such that
\begin{enumerate}
\item $V_1\cap V_2=\emptyset$ and $V_1\cup V_2=V$,
\item $G[V_1]$ and $G[V_2]$ are connected,
\item $|p(V_1)|,|p(V_2)|\leq c_p$ and $\max\{\frac{|V_1|}{|V_2|},\frac{|V_2|}{|V_1|}\}\leq c_s$, where $p(V_i)=\sum_{j\in V_i} p(j).$
\end{enumerate}
\end{definition}
\begin{remark}
Our techniques apply also to the case that $p(V)\neq0$.
In this case, the requirement 3 on $p(V_1)$ and $p(V_2)$ is
$|p(V_1)-p(V)/2|,|p(V_2)-p(V)/2|\leq c_p$, i.e., the excess supply/demand
is split approximately evenly between the two parts.
\end{remark}
We will concentrate on 2-connected and 3-connected graphs and
show that we can construct efficiently good partitions.
For most of the section we will focus on the case that $p(i)=\pm1, \forall i \in V$.
This case contains all the essential ideas.
All the techniques generalize to the case of arbitrary $p$,
and we will state the corresponding theorems.
We observed in Section 2 that if the graph is 2-connected and $p(i)=\pm1, \forall i \in V$ then there is always a connected partition
that is perfect with respect to the weight objective,
$p(V_1)=p(V_2)=0$, i.e., (3) is satisfied with $c_p=0$.
We know also from \cite{lovasz1977homology,gyori1976division}
that there is always a connected partition
that is perfect with respect to the size objective,
$|V_1|=|V_2|$, i.e., condition 3 is satisfied with $c_s=1$.
The following observations show that combining the two objectives makes
the problem more challenging.
If we insist on $c_p=0$, then $c_s$ cannot be bounded in general,
(it will be $\Omega(|V|)$),
and if we insist on $c_s=1$, then $c_p$ cannot be bounded.
The series-parallel graphs of Figure~\ref{fig:example_bound} provide simple counterexamples.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{Figures/BoundBalance.pdf}
\caption{Series-parallel graphs with $2s+1$ paths of length $4t+2$ used in Observations~\ref{obs:p=0} and \ref{obs:s=1}.}
\label{fig:example_bound}
\end{figure}
\begin{observation}\label{obs:p=0}
If $c_p=0$, then for any $c_s<|V|/2-1$, there exist a 2-connected graph $G$ such that the DBCP problem does not have a solution even when $\forall i, p(i)=\pm1$.
\end{observation}
\begin{proof}
In the graph depicted in Figure~\ref{fig:example_bound}, set $t=0$.
\end{proof}
\begin{observation}\label{obs:s=1}
If $c_s=1$, then for any $c_p<|V|/6$, there exist a 2-connected graph $G$ such that the DBCP problem does not have a solution even when $\forall i, p(i)=\pm1$.
\end{observation}
\begin{proof}
In the graph depicted in Figure~\ref{fig:example_bound}, set $s=1$.
\end{proof}
Thus, $c_p$ has to be at least 1 to have any hope for a bounded $c_s$.
We show in this section that $c_p=1$ suffices for all 2-connected graphs.
We first treat 3-connected graphs.
\subsection{3-Connected Graphs}\label{subsec:3-connected}
Let $G=(V,E)$ be a 3-connected graph.
Assume for the most of this section that $\forall i, p(i)=\pm1$ and $p(V)=0$
(we will state the results for general $p$ at the end).
We show that $G$ has a partition that is essentially perfect with
respect to both objectives, i.e., with $c_p=0$ and $c_s=1$.
We say ``essentially", because $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|$
imply that $|V_1|=|V_2|$ are even, and hence $V$ must be a multiple of 4.
If this is the case, then indeed we can find such a perfect partition.
If $|V|\equiv 2 (\mathrm{mod}~4)$ ($|V|$ has to be even since $p(V)=0$), then
we can find an `almost perfect' partition, one in which
$|p(V_1)|=|p(V_2)|=1$ and $|V_1|=|V_2|$ (or one in which
$p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|+2$).
We first treat the case that $G$ contains a triangle (i.e., cycle of length 3).
In the following Lemma, we use the embedding for $k$-connected graphs introduced in~\cite{linial1988rubber} and as described in Subsection~\ref{subsec:embedding}, to show that if $G$ is 3-connected with a triangle and all weights are $\pm 1$, then the DBCP problem has a perfect solution.
\begin{lemma}\label{lem:3-connected-tri}
If $G$ is 3-connected with a triangle, $\forall i, p(i)=\pm1$, and $|V|\equiv0(\mathrm{mod}~4)$, then there exists a solution to the DBCP problem
with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|$.
If $|V|\equiv2 (\mathrm{mod}~4)$, then there is a solution with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|+2$.
Moreover, this partition can be found in polynomial time.
\end{lemma}
\begin{proof}
Assume that $|V|\equiv0(\mathrm{mod}~4)$; the proof for the case $|V|\equiv2(\mathrm{mod}~4)$ is similar.
In~\cite{linial1988rubber} as described in Subsection~\ref{subsec:embedding}, it is proved that if $G$ is a $k$-connected graph, then for every $X\subset V$ with $|X|=k$, $G$ has a convex $X$-embedding in general position. Moreover, this embedding can be found by solving a set of linear equations of size $|V|$.
Now, assume $v,u,w\in V$ form a triangle in $G$. Set $X=\{v,u,w\}$. Using the theorem, $G$ has a convex $X$-embedding $f:V\rightarrow\mathbb{R}^2$ in general position. Consider a circle $\mathcal{C}$ around the triangle $f(u),f(v),f(w)$ in $\mathbb{R}^2$ as shown in an example in Fig.~\ref{fig:embedding_1}. Also consider a directed line $\mathcal{L}$ tangent to the circle $C$ at point $A$. If we project the nodes of $G$ onto the line $\mathcal{L}$, since the embedding is convex and also $\{u,v\},\{u,w\},\{w,v\}\in E$, the order of the nodes' projection gives an $st$-numbering between the first and the last node (notice that the first and last nodes are always from the set $X$). For instance in Fig.~\ref{fig:embedding_1}, the order of projections give an $st$-numbering between the nodes $u$ and $v$ in $G$. Hence, if we set $V_1$ to be the $|V|/2$ nodes whose projections come first and $V_2$ are the $|V|/2$ nodes whose projections come last, then $G[V_1]$ and $G[V_2]$ are both connected and $|V_1|=|V_2|=|V|/2$. The only thing that may not match is $p(V_1)$ and $p(V_2)$. Notice that for each directed line tangent to the circle $\mathcal{C}$, we can similarly get a partition such that $|V_1|=|V_2|=|V|/2$. So all we need is a point $D$ on the circle $\mathcal{C}$ such that if we partition based on the directed line tangent to $C$ at point $D$, then $p(V_1)=p(V_2)=0$. To find such a point, we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ ($AB$ is a diameter of the circle $\mathcal{C}$) and consider the resulting partition. Notice that if at point $A$, $p(V_1)>0$, then at point $B$ since $V_1$ and $V_2$ completely switch places compared to the partition at point $A$, $p(V_1)<0$. Hence, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ and keep it tangent to the circle, in the resulting partitions, $p(V_1)$ goes from some positive value to a non-positive value. Notice that the partition $(V_1,V_2)$ changes only if $\mathcal{L}$ passes a point $D$ on the circle such that at $D$, $\mathcal{L}$ is perpendicular to a line that connects $f(i)$ to $f(j)$ for some $i,j\in V$. Now, since the embedding is in general position, there are exactly two points on every line that connects two points $f(i)$ and $f(j)$, so $V_1$ changes at most by one node leaving $V_1$ and one node entering $V_1$ at each step as we move $\mathcal{L}$. Hence, $p(V_1)$ changes by either $\pm 2$ or $0$ value at each change. Now, since $|V|\equiv 0 (\mathrm{mod}~ 4)$, $p(V_1)$ has an even value in all the resulting partitions. Therefore, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$, there must be a point $D$ such that in the resulted partition $p(V_1)=p(V_2)=0$.
It is also easy to see that since $V_1$ may change only when a line that passes through 2 nodes of graph $G$ is perpendicular to $\mathcal{L}$, we can find $D$ in at most $O(|V|^2)$ steps.
\end{proof}
\begin{figure}[t]
\centering
\includegraphics[scale=0.75]{Figures/Embedding_1.pdf}
\caption{Proof of Lemma~\ref{lem:3-connected-tri}.}
\label{fig:embedding_1}
\end{figure}
When $G$ is a triangle-free 3-connected graph, however, the proof in Lemma~\ref{lem:3-connected-tri} cannot be directly used anymore. The reason is if for example $\{u,v\}\notin E$ and we project the nodes of $G$ onto the line $\mathcal{L}$, this time the order of the nodes projection does not give an $st$-numbering between the first and the last node if for example $u$ and $w$ are the first and last node, since some of the middle nodes may only be connected to $v$. To prove a similar result for triangle-free 3-connected case, we first provide the following two Lemmas. The main purpose of the following two Lemmas are to compensate for the triangle-freeness of $G$ in the proof of Lemma~\ref{lem:3-connected-tri}. The idea is to show that in every 3-connected graph, there is a triple $\{u,w,v\}\in V$, such that $\{u,w\},\{w,v\}\in E$ and in every partition that we get by the approach used in the proof of Lemma~\ref{lem:3-connected-tri}, if $u$ and $v$ are in $V_i$, so is a path between $u$ and $v$.
\begin{lemma}\label{lem:cut}
If $G$ is 3-connected, then there exists a set $\{u,v,w\}\in V$ and a partition of $V$ into $(V_1',V_2')$ such that:
\begin{enumerate}
\item $V_1'\cap V_2' =\emptyset$ and $V_1'\cup V_2'=V$,
\item $G[V_1']$ and $G[V_2']$ are connected,
\item $\{u,w\},\{v,w\}\in E$,
\item $w\in V_1'$, $u,v\in V_2'$,
\item $|V_2'|\leq |V|/2$.
\end{enumerate}
Moreover, such a partition and $\{u,v,w\}$ can be found in $O(|E|)$ time.
\end{lemma}
\begin{proof}
Using the algorithm presented in~\cite{cheriyan1988finding}, we can find a non-separating cycle $C_0$ in $G$ such that every node in $C_0$ has a neighbor in $G\backslash C_0$. Now, we consider two cases:
\begin{itemize}
\item[(i)] If $|C_0|\leq |V|/2+1$, then select any three consecutive nodes $(u,w,v)$ of $C_0$ and set $V_2'=C_0\backslash \{w\}$ and $V_1'=V\backslash V_2'$.
\item[(ii)] If $|C_0|> |V|/2+1$, since every node in $C_0$ has a neighbor in $G\backslash C_0$, there exists a node $w\in V\backslash C_0$ such that $|N(w)\cap C_0|\geq 2$. Select two nodes $u,v\in N(w)\cap C_0$. There exists a path $P$ in $C_0$ between $u$ and $v$ such that $|P|<|V|/2-1$. Set $V_2'=\{u,v\}\cup P$ and $V_1'=V\backslash V_2'$.\qedhere
\end{itemize}
\end{proof}
\begin{figure}[t]
\centering
\includegraphics[scale=0.9]{Figures/3-connected.pdf}
\caption{Proof of Lemma~\ref{lem:embedding} and Theorem~\ref{th:3-connected}.}
\label{fig:3-connected}
\end{figure}
\begin{lemma}\label{lem:embedding}
Given a partition $(V_1',V_2')$ of a 3-connected graph $G$ with properties described in Lemma~\ref{lem:cut}, $G$ has a convex $X$-embedding in general position with mapping $f:V\rightarrow \mathbb{R}^2$ such that:
\begin{enumerate}
\item $X=\{u,w,v\}$, $f(u)=(0,0)$, $f(v)=(1,0)$, and $f(w)=(0,1)$,
\item Every node $i$ in $V_1'$ is mapped to a point $(f_1(i),f_2(i))$ with $f_2(i)\geq 1/2$,
\item Every node $i$ in $V_2'$ is mapped to a point $(f_1(i),f_2(i))$ with $f_1(i)\geq f_2(i)$ and $f_1(i)+2f_2(i)\leq 1$.
\end{enumerate}
Moreover, such an embedding can be found in in polynomial time.
\end{lemma}
\begin{proof}[Sketch of the proof]
Set $X=\{v,u,w\}$. Using~\cite{linial1988rubber}, $G$ has a convex $X$-embedding in $\mathbb{R}^2$ in general position with mapping $f:V\rightarrow \mathbb{R}^2$ such that $f(u)=(0,0)$, $f(v)=(1,0)$, and $f(w)=(0,1)$. In the $X$-embedding of the nodes, we have a freedom to set the elasticity coefficient vector $\vec{c}$ to anything that we want (except a measure zero set of vectors). So for any edge $\{i,j\}\in G[V_1]\cup G[V_2]$, set $c_{ij}=g$; and for any $\{i,j\} \in E[V_1',V_2']$, set $c_{ij}=1$. Since both $G[V_1']$ and $G[V_2']$ are connected, as we increase $g$, nodes in $V_1'$ get closer to $w$ and nodes in $V_2'$ get closer to the line $uv$ (as $g\to \infty$, nodes in $V_1'$ get in the same position as $w$ and node in $V_2'$ get on the line $uv$). Hence, intuitively there exists a value $g$, for which all the nodes in $V_1$ are above line $\mathcal{L}_1$ and all the nodes in $V_2'$ are below the lines $\mathcal{L}_2$ and $\mathcal{L}_3$ as depicted in Fig.~\ref{fig:3-connected}. In the appendix we give the detailed proof and show that a $g$ with polynomially many bits suffices.
\end{proof}
Using Lemmas~\ref{lem:cut} and \ref{lem:embedding}, we are now able to prove that for any 3-connected graph $G$ such that all the weights are $\pm 1$, the DBCP problem has a solution for $c_p=0$ and $c_s=1$. The idea of the proof is similar to the proof of Lemma~\ref{lem:3-connected-tri}, however, we need to use Lemma~\ref{lem:cut} to find a desirable partition $(V_1',V_2')$ and then use this partition to find an embedding with properties as described in Lemma~\ref{lem:embedding}. By using this embedding, we
can show that in every partition that we obtain by the approach in the proof of Lemma~\ref{lem:3-connected-tri}, if $u$ and $v$ are in $V_i$, so is a path between $u$ and $v$. This implies then that $G[V_1]$ and $G[V_2]$ are connected. So we can use similar arguments as in the proof of Lemma~\ref{lem:3-connected-tri} to prove the following theorem (see the Appendix~\ref{sec:proof2} for the proof details).
\begin{theorem}\label{th:3-connected}
If $G$ is 3-connected, $\forall i, p(i)=\pm1$, and $|V|\equiv0(\mathrm{mod}~4)$,
then there exists a solution to the DBCP problem
with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|$.
If $|V|\equiv2(\mathrm{mod}~4)$, then there is a solution with $p(V_1)=p(V_2)=0$ and $|V_1|=|V_2|+2$.
Moreover, this partition can be found in polynomial time.
\end{theorem}
It is easy to check for a 3-connected graph $G$, by using the same approach as in the proof of Lemma~\ref{lem:3-connected-tri} and Theorem~\ref{th:3-connected}, that even when the weights are arbitrary (not necessarily $\pm 1$) and also $p(V)\neq 0$, we can still find a connected partition $(V_1,V_2)$ for $G$ such that
$|p(V_1)-p(V)/2|,|p(V_1)-p(V)/2| \leq \max_{i\in V}|p(i)|$ and $|V_1|=|V_2|$.
\begin{corollary}\label{cor:3-general}
If $G$ is 3-connected, then the DBCP problem (with arbitrary $p$, and not necessarily satisfying $p(V)= 0$) has a
solution for $c_p=\max_{i\in V}|p(i)|$ and $c_s=1$. Moreover, this solution can be found in polynomial time.
\end{corollary}
\subsection{2-Connected Graphs}\label{subsec:2-connected}
We first define a \emph{pseudo-path} between two nodes in a graph as below. The definition is inspired by the definition of the $st$-numbering.
\begin{definition}
A \emph{pseudo-path} between nodes $u$ and $v$ in $G=(V,E)$, is a sequence of nodes $v_1,\dots,v_t$ such that if $v_0= u$ and $v_{t+1}=v$, then
for any $1\leq i\leq t$, $v_i$ has neighbors $v_j$ and $v_k$ such that $j<i<k$.
\end{definition}
Using the pseudo-path notion, in the following lemma we show that if $G$ is 2-connected and has a separation pair such that none of the resulting components are too large, then the DBCP problem always has a solution for some $c_p=c_s=O(1)$. The idea used in the proof of this lemma is one of the building blocks of the proof for the general 2-connected graph case.
\begin{lemma}\label{lem:parallel}
Given a 2-connected graph $G$ and an integer $q>1$, if $\forall i: p(i)=\pm1$ and $G$ has a separation pair $\{u,v\}\subset V$ such that for every connected component $G_i=(V_i,E_i)$ of $G[V\backslash \{u,v\}]$, $|V_i|< (q-1)|V|/q$, then the DBCP problem has a solution for $c_p=1$, $c_s=q-1$, and it can be found in $O(|E|)$ time.
\end{lemma}
\begin{proof}
Assume for simplicity that $|V|$ is divisible by $q$. There is a separation pair $\{u,v\}\in V$ such that if $G_1,\dots,G_k$ are the connected components of $G\backslash \{u,v\}$, for any $i$, $|V_i|< (q-1)|V|/q$. Since $G$ is 2-connected, $G_1,\dots,G_k$ can be presented by pseudo-paths $P_1,\dots,P_k$ between $u$ and $v$. Assume $P_1,\dots,P_k$ are in increasing order based on their lengths. There exists two subsets of the pseudo-paths $S_1$ and $S_2$ such that $S_1\cap S_2=\emptyset$, $S_1\cup S_2=\{P_1,\dots,P_{k}\}$ and $\sum_{P_j\in S_i} |P_j| \geq |V|/q-1$ for $i=1,2$. The proof is very simple. If $|P_k|< |V|/q$, the greedy algorithm for the partition problem gives the desired partition of $\{P_1,\dots,P_{k}\}$. If $|P_k|\geq |V|/q$, since $|P_k|< (q-1)|V|/q$, $S_1=\{P_1,\dots,P_{k-1}\}$ and $S_2=\{P_k\}$ is the desired partition.
Now, if we put all the pseudo-paths in $S_1$ back to back, they will form a longer pseudo-path $Q_1$ between $u$ and $v$. Similarly, we can form another pseudo-path $Q_2$ from the paths in $S_2$. Without loss of generality we can assume $|Q_1|\geq |Q_2|$. From $u$, including $u$ itself, we count $|V|/q$ of the nodes in $Q_1$ towards $v$ and put them in a set $V'$. Without loss of generality, we can assume $p(V')\geq 0$. If $p(V')=0$, then $(V',V\backslash V')$ is a good partition and we are done. Hence, assume $p(V')>0$. We keep $V'$ fixed and make a new set $V''$ by continuing to add nodes from $Q_1$ to $V'$ one by one until we get to $v$.
If $p(V'')$ hits 0 as we add nodes one by one, we stop and $(V'',V\backslash V'')$ is a good partition and we are done. So, assume $V''=Q_1\cup \{u\}$ and $p(V'')>0$. Since $|Q_2\cup\{v\}|\geq |V|/q$, $|V''|\leq (q-1)|V|/q$. If $|V''| < (q-1)|V|/q$, we add nodes from $Q_2$ one by one toward $u$ until either $|V''|=0$ or $|V''| = (q-1)|V|/q$. If we hit 0 first (i.e., $p(V'')=0$) and $|V''|\leq (q-1)|V|/q$, define $V_1=V''\backslash\{u\}$, then $(V_1,V\backslash V_1)$ is a good partition. So assume $|V''|=(q-1)|V|/q$ and $p(V'')>0$. Define $V'''=V\backslash V''$. Since $p(V'')>0$ and $|V''|=(q-1)|V|/q$, then $p(V''')<0$ and $|V'''|=|V|/q$. Also notice that $V'''\subseteq Q_2$. We consider two cases. Either $|p(V')|\geq|p(V''')|$ or $|p(V')|<|p(V''')|$.
If $|p(V')|\geq|p(V''')|$, then if we start from $u$ and pick nodes one by one from $Q_1$ in order, we can get a subset $V'_1$ of $V'$ such that $|p(V'_1)|=|p(V''')|$. Hence, if we define $V_1=V_1'\cup V'''$, then $(V_1,V\backslash V_1)$ is a good partition.
If $|p(V')|<|p(V''')|$, then we can build a new set $V_1$ by adding nodes one by one from $Q_2$ to $V'$ until $P(V_1)=0$. It is easy to see that since $|p(V')|<|p(V''')|$, then $V_1\backslash V'\subseteq V'''$. Hence, $(V_1, V\backslash V_1)$ is a good partition.
\end{proof}
\begin{corollary}\label{cor:serpar}
If $G$ is a 2-connected series-parallel graph and $\forall i: p(i)=\pm1$, then the DBCP problem has a solution for $c_p=1,~c_s=2$, and the solution can be found in $O(|E|)$ time.
\end{corollary}
The graph in Figure~\ref{fig:example_bound} with $s=1$ shows that these parameters are
the best possible for series parallel graphs: if $c_p=O(1)$ then $c_s$ must be at least 2.
To generalize Lemma~\ref{lem:parallel} to all 2-connected graphs, we need to define the \emph{contractible} subgraph and the \emph{contraction} of a given graph as below.
\begin{definition}\label{def:contraction}
We say an induced subgraph $G_1$ of a 2-connected graph $G$ is \emph{contractible}, if there is a separating pair $\{u,v\}\subset V$ such that $G_1=(V_1,E_1)$ is a connected component of $G[V\backslash\{u,v\}]$. Moreover, if we replace $G_1$ by a weighted edge $e'$ with weight $w(e')=|V_1|$ between the nodes $u$ and $v$ in $G$ to obtain a smaller graph $G'$, we say $G$ is \emph{contracted} to $G'$.
\end{definition}
\begin{remark}
Notice that every contractible subgraph of a 2-connected graph $G$ can also be represented by a pseudo-path between its associated separating pair. We use this property in the proof of Theorem~\ref{th:2-connected}.
\end{remark}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.6\textwidth}
\includegraphics[scale=0.5]{Figures/Transformation_1.pdf}
\caption{}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\includegraphics[scale=0.5]{Figures/Transformation_2.pdf}
\caption{}
\end{subfigure}
\caption{Lemma~\ref{lem:2-connected-pair}.}
\label{fig:lem_2-connected-pair}
\end{figure}
Using the notion of the graph contraction, in the following lemma, we show that to partition a 2-connected graph, we can reduce it to one of two cases: either $G$ can be considered as a graph with a set of short pseudo-paths between two nodes, or it can be contracted into a 3-connected graph as illustrated in Fig.~\ref{fig:lem_2-connected-pair}.
\begin{lemma}\label{lem:2-connected-pair}
In every 2-connected graph $G=(V,E)$, given an integer $q\geq 3$, one of the following cases holds, and we can determine which in $O(|E|)$ time:
\begin{enumerate}
\item There is a separation pair $\{u,v\}\subset V$ such that if $G_1,\dots,G_k$ are the connected components of $G[V\backslash \{u,v\}]$, for all $i$, $|V_i|< (q-1)|V|/q$.
\item After a set of contractions, $G$ can be transformed into a 3-connected graph $G^*=(V^*,E^*)$ with weighted edges representing contracted subgraphs such that for every $e^*\in E^*$, $w(e^*)<|V|/q$.
\end{enumerate}
\end{lemma}
\begin{proof}
If there is no separation pairs in $G$, then $G$ is 3-connected and there is nothing left to prove. So assume $\{u,v\}\subset V$ is a separation pair and $G_1,\dots,G_k$ are the connected components of $G[V\backslash \{u,v\}]$. If $\forall i, |V_i|< (q-1)|V|/q$, we are done. So let's assume there is a connected component $G_j$ such that $|V_j|\geq (q-1)|V|/q$. Then for every $i\neq j$, $G_i$ can be contracted and represented by an edge of weight less than $|V|/q$ between $u$ and $v$. Now, we repeat the process by considering the weight of the edges in the size of each connected component. An example for each case is shown in Fig.~\ref{fig:lem_2-connected-pair} for $q=3$. We can find either a suitable separation pair as in case 1 or a
suitable contracted graph $G^*$ as in case 2 in linear time using the Hopcroft-Tarjan
algorithm for finding the triconnected components \cite{hopcroft1973dividing}.
\end{proof}
Using Lemma~\ref{lem:2-connected-pair} for $q=4$, then Lemma~\ref{lem:parallel}, and the idea of the proof for Theorem~\ref{th:3-connected}, we can prove that when $G$ is 2-connected and all $p(i)=\pm 1$, the DBCP problem has a solution for $c_p=1$ and $c_s=3$. There are some subtleties in adapting Lemma~\ref{lem:cut} for this case to account for the fact that the edges of $G^*$ are now weighted, and the partition $(V_1', V_2')$ has to take into account the edge weights.
We find a suitable convex embedding of the 3-connected graph $G^*$
and then embed the nodes of the contracted pseudopaths appropriately
along the segments corresponding to the weighted edges. Some care is needed
to carry out the argument of the 3-connected case, since as the line
tangent to the circle rotates, the order of the projections of many nodes may change at once, namely the nodes on an edge perpendicular to the rotating line.
The details of the proof are given in Appendix~\ref{sec:proof2}.
We have:
\begin{theorem}\label{th:2-connected}
If $G$ is 2-connected, $\forall i, p(i)=\pm1$, then the DBCP problem has a solution for $c_p=1$ and $c_s=3$. Moreover, this solution can be found in polynomial time.
\end{theorem}
Similar to Corollary~\ref{cor:3-general}, the approach used in the proof of Theorem~\ref{th:2-connected}, can also be used for the case when the weights are arbitrary (not necessarily $\pm 1$) and $p(V)\neq 0$. It is easy to verify that in this case, if $G$ is 2-connected, the DBCP problem has a
connected partition $(V_1,V_2)$ such that
$|p(V_1)-p(V)/2|, |p(V_2)-p(V)/2| \leq \max_{j\in V} |p(j)|$ and
$|V_1|, |V_2| \geq |V|/4$.
\begin{corollary}\label{cor:2-general}
If $G$ is 2-connected, then the DBCP problem (with general $p$ and not necessarily satisfying $p(V)= 0$) has a solution for $c_p=\max_{j\in V} |p(j)|$ and $c_s=3$. Moreover, this solution can be found in polynomial time.
\end{corollary}
\section{Graphs with Two Types of Nodes}\label{sec:blue_red}
Assume $G$ is a connected graph with nodes colored either red ($R\subseteq V$) or blue ($B\subseteq V$). Let $|V|=n$, $|R|=n_r$, and $|B|=n_b$.
If $G$ is 3-connected, set $p(i)=1$ if $i \in R$ and $p(i) =-1$ if $i \in B$.
Corollary~\ref{cor:3-general} implies then that there is always a connected partition $(V_1,V_2)$ of $V$ that splits both the blue and the red nodes evenly (assuming $n_r$ and $n_b$ are both even), i.e., such that $|V_1|=|V_2|$, $|R\cap V_1|=|R\cap V_2|$, and $|B\cap V_1|=|B\cap V_2|$. (If $n_r$ and/or $n_b$ are not even, then one side will
contain one more red or blue node.)
\begin{corollary}\label{cor:blue_red3}
Given a 3-connected graph $G$ with nodes colored either red ($R\subseteq V$) or blue ($B\subseteq V$). There is always a partition $(V_1,V_2)$ of $V$ such that $G[V_1]$ and $G[V_2]$ are connected, $|V_1|=|V_2|$, $|R\cap V_1|=|R\cap V_2|$, and $|B\cap V_1|=|B\cap V_2|$ (assuming $|R|$ and $|B|$ are both even). Such a partition can be
computed in polynomial time.
\end{corollary}
If $G$ is only 2-connected, we may not always get a perfect partition.
Assume wlog that $n_r \leq n_b$. If for every $v\in R$ and $u\in B$, we set $p(v)=1$ and $p(u)=-n_r/n_b$, Corollary~\ref{cor:2-general} implies that there is always a connected partition $(V_1,V_2)$ of $V$ such that both $|(|R\cap V_1|-n_r/n_b |B\cap V_1|)|\leq 1$ and $|(|R\cap V_2|-n_r/n_b |B\cap V_2|)|\leq 1$, and also $\max\{\frac{|V_1|}{|V_2|},\frac{|V_2|}{|V_1|}\}\leq 3$.
Thus, the ratio of red to blue nodes in each side $V_i$ differs from
the ratio $n_r/n_b$ in the whole graph by $O(1/n)$. Hence if the
numbers of red and blue nodes are $\omega(1)$, then the two types are
presented in both sides of the partition in approximately the same
proportion as in the whole graph.
\begin{corollary}
Given a 2-connected graph $G$ with nodes colored either red ($R\subseteq V$) or blue ($B\subseteq V$), and assume wlog $|R| \leq |B|$. We can always find
in polynomial time a partition $(V_1,V_2)$ of $V$
such that $G[V_1]$ and $G[V_2]$ are connected, $|V_1|,|V_2| \geq |V|/4$,
and the ratio of red to blue nodes in each side $V_i$ differs from the ratio $|R|/|B|$
in the whole graph by $O(1/n)$.
\end{corollary}
\section{Conclusion}\label{sec:conclusion}
In this paper, we introduced and studied the problem of partitioning a graph
into two connected subgraphs that satisfy simultaneouly two objectives:
(1) they balance the supply and demand within each side of the partition
(or more generally, for the case of $p(V) \neq 0$, they split approximately equally
the excess supply/demand between the two sides), and (2) the two sides are large and
have roughly comparable size (they are both $\Omega(|V|)$).
We showed that for 2-connected graphs it is always possible to achieve both objectives at the same time, and for 3-connected graphs there is a partition
that is essentially perfectly balanced in both objectives. Furthermore,
these partitions can be computed in polynomial time.
This is a paradigmatic bi-objective balancing problem.
We observed how it can be easily used to find a connected partition of a graph with two types of nodes that is balanced with respect to the sizes of both types.
Overall, we believe that the novel techniques used in this paper can be applied to partitioning heterogeneous networks in various contexts.
There are several interesting further directions that suggest themselves.
First, extend the theory and algorithms to find doubly balanced connected partitions to more than two parts.
Second, even considering only the supply/demand objective, does the analogue of the results
of Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division} for the connected $k$-way partitioning
of $k$-connected graphs with respect to size (which corresponds to $p(i)=1$) extend to the
supply/demand case ($p(i)=\pm1$) for $k>3$? And is there a polynomial algorithm
that constructs such a partition?
Finally, extend the results of Section \ref{sec:blue_red} to graphs with more than two types of nodes,
that is, can we partition (under suitable conditions) a graph with several types of nodes to two (or more)
large connected subgraphs that preserve approximately the diversity (the proportions of the types) of the whole
population?
\newpage
\section{Introduction}
Power Grid Islanding is an effective method to mitigate cascading failures in power grids~\cite{sun2002two}. The challenge is to partition the network into smaller connected components, called \emph{islands}, such that each island can operate independently for a while. In order for an island to operate, it is necessary that the power supply and demand at that island are almost equal.\footnote{If the supply and demand are not exactly equal but still relatively close, load shedding/generation curtailing can be used in order for the island to operate.} Equality of supply and demand in an island, however, may not be sufficient for its independent operation. It is also important that the infrastructure in that island has the physical capacity to safely transfer the power from the supply nodes to the demand nodes. When the island is large enough compared to the initial network, it is more likely that it has enough capacity. This problem has been studied in the power systems community but almost all the algorithms provided in the literature are heuristic methods that have been shown to be effective only by simulations~\cite{sun2002two,sanchez2014hierarchical,pahwa2013optimal,fan2012mixed}.
Motivated by this application, we formally introduce and study the Doubly Balanced Connected graph Partitioning (DBCP) problem:
Let $G=(V,E)$ be a connected graph with a weight (supply/demand) function $p:V\rightarrow \mathbb{Z}$ satisfying $p(V)=\sum_{j\in V} p(j)=0$. The objective is to partition $V$ into $(V_1,V_2)$ such that $G[V_1]$ and $G[V_2]$ are connected, $|p(V_1)|,|p(V_2)|\leq c_p$, and $\max\{\frac{|V_1|}{|V_2|},\frac{|V_2|}{|V_1|}\}\leq c_s$, for some constants $c_p$ and $c_s$. We also consider the case that $p(V)\neq 0$, in which the
excess supply/demand should be split roughly evenly.
The problem calls for a partition into two connected subgraphs that simultaneously balances two objectives, (1) the supply/demand within each part, and (2) the sizes of the parts.
The connected partitioning problem with only the size objective has been studied previously.
In the most well-known result, Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division} independently proved, using different methods, that every $k$-connected graph can be partitioned into $k$ arbitrarily sized connected subgraphs. However, neither of the proofs is constructive, and there are no known
polynomial-time algorithms to find such a partition for $k>3$.
For $k=2$, a linear time algorithm is provided in~\cite{suzuki1990linear} and for $k=3$
an $O(|V|^2)$ algorithm is provided in~\cite{wada1994efficient}.\footnote{For $k=2$, a much simpler approach than the one in~\cite{suzuki1990linear} is to use the $st$-numbering~\cite{lempel1967algorithm} for 2-connected graphs.}
The complexity of the problem with the size objective and related optimization problems
have been studied in~\cite{dyer,chlebikova1996approximating,chataigner2007approximation}
and there are various NP-hardness and inapproximability results.
Note that the size of the cut is not of any relevance here (so the
extensive literature on finding balanced partitions, not necessarily connected,
that minimize the cut is not relevant.)
The objective of balancing the supply/demand alone, when all $p(i)$ are $\pm1$, can also be seen as an extension for the objective of balancing the size (which corresponds to $p(i)=1$).
Our bi-objective problem of balancing both supply/demand and size, can be seen also
as an extension of the problem of finding a partition that balances the size for two types of nodes
simultaneously: Suppose the nodes of a graph are partitioned into red and blue nodes.
Find a partition of the graph into two large connected subgraphs that splits approximately evenly both
the red and the blue nodes.
We now summarize our results and techniques. Since the power grids are designed to withstand a single failure (``$N-1$" standard)~\cite{bienstock2016electrical}, and therefore 2-connected, our focus is mainly on the graphs that are at least 2-connected.
We first, in Section \ref{sec:BPGI}, study the connected partitioning problem with
only the supply/demand balancing objective, and show results that
parallel the results for balancing size alone, using similar techniques:
The problem is NP-hard in general.
For 2-connected graphs and weights $p(i)=\pm1$, there is always a perfectly balanced partition
and we can find it easily using an $st$-numbering. For 3-connected graphs and weights $p(i)=\pm1$,
there is a perfectly balanced partition into three connected graphs, and we can find it
using a nonseparating ear decomposition of 3-connected graphs~\cite{cheriyan1988finding} and similar ideas as in~\cite{wada1994efficient}.
The problem is more challenging when we deal with both balancing objectives, supply/demand and size.
This is the main focus and occupies the bulk of this paper.
Our main results are existence results and algorithms for 2- and 3-connected graphs.
It is easy to observe that we cannot achieve perfection in one objective ($c_p=0$ or $c_s =1$)
without sacrificing completely the other objective.
We show that allowing the supply/demand of the parts to be off balance by at most the weight of one node
suffices to get a partition that is roughly balanced also with respect to size.
First, in Section~\ref{subsec:3-connected}, we study the case of 3-connected graphs
since we use this later as the basis of handling 2-connected graphs.
We show that if $\forall i,~p(i)=\pm1$, there is a partition that is perfectly balanced
with respect to both objectives, if $|V|\equiv0 (\mathrm{mod}~4)$ (otherwise the
sizes are slightly off for parity reasons); for general $p$, the partition is perfect
in both objectives up to the weight of a single node. Furthermore, the partition can be constructed in
polynomial time. Our approach uses the convex embedding characterization of $k$-connectivity
studied by Linial, Lov\'{a}z, and Wigderson~\cite{linial1988rubber}. We need to adapt it for
our purposes so that the convex embedding also has certain desired geometric properties,
and for this purpose we use the nonseparating ear decomposition of 3-connected graphs
of~\cite{cheriyan1988finding} to obtain a suitable embedding.
Then, in Section~\ref{subsec:2-connected}, we analyze the case of 2-connected graphs.
We reduce it to two subcases: either (1) there is a separation pair that splits the graph
into components that are not very large, or (2) we can perform a series of
contractions to achieve a 3-connected graph whose edges represent
contracted subgraphs that are not too large. We provide a good partitioning algorithm
for case (1), and for case (2) we extend the algorithms for 3-connected graphs
to handle also the complications arising from edges representing contracted subgraphs.
Finally, in Section~\ref{sec:blue_red}, we briefly discuss the problem
of finding a connected partitioning of a graph with two types of nodes
that splits roughly evenly both types.
The body of the paper contains proofs and sketches for some of the results,
and Appendices~\ref{sec:proof1}, \ref{sec:proof2}, and \ref{sec:proof3} contain the remaining proofs.
Graph-theoretic background and definitions (e.g., the notions of $st$-numbering,
nonseparating ear decomposition, convex embedding of $k$-connected graphs)
are given in Appendix~\ref{sec:pre}.
\section{Definitions and Background}\label{sec:pre}
In this section, we provide a short overview of the definitions and tools used in our work. Most of the graph theoretical terms used in this paper are relatively standard and borrowed from~\cite{bondy2008graph} and \cite{west2001introduction}. All the graphs in this paper are loopless.
\subsection{Terms from Graph Theory}
\noindent\textbf{Cutpoints and Subgraphs:} A \emph{cutpoint} of a connected graph $G$ is a node whose deletion results in a disconnected graph.
Let $X$ and $Y$ be subsets of the nodes of a graph $G$. $G[X]$ denotes the subgraph of $G$ \emph{induced} by $X$. We denote by $E[X, Y]$ the set of edges of $G$ with one end in $X$ and the other end in $Y$. The neighborhood of a node $v$ is denoted $N(v)$.
\noindent\textbf{Connectivity:} The connectivity of a graph $G=(V,E)$ is the minimum size of a set $S\subset V$ such that $G\backslash S$ is not connected. A graph is $k$-connected if its connectivity is at least $k$.
\subsection{$st$-numbering of a Graph}\label{subsec:st-numbering}
Given any edge $\{s,t\}$ in a 2-connected graph $G$, an \emph{$st$-numbering} for $G$ is a numbering for the nodes in $G$ defined as follows~\cite{lempel1967algorithm}: the nodes of $G$ are numbered from 1 to $n$ so that $s$ receives number 1, node $t$ receives number $n$, and every node except $s$ and $t$ is adjacent both to a lower-numbered and to a higher-numbered node. It is shown in~\cite{even1976computing} that such a numbering can be found in $O(|V|+|E|)$.
\subsection{Series-Parallel Graphs
A Graph $G$ is \emph{series-parallel}, with terminals $s$ and $t$, if it can be produced by a sequence of the following operations:
\begin{enumerate}
\item Create a new graph, consisting of a single edge between $s$ and $t$.
\item Given two series parallel graphs, $X$ and $Y$ with terminals $s_X,t_X$ and $s_Y,t_Y$ respectively, form a new graph $G=P(X,Y)$ by identifying $s=s_X=s_Y$ and $t=t_X=t_Y$. This is known as the \emph{parallel composition} of $X$ and $Y$.
\item Given two series parallel graphs $X$ and $Y$, with terminals $s_X,t_X$ and $s_Y,t_Y$ respectively, form a new graph $G=S(X,Y)$ by identifying $s=s_X, t_X=s_Y$ and $t=t_Y$. This is known as the \emph{series composition} of $X$ and $Y$.
\end{enumerate}
It is easy to see that a series-parallel graph is 2-connected if, and only if, the last operation is a parallel composition.
\subsection{Nonseparating Induced Cycles and Ear Decomposition}\label{subsec:ear-decomposition}
Let $H$ be a subgraph of a graph $G$. An \emph{ear} of $H$ in $G$ is a nontrivial path in $G$ whose ends lie in $H$ but whose internal vertices do not. An ear decomposition of $G$ is a decomposition $G=P_0\cup\dots\cup P_k$ of the edges of $G$ such that $P_0$ is a cycle and $P_i$ for $i\geq 1$ is an ear of $P_0\cup P_1\cup\dots\cup P_{i-1}$ in $G$. It is known that every 2-connected graph has an ear decomposition (and vice-versa), and such a decomposition can be found in linear time.
A cycle $C$ is a \emph{nonseparating induced cycle} of $G$ if $G\backslash C$ is connected and $C$ has no chords. We say a cycle $C$ avoids a node $u$, if $u\notin C$.
\begin{theorem}[Tutte~\cite{tutte1963draw}] \label{th:tutte-3-connected}
Given a 3-connected graph $G(V,E)$ let $\{t,r\}$ be any edge of $G$ and let $u$ be any node of $G$, $r\neq u\neq t$. Then there is a nonseparating induced cycle of $G$ through $\{t,r\}$ and avoiding $u$.
\end{theorem}
Notice that since $G$ is 3-connected in the previous theorem, every node in $C$ has a neighbor in $G\backslash C$. Cheriyan and Maheshwari showed that the cycle in Theorem~\ref{th:tutte-3-connected} can be found in $O(E)$~\cite{cheriyan1988finding}. Moreover, they showed that any 3-connected graph $G$ has a nonseparating ear decomposition $G=P_0\cup\dots\cup P_k$ defined as follows: Let $V_i=V(P_0)\cup V(P_1)\dots \cup V(P_i)$, let $G_i=G[V_i]$ and $\overline{G}_i=G[V\backslash V_i]$. We say that $G=P_0\cup P_1\cup\dots\cup P_k$ is an ear decomposition through edge $\{t,r\}$ and avoiding vertex $u$ if the cycle $P_0$ contains edge $\{t,r\}$, and the last ear of length greater than one, say $P_m$, has $u$ as its only internal vertex. An ear decomposition $P_0\cup P_1\dots\cup P_k$ of graph $G$ through edge $\{t,r\}$ and avoiding vertex $u$ is a \emph{nonseparating ear decomposition} if for all $i$, $0\leq i<m$, graph $\overline{G}_i$ is connected and each internal vertex of ear $P_i$ has a neighbor in $\overline{G}_i$.
\begin{theorem}[Cheriyan and Maheshwari~\cite{cheriyan1988finding}]
Given an edge $\{t,r\}$ and a vertex $u$ of a 3-connected graph $G$, a nonseparating induced cycle of $G$ through $\{t,r\}$ and avoiding $u$, and
a nonseparating ear decomposition can be found in time $O(|V|+|E|)$.
\end{theorem}
\subsection{Partitioning of Graphs to Connected Subgraphs}
The following theorem is the main existing result in partitioning of graphs into connected subgraphs and is proved independently by Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division} by different methods.
\begin{theorem}[Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division}]
Let $G=(V,E)$ be a $k$-connected graph. Let $n=|V|,~ v_1,v_2,\dots,v_k\in V$ and let $n_1,n_2,\dots,n_k$ be positive integers satisfying $n_1+n_2+\dots+n_k=n$. Then, there exists a partition of $V$ into $(V_1,V_2\dots,V_k)$ satisfying $v_i\in V_i, |V_i|=n_i$, and $G[V_i]$ is connected for $i=1,2,\dots,k$.
\end{theorem}
Although the existence of such a partition has long been proved, there is no polynomial-time algorithm to find such a partition for $k>3$. For $k=2$, it is easy to find such partition using $st$-numbering. For $k=3$, Wada and Kawaguchi~\cite{wada1994efficient} provided an $O(n^2)$ algorithm using the nonseparating ear decomposition of $3$-connected graph.
\subsection{Convex Embedding of Graphs}\label{subsec:embedding}
In this subsection, we provide a short overview of the beautiful work by Linial, Lov\'{a}z, and Wigderson~\cite{linial1988rubber} on convex embedding of the $k$-connected graphs.
Let $Q=\{q_1,q_2,\dots,q_m\}$ be a finite set of points in $\mathbb{R}^d$. The convex hull $\text{conv}(Q)$ of $Q$ is the set of all points $\sum_{i=1}^m \lambda_i q_i$ with $\sum_{i=1} \lambda_i=1$. The rank of $Q$ is defined by $\text{rank}(Q)=1+\text{dim}(\text{conv}(Q))$. $Q$ is in general position if $\text{rank}(S)=d+1$ for every $(d+1)$-subset $S\subseteq Q$.
Let $G$ be a graph and $X\subset V$. A convex $X$-embedding of $G$ is any mapping $f:V\rightarrow \mathbb{R}^{|X|-1}$ such that for each $v\in V\backslash X$, $f(v)\in \mathrm{conv}(f(N(v)))$. We say that the convex embedding is in general position if the set $f(V)$ of the points is in general position.
\begin{theorem}[Linial, Lov\'{a}z, and Wigderson~\cite{linial1988rubber}]
Let $G$ be a graph on $n$ vertices and $1<k<n$. Then the following two conditions are equivalent:
\begin{enumerate}
\item $G$ is $k$-connected
\item For every $X\subset V$ with $|X|=k$, $G$ has a convex $X$-embedding in general position.
\end{enumerate}
\end{theorem}
Notice that the special case of the Theorem for $k=2$ asserts the existence of an $st$-numbering of a 2-connected graph. The proof of this theorem is inspired by physics. The embedding is found by letting the edges of the graph behave like ideal springs and letting its vertices settle. A formal summary of the proof ($1\rightarrow 2$) is as follows (for more details see~\cite{linial1988rubber}). For each $v_i\in X$, define $f(v_i)$ arbitrary in $\mathbb{R}^{k-1}$ such that $f(X)$ is in general position. Assign to every edge $(u,v)\in E$ a positive elasticity coefficient $c_{uv}$ and let $c\in \mathbb{R}^{|E|}$ be the vector of coefficients. It is proved in~\cite{linial1988rubber} that for almost any coefficient vector $c$, an embedding $f$ that minimizes the potential function $P=\sum_{\{u,v\}\in E} c_{uv}\|f(u)-f(v)\|^2$ provides a convex $X$-embedding in general position ($\|.\|$ is the Euclidean norm). Moreover, the embedding that minimizes $P$ can be computed as follows,
\begin{equation*}
f(v)=\frac{1}{c_v}\sum_{u\in N(v)} c_{uv}f(u)~\text{for all}~v\in V\backslash X,
\end{equation*}
in which $c_v=\sum_{u\in N(v)} c_{uv}$. Hence, the embedding can be found by solving a set of linear equations in at most $O(|V|^3)$ time (or matrix multiplication time).
\section{Missing Proofs from Section~\ref{sec:BCPI}}\label{sec:proof1}
\begin{proof}[Proof of Proposition~\ref{lem:BCPI_hard}]
We use the proof of \cite[Theorem 2]{chataigner2007approximation} with a modest change. The reduction is from the X3C problem~\cite{papadimitriou1982complexity}, which is a variant of the \emph{Exact Cover by 3-sets} and defined as follows: Given a set $X$ with $|X|=3q$ and a family $C$ of 3-element subsets of $X$ such that $|C|=3q$ and each element of $X$ appears in exactly 3 sets of $C$, decide whether $C$ contains an exact cover for $X$.
Given an instance $(X,C)$ of $X3C$, let $G=(V,E)$ be the graph with the vertex set $V=X\cup C\cup\{a,b\}$ and edge set $E=\bigcup_{j=1}^{3q}[\{C_jx_i|x_i\in C_j\}\cup\{C_ja\}\cup\{C_jb\}]$. Set $p(a)=2q$, $p(b)=9q^2+q$, $p(C_j)=-1$, and $p(x_i)=-3q$. It is easy to verify that $C$ contains an exact cover for $X$ if and only if the BCPI problem has a solution such that $p(V_1)=p(V_2)=0$. This shows the first claim.
For the second claim, attach to nodes $a$, $b$, and the $x_i$s, paths of length $2q$, $9q^2+q$, and $3q$, respectively, and set the supply/demand values of $a$, $b$, the $x_i$'s and the new nodes equal to $+1$ (for the paths for $a$ and $b$) or $-1$ (for the $x_i$'s).
\end{proof}
\medskip
\begin{proof}[Proof of Proposition~\ref{lem:2_poly}]
Clearly, part (2) follows immediately from part (1) because in this case, $p(V_1), p(V_2)$
are integer and $\max_{j\in V}|p(j)|/2 =1/2$.
To show part (1), pick two arbitrary nodes $u,v\in V$ with $p(u) p(v)>0$. Since we want to separate $u$ and $v$, we can assume wlog that initially $\{u,v\}\in G$. Since $G$ is 2-connected, an $st$-numbering between nodes $u$ and $v$ as $u=v_1,v_2,\dots,v_n=v$ can be found in $O(|V|+|E|)$~\cite{even1976computing}. Define $V_1^{(i)}:=\{v_1,v_2,\dots,v_i\}$. It is easy to see that $p(V_1^{(1)})=p(u)>0$ and $p(V_1^{(n-1)})=-p(v)<0$. Hence, there must exist an index $1\leq i^*<n$ such that $|p(V_1^{(i^*)})|>0$ and $|p(V_1^{(i^*+1)})|\leq 0$. Since $|p(V_1^{(i)})-p(V_1^{(i+1)})|= |p(i^*+1)|$, either $|p(V_1^{(i^*)})|\leq |p(i^*+1)|/2$ or $|p(V_1^{(i^*+1)})|\leq |p(i^*+1)|/2$; Accordingly set $V_1=V_1^{(i^*)}$ or $V_1=V_1^{(i^*+1)}$. Let $V_2=V\backslash V_1$. Hence, $(V_1,V_2)$ is a solution with $|p(V_1)|=|p(V_2)|\leq |p(i^*+1)/2|\leq \max_{j\in V}|p(j)|/2$. It is easy to see that $i^*$ can be found in $O(|V|)$.
\end{proof}
\medskip
\begin{proof}[Proof of Proposition~\ref{lem:3-connected-zero-p}]
Consider the case of general function $p$, and let $p_{\max} = \max_{j \in V} |p(j)|$.
We will show that we can find a solution such that $u \in V_1, v \in V_2, w \in V_3$ with $|p(V_1)|, |p(V_2)| \leq p_{\max}/2$.
Since $|p(V_3)| = |p(V_1) + p(V_2)|$ (recall $p(V)=0$),
this implies that $|p(V_3)| \leq p_{\max}$, and hence
$|p(V_1)|+ |p(V_2)| +|p(V_3)| \leq 2p_{\max}$.
Furthermore, if $ p(i) = \pm 1$ for all $i \in V$, hence $p_{\max}=1$,
then $|p(V_1)|, |p(V_2)| \leq p_{\max}/2$ implies that
$p(V_1) =p(V_2)=0$, and therefore also $p(V_3) = 0$.
Thus, both claims will follow.
Assume $u,v,w\in V$ and $p(u), p(v),p(w)>0$ (the case of negative $p(u), p(v),p(w)$
is symmetric).
Since we want to separate $u$ from $v$, we can assume without loss of generality that $\{u,v\}\in E$. Using~\cite{cheriyan1988finding}, there is a non-separating ear decomposition through the edge $\{u, v\}$ and avoiding node $w$.
Ignore the ears that do not contain any internal nodes, and let $Q_0\cup Q_1\cup\dots\cup Q_r$ be the decomposition consisting of the ears with nodes; we have $w\in Q_r$. Let $V_i=V(Q_0)\cup V(Q_1)\dots \cup V(Q_i)$, let $G_i=G[V_i]$ and $\overline{G}_i=G[V\backslash V_i]$. We distinguish two cases, depending on whether $p(V_0)\leq0$
or $p(V_0)>0$.
\begin{itemize}
\item[(i)] If $p(V_0)\leq0$, then consider an $st$-numbering between $u$ and $v$ in $V_0$, say $u=v_1,v_2,\dots,v_s=v$. Define $V_0^{(i)}=\{v_1,v_2,\dots,v_i\}$. Since $p(u),p(v)>0$ and $p(V_0)\leq0$, there must exist indices $1\leq i^*\leq j^*<s$ such that $p(V_0^{(i^*)})>0,~p(V_0^{(i^*+1)})\leq0$ and $p(V_0\backslash V_0^{(j^*+1)})>0,~p(V_0\backslash V_0^{(j^*)})\leq0$.
\begin{itemize}
\item[(a)] If $i^*=j^*$, since $p(V_0^{(i^*)})+ p(v_{i^*+1})+p(V_0\backslash V_0^{(i^*+1)})=p(V_0)<0$, we have $ p(V_0^{(i^*)})+p(V_0\backslash V_0^{(i^*+1)})\leq |p(v_{i^*+1})|$.
Now, one of the following three cases happens:
\begin{itemize}
\item[-] If $p(V_0^{(i^*)})\leq |p(v_{i^*+1})|/2$ and $p(V_0\backslash V_0^{(i^*+1)})\leq|p(v_{i^*+1})|/2$, then it is easy to see that $V_1=V_0^{(i^*)}$, $V_2=V_0\backslash V_0^{(i^*+1)}$, and $V_3=V\backslash (V_1\cup V_2)$ is a good partition.
\item[-] If $p(V_0^{(i^*)})> |p(v_{i^*+1})|/2$ and $p(V_0\backslash V_0^{(i^*+1)})\leq|p(v_{i^*+1})|/2$, then $p(V_0^{(i^*)})+p(v_{i^*+1})=p(V_0^{(i^*+1)})\leq |p(v_{i^*+1})|/2$. Hence, $V_1=V_0^{(i^*+1)}$, $V_2=V_0\backslash V_0^{(i^*+1)}$, and $V_3=V\backslash V_0$ is a good partition.
\item[-] If $p(V_0^{(i^*)})\leq |p(v_{i^*+1})|/2$ and $p(V_0\backslash V_0^{(i^*+1)})>|p(v_{i^*+1})|/2$, then $p(V_0\backslash V_0^{(i^*+1)})+p(v_{i^*+1})= p(V_0\backslash V_0^{(i^*)})\leq|p(v_{i^*+1})|/2$. Hence, $V_1=V_0^{(i^*)}$, $V_2=V_0\backslash V_0^{(i^*)}$, and $V_3=V\backslash V_0$ is a good partition.
\end{itemize}
\item[(b)] If $i^*<j^*$, then either $p(V_0^{(i^*)})\leq|p(v_{i^*+1})|/2$ or $|p(V_0^{(i^*+1)})|\leq|p(v_{i^*+1})|/2$, accordingly set $V_1=V_0^{(i^*)}$ or $V_1=V_0^{(i^*+1)}$. Similarly, either $p(V_0\backslash V_0^{(j^*+1)})\leq |p(v_{j^*+1})|/2$ or $|p(V_0\backslash V_0^{(j^*)})|\leq |p(v_{j^*+1})|/2$, so accordingly set $V_2=V_0\backslash V_0^{(j^*+1)}$ or $V_2=V_0\backslash V_0^{(j^*)}$. Set $V_3=V\backslash (V_1\cup V_2)$. It is easy to check that $(V_1,V_2,V_3)$ is a good partition.
\end{itemize}
\item[(ii)] If $p(V_0)>0$, then since $p(w)>0$ and therefore $p(V_{r-1})<0$, there must exist an index $0\leq j<r-1$ such that $p(V_j)>0$ and $p(V_{j+1})\leq0$. Consider an $st$-numbering between $u$ and $v$ in $G[V_j]$ as $u=v_1,v_2,\dots,v_s=v$ and define $V_j^{(i)}=\{v_1,v_2,\dots,v_i\}$. The
ear $Q_{j+1}$ is a path of new nodes $q_1,q_2,\dots, q_t$
attached to two (distinct) nodes $v_x, v_y$ of $G[V_j]$ through
edges $\{v_x,q_1\},\{q_t,v_y\}\in E$; assume wlog that
$1\leq x<y\leq s$.
For simplicity, we will use below $Q_{j+1}$ to denote also the
set $\{q_1,q_2,\dots, q_t\}$ of internal (new) nodes of the ear.
Also define $Q_{j+1}^{(i)}=\{q_1,q_2,\dots,q_i\}$ and $Q_{j+1}^{(0)}=\emptyset$.
One of the following cases must happen:
\begin{itemize}
\item[(a)] Suppose there is an index $1\leq i^*<(y-1)$ such that $p(V_j^{(i^*)})>0$ and $p(V_j^{(i^*+1)})\leq0$ or there is an index $x+1<i^*< s$ such that $p(V_j\backslash V_j^{(i^*-1)})>0$ and $p(V_j\backslash V_j^{(i^*)})\leq0$.
Let's assume there is an index $1\leq i^*< (y-1)$, such that $p(V_j^{(i^*)})>0$ and $p(V_j^{(i^*+1)})\leq0$ (the other case is exactly similar). Then either $p(V_j^{(i^*)})\leq|p(v_{i^*+1})|/2$ or $|p(V_j^{(i^*+1)})|\leq|p(v_{i^*+1})|/2$, accordingly set either $V_1=V_j^{(i^*)}$ or $V_1=V_j^{(i^*+1)}$. Set $V_2'=V_j\backslash V_1$. One of the following cases happens:
\begin{itemize}
\item[-] If $V_1=V_j^{(i^*)}$ and $p(V_2')\leq 0$, then since $p(V_j^{(i^*+1)})\leq0$, we have $p(V_j\backslash V_j^{(i^*+1)})>0$. Hence, $p(V_2'\backslash \{v_{i^*+1}\})>0$. So, it is either $|p(V_2')|\leq |p(v_{i^*+1})|/2$ or $p(V_2'\backslash \{v_{i^*+1}\})\leq |p(v_{i^*+1})|/2$. Now if $p(V_2'\backslash \{v_{i^*+1}\})\leq |p(v_{i^*+1})|/2$, since also $p(V_1)\leq |p(v_{i^*+1})|/2$, $p(V_j)\leq 0$ which contradicts with the assumption. Therefore, $|p(V_2')|\leq |p(v_{i^*+1})|/2$. Set $V_2=V_2'$ and $V_3=V\backslash (V_1\cup V_2)$. It is easy to check that $(V_1,V_2,V_3)$ is a good partition.
\item[-] If $V_1=V_j^{(i^*)}$ and $p(V_2')> 0$, then since $p(V_j\cup Q_{j+1})<0$, there is an index $0< t^*\leq t$, such that $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*)}))>0$ and $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*-1)}))\leq 0$. Hence, either $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*)}))\leq |p(q_{t^*})|/2$ or $|p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*-1)}))|\leq |p(q_{t^*})|/2$, accordingly set $V_2=V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*)})$ or $V_2=V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*-1)})$. Set $V_3=V\backslash (V_1\cup V_2)$. It is easy to see that $(V_1,V_2,V_3)$ is a good partition.
\item[-] If $V_1=V_j^{(i^*+1)}$, then since $p(V_1)\leq 0$, we have $p(V_2')>0$. The rest is exactly like the previous case when $V_1=V_j^{(i^*)}$ and $p(V_2')> 0$.
\end{itemize}
\item[(b)] Suppose that for every $1\leq i<y$, $p(V_j^{(i)})>0$ and for every $x<i< s$, $p(V_j\backslash V_j^{(i)})>0$. Set $V_1'=V_j^{(y-1)}$ and $V_2'=V_j\backslash V_1'$. Based on the assumption $p(V_1'),p(V_2')>0$. Since $p(V_{j+1})\leq 0$, there are indices $0\leq i^*\leq j^*<t$ such that $p(V_1'\cup Q_{j+1}^{(i^*)})>0$, $p(V_1'\cup Q_{j+1}^{(i^*+1)})\leq0$ and $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(j^*+1)}))>0$, $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(j^*)}))\leq0$. The rest of the proof is similar to case (i) when $p(V_0)\leq 0$.
\end{itemize}
\end{itemize}
\end{proof}
\section{Missing Proofs from Section~\ref{sec:BPGI}}\label{sec:proof2}
\begin{proof}[Proof of Lemma~\ref{lem:embedding}]
Set $X=\{v,u,w\}$. Using~\cite{linial1988rubber}, $G$ has a convex $X$-embedding in $\mathbb{R}^2$ in general position with mapping $f:V\rightarrow \mathbb{R}^2$ such that $f(u)=(0,0)$, $f(v)=(1,0)$, and $f(w)=(0,1)$. In the $X$-embedding of the nodes, we have a freedom to set the elasticity coefficient vector $\vec{c}$ to anything that we want (except a measure zero set of vectors). So for any edge $\{i,j\}\in G[V_1']\cup G[V_2']$, set $c_{ij}=g$; and for any $\{i,j\} \in E[V_1',V_2']$, set $c_{ij}=1$.
Assume $\mathcal{L}_1$ is the line $y=0.5$, $\mathcal{L}_2$ is the line $x+2y=1$, and $\mathcal{L}_3$ is the line $x=y$.
First, we show that there exist a $g$ for which all the nodes in $V_1'$ will be embedded above the line $\mathcal{L}_1$. To show this, from~\cite{linial1988rubber}, we know the embedding is such that it minimizes the total potential $P(f,\vec{c})= \sum_{\{i,j\}\in E} c_{ij}\|f(i)-f(j)\|^2$. Notice that we can independently minimize $P$ on $x$-axis values and $y$-axis values as below:
\begin{eqnarray*}
\min_{f} P &=& \min_{f_1} P_x +\min_{f_2} P_y\\
&=&\min_{f_1} \sum_{\{i,j\}\in E} c_{ij} (f_1(i)-f_1(j))^2+ \min_{f_2} \sum_{\{i,j\}\in E} c_{ij} (f_2(i)-f_2(j))^2
\end{eqnarray*}
Now, notice that if we place all the nodes in $V_1'$ at point (0,1) and all the nodes in $V_2'$ on the line $uv$, then $P_y\leq |E|$. Hence, if $f_2$ minimizes $P_y$, then $P_y(f_2,c)\leq |E|$. Set $g \geq 4|V|^2 |E|$. We show that if $f_2$ minimizes $P_y$, then for all edges $\{i,j\}\in G[V_1']\cup G[V_2']$, $(f_2(i)-f_2(j))^2\leq1/(4|V|^2)$. By contradiction, assume there is an edge $\{i,j\}\in G[V_1']\cup G[V_2']$ such that $(f_2(i)-f_2(j))^2>1/(4|V|^2)$. Then, $c_{ij}(f_2(i)-f_2(j))^2=g (f_2(i)-f_2(j))^2 >|E|$. Hence, $P_y(f_2,c)> |E|$ which contradicts with the fact the $f_2$ minimizes $P_y$. Therefore, if $g\geq4|V|^2 |E|$, then for all $\{i,j\}\in G[V_1']\cup G[V_2']$, $|f_2(i)-f_2(j)|\leq1/(2|V|)$. Now, since $G[V_1']$ is connected, all the nodes in $V_1'$ are connected to $w$ with a path of length (in number of hops) less than $|V|-1$. Hence, using the triangle inequality, for all $i\in V_1'$:
\begin{equation*}
|f_2(w)-f_2(i)|\leq (|V|-1)/(2|V|)<1/2\Rightarrow |1-f_2(i)|< 1/2,
\end{equation*}
which means that all the nodes in $V_1'$ are above $\mathcal{L}_1$.
With the very same argument, if $g\geq t^2 |V|^2 |E|$, then for all $i\in V_2'$, $f_2(i)<1/t$.
Now, we want to prove that there is a $g$ such that all the nodes in $V_2'$ will be embedded below the lines $\mathcal{L}_2$ and $\mathcal{L}_3$. Define $n_1(i):=|N(i)\cap V_1'|$ and $n_2(i):=|N(i)\cap V_2'|$. From~\cite{linial1988rubber}, we know the embedding is such that for all $i\in V\backslash\{u,v,w\}$, $f(i)=1/c_i\sum_{j\in N(i)}c_{ij}f(j)$, where $c_j=\sum_{j\in N(i)}c_{ij}f(j)$. Since $G[V'_2]$ is connected, for any $i\in V_2'$ there is a path $i=v_1,v_2,\dots,v_r=v$ in $V_2'$. Using this ordering:
\begin{eqnarray*}
&&\begin{cases}
f_1(v_j)\geq \frac{1}{n_2(v_j)g+n_1(v_j)}g f_1(v_{j+1})\geq (1/|V|)f_1(v_{j+1}),& \forall j\in \{1,\dots,r-1\}\\
f_1(v_r)=f_1(v)=1
\end{cases}
\\&&\Rightarrow \forall i\in V_2'\backslash\{u,v\}:~ f_1(i)\geq (1/|V|)^{r}\geq(1/|V|)^{|V|}.
\end{eqnarray*}
On the other hand, from the previous part, if we set $g\geq |V|^{2|V|+2}|E|$, then for all $i\in V_2'$, $f_2(i)\leq (1/|V|)^{|V|}$. Hence, for all $i\in V_2'$, $f_2(i)\leq f_1(i)$, which means that all the nodes in $V_2'$ will be placed below the line $\mathcal{L}_3$.
With the very same idea, we show that there exist a $g$ for which all the nodes in $V_2'$ will be placed below the line $\mathcal{L}_2$. Since $G[V'_2]$ is connected, for any $i\in V_2'$ there is a path $u=u_1,v_2,\dots,u_t=i$ in $V_2'$. Notice that for all $i\in V\backslash\{u,v,w\}$, $1-f_1(i)=1/c_i\sum_{j\in N(i)}c_{ij}(1-f_1(j))$. Hence, since $\forall j\in V: f_1(j)\leq 1$, we have,
\begin{eqnarray*}
&&\begin{cases}
1-f_1(u_j)\geq \frac{1}{n_2(u_j)g+n_1(u_j)}g (1-f_1(u_{i-1}))\geq (1/|V|)(1-f_1(u_{i-1})),& \forall j\in \{2,\dots,t\}\\
1-f_1(u)=1-f_1(u_1)=1
\end{cases}
\\&& \Rightarrow \forall i\in V_2'\backslash\{u,v\}:~1-f_1(i)\geq (1/|V|)^{t}\geq (1/|V|)^{|V|}.
\end{eqnarray*}
From the previous part, if we set $g\geq 4 |V|^{2|V|+2}|E|$, then for all $i\in V_2'$, $f_2(i)\leq 1/2(1/|V|)^{|V|}$. Hence, for $i\in V_2'$, $f_1(i)+2f_2(i)\leq 1$, which means that all the nodes in $V_2'$ will be placed below the line $\mathcal{L}_3$. Therefore, if we set $g\geq 4|V|^{2|V|+2}|E|$, then we will get an embedding as depicted in Fig.~\ref{fig:3-connected}.
Note that a polynomial number of bits suffices for $g$
Notice that if $\vec{c}$ is a ``good" vector, then so is $\vec{c}+\vec{\epsilon}$ in which $\vec{\epsilon}$ is a vector with very small Euclidean norm. Hence, we can always find a ``good" vector $\vec{c}$ which result in a $X$-embedding in general position.
\end{proof}
\medskip
\begin{proof}[Proof of Theorem~\ref{th:3-connected}]
Assume that $|V|\equiv0(\mathrm{mod}~4)$; the proof for the case $|V|\equiv2(\mathrm{mod}~4)$ is similar. Using Lemma~\ref{lem:cut}, we can find $\{u,v,w\}\in V$ and a partition $(V_1',V_2')$ of $V$ with properties described in the Lemma. Set $X=\{u,v,w\}$. Using Lemma~\ref{lem:embedding}, we can find a convex $X$-embedding of $G$ in general position with properties described in the Lemma as depicted in Fig.~\ref{fig:3-connected}.
The rest of the proof is very similar to the proof of Lemma~\ref{lem:3-connected-tri}. We consider again a circle $\mathcal{C}$ around $f(u),f(v),f(w)$ in $\mathbb{R}^2$ as shown in Fig.~\ref{fig:3-connected}. Also consider a directed line $\mathcal{L}$ tangent to the circle $C$ at point $A$. If we project the nodes of $G$ onto the line $\mathcal{L}$, this time the order of the nodes projection gives an $st$-numbering between the first and the last node only if $u$ and $v$ are the first and last node. However, if we set $V_1$ to be the $|V|/2$ nodes whose projections come first and $V_2$ are the $|V|/2$ nodes whose projections come last, then $G[V_1]$ and $G[V_2]$ are both connected even when $u$ and $v$ are not the first and last nodes. The reason lies on the special embedding that we considered here. Assume for example $w$ and $v$ are the first and the last projected nodes, and $V_1$ and $V_2$ are set of the $|V|/2$ nodes which projections come first and last, respectively. Two cases might happen:
\begin{itemize}
\item[(i)] If $u,w\in V_1$ and $v\in V_2$, then since $\{u,w\}\in E$, both $G[V_1]$ and $G[V_2]$ are connected because of the properties of the embedding.
\item[(ii)] If $w\in V_1$ and $u,v\in V_2$, since $|V_2'|\leq |V|/2$ and $|V_2|=|V|/2$, then either $V_2=V_2'$ or $V_2 \cap V_1' \neq \emptyset$.
If $V_2= V_2'$, and hence $V_1=V_1'$ then there is nothing to prove.
So assume there is a node $z\in V_2 \cap V_1'$.
From the properties of the embedding, the triangle $\{z,u,v\}$
contains all the nodes of $V_2'$.
Since $\{z,u,v\}\in V_2$, and $V_2$ contains all the nodes that are on a same side of a halfplane, we should also have $V_2'\subset V_2$. Now, from the properties of the embedding, it is easy to see that every node in $V_2$ has a path either to $u$ or $v$. Since $V_2'\subset V_2$, there is also a path between $u$ and $v$. Thus, $G[V_2]$ is connected. From the properties of the embedding, $G[V_1]$ is connected as before.
\end{itemize}
The rest of the proof is exactly the same as the proof of Lemma~\ref{lem:3-connected-tri}. We move $\mathcal{L}$ from being tangent at point $A$ to point $B$ ($AB$ is a diameter of the circle $\mathcal{C}$) and consider the resulting partition. Notice that if at point $A$, $p(V_1)>0$, then at point $B$ since $V_1$ and $V_2$ completely switch places compared to the partition at point $A$, $p(V_1)<0$. Hence, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ and keep it tangent to the circle, in the resulted partitions, $p(V_1)$ goes from some positive value to a negative value. Notice that the partition $(V_1,V_2)$ changes only if $\mathcal{L}$ passes a point $D$ on the circle such that at $D$, $\mathcal{L}$ is perpendicular to a line that connects $f(i)$ to $f(j)$ for $i,j\in V$. Now, since the embedding is in general position, there are exactly two points on every line that connects two points $f(i)$ and $f(j)$, so $V_1$ changes at most by one node leaving $V_1$ and one node entering $V_1$. Hence, $p(V_1)$ changes by either $\pm 2$ or $0$ value at each change. Now, since $|V|\equiv 0 (\mathrm{mod}~ 4)$, $p(V_1)$ has an even value in all the resulting partitions. Therefore, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$, there should be a point $D$ such that in the resulted partition $p(V_1)=p(V_2)=0$.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary~\ref{cor:serpar}]
Every series-parallel graph $G$ has a separation pair $\{u,v\}$
such that every connected component of $G[V\backslash \{u,v\}]$ has less that $2|V|/3$ nodes,
and furthermore, such a separation pair can be found in linear time.
To see this, consider the derivation tree $T$ of the construction of $G$.
The root of $T$ corresponds to $G$, the leaves correspond to the edges, and every internal node
$i$ corresponds to a subgraph $G_i=(V_i,E_i)$ that is the series or parallel composition of the
subgraphs corresponding to its children.
Starting at the root of $T$, walk down the tree following always the edge to the child corresponding to
a subgraph with the maximum number of nodes until the number of nodes becomes $\leq 2|V|/3$.
Thus, we arrive at a node $i$ of the tree such that $|V_i| > 2|V|/3$
and $|V_j| \leq 2|V|/3$ for all children $j$ of $i$.
Let $u_i, v_i$ be the terminals of $G_i$. Note that $u_i, v_i$ separate all
the nodes of $G_i$ from all the nodes that are not in $G_i$.
Since $|V_i| > 2|V|/3$, we have $|V \backslash V_i| < |V|/3$.
If $G_i$ is the parallel composition of the graphs corresponding to the children of $i$,
then the separation pair $\{u_i, v_i\}$ has the desired property,
i.e. all the components of $G[V\backslash \{u,v\}]$ have less than $2|V|/3$ nodes.
Suppose $G_i$ is the series composition of the graphs $G_j$, $G_k$
corresponding to the children $j,k$ of $i$, and let $w$ be the common
terminal of $G_j$, $G_k$; thus, $G_i$ has terminals $u_i, w$,
and $G_k$ has terminals $w, v_i$.
Assume wlog that $|V_j| \geq |V_k|$.
Then $|V|/3 < |V_j| \leq 2|V|/3$.
The pair $\{u_i,w\}$ of terminals of $G_j$ separates all the nodes
of $V_j \backslash \{u_i,w\}$ from all the nodes of $V \backslash V_j$,
and both these sets have less than $2|V|/3$ nodes.
Thus, $\{u_i,w\}$ has the required property.
\end{proof}
\medskip
\begin{proof}[Proof of Theorem~\ref{th:2-connected}]
Using Lemma~\ref{lem:2-connected-pair} for $q=4$, we consider two cases:
\begin{itemize}
\item[(i)] There is a separation pair $\{u,v\}\in V$ such that if $G_1,\dots,G_k$ are the connected components of $G\backslash \{u,v\}$, for any $i$, $|V_i|< 3|V|/4$. In this case Lemma~\ref{lem:parallel} for $q=4$ proves the theorem.
\item[(ii)] After a set of contractions, $G$ can be transformed into a 3-connected graph $G^*=(V^*,E^*)$ with weighted edges such that for any edge $e^*\in E^*$, $w(e^*)<|V|/4$. In this case the proof is similar to the proof of Theorem~\ref{th:3-connected}. Notice that if $G^*$ contains a triangle then the proof is much simpler as in the proof of Lemma~\ref{lem:3-connected-tri} but here to avoid repetition, we use the approach in the proof of Theorem~\ref{th:3-connected} and prove the theorem once for all cases of $G^*$.
Recall that for each edge in $G^*$, its weight represents the number of nodes in the contracted subgraph that it represents. So if $e^*\in E^*$ represents an induced subgraph of $G$ with $l$ nodes, then $w(e^*)=l$; and if $e^*\in E\cap E^*$ then $w(e^*)=0$. Using Lemma~\ref{lem:cut}, we can find $\{u,v,w\}\in V^*$ and a partition $(V_1^*,V_2^*)$ of $V^*$ with properties described in the Lemma. Since in $G^*$ edges have weights that actually represent nodes in $G$, we want a partition such that we also have $|V^*_2|+\sum_{e\in G[V_2^*]}w(e)\leq |V|/2$. Notice that we can modify the proof of Lemma~\ref{lem:cut} to take into account the weights for the edges and find a partition such that $|V^*_2|+\sum_{e\in G[V_2^*]}w(e)\leq |V|/2$ as follows. Again, using the algorithm presented in~\cite{cheriyan1988finding}, we can find a non-separating proper cycle $C_0$ in $G^*$ such that every node in $C_0$ has a neighbor in $G^*\backslash C_0$. We consider three cases:
\begin{itemize}
\item[(a)] If $|C_0|+\sum_{e\in G[C_0]}w(e)\leq |V|/2+1$, then the proof is as in the proof of Lemma~\ref{lem:cut}. Select any three consecutive nodes $(u,w,v)$ of $C_0$ and set $V_2^*=C_0\backslash \{w\}$ and $V_1^*=V^*\backslash V_2^*$.
\item[(b)] If $|C_0|>|V^*|/2$ and $|C_0|+\sum_{e\in G[C_0]}w(e)> |V|/2+1$, then as in the proof of Lemma~\ref{lem:cut}, since $|C_0|>|V^*\backslash C_0|$ and every node is $C_0$ has a neighbor in $V^*\backslash C_0$, there exist a node $w\in V^*\backslash C_0$ such that $|N(w)\cap C_0|\geq 2$. Select two nodes $u,v\in N(w)\cap C_0$. There exists a path $P$ in $C_0$ between $u$ and $v$ such that $|P|+\sum_{e\in G[P]}w(e)<|V|/2-1$. Set $V_2^*=\{u,v\}\cup P$ and $V_1^*=V^*\backslash V_2^*$.
\item[(c)] If $|C_0|\leq |V^*|/2$ and $|C_0|+\sum_{e\in G[C_0]}w(e)> |V|/2+1$, then if there exist a node $w\in V^*\backslash C_0$ such that $|N(w)\cap C_0|\geq 2$, the proof is as in the previous part. So assume for all $w\in V^*\backslash C_0$, $|N(w)\cap C_0|\leq 1$. Assume $w_1,w_2,\dots,w_t\in V^*\backslash C_0$ are all the nodes with $|N(w_i)\cap C_0|=1$. We show that there is a $1\leq i\leq t$, such that $w_i$ is not a cut-point of $G^*[V^*\backslash C_0]$. Let $T$ be a spanning tree of $G^*[V^*\backslash C_0]$. If there is a $1\leq i\leq t$, such that $w_i$ is a leaf of $T$, then $w_i$ is not a cut-point of $G^*[V^*\backslash C_0]$ and there is nothing left to prove. So assume none of $w_i$s is a leaf of $T$. Suppose $1\leq p,q\leq t$ are such that the path between $w_p$ and $w_q$ in $T$ is the longest between all pairs of $w_i$s. We show that $w_p$ and $w_q$ cannot be cut-points of $G^*[V^*\backslash C_0]$. Assume for example $w_p$ is a cut-point of $G^*[V^*\backslash C_0]$ and $G_{\bar{q}}^*$ is a connected component of $G^*[V^*\backslash C_0]\backslash\{w_p\}$ such that $w_q\notin G_{\bar{q}}^*$. Since the path between $w_p$ and $w_q$ in $T$ is the longest between all pairs of $w_i$s, $\forall 1\leq i\leq t: w_i\notin G_{\bar{q}}^*$, otherwise we can find a longer path from $w_q$ to some other $w_i$. On the other hand, if $\forall 1\leq i\leq t: w_i\notin G_{\bar{q}}$, in $G^*\backslash \{w_p\}$ the cycle $C_0$ is disconnected from $G_{\bar{q}}^*$ which contradicts with the 3-connectedness of $G^*$. Hence, there should be at least a noncut-point $w\in V^*\backslash C_0$ with $|N(w)\cap C_0|=1$. Since $G^*$ is 3-connected, each node has degree at least 3. Hence, there are two nodes $u,v\in N(w)\cap (V^*\backslash C_0)$. Set $V_1^*=C_0\cup\{w\}$ and $V_2^*=V^*\backslash V_1^*$. $(V_1^*,V_2^*)$ is a desirable partition for $V^*$.
\end{itemize}
Hence, we can find $\{u,v,w\}\in V^*$ and a partition $(V_1^*,V_2^*)$ of $V^*$ with properties described in the Lemma~\ref{lem:cut} as well as having $|V^*_2|+\sum_{e\in G[V_2^*]}w(e)\leq |V|/2$. Set $X=\{u,v,w\}$. Using Lemma~\ref{lem:embedding}, $G^*$ has a convex $X$-embedding in general position like $f^*:V^*\rightarrow \mathbb{R}^2$ as described in the lemma and depicted in Fig.~\ref{fig:3-connected}. Now, from this embedding, we get a convex $X$-embedding for $G$ like $f:V\rightarrow \mathbb{R}^2$ as follows. For any $i\in V\cap V^*$, $f(i)=f^*(i)$. For any edge $\{i,j\}\in E^*$ such that $\{i,j\}$ represents an induced subgraph of $G$, we represent it by a pseudo-path between $i$ and $j$ in $G$ like $P$ and place the nodes in $P$ in order on random places on the line segment that connects $f(i)$ to $f(j)$. If the edge $\{i,j\}\in E^*$ is between a node in $V_1^*$ and a node in $V_2^*$ and represents a pseudo-path $P$ in $G$ , we place the nodes in $P$ in order on random places on the segment that connects $f(i)$ to $f(j)$ but above the line $\mathcal{L}_1$. Hence, by this process, we get a convex $X$-embedding for $G$ which is in general position (almost surely) except for the nodes that are part of a pseudo-path (which we know have length less than $|V|/4$). Notice that if $V_2'\subset V$ contains all the nodes in $V_2^*$ and all the nodes that are part of a pseudo-path between the nodes of $V_2^*$ (represented by weighted edges in $G^*$) in $G$, and $V_1'=V\backslash V_2'$, then $(V_1',V_2')$ is partition of $G$ with all the properties of Lemma~\ref{lem:cut}. Moreover, the $X$-embedding $f$ of $G$ that we derived from $f^*$ for $G^*$, has all the properties of Lemma~\ref{lem:embedding} for the partition $(V_1',V_2')$.
The rest of the proof is similar to the proof of Theorem~\ref{th:3-connected}. We consider again a circle $\mathcal{C}$ around $f(u),f(v),f(w)$ in $\mathbb{R}^2$ as shown in Fig.~\ref{fig:3-connected}. Also consider a directed line $\mathcal{L}$ tangent to the circle $C$ at point $A$ and project nodes of $G$ onto the line $\mathcal{L}$. With the same argument as in the proof of Theorem~\ref{th:3-connected}, since the embedding $f$ has the properties in Lemma~\ref{lem:cut} and \ref{lem:embedding}, if we set $V_1$ to be the $|V|/2$ nodes whose projections come first and $V_2$ are the $|V|/2$ nodes whose projections come last, then $G[V_1]$ and $G[V_2]$ are both connected. If $|p(V_1)|\leq 1$, then $(V_1,V_2)$ is a good partition and there is nothing left to prove. Otherwise, we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ ($AB$ is a diameter of the circle $\mathcal{C}$) and consider the resulting partition. Notice that if at point $A$, $p(V_1)>0$, then at point $B$ since $V_1$ and $V_2$ completely switch places compared to the partition at point $A$, $p(V_1)<0$. Hence, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ and keep it tangent to the circle, in the resulting partitions, $p(V_1)$ goes from some positive value to a negative value. Notice that the partition $(V_1,V_2)$ changes only if $\mathcal{L}$ passes a point $D$ on the circle such that at $D$, $\mathcal{L}$ is perpendicular to a line that connects $f(i)$ to $f(j)$ for $i,j\in V$. Now, since the embedding is in general position except for few lines that contain a pseudo-path, if $\{i,j\}\notin E\cup E^*$ or $\{i,j\}\in E\cap E^*$ there are exactly two points on the line that connects $f(i)$ and $f(j)$, so $V_1$ changes at most by one node leaving $V_1$ and one node entering $V_1$. Thus, in this case $p(V_1)$ changes by at most 2. However, if $\{i,j\}\in E^*$, then since there can be at most $|V|/4-1$ points on the line that connects $f(i)$ to $f(j)$, $p(V_1)$ may change by $|V|/4-1$. So let's consider that $\mathcal{L}$ is perpendicular to the line that connects $f(i)$ and $f(j)$ at point $D$ and $p(V_1)>0$ slightly before $\mathcal{L}$ passes point $D$ and $p(V_1)\leq 0$ slightly after $\mathcal{L}$ passes point $D$. We consider 2 cases:
\begin{itemize}
\item[(a)] Suppose $\{i,j\}\notin E\cup E^*$ or $\{i,j\}\in E\cap E^*$. In this case since $V_1$ changes by at most $2$, therefore, $p(V_1)$ slightly after $\mathcal{L}$ passes $D$ is either $0$ or $-1$. Hence, the partition $(V_1,V_2)$ that we get after $\mathcal{L}$ passes $D$ is a good partition with $|V_1|=|V_2|=|V|/2$.
\item[(b)] Suppose $\{i,j\}\in E^*$. Let $(V_{b1},V_{b2})$ be the partition slightly before $\mathcal{L}$ passes $D$ and let $(V_{a1},V_{a2})$ be the partition slightly after. Assume $P_0$ is the pseudo-path between $i$ and $j$ and $P=P_0\cup\{i,j\}$. It is easy to see that $(V_{b1}\backslash P,V_{b2}\backslash P)=(V_{a1}\backslash P,V_{a2}\backslash P)$.
If $p(V_{b1}\backslash P)=0$, since $|P_0|<|V|/4$, then $V_1=V_{b1}\backslash P$ and $V_2= V\backslash V_1$ is a good partition and there is nothing left to prove. So either $p(V_{b1}\backslash P)> 0$ or $p(V_{b1}\backslash P)< 0$.
If $p(V_{b1}\backslash P)> 0$, since $p(V_{a1})\leq 0$ and $V_{b1}\backslash P=V_{a1}\backslash P$, we can add a set of nodes from $P\cap V_{a1}$, say $P'$, to $V_{a1}\backslash P$ to get $p((V_{a1}\backslash P)\cup P')=0$. Now since $|P_0|<|V|/4$, $V_1=(V_{a1}\backslash P)\cup P'$ and $V_2=V\backslash V_1$ is a good partition.
If $p(V_{b1}\backslash P)< 0$, since $p(V_{b1})> 0$, we can add a set of nodes from $P\cap V_{b1}$, say $P'$, to $V_{b1}\backslash P$ to get $p((V_{b1}\backslash P)\cup P')=0$. Now since $|P_0|<|V|/4$, $V_1=(V_{b1}\backslash P)\cup P'$ and $V_2=V\backslash V_1$ is a good partition.
\end{itemize}
\end{itemize}
\end{proof}
\medskip
\section{Missing Proofs from Section~\ref{sec:blue_red}}\label{sec:proof3}
\begin{proof}[Proof of Corollary~\ref{cor:blue_red3}]
Suppose without loss of generality that $n_r \geq n_b$ and
let $n_r-n_b=2t$ and $n_r+ n_b =n =2m$.
Set $p(i)=1$ for $i \in R$ and $p(i)=-1$ for $i \in B$.
Then $p(V) =2t$.
From the equations, we have $n_r=m+t$ and $n_b= m-t$.
From Corollary ~\ref{cor:3-general} we can find a partition $(V_1,V_2)$
such that $|V_1|=|V_2|$ and $|p(V_1)-p(V)/2|, |p(V_1)-p(V)/2| \leq 1$.
Let $r_1 = |R \cap V_1|$ and $b_1 = |B \cap V_1|$.
We have $r_1 + b_1 = n/2 =m$ and $t-1 \leq r_1 - b_1 \leq t+1$.
Therefore, $(m+t)/2 - (1/2) \leq r_1 \leq (m+t)/2 +(1/2)$.
Since $r_1$ is an integer and $n_r=m+t$ is even,
it follows that $r_1 = (m+t)/2 = n_r/2$.
Hence, $b_1 = (m-t)/2 = n_b/2$.
Therefore, $V_2$ also contains $n_r/2$ red nodes and
$n_b/2$ blue nodes.
\end{proof}
\section{Introduction}
Power Grid Islanding is an effective method to mitigate cascading failures in power grids~\cite{sun2002two}. The challenge is to partition the network into smaller connected components, called \emph{islands}, such that each island can operate independently for a while. In order for an island to operate, it is necessary that the power supply and demand at that island are almost equal.\footnote{If the supply and demand are not exactly equal but still relatively close, load shedding/generation curtailing can be used in order for the island to operate.} Equality of supply and demand in an island, however, may not be sufficient for its independent operation. It is also important that the infrastructure in that island has the physical capacity to safely transfer the power from the supply nodes to the demand nodes. When the island is large enough compared to the initial network, it is more likely that it has enough capacity. This problem has been studied in the power systems community but almost all the algorithms provided in the literature are heuristic methods that have been shown to be effective only by simulations~\cite{sun2002two,sanchez2014hierarchical,pahwa2013optimal,fan2012mixed}.
Motivated by this application, we formally introduce and study the Doubly Balanced Connected graph Partitioning (DBCP) problem:
Let $G=(V,E)$ be a connected graph with a weight (supply/demand) function $p:V\rightarrow \mathbb{Z}$ satisfying $p(V)=\sum_{j\in V} p(j)=0$. The objective is to partition $V$ into $(V_1,V_2)$ such that $G[V_1]$ and $G[V_2]$ are connected, $|p(V_1)|,|p(V_2)|\leq c_p$, and $\max\{\frac{|V_1|}{|V_2|},\frac{|V_2|}{|V_1|}\}\leq c_s$, for some constants $c_p$ and $c_s$. We also consider the case that $p(V)\neq 0$, in which the
excess supply/demand should be split roughly evenly.
The problem calls for a partition into two connected subgraphs that simultaneously balances two objectives, (1) the supply/demand within each part, and (2) the sizes of the parts.
The connected partitioning problem with only the size objective has been studied previously.
In the most well-known result, Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division} independently proved, using different methods, that every $k$-connected graph can be partitioned into $k$ arbitrarily sized connected subgraphs. However, neither of the proofs is constructive, and there are no known
polynomial-time algorithms to find such a partition for $k>3$.
For $k=2$, a linear time algorithm is provided in~\cite{suzuki1990linear} and for $k=3$
an $O(|V|^2)$ algorithm is provided in~\cite{wada1994efficient}.\footnote{For $k=2$, a much simpler approach than the one in~\cite{suzuki1990linear} is to use the $st$-numbering~\cite{lempel1967algorithm} for 2-connected graphs.}
The complexity of the problem with the size objective and related optimization problems
have been studied in~\cite{dyer,chlebikova1996approximating,chataigner2007approximation}
and there are various NP-hardness and inapproximability results.
Note that the size of the cut is not of any relevance here (so the
extensive literature on finding balanced partitions, not necessarily connected,
that minimize the cut is not relevant.)
The objective of balancing the supply/demand alone, when all $p(i)$ are $\pm1$, can also be seen as an extension for the objective of balancing the size (which corresponds to $p(i)=1$).
Our bi-objective problem of balancing both supply/demand and size, can be seen also
as an extension of the problem of finding a partition that balances the size for two types of nodes
simultaneously: Suppose the nodes of a graph are partitioned into red and blue nodes.
Find a partition of the graph into two large connected subgraphs that splits approximately evenly both
the red and the blue nodes.
We now summarize our results and techniques. Since the power grids are designed to withstand a single failure (``$N-1$" standard)~\cite{bienstock2016electrical}, and therefore 2-connected, our focus is mainly on the graphs that are at least 2-connected.
We first, in Section \ref{sec:BPGI}, study the connected partitioning problem with
only the supply/demand balancing objective, and show results that
parallel the results for balancing size alone, using similar techniques:
The problem is NP-hard in general.
For 2-connected graphs and weights $p(i)=\pm1$, there is always a perfectly balanced partition
and we can find it easily using an $st$-numbering. For 3-connected graphs and weights $p(i)=\pm1$,
there is a perfectly balanced partition into three connected graphs, and we can find it
using a nonseparating ear decomposition of 3-connected graphs~\cite{cheriyan1988finding} and similar ideas as in~\cite{wada1994efficient}.
The problem is more challenging when we deal with both balancing objectives, supply/demand and size.
This is the main focus and occupies the bulk of this paper.
Our main results are existence results and algorithms for 2- and 3-connected graphs.
It is easy to observe that we cannot achieve perfection in one objective ($c_p=0$ or $c_s =1$)
without sacrificing completely the other objective.
We show that allowing the supply/demand of the parts to be off balance by at most the weight of one node
suffices to get a partition that is roughly balanced also with respect to size.
First, in Section~\ref{subsec:3-connected}, we study the case of 3-connected graphs
since we use this later as the basis of handling 2-connected graphs.
We show that if $\forall i,~p(i)=\pm1$, there is a partition that is perfectly balanced
with respect to both objectives, if $|V|\equiv0 (\mathrm{mod}~4)$ (otherwise the
sizes are slightly off for parity reasons); for general $p$, the partition is perfect
in both objectives up to the weight of a single node. Furthermore, the partition can be constructed in
polynomial time. Our approach uses the convex embedding characterization of $k$-connectivity
studied by Linial, Lov\'{a}z, and Wigderson~\cite{linial1988rubber}. We need to adapt it for
our purposes so that the convex embedding also has certain desired geometric properties,
and for this purpose we use the nonseparating ear decomposition of 3-connected graphs
of~\cite{cheriyan1988finding} to obtain a suitable embedding.
Then, in Section~\ref{subsec:2-connected}, we analyze the case of 2-connected graphs.
We reduce it to two subcases: either (1) there is a separation pair that splits the graph
into components that are not very large, or (2) we can perform a series of
contractions to achieve a 3-connected graph whose edges represent
contracted subgraphs that are not too large. We provide a good partitioning algorithm
for case (1), and for case (2) we extend the algorithms for 3-connected graphs
to handle also the complications arising from edges representing contracted subgraphs.
Finally, in Section~\ref{sec:blue_red}, we briefly discuss the problem
of finding a connected partitioning of a graph with two types of nodes
that splits roughly evenly both types.
The body of the paper contains proofs and sketches for some of the results,
and Appendices~\ref{sec:proof1}, \ref{sec:proof2}, and \ref{sec:proof3} contain the remaining proofs.
Graph-theoretic background and definitions (e.g., the notions of $st$-numbering,
nonseparating ear decomposition, convex embedding of $k$-connected graphs)
are given in Appendix~\ref{sec:pre}.
\section{Definitions and Background}\label{sec:pre}
In this section, we provide a short overview of the definitions and tools used in our work. Most of the graph theoretical terms used in this paper are relatively standard and borrowed from~\cite{bondy2008graph} and \cite{west2001introduction}. All the graphs in this paper are loopless.
\subsection{Terms from Graph Theory}
\noindent\textbf{Cutpoints and Subgraphs:} A \emph{cutpoint} of a connected graph $G$ is a node whose deletion results in a disconnected graph.
Let $X$ and $Y$ be subsets of the nodes of a graph $G$. $G[X]$ denotes the subgraph of $G$ \emph{induced} by $X$. We denote by $E[X, Y]$ the set of edges of $G$ with one end in $X$ and the other end in $Y$. The neighborhood of a node $v$ is denoted $N(v)$.
\noindent\textbf{Connectivity:} The connectivity of a graph $G=(V,E)$ is the minimum size of a set $S\subset V$ such that $G\backslash S$ is not connected. A graph is $k$-connected if its connectivity is at least $k$.
\subsection{$st$-numbering of a Graph}\label{subsec:st-numbering}
Given any edge $\{s,t\}$ in a 2-connected graph $G$, an \emph{$st$-numbering} for $G$ is a numbering for the nodes in $G$ defined as follows~\cite{lempel1967algorithm}: the nodes of $G$ are numbered from 1 to $n$ so that $s$ receives number 1, node $t$ receives number $n$, and every node except $s$ and $t$ is adjacent both to a lower-numbered and to a higher-numbered node. It is shown in~\cite{even1976computing} that such a numbering can be found in $O(|V|+|E|)$.
\subsection{Series-Parallel Graphs
A Graph $G$ is \emph{series-parallel}, with terminals $s$ and $t$, if it can be produced by a sequence of the following operations:
\begin{enumerate}
\item Create a new graph, consisting of a single edge between $s$ and $t$.
\item Given two series parallel graphs, $X$ and $Y$ with terminals $s_X,t_X$ and $s_Y,t_Y$ respectively, form a new graph $G=P(X,Y)$ by identifying $s=s_X=s_Y$ and $t=t_X=t_Y$. This is known as the \emph{parallel composition} of $X$ and $Y$.
\item Given two series parallel graphs $X$ and $Y$, with terminals $s_X,t_X$ and $s_Y,t_Y$ respectively, form a new graph $G=S(X,Y)$ by identifying $s=s_X, t_X=s_Y$ and $t=t_Y$. This is known as the \emph{series composition} of $X$ and $Y$.
\end{enumerate}
It is easy to see that a series-parallel graph is 2-connected if, and only if, the last operation is a parallel composition.
\subsection{Nonseparating Induced Cycles and Ear Decomposition}\label{subsec:ear-decomposition}
Let $H$ be a subgraph of a graph $G$. An \emph{ear} of $H$ in $G$ is a nontrivial path in $G$ whose ends lie in $H$ but whose internal vertices do not. An ear decomposition of $G$ is a decomposition $G=P_0\cup\dots\cup P_k$ of the edges of $G$ such that $P_0$ is a cycle and $P_i$ for $i\geq 1$ is an ear of $P_0\cup P_1\cup\dots\cup P_{i-1}$ in $G$. It is known that every 2-connected graph has an ear decomposition (and vice-versa), and such a decomposition can be found in linear time.
A cycle $C$ is a \emph{nonseparating induced cycle} of $G$ if $G\backslash C$ is connected and $C$ has no chords. We say a cycle $C$ avoids a node $u$, if $u\notin C$.
\begin{theorem}[Tutte~\cite{tutte1963draw}] \label{th:tutte-3-connected}
Given a 3-connected graph $G(V,E)$ let $\{t,r\}$ be any edge of $G$ and let $u$ be any node of $G$, $r\neq u\neq t$. Then there is a nonseparating induced cycle of $G$ through $\{t,r\}$ and avoiding $u$.
\end{theorem}
Notice that since $G$ is 3-connected in the previous theorem, every node in $C$ has a neighbor in $G\backslash C$. Cheriyan and Maheshwari showed that the cycle in Theorem~\ref{th:tutte-3-connected} can be found in $O(E)$~\cite{cheriyan1988finding}. Moreover, they showed that any 3-connected graph $G$ has a nonseparating ear decomposition $G=P_0\cup\dots\cup P_k$ defined as follows: Let $V_i=V(P_0)\cup V(P_1)\dots \cup V(P_i)$, let $G_i=G[V_i]$ and $\overline{G}_i=G[V\backslash V_i]$. We say that $G=P_0\cup P_1\cup\dots\cup P_k$ is an ear decomposition through edge $\{t,r\}$ and avoiding vertex $u$ if the cycle $P_0$ contains edge $\{t,r\}$, and the last ear of length greater than one, say $P_m$, has $u$ as its only internal vertex. An ear decomposition $P_0\cup P_1\dots\cup P_k$ of graph $G$ through edge $\{t,r\}$ and avoiding vertex $u$ is a \emph{nonseparating ear decomposition} if for all $i$, $0\leq i<m$, graph $\overline{G}_i$ is connected and each internal vertex of ear $P_i$ has a neighbor in $\overline{G}_i$.
\begin{theorem}[Cheriyan and Maheshwari~\cite{cheriyan1988finding}]
Given an edge $\{t,r\}$ and a vertex $u$ of a 3-connected graph $G$, a nonseparating induced cycle of $G$ through $\{t,r\}$ and avoiding $u$, and
a nonseparating ear decomposition can be found in time $O(|V|+|E|)$.
\end{theorem}
\subsection{Partitioning of Graphs to Connected Subgraphs}
The following theorem is the main existing result in partitioning of graphs into connected subgraphs and is proved independently by Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division} by different methods.
\begin{theorem}[Lov\'{a}z and Gyori~\cite{lovasz1977homology,gyori1976division}]
Let $G=(V,E)$ be a $k$-connected graph. Let $n=|V|,~ v_1,v_2,\dots,v_k\in V$ and let $n_1,n_2,\dots,n_k$ be positive integers satisfying $n_1+n_2+\dots+n_k=n$. Then, there exists a partition of $V$ into $(V_1,V_2\dots,V_k)$ satisfying $v_i\in V_i, |V_i|=n_i$, and $G[V_i]$ is connected for $i=1,2,\dots,k$.
\end{theorem}
Although the existence of such a partition has long been proved, there is no polynomial-time algorithm to find such a partition for $k>3$. For $k=2$, it is easy to find such partition using $st$-numbering. For $k=3$, Wada and Kawaguchi~\cite{wada1994efficient} provided an $O(n^2)$ algorithm using the nonseparating ear decomposition of $3$-connected graph.
\subsection{Convex Embedding of Graphs}\label{subsec:embedding}
In this subsection, we provide a short overview of the beautiful work by Linial, Lov\'{a}z, and Wigderson~\cite{linial1988rubber} on convex embedding of the $k$-connected graphs.
Let $Q=\{q_1,q_2,\dots,q_m\}$ be a finite set of points in $\mathbb{R}^d$. The convex hull $\text{conv}(Q)$ of $Q$ is the set of all points $\sum_{i=1}^m \lambda_i q_i$ with $\sum_{i=1} \lambda_i=1$. The rank of $Q$ is defined by $\text{rank}(Q)=1+\text{dim}(\text{conv}(Q))$. $Q$ is in general position if $\text{rank}(S)=d+1$ for every $(d+1)$-subset $S\subseteq Q$.
Let $G$ be a graph and $X\subset V$. A convex $X$-embedding of $G$ is any mapping $f:V\rightarrow \mathbb{R}^{|X|-1}$ such that for each $v\in V\backslash X$, $f(v)\in \mathrm{conv}(f(N(v)))$. We say that the convex embedding is in general position if the set $f(V)$ of the points is in general position.
\begin{theorem}[Linial, Lov\'{a}z, and Wigderson~\cite{linial1988rubber}]
Let $G$ be a graph on $n$ vertices and $1<k<n$. Then the following two conditions are equivalent:
\begin{enumerate}
\item $G$ is $k$-connected
\item For every $X\subset V$ with $|X|=k$, $G$ has a convex $X$-embedding in general position.
\end{enumerate}
\end{theorem}
Notice that the special case of the Theorem for $k=2$ asserts the existence of an $st$-numbering of a 2-connected graph. The proof of this theorem is inspired by physics. The embedding is found by letting the edges of the graph behave like ideal springs and letting its vertices settle. A formal summary of the proof ($1\rightarrow 2$) is as follows (for more details see~\cite{linial1988rubber}). For each $v_i\in X$, define $f(v_i)$ arbitrary in $\mathbb{R}^{k-1}$ such that $f(X)$ is in general position. Assign to every edge $(u,v)\in E$ a positive elasticity coefficient $c_{uv}$ and let $c\in \mathbb{R}^{|E|}$ be the vector of coefficients. It is proved in~\cite{linial1988rubber} that for almost any coefficient vector $c$, an embedding $f$ that minimizes the potential function $P=\sum_{\{u,v\}\in E} c_{uv}\|f(u)-f(v)\|^2$ provides a convex $X$-embedding in general position ($\|.\|$ is the Euclidean norm). Moreover, the embedding that minimizes $P$ can be computed as follows,
\begin{equation*}
f(v)=\frac{1}{c_v}\sum_{u\in N(v)} c_{uv}f(u)~\text{for all}~v\in V\backslash X,
\end{equation*}
in which $c_v=\sum_{u\in N(v)} c_{uv}$. Hence, the embedding can be found by solving a set of linear equations in at most $O(|V|^3)$ time (or matrix multiplication time).
\section{Missing Proofs from Section~\ref{sec:BCPI}}\label{sec:proof1}
\begin{proof}[Proof of Proposition~\ref{lem:BCPI_hard}]
We use the proof of \cite[Theorem 2]{chataigner2007approximation} with a modest change. The reduction is from the X3C problem~\cite{papadimitriou1982complexity}, which is a variant of the \emph{Exact Cover by 3-sets} and defined as follows: Given a set $X$ with $|X|=3q$ and a family $C$ of 3-element subsets of $X$ such that $|C|=3q$ and each element of $X$ appears in exactly 3 sets of $C$, decide whether $C$ contains an exact cover for $X$.
Given an instance $(X,C)$ of $X3C$, let $G=(V,E)$ be the graph with the vertex set $V=X\cup C\cup\{a,b\}$ and edge set $E=\bigcup_{j=1}^{3q}[\{C_jx_i|x_i\in C_j\}\cup\{C_ja\}\cup\{C_jb\}]$. Set $p(a)=2q$, $p(b)=9q^2+q$, $p(C_j)=-1$, and $p(x_i)=-3q$. It is easy to verify that $C$ contains an exact cover for $X$ if and only if the BCPI problem has a solution such that $p(V_1)=p(V_2)=0$. This shows the first claim.
For the second claim, attach to nodes $a$, $b$, and the $x_i$s, paths of length $2q$, $9q^2+q$, and $3q$, respectively, and set the supply/demand values of $a$, $b$, the $x_i$'s and the new nodes equal to $+1$ (for the paths for $a$ and $b$) or $-1$ (for the $x_i$'s).
\end{proof}
\medskip
\begin{proof}[Proof of Proposition~\ref{lem:2_poly}]
Clearly, part (2) follows immediately from part (1) because in this case, $p(V_1), p(V_2)$
are integer and $\max_{j\in V}|p(j)|/2 =1/2$.
To show part (1), pick two arbitrary nodes $u,v\in V$ with $p(u) p(v)>0$. Since we want to separate $u$ and $v$, we can assume wlog that initially $\{u,v\}\in G$. Since $G$ is 2-connected, an $st$-numbering between nodes $u$ and $v$ as $u=v_1,v_2,\dots,v_n=v$ can be found in $O(|V|+|E|)$~\cite{even1976computing}. Define $V_1^{(i)}:=\{v_1,v_2,\dots,v_i\}$. It is easy to see that $p(V_1^{(1)})=p(u)>0$ and $p(V_1^{(n-1)})=-p(v)<0$. Hence, there must exist an index $1\leq i^*<n$ such that $|p(V_1^{(i^*)})|>0$ and $|p(V_1^{(i^*+1)})|\leq 0$. Since $|p(V_1^{(i)})-p(V_1^{(i+1)})|= |p(i^*+1)|$, either $|p(V_1^{(i^*)})|\leq |p(i^*+1)|/2$ or $|p(V_1^{(i^*+1)})|\leq |p(i^*+1)|/2$; Accordingly set $V_1=V_1^{(i^*)}$ or $V_1=V_1^{(i^*+1)}$. Let $V_2=V\backslash V_1$. Hence, $(V_1,V_2)$ is a solution with $|p(V_1)|=|p(V_2)|\leq |p(i^*+1)/2|\leq \max_{j\in V}|p(j)|/2$. It is easy to see that $i^*$ can be found in $O(|V|)$.
\end{proof}
\medskip
\begin{proof}[Proof of Proposition~\ref{lem:3-connected-zero-p}]
Consider the case of general function $p$, and let $p_{\max} = \max_{j \in V} |p(j)|$.
We will show that we can find a solution such that $u \in V_1, v \in V_2, w \in V_3$ with $|p(V_1)|, |p(V_2)| \leq p_{\max}/2$.
Since $|p(V_3)| = |p(V_1) + p(V_2)|$ (recall $p(V)=0$),
this implies that $|p(V_3)| \leq p_{\max}$, and hence
$|p(V_1)|+ |p(V_2)| +|p(V_3)| \leq 2p_{\max}$.
Furthermore, if $ p(i) = \pm 1$ for all $i \in V$, hence $p_{\max}=1$,
then $|p(V_1)|, |p(V_2)| \leq p_{\max}/2$ implies that
$p(V_1) =p(V_2)=0$, and therefore also $p(V_3) = 0$.
Thus, both claims will follow.
Assume $u,v,w\in V$ and $p(u), p(v),p(w)>0$ (the case of negative $p(u), p(v),p(w)$
is symmetric).
Since we want to separate $u$ from $v$, we can assume without loss of generality that $\{u,v\}\in E$. Using~\cite{cheriyan1988finding}, there is a non-separating ear decomposition through the edge $\{u, v\}$ and avoiding node $w$.
Ignore the ears that do not contain any internal nodes, and let $Q_0\cup Q_1\cup\dots\cup Q_r$ be the decomposition consisting of the ears with nodes; we have $w\in Q_r$. Let $V_i=V(Q_0)\cup V(Q_1)\dots \cup V(Q_i)$, let $G_i=G[V_i]$ and $\overline{G}_i=G[V\backslash V_i]$. We distinguish two cases, depending on whether $p(V_0)\leq0$
or $p(V_0)>0$.
\begin{itemize}
\item[(i)] If $p(V_0)\leq0$, then consider an $st$-numbering between $u$ and $v$ in $V_0$, say $u=v_1,v_2,\dots,v_s=v$. Define $V_0^{(i)}=\{v_1,v_2,\dots,v_i\}$. Since $p(u),p(v)>0$ and $p(V_0)\leq0$, there must exist indices $1\leq i^*\leq j^*<s$ such that $p(V_0^{(i^*)})>0,~p(V_0^{(i^*+1)})\leq0$ and $p(V_0\backslash V_0^{(j^*+1)})>0,~p(V_0\backslash V_0^{(j^*)})\leq0$.
\begin{itemize}
\item[(a)] If $i^*=j^*$, since $p(V_0^{(i^*)})+ p(v_{i^*+1})+p(V_0\backslash V_0^{(i^*+1)})=p(V_0)<0$, we have $ p(V_0^{(i^*)})+p(V_0\backslash V_0^{(i^*+1)})\leq |p(v_{i^*+1})|$.
Now, one of the following three cases happens:
\begin{itemize}
\item[-] If $p(V_0^{(i^*)})\leq |p(v_{i^*+1})|/2$ and $p(V_0\backslash V_0^{(i^*+1)})\leq|p(v_{i^*+1})|/2$, then it is easy to see that $V_1=V_0^{(i^*)}$, $V_2=V_0\backslash V_0^{(i^*+1)}$, and $V_3=V\backslash (V_1\cup V_2)$ is a good partition.
\item[-] If $p(V_0^{(i^*)})> |p(v_{i^*+1})|/2$ and $p(V_0\backslash V_0^{(i^*+1)})\leq|p(v_{i^*+1})|/2$, then $p(V_0^{(i^*)})+p(v_{i^*+1})=p(V_0^{(i^*+1)})\leq |p(v_{i^*+1})|/2$. Hence, $V_1=V_0^{(i^*+1)}$, $V_2=V_0\backslash V_0^{(i^*+1)}$, and $V_3=V\backslash V_0$ is a good partition.
\item[-] If $p(V_0^{(i^*)})\leq |p(v_{i^*+1})|/2$ and $p(V_0\backslash V_0^{(i^*+1)})>|p(v_{i^*+1})|/2$, then $p(V_0\backslash V_0^{(i^*+1)})+p(v_{i^*+1})= p(V_0\backslash V_0^{(i^*)})\leq|p(v_{i^*+1})|/2$. Hence, $V_1=V_0^{(i^*)}$, $V_2=V_0\backslash V_0^{(i^*)}$, and $V_3=V\backslash V_0$ is a good partition.
\end{itemize}
\item[(b)] If $i^*<j^*$, then either $p(V_0^{(i^*)})\leq|p(v_{i^*+1})|/2$ or $|p(V_0^{(i^*+1)})|\leq|p(v_{i^*+1})|/2$, accordingly set $V_1=V_0^{(i^*)}$ or $V_1=V_0^{(i^*+1)}$. Similarly, either $p(V_0\backslash V_0^{(j^*+1)})\leq |p(v_{j^*+1})|/2$ or $|p(V_0\backslash V_0^{(j^*)})|\leq |p(v_{j^*+1})|/2$, so accordingly set $V_2=V_0\backslash V_0^{(j^*+1)}$ or $V_2=V_0\backslash V_0^{(j^*)}$. Set $V_3=V\backslash (V_1\cup V_2)$. It is easy to check that $(V_1,V_2,V_3)$ is a good partition.
\end{itemize}
\item[(ii)] If $p(V_0)>0$, then since $p(w)>0$ and therefore $p(V_{r-1})<0$, there must exist an index $0\leq j<r-1$ such that $p(V_j)>0$ and $p(V_{j+1})\leq0$. Consider an $st$-numbering between $u$ and $v$ in $G[V_j]$ as $u=v_1,v_2,\dots,v_s=v$ and define $V_j^{(i)}=\{v_1,v_2,\dots,v_i\}$. The
ear $Q_{j+1}$ is a path of new nodes $q_1,q_2,\dots, q_t$
attached to two (distinct) nodes $v_x, v_y$ of $G[V_j]$ through
edges $\{v_x,q_1\},\{q_t,v_y\}\in E$; assume wlog that
$1\leq x<y\leq s$.
For simplicity, we will use below $Q_{j+1}$ to denote also the
set $\{q_1,q_2,\dots, q_t\}$ of internal (new) nodes of the ear.
Also define $Q_{j+1}^{(i)}=\{q_1,q_2,\dots,q_i\}$ and $Q_{j+1}^{(0)}=\emptyset$.
One of the following cases must happen:
\begin{itemize}
\item[(a)] Suppose there is an index $1\leq i^*<(y-1)$ such that $p(V_j^{(i^*)})>0$ and $p(V_j^{(i^*+1)})\leq0$ or there is an index $x+1<i^*< s$ such that $p(V_j\backslash V_j^{(i^*-1)})>0$ and $p(V_j\backslash V_j^{(i^*)})\leq0$.
Let's assume there is an index $1\leq i^*< (y-1)$, such that $p(V_j^{(i^*)})>0$ and $p(V_j^{(i^*+1)})\leq0$ (the other case is exactly similar). Then either $p(V_j^{(i^*)})\leq|p(v_{i^*+1})|/2$ or $|p(V_j^{(i^*+1)})|\leq|p(v_{i^*+1})|/2$, accordingly set either $V_1=V_j^{(i^*)}$ or $V_1=V_j^{(i^*+1)}$. Set $V_2'=V_j\backslash V_1$. One of the following cases happens:
\begin{itemize}
\item[-] If $V_1=V_j^{(i^*)}$ and $p(V_2')\leq 0$, then since $p(V_j^{(i^*+1)})\leq0$, we have $p(V_j\backslash V_j^{(i^*+1)})>0$. Hence, $p(V_2'\backslash \{v_{i^*+1}\})>0$. So, it is either $|p(V_2')|\leq |p(v_{i^*+1})|/2$ or $p(V_2'\backslash \{v_{i^*+1}\})\leq |p(v_{i^*+1})|/2$. Now if $p(V_2'\backslash \{v_{i^*+1}\})\leq |p(v_{i^*+1})|/2$, since also $p(V_1)\leq |p(v_{i^*+1})|/2$, $p(V_j)\leq 0$ which contradicts with the assumption. Therefore, $|p(V_2')|\leq |p(v_{i^*+1})|/2$. Set $V_2=V_2'$ and $V_3=V\backslash (V_1\cup V_2)$. It is easy to check that $(V_1,V_2,V_3)$ is a good partition.
\item[-] If $V_1=V_j^{(i^*)}$ and $p(V_2')> 0$, then since $p(V_j\cup Q_{j+1})<0$, there is an index $0< t^*\leq t$, such that $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*)}))>0$ and $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*-1)}))\leq 0$. Hence, either $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*)}))\leq |p(q_{t^*})|/2$ or $|p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*-1)}))|\leq |p(q_{t^*})|/2$, accordingly set $V_2=V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*)})$ or $V_2=V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(t^*-1)})$. Set $V_3=V\backslash (V_1\cup V_2)$. It is easy to see that $(V_1,V_2,V_3)$ is a good partition.
\item[-] If $V_1=V_j^{(i^*+1)}$, then since $p(V_1)\leq 0$, we have $p(V_2')>0$. The rest is exactly like the previous case when $V_1=V_j^{(i^*)}$ and $p(V_2')> 0$.
\end{itemize}
\item[(b)] Suppose that for every $1\leq i<y$, $p(V_j^{(i)})>0$ and for every $x<i< s$, $p(V_j\backslash V_j^{(i)})>0$. Set $V_1'=V_j^{(y-1)}$ and $V_2'=V_j\backslash V_1'$. Based on the assumption $p(V_1'),p(V_2')>0$. Since $p(V_{j+1})\leq 0$, there are indices $0\leq i^*\leq j^*<t$ such that $p(V_1'\cup Q_{j+1}^{(i^*)})>0$, $p(V_1'\cup Q_{j+1}^{(i^*+1)})\leq0$ and $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(j^*+1)}))>0$, $p(V_2'\cup (Q_{j+1}\backslash Q_{j+1}^{(j^*)}))\leq0$. The rest of the proof is similar to case (i) when $p(V_0)\leq 0$.
\end{itemize}
\end{itemize}
\end{proof}
\section{Missing Proofs from Section~\ref{sec:BPGI}}\label{sec:proof2}
\begin{proof}[Proof of Lemma~\ref{lem:embedding}]
Set $X=\{v,u,w\}$. Using~\cite{linial1988rubber}, $G$ has a convex $X$-embedding in $\mathbb{R}^2$ in general position with mapping $f:V\rightarrow \mathbb{R}^2$ such that $f(u)=(0,0)$, $f(v)=(1,0)$, and $f(w)=(0,1)$. In the $X$-embedding of the nodes, we have a freedom to set the elasticity coefficient vector $\vec{c}$ to anything that we want (except a measure zero set of vectors). So for any edge $\{i,j\}\in G[V_1']\cup G[V_2']$, set $c_{ij}=g$; and for any $\{i,j\} \in E[V_1',V_2']$, set $c_{ij}=1$.
Assume $\mathcal{L}_1$ is the line $y=0.5$, $\mathcal{L}_2$ is the line $x+2y=1$, and $\mathcal{L}_3$ is the line $x=y$.
First, we show that there exist a $g$ for which all the nodes in $V_1'$ will be embedded above the line $\mathcal{L}_1$. To show this, from~\cite{linial1988rubber}, we know the embedding is such that it minimizes the total potential $P(f,\vec{c})= \sum_{\{i,j\}\in E} c_{ij}\|f(i)-f(j)\|^2$. Notice that we can independently minimize $P$ on $x$-axis values and $y$-axis values as below:
\begin{eqnarray*}
\min_{f} P &=& \min_{f_1} P_x +\min_{f_2} P_y\\
&=&\min_{f_1} \sum_{\{i,j\}\in E} c_{ij} (f_1(i)-f_1(j))^2+ \min_{f_2} \sum_{\{i,j\}\in E} c_{ij} (f_2(i)-f_2(j))^2
\end{eqnarray*}
Now, notice that if we place all the nodes in $V_1'$ at point (0,1) and all the nodes in $V_2'$ on the line $uv$, then $P_y\leq |E|$. Hence, if $f_2$ minimizes $P_y$, then $P_y(f_2,c)\leq |E|$. Set $g \geq 4|V|^2 |E|$. We show that if $f_2$ minimizes $P_y$, then for all edges $\{i,j\}\in G[V_1']\cup G[V_2']$, $(f_2(i)-f_2(j))^2\leq1/(4|V|^2)$. By contradiction, assume there is an edge $\{i,j\}\in G[V_1']\cup G[V_2']$ such that $(f_2(i)-f_2(j))^2>1/(4|V|^2)$. Then, $c_{ij}(f_2(i)-f_2(j))^2=g (f_2(i)-f_2(j))^2 >|E|$. Hence, $P_y(f_2,c)> |E|$ which contradicts with the fact the $f_2$ minimizes $P_y$. Therefore, if $g\geq4|V|^2 |E|$, then for all $\{i,j\}\in G[V_1']\cup G[V_2']$, $|f_2(i)-f_2(j)|\leq1/(2|V|)$. Now, since $G[V_1']$ is connected, all the nodes in $V_1'$ are connected to $w$ with a path of length (in number of hops) less than $|V|-1$. Hence, using the triangle inequality, for all $i\in V_1'$:
\begin{equation*}
|f_2(w)-f_2(i)|\leq (|V|-1)/(2|V|)<1/2\Rightarrow |1-f_2(i)|< 1/2,
\end{equation*}
which means that all the nodes in $V_1'$ are above $\mathcal{L}_1$.
With the very same argument, if $g\geq t^2 |V|^2 |E|$, then for all $i\in V_2'$, $f_2(i)<1/t$.
Now, we want to prove that there is a $g$ such that all the nodes in $V_2'$ will be embedded below the lines $\mathcal{L}_2$ and $\mathcal{L}_3$. Define $n_1(i):=|N(i)\cap V_1'|$ and $n_2(i):=|N(i)\cap V_2'|$. From~\cite{linial1988rubber}, we know the embedding is such that for all $i\in V\backslash\{u,v,w\}$, $f(i)=1/c_i\sum_{j\in N(i)}c_{ij}f(j)$, where $c_j=\sum_{j\in N(i)}c_{ij}f(j)$. Since $G[V'_2]$ is connected, for any $i\in V_2'$ there is a path $i=v_1,v_2,\dots,v_r=v$ in $V_2'$. Using this ordering:
\begin{eqnarray*}
&&\begin{cases}
f_1(v_j)\geq \frac{1}{n_2(v_j)g+n_1(v_j)}g f_1(v_{j+1})\geq (1/|V|)f_1(v_{j+1}),& \forall j\in \{1,\dots,r-1\}\\
f_1(v_r)=f_1(v)=1
\end{cases}
\\&&\Rightarrow \forall i\in V_2'\backslash\{u,v\}:~ f_1(i)\geq (1/|V|)^{r}\geq(1/|V|)^{|V|}.
\end{eqnarray*}
On the other hand, from the previous part, if we set $g\geq |V|^{2|V|+2}|E|$, then for all $i\in V_2'$, $f_2(i)\leq (1/|V|)^{|V|}$. Hence, for all $i\in V_2'$, $f_2(i)\leq f_1(i)$, which means that all the nodes in $V_2'$ will be placed below the line $\mathcal{L}_3$.
With the very same idea, we show that there exist a $g$ for which all the nodes in $V_2'$ will be placed below the line $\mathcal{L}_2$. Since $G[V'_2]$ is connected, for any $i\in V_2'$ there is a path $u=u_1,v_2,\dots,u_t=i$ in $V_2'$. Notice that for all $i\in V\backslash\{u,v,w\}$, $1-f_1(i)=1/c_i\sum_{j\in N(i)}c_{ij}(1-f_1(j))$. Hence, since $\forall j\in V: f_1(j)\leq 1$, we have,
\begin{eqnarray*}
&&\begin{cases}
1-f_1(u_j)\geq \frac{1}{n_2(u_j)g+n_1(u_j)}g (1-f_1(u_{i-1}))\geq (1/|V|)(1-f_1(u_{i-1})),& \forall j\in \{2,\dots,t\}\\
1-f_1(u)=1-f_1(u_1)=1
\end{cases}
\\&& \Rightarrow \forall i\in V_2'\backslash\{u,v\}:~1-f_1(i)\geq (1/|V|)^{t}\geq (1/|V|)^{|V|}.
\end{eqnarray*}
From the previous part, if we set $g\geq 4 |V|^{2|V|+2}|E|$, then for all $i\in V_2'$, $f_2(i)\leq 1/2(1/|V|)^{|V|}$. Hence, for $i\in V_2'$, $f_1(i)+2f_2(i)\leq 1$, which means that all the nodes in $V_2'$ will be placed below the line $\mathcal{L}_3$. Therefore, if we set $g\geq 4|V|^{2|V|+2}|E|$, then we will get an embedding as depicted in Fig.~\ref{fig:3-connected}.
Note that a polynomial number of bits suffices for $g$
Notice that if $\vec{c}$ is a ``good" vector, then so is $\vec{c}+\vec{\epsilon}$ in which $\vec{\epsilon}$ is a vector with very small Euclidean norm. Hence, we can always find a ``good" vector $\vec{c}$ which result in a $X$-embedding in general position.
\end{proof}
\medskip
\begin{proof}[Proof of Theorem~\ref{th:3-connected}]
Assume that $|V|\equiv0(\mathrm{mod}~4)$; the proof for the case $|V|\equiv2(\mathrm{mod}~4)$ is similar. Using Lemma~\ref{lem:cut}, we can find $\{u,v,w\}\in V$ and a partition $(V_1',V_2')$ of $V$ with properties described in the Lemma. Set $X=\{u,v,w\}$. Using Lemma~\ref{lem:embedding}, we can find a convex $X$-embedding of $G$ in general position with properties described in the Lemma as depicted in Fig.~\ref{fig:3-connected}.
The rest of the proof is very similar to the proof of Lemma~\ref{lem:3-connected-tri}. We consider again a circle $\mathcal{C}$ around $f(u),f(v),f(w)$ in $\mathbb{R}^2$ as shown in Fig.~\ref{fig:3-connected}. Also consider a directed line $\mathcal{L}$ tangent to the circle $C$ at point $A$. If we project the nodes of $G$ onto the line $\mathcal{L}$, this time the order of the nodes projection gives an $st$-numbering between the first and the last node only if $u$ and $v$ are the first and last node. However, if we set $V_1$ to be the $|V|/2$ nodes whose projections come first and $V_2$ are the $|V|/2$ nodes whose projections come last, then $G[V_1]$ and $G[V_2]$ are both connected even when $u$ and $v$ are not the first and last nodes. The reason lies on the special embedding that we considered here. Assume for example $w$ and $v$ are the first and the last projected nodes, and $V_1$ and $V_2$ are set of the $|V|/2$ nodes which projections come first and last, respectively. Two cases might happen:
\begin{itemize}
\item[(i)] If $u,w\in V_1$ and $v\in V_2$, then since $\{u,w\}\in E$, both $G[V_1]$ and $G[V_2]$ are connected because of the properties of the embedding.
\item[(ii)] If $w\in V_1$ and $u,v\in V_2$, since $|V_2'|\leq |V|/2$ and $|V_2|=|V|/2$, then either $V_2=V_2'$ or $V_2 \cap V_1' \neq \emptyset$.
If $V_2= V_2'$, and hence $V_1=V_1'$ then there is nothing to prove.
So assume there is a node $z\in V_2 \cap V_1'$.
From the properties of the embedding, the triangle $\{z,u,v\}$
contains all the nodes of $V_2'$.
Since $\{z,u,v\}\in V_2$, and $V_2$ contains all the nodes that are on a same side of a halfplane, we should also have $V_2'\subset V_2$. Now, from the properties of the embedding, it is easy to see that every node in $V_2$ has a path either to $u$ or $v$. Since $V_2'\subset V_2$, there is also a path between $u$ and $v$. Thus, $G[V_2]$ is connected. From the properties of the embedding, $G[V_1]$ is connected as before.
\end{itemize}
The rest of the proof is exactly the same as the proof of Lemma~\ref{lem:3-connected-tri}. We move $\mathcal{L}$ from being tangent at point $A$ to point $B$ ($AB$ is a diameter of the circle $\mathcal{C}$) and consider the resulting partition. Notice that if at point $A$, $p(V_1)>0$, then at point $B$ since $V_1$ and $V_2$ completely switch places compared to the partition at point $A$, $p(V_1)<0$. Hence, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ and keep it tangent to the circle, in the resulted partitions, $p(V_1)$ goes from some positive value to a negative value. Notice that the partition $(V_1,V_2)$ changes only if $\mathcal{L}$ passes a point $D$ on the circle such that at $D$, $\mathcal{L}$ is perpendicular to a line that connects $f(i)$ to $f(j)$ for $i,j\in V$. Now, since the embedding is in general position, there are exactly two points on every line that connects two points $f(i)$ and $f(j)$, so $V_1$ changes at most by one node leaving $V_1$ and one node entering $V_1$. Hence, $p(V_1)$ changes by either $\pm 2$ or $0$ value at each change. Now, since $|V|\equiv 0 (\mathrm{mod}~ 4)$, $p(V_1)$ has an even value in all the resulting partitions. Therefore, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$, there should be a point $D$ such that in the resulted partition $p(V_1)=p(V_2)=0$.
\end{proof}
\medskip
\begin{proof}[Proof of Corollary~\ref{cor:serpar}]
Every series-parallel graph $G$ has a separation pair $\{u,v\}$
such that every connected component of $G[V\backslash \{u,v\}]$ has less that $2|V|/3$ nodes,
and furthermore, such a separation pair can be found in linear time.
To see this, consider the derivation tree $T$ of the construction of $G$.
The root of $T$ corresponds to $G$, the leaves correspond to the edges, and every internal node
$i$ corresponds to a subgraph $G_i=(V_i,E_i)$ that is the series or parallel composition of the
subgraphs corresponding to its children.
Starting at the root of $T$, walk down the tree following always the edge to the child corresponding to
a subgraph with the maximum number of nodes until the number of nodes becomes $\leq 2|V|/3$.
Thus, we arrive at a node $i$ of the tree such that $|V_i| > 2|V|/3$
and $|V_j| \leq 2|V|/3$ for all children $j$ of $i$.
Let $u_i, v_i$ be the terminals of $G_i$. Note that $u_i, v_i$ separate all
the nodes of $G_i$ from all the nodes that are not in $G_i$.
Since $|V_i| > 2|V|/3$, we have $|V \backslash V_i| < |V|/3$.
If $G_i$ is the parallel composition of the graphs corresponding to the children of $i$,
then the separation pair $\{u_i, v_i\}$ has the desired property,
i.e. all the components of $G[V\backslash \{u,v\}]$ have less than $2|V|/3$ nodes.
Suppose $G_i$ is the series composition of the graphs $G_j$, $G_k$
corresponding to the children $j,k$ of $i$, and let $w$ be the common
terminal of $G_j$, $G_k$; thus, $G_i$ has terminals $u_i, w$,
and $G_k$ has terminals $w, v_i$.
Assume wlog that $|V_j| \geq |V_k|$.
Then $|V|/3 < |V_j| \leq 2|V|/3$.
The pair $\{u_i,w\}$ of terminals of $G_j$ separates all the nodes
of $V_j \backslash \{u_i,w\}$ from all the nodes of $V \backslash V_j$,
and both these sets have less than $2|V|/3$ nodes.
Thus, $\{u_i,w\}$ has the required property.
\end{proof}
\medskip
\begin{proof}[Proof of Theorem~\ref{th:2-connected}]
Using Lemma~\ref{lem:2-connected-pair} for $q=4$, we consider two cases:
\begin{itemize}
\item[(i)] There is a separation pair $\{u,v\}\in V$ such that if $G_1,\dots,G_k$ are the connected components of $G\backslash \{u,v\}$, for any $i$, $|V_i|< 3|V|/4$. In this case Lemma~\ref{lem:parallel} for $q=4$ proves the theorem.
\item[(ii)] After a set of contractions, $G$ can be transformed into a 3-connected graph $G^*=(V^*,E^*)$ with weighted edges such that for any edge $e^*\in E^*$, $w(e^*)<|V|/4$. In this case the proof is similar to the proof of Theorem~\ref{th:3-connected}. Notice that if $G^*$ contains a triangle then the proof is much simpler as in the proof of Lemma~\ref{lem:3-connected-tri} but here to avoid repetition, we use the approach in the proof of Theorem~\ref{th:3-connected} and prove the theorem once for all cases of $G^*$.
Recall that for each edge in $G^*$, its weight represents the number of nodes in the contracted subgraph that it represents. So if $e^*\in E^*$ represents an induced subgraph of $G$ with $l$ nodes, then $w(e^*)=l$; and if $e^*\in E\cap E^*$ then $w(e^*)=0$. Using Lemma~\ref{lem:cut}, we can find $\{u,v,w\}\in V^*$ and a partition $(V_1^*,V_2^*)$ of $V^*$ with properties described in the Lemma. Since in $G^*$ edges have weights that actually represent nodes in $G$, we want a partition such that we also have $|V^*_2|+\sum_{e\in G[V_2^*]}w(e)\leq |V|/2$. Notice that we can modify the proof of Lemma~\ref{lem:cut} to take into account the weights for the edges and find a partition such that $|V^*_2|+\sum_{e\in G[V_2^*]}w(e)\leq |V|/2$ as follows. Again, using the algorithm presented in~\cite{cheriyan1988finding}, we can find a non-separating proper cycle $C_0$ in $G^*$ such that every node in $C_0$ has a neighbor in $G^*\backslash C_0$. We consider three cases:
\begin{itemize}
\item[(a)] If $|C_0|+\sum_{e\in G[C_0]}w(e)\leq |V|/2+1$, then the proof is as in the proof of Lemma~\ref{lem:cut}. Select any three consecutive nodes $(u,w,v)$ of $C_0$ and set $V_2^*=C_0\backslash \{w\}$ and $V_1^*=V^*\backslash V_2^*$.
\item[(b)] If $|C_0|>|V^*|/2$ and $|C_0|+\sum_{e\in G[C_0]}w(e)> |V|/2+1$, then as in the proof of Lemma~\ref{lem:cut}, since $|C_0|>|V^*\backslash C_0|$ and every node is $C_0$ has a neighbor in $V^*\backslash C_0$, there exist a node $w\in V^*\backslash C_0$ such that $|N(w)\cap C_0|\geq 2$. Select two nodes $u,v\in N(w)\cap C_0$. There exists a path $P$ in $C_0$ between $u$ and $v$ such that $|P|+\sum_{e\in G[P]}w(e)<|V|/2-1$. Set $V_2^*=\{u,v\}\cup P$ and $V_1^*=V^*\backslash V_2^*$.
\item[(c)] If $|C_0|\leq |V^*|/2$ and $|C_0|+\sum_{e\in G[C_0]}w(e)> |V|/2+1$, then if there exist a node $w\in V^*\backslash C_0$ such that $|N(w)\cap C_0|\geq 2$, the proof is as in the previous part. So assume for all $w\in V^*\backslash C_0$, $|N(w)\cap C_0|\leq 1$. Assume $w_1,w_2,\dots,w_t\in V^*\backslash C_0$ are all the nodes with $|N(w_i)\cap C_0|=1$. We show that there is a $1\leq i\leq t$, such that $w_i$ is not a cut-point of $G^*[V^*\backslash C_0]$. Let $T$ be a spanning tree of $G^*[V^*\backslash C_0]$. If there is a $1\leq i\leq t$, such that $w_i$ is a leaf of $T$, then $w_i$ is not a cut-point of $G^*[V^*\backslash C_0]$ and there is nothing left to prove. So assume none of $w_i$s is a leaf of $T$. Suppose $1\leq p,q\leq t$ are such that the path between $w_p$ and $w_q$ in $T$ is the longest between all pairs of $w_i$s. We show that $w_p$ and $w_q$ cannot be cut-points of $G^*[V^*\backslash C_0]$. Assume for example $w_p$ is a cut-point of $G^*[V^*\backslash C_0]$ and $G_{\bar{q}}^*$ is a connected component of $G^*[V^*\backslash C_0]\backslash\{w_p\}$ such that $w_q\notin G_{\bar{q}}^*$. Since the path between $w_p$ and $w_q$ in $T$ is the longest between all pairs of $w_i$s, $\forall 1\leq i\leq t: w_i\notin G_{\bar{q}}^*$, otherwise we can find a longer path from $w_q$ to some other $w_i$. On the other hand, if $\forall 1\leq i\leq t: w_i\notin G_{\bar{q}}$, in $G^*\backslash \{w_p\}$ the cycle $C_0$ is disconnected from $G_{\bar{q}}^*$ which contradicts with the 3-connectedness of $G^*$. Hence, there should be at least a noncut-point $w\in V^*\backslash C_0$ with $|N(w)\cap C_0|=1$. Since $G^*$ is 3-connected, each node has degree at least 3. Hence, there are two nodes $u,v\in N(w)\cap (V^*\backslash C_0)$. Set $V_1^*=C_0\cup\{w\}$ and $V_2^*=V^*\backslash V_1^*$. $(V_1^*,V_2^*)$ is a desirable partition for $V^*$.
\end{itemize}
Hence, we can find $\{u,v,w\}\in V^*$ and a partition $(V_1^*,V_2^*)$ of $V^*$ with properties described in the Lemma~\ref{lem:cut} as well as having $|V^*_2|+\sum_{e\in G[V_2^*]}w(e)\leq |V|/2$. Set $X=\{u,v,w\}$. Using Lemma~\ref{lem:embedding}, $G^*$ has a convex $X$-embedding in general position like $f^*:V^*\rightarrow \mathbb{R}^2$ as described in the lemma and depicted in Fig.~\ref{fig:3-connected}. Now, from this embedding, we get a convex $X$-embedding for $G$ like $f:V\rightarrow \mathbb{R}^2$ as follows. For any $i\in V\cap V^*$, $f(i)=f^*(i)$. For any edge $\{i,j\}\in E^*$ such that $\{i,j\}$ represents an induced subgraph of $G$, we represent it by a pseudo-path between $i$ and $j$ in $G$ like $P$ and place the nodes in $P$ in order on random places on the line segment that connects $f(i)$ to $f(j)$. If the edge $\{i,j\}\in E^*$ is between a node in $V_1^*$ and a node in $V_2^*$ and represents a pseudo-path $P$ in $G$ , we place the nodes in $P$ in order on random places on the segment that connects $f(i)$ to $f(j)$ but above the line $\mathcal{L}_1$. Hence, by this process, we get a convex $X$-embedding for $G$ which is in general position (almost surely) except for the nodes that are part of a pseudo-path (which we know have length less than $|V|/4$). Notice that if $V_2'\subset V$ contains all the nodes in $V_2^*$ and all the nodes that are part of a pseudo-path between the nodes of $V_2^*$ (represented by weighted edges in $G^*$) in $G$, and $V_1'=V\backslash V_2'$, then $(V_1',V_2')$ is partition of $G$ with all the properties of Lemma~\ref{lem:cut}. Moreover, the $X$-embedding $f$ of $G$ that we derived from $f^*$ for $G^*$, has all the properties of Lemma~\ref{lem:embedding} for the partition $(V_1',V_2')$.
The rest of the proof is similar to the proof of Theorem~\ref{th:3-connected}. We consider again a circle $\mathcal{C}$ around $f(u),f(v),f(w)$ in $\mathbb{R}^2$ as shown in Fig.~\ref{fig:3-connected}. Also consider a directed line $\mathcal{L}$ tangent to the circle $C$ at point $A$ and project nodes of $G$ onto the line $\mathcal{L}$. With the same argument as in the proof of Theorem~\ref{th:3-connected}, since the embedding $f$ has the properties in Lemma~\ref{lem:cut} and \ref{lem:embedding}, if we set $V_1$ to be the $|V|/2$ nodes whose projections come first and $V_2$ are the $|V|/2$ nodes whose projections come last, then $G[V_1]$ and $G[V_2]$ are both connected. If $|p(V_1)|\leq 1$, then $(V_1,V_2)$ is a good partition and there is nothing left to prove. Otherwise, we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ ($AB$ is a diameter of the circle $\mathcal{C}$) and consider the resulting partition. Notice that if at point $A$, $p(V_1)>0$, then at point $B$ since $V_1$ and $V_2$ completely switch places compared to the partition at point $A$, $p(V_1)<0$. Hence, as we move $\mathcal{L}$ from being tangent at point $A$ to point $B$ and keep it tangent to the circle, in the resulting partitions, $p(V_1)$ goes from some positive value to a negative value. Notice that the partition $(V_1,V_2)$ changes only if $\mathcal{L}$ passes a point $D$ on the circle such that at $D$, $\mathcal{L}$ is perpendicular to a line that connects $f(i)$ to $f(j)$ for $i,j\in V$. Now, since the embedding is in general position except for few lines that contain a pseudo-path, if $\{i,j\}\notin E\cup E^*$ or $\{i,j\}\in E\cap E^*$ there are exactly two points on the line that connects $f(i)$ and $f(j)$, so $V_1$ changes at most by one node leaving $V_1$ and one node entering $V_1$. Thus, in this case $p(V_1)$ changes by at most 2. However, if $\{i,j\}\in E^*$, then since there can be at most $|V|/4-1$ points on the line that connects $f(i)$ to $f(j)$, $p(V_1)$ may change by $|V|/4-1$. So let's consider that $\mathcal{L}$ is perpendicular to the line that connects $f(i)$ and $f(j)$ at point $D$ and $p(V_1)>0$ slightly before $\mathcal{L}$ passes point $D$ and $p(V_1)\leq 0$ slightly after $\mathcal{L}$ passes point $D$. We consider 2 cases:
\begin{itemize}
\item[(a)] Suppose $\{i,j\}\notin E\cup E^*$ or $\{i,j\}\in E\cap E^*$. In this case since $V_1$ changes by at most $2$, therefore, $p(V_1)$ slightly after $\mathcal{L}$ passes $D$ is either $0$ or $-1$. Hence, the partition $(V_1,V_2)$ that we get after $\mathcal{L}$ passes $D$ is a good partition with $|V_1|=|V_2|=|V|/2$.
\item[(b)] Suppose $\{i,j\}\in E^*$. Let $(V_{b1},V_{b2})$ be the partition slightly before $\mathcal{L}$ passes $D$ and let $(V_{a1},V_{a2})$ be the partition slightly after. Assume $P_0$ is the pseudo-path between $i$ and $j$ and $P=P_0\cup\{i,j\}$. It is easy to see that $(V_{b1}\backslash P,V_{b2}\backslash P)=(V_{a1}\backslash P,V_{a2}\backslash P)$.
If $p(V_{b1}\backslash P)=0$, since $|P_0|<|V|/4$, then $V_1=V_{b1}\backslash P$ and $V_2= V\backslash V_1$ is a good partition and there is nothing left to prove. So either $p(V_{b1}\backslash P)> 0$ or $p(V_{b1}\backslash P)< 0$.
If $p(V_{b1}\backslash P)> 0$, since $p(V_{a1})\leq 0$ and $V_{b1}\backslash P=V_{a1}\backslash P$, we can add a set of nodes from $P\cap V_{a1}$, say $P'$, to $V_{a1}\backslash P$ to get $p((V_{a1}\backslash P)\cup P')=0$. Now since $|P_0|<|V|/4$, $V_1=(V_{a1}\backslash P)\cup P'$ and $V_2=V\backslash V_1$ is a good partition.
If $p(V_{b1}\backslash P)< 0$, since $p(V_{b1})> 0$, we can add a set of nodes from $P\cap V_{b1}$, say $P'$, to $V_{b1}\backslash P$ to get $p((V_{b1}\backslash P)\cup P')=0$. Now since $|P_0|<|V|/4$, $V_1=(V_{b1}\backslash P)\cup P'$ and $V_2=V\backslash V_1$ is a good partition.
\end{itemize}
\end{itemize}
\end{proof}
\medskip
\section{Missing Proofs from Section~\ref{sec:blue_red}}\label{sec:proof3}
\begin{proof}[Proof of Corollary~\ref{cor:blue_red3}]
Suppose without loss of generality that $n_r \geq n_b$ and
let $n_r-n_b=2t$ and $n_r+ n_b =n =2m$.
Set $p(i)=1$ for $i \in R$ and $p(i)=-1$ for $i \in B$.
Then $p(V) =2t$.
From the equations, we have $n_r=m+t$ and $n_b= m-t$.
From Corollary ~\ref{cor:3-general} we can find a partition $(V_1,V_2)$
such that $|V_1|=|V_2|$ and $|p(V_1)-p(V)/2|, |p(V_1)-p(V)/2| \leq 1$.
Let $r_1 = |R \cap V_1|$ and $b_1 = |B \cap V_1|$.
We have $r_1 + b_1 = n/2 =m$ and $t-1 \leq r_1 - b_1 \leq t+1$.
Therefore, $(m+t)/2 - (1/2) \leq r_1 \leq (m+t)/2 +(1/2)$.
Since $r_1$ is an integer and $n_r=m+t$ is even,
it follows that $r_1 = (m+t)/2 = n_r/2$.
Hence, $b_1 = (m-t)/2 = n_b/2$.
Therefore, $V_2$ also contains $n_r/2$ red nodes and
$n_b/2$ blue nodes.
\end{proof}
|
\section*{Acknowledgements}
\noindent We express our gratitude to our colleagues in the CERN
accelerator departments for the excellent performance of the LHC. We
thank the technical and administrative staff at the LHCb
institutes. We acknowledge support from CERN and from the national
agencies: CAPES, CNPq, FAPERJ and FINEP (Brazil); NSFC (China);
CNRS/IN2P3 (France); BMBF, DFG and MPG (Germany); INFN (Italy);
FOM and NWO (The Netherlands); MNiSW and NCN (Poland); MEN/IFA (Romania);
MinES and FASO (Russia); MinECo (Spain); SNSF and SER (Switzerland);
NASU (Ukraine); STFC (United Kingdom); NSF (USA).
We acknowledge the computing resources that are provided by CERN, IN2P3 (France), KIT and DESY (Germany), INFN (Italy), SURF (The Netherlands), PIC (Spain), GridPP (United Kingdom), RRCKI and Yandex LLC (Russia), CSCS (Switzerland), IFIN-HH (Romania), CBPF (Brazil), PL-GRID (Poland) and OSC (USA). We are indebted to the communities behind the multiple open
source software packages on which we depend.
Individual groups or members have received support from AvH Foundation (Germany),
EPLANET, Marie Sk\l{}odowska-Curie Actions and ERC (European Union),
Conseil G\'{e}n\'{e}ral de Haute-Savoie, Labex ENIGMASS and OCEVU,
R\'{e}gion Auvergne (France), RFBR and Yandex LLC (Russia), GVA, XuntaGal and GENCAT (Spain), Herchel Smith Fund, The Royal Society, Royal Commission for the Exhibition of 1851 and the Leverhulme Trust (United Kingdom).
\section*{Tabulated results and correlation matrices}
\label{appendix_results}
The FSR corrections used in this analysis are given in Tables~\ref{tab:fsr1},~\ref{tab:fsr3}, and~\ref{tab:fsr2}. The bins are indexed in increasing rapidity, \ensuremath{\phi^{*}_\eta}\xspace and transverse momentum, and the same binning schemes as in Refs.~\cite{LHCb-PAPER-2015-001,LHCb-PAPER-2015-003,LHCb-PAPER-2015-049} are used. The bin index scheme defined in Tables~\ref{tab:fsr1},~\ref{tab:fsr3}, and~\ref{tab:fsr2} is used throughout the appendix. The differential cross-section results are tabulated in Tables~\ref{tab:rapidity},~\ref{tab:phist} and~\ref{tab:pt}. The correlation matrices are given in Tables~\ref{tab:corr_rap},~\ref{tab:corr_rap_electron},~\ref{tab:corr_ps_mu},~\ref{tab:corr_ps_e}, and~\ref{tab:corr_pt}.
\begin{table}[!h]
\begin{center}
\caption{The FSR correction applied as a function of the boson rapidity.\label{tab:fsr1}}
\begin{tabular}{cc|cc}
Bin index & Bin range& $f_\text{FSR}^{\Pmu\Pmu}$ & $f_\text{FSR}^{\Pe\Pe}$\\\hline
1&2.000-2.125&1.016$\pm$0.005&1.034$\pm$0.003\\
2&2.125-2.250&1.017$\pm$0.004&1.037$\pm$0.005\\
3&2.250-2.375&1.021$\pm$0.002&1.040$\pm$0.002\\
4&2.375-2.500&1.018$\pm$0.002&1.041$\pm$0.002\\
5&2.500-2.625&1.023$\pm$0.003&1.043$\pm$0.002\\
6&2.625-2.750&1.022$\pm$0.003&1.044$\pm$0.004\\
7&2.750-2.875&1.022$\pm$0.002&1.047$\pm$0.004\\
8&2.875-3.000&1.023$\pm$0.003&1.048$\pm$0.002\\
9&3.000-3.125&1.026$\pm$0.002&1.051$\pm$0.002\\
10&3.125-3.250&1.026$\pm$0.002&1.051$\pm$0.002\\
11&3.250-3.375&1.025$\pm$0.004&1.055$\pm$0.001\\
12&3.375-3.500&1.026$\pm$0.005&1.053$\pm$0.003\\
13&3.500-3.625&1.027$\pm$0.002&1.049$\pm$0.005\\
14&3.625-3.750&1.024$\pm$0.002&1.051$\pm$0.007\\
15&3.750-3.875&1.021$\pm$0.003&1.045$\pm$0.004\\
16&3.875-4.000&1.019$\pm$0.019&1.038$\pm$0.011\\
17&4.000-4.250&1.034$\pm$0.014&1.061$\pm$0.013\\
18&4.250-4.500&1.046$\pm$0.119&
\end{tabular}
\end{center}
\end{table}
\begin{table}[!t]
\begin{center}
\caption{The FSR correction applied as a function of \ensuremath{\phi^{*}_\eta}\xspace.\label{tab:fsr3}}
\begin{tabular}{cc|cc}
Bin index &Bin range& $f_\text{FSR}^{\Pmu\Pmu}$ & $f_\text{FSR}^{\Pe\Pe}$\\\hline
1&0.00-0.01&1.034$\pm$0.002&1.057$\pm$0.002\\
2&0.01-0.02&1.035$\pm$0.002&1.057$\pm$0.001\\
3&0.02-0.03&1.028$\pm$0.001&1.054$\pm$0.001\\
4&0.03-0.05&1.027$\pm$0.002&1.050$\pm$0.002\\
5&0.05-0.07&1.022$\pm$0.002&1.048$\pm$0.001\\
6&0.07-0.10&1.018$\pm$0.003&1.041$\pm$0.002\\
7&0.10-0.15&1.015$\pm$0.004&1.040$\pm$0.004\\
8&0.15-0.20&1.016$\pm$0.001&1.038$\pm$0.003\\
9&0.20-0.30&1.012$\pm$0.003&1.039$\pm$0.002\\
10&0.30-0.40&1.014$\pm$0.003&1.042$\pm$0.003\\
11&0.40-0.60&1.017$\pm$0.005&1.042$\pm$0.002\\
12&0.60-0.80&1.021$\pm$0.004&1.044$\pm$0.007\\
13&0.80-1.20&1.027$\pm$0.010&1.044$\pm$0.004\\
14&1.20-2.00&1.028$\pm$0.008&1.048$\pm$0.007\\
15&2.00-4.00&1.002$\pm$0.041&1.080$\pm$0.023
\end{tabular}
\end{center}
\end{table}
\begin{table}[!t]
\begin{center}
\caption{The FSR correction applied as a function of the boson transverse momentum.\label{tab:fsr2}}
\begin{tabular}{cc|c}
Bin index & Bin range [GeV] &$f_\text{FSR}^{\Pmu\Pmu}$\\\hline
1&0.0-2.2&1.090$\pm$0.004\\
2&2.2-3.4&1.075$\pm$0.002\\
3&3.4-4.6&1.062$\pm$0.003\\
4&4.6-5.8&1.045$\pm$0.003\\
5&5.8-7.2&1.029$\pm$0.001\\
6&7.2-8.7&1.014$\pm$0.005\\
7&8.7-10.5&1.002$\pm$0.007\\
8&10.5-12.8&0.990$\pm$0.008\\
9&12.8-15.4&0.984$\pm$0.005\\
10&15.4-19.0&0.976$\pm$0.008\\
11&19.0-24.5&0.980$\pm$0.005\\
12&24.5-34.0&1.007$\pm$0.002\\
13&34.0-63.0&1.035$\pm$0.001\\
14&63.0-270.0&1.064$\pm$0.004
\end{tabular}
\end{center}
\end{table}
\begin{sidewaystable}
\begin{center}
\caption{The measured differential cross-sections as a function of the boson rapidity. The first uncertainty is due to the size of the dataset, the second is due to experimental systematic uncertainties, and the third is due to the luminosity.}
\label{tab:rapidity}
\begin{tabular}{c|ccccccc|ccccccc}
Bin index & \multicolumn{7}{c|}{$\rm{d}\ensuremath{\sigma_{\PZ}}\xspace^{\Pmu\Pmu}/\rm{d}y_\PZ$ [pb]} & \multicolumn{7}{c}{$\rm{d}\ensuremath{\sigma_{\PZ}}\xspace^{\rm{ee}}/\rm{d}y_\PZ$ [pb]}\\\hline
1&14.2&$\pm$&0.7&$\pm$&0.5&$\pm$&0.6&11.8&$\pm$&1.3&$\pm$&0.7&$\pm$&0.5\\
2&41.9&$\pm$&1.2&$\pm$&1.2&$\pm$&1.6&42.1&$\pm$&2.2&$\pm$&1.6&$\pm$&1.6\\
3&65.2&$\pm$&1.5&$\pm$&1.8&$\pm$&2.5&66.1&$\pm$&2.5&$\pm$&2.1&$\pm$&2.6\\
4&91.3&$\pm$&1.8&$\pm$&2.3&$\pm$&3.6&87.9&$\pm$&2.9&$\pm$&2.6&$\pm$&3.4\\
5&108.0&$\pm$&2.0&$\pm$&2.7&$\pm$&4.2&95.8&$\pm$&3.0&$\pm$&2.8&$\pm$&3.7\\
6&121.4&$\pm$&2.1&$\pm$&3.0&$\pm$&4.7&118.5&$\pm$&3.3&$\pm$&3.4&$\pm$&4.6\\
7&136.0&$\pm$&2.2&$\pm$&3.3&$\pm$&5.3&133.3&$\pm$&3.6&$\pm$&3.7&$\pm$&5.2\\
8&140.8&$\pm$&2.2&$\pm$&3.4&$\pm$&5.5&141.3&$\pm$&3.7&$\pm$&3.9&$\pm$&5.5\\
9&145.5&$\pm$&2.3&$\pm$&3.5&$\pm$&5.7&151.2&$\pm$&4.0&$\pm$&4.2&$\pm$&5.9\\
10&144.0&$\pm$&2.3&$\pm$&3.4&$\pm$&5.6&133.6&$\pm$&3.9&$\pm$&3.7&$\pm$&5.2\\
11&137.1&$\pm$&2.2&$\pm$&3.3&$\pm$&5.3&129.6&$\pm$&4.1&$\pm$&3.7&$\pm$&5.1\\
12&121.8&$\pm$&2.1&$\pm$&3.0&$\pm$&4.8&116.5&$\pm$&4.0&$\pm$&3.4&$\pm$&4.5\\
13&100.4&$\pm$&1.9&$\pm$&2.4&$\pm$&3.9&93.5&$\pm$&3.8&$\pm$&2.9&$\pm$&3.6\\
14&75.2&$\pm$&1.7&$\pm$&1.8&$\pm$&2.9&63.8&$\pm$&3.7&$\pm$&2.2&$\pm$&2.5\\
15&57.9&$\pm$&1.5&$\pm$&1.5&$\pm$&2.3&58.6&$\pm$&3.7&$\pm$&2.4&$\pm$&2.3\\
16&41.1&$\pm$&1.2&$\pm$&1.3&$\pm$&1.6&34.7&$\pm$&4.0&$\pm$&1.9&$\pm$&1.4\\
17&18.4&$\pm$&0.6&$\pm$&0.6&$\pm$&0.7&18.8&$\pm$&3.2&$\pm$&1.6&$\pm$&0.7\\
18&2.6&$\pm$&0.2&$\pm$&0.3&$\pm$&0.1&
\end{tabular}
\end{center}
\end{sidewaystable}
\begin{sidewaystable}
\begin{center}
\caption{The measured differential cross-sections as a function of \ensuremath{\phi^{*}_\eta}\xspace. The first uncertainty is due to the size of the dataset, the second is due to experimental systematic uncertainties, and the third is due to the luminosity.}
\label{tab:phist}
\begin{tabular}{c|ccccccc| ccccccc}
Bin index & \multicolumn{7}{c|}{$\rm{d}\ensuremath{\sigma_{\PZ}}\xspace^{\Pmu\Pmu}/\rm{d}\ensuremath{\phi^{*}_\eta}\xspace$ [pb]} & \multicolumn{7}{c}{$\rm{d}\ensuremath{\sigma_{\PZ}}\xspace^{\rm{ee}}/\rm{d}\ensuremath{\phi^{*}_\eta}\xspace$ [pb]}\\\hline
1&1873&$\pm$&29&$\pm$&45&$\pm$&73&1725&$\pm$&49&$\pm$&48&$\pm$&67\\
2&1741&$\pm$&28&$\pm$&42&$\pm$&68&1696&$\pm$&49&$\pm$&48&$\pm$&66\\
3&1635&$\pm$&27&$\pm$&39&$\pm$&64&1549&$\pm$&47&$\pm$&44&$\pm$&60\\
4&1330&$\pm$&17&$\pm$&32&$\pm$&52&1296&$\pm$&30&$\pm$&35&$\pm$&51\\
5&983&$\pm$&15&$\pm$&24&$\pm$&38&955&$\pm$&26&$\pm$&27&$\pm$&37\\
6&722&$\pm$&10&$\pm$&17&$\pm$&28&730&$\pm$&19&$\pm$&20&$\pm$&28\\
7&471&$\pm$&7&$\pm$&11&$\pm$&18&432&$\pm$&11&$\pm$&12&$\pm$&17\\
8&300&$\pm$&5&$\pm$&7&$\pm$&12&300&$\pm$&10&$\pm$&9&$\pm$&12\\
9&160.4&$\pm$&2.7&$\pm$&3.8&$\pm$&6.3&152.4&$\pm$&4.7&$\pm$&4.4&$\pm$&5.9\\
10&81.2&$\pm$&1.9&$\pm$&1.9&$\pm$&3.2&82.6&$\pm$&3.6&$\pm$&2.7&$\pm$&3.2\\
11&38.0&$\pm$&0.9&$\pm$&0.9&$\pm$&1.5&34.0&$\pm$&1.7&$\pm$&1.1&$\pm$&1.3\\
12&14.72&$\pm$&0.58&$\pm$&0.36&$\pm$&0.57&14.71&$\pm$&1.01&$\pm$&0.63&$\pm$&0.57\\
13&6.21&$\pm$&0.27&$\pm$&0.16&$\pm$&0.24&4.94&$\pm$&0.43&$\pm$&0.23&$\pm$&0.19\\
14&1.289&$\pm$&0.086&$\pm$&0.043&$\pm$&0.050&1.213&$\pm$&0.148&$\pm$&0.080&$\pm$&0.047\\
15&0.190&$\pm$&0.021&$\pm$&0.009&$\pm$&0.007&0.201&$\pm$&0.042&$\pm$&0.021&$\pm$&0.008
\end{tabular}
\end{center}
\end{sidewaystable}
\begin{table}[!t]
\begin{center}
\caption{The measured differential cross-sections as a function of \mbox{$p_{\rm T}$}\xspace. The first uncertainty is due to the size of the dataset, the second is due to experimental systematic uncertainties, and the third is due to the luminosity.}
\label{tab:pt}
\begin{tabular}{c|ccccccc}
Bin index & \multicolumn{7}{c}{$\rm{d}\ensuremath{\sigma_{\PZ}}\xspace^{\Pmu\Pmu}/\rm{d}{p_\text{T, Z}}$ [pb / GeV]}\\\hline
1&5.55&$\pm$&0.11&$\pm$&0.15&$\pm$&0.22\\
2&11.01&$\pm$&0.21&$\pm$&0.29&$\pm$&0.43\\
3&11.36&$\pm$&0.21&$\pm$&0.30&$\pm$&0.44\\
4&11.06&$\pm$&0.21&$\pm$&0.29&$\pm$&0.43\\
5&9.93&$\pm$&0.18&$\pm$&0.26&$\pm$&0.39\\
6&8.86&$\pm$&0.16&$\pm$&0.23&$\pm$&0.35\\
7&7.22&$\pm$&0.13&$\pm$&0.19&$\pm$&0.28\\
8&6.48&$\pm$&0.11&$\pm$&0.18&$\pm$&0.25\\
9&5.28&$\pm$&0.09&$\pm$&0.14&$\pm$&0.21\\
10&4.29&$\pm$&0.07&$\pm$&0.12&$\pm$&0.17\\
11&2.88&$\pm$&0.05&$\pm$&0.08&$\pm$&0.11\\
12&1.760&$\pm$&0.029&$\pm$&0.046&$\pm$&0.069\\
13&0.709&$\pm$&0.011&$\pm$&0.018&$\pm$&0.028\\
14&0.0376&$\pm$&0.0009&$\pm$&0.0010&$\pm$&0.0015
\end{tabular}
\end{center}
\end{table}
\begin{sidewaystable}
\caption{The correlation matrix for the differential cross-section measurement as a function of \PZ boson rapidity, for the dimuon final state, excluding the luminosity uncertainty, which is fully correlated between bins.}
\begin{tabular}{c | c c c c c c c c c c c c c c c c c c}
Bin index & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 \\ \hline
1 & 1.00& & & & & & & & & & & & & & & & & \\
2 & 0.37&1.00& & & & & & & & & & & & & & & & \\
3 & 0.35&0.50&1.00& & & & & & & & & & & & & & & \\
4 & 0.35&0.51&0.57&1.00& & & & & & & & & & & & & & \\
5 & 0.35&0.50&0.57&0.62&1.00& & & & & & & & & & & & & \\
6 & 0.34&0.50&0.57&0.62&0.64&1.00& & & & & & & & & & & & \\
7 & 0.34&0.50&0.57&0.63&0.65&0.67&1.00& & & & & & & & & & & \\
8 & 0.33&0.49&0.57&0.62&0.64&0.67&0.68&1.00& & & & & & & & & & \\
9 & 0.33&0.48&0.57&0.61&0.63&0.66&0.67&0.69&1.00& & & & & & & & & \\
10 & 0.32&0.47&0.55&0.61&0.64&0.66&0.68&0.68&0.67&1.00& & & & & & & & \\
11 & 0.31&0.45&0.53&0.59&0.62&0.64&0.66&0.67&0.67&0.68&1.00& & & & & & & \\
12 & 0.28&0.42&0.48&0.57&0.59&0.61&0.63&0.62&0.61&0.65&0.64&1.00& & & & & & \\
13 & 0.28&0.41&0.47&0.54&0.57&0.59&0.61&0.61&0.60&0.64&0.63&0.63&1.00& & & & & \\
14 & 0.26&0.38&0.43&0.50&0.53&0.54&0.57&0.56&0.56&0.59&0.59&0.59&0.58&1.00& & & & \\
15 & 0.23&0.34&0.39&0.46&0.48&0.49&0.51&0.51&0.50&0.55&0.54&0.55&0.54&0.51&1.00& & & \\
16 & 0.19&0.28&0.31&0.37&0.39&0.39&0.41&0.41&0.39&0.44&0.44&0.46&0.45&0.43&0.42&1.00& & \\
17 & 0.19&0.27&0.31&0.36&0.38&0.39&0.40&0.40&0.40&0.44&0.44&0.45&0.44&0.42&0.41&0.35&1.00& \\
18 & 0.05&0.08&0.09&0.11&0.11&0.11&0.12&0.12&0.11&0.13&0.13&0.14&0.14&0.14&0.14&0.12&0.14&1.00
\label{tab:corr_rap}
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}
\caption{The correlation matrix for the differential cross-section measurements as a function of the \PZ boson rapidity, for the dielectron final state, excluding the luminosity uncertainty, which is fully correlated between bins.}
\begin{tabular}{c | c c c c c c c c c c c c c c c c c }
Bin index & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 & 16 & 17 \\ \hline
1 & 1.00 & & & & & & & & & & & & & & & & \\
2 & 0.07 & 1.00 & & & & & & & & & & & & & & & \\
3 & 0.09 & 0.19 & 1.00 & & & & & & & & & & & & & & \\
4 & 0.09 & 0.17 & 0.22 & 1.00 & & & & & & & & & & & & & \\
5 & 0.11 & 0.22 & 0.28 & 0.26 & 1.00 & & & & & & & & & & & & \\
6 & 0.12 & 0.24 & 0.30 & 0.28 & 0.35 & 1.00 & & & & & & & & & & & \\
7 & 0.12 & 0.24 & 0.31 & 0.29 & 0.36 & 0.39 & 1.00 & & & & & & & & & & \\
8 & 0.12 & 0.24 & 0.31 & 0.29 & 0.37 & 0.39 & 0.40 & 1.00 & & & & & & & & & \\
9 & 0.12 & 0.24 & 0.31 & 0.28 & 0.36 & 0.39 & 0.40 & 0.40 & 1.00 & & & & & & & & \\
10 & 0.11 & 0.23 & 0.29 & 0.27 & 0.34 & 0.37 & 0.38 & 0.38 & 0.38 & 1.00 & & & & & & & \\
11 & 0.11 & 0.22 & 0.28 & 0.26 & 0.33 & 0.35 & 0.36 & 0.36 & 0.36 & 0.34 & 1.00 & & & & & & \\
12 & 0.10 & 0.21 & 0.26 & 0.24 & 0.31 & 0.33 & 0.34 & 0.34 & 0.34 & 0.32 & 0.31 & 1.00 & & & & & \\
13 & 0.09 & 0.18 & 0.23 & 0.22 & 0.27 & 0.29 & 0.30 & 0.30 & 0.30 & 0.29 & 0.27 & 0.26 & 1.00 & & & & \\
14 & 0.07 & 0.14 & 0.18 & 0.17 & 0.21 & 0.23 & 0.23 & 0.24 & 0.23 & 0.22 & 0.21 & 0.20 & 0.18 & 1.00 & & & \\
15 & 0.07 & 0.14 & 0.17 & 0.16 & 0.20 & 0.22 & 0.22 & 0.23 & 0.22 & 0.21 & 0.20 & 0.19 & 0.17 & 0.13 & 1.00 & & \\
16 & 0.04 & 0.08 & 0.11 & 0.10 & 0.13 & 0.13 & 0.14 & 0.14 & 0.14 & 0.13 & 0.13 & 0.12 & 0.10 & 0.08 & 0.08 & 1.00 & \\
17 & 0.03 & 0.06 & 0.07 & 0.07 & 0.09 & 0.09 & 0.10 & 0.10 & 0.10 & 0.09 & 0.09 & 0.08 & 0.07 & 0.06 & 0.05 & 0.03 & 1.00
\label{tab:corr_rap_electron}
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}
\caption{The correlation matrix for the differential cross-section measurement as a function of \ensuremath{\phi^{*}_\eta}\xspace, for the dimuon final state, excluding the luminosity uncertainty, which is fully correlated between bins.}
\begin{tabular}{c | c c c c c c c c c c c c c c c}
Bin index & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 \\ \hline
1 & 1.00 & & & & & & & & & & & & & & \\
2 & 0.69 & 1.00 & & & & & & & & & & & & & \\
3 & 0.68 & 0.66 & 1.00 & & & & & & & & & & & & \\
4 & 0.73 & 0.72 & 0.71 & 1.00 & & & & & & & & & & & \\
5 & 0.70 & 0.69 & 0.67 & 0.73 & 1.00 & & & & & & & & & & \\
6 & 0.71 & 0.70 & 0.69 & 0.74 & 0.71 & 1.00 & & & & & & & & & \\
7 & 0.70 & 0.69 & 0.70 & 0.74 & 0.70 & 0.72 & 1.00 & & & & & & & & \\
8 & 0.67 & 0.65 & 0.66 & 0.70 & 0.66 & 0.68 & 0.69 & 1.00 & & & & & & & \\
9 & 0.68 & 0.67 & 0.66 & 0.71 & 0.68 & 0.69 & 0.69 & 0.65 & 1.00 & & & & & & \\
10 & 0.59 & 0.58 & 0.57 & 0.62 & 0.59 & 0.60 & 0.60 & 0.57 & 0.58 & 1.00 & & & & & \\
11 & 0.56 & 0.54 & 0.56 & 0.60 & 0.55 & 0.57 & 0.59 & 0.56 & 0.55 & 0.48 & 1.00 & & & & \\
12 & 0.43 & 0.42 & 0.42 & 0.45 & 0.42 & 0.44 & 0.44 & 0.42 & 0.42 & 0.36 & 0.36 & 1.00 & & & \\
13 & 0.40 & 0.40 & 0.38 & 0.41 & 0.41 & 0.41 & 0.40 & 0.38 & 0.39 & 0.34 & 0.31 & 0.24 & 1.00 & & \\
14 & 0.25 & 0.23 & 0.27 & 0.28 & 0.23 & 0.26 & 0.30 & 0.28 & 0.26 & 0.21 & 0.26 & 0.18 & 0.12 & 1.00 & \\
15 & 0.16 & 0.15 & 0.17 & 0.18 & 0.15 & 0.17 & 0.18 & 0.17 & 0.16 & 0.14 & 0.15 & 0.11 & 0.09 & 0.10 & 1.00
\label{tab:corr_ps_mu}
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}
\caption{The correlation matrix for the differential cross-section measurements as a function of \ensuremath{\phi^{*}_\eta}\xspace, for the dielectron final state, excluding the luminosity uncertainty, which is fully correlated between bins.}
\begin{tabular}{c | c c c c c c c c c c c c c c c}
Bin index & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 \\\hline
1 & 1.00 & & & & & & & & & & & & & & \\
2 & 0.36 & 1.00 & & & & & & & & & & & & & \\
3 & 0.35 & 0.35 & 1.00 & & & & & & & & & & & & \\
4 & 0.45 & 0.45 & 0.43 & 1.00 & & & & & & & & & & & \\
5 & 0.37 & 0.37 & 0.36 & 0.46 & 1.00 & & & & & & & & & & \\
6 & 0.39 & 0.38 & 0.37 & 0.48 & 0.39 & 1.00 & & & & & & & & & \\
7 & 0.39 & 0.38 & 0.37 & 0.48 & 0.39 & 0.41 & 1.00 & & & & & & & & \\
8 & 0.34 & 0.34 & 0.33 & 0.42 & 0.35 & 0.36 & 0.36 & 1.00 & & & & & & & \\
9 & 0.35 & 0.34 & 0.33 & 0.43 & 0.35 & 0.37 & 0.37 & 0.33 & 1.00 & & & & & & \\
10 & 0.27 & 0.27 & 0.26 & 0.34 & 0.28 & 0.29 & 0.29 & 0.25 & 0.26 & 1.00 & & & & & \\
11 & 0.25 & 0.25 & 0.24 & 0.31 & 0.26 & 0.26 & 0.27 & 0.23 & 0.24 & 0.19 & 1.00 & & & & \\
12 & 0.19 & 0.19 & 0.18 & 0.23 & 0.19 & 0.20 & 0.20 & 0.18 & 0.18 & 0.14 & 0.13 & 1.00 & & & \\
13 & 0.15 & 0.15 & 0.15 & 0.19 & 0.15 & 0.16 & 0.16 & 0.14 & 0.14 & 0.11 & 0.10 & 0.08 & 1.00 & & \\
14 & 0.11 & 0.11 & 0.11 & 0.14 & 0.11 & 0.12 & 0.12 & 0.10 & 0.11 & 0.08 & 0.08 & 0.06 & 0.05 & 1.00 & \\
15 & 0.08 & 0.08 & 0.08 & 0.10 & 0.08 & 0.09 & 0.09 & 0.08 & 0.08 & 0.06 & 0.06 & 0.04 & 0.03 & 0.02 & 1.00
\label{tab:corr_ps_e}
\end{tabular}
\end{sidewaystable}
\begin{sidewaystable}
\caption{The correlation matrix for the differential cross-section measurements as a function of boson \mbox{$p_{\rm T}$}\xspace, for the dimuon final state, excluding the luminosity uncertainty, which is fully correlated between bins. }
\begin{tabular}{c | c c c c c c c c c c c c c c }
Bin index & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 \\ \hline
1 & 1.00 & & & & & & & & & & & & & \\
2 & 0.54 & 1.00 & & & & & & & & & & & & \\
3 & 0.54 & 0.55 & 1.00 & & & & & & & & & & & \\
4 & 0.53 & 0.54 & 0.55 & 1.00 & & & & & & & & & & \\
5 & 0.55 & 0.55 & 0.55 & 0.56 & 1.00 & & & & & & & & & \\
6 & 0.54 & 0.56 & 0.56 & 0.55 & 0.56 & 1.00 & & & & & & & & \\
7 & 0.53 & 0.53 & 0.54 & 0.55 & 0.56 & 0.54 & 1.00 & & & & & & & \\
8 & 0.52 & 0.54 & 0.55 & 0.55 & 0.55 & 0.55 & 0.54 & 1.00 & & & & & & \\
9 & 0.51 & 0.52 & 0.54 & 0.55 & 0.55 & 0.53 & 0.55 & 0.55 & 1.00 & & & & & \\
10 & 0.53 & 0.54 & 0.55 & 0.55 & 0.57 & 0.54 & 0.55 & 0.55 & 0.56 & 1.00 & & & & \\
11 & 0.53 & 0.55 & 0.57 & 0.56 & 0.56 & 0.56 & 0.55 & 0.57 & 0.57 & 0.56 & 1.00 & & & \\
12 & 0.53 & 0.56 & 0.57 & 0.56 & 0.56 & 0.56 & 0.55 & 0.58 & 0.57 & 0.57 & 0.60 & 1.00 & & \\
13 & 0.54 & 0.55 & 0.57 & 0.58 & 0.58 & 0.56 & 0.58 & 0.58 & 0.60 & 0.59 & 0.60 & 0.61 & 1.00 & \\
14 & 0.42 & 0.44 & 0.45 & 0.47 & 0.47 & 0.44 & 0.47 & 0.48 & 0.50 & 0.49 & 0.50 & 0.51 & 0.54 & 1.00
\label{tab:corr_pt}
\end{tabular}
\end{sidewaystable}
\section{Dataset and event selection}
\label{sec:Selection}
This analysis uses a dataset corresponding to an integrated luminosity of $294\pm11$\ensuremath{\mbox{\,pb}^{-1}}\xspace recorded by the LHCb experiment in pp collisions at $\sqrt{s} = 13$\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace. This integrated luminosity is determined using the beam-imaging techniques described in Ref.~\cite{LHCb-PAPER-2014-047}. Candidates are selected by requiring two high \mbox{$p_{\rm T}$}\xspace muons or electrons of opposite charge. Additional requirements are then made to select pure samples; these and the resulting purity are now discussed in turn for the dimuon and dielectron final states.
\subsection{Dimuon final state}
The fiducial requirements outlined in Sect.~\ref{sec:Introduction} are applied as selection criteria for the dimuon final state. In addition, the two tracks are required to satisfy quality criteria and to be identified as muons. At least one of the muons is required to be responsible for the event passing the hardware and software stages of the trigger. The number of selected $\PZ\ensuremath{\rightarrow}\xspace\Pmu\Pmu$ candidates is $43\,643$.
Five sources of background are investigated: heavy flavour hadron decays, misidentified hadrons, $\PZ\ensuremath{\rightarrow}\xspace\Ptau\Ptau$ decays, {\ensuremath{\tquark\tquarkbar}}\xspace events, and WW events. Similar techniques to those used in previous analyses are applied to quantify the contribution of each source~\cite{LHCb-PAPER-2015-001,LHCb-PAPER-2015-049}.
The contribution where at least one muon is produced by the decay of heavy flavour particles is studied by selecting sub-samples where this contribution is enhanced, either by requiring that the muons are not spatially isolated from other activity in the event, or by requiring that the muons are not consistent with a common production point. Studies on these two sub-samples are consistent, and the background contribution is estimated to be $180\pm50$ events.
The contribution from misidentified hadrons is evaluated from the probability with which hadrons are incorrectly identified as muons, and is determined to be $100\pm 13$ events. Following Refs.~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2015-001,LHCb-PAPER-2015-049}, this evaluation is made with randomly triggered data.
An alternative estimate of the contribution from these sources is found by selecting events where both muons have the same charge, but pass all other selection criteria. The assumption that the charges of the selected muons are uncorrelated for these sources is validated by confirming that the same-sign event yield is compatible with the opposite-sign event yield in background-enriched regions.
The overall number of same-sign events is 198, with the numbers of $\Pmu^+\Pmu^+$ and $\Pmu^-\Pmu^-$ candidates statistically compatible with each other.
The difference between this number and the sum of the hadron misidentification and heavy-flavour contributions is assigned as an additional uncertainty on the purity estimate.
The contribution from $\PZ\ensuremath{\rightarrow}\xspace\Ptau\Ptau$ decays where both \Ptau leptons subsequently decay to muons is estimated from \pythia8 simulation to be $30\pm10$ events. The background from muons produced in top-quark decays is determined from simulation normalised using the measurement of the cross-section for top-pair production measured at the ATLAS experiment~\cite{ATLAS-CONF-2015-049}, and is estimated to be $28\pm10$ events. The background from WW decays is also determined from the simulation and found to be negligible. Overall, the purity of the dataset is estimated to be $\rho^{\Pmu\Pmu} = (99.2\pm0.2)\%$, consistent with purity estimates found in previous LHCb measurements at lower centre-of-mass energies~\cite{LHCb-PAPER-2015-001,LHCb-PAPER-2015-049}. As in these previous measurements, no significant variation of the purity is found as a function of the kinematic variables studied, and so the purity is treated as constant. A systematic uncertainty associated with this assumption is discussed in Sect.~\ref{sec:Systematics}.
\subsection{Dielectron final state}
The dielectron final state requires two opposite-sign electron candidates, using the same selection criteria based on calorimeter energy deposits as previous LHCb analyses~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2015-003}. Electron candidates are required to have $\mbox{$p_{\rm T}$}\xspace > 20$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace and $2.0<\eta<4.5$. A loose requirement is made on the dielectron invariant mass, $m(\Pe\Pe) > 40$~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, since many events where the dielectron system is produced with an invariant mass above 60~GeV may be reconstructed at lower mass due to bremsstrahlung. Effects arising from the difference between the fiducial acceptance and the selection requirements will be discussed in Sect.~\ref{sec:acc}. At least one of the electrons is required to be responsible for the event passing the hardware and software stages of the LHCb trigger. In total $16\,395$ candidates are selected.
Backgrounds are determined using similar techniques as in previous analyses~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2015-003}. A sample of same-sign $\Pe^\pm \Pe^\pm$ combinations, otherwise subject to the same selection criteria as the standard dataset, is used to provide a data-based estimate of the largest backgrounds. Hadrons that shower early in the ECAL and fake the signature of an electron are expected to be the dominant background, and should contribute roughly equally to same-sign and opposite-sign pairs. The contribution from heavy-flavour decays is also expected to contribute approximately equally to same-sign and opposite-sign datasets, and is much smaller than the background due to misidentified hadrons. Overall, $1\,255$ candidate same-sign events are selected, with no significant difference observed between the $\Pe^+\Pe^+$ and $\Pe^-\Pe^-$ datasets. In order to ascertain the reliability of this procedure, a hadron-enriched sample is selected by requiring that one of the electron candidates is associated with a significant energy deposit in the HCAL, suggesting that it is likely to be a misidentified hadron. The numbers of same-sign and opposite-sign pairs satisfying these requirements are found to agree within 6.2\%. Consequently a 6.2\% uncertainty is assigned to the estimated yield of background events, which corresponds to a 0.5\% uncertainty on the signal yield. In addition, simulated background datasets of $\PZ\ensuremath{\rightarrow}\xspace\Ptau\Ptau$ decays, {\ensuremath{\tquark\tquarkbar}}\xspace events and WW events are generated~\cite{Sjostrand:2007gs,*Sjostrand:2006za} and studied similarly to the dimuon final state. These all contribute at the level of 0.1\% or less.
The overall purity of the electron dataset is found to be $\rho^{\Pe\Pe} = (92.2\pm0.5)\%$.
\section{Cross-section measurement}
\label{sec:Method}
The \PZ boson production cross-section is measured in bins of $y_\PZ$, \ensuremath{\phi^{*}_\eta}\xspace, and, for the dimuon final state, in bins of the boson \mbox{$p_{\rm T}$}\xspace. For the dimuon final state the efficiency is obtained from per-event weights that depend on the kinematics of the muons, whereas for the dielectron final state the reconstruction and detection efficiencies are evaluated within each bin of the distribution. These approaches are validated using simulation.
The cross-section for the dimuon final state in a particular bin $i$ is determined as
\begin{align*}
\ensuremath{\sigma_{\PZ}}\xspace^{\Pmu\Pmu}(i) =& \frac{1}{L}\, \rho^{\Pmu\Pmu} \ensuremath{f_{\textrm{\fsr}}}\xspace^{\Pmu\Pmu}(i) f_\text{unf}^{\Pmu\Pmu}(i) \sum\limits_{j=1}^{N_{\PZ}^{\Pmu\Pmu}(i)} \frac{1}{\varepsilon(\Pmu^{+}_j, \Pmu^{-}_j)},
\end{align*}
\noindent where the index $j$ runs over the candidates contributing to the bin, with the total number of candidates in the bin denoted by $N_{\PZ}^{\Pmu\Pmu}(i)$. The total reconstruction and detection efficiency for a given event $j$, $\varepsilon(\Pmu^{+}_j,\Pmu^{-}_j)$, depends on the kinematics of each muon. The correction factors for final-state radiation (FSR) are denoted by $\ensuremath{f_{\textrm{\fsr}}}\xspace^{\Pmu\Pmu}(i)$. Corrections for resolution effects that cause bin-to-bin migrations, where applicable, which do not change the fiducial cross-section, are denoted by $f_\text{unf}^{\Pmu\Pmu}(i)$. Migration of events in and out of the overall LHCb fiducial acceptance is negligible. The purity, introduced earlier, is denoted $\rho^{\Pmu\Pmu}$. The integrated luminosity is denoted by $L$.
For the dielectron final state the cross-section in a particular bin is determined as
\begin{align*}
\ensuremath{\sigma_{\PZ}}\xspace^{\Pe\Pe}(i) =&\frac{1}{L}\, \rho^{ee}(i) \ensuremath{f_{\textrm{\fsr}}}\xspace^{{\Pe\Pe}}(i) f^{\Pe\Pe}_\text{MZ}(i) \frac{N_{\PZ}^{\Pe\Pe}(i)}{\varepsilon^{\Pe\Pe}(i)},
\end{align*}
where $N_{\PZ}^{\Pe\Pe}(i)$ denotes the number of candidates in bin $i$. The efficiency associated with reconstructing the dielectron final state in bin $i$ is $\varepsilon^{\Pe\Pe}(i)$ and the purity is $\rho^{\Pe\Pe}$. The correction for FSR from the electrons is denoted $\ensuremath{f_{\textrm{\fsr}}}\xspace^{\Pe\Pe}(i)$, while $f_\text{MZ}^{\Pe\Pe}(i)$ corrects the measurement for migrations in the dielectron invariant mass into and out of the fiducial region.
For both final states the total cross-section is obtained by summing over $i$. The various correction factors are discussed below.
\subsection{Efficiency determination}
\label{sec:eff}
For the measurement in the dimuon final state, candidates are assigned a weight associated with the probability of reconstructing each muon, and the correction for any inefficiency is applied on an event-by-event basis. Muon reconstruction efficiencies are determined directly from data using the same tag-and-probe techniques as applied in previous LHCb measurements of high-\mbox{$p_{\rm T}$}\xspace muons~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2014-033,LHCb-PAPER-2015-001,LHCb-PAPER-2015-049}. Averaged over the muon kinematic distributions, the track reconstruction efficiency is determined to be 95\%, the muon identification efficiency is determined to be 95\% and the single muon trigger efficiency is 80\%. Since either muon can be responsible for the event passing the trigger, the overall efficiency with which candidates pass the trigger is higher, on average 95\%. These efficiencies are determined as a function of the muon pseudorapidity. Efficiency measurements as a function of other variables, such as the muon \mbox{$p_{\rm T}$}\xspace and the detector occupancy, are studied as a cross-check, with no significant change in the final results.
For the measurement in the dielectron final state, electron reconstruction efficiencies are determined from data and simulation for each bin of the measurement, using the same techniques applied in previous LHCb measurements of $\PZ\rightarrow\Pe\Pe$ production~\cite{LHCb-PAPER-2012-036,LHCb-PAPER-2015-003}. The use of different techniques to determine efficiencies to those applied in the muon channel provides uncorrelated systematic uncertainties between the two measurements. The efficiency for electrons is factorised into similar components to those applied in the dimuon analysis, though one extra effect is considered. The GEC efficiency determines the probability that the dielectron candidates pass the GEC present in the hardware trigger. There is no such requirement in the dimuon trigger.
The GEC efficiency for dielectron data is determined from the dimuon data, correcting for small differences in the detector response to muons and electrons.
The average GEC efficiency is 79\% and exhibits a weak dependence on rapidity and \ensuremath{\phi^{*}_\eta}\xspace. The trigger efficiency is determined directly from data using a tag-and-probe method, and is typically 93\%. The efficiency with which both electrons are identified by the calorimetry is typically 78\% and is determined from simulation that has been calibrated with data. This efficiency exhibits a significant dependence on the boson rapidity, since the LHCb calorimeter acceptance only extends as far as $\eta \approx 4.25$. The track reconstruction and kinematic efficiency describes the efficiency with which electrons that are in the fiducial region are reconstructed with $\mbox{$p_{\rm T}$}\xspace > 20\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$. It corrects both for failure to reconstruct a track and for incomplete bremsstrahlung recovery incorrectly reconstructing electrons with \mbox{$p_{\rm T}$}\xspace below the 20~GeV threshold. This is also determined from simulation calibrated to data, and is on average $48\%$.
\subsection{Resolution effects}
\label{sec:unf}
The excellent angular resolution of the LHCb detector in comparison to the bin widths means that no significant bin-to-bin migrations occur in the \ensuremath{\phi^{*}_\eta}\xspace or $y_\PZ$ distributions for either the dimuon or dielectron final states. In addition, net migration in and out of the overall LHCb angular acceptance is negligible. However, small migrations in the boson \mbox{$p_{\rm T}$}\xspace distribution measured using the dimuon final state are expected at low transverse momenta. These effects are typically of similar size to the statistical uncertainty in each bin. This distribution is therefore unfolded to correct for the impact of these migrations using multiplicative correction factors (defined above as $f_\text{unf}^{\Pmu\Pmu}$) determined for each bin from simulation.
\subsection{Final-state radiation corrections}
\label{sec:FSR}
The data are corrected for the effect of FSR from the leptons, allowing comparison of electron and muon final states. The correction in each bin of the measured differential distributions is taken as the average of the values determined using \mbox{\textsc{Herwig++}}\xspace~\cite{GEN-HERWIGPP} and \pythia8~\cite{Sjostrand:2007gs,*Sjostrand:2006za}. The two generators typically agree at the per-mille level; the mean correction is about 2\% for muons and 5\% for electrons, but dependence is seen as functions of the different kinematic variables studied. The strongest variation is seen as a function of the boson \mbox{$p_{\rm T}$}\xspace, where the correction varies over the distribution by about 10\%. The corrections applied are tabulated in the \hyperref[appendix_results]{appendix}.
\subsection{Acceptance corrections}
\label{sec:acc}
The acceptance correction $f_\text{MZ}^{\text{ee}}$ is applied for electrons to correct for events which pass the selection but are not in the fiducial acceptance in dilepton mass. This correction factor, typically 0.97, is determined from simulation as in previous analyses~\cite{LHCb-PAPER-2012-036,LHCb-PAPER-2015-003} and cross-checked using data.
No correction is applied for muons, where the fiducial acceptance is identical to the kinematic requirement in the acceptance, and where the experimental resolution is sufficient such that net migrations in and out of the acceptance due to experimental resolution are negligible.
\subsection{Measuring fiducial cross-sections}
The fiducial cross-sections are determined by integrating over the $y_{_\PZ}$ distributions. Since no candidates in the bin $4.25<y_{_\PZ}<4.50$ are observed for electrons, a correction for this bin is evaluated using FEWZ~\cite{GEN-FEWZ3}. This correction is found to be 0.7~pb. The fraction of the fiducial cross-section expected in the bin determined using \pythia8 simulation~\cite{Sjostrand:2007gs,*Sjostrand:2006za} is consistent with this estimate to within 0.1~pb. This is assigned as the uncertainty associated with the contribution from this bin to the fiducial cross-section measured in the dielectron final state. Consistent results are obtained when integrating over \ensuremath{\phi^{*}_\eta}\xspace or \mbox{$p_{\rm T}$}\xspace.
\section{Detector and simulation}
\label{sec:Detector}
The \mbox{LHCb}\xspace detector~\cite{Alves:2008zz,LHCb-DP-2014-002} is a single-arm forward
spectrometer covering the \mbox{pseudorapidity} range $2<\eta <5$, primarily
designed for the study of particles containing b or c~quarks. The detector includes a high-precision tracking system
consisting of a silicon-strip vertex detector surrounding the pp
interaction region, a large-area silicon-strip detector located
upstream of a dipole magnet with a bending power of about
$4{\mathrm{\,Tm}}$, and three stations of silicon-strip detectors and straw
drift tubes placed downstream of the magnet.
The tracking system provides a measurement of momentum, \mbox{$p$}\xspace, of charged particles with
a relative uncertainty that varies from 0.5\% at low momentum to 1.0\% at 200\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
The minimum distance of a track to a primary vertex, the impact parameter, is measured with a resolution of $(15+29/\mbox{$p_{\rm T}$}\xspace)\ensuremath{{\,\upmu\rm m}}\xspace$,
where the \mbox{$p_{\rm T}$}\xspace is measured in\,\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
Different types of charged hadrons are distinguished using information
from two ring-imaging Cherenkov detectors.
Photons, electrons and hadrons are identified by a calorimeter system consisting of
scintillating-pad (SPD) and preshower (PS) detectors, an electromagnetic
calorimeter (ECAL) and a hadronic calorimeter (HCAL). Muons are identified by a
system composed of alternating layers of iron and multiwire
proportional chambers.
The online event selection is performed by a trigger,
which consists of a hardware stage, based on information from the calorimeter and muon
systems, followed by a software stage, which applies a full event
reconstruction. The analysis described here uses triggers designed to select events containing at least one muon or at least one electron. The hardware trigger used for these studies requires that a candidate muon has $\mbox{$p_{\rm T}$}\xspace > 6$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace or that a candidate electron has transverse energy $\mbox{$E_{\rm T}$}\xspace > 2.28$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace. Global event cuts (GEC) are applied in the electron trigger in order to prevent events with high occupancy from dominating the processing time: events only pass the electron trigger if they contain fewer than 450 hits in the SPD detector. No such requirement is made within the muon trigger. The software trigger used here selects events containing a muon candidate with $\mbox{$p_{\rm T}$}\xspace >12.5$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, or an electron candidate with $\mbox{$p_{\rm T}$}\xspace >15$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace.
The main challenge with electron reconstruction at LHCb is the energy measurement. The calorimeters at LHCb are optimised for the study of low \mbox{$E_{\rm T}$}\xspace physics, and individual cells saturate for transverse energies greater than approximately 10\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace. Electron reconstruction at LHCb therefore relies on accurate tracking measurements to determine the electron momentum. However, bremsstrahlung photons are often emitted as an electron traverses the LHCb detector, so the measured momentum does not directly correspond to the momentum of the electron produced in the proton-proton collision. These photons are often collinear with the electron and are detected in the same saturated calorimeter cell so that recovery of this emitted photon energy is incomplete. Consequently LHCb accurately determines the direction of electrons, but tends to underestimate their energy by a variable amount, typically around 25\%. Despite these challenges, the excellent angular resolution of electrons provided by the LHCb detector means that measurements using the dielectron final state can be used to complement analyses of angular variables such as rapidity and \ensuremath{\phi^{*}_\eta}\xspace in the dimuon final state~\cite{LHCb-PAPER-2012-036,LHCb-PAPER-2015-003}.
Simulated pp collisions for the study of reconstruction effects are generated using
\pythia8~\cite{Sjostrand:2007gs,*Sjostrand:2006za} with a specific \mbox{LHCb}\xspace
configuration~\cite{LHCb-PROC-2010-056}. Decays of hadronic particles
are described by \mbox{\textsc{EvtGen}}\xspace~\cite{Lange:2001uf}, in which final-state
radiation is modelled using \mbox{\textsc{Photos}}\xspace~\cite{Golonka:2005pn}. The
interaction of the generated particles with the detector, and its response,
are implemented using the \mbox{\textsc{Geant4}}\xspace toolkit~\cite{Allison:2006ve, *Agostinelli:2002hh} as described in Ref.~\cite{LHCb-PROC-2011-006}.
The results reported in this article are compared to fixed-order predictions calculated within perturbative quantum chromodynamics (pQCD) determined using the FEWZ~3.1 generator~\cite{GEN-FEWZ3} at $\mathcal{O}(\alpha_s^2)$, where ${\ensuremath{\alpha_s}}\xspace$ is the coupling strength of the strong force. These predictions do not include electroweak corrections. Predictions are made using MMHT14~\cite{PDF-MMHT14}, NNPDF3.0~\cite{PDF-NNPDF30}, and CT14~\cite{PDF-CT14} PDF sets. In all cases, the factorisation and renormalisation scales are set to the \PZ boson mass. Uncertainties on the fixed-order predictions are evaluated by varying the factorisation and renormalisation scales independently using the seven-point scale variation prescription~\cite{Hamilton:2013fea}, and combining this effect in quadrature with the 68\% CL uncertainties associated with the PDF sets and the value of ${\ensuremath{\alpha_s}}\xspace$. The results are also compared to predictions using the Monash 2013 tune of \pythia8~\cite{Sjostrand:2007gs,*Sjostrand:2006za,Skands:2014pea} and an updated version of the LHCb-specific \pythia8 tune~\cite{LHCb-PROC-2010-056}. In addition, results are compared to predictions from \mbox{\textsc{Powheg}}\xspace~\cite{GEN-POWHEG,Alioli:2008gx} at $\mathcal{O}({\ensuremath{\alpha_s}}\xspace)$ using the NNPDF3.0 PDF set, with the showering implemented using \pythia8. These predictions are calculated using the default \mbox{\textsc{Powheg}}\xspace settings and the \pythia8 Monash 2013 tune. The \PZ differential cross-section results are also compared to simulated datasets produced using {\textsc{MadGraph5}}\_aMC@NLO\cite{Alwall:2014hca}. Different schemes are used to match and merge these samples. The MLM~\cite{mlm} sample has leading-order accuracy for the emission of zero, one or two jets; the FxFx~\cite{Frederix:2012ps} sample has next-to-leading-order (NLO) accuracy for zero- or one-jet emission; and the UNLOPS~\cite{Lonnblad:2012ix} sample is accurate at NLO for zero- or one-jet emission and accurate at LO for two-jet emission. Higher jet multiplicities are generated by a parton shower, implemented here using the Monash 2013 tune for \pythia8.
\section{Introduction}
\label{sec:Introduction}
Measurements are reported of \PZ boson production\footnote{The label \PZ boson is defined to include the effects of virtual photon production and interference terms. The terms electron and muon are also used to refer to both matter and anti-matter species of the particles.} at the LHCb experiment in proton-proton collisions at $\sqrt{s}=13$\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace. The analysis uses a dataset corresponding to an integrated luminosity of $294\pm11$\ensuremath{\mbox{\,pb}^{-1}}\xspace and considers events where the boson decays either to a dimuon or a dielectron final state. The two final states offer statistically independent samples with largely independent systematic uncertainties. The analysis is performed using similar methods to previous LHCb measurements of electroweak boson production at lower pp collision energies~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2012-036,LHCb-PAPER-2015-001,LHCb-PAPER-2015-003,LHCb-PAPER-2015-049}. The LHCb detector measures particle production in the forward region; the ATLAS and CMS collaborations have reported similar measurements at $\sqrt{s}=13$\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace~\cite{Aad:2016naf,CMS-PAS-SMP-15-011} in a different kinematic region.
Measurements of electroweak gauge boson production are benchmark tests of Standard Model processes at hadron colliders, and are of interest for constraining the parton distribution functions (PDFs) that describe the structure of the proton. Because of the longitudinal boost required for a \PZ boson to be produced in the forward region, LHCb results are particularly sensitive to effects at low and high values of Bjorken-$x$\cite{Thorne:2008am}, and have been used to constrain global PDF fits~\cite{PDF-MMHT14,PDF-NNPDF30,PDF-CT14}. The $\sqrt{s} = 13$~\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace pp collisions allow LHCb to access lower values of $x$ than previous measurements at 7 and 8~TeV. In addition, the boson transverse momentum (\mbox{$p_{\rm T}$}\xspace) and $\phi^*_\eta$ distributions can be used to test Monte Carlo modelling of additional higher-order radiation that arises from quantum chromodynamics (QCD). The $\phi^*_\eta$ variable~\cite{Banfi:2010cf} is defined as $\phi^*_\eta \equiv \tan(\phi_\text{acop} / 2) / \cosh(\Delta\eta /2)$, where the acoplanarity angle $\phi_\text{acop} \equiv \pi - \Delta\phi$ depends on the difference in azimuthal angle of the two leptons, $\Delta\phi$, and $\Delta\eta$ is the difference in pseudorapidity of the two leptons. This variable probes similar physics to that probed by the boson transverse momentum, but with better experimental resolution.
The fiducial region used for the results presented here is the same as in
previous measurements of \PZ boson production at LHCb~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2012-029,LHCb-PAPER-2012-036,LHCb-PAPER-2015-001,LHCb-PAPER-2015-003,LHCb-PAPER-2015-049}. Both final-state leptons are required to have $\mbox{$p_{\rm T}$}\xspace > 20$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace and pseudorapidity $2.0\nolinebreak<\nolinebreak\eta\nolinebreak<\nolinebreak4.5$.\footnote{This article uses natural units with $c=1$.} The invariant mass of the dilepton pair, $m(\ell\ell)$, is required to be in the range $60\nolinebreak<\nolinebreak m(\ell\ell)\nolinebreak<\nolinebreak120$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace. The measurements are corrected for final-state radiation to the Born level in quantum electrodynamics (QED), allowing direct comparison of the results in the muon and electron final states, which are reported separately in bins of the boson rapidity, $y_\PZ$, of $\phi^*_\eta$ and, using the dimuon events, as a function of the boson \mbox{$p_{\rm T}$}\xspace. Cross-sections integrated over the fiducial region (fiducial cross-sections) are also determined using both final states. These are then averaged into a single measurement of the $\PZ\rightarrow \ell\ell$ fiducial cross-section in $\sqrt{s} = 13$~TeV pp collisions.
\section{Systematic uncertainties}
\label{sec:Syst}
The systematic uncertainties associated with the measurement are estimated using the same techniques as in previous analyses~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2015-001,LHCb-PAPER-2015-003,LHCb-PAPER-2015-049}. The contributions from different sources are combined in quadrature. The uncertainties on the fiducial cross-section measurement are summarised in Table~\ref{tab:systsum}.
For both muons and electrons, the statistical precisions of the efficiencies are assigned as systematic uncertainties. For muons, the accuracy of the tag-and-probe methods used to determine efficiencies is tested in simulation, and efficiencies calculated using the tag-and-probe method are generally found to match simulated efficiencies at the per-mille level, with the largest difference arising from the determination of the track reconstruction efficiency. An uncertainty of 1\% is assigned to this efficiency for each muon. The method of treating each muon independently and applying the efficiencies as a function of the muon pseudorapidity is also studied in simulation, and is found to be accurate to better than 0.6\%. This is also assigned as a systematic uncertainty. For electrons, the accuracy of the method used to determine the trigger efficiency is studied by applying it to the simulated dataset and comparing the resulting efficiencies to those directly determined in the same dataset: no bias is observed, and no additional uncertainty is assigned. For the electron track reconstruction efficiency the relative performance in data and simulation is studied using a tag-and-probe method and an uncertainty of 1.6\% is assigned. The uncertainty associated with potential mismodelling of the electron identification efficiency is determined by comparing between data and simulation the distributions of calorimeter energy deposits used to identify electrons. The impact of any mismodelling is propagated through the measurement, and an uncertainty of 1.3\% is assigned. Apart from the uncertainties arising from the statistical precision of the efficiency evaluation, these uncertainties are treated as fully correlated between bins. Since the efficiencies are determined using different methods for muons and electrons these uncertainties are taken as uncorrelated between the dimuon and dielectron final states.
The uncertainties on the purity estimates described in Sect.~\ref{sec:Selection} introduce uncertainties on the overall cross-sections of $0.2\%$ for muons and $0.5\%$ uncertainty for electrons, treated as correlated between all bins. For the muon analysis, the purity is assumed to be uniform across all bins. To evaluate the uncertainty associated with this assumption, the purity is allowed to vary in each bin, with the change from the nominal result providing an additional uncertainty at the per-mille level for the differential measurement.
The statistical uncertainty on the FSR corrections is treated as a systematic uncertainty on the corrections. This is combined in quadrature with the difference between the corrections derived using the \mbox{\textsc{Herwig++}}\xspace~\cite{GEN-HERWIGPP} and \pythia8~\cite{Sjostrand:2007gs,*Sjostrand:2006za} simulated datasets. The uncertainties on the FSR corrections are taken as uncorrelated between all bins.
The dimuon analysis is repeated using a momentum scale calibration and detector alignment determined from $\PZ\rightarrow\Pmu\Pmu$ events, in a similar approach to that documented in Ref.~\cite{LHCb-PAPER-2015-039}. The impact on the measured total cross-section and the differential $y_\PZ$ and \ensuremath{\phi^{*}_\eta}\xspace measurements is negligible. The mean effect in any bin of transverse momentum is typically 1\% and is not statistically significant. However this is assigned as an additional uncertainty on the differential cross-section in each bin of transverse momentum. While the \PZ boson transverse momentum distribution is not measured in the dielectron final state, the momentum scale plays a larger role in the analysis of the dielectron final state due to the significant effect of bremsstrahlung and migrations in electron \mbox{$p_{\rm T}$}\xspace across the 20~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace threshold. The impact of the scale around this threshold is evaluated in the same way as in previous $\PZ\rightarrow\Pe\Pe$ analyses at LHCb~\cite{LHCb-PAPER-2012-008,LHCb-PAPER-2015-003}. A fit to the $\min[\mbox{$p_{\rm T}$}\xspace(\Pe^+),\mbox{$p_{\rm T}$}\xspace(\Pe^-)]$ spectrum returns a momentum scale correction factor of $1.000\pm0.005$ for simulation. Propagating this uncertainty on the electron momentum scale onto the cross-section measurement yields an uncertainty of about $0.6\%$, which is treated as correlated between all bins.
\label{sec:Systematics}
\begin{table}[!t]
\begin{center}
\caption{Summary of the relative uncertainties on the \PZ boson total cross-section.}
\label{tab:systsum}
\begin{tabular}{lcc}
\textbf{Source} & \multicolumn{1}{c}{$\Delta\ensuremath{\sigma_{\PZ}}\xspace^{\Pmu\Pmu}$ [\%]} & \multicolumn{1}{c}{$\Delta\ensuremath{\sigma_{\PZ}}\xspace^{\Pe\Pe} [\%]$} \\ \hline
Statistical & 0.5 & 0.9\\\hline
Reconstruction efficiencies & 2.4& 2.4\\
Purity & 0.2 & 0.5\\
FSR & 0.1 & 0.2\\\hline
Total systematic (excl. lumi.) & 2.4 & 2.5\\
Luminosity & {3.9} & 3.9\\
\end{tabular}
\end{center}
\end{table}
The transverse momentum distribution is unfolded to account for potential migration of events between bins arising from the experimental resolution using correction factors in each bin. A systematic uncertainty on this approach is set by considering the Bayesian method~\cite{DAGOSTINI,ROOUNFOLD} with two iterations as an alternative. The difference between the two approaches is at the per-mille level in each bin and is assigned as the uncertainty. As in previous analyses~\cite{LHCb-PAPER-2015-001,LHCb-PAPER-2015-049}, the unfolding is studied using different models of the underlying distribution, and no significant additional variation is observed.
The only uncertainty treated as correlated between the muon and electron final states is the one associated with the luminosity determination. This uncertainty is determined to be 3.9\% following the procedures used in Ref.~\cite{LHCb-PAPER-2014-047}. The uncertainty on the FSR correction may also be correlated, but is sufficiently small for the effects of such correlation to be negligible. The measurement is performed for the nominal centre-of-mass energy of the colliding beams. This energy was determined to an accuracy of 0.65\% for the 4\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace proton beams used in earlier LHC operations~\cite{BEAM}. No studies have yet been published for the 6.5~\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace proton beams used here, but for calculations performed using the FEWZ generator~\cite{GEN-FEWZ3} at NNLO in pQCD, a 0.65\% shift in the beam and collision energy would correspond to a shift in the fiducial cross-section of 0.9\%. This is not assigned as an additional uncertainty. The correlation matrices for the measurements of the differential cross-section as a function of the \PZ boson rapidity are given in the \hyperref[appendix_results]{appendix}.
\section{Results}
\label{sec:Results}
The inclusive \PZ boson cross-section for decays to a dilepton final state with the dilepton invariant mass in the range $60 < m(\ell\ell) < 120$~\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace, and where the leptons have $p_\text{T}\nolinebreak>\nolinebreak20$\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace and $2.0<\eta<4.5$, is measured in $\sqrt{s} = 13\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ pp collisions to be
\begin{align*}
\ensuremath{\sigma_{\PZ}}\xspace^{\Pmu\Pmu} &= 198.0 \pm 0.9 \pm 4.7 \pm 7.7\ensuremath{\rm \,pb}\xspace,\\
\ensuremath{\sigma_{\PZ}}\xspace^{\Pe\Pe} &= 190.2 \pm 1.7 \pm 4.7 \pm 7.4\ensuremath{\rm \,pb}\xspace.
\end{align*}
\noindent The first uncertainties quoted are statistical, the second arise from systematic effects, and the third are due to the accuracy of the luminosity determination. This cross-section is determined at the Born level in QED. Taking the luminosity uncertainty to be fully correlated, the two measurements are consistent at the level of $1.1\,\sigma$, and are linearly combined to give
\begin{align*}
\ensuremath{\sigma_{\PZ}}\xspace^{\ell\ell} &= 194.3 \pm 0.9 \pm 3.3 \pm 7.6\ensuremath{\rm \,pb}\xspace,
\end{align*}
where the combination minimises the sum of the statistical and systematic uncertaintes in quadrature. The integrated cross-section in the fiducial acceptance and the differential measurement as a function of the \PZ boson rapidity are compared in Figs.~\ref{fig:totZ} and~\ref{fig:rapZ} to the fixed-order predictions for both dimuon and dielectron final states. The measured differential cross-sections are tabulated in the \hyperref[appendix_results]{appendix}. Fixed-order predictions describe the LHCb data well for a range of PDF sets. The measured differential cross-section is slightly larger than the next-to-next-to-leading order pQCD predictions at lower rapidities, in line with observations in Ref.~\cite{CMS-PAS-SMP-15-011}. The differences between the PDF sets, and the PDF uncertainties, are larger than those at lower values of $\sqrt{s}$. Larger LHCb datasets with the uncertainty on the luminosity determination reduced to the level of previous studies (1.2\%) should significantly constrain the PDFs. The differential cross-sections as a function of \mbox{$p_{\rm T}$}\xspace and \ensuremath{\phi^{*}_\eta}\xspace, normalised to the total cross-section, are shown in Figs.~\ref{fig:psZ},~\ref{fig:ptZ} and~\ref{fig:mm}. Since the largest systematic effects are independent of these variables, systematic uncertainties largely cancel when these distributions are normalised, and the uncertainties on the normalised distributions are dominated by the statistical components. The LHCb data agree better with \pythia8 predictions than with \mbox{\textsc{Powheg}}\xspace+~\mbox{\textsc{Pythia}}\xspace\nolinebreak8 predictions, as seen also in previous analyses~\cite{LHCb-PAPER-2015-001,LHCb-PAPER-2012-036}. The LHCb specific tune of \pythia8 does not describe the data significantly better than the Monash 2013 tune. In addition, the data do not favour a particular matching and merging scheme generated using {\textsc{MadGraph5}}\_aMC@NLO.
\section{Conclusions}
\label{sec:Conclusions}
The \PZ production cross-section measured in pp collisions at $\sqrt{s} = 13$\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace is presented using LHCb events where the $\PZ$ boson decays to two muons or two electrons. The cross-section is measured in a fiducial acceptance defined by lepton pseudorapidity in the range $2.0 < \eta < 4.5$, transverse momentum $\mbox{$p_{\rm T}$}\xspace > 20\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$, and dilepton invariant mass in the range $60<m(\ell\ell)<120\ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace$. The cross-section is measured to be
\begin{align*}
\ensuremath{\sigma_{\PZ}}\xspace^{\ell\ell} &= 194.3 \pm 0.9 \pm 3.3 \pm 7.6\ensuremath{\rm \,pb}\xspace,
\end{align*}
\noindent where the uncertainties are due to the size of the dataset, systematic effects, and the luminosity determination respectively.
In addition, the measurement is performed in bins of the \PZ boson rapidity, transverse momentum and \ensuremath{\phi^{*}_\eta}\xspace. The measurement is compared to theoretical predictions calculated at $\mathcal{O}(\alpha_{s}^2)$ in pQCD as a function of the boson rapidity. The results do not favour any specific parton distribution function, but the differences between the PDF sets suggest that, with more data and a reduction in the uncertainty associated with the luminosity determination, LHCb results will significantly constrain the PDFs. The \ensuremath{\phi^{*}_\eta}\xspace and boson transverse momentum distributions are compared to theoretical predictions that model higher orders in pQCD in different ways. No significant deviations are seen between the data and the Standard Model.
\pagebreak
\begin{figure}[!b]
\begin{center}
\includegraphics[width=.75\textwidth]{figures/Fig1.pdf}
\caption{The fiducial cross-section compared between theory and data. The bands correspond to the average of the dimuon and dielectron final states, with the inner band corresponding to the statistical uncertainty and the outer band corresponding to the total uncertainty. The top three points correspond to $\mathcal{O}(\alpha_{s}^2)$ predictions with different PDF sets. The inner error bars on these points are due to the PDF uncertainty, with the outer error bars giving the contribution of all uncertainties. The bottom points correspond to the LHCb measurements in the dielectron and dimuon final states and their average, with the inner error bar showing the statistical uncertainty and the outer error bar the total uncertainty.}
\label{fig:totZ}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=.75\textwidth]{figures/Fig2top.pdf}
\includegraphics[width=.75\textwidth]{figures/Fig2bottom.pdf}
\caption{The differential cross-section as a function of the \PZ boson rapidity, compared between theory and data. The bands correspond to the data, with the inner band corresponding to the statistical uncertainty and the outer band corresponding to the total uncertainty. The points correspond to $\mathcal{O}(\alpha_{s}^2)$ predictions with different PDF sets. The inner error bars on these points are due to the PDF uncertainty, with the outer error bars giving the contribution of all uncertainties. The different predictions are displaced horizontally within bins to enable ease of comparison. The upper plot shows the differential cross-section, and the lower plot shows the same information as ratios to the central values of the NNPDF3.0 predictions.}
\label{fig:rapZ}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=.75\textwidth]{figures/Fig3top.pdf}
\includegraphics[width=.75\textwidth]{figures/Fig3bottom.pdf}
\caption{The normalised differential cross-section as a function of the \PZ boson $\phi^{*}_\eta$, compared between theory and data. The bands correspond to the data, with the inner band corresponding to the statistical uncertainty and the outer band corresponding to the total uncertainty. The points correspond to the theoretical predictions from the different generators and tunes. The different predictions are displaced horizontally within bins to enable ease of comparison. The upper plot shows the normalised differential cross-section, and the lower plot shows the same information as ratios to the central values of the predictions produced using the Monash 2013 tune of \pythia8. The uncertainties on the theoretical predictions, visible at high \ensuremath{\phi^{*}_\eta}\xspace, are statistical.}
\label{fig:psZ}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=.75\textwidth]{figures/Fig4top.pdf}
\includegraphics[width=.75\textwidth]{figures/Fig4bottom.pdf}
\caption{The normalised differential cross-section as a function of the \PZ boson transverse momentum, compared between theory and data. The bands correspond to the data, with the inner band corresponding to the statistical uncertainty and the outer band corresponding to the total uncertainty. The points correspond to the theoretical predictions from the different generators and tunes. The different predictions are displaced horizontally within bins to enable ease of comparison. The upper plot shows the normalised differential cross-section, and the lower plot shows the same information as ratios to the central values of the predictions produced using the Monash 2013 tune of \pythia8.}
\label{fig:ptZ}
\end{center}
\end{figure}
\begin{figure}[!t]
\begin{center}
\includegraphics[width=.75\textwidth]{figures/Fig5top.pdf}
\includegraphics[width=.75\textwidth]{figures/Fig5bottom.pdf}
\caption{The ratio of the normalised differential cross-sections to the predictions evaluated using the FxFx scheme. The bands correspond to the data, with the inner band corresponding to the statistical uncertainty and the outer band corresponding to the total uncertainty. The different predictions are displaced horizontally within bins to enable ease of comparison. Alternative schemes give different predictions, shown as points.All predictions are generated using {\textsc{MadGraph5}}\_aMC@NLO. The uncertainties on the theoretical predictions are statistical. The upper plot shows the $\ensuremath{\phi^{*}_\eta}\xspace$ distribution, and the lower plot shows the \mbox{$p_{\rm T}$}\xspace distribution.}
\label{fig:mm}
\end{center}
\end{figure}
|
\section{Introduction}
\label{sec:intro}
Interpolating signals defined on graphs is a basic problem that has found numerous applications in a variety of fields such as sensor networks, data classification, brain-imaging and recommendation systems \cite{hagmann2008mapping, sandryhaila2013discrete, sandryhaila2014discrete,hoche2008fast} The aim of interpolation is to find missing values of a graph signal from its values on a subset of the nodes assuming a particular signal model e.g. band limitedness in the GFT (Graph Fourier Transform) domain, Sparsity, and etc. Different algorithms have been proposed for this problem so far \cite{narang2013signal, narang2013localized,segarra2015interpolation,belkin2004semi,chen2009fast,grady2010anisotropic} (See \cite{shuman2013emerging}for an extensive review).
The K-Nearest Neighbor (KNN) method proposed in \cite{chen2009fast} is a basic technique that uses an efficient Lanczos procedure for recursive data partitioning and reconstructs the unknown signal values using a weighted combination of the known values on the k-nearest nodes \cite{grady2010anisotropic}. It is known that KNN overlooks the dependencies existing between the known samples. However, more computationally demanding algorithms have been proposed by \cite{narang2013signal,belkin2004semi,grady2010anisotropic} that take more similarity factors into account and thus provide more accurate estimates.
The method proposed by \cite{narang2013localized} shows improved performance regarding both accuracy and computational efficiency and serves as a benchmark for performance comparisons in this research. \cite{narang2013localized} proposes a Regularization Based Method (RBM) in order to minimize a cost function consisting of both signal smoothness and the square reconstruction error. Furthermore, it proposes the Iterative Least Square Reconstruction (ILSR) method for graph signal reconstruction based on band-limitedness. \cite{narang2013localized} also provides a comprehensive comparison between the performance of state-of-the-art interpolation methods for the application of recommendation systems working on three benchmark datasets of Movielens \cite{WinNT}, Jester \cite{WinNT2} and Books \cite{WinNT3}.
The Iterative Weighting Reconstruction (IWR) and Iterative Propagation Reconstruction (IPR) methods were proposed by \cite{wang2015local} to reconstruct band-limited graph signals by the idea of division to sub-graphs. Compared to ILSR in \cite{narang2013localized}, these methods achieve improved convergence rates, however the partitioning technique creates isolated local-sets which leads to reluctant sampling vertices.
In \cite{shuman2013emerging}, a three layer cluster division is proposed which is similar to \cite{wang2015local} but reduces the sampling rate by removing the isolated vertex sets.
\textbf{\textit{Contributions:}} As observed above, a major presumption that has been considered in many previous works on graph signal interpolation \cite{narang2013signal,narang2013localized,segarra2015interpolation,shuman2013emerging,anis2015asymptotic} is that the signal defined on the graph is band-limited and there are a few prior works that assume sparsity \cite{zhu2012graph,marques2016sampling}. In this work, we consider the graph signal to be sparse/compressible rather than band-limited in the GFT domain i.e. it has a few non-negligible coefficients spread along the whole GFT range without prior knowledge of their locations. We propose the Iterative Method with Adaptive Thresholding for Graph Interpolation (IMATGI) for sparsity promoting reconstruction of graph signals. We provide the convergence analysis for the proposed method and show its efficient performance by simulations. Another key contribution of this work is that we show (by extensive simulations on the benchmark datasets used by \cite{narang2013signal,narang2013localized},\cite{hofmann2005collaborative,ziegler2005improving,goldberg2001eigentaste} that applying the sparse signal assumption by IMATGI significantly improves the interpolation performance in the widely desirable application of recommendation systems. This observation brings us to the conclusion that the natural Movies, Jokes, Books and etc. datasets better match the sparse signal assumption rather than the classic band-limitedness.
\textbf{\textit{Notations}}: Throughout this paper, we denote scalar values and vectors by italic and regular lowercase letters, respectively. Matrices and sets are denoted by boldface and regular uppercase letters. Finally, calligraphic letters denote mathematical operators and $E\{.\}, (.)^t$ and $||.||_2$ are expected value, matrix transposition and the second vector norm, respectively.
The rest of this paper is organized as follows. Section \ref{sec:matmodel} introduces the proposed IMATGI algorithm. Section \ref{sec:converge} analytically discusses the reconstruction capability of IMATGI. Section \ref{sec:sim} includes the simulation results and performance comparisons and finally section \ref{sec:con} concludes the paper.
For further reproduction of the reported simulation results, MATLAB codes are made available on $ee.sharif.edu/\sim boloursaz$.
\section{The proposed IMATGI algorithm}
\label{sec:matmodel}
In this subsection, we present the proposed Iterative Method with Adaptive Thresholding for Graph Interpolation (IMATGI) algorithm. This algorithm assumes that the underlying graph signal is sparse/compressible in the Graph Fourier Domain (GFT) and gradually extract the significant signal components by iterative thresholding of the estimated signal with a decreasing threshold. This technique is inspired by the previous findings on sparsity promoting reconstruction of regular signals from missing samples \cite{marvasti2012unified}.
Consider an undirected graph $G=(V,E)$ with $V$ as the set of vertices and $E$ as edges. Denote by $\textbf{L}$ the symmetric normalized Laplacian matrix for this graph as defined by \cite{narang2013localized}. Now, decompose $\mathbf{L=U\Lambda{U}^t}$ in which $\mathbf{\Lambda}=\mathop{\mathrm{diag}}({\mathit{\lambda}_{1},\mathit{\lambda}_{2},\cdots,\mathit{\lambda}_{n}})$ is a diagonal matrix of non-negative eigenvalues and $\mathbf{U}=[{u}_1,{u}_2,\cdots,{u}_n]$ is a unitary matrix containing the corresponding eigenvectors.
Now, define the corresponding graph signal as a function $\text f:V\rightarrow R$ and denote it by the vector $\text f\in R^N $where the $i$th component represents the signal value on the $i$th vertex. Considering the eigenvectors $u_i$ as the basis vectors and the corresponding eigenvalues $\lambda_i$ as the graph frequencies (as defined by [12]), this signal can be transformed into the Graph Fourier Transform (GFT) domain by ${\hat{{\text f}}=\textbf{U}^t\text f}$.
In the graph interpolation problem, the signal entries are known only on a subset of nodes $\textbf{S}$ and we aim to interpolate the unknown signal values on $\textbf{S}^c$. Define the diagonal sampling matrix $\mathbf{S_{N\times N}}=\mathop{\mathrm{diag}}(s_{1},s_{2},\cdots,s_{N})$ in which $i$th diagonal element is defined by:
\begin{eqnarray}
\label{eq:sol_1}
s_i= \bigg\{{\genfrac{}{}{0pt}{}{1\quad if\quad i\in \textbf{S}}{0\quad if \quad i\not\in \textbf{S}}}
\end{eqnarray}
Hence, the sub-sampled signal is given by $\text f_s=\textbf{S}\text f$. Utilizing this notation, the proposed IMATGI algorithm is presented in Table~\ref{tab:t1}. The idea of reconstructing sparse graph signals by sequential thresholdings according to the IMATGI update rule (\ref{eq:sol}) is a rational guess that is inspired by the Iterative Hard Thresholding (IHT) \cite{IHT1} and Iterative Method with Adaptive Thresholding (IMAT) \cite{marvasti2012unified} algorithms in the literature of sparsity promoting reconstruction of regular signals from missing samples.
In Table~\ref{tab:t1}, $\text f$ and $\text f_k$ denote the original signal and its reconstructed version at the $k$th algorithm iteration. $\mathit{\lambda}$ is the relaxation parameter that controls the convergence rate of the algorithm and $\mathcal T(.)$ denotes the thresholding operator.
The thresholding block operates elementwise on the input vector and sets the vector entries with absolute values below the threshold to zero. The threshold value $t(k)$ is decreased exponentially by $t(k)=\beta e^{-\alpha k}$ where $k$ is the iteration number. The algorithm parameters $\lambda$,$\beta$,$\alpha$ are optimized empirically for fastest convergence.
\begin{table}[H]
\centering
\caption {Stepwise presentation for IMATGI } \label{tab:t1}
\begin{tabular}{|l|l|}
\hline
\multicolumn{1}{|c|}{IMATGI Algorithm for Sparse Signal Reconstruction on Graphs} \\
\hline \\
\textbf {Require: $G0 = (V,E)$} \\
\quad~~\llap{\textbullet}~~{Compute normalized Laplacian matrix} \\
\textbf{Inputs}: \\
\quad~~\llap{\textbullet}~~ $\mathbf{S_{N\times N}}$: The sampling matrix\\
\quad~~\llap{\textbullet}~~ \textit{$\epsilon$}: Stopping criteria\\
\quad~~\llap{\textbullet}~~ $\mathbf{(\alpha,\beta,\lambda)}$: Algorithm Parameters \\
\textbf{Output}: \\
\quad~~\llap{\textbullet}~~ {$ {\tilde{\text{f}}_{N\times 1}}$}: The reconstructed signal\\
\textbf{Algorithm}: \\
\quad~~\llap{\textbullet}~~ Initialization {$\tilde{\text{f}} =\text f_1=\textbf{S}\text {f},\text {f}_0=0_{N \times 1}, k=1$}\\
\quad~~\llap{\textbullet}~~ While $(||\text{f}_k-\text{f}_{k-1}||>\epsilon)$ \\
\quad \quad \quad -~Calculate the threshold as: $t(k)=\beta e^{-\alpha k}$ \\
\quad \quad \quad -~Perform the thresholding as: $\text{g}_k=\textbf{U}(\mathcal{T}(\textbf{U}^t\text{f}_k))$ \\
\quad \quad \quad -~Perform the recursion as:\\
\begin{tabular}{p{8cm}}
\tableequation
{\quad \quad\quad \quad \quad \text{f}_{k+1}=(\textbf{I}_{N\times N}-\lambda\textbf S)\text{g}_k+\lambda \text{f}_s \label{eq:sol}} \\
\end{tabular}\\
\quad \quad \quad -~ ${\tilde{\text{f}}}=\text{f}_{k+1} $ \\
\quad \quad \quad -~$k=k+1$ \\
\quad~~\llap{\textbullet}~~ End While\\
\hline
\end{tabular}
\end{table}
\section{CONVERGENCE ANALYSIS}
\label{sec:converge}
In this section, we discuss convergence of the proposed IMATGI algorithm. To proceed, we need to prove the following lemma.
\textbf{Lemma 1}: Let\textquotesingle s denote the GFT of the sub-sampled graph signal by $\hat{\text{f}}_s=\textbf U^t\textbf S \text{f}$. Also, assume that the diagonal elements of the sampling matrix $\textbf{S}$ are independent identically distributed (iid) random variables coming from Bernoulli(p) distribution ($s_i\sim Bernoulli(p),\forall i$). We have:
\begin{align}
\label{eq:sol_2}
\quad\quad \quad \quad E\{\hat {\text{f}}_s\} &= p\hat{\text{f}}\nonumber\\
E\{trace((\hat {\text{f}}_s-p\hat {\text{f}})(\hat {\text{f}}_s-p\hat {\text{f}})^t)\}&=(p-p^2)\epsilon
\end{align}
in which $\epsilon$ is the energy of the graph signal defined as $ \epsilon={\text{f}}^t {\text{f}}$.
\begin{proof} [\textbf{Proof:}\nopunct]
For the first equation we have:
\begin{eqnarray}
\label{eq:sol_3}
E\{\hat {\text{f}}_s\}= E\{\mathbf{U^t}\mathbf{S}{\text{f}}\}=\mathbf U^tE\{\textbf{S}\}{\text{f}}=\mathbf{U^t}(p\textbf{I}){\text{f}}=p\mathbf{U^t}{\text{f}}=p\hat {\text{f}}
\end{eqnarray}
For the second equation we write (\ref{eq:sol_3}):
\begin{eqnarray}
\label{eq:sol_4}
\hat {{\text{f}}_s}-p\hat {\text{f}}= \mathbf{U^t}\mathbf{S}{\text{f}}-p\mathbf{U^t}{\text{f}}=\mathbf{U^t}(\textbf{S}-p\textbf{I}){\text{f}}
\end{eqnarray}
Now substituting (\ref{eq:sol_4}) we get:
\begin{align}
\label{eq:sol_5}
&E\{trace((\hat {\text{f}}_s-p\hat {\text{f}})(\hat {\text{f}}_s-p\hat {\text{f}})^t)\}\nonumber\\&=E\{trace(\mathbf{U^t}(\mathbf {S}-p\textbf{I}){\text{f}}{\text{f}}^t(\mathbf {S}-p\textbf{I})^t\mathbf{U})\}\nonumber\\&=E\{trace((\mathbf {S}-p\textbf{I}){\text{f}}{\text{f}}^t(\mathbf {S}-p\textbf{I})^t)\}\nonumber\\&=trace(E\{(\mathbf {S}-p\textbf{I})^t(\mathbf {S}-p\textbf{I}){\text{f}}{\text{f}}^t\})
\end{align}
Now note that $\textbf{S}-p\textbf{I}$ is a diagonal matrix with $E\{(\textbf{S}-p\textbf{I})^t (\textbf{S}-p\textbf{I})\}=p(1-p)\textbf{I}$, hence we have:
\begin{align}
\label{eq:sol_6}
&E\{trace((\hat {\text{f}}_s-p\hat {\text{f}})(\hat {\text{f}}_s-p\hat {\text{f}})^t)\}\nonumber\\&=p(1-p) trace({\text{f}}{\text{f}}^t)=p(1-p)\epsilon
\end{align}
\end{proof}
\textbf{Theorem 1}: Under the assumptions of Lemma 1 (i.e. $s_i\sim Bernoulli(p),\forall i$), and considering the IMATGI reconstruction formula given in Table~\ref{tab:t1}, $\lim_{k\to\infty} \hat {\text{f}}_k$is an unbiased estimator of $\hat {\text{f}}$ for $0<\lambda<2/p$.
\begin{proof} [\textbf{Proof:}\nopunct]
To prove this theorem, we need to show that $E\{\lim_{k\to\infty}\hat {\text{f}}_k\}=\hat {\text{f}} $ or equivalently $\lim_{k\to\infty}E\{\hat {\text{f}}_k\}=\hat {\text{f}}$. ̂To this end, we define the sequence of error vectors $e_k=\hat {\text{f}}-E\{\hat {\text{f}}_k\}$ and show that each element of $e_k$ forms a geometric progression with common ratio $r=1-\lambda p$. Hence, if $0<\lambda<2/p$ the IMATGI reconstruction technique converges linearly (of order 1) to the original graph signal in the mean.
Starting the algorithm from a zero initial condition, we have $\hat {\text{f}}_0=0$ and hence $e_0=\hat {\text{f}}$. Also from the basic IMATGI recursion (\ref{eq:sol}) we have $\hat {\text{f}}_1=\lambda \hat {\text{f}}_s $ and hence from Lemma 1 we get:
\begin{eqnarray}
\label{eq:sol_7}
\text{e}_1=\hat {\text{f}}-E\{\hat {\text{f}}_1\}=(1-\lambda p)\hat {\text{f}}
\end{eqnarray}
Now note that the basic IMATGI recursion can be rewritten in transform domain as:
\begin{eqnarray}
\label{eq:sol_8}
\hat {\text{f}}_{k+1}=\textbf U^t(\textbf I_{N \times N}-\lambda \textbf S){\text{g}}_k+\lambda \hat {\text{f}}_s
\end{eqnarray}
Taking expected value from both sides of (\ref{eq:sol_8}) and utilizing Lemma 1 yields:
\begin{align}
\label{eq:sol_9}
E\{\hat {\text{f}}_{k+1}\}&=E\{\textbf U^t(\textbf I_{N \times N}-\lambda \textbf S){\text{g}}_k\}+\lambda p \hat {\text{f}} \nonumber\\&=\textbf U^t (1-\lambda p )\textbf I E\{\textbf U (\mathcal{T}(\textbf U^t {\text{f}}_k))\}+\lambda p \hat {\text{f}}\nonumber\\&=(1-\lambda p)E\{\mathcal{T}(\textbf U^t {\text{f}}_k)\}+\lambda p \hat {\text{f}}
\end{align}
Utilizing (\ref{eq:sol_9}) we get:
\begin{align}
\label{eq:sol_10}
\text{e}_{k+1}(i)&=\hat {\text{f}}-E\{\hat {\text{f}}_{k+1}\} \nonumber\\&=(1-\lambda p)\hat {\text{f}}-(1-\lambda p)E\{\mathcal{T}(\textbf U^t {\text{f}}_k)\}\nonumber\\&=(1-\lambda p)(\hat {\text{f}}-E\{\mathcal{T}(\hat {\text{f}}_k)\})
\end{align}
Now let\textquotesingle s take an elementwise look at the final equation (\ref{eq:sol_10}). Denote the $i$th element of the original signal, the estimated signal and the error vector by $\hat {\text{f}}(i),\hat {\text{f}}_k(i)$ and $e_{k+1} (i)$, respectively. We get (\ref{eq:sol_11}):
\begin{eqnarray}
\label{eq:sol_11}
\text{e}_{k+1}(i)=(1-\lambda p)(\hat {\text{f}}(i)-E\{\mathcal{T}(\hat {\text{f}}_k(i))\})
\end{eqnarray}
Now if $|\hat {\text{f}}_k(i)|\geqslant t(k)$, this element successfully passes the threshold. In this case we can omit the thresholding operator from the right side of (\ref{eq:sol_11}) and we get (\ref{eq:sol_12}):
\begin{align}
\label{eq:sol_12}
\text{e}_{k+1}(i)&=(1-\lambda p)(\hat {\text{f}}(i)-E\{\hat {\text{f}}_k(i)\})\nonumber\\&=(1-\lambda p)\text{e}_{k}(i)
\end{align}
On the other hand, if $|\hat {\text{f}}_k(i)|<t(k)$ then $\hat {\text{f}}_k(i)$ does not pass through the threshold and we have $E\{\mathcal{T}(\hat {\text{f}}_k)\}=0$ and it is obvious from (\ref{eq:sol_10}) that:
\begin{eqnarray}
\label{eq:sol_13}
\text{e}_{k+1}(i)=(1-\lambda p)\hat {\text{f}}(i)
\end{eqnarray}
Hence, once a vector element passes through the threshold in a specific iteration, its corresponding error sequence converges linearly to zero provided that $0<\lambda<2/p$. As the threshold is strictly decreasing and approaches zero as $k\rightarrow \infty$, all vector elements will eventually pass through the threshold and the proof is complete.
\end{proof}
In order to guarantee perfect reconstruction/convergence of the IMATGI algorithm, we also need to show that the variance of this unbiased estimator approaches zero as $k\rightarrow \infty$. Theorem 2 explains this variance fluctuation issue as $k$ approaches infinity. Before proceeding to the formal statement of Theorem 2, let’s define the support for the sparse graph signal $ {\text{f}}$ as the set of all non-zero elements in its GFT representation as $Supp=\{j|\hat {\text{f}}(j)\neq 0\}$.
\textbf{Theorem 2}: Under the assumptions of Lemma 1 (i.e. $s_i\sim Bernoulli(p),\forall i$), if the GFT component $\hat {\text{f}}_k(i) $ passes through the threshold in the $k$th iteration of the IMATGI algorithm, this decreases the estimation variance defined as $\sigma_k^2=E\{trace((\hat {\text{f}}_k -E\{\hat {\text{f}}_k\})(\hat {\text{f}}_k-E\{\hat {\text{f}}_k\})^t)\}$ if $i\in Supp$ and increases the variance for $i\not\in Supp$
\begin{proof} [\textbf{Proof:}\nopunct]
Let\textquotesingle s partition the set of all GFT components passed through the threshold at the $k$th iteration as $Supp_k= {\text{Q}}_k\bigcup {\text{L}}_k$ in which $ {\text{Q}}_k$ represents the set of GFT components present in the original signal support (Supp) and $ {\text{L}}_k$ denotes the rest. Correspondingly, decompose $ {\text{g}}=\textbf U(\mathcal{T}(\textbf U^t {\text{f}}_k ))$ as:
\begin{eqnarray}
\label{eq:sol_14}
{\text{g}}_k= {\text{q}}_k+ {\text{l}}_k
\end{eqnarray}
In which $ {\text{q}}_k$ is the portion due to the support components and $ {\text{l}}_k$ is due to the non-support portion passed mistakenly through the threshold. Similarly, let\textquotesingle s decompose $ {\text{f}}$ as the sum of its reconstructed portion $ {\text{q}}_k$ and a residual $ {\text{r}}_k$ as
\begin{eqnarray}
\label{eq:sol_15}
{\text{f}}= {\text{q}}_k+ {\text{r}}_k
\end{eqnarray}
Now, substituting (\ref{eq:sol_14}) and (\ref{eq:sol_15}) in (\ref{eq:sol}) gives:
\begin{align}
\label{eq:sol_16}
& {\text{f}}_{k+1}=(\textbf I_{N\times N}-\lambda \textbf S) {\text{g}}_k+\lambda {\text{f}}_s\nonumber\\&=(\textbf I_{N\times N}-\lambda \textbf S)( {\text{q}}_k+ {\text{l}}_k)+\lambda \textbf{S}( {\text{q}}_k+ {\text{r}}_k)\nonumber\\&=\lambda \textbf{S} {\text{r}}_k-\lambda \textbf{S} {\text{l}}_k+ {\text{q}}_k+ {\text{l}}_k
\end{align}
The last two terms in (\ref{eq:sol_16}) $( {\text{q}}_k \ and \ {\text{l}}_k)$ are not sub- sampled and hence do not contribute to the estimation variance $\sigma_{k+1}^2$. Utilizing Lemma 1, we can compute $\sigma_{k+1}^2$ due to the sub-sampled terms by (\ref{eq:lost})
\begin{align}
\label{eq:lost}
\sigma_{k+1}^2&=E\{trace((\hat {\text{f}}_{k+1}-E\{\hat {\text{f}}_{k+1}\})(\hat {\text{f}}_{k+1}-E\{\hat {\text{f}}_{k+1}\})^t)\}\nonumber\\&=\lambda^2(p-p^2)\epsilon_{ {\text{l}}_k}+\lambda^2(p-p^2)\epsilon_{ {\text{r}}_k}
\end{align}
In which $\epsilon_{ {\text{r}}_k}= {\text{r}}_k^t {\text{r}}_k$ and $\epsilon_{ {\text{l}}_k}= {\text{l}}_k^t {\text{l}}_k$ denote the energies of the residual and the portion due to the non-support components mistakenly passed through the threshold. As each mistakenly passed component $i\not\in Supp$ increases $\epsilon_{ {\text{l}}_k}$, it will consequently increase the spectrum variance. Similarly, for a correctly passed signal component $i\in Supp$, $\epsilon_{ {\text{r}}_k}$ and consequently the spectrum variance $\sigma_{k+1}^2$ is decreased. The above discussion completes the proof.
\end{proof}
\textbf{Remark 1}: As stated previously, due to the non-zero spectrum variance, $\hat {\text{f}}_k(i)$ is generally non-zero for $i\not\in Supp$. Hence, the threshold parameters must be adjusted such that the threshold value always keeps above the standard deviation at the $k$th iteration (e.g. $t(k)\geq \gamma\sigma_k$, $\gamma>1$) to prevent the algorithm from picking up incorrect GFT components. In this case, $\epsilon_{ {\text{l}}_k}=0$ and the estimation variance is decreasing in each iteration $\sigma^2_{k+1}\leq\sigma_k^2$.
\textbf{Corollary 1}: Considering Theorem 1, we conclude that the IMATGI estimation bias approaches zero as k approaches infinity. On the other hand, the variance of the IMATGI estimation is decreasing provided that the condition in Remark 1 $(\epsilon_{ {\text{l}}_k}=0,\forall k)$ always holds. Now considering the fact that the Mean Square Error (MSE) of the estimator is given by (\ref{eq:sol_17})
\begin{align}
\label{eq:sol_17}
&MSE_k=E\{trace((\hat {\text{f}}_k-\hat {\text{f}})(\hat {\text{f}}_k-E\{\hat {\text{f}}\})^t)\}\nonumber\\&=trace((E\{\hat {\text{f}}_k\}-\hat {\text{f}})(E\{\hat {\text{f}}_k\}-\hat {\text{f}})^t)+\sigma_k^2
\end{align}
As both terms in (\ref{eq:sol_17}) are decreasing, we conclude that $MSE_k$ is also decreasing. In other words, if the condition in Remark 1 holds, the IMATGI algorithm strictly decreases the Mean Square Error (MSE) of the estimated signal in each iteration. As MSE is a convex cost function, these sequential decreases converge to the unique global minimum.
Concluding this section, we have shown that perfect reconstruction by the proposed IMATGI algorithm is possible if the threshold parameters ($\alpha, \beta$) are selected such that the condition in remark 1 is satisfied, $0<\lambda<2/p$ and $k\rightarrow \infty$.
\section{Simulation results}
\label{sec:sim}
In this section we demonstrate efficient performance of the proposed IMATGI algorithm by simulations on both randomly generated sparse signals and three benchmark data sets used in recommendation systems.\\
\textbf{A. Generic Sparse Signals} \\
In order to fairly evaluate the performance of the proposed algorithm, we calculate and report the reconstruction SNR as (\ref{eq:sol_18}):
\begin{eqnarray}
\label{eq:sol_18}
SNR=\frac{\vert\vert {\text{f}}\vert\vert^2_2}{\vert\vert {\text{f}}-\hat {\text{f}}\vert\vert^2_2}
\end{eqnarray}
where $ {\text{f}}$ and $\hat {\text{f}}$ denote the original and reconstructed graph signals, respectively.
In this simulation scenario, we generate a graph with $N=1000$ randomly located nodes and edges similar to \cite{narang2013localized}. The signal entries $ {\text{f}}(i)$ associated with each node are taken from the uniform random variable $U(0,1)$.
Now, define the number of sparse GFT components as $k$. In order to enforce sparsity of the generated signal in the GFT domain, we project the random signal onto the GFT domain $(\hat {\text{f}}=\textbf{U}^t {\text{f}})$, keep k entries with largest absolute values and set all the other GFT components to zero.
To study the reconstruction performance of the proposed algorithm, we randomly sub-sample these generic $k$-sparse signals utilizing the sampling matrix $\textbf{S}$ at rates ranging from p=0.45 to p=0.65. We sweep the sparsity factor $\frac{k}{N}$ from 10\% to 60\%. For each sparsity factor and sampling rate, we repeatly generate 100 random $k$-sparse signals, sample randomly at rate $p$, reconstruct using 20 iterations of the proposed IMATGI algorithm and report the average achieved SNR in Fig.\ref{fig:sysmodel31}.
As observed in Fig.\ref{fig:sysmodel31}, all curves experience a sudden knee-like fall in reconstruction SNR as the sparsity factor increases. This fall is considered as the boundary between successful and unsuccessful reconstruction. As expected, the simulation results reveal that as the sampling rate increases, the algorithm can successfully reconstruct less sparse signals.
Note that as the number of iterations are bounded to 20 and the threshold parameters are not ideally trained in order to guarantee that Remark 1 holds, perfect reconstruction (infinite SNR value) is not observed in these simulations even on the left hand of the knee-like fall.\\
\textbf{B. Recommendation Systems } \\
In this scenario we compare the performance of the proposed IMATGI algorithm with the previously proposed graph interpolation methods in the widely desirable application of recommendation systems. To this end, we apply IMATGI on three benchmark datasets widely used for performance evaluations in recommendation systems \cite{WinNT,WinNT2,WinNT3}. To have a fair comparison between the performances of different methods, we report the normalized reconstruction RMSE values achieved (as defined by \cite{narang2013localized}) in Table~\ref{tab:t2}.
Following an approach similar to \cite{narang2013localized}, each dataset is reduced to a $100K$ randomly selected user-item sub-dataset and split into 5 fold cross-validation sets. In each iteration we use four subsets for training (i.e. is computing the graph and signal values) and the last subset for testing the performance of the algorithm \cite{narang2013localized}.
Table~\ref{tab:t2} reports the RMSE values achieved by the proposed IMATGI algorithm along with the previously reported results for the other methods. As observed in this table, IMATGI improves the reconstruction performance in comparison with the literature. This is due to the fact that IMATGI utilizes the more general assumption of sparsity rather than bandlimitedness of the underlying graph signals. In fact, in this scenario, we observe that the real signals that arise in the application of recommendation systems are rather sparse than bandlimited (i.e. they have a few non-zero GFT components that may be located far apart from each other rather than condensed in a specific spectral range.)
\begin{figure}[H]
\centering
\includegraphics[scale=.7]{./3.jpg}
\caption{ The Reconstruction Performance for IMATGI}
\label{fig:sysmodel31}
\end{figure}
\section{Conclusion}
\label{sec:con}
In this paper we proposed the Iterative Method with Adaptive Thresholding for Graph Interpolation (IMATGI) algorithm for sparsity promoting interpolation of signals defined on graphs. We provided a formal convergence analysis for the proposed IMATGI algorithm and finally demonstrated its efficient reconstruction performance on both generic sparse data and the benchmark datasets for recommendation systems.
\begin{table}[!h]
\tiny
\caption {RMSE Performance Comparison between Different Graph Interpolation Techniques for Recommendation Systems} \label{tab:t2}
\centering
\begin{tabular}{ |p{.85cm}|p{.65cm}|p{.65cm}|p{.65cm}|p{.65cm}|p{.65cm}|p{.65cm}|p{.65cm}| }
\hline
Dataset& KNN & PMF & RBM & IRBM & LSR & ILSR & IMATGI\\ \hline
Movielens\cite{WinNT} & 0.2482 & 0.2513 & 0.2414 & 0.2450 & 0.2514 & 0.2466 & \textbf{0.2406} \\ \hline
Jester\cite{WinNT2} & 0.2348 & 0.2299 & 0.2304 & 0.2341 & 0.2344 &0.2315&\textbf{0.2130 } \\ \hline
BX-books\cite{WinNT3} & 0.2677 &0.2093 & 0.1966 & 0.2138& 0.2651&0.2828&\textbf{0.1790} \\\hline
\end{tabular}
\end{table}
\section{Acknowledgment}
Mahdi Boloursaz Mashhadi is supported by grants from Sharif University of Technology (SUT) and the Iranian National Science Foundation (INSF).
\bibliographystyle{ieeetr}
|
\section{Introduction}\label{intro}
The standard model of the particle physics~(SM) can explain
most of results obtained from experiments.
However, the smallness of the electroweak scale is one of the theoretical mysteries when one compares it with the Planck scale.
If SM is the complete theory to describe our nature, it tells us that the nature might be unnatural because SM itself does not answer why the Higgs mass is tuned to the electroweak scale against quantum corrections,
which are induced through the fundamental scale such as the Planck scale.
Even if the Higgs mass is protected by some symmetry, such as supersymmetry,
some level of tuning against the mass is
required below the symmetry breaking scale.
Recently, a novel mechanism, called the cosmological relaxation, has been proposed in~\cite{Graham:2015cka}, which solves the hierarchy problem by the dynamics of an additional scalar field called relaxion. The relaxion has a (discrete) shift symmetry, but it is explicitly broken at a scale $M$ larger than the electroweak scale with a dimensionless coupling $g$. In this scenario, the mass of the Higgs field is
dependent of the relaxion and is
dynamically determined by the relaxion field. Further, the relaxion potential depends on the Higgs field value
through the instanton, i.e, through the coupling between the Higgs and
(hidden) quark-pair condensation scale $\Lambda_c$,
which is chiral-rotated by the relaxion-dependent phase.
By this reciprocal relation with the relaxion, the Higgs vacuum expectation value (VEV) is set to the electroweak scale in the early period of the inflation in the original literature.
The relaxion mechanism is attractive to solve the hierarchy problem, but it requires some severe constraints on the inflation scenario. To fix the field value of the relaxion by the periodic potential coming from the instanton during inflation, the Hubble expansion rate should be small. For instance, in a benchmark model,
we find $H_{\rm inf}\sim\mathcal{O}(10^{-5})$ GeV~\cite{Graham:2015cka}, which is much smaller than the upper bound by the observations of the cosmic microwave background with $H_{\rm inf}\simeq10^{13}{\rm GeV}$~\cite{Ade:2015lrj}.
After the inflation, inflaton decays into the SM particles and then makes thermal bath with a reheating temperature $T_{\rm R}$. Since the periodic potential is made by the non-perturbative effects through (hidden) quarks, it disappears when the chiral symmetry of the quarks are restored at a temperature higher than $\Lambda_c$. Thus, for the relaxion to be fixed even after the inflation, naively $T_R$ needs to be lower than the condensation scale $\Lambda_c$. These requirements
to accomplish the relaxion mechanism severely restrict the possible inflation models.
In this paper, we discuss conditions
to relax the models from such constraints exhibited
in Ref.~\cite{Graham:2015cka}, assuming that
there exists relaxion during a high scale inflation.
Therefore, we generalize the scenario by considering
various possible interaction between relaxion and other sector including
(quantum) gravity.
Among of them, we especially take into account two effects: the Hubble induced mass and thermal one. We discuss whether the two masses change the relaxion dynamics and make the scenario compatible with the large Hubble expansion rate or high reheating temperature. In our scenario, the relaxion mechanism takes place after inflation, which is another different point from the original scenario~\cite{Graham:2015cka}.
The remaining of this paper is organized as follows.
In Sec.~\ref{Sec2}, we review the relaxion model briefly.
In Sec.~\ref{model}, we show the relaxion model with the Hubble mass and thermal one. Then, in Sec.~\ref{dyn}, we discuss the relaxion mechanism realized after inflation and note that scenarios are classified into two ways, depending on the time of the reheating. We discuss the relaxion dynamics in each case. Finally, we summarize this paper and discuss remaining issues in Sec.~\ref{concl}.
\section{Brief review of the relaxion model}\label{Sec2}
We consider a system with a scalar field $\phi$, called relaxion. The relaxion has a discrete symmetry under $\phi\to \phi+2\pi f$.
$f$ is the decay constant of $\phi$.
This discrete symmetry is broken, and its explicit breaking is characterized by a dimensionless constant $g$. In this paper, we assume that a breaking into the discrete symmetry occurs during or before inflation.
The potential for the Higgs field $h$ and the relaxion $\phi$
is given with the breaking parameters of the discrete symmetry by
\begin{align}\label{V}
V=&V(g\phi/M)+(M^2-gM\phi+\cdots)|h|^2\nonumber\\
&+\lambda|h|^4+\Lambda_c^3|h|\cos\left(\frac{\phi}{f}\right),
\end{align}
where $\lambda>0$ is the Higgs quartic coupling.
$M$ is the cutoff scale of our effective field theory, and this Higgs mass should be regarded as the value renormalized above the scale $M$.
In general, $M$ can be smaller than a fundamental scale of a theory,
such as the Planck scale or string one.
For example, in supersymmetric cases,
$M$ would be the supersymmetry breaking scale.
The potential $V(g\phi/M)$ is given by series of $g\phi/M$:
\begin{align}\label{lin}
V(g\phi/M)=-M^3g\phi+\frac{1}{2}M^2g^2\phi^2+\cdots.
\end{align}
The last term in (\ref{V}) comes from the fermion condensation
caused by a hidden strongly coupled gauge theory, whose dynamical scale is
denoted by $\Lambda_c$.\footnote
This condensation could be identified with the SM quark condensation in QCD.
However, in general, such a term may be originated from other hidden strong dynamics. In this paper, we only assume that the effective potential is proportional to the Higgs VEV $\langle h\rangle$.
}
We have assumed that this term is proportional to $h$ for simplicity.
In general, however, this term can depend on $h^2$ \cite{Graham:2015cka}.
This term becomes important to fix the relaxion value in the electroweak vacuum
as discussed below.
Note that the potential is natural in the view point of the shift symmetry~\cite{'tHooft:1979bh}:
We then find $g \lesssim 1$ and $\Lambda_c/M \lesssim 1$.
Due to the linear potential in (\ref{lin}), the relaxion slowly rolls down toward larger field value during inflation, and after long time, $\phi$ crosses the point $\phi\sim M/g$, where the Higgs mass becomes tachyonic. Then, the Higgs field acquires the VEV, and the last term in (\ref{V}) becomes effective. By the periodic term the relaxion is fixed, equivalently the Higgs mass is stabilized at the value much smaller than the scale $M$.
This mechanism is stable against quantum corrections,
because the parameter $g$ is expected to be small through the
naturalness argument.
Next, we focus on the condensation scale $\Lambda_c$. The hight of the periodic potential is proportional to
$4\pi f_{\pi'}^3m_N$. Here $f_{\pi'} (\sim \Lambda_c)$ is the chiral symmetry breaking scale and
$m_N $ is mass of the lightest fermion which
is charged under the strongly coupled gauge theory. Here, $m_N < f_{\pi'}$.
For the successful relaxion mechanism, the periodic potential of the relaxion should sensitively emerge as the Higgs field value evolves to around electroweak scale.
Thus, $m_N$ is required to be sensitively determined by the Higgs VEV.
According to argument on technical naturalness in Ref.~\cite{Graham:2015cka},
we set the condensation scale to around the electroweak scale as
\begin{equation}\label{eq:scale}
\Lambda_c\sim\braket{h}\sim{\cal O} (10^2)\,{\rm GeV}.
\end{equation}
\section{Relaxion model with the Hubble and thermal masses}\label{model}
As mentioned in Introduction, we consider the relaxion model (\ref{V}) in the scenario that the relaxation of the Higgs mass takes place {\it after inflation}. This is the essential difference from the original model~\cite{Graham:2015cka}, in which the relaxation mechanism is entirely realized {\it during inflation}.
In general, the initial value of the relaxion after inflation can be arbitrary due to the quantum fluctuation of the relaxion $\phi$, if $\phi$ is lighter than the Hubble scale during inflation $H_{\rm inf}$. Since the Higgs mass is determined by the field value of the relaxion as shown in (\ref{V}),
the fluctuations of $\phi$ leads to the inhomogeneity of the Higgs VEV among each patch of the Universe.
In order to avoid such a problematic situation, we consider the Hubble induced mass for the relaxion given by
\begin{align}\label{Hmass}
\Delta V=\frac{1}{2}c\,H^2(\phi^2-\phi_{H}(t))^2,
\end{align}
where $c>0$ is a dimensionless constant, $H$ is the Hubble parameter, and $\phi_H$ is a function of time $t$ in general.
If the coefficient is larger than unity as $c>1$,
the fluctuations of the relaxion during inflation are suppressed
by the Hubble induced mass. Then the relaxion value remains homogeneous
in the Universe. As a consequence isocurvature perturbation of the relaxion during inflation is suppressed \cite{Dine:2004cq}.
To realize our scenario, $\phi_H(t)$ should satisfy
\begin{align}
\phi_{H}(t)<\frac{M}{g}
\end{align}
not to overshoot the electroweak vacuum during and after inflation.
For simplicity, we assume $\phi_H(t)=0$ in this paper.
There are some possibilities to realize the Hubble induced mass~(\ref{Hmass})
through (quantum) gravity: A coupling between an inflaton and the relaxion
$\frac{V_{\rm inf}}{2M_*^2}\phi^2$ gives the term~(\ref{Hmass}) with
$c=\frac{3 M_{\rm pl}^2}{M_*^2}$ and $\phi_H(t)=0$,
where $V_{\rm inf}$ is the inflaton potential,
and $M_{\rm pl}(\sim 2.4\times 10^{18}$GeV) is the reduced Planck scale.
Here, $M_*$ is thought to be a scale where quantum gravity effects becomes relevant, because they would violate the global (continuous) shift symmetry \cite{Misner:1957mt}-\cite{Banks:2010zn}
in addition to the parameter of $g$.
In the string theory, $M_*$ is expected to be the string scale.
For instance, we can find $M_* \sim M_{\rm pl}/\sqrt{{\cal V}}$,
in which ${\cal V} $ is the volume of the extra dimension in the string length unit.
Then one finds $c \sim {\cal V} > 1$ in a large volume limit
\cite{Balasubramanian:2005zx, Higaki:2012ba}.\footnote{
If such a Hubble induced mass is generated by (stingy) instantons \cite{Blumenhagen:2009qh},
we may find the coupling $\tilde{g}V_{\rm inf}\phi^2/(2M_*^2)$, where $\tilde{g} < 1$
depends on such instanton and
may be different from $g$.
Thus, one finds $c \sim \tilde{g}{\cal V} > 1$ for large volume cases where ${\cal V} > 1/\tilde{g}$.
}
The non-minimal coupling to gravity $\xi \phi^2R$ also gives rise to the Hubble induced mass effectively: $c \sim \xi $.
Note that the Higgs quartic coupling $\lambda$ in (\ref{V}) can be negative due to the running effect, but we assume that $\lambda$ remains positive even at the inflation scale, which can be achieved with some additional scalar contribution to the running. The Higgs field $h$ also should be stabilized during inflation to avoid the inhomogeneity of the value of $h$. Such a situation can be circumvented if $M>H$ during inflation or there exists the Hubble induced mass of the Higgs field $\tilde{c}H^2|h|^2$, where $\tilde{c}>0$.
After inflation, inflaton decays into
the SM particles. When the cosmological expansion rate drops bellow the decay rate, the reheating completes. Then, the decay products make a thermal bath with a reheating temperature $T_{\rm R}$. By the inflaton decay, the Hubble induced mass term rapidly decreases and becomes irrelevant for the dynamics of the relaxion, but alternatively the thermal bath affects the dynamics. We take into account this effect.\footnote
Thermal effects on the relaxion model are also discussed in Ref.~\cite{Hardy:2015laa}.
We do not address detailed issues on the thermalization after inflation, but just add the thermal mass term of $\phi$ in an ad hoc way,
\begin{align}\label{eq:Tmass}
\Delta V=\frac{1}{2}\alpha g^2 T^2 (\phi-\phi_{T}(t))^2,
\end{align}
where $\alpha$ is a real constant\footnote{
Here, we treat $\alpha$ as a constant for simplicity. For more precise discussion of the thermal effect, we should take into account dissipation effects on the dynamics of the relaxion.
} normalized by $g^2$,
$T$ denotes the temperature of thermal bath, and $\phi_T(t)$ is a function of time. As in the case of the Hubble induced mass, we need to require the following condition,
\begin{align}
\phi_T(t)<\frac{M}{g}.
\end{align}
For simplicity, we assume $\phi_T(t)=0$ in the following discussion.
As in the case of Hubble induced mass,
the Higgs VEV is stabilized around the origin and becomes homogeneous in the presence of a temperature dependent mass term $T^2 |h|^2$ for $T \gtrsim M$.
In our scenario, the field value of the relaxion after inflation follows along the minimum of the potential determined by the Hubble induced mass and (or) the thermal one. When the Hubble expansion rate decreases so that $cH^2\lesssim g^2\alpha T^2$, the minimum is determined by the thermal mass. In this paper, we mainly focus on the case that the Hubble induced mass is given by a coupling to the inflaton $(V_{\rm inf}/2M_{\ast}^2)\phi^2$, then the mass term exponentially decreases as the reheating completes, $T\simeq T_R$. In this case, at $T=T_R$ the thermal mass dominates over the Hubble induced one (we assume that the temperature of the dilution plasma before reheating is sufficiently low).
As seen below, the relaxation of the Higgs mass starts during the reheating
process where the inflaton oscillation dominates the Universe
for a lower $T_R$, or in the radiation dominated epoch after the reheating
for a higher $T_R$.
\section{Relaxion dynamics}\label{dyn}
We discuss the relaxion dynamics during and after inflation
in the presence of the Hubble induced mass and thermal one.
Due to the Hubble induced mass term
$\frac{1}{2}cH^2\phi^2$,
the relaxion $\phi$ can be stabilized at a local minimum $\phi_{\min}$,
\begin{align}
\phi_{\min}
\sim\frac{gM^2}{cH^2+g^2M^2}M
\sim \frac{gM^2}{cH^2}M .
\label{Hdom}
\end{align}
We require the following condition during inflation,
\begin{align}
\frac{g^2M^2}{cH_{\rm inf}^2}\ll 1,
\end{align}
which ensures the validity of the expansion with respect to $g\phi$, that is, the relaxion is stabilized to a much smaller value than the critical value $\simeq M/g$.
The relaxion $\phi$ continues to stay at the minimum determined by the Hubble induced mass term even after the end of inflation, as long as
the Hubble induced mass dominates over other contributions to the relaxion mass.
During reheating epoch, inflaton oscillates around its minimum, and the inflaton energy density decreases as time passes. Because of this decreasing, the Hubble parameter decreases, and then $\phi$ moves toward a larger value along the time dependent minimum (\ref{Hdom}).
In our model, the relaxion mass can be dominated by the Hubble induced mass or the thermal one just before the relaxation mechanism takes place.
The potential energy of inflaton decreases soon after reheating, then
the thermal effect can dominate the relaxion mass, depending on temperature.
For a lower temperature, the Hubble induced mass may dominate it.
We classify the two cases by the reheating temperature as (A) $T_{\rm R}>\Lambda_c$ and (B) $T_{\rm R}<\Lambda_c$, and then discuss each one.
\subsection{The case (A): $T_{\rm R}>\Lambda_c$}
First, we consider the case in which $T_{\rm R}>\Lambda_c$.
When inflation ends before the relaxion mechanism takes place,
the Hubble parameter becomes small.
Then, the Hubble induced mass term becomes ineffective for the relaxation
owing to a coupling to the inflaton.
Instead, the thermal mass term gives significant effects to the relaxion dynamics,
when the inflaton decay reheats the Universe after inflation;
potential becomes
\begin{align}
V \sim-M^3g\phi+ \frac{1}{2}M^2g^2\phi^2+\frac{1}{2}\alpha g^2T^2\phi^2
\end{align}
for $T>\Lambda_c$. This potential is deformed adiabatically as the temperature decreases,
and the field value in the time-dependent minimum $\phi_{\rm min}$
gradually increases due to the decrease of the temperature as
\begin{align}
\phi_{\rm min}\simeq \frac{M^3}{g(\alpha T^2+M^2)}.\label{Tdom}
\end{align}
The thermal mass becomes larger than the Hubble expansion rate for the relativistic degrees of freedom in the thermal bath being around $g_{\ast}={\cal O}(10^2)$
as in the SM and for a large $\alpha~(\gtrsim 10^2)$, which
we will discuss below.
Thus, the field value of the relaxion settles down to and follows along the temporal minimum (\ref{Tdom}) even in the presence of the Hubble friction.
\footnote
The energy density of the oscillation deviated from the temporal minimum $\Delta \equiv \phi-\phi_{\rm min}$ decrease as like the radiation components, $\rho_{\Delta}\propto a^{-4}$, since the number density of the oscillator decreases by $a^{-3}$ with the decreasing of the thermal mass by $a^{-1}$. On the other hand, the kinetic energy of $\phi_{\rm min}$ does not decrease during radiation dominated epoch as $(\dot{\phi}_{\rm min})^2\sim M^6 H^2/(g^2\alpha^2 T^4)
\sim{\rm const}$. Thus, the energy of the oscillation becomes sub-dominant after sufficient expansion of the Universe.
As $T$ becomes lower, $\phi$ goes down to the larger field value, and finally reaches the critical value
\begin{align}
\phi_c=\frac{M}{g}
\end{align}
around which the Higgs mass becomes tachyonic.
We define the critical temperature $T_c$ around which the thermal mass becomes comparable to the zero temperature mass $g^2 M^2$ and $\phi_{\rm min}$ approaches $\phi_c$:
\begin{align}
T_c\equiv \frac{M}{\sqrt{\alpha}}.
\end{align}
For the efficient transition of the instanton effect to make the periodic potential at this time, the temperature should be smaller than the condensation scale
$T_c\lesssim\Lambda_c$,
\footnote{If the relaxion is in the thermal equilibrium with a heat bath at a very high
temperature $T$, the relaxion is fluctuated with $\delta\phi\sim T$. To avoid the transition over the periodic potential by this fluctuation, we require $\alpha g^2T^2\delta\phi^2\sim\alpha g^2T^4\lesssim\Lambda_c^3\braket{h}$ at $T\simeq T_c$, which reduces to
$\frac{g^2}{\alpha}M^4\lesssim \Lambda_c^3\braket{h}$.
However, this condition is weaker than $T_c \lesssim \Lambda_c$.
}
which gives a lower limit for the dimensionless coupling as
\begin{equation}\label{eq:cond_al}
\alpha\gtrsim\lrp{\frac{M}{\Lambda_c}}^2 \gtrsim 1.
\end{equation}
Otherwise, the relaxion does not settle down to the local minimum where the electroweak scale is dynamically realized through the periodic potential.
For the success of the Big Bang Nucleosynthesis (BBN), the critical temperature needs to be larger than ${\rm MeV}$, giving a upper bound on the dimensionless parameter as
\begin{equation}\label{eq:al_upp}
\alpha\lesssim\lrp{\frac{M}{{\rm MeV}}}^2.
\end{equation}
To fix the field value of the relaxion by the periodic term,
the slope of the periodic potential should be comparable to the linear term:
\begin{equation}\label{Acond}
M^3g\simeq \frac{\Lambda_c^3\braket{h}}{f}\left|\sin\lrp{\frac{\phi_c}{f}}\right|,
\end{equation}
where $\braket{h}$ is the electroweak VEV of the Higgs field.
Since the relaxion dynamically moves following the minimum of the potential (\ref{Tdom}) as temperature drops, the relaxion has kinetic energy with $(\dot{\phi}_{\rm min})^2\sim\lrp{M^3/(g\alpha M_{\rm pl})}^2$. For the relaxion to be stopped by the periodic potential, the kinetic energy should be smaller than $\Lambda_c^3\braket{h}$ as
\begin{equation}\label{eq:cond_veloci}
\frac{1}{g^2\alpha^2}\frac{M^6}{M_{\rm p}^2}\lesssim\Lambda_c^3\braket{h}.
\end{equation}
We have obtained four conditions on the relaxion parameters as (\ref{eq:cond_al}),~(\ref{eq:al_upp}),~(\ref{Acond}),
and (\ref{eq:cond_veloci}), which are summarized as
\begin{equation}\label{eq:sum_cond}
\left\{
\begin{aligned}
&\lrp{\frac{M}{\Lambda_c}}^2\lesssim\alpha\lesssim\lrp{\frac{M}{\rm MeV}}^2,\\
&\alpha\gtrsim\frac{1}{g}\frac{M^3}{M_{\rm pl}\lrp{\Lambda_c^3\braket{h}}^{1/2}},\\
&g\simeq\frac{\Lambda_c^3\braket{h}}{M^3f}
\left|\sin\lrp{\frac{M}{gf}}\right|.
\end{aligned}
\right.
\end{equation}
Note that we can rewrite the last equation as
$1 \sim \left(\frac{\Lambda_c}{M}\right)^4
\left(\frac{\langle h\rangle}{\Lambda_c}\right)
\left(\frac{M}{gf}\right) \left|\sin\lrp{\frac{M}{gf}}\right|$.
Since $M \gtrsim \Lambda_c$ and $\Lambda_c \sim \langle h \rangle$,
the argument of the sine function in the relation needs to be larger than unity to have a solution. By an approximation $|\sin \left(\frac{M}{gf}\right)|\simeq1$, we can rewrite the last equation as
\begin{equation}\label{eq:appro_rela}
g\simeq \lrp{\frac{\Lambda_c}{M}}^{3} \lrp{\frac{\braket{h}}{f}}.
\end{equation}
The condition in the first line of (\ref{eq:sum_cond}) is independent of any choice of $g$ and $f$, and it requires $\alpha$ to be larger than unity. A large value of $\alpha$ could be achieved in the case with the large numbers of the species mediating between the relaxion and the thermal bath, or
$g^2 \alpha$ would be nothing to do with $g$ (or $M$) essentially.
With any value of $M$, we should require at least
\begin{equation}\label{eq:lowest_al}
\alpha\simeq\lrp{\frac{M}{\Lambda_c}}^2.
\end{equation}
This lowest hierarchy is achieved for the decay constant
$f/M_{\rm pl}\lesssim\lrp{\braket{h}/\Lambda_c}^{3/2}\lrp{\Lambda_c/M}^{4}$,
which is obtained by substitution of (\ref{eq:appro_rela}) and (\ref{eq:lowest_al}) into the condition in the second line of (\ref{eq:sum_cond}).
Using the approximated relation (\ref{eq:appro_rela}), we can rewrite the condition in the third line of (\ref{eq:sum_cond}) (corresponding to (\ref{eq:cond_veloci})) as
\begin{equation}\label{eq:cond_veloci_v2}
\alpha\gtrsim\lrp{\frac{M}{\Lambda_c}}^6\frac{f}{M_{\rm pl}}\lrp{\frac{\Lambda_c}{\braket{h}}}^{3/2}.
\end{equation}
Since $\alpha$ is limited from above as shown in the first line of (\ref{eq:sum_cond}), (\ref{eq:cond_veloci_v2}) gives a upper limit on the decay constant as
\begin{equation}\label{eq:cond_f}
\frac{f}{M_{\rm pl}}\lesssim \lrp{\frac{\braket{h}}{\Lambda_c}}^{3/2}\lrp{\frac{\Lambda}{\rm MeV}}^2
\lrp{\frac{\Lambda_c}{M}}^{4}.
\end{equation}
\subsection{The case (B): $T_{\rm R}<\Lambda_c$}
In the case where $T_{\rm R}<\Lambda_c$,
we consider that the energy of inflaton oscillation
dominates the Universe even when the relaxion reaches the critical point $\phi_c$
at which the Higgs mass vanishes. In this case, the temporal minimum of the relaxion is determined by the Hubble induced mass term as~(\ref{Hdom}) instead of the thermal one. When $H=H_c\sim (g/\sqrt{c})M$, the temporal minimum reaches $\phi=\phi_c$. At this point, the minimization condition $\partial_\phi V=0$ reads
\begin{align}\label{eq:relation_H}
gf\simeq \left( \frac{\Lambda_c}{M}\right)^3\braket{h}\sin\lrp{\frac{\phi_c}{f}}.
\end{align}
The quarks of the SM need to obtain their observed masses for the successful BBN.
The temperature at $H=H_c$ should be larger than $\mathcal{O}(1){\rm MeV}$ around which the BBN starts. This requirement gives an upper bound on the dimensionless parameter $c$ as
\begin{equation}
c<
10^{14}\times\lrp{\frac{g_{\ast}}{100}}^{-1}\lrp{\frac{g}{10^{-20}}}^2\lrp{\frac{M}{10{\rm TeV}}}^2.
\end{equation}
As discussed in the case of (A), the dynamical change of the potential minimum
gives kinetic energy for relaxion. In present case, the deformation of the potential is due to the decrease of the Hubble expansion rate. To stop the relaxion by the periodic potential, this kinetic energy should be smaller than the scale of the potential $\Lambda_c^3\braket{h}$. Now since the potential minimum of the relaxion is determined by the Hubble induced mass as $\phi_{\rm min}\sim gM^3/(cH^2)$, the kinetic energy when the relaxion mechanism takes place at $H\sim H_c$ is estimated as $(\dot{\phi}_{\rm min})^2\sim M^4/c.$ Thus, we obtain a condition on the dimensionless parameter $c$ as
\begin{equation}\label{eq:cond_c}
c> \left(\frac{M}{\Lambda_c}\right)^4\left(\frac{\Lambda_c}{\braket{h}}\right) \gtrsim 1.
\end{equation}
Thus, to achieve the relaxion mechanism with the low reheating temperature, the coefficient of the Hubble induced mass term needs to be much larger than unity.
As discussed in Sec.~\ref{model}, one of the possibilities to realize it is that the shift symmetry is broken by quantum gravity effects.
If this is the case, $c\gg 1$ implies that the cutoff scale of such effects would
be smaller than $M_{\rm pl}$.
\\
In this section we have discussed the dynamics of the relaxion especially focusing on the time when the relaxion reaches the critical point $\phi_c = M/g$,
and then obtained the constraints of (\ref{eq:sum_cond})
for $T_R > \Lambda_c$,
and them of (\ref{eq:relation_H}) and (\ref{eq:cond_c}) for $T_R < \Lambda_c$.
In the case where
$T_R>\Lambda_c$, the thermal mass term stabilizes the relaxion field.
Then, a large $\alpha$ is required.
For example, the scale $M$ is around $10~{\rm TeV}$ and $\Lambda_c\sim 100$ GeV~(\ref{eq:scale}), the hierarchy is required with $\alpha_{\rm min}\sim10^4$. In this case, the relaxion mechanism is realized e.g by $(g,\,f)\sim(10^{-13},\,10^{9}\,{\rm GeV})$. With larger hierarchy, the relaxation
is realized with larger decay constants. For examples, $( g,~f)\sim(10^{-16},10^{12}\,{\rm GeV}),\,(10^{-20},10^{16}\,{\rm GeV})$. In the case where
$T_R<\Lambda_c$, the Hubble induced mass stabilizes the relaxion field.
In this case, a sizable $c$ is required.
For $M\sim10~{\rm TeV}$, the dimensionless parameter $c$ needs to be larger than $10^8$.
\section{Conclusion and discussion}\label{concl}
Recently the relaxion mechanism to solve the Higgs hierarchy problem was proposed~\cite{Graham:2015cka}.
In this paper, we have studied
whether the mechanism is accomplished with the large Hubble expansion rate of inflation or high reheating temperature. Unlike the original one, we have discussed the scenario that the relaxion mechanism takes place after inflation, and
then the Higgs VEV settles down to the electroweak scale.
To achieve the scenario, we have taken into account the effects of the
Hubble induced mass or thermal one on the dynamics of the relaxion. In (\ref{Hmass}) and (\ref{eq:Tmass}), we have defined the masses with dimensionless parameters $c$ and $\alpha$. Then, by discussing the cosmological scenario of the relaxion, we have obtained the constraints on the parameters as
(\ref{eq:cond_al}) and (\ref{eq:cond_c}).
From these constraints we can see that the dimensionless parameters need to be larger than unity for $M>\Lambda_c$. Therefore, to accomplish the relaxion mechanism with the Hubble induced mass or thermal one,
there should exist additional shift symmetry breaking terms
in different ways from those given by the combination of $g \phi/M$.
To complete our scenario, it is necessary to consider UV completions
and also the phenomenological aspects of the model.
There exist several origins of the shift symmetry breaking
relevant to a sizable Hubble induced mass and thermal one.
It is also important to realize a large field excursion of the relaxion
\cite{Choi:2015fiu,Kaplan:2015fuy,Ibanez:2015fcv}. The relaxion can be a part of the dark matter component. Along the line of Ref.~\cite{Kobayashi:2016bue},
it is necessary to discuss whether the relic abundance of relaxion in our scenario can be consistent with the present data.
The testability in collider physics would also be interesting as, for instance, in Refs.~\cite{Graham:2015cka,Batell:2015fma}.
Finally, we comment on the breaking of the (continuous)
shift symmetry of the relaxion before or during inflation.
Throughout this paper, we have assumed that
an approximate continuous global symmetry is spontaneously broken down
before or during inflation and the relaxion exists then as its Nambu-Goldstone mode.
(Note that isocurvature perturbation during inflation is suppressed
in the presence of a sizable
Hubble induced mass in our scenario \cite{Dine:2004cq}.)
In contrast, if the spontaneous symmetry breaking occurs after inflation,
the relaxion scenario might lead to contradictions with the observed Universe.
After the symmetry breaking,
the relaxion field would take random values in each patch of the Universe.
The random values can be much larger than the scale $M/g$ owing to
an almost flat potential with monodromy $g\phi/M$. Since we have discussed the relaxion dynamics based on an effective field theory, we can not control the higher terms of the relaxion potential (\ref{lin}) for $\phi \gg M/g$, where there might exist local minima,
and the relaxion would be trapped in them.
For discussion of this issue, we have to determine a UV theory. Further,
by the symmetry breaking,
topological defects would be formed and might cause serious problems for cosmology such as the domination of the domain walls even if there is a bias.
\section*{Acknowledgement}
This work is supported by MEXT-Supported Program for the Strategic Research Foundation at Private Universities,``Topological Science, Grant Number S1511006
(T.H, N.T. and Y.Y.) and JSPS KAKENHI Grant Number 26247042 (T.H.).
|
\section{Introduction}
The cosmic star-formation history (SFH) describes the evolution of star formation in galaxies across cosmic time. It is well summarized by the so-called ``Lilly-Madau'' plot \citep{lilly95,madau96}, which shows the redshift evolution of the star-formation rate (SFR) density, i.e., the total SFR in galaxies in a comoving volume of the universe. The SFR density increases from an early epoch ($z>8$) up to a peak ($z\sim 2$) and then declines by a factor $\sim 20$ down to present day \citep[see][for a recent review]{madau14}.
Three key quantities are likely to drive this evolution: the growth rate of dark matter halos, the gas content of galaxies (i.e., the availability of fuel for star formation), and the efficiency at which gas is transformed into stars. Around $z$=$2$, the mass of halos can grow by a factor of $>2$ in a Gyr; by $z\approx0$, the mass growth rate has dropped by an order of magnitude \citep[e.g.,][]{griffen16}. How does the halo growth rate affect the gas resupply of galaxies? Do galaxies at $z\sim 2$ harbor larger reservoirs of gas? Are they more effective at high redshift in forming stars from their gas reservoirs, possibly as a consequence of different properties of the interstellar medium, or do they typically have more disturbed gas kinematics due to gravitational interactions?
To address some of these questions, we need a census of the dense gas stored in galaxies and available to form new stars as a function of cosmic time, i.e., the total mass of gas in galaxies per comoving volume [$\rho$(gas)]. The statistics of Ly$\alpha$ absorbers (associated with atomic hydrogen, H{\sc i}) along the line of sight toward bright background sources provide us with a measure of $\rho$(H{\sc i}). This appears to be consistent with being constant (within a $\sim$30\% fluctuation) from redshift $z=0.3$ to $z\sim 5$ \citep[see, e.g.,][]{crighton15}, possibly as a result of the balance between gas inflows and outflows in low-mass galaxies \citep{lagos14} and of the on-going gas resupply from the intergalactic medium \citep{lagos11}. However, beyond the local universe, little information currently exists on the amount of molecular gas that is stored in galaxies, $\rho$(H$_2$), which is the immediate fuel for star formation (e.g., see review by \citealt{carilli13}).
Attempts have been made to infer the mass of molecular gas in distant targeted galaxies indirectly from the measurement of their dust emission, via dust--to--gas scaling relations \citep{magdis11,magdis12,scoville14,scoville15,groves15}. But a more direct route is to derive it from the observations of rotational transitions of $^{12}$CO (hereafter, CO), the second most abundant molecule in the universe (after H$_2$). As the second approach is most demanding in terms of telescope time, it has been traditionally applied only with extreme, infrared (IR) luminous sources (e.g., \citealt{bothwell13}; these however account for only 10-20\% of the total SFR budget in the universe; see, \citealt{rodighiero11,magnelli13,gruppioni13,casey14}), or on samples of galaxies pre-selected based on their stellar mass and/or SFR \citep[e.g.,][]{daddi10a,daddi10b,daddi15,tacconi10,tacconi13,genzel10,genzel15,bolatto15}. These observations have been instrumental in shaping our understanding of the molecular gas properties in high-$z$ galaxies.
Through the observation of multiple CO transitions for single galaxies, the CO excitation has been constrained in a variety of systems \citep{weiss07,riechers11,bothwell13,spilker14,daddi15}. Most remarkably, various studies showed that $M_*$- and SFR-selected galaxies at $z>0$ tend to host much larger molecular gas reservoirs than typically observed in local galaxies for a given stellar mass ($M_*$) suggesting that an evolution in the gas fraction $f_{\rm gas}=M_{\rm H2}/(M_*+M_{\rm H2})$ occurs through cosmic time \citep{daddi10a,riechers10,tacconi10,tacconi13,genzel10,genzel15,geach11,magdis12,magnelli12}.
For molecular gas observations to constrain $\rho$(H$_2$) as a function of cosmic time, we need to sample the CO luminosity function in various redshift bins. CO is the second most abundant molecule in the universe (after H$_2$) and therefore is an excellent tracer of the molecular phase of the gas. The CO(1-0) ground transition has an excitation temperature of only $T_{\rm ex}=5.5$\,K, i.e., the molecule is excited in virtually any galactic environment. Other low-J CO lines may be of practical interest, as these levels remain significantly excited in star-forming galaxies; and thus, the associated lines [CO(2-1), CO(3-2), CO(4-3)] are typically brighter and easier to detect than the ground state transition CO(1-0). There have been various predictions of the CO luminosity functions both for the J=1$\rightarrow$0 transition and for intermediate and high-J lines, using either theoretical models \citep[e.g.,][]{obreschkow09a,obreschkow09b,lagos11,lagos12,lagos14,popping14a,popping14b,popping16} or empirical relations \citep[e.g.,][]{sargent12,sargent14,dacunha13,vallini16}.
Theoretical models typically rely on semi-analytical estimates of the budget of gas in galaxies (e.g., converting H{\sc i} into H$_2$ assuming a pressure-based argument, as in \citealt{blitz06}; via metallicity-based arguments, as in \citealt{gnedin10,gnedin11}; or based on the intensity of the radiation field and the gas properties, as in \citealt{krumholz08,krumholz09}), and inferring the CO luminosity and excitation via radiative transfer models. These models broadly agree on the dependence of $\rho$(H$_2$) on $z$, at least up to $z\sim 2$, but widely differ in the predicted CO luminosity functions, in particular for intermediate and high J transitions, where details on the treatment of the CO excitation become critical. For example, the models by \citet{lagos12} predict that the knee of the CO(4-3) luminosity function lies at $L'\approx 5\times10^8$\,K~km\,s$^{-1}$\,pc$^2${} at $z\sim 3.8$, while the models by \citet{popping16} place the knee at a luminosity about 10 times brighter. Such a spread in the predictions highlight the lack of observational constraints to guide the theoretical assumptions.
This study aims at providing observational constraints on the CO luminosity functions and cosmic density of molecular gas via the `molecular deep field' approach. We perform a scan over a large range of frequency ($\Delta \nu/\nu\approx25-30$\,\%) in a region of the sky, and ``blindly'' search for molecular gas tracers at any position and redshift. By focusing on a blank field, we avoid the biases due to pre-selection of sources. This method naturally provides us with a well-defined cosmic volume where to search for CO emitters, thus leading to direct constraints on the CO luminosity functions. Our first pilot experiment with the IRAM Plateau de Bure Interferometer \citep[PdBI; see][]{decarli14} led to the first, weak constraints on the CO luminosity functions at $z>0$ \citep{walter14}. The modest sensitivity (compared with the expected knee of the CO luminosity functions) resulted in large Poissonian uncertainties. These can be reduced now, thanks to the Atacama Large Millimeter/Sub-millimeter Array (ALMA).
We obtained ALMA Cycle 2 observations to perform two spatially coincident molecular deep fields, at 3mm and 1mm respectively, in a region of the Hubble Ultra Deep Field \citep[UDF,][]{beckwith06}. The data set of our ALMA Spectroscopic Survey (ASPECS) is described in detail in Paper I of this series \citep{walter16}. Compared with the aforementioned PdBI effort, we now reach a factor of 3--4 better sensitivity, which allows us to sample the expected knee of the CO luminosity functions over a large range of transitions. Furthermore, the combination of band 3 and 6 offers us direct constraints on the CO excitation of the observed sources, thus allowing us to infer the corresponding CO(1-0) emission, and therefore $\rho$(H$_2$). The collapsed cube of the 1mm observations also yields one of the deepest dust continuum observations ever obtained \citep[Paper II of this series,][]{aravena16a}, which we can use to compare the $\rho$(H$_2$) estimates based on CO and the $\rho$(gas) estimates based on the dust emission.
This paper is organized as follows: In Sec.~\par\noindent\hangindent 20 pt{sec_observations} we summarize the observations and the properties of the data set. In Sec.~\par\noindent\hangindent 20 pt{sec_analysis} we describe how we derive our constraints on the CO luminosity functions and on $\rho$(H$_2$) and $\rho$(gas). In Sec.~\par\noindent\hangindent 20 pt{sec_discussion} we discuss our results. Throughout the paper we assume a standard $\Lambda$CDM cosmology with $H_0=70$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{\rm m}=0.3$ and $\Omega_{\Lambda}=0.7$ \citep[broadly consistent with the measurements by the][]{planck15}.
\section{Observations}\label{sec_observations}
The data set used in this study consists of two frequency scans at 3mm (band 3) and 1mm (band 6) obtained with ALMA in the UDF centered at RA = 03:32:37.900, Dec = --27:46:25.00 (J2000.0). Details on the observations and data reduction are presented in Paper I, but the relevant information is briefly summarized here. The 3mm scan covers the 84-115 GHz range with a single spatial pointing. The primary beam of the 12m ALMA antennas is $\sim75''$ at 84\,GHz and $\sim 54''$ at 115\,GHz. The typical RMS noise is 0.15\,mJy\,beam$^{-1}$ per 20\,MHz channel. The 1mm scan encompasses the frequency window 212-272 GHz. In order to sample a similar area as in the 3mm scan, given the smaller primary beam ($\sim 26''$), we performed a 7 point mosaic. The typical depth of the data is $\sim 0.5$\,mJy\,beam$^{-1}$ per 30 MHz channel. The synthesized beams are $\sim 3.5''\times 2.0''$ at 3mm and $\sim 1.5''\times 1.0''$ at 1mm.
Fig.~\par\noindent\hangindent 20 pt{fig_lum_lim} shows the redshift ranges and associated luminosity limits reached for various transitions in the two bands. The combination of band 3 and band 6 provides virtually-complete CO redshift coverage. The luminosity limits are computed assuming 5-$\sigma$ significance, a line width of 200\,km\,s$^{-1}${}, and unresolved emission at the angular resolution of our data. At $z\mathrel{\rlap{\lower 3pt \hbox{$\sim$}} \raise 2.0pt \hbox{$>$}} 1.5$, the luminosity limit (expressed as a velocity-integrated temperature over the beam, which is constant for all CO transitions in the case of thermalized emission) is roughly constant as a function of redshift for different CO transitions as well as for [C\,{\sc ii}]{}: $\sim 2\times 10^9$\,K~km\,s$^{-1}$\,pc$^2${}.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_lim_lum.png}\\
\caption{Redshift coverage and luminosity limit reached in our 1mm and 3mm scans, for various CO transitions and for the [C\,{\sc ii}]{} line. The (5-$\sigma$) limits plotted here are computed assuming point-source emission, and are based on the observed noise per channel, scaled for a line width of 200 km\,s$^{-1}${}. The combination of band 3\&6 offers a virtually-complete CO redshift coverage. The luminosity limit (expressed as velocity-integrated temperature) is roughly constant at $z\mathrel{\rlap{\lower 3pt \hbox{$\sim$}} \raise 2.0pt \hbox{$>$}} 1.5$. The depth of our observations is sufficient to sample the typical knee of the expected CO luminosity functions ($L'\sim5\times10^9$\,K~km\,s$^{-1}$\,pc$^2${}). }
\label{fig_lum_lim}
\end{figure}
\section{Analysis}\label{sec_analysis}
Given the blank field approach of ASPECS, with no pre-selection on the targeted sources, we have a well-defined, volume-limited sample of galaxies at various redshifts where we can search for CO emission. We first concentrated on the ``blind'' CO detections presented in Paper I (Tab.~2), and then include the information from galaxies with a known redshift. This provides us with direct constraints on the CO luminosity function in various redshift bins. We then use these constraints to infer the CO(1-0) luminosity functions in various redshift bins, and therefore the H$_2$ mass ($M_{\rm H2}$) budget in galaxies throughout cosmic time.
\subsection{CO detections}
\subsubsection{Blind detections}\label{sec_blind}
\begin{table*}
\caption{{\rm Catalogue of the line candidates discovered with the blind line search.
(1) Line ID. (2-3) Right ascension and declination (J2000).
(4) Fidelity level at the S/N of the line candidate.
(5) Completeness at the luminosity of the line candidate.
(6) Is there an optical/near-IR counterpart?
(7) Notes on line identification:
{\em i}- Multiple lines detected in the ASPECS cubes;
{\em ii}- Lack of other lines in the ASPECS cubes;
{\em iii}- Absence of optical/near-IR counterpart suggests high $z$;
{\em iv}- Supported by (a) spectroscopic, (b) grism, or (c) photometric redshift.
(8) Possible line identification. A cardinal number indicates the upper J level of a CO transition.
(9) CO redshift corresponding to the adopted line identification.
(10) Line luminosity, assuming the line identification in col.(8). The uncertainties are propagated from the uncertainties in the line flux measurement.
(11) Molecular gas mass $M_{\rm H2}$ as derived from the observed CO luminosity (see eq.~\par\noindent\hangindent 20 pt{eq_H2}), only for J$<$5 CO lines.
}} \label{tab_lines}
\begin{center}
\begin{tabular}{ccccccccccc}
\hline
ASPECS ID & RA & Dec & Fid. & $C$ & C.part? & Notes & Line & $z_{\rm CO}$ & $L'$ & $M_{\rm H2}$ \\
& & & & & & &ident.& &[$10^8$K~km\,s$^{-1}$\,pc$^2$]& [$10^8$M$_\odot$] \\
(1)& (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) & (11) \\
\hline
\multicolumn{11}{c}{3mm}\\
3mm.1 & 03:32:38.52 & --27:46:34.5 & 1.00 & 1.00 & Y & {\em i, iv}(b) & 3 & 2.5442 & $240.4\pm1.0$ & $2061\pm 9$ \\
3mm.2 & 03:32:39.81 & --27:46:11.6 & 1.00 & 1.00 & Y & {\em i, iv}(a) & 2 & 1.5490 & $136.7\pm2.1$ & $ 648\pm10$ \\
3mm.3 & 03:32:35.55 & --27:46:25.7 & 1.00 & 0.85 & Y & {\em iv}(a) & 2 & 1.3823 & $33.7\pm0.7$ & $ 160\pm 3$ \\
3mm.4 & 03:32:40.64 & --27:46:02.5 & 1.00 & 0.85 & N & {\em ii} & 3 & 2.5733 & $45.8\pm1.0$ & $ 393\pm 9$ \\
& & & & & & & 4 & 4.0413 & $92.2\pm2.8$ & $1071\pm33$ \\
& & & & & & & 5 & 5.3012 & $89.5\pm2.7$ & --- \\
3mm.5 & 03:32:35.48 & --27:46:26.5 & 0.87 & 0.85 & Y & {\em iv}(a) & 2 & 1.0876 & $28.3\pm0.9$ & $ 134\pm 4$ \\
3mm.6 & 03:32:35.64 & --27:45:57.6 & 0.86 & 0.85 & N & {\em ii, iii} & 3 & 2.4836 & $72.8\pm1.0$ & $ 624\pm 9$ \\
& & & & & & & 4 & 3.6445 & $77.3\pm1.0$ & $ 898\pm12$ \\
& & & & & & & 5 & 4.8053 & $76.2\pm1.0$ & --- \\
3mm.7 & 03:32:39.26 & --27:45:58.8 & 0.86 & 0.85 & N & {\em ii, iii} & 3 & 2.4340 & $25.9\pm1.0$ & $ 222\pm 9$ \\
& & & & & & & 4 & 3.5784 & $27.6\pm1.0$ & $ 321\pm12$ \\
& & & & & & & 5 & 4.7227 & $27.3\pm1.0$ & --- \\
3mm.8 & 03:32:40.68 & --27:46:12.1 & 0.76 & 0.85 & N & {\em ii, iii} & 3 & 2.4193 & $58.6\pm0.9$ & $ 502\pm 8$ \\
& & & & & & & 4 & 3.5589 & $62.6\pm1.0$ & $ 727\pm12$ \\
& & & & & & & 5 & 4.6983 & $62.0\pm1.0$ & --- \\
3mm.9 & 03:32:36.01 & --27:46:47.9 & 0.74 & 0.85 & N & {\em ii, iii} & 3 & 2.5256 & $30.5\pm1.0$ & $ 261\pm 9$ \\
& & & & & & & 4 & 3.7006 & $32.3\pm1.0$ & $ 375\pm12$ \\
& & & & & & & 5 & 4.8754 & $31.8\pm1.0$ & --- \\
3mm.10 & 03:32:35.66 & --27:45:56.8 & 0.61 & 0.85 & Y & {\em ii, iv}(b) & 3 & 2.3708 & $70.4\pm0.9$ & $ 603\pm 8$ \\
\hline
\multicolumn{11}{c}{1mm}\\
1mm.1$^*$ & 03:32:38.54 & --27:46:34.5 & 1.00 & 1.00 & Y & {\em i, iv}(b) & 7 & 2.5439 & $48.02\pm0.37$ & --- \\
1mm.2$^*$ & 03:32:38.54 & --27:46:34.5 & 1.00 & 1.00 & Y & {\em i, iv}(a) & 8 & 2.5450 & $51.42\pm0.23$ & --- \\
1mm.3 & 03:32:38.54 & --27:46:31.3 & 0.93 & 0.85 & Y & {\em iv}(b) & 3 & 0.5356 & $3.66\pm0.08$ & $ 31\pm1$ \\
1mm.4 & 03:32:37.36 & --27:46:10.0 & 0.85 & 0.65 & N & {\em i} & [C\,{\sc ii}] & 6.3570 & $12.49\pm0.23$ & --- \\
1mm.5 & 03:32:38.59 & --27:46:55.0 & 0.79 & 0.75 & N & {\em ii} & 4 & 0.7377 & $12.95\pm0.09$ & $150\pm1$ \\
& & & & & & & [C\,{\sc ii}] & 6.1632 & $31.84\pm0.22$ & --- \\
1mm.6 & 03:32:36.58 & --27:46:50.1 & 0.78 & 0.75 & Y & {\em iv}(c) & 4 & 1.0716 & $21.45\pm0.15$ & $249\pm2$ \\
& & & & & & & 5 & 1.5894 & $29.12\pm0.21$ & --- \\
& & & & & & & 6 & 2.1070 & $33.68\pm0.24$ & --- \\
1mm.7 & 03:32:37.91 & --27:46:57.0 & 0.77 & 1.00 & N & {\em ii, iii} & 4 & 0.7936 & $37.53\pm0.10$ & $436\pm1$ \\
& & & & & & & [C\,{\sc ii}] & 6.3939 & $84.01\pm0.23$ & --- \\
1mm.8 & 03:32:37.68 & --27:46:52.6 & 0.71 & 0.72 & N & {\em ii, iii} & [C\,{\sc ii}] & 7.5524 & $23.22\pm0.24$ & --- \\
1mm.9 & 03:32:36.14 & --27:46:37.0 & 0.63 & 0.75 & N & {\em ii, iii} & 4 & 0.8509 & $8.21\pm0.12$ & $ 95\pm1$ \\
& & & & & & & [C\,{\sc ii}] & 6.6301 & $16.84\pm0.25$ & --- \\
1mm.10 & 03:32:37.08 & --27:46:19.9 & 0.62 & 0.75 & N & {\em ii, iii} & 4 & 0.9442 & $14.74\pm0.18$ & $171\pm2$ \\
& & & & & & & 6 & 1.9160 & $25.05\pm0.30$ & --- \\
& & & & & & & [C\,{\sc ii}] & 7.0147 & $26.59\pm0.32$ & --- \\
1mm.11 & 03:32:37.71 & --27:46:41.0 & 0.61 & 0.85 & N & {\em ii, iii} & 3 & 0.5502 & $4.84\pm0.09$ & $ 41\pm1$ \\
& & & & & & & [C\,{\sc ii}] & 7.5201 & $16.25\pm0.30$ & --- \\
\hline
\end{tabular}
\begin{tablenotes}
\small
\item $^*$ Not used for deriving the H$_2$ mass for this source, as a lower-J line is available.
\end{tablenotes}
\end{center}
\end{table*}
In Paper I, we describe our ``blind search'' of CO emission purely based on the ALMA data (i.e., with no support from ancillary data at other wavelengths)\footnote{The code for the blind search of line candidates is publicly available at \textsf{http://www.mpia.de/homes/decarli/ASPECS/findclumps.cl}.}. In brief, we perform a floating average of consecutive frequency channels in bins of $\sim$50--300\,km\,s$^{-1}${} in the imaged cubes. For each averaged image, we compute the map rms and select peaks based on their S/N. A search for negative (= noise) peaks allows us to quantify the fidelity of our line candidates based on their S/N, and the injection of mock lines allows us to assess the level of completeness of our search as a function of various line parameters, including the line luminosity. The final catalogue consists of 10 line candidates from the 3mm cube, and 11 from the 1mm cube. We use a Gaussian fit of the candidate spectra to estimate the line flux, width, and frequency (see Tab.~2 of Paper I), and we investigate the available optical/near-IR images to search for possible counterparts.
The line identification (and therefore, the redshift association) requires a number of stpng, similar to our earlier study of the HDF--N \citep{decarli14}, which are as follows:
\begin{itemize}
\item[{\em i-}] We inspect the cubes at the position of each line candidate, and search for multiple lines. If multiple lines are found, the redshift should be uniquely defined. Since $\nu_{\rm CO[J-(J-1)]}\approx {\rm J}\,\nu_{\rm CO(1-0)}$, some ambiguity may still be in place [e.g., two lines with a frequency ratio of 2 could be CO(2-1) and CO(4-3), or CO(3-2) and CO(6-5)]. In these cases, the following stpng allow us to break the degeneracy.
\item[{\em ii-}] The absence of multiple lines can then be used to exclude some redshift identification. E.g., lines with similar J should show similar fluxes, under reasonable excitation conditions. If we identify a bright line as, e.g., CO(5-4), we expect to see a similarly-luminous CO(4-3) line (if this falls within the coverage of our data set). If that is not the case, we can exclude this line identification.
\item[{\em iii-}] The exquisite depth of the available multi-wavelength data allows us to detect the starlight emission of galaxies with stellar mass $M_{*}\sim10^8$\,M$_\odot${} at almost all $z<2$. In the absence of an optical / near-IR counterpart, we thus exclude redshift identification that would locate the source at $z<2$.
\item[{\em iv-}] In the presence of an optical/near-IR counterpart, the line identification is guided by the availability of optical redshift estimates. Optical spectroscopy \citep[e.g., see the compilations by][]{lefevre05,coe06,skelton14,morris15} is considered secure (typical uncertainties are in the order of a few hundred km\,s$^{-1}${}). When not available, we rely on {\em HST} grism data \citep{morris15,momcheva16}, or photometric redshifts \citep{coe06,skelton14}.
\end{itemize}
Ten out of 21 blindly-selected lines are uniquely identified in this way. A bootstrap analysis is then adopted to account for the remaining uncertainties in the line identification: To each source, we assign a redshift probability distribution which is proportional to the comoving volume in the redshift bins sampled with all the possible line identifications. We then run 1000 extractions of the redshift values picked from their probability distributions and compute the relevant quantities (line luminosities, inferred molecular masses, contribution to the cosmic density of molecular gas) in each case. The results are then averaged among all the realizations. The line identifications and associated redshifts are listed in Tab.~\par\noindent\hangindent 20 pt{tab_lines}.
\begin{figure*}
\includegraphics[width=0.99\columnwidth]{fig_completeness_Lum_3mm.png}
\includegraphics[width=0.99\columnwidth]{fig_completeness_Lum_1mm.png}\\
\caption{Luminosity limit reached in our 3mm and 1mm scans, for various CO transitions. The completeness is computed as the number of mock lines retrieved by our blind search analysis divided by the number of ingested mock lines, and here it is plotted as a function of the line luminosity. The 50\% limits, marked as dashed vertical lines, are typically met at $L'=(3-6) \times 10^9$\,K~km\,s$^{-1}$\,pc$^2${} at 3mm for any J$>$1, and at $L'=(4-8) \times 10^8$\,K~km\,s$^{-1}$\,pc$^2${} at 1mm for any J$>$3. The J=1 and 3 cases in the 3mm and 1mm cubes show a broader distribution towards lower luminosity limits due to the wide spread of luminosity distance for these transitions within the frequency ranges of our observations.}
\label{fig_completeness}
\end{figure*}
To compute the contribution of each line candidate to the CO luminosity functions and to the cosmic budget of molecular gas mass in galaxies, we need to account for the fidelity (i.e., the reliability of a line candidate against false-positive detections) and completeness (i.e., the fraction of line candidates that we retrieve as a function of various line parameters) of our search. For the fidelity, we infer the incidence of false-positive detections from the statistics of negative peaks in the cubes as a function of the line S/N, as described in Sec.~3.1.1 of Paper I. Fig.~\par\noindent\hangindent 20 pt{fig_completeness} shows the completeness of our line search as a function of the line luminosity. This is obtained by creating a sample of 2500 mock lines (as point sources), with a uniform distribution of frequency, peak flux density, width, and position within the primary beam. Under the assumption of observing a given transition [e.g., CO(3-2)], we convert the input frequency into redshift, and the integrated line flux ($F_{\rm line}$) from the peak flux density and width. We then compute line luminosities for all the mock input lines as:
\begin{equation}\label{eq_L1def}
\frac{L'}{\rm K\,km\,s^{-1}\,pc^2}=\frac{3.25\times 10^7}{(1+z)^3} \, \frac{F_{\rm line}}{\rm Jy\,km\,s^{-1}} \left(\frac{\nu_{\rm obs}}{\rm GHz}\right)^{-2} \left(\frac{D_{\rm L}}{\rm Mpc}\right)^2
\end{equation}
where $\nu_{\rm obs}$ is the observed frequency of the line, and $D_{\rm L}$ is the luminosity distance \citep[see, e.g.,][]{solomon97}. Finally, we run our blind line search algorithm and display the fraction of retrieved-to-input lines as a function of the input line luminosity. Our analysis is 50\% complete down to line luminosities of $(4-6) \times 10^9$\,K~km\,s$^{-1}$\,pc$^2${} at 3mm for any J$>$1, and $(1-6) \times 10^8$\,K~km\,s$^{-1}$\,pc$^2${} at 1mm for any J$>$3, in the area corresponding to the primary beam of the 3mm observations. The completeness distributions as a function of line luminosity in the J=1 case (at 3mm) and the J=3 case (at 1mm) show long tails towards lower luminosities due to the large variations of $D_{\rm L}$ within our scans for these lines (see also Fig.~\par\noindent\hangindent 20 pt{fig_lum_lim}).
The levels of fidelity and completeness at the S/N and luminosity of the line candidates in our analysis is reported in Tab.~\par\noindent\hangindent 20 pt{tab_lines}. At low S/N, flux boosting might bias our results high, through effectively overestimating the impact of a few intrinsically bright sources against many fainter ones scattered above our detection threshold by the noise. However, the relatively high S/N ($>$5) of our line detections, and the statistiscal corrections for missed lines that are scattered below our detection threshold, and for spurious detections, make the impact of flux boosting negligible in our analysis.
\subsubsection{CO line stack}
We can improve the sensitivity of our CO search beyond our `blind' CO detections by focusing on those galaxies where an accurate redshift is available via optical/near-IR spectroscopy. Slit spectroscopy typically leads to uncertainties of a few 100 km\,s$^{-1}${}, while grism spectra from the 3D-{\em HST} \citep{momcheva16} have typical uncertainties of $\sim 1000$ km\,s$^{-1}${} due to the coarser resolution and poorer S/N. By combining the available spectroscopy, we construct a list of 42 galaxies for which a slit or grism redshift information is available \citep{lefevre05,coe06,skelton14,morris15,momcheva16} within $37.5''$ from our pointing center (this corresponds to the area of the primary beam at the low-frequency end of the band 3 scan). Out of these, 36 galaxies have a redshift for which one or more J$<$5 CO transitions have been covered in our frequency scans. We extract the 3mm and 1mm spectra of all these sources, and we stack them with a weighted-average. As weights, we used the inverse of the variance of the spectral noise. This is the pixel rms of each channel map, corrected a posteriori for the primary beam attenuation at the source position. As Fig.~\par\noindent\hangindent 20 pt{fig_zspec_stack} shows, no obvious line is detected above a S/N=3. If we integrate the signal in a 1000\,km\,s$^{-1}${} wide bin centered on the rest-frame frequency of the lines, we retrieve a $\sim 2$-$\sigma$ detection of the CO(2-1) and CO(4-3) lines (corresponding to average line fluxes of $\sim 0.006$\,Jy\,km\,s$^{-1}${} and $\sim 0.010$\,Jy\,km\,s$^{-1}${} respectively). However, given that their low significance, and that they are drawn from a relatively sparse sample, we opt not to include them in the remainder of the analysis, until we are able to significantly expand the list of sources with secure optical/near-IR redshifts. This will be possible thanks to the advent of integral field spectroscopy units with large field of view, like MUSE, which will provide spectra (and therefore redshifts) for hundreds of galaxies in our pointing.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_zspec_stack.png}\\
\caption{Stacked mm spectrum of the sources in our field with optical/near-IR redshifts. The adopted spectral bin is 70\,km\,s$^{-1}${} wide. The 1-$\sigma$ uncertainties are shown as grey lines. We highlight the $\pm$500\,km\,s$^{-1}${} range where the stacked flux is integrated. We also list the number of sources entering each stack. No clear detection is reported in any of the stacked transitions.}
\label{fig_zspec_stack}
\end{figure}
\subsection{CO luminosity functions}
The CO luminosity functions are constructed as follows:
\begin{equation}\label{eq_LF}
\Phi (\log L_i)=\frac{1}{V}\,\sum_{j=1}^{N_i}\frac{{\rm Fid}_j}{C_j}
\end{equation}
Here, $N_i$ is the number of galaxies with a CO luminosity falling into the luminosity bin $i$, defined as the luminosity range between $\log L_i-0.5$ and $\log L_i+0.5$, while $V$ is the volume of the universe sampled in a given transition. Each entry $j$ is down-weighted according to the fidelity (Fid$_j$) and up-scaled according to the completeness ($C_j$) of the $j$-th line. As described in Paper I, the fidelity at a given S/N is defined as $(N_{\rm pos}-N_{\rm neg})/N_{\rm pos}$, where $N_{\rm pos/neg}$ is the number of positive and negative lines with said S/N. This definition of the fidelity allows us to statistically subtract the false positive line candidates from our blind selection. The uncertainties on $\Phi(\log L_i)$ are set by the Poissonian errors on $N_i$, according to \citet{gehrels86}\footnote{According to \citet{cameron11}, the binomial confidence intervals in \citet{gehrels86} might be overestimated in the low-statistics regime compared to a fully Bayesian treatment of the distributions. A similar effect is possibly in place for Poissonian distributions, although a formal derivation is beyond the scope of this work. Here we conservatively opt to follow the classical \citet{gehrels86} method.}. We consider the confidence level corresponding to 1-$\sigma$. We include the uncertainties associated with the line identification and the errors from the flux measurements in the bootstrap analysis described in Sec.~\par\noindent\hangindent 20 pt{sec_blind}. Given that all our blind sources have S/N$>$5 by construction, and the number of entries is typically of a few sources per bin, Poissonian uncertainties always dominate. The results of the bootstrap are averaged in order to produce the final luminosity functions.
The CO luminosity functions obtained in this way are shown in Fig.~\par\noindent\hangindent 20 pt{fig_co_lf}. For comparison, we include the predictions based on semi-analytical models by \citet{lagos12} and \citet{popping16} and on empirical IR luminosity function of {\em Herschel} sources by \citet{vallini16}, as well as the constraints obtained by the earlier study of the HDF--N \citep{walter14}.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_co_lf.png}\\
\caption{CO luminosity functions in various redshift bins. The constraints from our ALMA UDF project are marked as red squares, with the vertical size of the box showing the Poissonian uncertainties. The results of the HDF study by \citet{walter14} are shown as cyan boxes, with error bars marking the Poissonian uncertainties. Semianalytical models by \citet{lagos12} and \citet{popping16} as well as the empirical predictions by \citet{vallini16} are shown for comparison. Our ALMA observations reach the depth required to sample the expected knee of the luminosity functions in most cases \citep[the only exception being the $\langle z \rangle=3.80$ bin when compared with the predictions by][]{lagos12}. Our observations reveal an excess of CO-luminous sources at the bright end of the luminosity function, especially in the 3mm survey, with respect to the predictions. Such an excess is not observed in the 1mm, suggesting that the CO excitation is typically modest compared to the models shown here.}
\label{fig_co_lf}
\end{figure}
\begin{figure*}
\includegraphics[width=0.99\textwidth]{fig_co10_lf.png}\\
\caption{CO(1-0) luminosity functions in various redshift bins. The constraints from ASPECS are marked as red squares, with the vertical size of each box showing the uncertainties. The results from the 3mm scan with PdBI by \citet{walter14} are shown as cyan boxes, with error bars marking the Poissonian uncertainties. The observed CO(1-0) luminosity functions of local galaxies by \citet{keres03} and \citet{boselli14} are shown as red circles and orange diamonds in the first panel, respectively, and as grey points for comparison in all the other panels. The intensity mapping constraints from \citet{keating16} are shown as a shaded yellow area. Semi-analytical models by \citet{lagos12} and \citet{popping16} as well as the empirical predictions by \citet{vallini16} are shown for comparison. The mass function scale shown in the top assumes a fixed $\alpha_{\rm CO}=3.6$\,M$_\odot${}(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$. Our results agree with the predictions at $z<1$, and suggest that an excess of bright sources with respect to both the empirical predictions by \citet{vallini16} and the models by \citet{lagos12} appears at $z>1$.}
\label{fig_co10_lf}
\end{figure*}
Our observations reach the knee of the luminosity functions in almost all redshift bins. The only exception is the CO(4-3) transition in the $\langle z \rangle=3.80$ bin, for which the models by \citet{lagos12} place the knee approximately one order of magnitude below that predicted by \citet{popping16}, thus highlighting the large uncertainties in the state-of-the-art predictions of gas content and CO excitation, especially at high redshift. In particular, these two approaches differ in the treatment of the radiative transfer and CO excitation in a number of ways: 1) \citet{lagos12} adopt a single gas density value for each galaxy, whereas \citet{popping16} construct a density distribution for each galaxy, and assume a log-normal density distribution for the gas within clouds; 2) \citet{lagos12} include heating from both UV and X-rays (although the latter might be less critical for the purposes of this paper), while \citet{popping16} only consider the UV contribution to the heating; 3) the CO chemistry in \citet{lagos12} is set following the \textsf{UCL\_PDR} photo-dissociation region code \citep{bell06,bell07}, and in \citet{popping16} it is based on a fit to results from the \citet{wolfire10} photo-dissociation region code; 4) the CO excitation in \citet{lagos12} is also based on the \textsf{UCL\_PDR} code, while \citet{popping16} adopt a customized escape probability code for the level population; 5) the typical $\alpha_{\rm CO}$ in the \citet{lagos12} models is higher than in \citet{popping16}, although the exact value of $\alpha_{\rm CO}$ in both models changes from galaxy to galaxy [i.e., the CO(1-0) luminosity functions do not translate into H$_2$ mass functions with a simple scaling].
Our observations shown in Fig.~\par\noindent\hangindent 20 pt{fig_co_lf} indicate that an excess of CO-bright sources with respect to semi-analytical models might be in place. This is apparent in the 3mm data. However, the same excess is not observed in the 1mm band. In particular, in the $\langle z \rangle = 1.43$ bin, the lack of bright CO(5-4) lines [compared to the brighter CO(2-1) emission reported here] suggests that the CO excitation is typically modest.
Such apparent low CO excitation is supported by the detailed analysis of a few CO-bright sources presented in a companion paper \citep[Paper IV of this series,][]{decarli16}. These findings guide our choice of a low-excitation template to convert the observed J$>$1 luminosities into CO(1-0). In the next stpng of our analysis, we refer to the template of CO excitation of main sequence galaxies by \citet{daddi15}: If $r_{\rm J1}$ is the temperature ratio between the CO(J-[J-1]) and the CO(1-0) transitions, we adopt $r_{\rm J1}$=$0.76\pm0.09$, $0.42\pm0.07$, $0.23\pm0.04$ for J=2,3,5. In the case of CO(4-3) (which is not part of the template), we interpolate the models shown in the left-hand panel of Fig.~10 in \citet{daddi15}, yielding $r_{41}$=$0.31\pm0.06$, where we conservatively assume a 20\% uncertainty. Each line luminosity is then converted into CO(1-0) as:
\begin{equation}\label{eq_co10}
\log L_{\rm CO(1-0)}'=\log L_{\rm CO(J-[J-1])}'-\log r_{\rm J1}
\end{equation}
The uncertainties in the excitation correction are included in the bootstrap analysis described in Sec.~\par\noindent\hangindent 20 pt{sec_blind}. Based on these measurements, we derive CO(1-0) luminosity functions following eq.~\par\noindent\hangindent 20 pt{eq_LF}. The results are shown in Fig.~\par\noindent\hangindent 20 pt{fig_co10_lf}. Compared to Fig.~\par\noindent\hangindent 20 pt{fig_co_lf}, we have removed the $\langle z \rangle = 1.43$ bin from the 1mm data as the CO(2-1) line at 3mm is observed in practically the same redshift range and is subject to smaller uncertainties related to CO excitation corrections. Our observations succeed in sampling the predicted knee of the CO(1-0) luminosity functions at least up to $z\sim3$. Our measurements reveal that the knee of the CO(1-0) luminosity function shifts toward higher luminosities as we move from $z\approx 0$ \citep{keres03,boselli14} to $z\sim 2$. Our results agree with the model predictions at $z<1$. However, at $z>1$ they suggest an excess of CO--luminous sources, compared to the current models.
This result is robust against CO excitation uncertainties: For example, it is already apparent in the $\langle z\rangle =1.43$ bin, where we covered the CO(2-1) line in our 3mm cube; this line is typically close to be thermalized in star forming galaxies, so excitation corrections are small. Our result is also broadly consistent with the findings by \citet{keating16}, based on a CO(1-0) intensity mapping study at $z=2-3$, that is unaffected by CO excitation.
\subsection{Cosmic H$_2$ mass density}
To derive H$_2$ masses, and the evolution of the cosmic H$_2$ mass density, we now convert the CO(1-0) luminosities into molecular gas masses $M_{\rm H2}$:
\begin{equation}\label{eq_H2}
M_{\rm H2}=\alpha_{\rm CO} \, L_{\rm CO(1-0)}'
\end{equation}
The conversion factor $\alpha_{\rm CO}$ implicitly assumes that CO is optically thick. The value of $\alpha_{\rm CO}$ critically depends on the metallicity of the interstellar medium \citep[see][for a review]{bolatto13}. A galactic value $\alpha_{\rm CO} = 3-6$ M$_\odot$\,(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$ is expected for most of non-starbursting galaxies with metallicities $Z\mathrel{\rlap{\lower 3pt \hbox{$\sim$}} \raise 2.0pt \hbox{$>$}} 0.5$ Z$_\odot$ \citep{wolfire10,glover11,feldmann12}. At $z\sim 0.1$, this is the case for the majority of main-sequence galaxies with $M_{*}>10^9$\,M$_\odot${} \citep{tremonti04}. This seems to hold even at $z\sim 3$, if one takes into account the SFR dependence of the mass-metallicity relation \citep{mannucci10}. Following \citet{daddi10a}, we thus assume $\alpha_{\rm CO} = 3.6$ M$_\odot$\,(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$ for all the sources in our sample. In section \par\noindent\hangindent 20 pt{sec_discussion} we discuss how our results would be affected by relaxing this assumption.
Next, we compute the cosmic density of molecular gas in galaxies, $\rho$(H$_2$):
\begin{equation}\label{eq_rhoH2}
\rho ({\rm H_2})=\frac{1}{V}\,\sum_{i} \sum_{j=1}^{N_i}\frac{M_{i,j} \, P_j}{C_j}
\end{equation}
where $M_{i,j}$ is a compact notation for $M_{\rm H2}$ of the $j$-th galaxy in the mass bin $i$, and the index $i$ cycles over all the mass bins. As for $\Phi$, the uncertainties on $\rho$(H$_2$) are dominated by the Poissonian errors. Our findings are shown in Fig.~\par\noindent\hangindent 20 pt{fig_rhoH2_z} and are summarized in Tab.~\par\noindent\hangindent 20 pt{tab_rhoH2}. We note that the measurements presented here are only based on the observed part of the luminosity function. Therefore, we do not attempt to correct for undetected galaxies in lower luminosity bins given the large uncertainties in the individual luminosity bins and the unknown intrinsic shape of the CO luminosity function.
From Fig.~\par\noindent\hangindent 20 pt{fig_rhoH2_z}, it is clear that there is an evolution in the molecular gas content of galaxies with redshift, in particular compared with the $z=0$ measurements by \citet{keres03} [$\rho$(H$_2$)=$(2.2\pm0.8)\times 10^7$\,M$_\odot${}\,Mpc$^{-3}$] and \citet{boselli14} [$\rho$(H$_2$)$= (1.2\pm0.2)\times 10^7$\,M$_\odot${}\,Mpc$^{-3}$]. The global amount of molecular gas stored in galaxies at the peak epoch of galaxy assembly is 3--10 times larger than at the present day. This evolution can be followed up to $z\sim4.5$, i.e., 90\% of the age of the universe. This trend agrees with the initial findings using PdBI \citep{walter14}. Our results are consistent with the constraints on $\rho$(H$_2$) at $z\sim 2.6$ based on the CO(1-0) intensity mapping experiment by \citet{keating16}\footnote{For a CO intensity mapping experiment based on the ASPECS data, see \citet{carilli16}.}: By assuming a linear relation between the CO luminosity of galaxies and their dark matter halo mass, they interpret their CO power spectrum constraint in terms of $\rho$(H$_2$) $<2.6\times10^{8}$\,M$_\odot${}\,Mpc$^{-1}$ (at 1-$\sigma$). They further tighten the constraint on $\rho$(H$_2$) by assuming that the $L_{\rm CO}$--dark matter halo mass relation has a scatter of 0.37 dex (a factor $\approx 2.3$), which translates into $\rho$(H$_2$)=$1.1_{-0.4}^{+0.7}\times10^8$\,M$_\odot${}\,Mpc$^{-1}$, in excellent agreement with our measurement. Our findings are also consistent with the global increase of the gas fraction as a function of redshift found in targeted observations \citep[e.g.,][]{daddi10a,riechers10,tacconi10,tacconi13,genzel10,genzel15,geach11,magdis12,magnelli12}, although we find a large variety in the gas fraction in individual sources \citep[see][]{decarli16}. Our results are also in general agreement with the expectations from semi-analytical models \citep{obreschkow09a,obreschkow09b,lagos11,lagos12,popping14a,popping14b} and from empirical predictions \citep{sargent12,sargent14}. From the present data, there is an indication for a decrease of $\rho$(H$_2$) at $z>3$, as suggested by some models\footnote{The $\rho$(H$_2$) value at $z>3$ in the models by \citet{popping16} is lower than in the predictions in \citet{lagos11}. This might be surprising as the CO(1-0) luminosity function in the former exceeds the one of the latter, especially at high redshift (see Fig.~\par\noindent\hangindent 20 pt{fig_co10_lf}). This discrepancy is explained with the non-trivial galaxy--to--galaxy variations of $\alpha_{\rm CO}$ in the two models.}. A larger sample of $z>3$ CO emitters with spectroscopically--confirmed redshifts, and covering more cosmic volume, is required in order to explore this redshift range.
\begin{figure*}
\begin{center}
\includegraphics[width=0.79\textwidth]{fig_rhoH2_z.png}\\
\caption{Comoving cosmic mass density of molecular gas in galaxies $\rho$(H$_2$) as a function of redshift, based on our molecular survey in the UDF. Our ASPECS constraints are displayed as red boxes. The vertical size indicates our uncertainties (see text for details). Our measurements are not extrapolated to account for the faint end of the molecular gas mass function. Since our observations sample the expected knee of the CO luminosity functions in the redshift bins of interest, the correction is expected to be small ($<2\times$). Semi-analytical model predictions by \citet{obreschkow09a,obreschkow09b}, \citet{lagos12} and \citet{popping14a,popping14b} are shown as lines; the empirical predictions by \citet{sargent14} are plotted as a grey area; the constraints by \citet{keating16} are displayed with triangles; the PdBI constraints \citep{walter14} are represented by cyan boxes. Our ALMA observations show an evolution in the cosmic density of molecular gas up to $z\sim 4.5$. The global molecular content of galaxies at the peak of galaxy formation appears 3--10$\times$ higher than in galaxies in the local universe, although large uncertainties remain due to the limited area that is covered.
}
\label{fig_rhoH2_z}\end{center}
\end{figure*}
\begin{table*}
\caption{\rm Redshift ranges covered in the molecular line scans, the corresponding comoving volume, the number of galaxies in each bin (accounting for different line identifications), and our constraints on the molecular gas content in galaxies $\rho$(H$_2$) and $\rho$(ISM).} \label{tab_rhoH2}
\begin{center}
\begin{tabular}{cccccccccccc}
\hline
Transition & $\nu_0$ & $z_{\rm min}$ & $z_{\rm max}$ & $\langle z \rangle$ & Volume & N(H$_2$) & log $\rho_{\rm min}$(H$_2$) & log $\rho_{\rm max}$(H$_2$) & N(ISM) & log $\rho_{\rm min}$(ISM) & log $\rho_{\rm max}$(ISM) \\
& [GHz] & & & & [Mpc$^3$] & & [M$_\odot${}\,Mpc$^{-3}$] & [M$_\odot${}\,Mpc$^{-3}$] & & [M$_\odot${}\,Mpc$^{-3}$] & [M$_\odot${}\,Mpc$^{-3}$] \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) &(11)& (12) \\
\hline
\multicolumn{12}{c}{1mm (212.032--272.001 GHz)}\\
CO(3-2) & 345.796 & 0.2713 & 0.6309 & 0.4858 & 314 & 1--2 & 6.56 & 7.76 & 2 & 6.36 & 7.18 \\
CO(4-3) & 461.041 & 0.6950 & 1.1744 & 0.9543 & 1028 & 0--5 & 6.83 & 7.73 & 5 & 7.13 & 7.60 \\
\hline
\multicolumn{12}{c}{3mm (84.176--114.928 GHz)}\\
CO(2-1) & 230.538 & 1.0059 & 1.7387 & 1.4277 & 1920 & 3 & 7.53 & 8.09 & 13 & 7.50 & 7.77 \\
CO(3-2) & 345.796 & 2.0088 & 3.1080 & 2.6129 & 3363 & 2--7 & 7.69 & 8.28 & 6 & 7.04 & 7.46 \\
CO(4-3) & 461.041 & 3.0115 & 4.4771 & 3.8030 & 4149 & 0--5 & 5.53 & 7.58 & 0 & -- & 6.21 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Estimates from dust continuum emission}
In Fig.~\par\noindent\hangindent 20 pt{fig_rhoH2_co_ism} we compare the constraints on $\rho$(H$_2$) inferred from CO with those on $\rho$(ISM) derived from the dust continuum in our observations of the UDF. These are derived following \citet{scoville14}. In brief, for each 1mm continuum source (see the companion paper \citealt{aravena16a}), the ISM mass is computed as:
\begin{equation}\label{eq_scoville}
\frac{M_{\rm ISM}}{10^{10}\,{\rm M_\odot}}=\frac{1.78}{(1+z)^{4.8}}\,\,\frac{S_\nu}{\rm mJy} \, \left(\frac{\nu}{\rm 350\,GHz}\right)^{-3.8} \, \frac{\Gamma_0}{\Gamma_{\rm RJ}} \, \left(\frac{D_{\rm L}}{\rm Gpc}\right)^2
\end{equation}
where $S_\nu$ is the observed continuum flux density, $\nu$ is the observing frequency (here, we adopt $\nu=242$\,GHz as the central frequency of the continuum image), $\Gamma_{\rm RJ}$ is a unitless correction factor that accounts for the deviation from the $\nu^2$ scaling of the Rayleigh-Jeans tail, $\Gamma_0=0.71$ is the tuning value obtained at low-$z$, and $D_{\rm L}$ is the luminosity distance \citep[see eq.~12 in][]{scoville14}. The dust temperature (implicit in the definition of $\Gamma_{\rm RJ}$), is set to 25\,K. The ISM masses obtained via eq.~\par\noindent\hangindent 20 pt{eq_scoville} for each galaxy detected in the continuum \citep[see][]{aravena16a} are then split in the same redshift bins used for the CO-based estimates, and summed. We include here all the sources detected down to S/N=3 in the 1mm continuum. Poissonian uncertainties are found again to dominate the estimates of $\rho$ (if model uncertainties are neglected). The values of $\rho$(ISM) obtained in this way are reported in Tab.~\par\noindent\hangindent 20 pt{tab_rhoH2}. We find that the ISM mass density estimates are roughly consistent (within the admittedly large uncertainties) with the CO-based estimates in the lower redshift bins ($z\sim 0.5$, $0.95$, and $1.4$), while discrepancies are found at $z>2$, where $\rho$(H$_2$) estimates based on CO tend to be larger than $\rho$(ISM) estimates based on dust. \citet{scoville15} present a different calibration of the recipe that would shift the dust-based mass estimates up by a factor $1.5$. However, even applying the more recent calibration would not be sufficient to significantly mitigate the discrepancy between CO-based and dust-based estimates of the gas mass at high redshift. In \citet{aravena16a} we show that all of our 1mm continuum sources detected at $>$3.5-$\sigma$ (except one) are at $z<2$. On the other hand, the redshift distribution of CO-detected galaxies in our sample extends well beyond $z$=2, thus leading to the discrepancy in the $\rho$ estimates at high redshift. Possible explanations for this difference might be related to the dust temperature and opacity, and to the adopted $\alpha_{\rm CO}$. A higher dust temperature in high-$z$ galaxies ($>40$\,K) would shift the dust emission towards higher frequencies, thus explaining the comparably lower dust emission observed at 1mm (at a fixed IR luminosity). Moreover, at $z=4$ our 1mm continuum observations sample the rest-frame $\sim 250$\,$\mu$m range, where dust might turn optically thick (thus leading to underestimates of the dust emission). Finally, we might be over-estimating molecular gas masses at high $z$ if the $\alpha_{\rm CO}$ factor is typically closer to the ULIRG/starburst value [$\alpha_{\rm CO}\approx 0.8$\,M$_\odot${}(K~km\,s$^{-1}$\,pc$^2$)$^{-1}$, see \citealt{daddi10b,bolatto13}]. However, the observed low CO excitation and faint IR luminosity do not support the ULIRG scenario for our high-$z$ galaxies. Furthermore, any metallicity evolution would yield a higher $\alpha_{\rm CO}$ at high $z$, instead of a lower one. In Paper IV we discuss the discrepancy between dust- and CO-based gas masses on a source-by-source basis.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_rhoH2_CO_ism.png}\\
\caption{Comparison between the CO-derived estimates of $\rho$(H$_2$) and the 1mm dust continuum-based estimates of $\rho$(ISM). The galaxies are binned in the same redshift bins as presented in Fig.~\par\noindent\hangindent 20 pt{fig_rhoH2_z}, as indicated by the color of the symbols. The one--to--one case is shown as a dashed line. The dust--based estimates agree with the CO-based estimates at $z<2$, but they seem to fall below line of unity case at higher redshifts.}
\label{fig_rhoH2_co_ism}
\end{figure}
\section{Summary and Discussion}\label{sec_discussion}
In this paper we use our ALMA molecular scans of the {\em Hubble} UDF in band~3 and band~6 to place blind constraints on the CO luminosity function up to $z\sim 4.5$. We provide constraints on the evolution of the cosmic molecular gas density as a function of redshift. This study is based on galaxies that have been blindly selected through their CO emission, and not through any other multi--wavelength property. The CO number counts have been corrected for by two parameters, {\em fidelity} and {\em completeness}, which take into account the number of false positive detections due to noise peaks and the fraction of lines that our algorithm successfully recovers in our data cubes from a parent population of known (artificial) lines.
We start by constructing CO luminosity functions for the respective rotational transitions of CO for both the 3\,mm and 1\,mm observations. We compare these measurements to models that also predict CO luminosities in various rotational transitions, i.e. no assumptions were made in comparing our measurements to the models. This comparison shows that our derived CO luminosity functions lie above the predictions in the 3\,mm band. On the other hand, in the 1\,mm band our measurements are comparable to the models. Together this implies that the observed galaxies are more gas--rich than currently attributed for in the models, but with lower excitation.
Accounting for a CO excitation characteristic of main--sequence galaxies at $z\sim 1$--2, we derive the CO luminosity function of the ground--transition of CO (J=1--0) from our observations. We do so only up the J=4 transition of CO, to ensure that our results are not too strongly affected by the excitation corrections that would dominate the analysis at higher J. We find an evolution in the CO(1-0) luminosity function compared with observations in the local universe, with an excess of CO-emitting sources at the bright end of the luminosity functions. This is in general agreement with first constraints on the CO intensity mapping from the literature. This evolution exceeds what is predicted by the current models. This discrepancy appear to be a common trait of models of galaxy formation: galaxies with $M_*>10^{10}$\,M$_\odot${} at $z=2-3$ are predicted to be 2--3 times less star forming than observed (see, e.g., the recent review by \citealt{somerville15}), and similarly less gas--rich (see the analysis in \citealt{popping15a,popping15b}).
The sensitivity of the ALMA observations reaches below the knee of the predicted CO luminosity functions (around 5$\times$10$^{9}$\,K~km\,s$^{-1}$\,pc$^2$) at all redshifts. We convert our luminosity measurements into molecular gas masses via a `Galactic' conversion factor. By summing the molecular gas masses obtained at each redshift, we obtain an estimate of the cosmic density of molecular gas in galaxies, $\rho$(H$_2$). Given the admittedly large uncertainties (mainly due to Poisson errors), and the unknown shape of the intrinsic CO luminosity functions, we do not extrapolate our measurements outside the range of CO luminosities (i.e., H$_2$ masses) covered in our survey.
We find an increase (factor of 3--10) of the cosmic density of molecular gas from $z\sim 0$ to $z\sim$ 2--3, albeit with large uncertainties given the limited statistics. This is consistent with previous findings that the gas mass fraction increases with redshift \citep[see, e.g.,][]{tacconi10,tacconi13,magdis12}. However our measurements have been derived in a completely different fashion, by simply counting the molecular gas that is present in a given cosmic volume, without any prior knowledge of the general galaxy population in the field. In this respect, our constraints on $\rho$(H$_2$) are actually lower limits, in the sense that they do not recover the full extent of the luminosity function. However, a) we do sample the predicted knee of the luminosity function in most of the redshift bins, suggesting that we recover a large part ($>$50\%) of the total CO luminosity per comoving volume; b) the fraction of the CO luminosity function missed because of our sensitivity cut is likely larger at higher redshift, i.e., correcting for the contribution of the faint end would make the evolution in $\rho$(H$_2$) even steeper.
We have also derived the molecular gas densities using the dust emission as a tracer for the molecular gas, following \citet{scoville14,scoville15}. The molecular gas densities derived from dust emission are generally smaller than but broadly consistent with those measured from CO at $z<2$, but that they might fall short at reproducing the predicted gas mass content of galaxies at $z>2$.
Our analysis demonstrates that CO-based gas mass estimates result in 3--10 times higher gas masses in galaxies at $z\sim2$ than in the local universe. The history of cosmic SFR \citep{madau14} appears to at least partially follow the evolution in molecular gas supply in galaxies. The remaining difference between the evolution of the SFR density (a factor of $\sim 20$) and the one of molecular gas (a factor of 3--10) may due to the shortened depletion time scales. A further contribution to this difference may be ascribed to cosmic variance. The UDF in general (and therefore also the region studied here) is found to be underdense at $z>3$ \citep[e.g., Fig.~14 in][]{beckwith06} and in IR-bright sources \citep{weiss09}.
The impact of cosmic variance can be estimated empirically from the comparison with the number counts of sources detected in the dust continuum \citep{aravena16a}, or analytically from the variance in the dark matter structures, coupled with the clustering bias of a given galaxy population \citep[see, e.g.,][]{somerville04}. \citet{trenti08} provide estimates of the cosmic variance as a function of field size, halo occupation fraction, survey completeness, and number of sources in a sample. For a $\Delta z=1$ bin centered at $z=2.5$, a 100\% halo occupation fraction and 5 sources detected over 1 arcmin$^2$ (i.e., roughly mimicing the $z\sim2.5$ bin in our analysis), the fractional uncertainty in the number counts due to cosmic variance is $\sim 20$\% ($\sim 60$\% if we include Poissonian fluctuations).
Already a factor 5 increase in target area (resulting in a field that is approximately the size of the {\em Hubble} eXtremely Deep Field, \citealt{illingworth13}), at similar depth, would beat down the uncertainties significantly ($\mathrel{\rlap{\lower 3pt \hbox{$\sim$}} \raise 2.0pt \hbox{$<$}} 30$\,\%, including Poissonian fluctuations). With ALMA now being fully operational, such an increase in areal coverage appears to be within reach.
\acknowledgements
We thank the anonymous referee for excellent feedback that improved the quality of the paper.
FW, IRS, and RJI acknowledge support through ERC grants COSMIC--DAWN, DUSTYGAL, and COSMICISM, respectively. M.A. acknowledges partial support from FONDECYT through grant 1140099. DR acknowledges support from the National Science Foundation under grant number AST-1614213 to Cornell University. FEB and LI acknowledge Conicyt grants Basal-CATA PFB--06/2007 and Anilo ACT1417. FEB also acknowledge support from FONDECYT Regular 1141218 (FEB), and the Ministry of Economy, Development, and Tourism's Millennium Science Initiative through grant IC120009, awarded to The Millennium Institute of Astrophysics, MAS. IRS also acknowledges support from STFC (ST/L00075X/1) and a Royal Society / Wolfson Merit award. Support for RD and BM was provided by the DFG priority program 1573 `The physics of the interstellar medium'. AK and FB acknowledge support by the Collaborative Research Council 956, sub-project A1, funded by the Deutsche Forschungsgemeinschaft (DFG). PI acknowledges Conict grants Basal-CATA PFB--06/2007 and Anilo ACT1417. RJA was supported by FONDECYT grant number 1151408.
This paper makes use of the following ALMA data: \dataset[ ADS/JAO.ALMA\# 2013.1.00146.S and 2013.1.00718.S.]{https://almascience.nrao.edu/aq/}. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The 3mm-part of the ASPECS project had been supported by the German ARC.
|
\section{Introduction}
The theory of general relativity by Einstein represents a wonderful combination of the
theory of gravitation and geometry, resulting in great formal beauty and mathematical
elegance. Einstein has taken both the general coordinate invariance (diffeomorphism invariance)
and the equivalence principle as the fundamental principle of his gravitational theory.
With the help of the Riemannian geometry, the only two fundamental principles fix the physical
content of general relativity almost completely and provide us with a playground for discussing
various cosmological aspects in the universe. Not to mention the recent discovery of
gravitational wave \cite{LIGO}, we have thus far watched an overwhelming success of Einstein's general
relativity both experimentally and theoretically.
Nevertheless, considerable efforts have been made in order to construct its alternative
theories from several reasons. This trend may be justified insofar as the unimodular gravity
is concerned \cite{Einstein}-\cite{Padilla} since the cosmological constant problem \cite{Weinberg},
which is one of the most difficult and important problems in modern theoretical physics,
might be solved within this class of the gravitational theory.
Among some aspects of the cosmological constant problem, we are mainly interested in the issue of radiative
instability of the cosmological constant: the necessity of fine-tuning the value of the cosmological
constant every time the higher-order loop corrections are added in perturbation theory. To resolve
this problem, unimodular gravity \cite{Einstein}-\cite{Padilla} has been put forward where the vacuum energy
and $\it{a \ forteriori}$ all potential energy are decoupled from gravity since in the unimodular condition
$\sqrt{-g} = 1$, the potential energy cannot couple to gravity at the action level. In this approach,
the value of the cosmological constant is not predicted theoretically but fixed by an initial condition.
\footnote{Recently, we have established a topological model where the Newton's constant is determined
by an initial condition \cite{Oda1}-\cite{Oda3}.}
However, in quantum field theories the unimodular condition must be properly implemented via the Lagrange
multiplier field. Then, radiative corrections modify the Lagrange multiplier field, which corresponds to
the cosmological constant in unimodular gravity, thereby rendering its initial value radiatively unstable.
To diminish the contribution of the radiative corrections to the cosmological constant, the Weyl symmetry, or equivalently,
the local conformal symmetry, could be added to the volume preserving diffeomorphisms, or equivalently, the transverse
diffeomorphisms (TDiff) of unimodular gravity \cite{Izawa}-\cite{Oda0}. We will henceforth call such the theory
the Weyl-transverse (WTDiff) gravity.
One of the purposes in this article is to study classical solutions in the WTDiff gravity. Even if the WTDiff
gravity is equivalent to general relativity at the classical level, as long as we know, nobody has explicitly
derived classical solutions within the framework of the WTDiff gravity. In particular, we wish to investigate
whether the Schwarzschild metric is included in the classical solutions of the WTDiff gravity. The Schwarzschild
solution is of particular importance since it corresponds to the basic one-body problem of classical astronomy.
Indeed, many of the reliable experimental verifications of Einstein equations are based on the Schwarzschild
line element.
This paper is organised as follows: In Section 2, we review the WTDiff gravity. To this aim, we start with
the conformally invariant scalar-tensor gravity, and then fix the gauge symmetries by different gauge
conditions \cite{Oda0}. One gauge condition leads to general relativity while the other gauge condition produces the
WTDiff gravity. This fact shows the classical equivalence between general relativity and the WTDiff gravity even if
local symmetries in the both theories are different. Moreover, we derive equations of motion, and check
that they are invariant under the Weyl transformation and the TDiff. In Section 3, we solve the equations of motion
of the WTDiff gravity in the static and spherically symmetric ansatz. We show that the Schwarzschild metric
in the Cartesian coordinate system is in fact a classical solution.
The final section is devoted to discussions.
\section{The Weyl-transverse (WTDiff) gravity}
We will start with the action of the Weyl-transverse (WTDiff) gravity in a class of unimodular gravity
in a general $n$ dimensional space-time \cite{Alvarez1}-\cite{Oda0}, which is given by
\footnote{We follow notation and conventions by Misner et al.'s textbook \cite{MTW}, for instance,
the flat Minkowski metric $\eta_{\mu\nu} = diag(-, +, +, +)$, the Riemann curvature tensor
$R^\mu \ _{\nu\alpha\beta} = \partial_\alpha \Gamma^\mu_{\nu\beta} - \partial_\beta \Gamma^\mu_{\nu\alpha}
+ \Gamma^\mu_{\sigma\alpha} \Gamma^\sigma_{\nu\beta} - \Gamma^\mu_{\sigma\beta} \Gamma^\sigma_{\nu\alpha}$,
and the Ricci tensor $R_{\mu\nu} = R^\alpha \ _{\mu\alpha\nu}$.
The reduced Planck mass is defined as $M_p = \sqrt{\frac{c \hbar}{8 \pi G}} = 2.4 \times 10^{18} GeV$.
Throughout this article, we adopt the reduced Planck units where we set $c = \hbar = M_p = 1$.
In this units, all quantities become dimensionless.
Finally, note that in the reduced Planck units, the Einstein-Hilbert Lagrangian density takes the form
${\cal L}_{EH} = \frac{1}{2} \sqrt{-g} R$.}
\begin{eqnarray}
S &=& \int d^n x \ {\cal L} \nonumber\\
&=& \frac{1}{2} \int d^n x \ |g|^{\frac{1}{n}} \left[ R + \frac{(n-1)(n-2)}{4n^2} \frac{1}{|g|^2}
g^{\mu\nu} \partial_\mu |g| \partial_\nu |g| \right],
\label{WTDiff Action 1}
\end{eqnarray}
where we have defined as $g = \det g_{\mu\nu} < 0$. This action (\ref{WTDiff Action 1}) turns out to be
invariant under not the full group of diffeomorphisms (Diff) but only the transverse diffeomorphisms (TDiff).
Moreover, it is worthwhile to notice that in spite of the existence of an explicit mass scale (the reduced
Planck mass $M_p = 1$ emerges in the overall constant $\frac{1}{2} M_p^{n-2}$ of the action (\ref{WTDiff Action 1})),
this action is also invariant under Weyl transformation. Actually, under the Weyl transformation
\begin{eqnarray}
g_{\mu\nu} \rightarrow g^\prime_{\mu\nu} = \Omega^2(x) g_{\mu\nu},
\label{Weyl transf}
\end{eqnarray}
the Lagrangian density in (\ref{WTDiff Action 1}) is changed as
\begin{eqnarray}
{\cal L}^\prime = {\cal L} - (n-1) \partial_\mu \left( |g|^{\frac{1}{n}}
g^{\mu\nu} \frac{1}{\Omega} \partial_\nu \Omega \right).
\label{L'}
\end{eqnarray}
In what follows, let us explain how to derive the action (\ref{WTDiff Action 1}) by beginning with
the conformally invariant scalar-tensor gravity since this derivation makes it possible to
clarify the equivalence between general relativity and the WTDiff gravity and derive equations
of motion of the WTDiff gravity in a concise manner.
The conformally invariant scalar-tensor gravity in $n$ space-time dimensions takes the form
\footnote{This conformally invariant gravity theory has a wide application in phenomenology
and cosmology \cite{Oda4}-\cite{Oda7}.}
\begin{eqnarray}
S = \int d^n x \ \sqrt{-g} \left[ \frac{n-2}{8(n-1)} \varphi^2 R + \frac{1}{2}
g^{\mu\nu} \partial_\mu \varphi \partial_\nu \varphi \right],
\label{Cof-inv S-T Action 1}
\end{eqnarray}
which is invariant under the Weyl transformation (\ref{Weyl transf}) of the metric tensor in addition to
the ghost-like scalar field $\varphi$ as
\begin{eqnarray}
\varphi \rightarrow \varphi^\prime = \Omega^{- \frac{n-2}{2}}(x) \varphi.
\label{Scalar Weyl transf}
\end{eqnarray}
The gauge condition $\varphi = 2 \sqrt{\frac{n-1}{n-2}}$ for the Weyl symmetry leads to the well-known
Einstein-Hilbert action of general relativity. On the other hand, the gauge condition
$\varphi = 2 \sqrt{\frac{n-1}{n-2}} |g|^{- \frac{n-2}{4n}}$ for the longitudinal diffeomorphism
results in the action (\ref{WTDiff Action 1}) of the WTDiff gravity. Thus, the WTDiff gravity is at least
classically equivalent to general relativity since the both actions are obtained via the different
choices of gauge condition from the same action (\ref{Cof-inv S-T Action 1}).
Here it is worth stressing that the latter gauge condition
is not for the Weyl transformation but for the longitudinal diffeomorphism. Actually, it is easy to see that
the latter gauge condition is invariant under the Weyl transformation but breaks the longitudinal
diffeomorphism.
In this context, it is useful to comment on the transverse diffeomorphisms and the unimodular condition.
First, let us start with the diffeomorphism invariance. Under the general coordinate transformation
or Diff, the metric tensor transforms as
\begin{eqnarray}
g_{\mu\nu}(x) \rightarrow g_{\mu\nu}^\prime(x^\prime) = \frac{\partial x^\alpha}{\partial x^{\mu \prime}}
\frac{\partial x^\beta}{\partial x^{\nu \prime}} g_{\alpha\beta}(x) \equiv J^\alpha_{\mu \prime}
J^\beta_{\nu \prime} g_{\alpha\beta}(x),
\label{Diff}
\end{eqnarray}
where the Jacobian matrix $J^\alpha_{\mu \prime}$, which is defined as $J^\alpha_{\mu \prime} =
\frac{\partial x^\alpha}{\partial x^{\mu \prime}}$, was introduced. Denoting the determinant of the Jacobian matrix as
$J = \det J^\alpha_{\mu \prime} = \det \frac{\partial x^\alpha}{\partial x^{\mu \prime}}$, taking the
determinant of Eq. (\ref{Diff}) gives us
\begin{eqnarray}
g^\prime(x^\prime) = J^2(x) g(x).
\label{J}
\end{eqnarray}
Then, the transverse diffeomorphisms (TDiff), or equivalently, the volume preserving diffeomorphisms,
are defined as a subgroup of the full diffeomorphisms such that the determinant of the Jacobian matrix is the unity
\begin{eqnarray}
J(x) = 1.
\label{Unimodular J}
\end{eqnarray}
With this conditon (\ref{Unimodular J}), the volume element is preserved under the Diff, and Eq. (\ref{J}) shows
that $g(x)$ is a dimensionless scalar field. The existence of such the dimensionless scalar field in the TDiff gravity,
which is a gravitational theory based on the TDiff instead of the Diff or the WTDiff, is thought to be a defect
since one cannot exclude any terms with the form of a polynomial of $g(x)$ from the action by the fundamental principles
of QFT's \cite{Oda0}. In the infinitesimal form of diffeomorphisms $x^\mu \rightarrow x^{\mu \prime} = x^\mu - \xi^\mu(x)$,
the TDiff can be expressed by
\begin{eqnarray}
\partial_\mu \xi^\mu = 0.
\label{TDiff}
\end{eqnarray}
The unimodular condition is of course a distinct notion from the TDiff, but is closely related to each other.
The unimodular condition is defined as
\begin{eqnarray}
g(x) = -1.
\label{Unimodularity}
\end{eqnarray}
This condition, together with Eq. (\ref{J}), implies the TDiff because of (\ref{Unimodular J}). (We have
assumed $J > 0$.) Also note that the unimodular condition (\ref{Unimodularity}) yields the condition such that
the variation of the metric tensor is traceless
\begin{eqnarray}
g^{\mu\nu} \delta g_{\mu\nu} = 0.
\label{Traceless variation}
\end{eqnarray}
In the case of the conformally invariant scalar-tensor gravity, one can construct a Weyl invariant metric
\begin{eqnarray}
\hat g_{\mu\nu} = \left( \frac{1}{2} \sqrt{\frac{n-2}{n-1}} \varphi \right)^{\frac{4}{n-2}} g_{\mu\nu}.
\label{Weyl invariant metric}
\end{eqnarray}
In taking the gauge condition $\varphi = 2 \sqrt{\frac{n-1}{n-2}} |g|^{- \frac{n-2}{4n}}$, this metric
is reduced to the form
\begin{eqnarray}
\hat g_{\mu\nu} = |g|^{-\frac{1}{n}} g_{\mu\nu}.
\label{Weyl invariant metric 2}
\end{eqnarray}
The metric tensor (\ref{Weyl invariant metric 2}) satisfies the unimodular condition (\ref{Unimodularity}),
so that because of the equation (\ref{Traceless variation}), the resultant equations of motion stemming from
the WTDiff gravity action (\ref{WTDiff Action 1}), become the traceless equations as shown shortly.
Armed with the knowledge of the TDiff, Diff and the unimodular condition, we are ready to show explicitly that the action
(\ref{WTDiff Action 1}) of the WTDiff gravity is indeed invariant under not the Diff but the TDiff.
For this purpose, let us perform the Diff in the Lagrangian density of (\ref{WTDiff Action 1}) whose result
is given by
\begin{eqnarray}
{\cal L}^\prime(x^\prime)
= \frac{1}{2} |J^2 g|^{\frac{1}{n}} \left[ R + \frac{(n-1)(n-2)}{4n^2} \frac{1}{|g|^2}
g^{\mu\nu} (\partial_\mu |g| + \frac{2 |g|}{J} \partial_\mu J)
(\partial_\nu |g| + \frac{2 |g|}{J} \partial_\nu J) \right].
\label{Diff of WTDiff Action}
\end{eqnarray}
It is obvious that the Lagrangian density ${\cal L}$ is not invariant under the Diff owing to
the presence of the terms with $J$, but when $J = 1$ as in (\ref{Unimodular J}) in the case of
the TDiff, the Lagrangian density ${\cal L}$ becomes invariant, which means that the TDiff are
in fact a symmetry of the action (\ref{WTDiff Action 1}) of the WTDiff gravity.
Next, we will derive the equations of motion for the WTDiff gravity (\ref{WTDiff Action 1}).
A method of the derivation is to work with the action (\ref{Cof-inv S-T Action 1}) of the
conformally invariant scalar-tensor gravity, derive its equations of motion, and then
substitute the gauge condition $\varphi = 2 \sqrt{\frac{n-1}{n-2}} |g|^{- \frac{n-2}{4n}}$
into them. After some calculations, it turns out that the action (\ref{Cof-inv S-T Action 1}) produces
the equations of motion for $g_{\mu\nu}$ and $\varphi$, respectively
\begin{eqnarray}
\frac{n-2}{8(n-1)} \left[ \varphi^2 G_{\mu\nu} + ( g_{\mu\nu} \Box - \nabla_\mu \nabla_\nu )
(\varphi^2)\right] = \frac{1}{4} g_{\mu\nu} \partial_\rho \varphi \partial^\rho \varphi
- \frac{1}{2} \partial_\mu \varphi \partial_\nu \varphi,
\label{Eq of motion of conf-ST 1}
\end{eqnarray}
and
\begin{eqnarray}
\frac{n-2}{4(n-1)} \varphi R = \Box \varphi,
\label{Eq of motion of conf-ST 2}
\end{eqnarray}
where $G_{\mu\nu} = R_{\mu\nu} - \frac{1}{2} g_{\mu\nu} R$ is the Einstein tensor
and $\Box \varphi = g^{\mu\nu} \nabla_\mu \nabla_\nu \varphi$. It is well-known that
the conformally invariant scalar-tensor gravity can be obtained from the Einstein-Hilbert action
via the Weyl-invariant metric $\hat g_{\mu\nu} \propto \varphi^{\frac{4}{n-2}} g_{\mu\nu}$, so the equation of
motion (\ref{Eq of motion of conf-ST 2}) for the $\it{spurion}$ field $\varphi$ should be not independent of
the equations of motion (\ref{Eq of motion of conf-ST 1}) for the metric tensor. In fact, taking the trace
part of Eq. (\ref{Eq of motion of conf-ST 1}) naturally leads to Eq. (\ref{Eq of motion of conf-ST 2}).
Thus, it is sufficient to take only the equations of motion (\ref{Eq of motion of conf-ST 1}) into consideration.
Substituting the gauge condition $\varphi = 2 \sqrt{\frac{n-1}{n-2}} |g|^{- \frac{n-2}{4n}}$ into
Eq. (\ref{Eq of motion of conf-ST 1}) gives us the equations of motion for the WTDiff gravity
\begin{eqnarray}
G_{\mu\nu}^T = \Delta_{\mu\nu}^T,
\label{Eq of motion of WTDiff}
\end{eqnarray}
where $G_{\mu\nu}^T \equiv R_{\mu\nu} - \frac{1}{n} g_{\mu\nu} R$ is the $\it{traceless}$ Einstein
tensor and $\Delta_{\mu\nu}^T$ is also a traceless object defined by
\begin{eqnarray}
\Delta_{\mu\nu}^T &=& \frac{(n-2)(2n-1)}{4n^2} \left[ \frac{1}{|g|^2} \partial_\mu |g| \partial_\nu |g|
- \frac{1}{n} g_{\mu\nu} \frac{1}{|g|^2} (\partial_\rho |g|)^2 \right] \nonumber\\
&-& \frac{n-2}{2n} \left[ \frac{1}{|g|} D_\mu D_\nu |g|
- \frac{1}{n} g_{\mu\nu} \frac{1}{|g|} D_\rho D^\rho |g| \right],
\label{Delta}
\end{eqnarray}
where we have defined $D_\mu D_\nu |g| = \partial_\mu \partial_\nu |g| - \Gamma^\rho_{\mu\nu} \partial_\rho |g|$.
The explicit existence of $g$ in $\Delta_{\mu\nu}^T$ clearly indicates that the equations of motion for
the WTDiff gravity are not invariant under the full Diff. Finally, note that as mentioned before,
Eq. (\ref{Eq of motion of WTDiff}) is purely a traceless equation.
The equations of motion for the WTDiff gravity, (\ref{Eq of motion of WTDiff}), are derived by starting with
the action of the conformally invariant scalar-tensor gravity which is invariant under both the Weyl transformation
and the Diff, but the gauge condition breaks Diff down to TDiff. Thus, the equations of motion (\ref{Eq of motion of WTDiff})
should be invariant under both the Weyl transformation and the TDiff. Let us demonstrate this fact by an explicit calculation.
Under the Weyl transformation (\ref{Weyl transf}), the traceless Einstein tensor $G_{\mu\nu}^T$
and $\Delta_{\mu\nu}^T$ are transformed by the same quantity
\begin{eqnarray}
G_{\mu\nu}^{T \prime} &=& G_{\mu\nu}^T + A_{\mu\nu}^T, \nonumber\\
\Delta_{\mu\nu}^{T \prime} &=& \Delta_{\mu\nu}^T + A_{\mu\nu}^T,
\label{Weyl transf of Eq}
\end{eqnarray}
where $A_{\mu\nu}^T$ is defined as
\begin{eqnarray}
A_{\mu\nu}^T = 2(n-2) \frac{1}{\Omega^2} \left[ \partial_\mu \Omega \partial_\nu \Omega
- \frac{1}{n} g_{\mu\nu} (\partial_\rho \Omega)^2 \right]
-(n-2) \frac{1}{\Omega} \left[ \nabla_\mu \nabla_\nu \Omega
- \frac{1}{n} g_{\mu\nu} \nabla_\rho \nabla^\rho \Omega \right].
\label{A}
\end{eqnarray}
It is therefore obvious that Eq. (\ref{Eq of motion of WTDiff}) is invariant under the Weyl transformation.
Next, let us perform the general coordinate transformation to Eq. (\ref{Eq of motion of WTDiff})
whose result is described as
\begin{eqnarray}
G_{\mu\nu}^{T \prime} - \Delta_{\mu\nu}^{T \prime}
&=& J_{\mu \prime}^\alpha J_{\nu \prime}^\beta \Biggl\{ G_{\alpha\beta}^T - \Delta_{\alpha\beta}^T
+ \frac{n-2}{2n} \biggl[ \frac{1}{n} \frac{1}{J|g|} (\partial_\alpha J \partial_\beta |g|
+ \partial_\beta J \partial_\alpha |g|)
\nonumber\\
&+& \frac{2(1-n)}{n} \frac{1}{J^2} \partial_\alpha J \partial_\beta J
+ \frac{2}{J} D_\alpha D_\beta J \biggr]
- \frac{n-2}{n^2} \biggl[ \frac{1}{n} \frac{1}{J|g|} \partial_\rho J \partial^\rho |g|
\nonumber\\
&+& \frac{1-n}{n} \frac{1}{J^2} (\partial_\rho J)^2 + \frac{1}{J} D_\rho D^\rho J \biggr]
g_{\alpha\beta} \Biggr\}.
\label{Diff of Eq}
\end{eqnarray}
From this expression, we see that (\ref{Eq of motion of WTDiff}) is not invariant under the Diff, but with $J=1$,
that is, under the TDiff, it becomes invariant. In this way, we have shown that Eq. (\ref{Eq of motion of WTDiff})
is invariant under the Weyl transformation as well as the TDiff.
\section{Schwarzschild solution}
In this section, we wish to show that the Schwarzschild metric is a classical solution to the equations
of motion of the WTDiff gravity, (\ref{Eq of motion of WTDiff}). Before doing so, let us study a general
feature of Eq. (\ref{Eq of motion of WTDiff}).
In attempting to analyse a structure of classical solutions to Eq. (\ref{Eq of motion of WTDiff}),
we soon realize that a notable feature of Eq. (\ref{Eq of motion of WTDiff}) is that $G_{\mu\nu}^T
= R_{\mu\nu} - \frac{1}{n} g_{\mu\nu} R$ in the LHS has a beautiful geometrical structure
whereas $\Delta_{\mu\nu}^T$ in the RHS has a ugly expression, and the presence of the metric determinant $g$ and
its derivative $D_\mu D_\nu |g|$ reflects the fact that the equations of motion are not invariant
under the Diff, but only the TDiff. In this respect, note that $D_\mu D_\nu |g|$ has a bizarre
transformation property. Thus, it is natural to fix the Weyl symmetry first by the gauge condition
\begin{eqnarray}
g = -1,
\label{g=-1}
\end{eqnarray}
which is nothing but the unimodular condition (\ref{Unimodularity}). Since the unimodular condition
naturally yields Eq. (\ref{Unimodular J}) as mentioned before, this gauge condition does not break
the TDiff.
Then, $\Delta_{\mu\nu}^T$ in the RHS of Eq. (\ref{Eq of motion of WTDiff}) trivially vanishes
so that we have the equations
\begin{eqnarray}
G_{\mu\nu}^T \equiv R_{\mu\nu} - \frac{1}{n} g_{\mu\nu} R = 0.
\label{Einstein spaces}
\end{eqnarray}
The space-time defined by Eq. (\ref{Einstein spaces}) is called Einstein spaces in four dimensions
and the study of the Riemannian spaces which are conformally related to Einstein spaces, has been
addressed for a long time \cite{Kozameh}. Together with the Bianchi identity, Eq. (\ref{Einstein spaces})
leads to
\begin{eqnarray}
\frac{n-2}{2n} \nabla_\mu R = 0,
\label{Einstein spaces 2}
\end{eqnarray}
implying the constant curvature spaces except in two dimensions.
Now we are willing to demonstrate that the Schwarzschild metric in the Cartesian coordinate system
is a classical solution to the equations of motion of the WTDiff gravity, (\ref{Eq of motion of WTDiff}).
Since we take the gauge condition (\ref{g=-1}) for the Weyl transformation, classical solutions
in which we are interested belong to a subgroup of Einstein spaces where the gauge condition (\ref{g=-1})
is imposed as an additional condition.
We wish to look for a gravitational field outside an isolated, static, spherically symmetric object
with mass $M$. In the far region from the isolated object, we assume that the metric tensor is in an asymptotically
Lorentzian form
\begin{eqnarray}
g_{\mu\nu} \rightarrow \eta_{\mu\nu} + {\cal {O}}\left(\frac{1}{r^{n-3}}\right),
\label{BC}
\end{eqnarray}
where $\eta_{\mu\nu}$ is the Minkowski metric, and the radial coordinate $r$ is defined as
\begin{eqnarray}
r = \sqrt{(x^1)^2 + (x^2)^2 + \cdots + (x^{n-1})^2} = \sqrt{(x^i)^2},
\label{radial}
\end{eqnarray}
with $i$ running over spatial coordinates ($i = 1, 2, \cdots, n-1$).
Let us recall that the most spherically symmetric line element in $n$ space-time dimensions
reads
\begin{eqnarray}
d s^2 = - A(r) d t^2 + B(r) (x^i d x^i)^2 + C(r) (d x^i)^2 + D(r) d t \ x^i d x^i,
\label{Line element 1}
\end{eqnarray}
where $A(r)$ and $C(r)$ are positive functions depending on only $r$. Requiring
the invariance under the time reversal $t \rightarrow -t$ leads to $D = 0$. As is well-known,
we can set $C(r) = 1$ by redefining the radial coordinate $r$ \cite{Adler}. Thus, the line
element under consideration takes the form in the Cartesian coordinate system
\begin{eqnarray}
d s^2 = - A(r) d t^2 + (d x^i)^2 + B(r) (x^i d x^i)^2.
\label{Line element 2}
\end{eqnarray}
From this line element (\ref{Line element 2}), the non-vanishing components of the metric tensor
are given by
\begin{eqnarray}
g_{tt} = - A, \quad g_{ij} = \delta_{ij} + B x^i x^j,
\label{Metric}
\end{eqnarray}
and the components of its inverse matrix are
\begin{eqnarray}
g^{tt} = - \frac{1}{A}, \quad g^{ij} = \delta^{ij} - \frac{B}{1 + B r^2} x^i x^j.
\label{Inverse Metric}
\end{eqnarray}
Moreover, using these components of the metric tensor, the affine connection is calculated to be
\begin{eqnarray}
\Gamma^t_{ti} &=& \frac{A^\prime}{2A} \frac{x^i}{r}, \quad
\Gamma^i_{tt} = \frac{A^\prime}{2(1 + B r^2)} \frac{x^i}{r}, \nonumber\\
\Gamma^i_{jk} &=& \frac{1}{2(1 + B r^2)} \frac{x^i}{r} ( 2B r \delta_{jk} + B^\prime x^j x^k ),
\label{Affine}
\end{eqnarray}
where we have defined $A^\prime = \frac{dA}{dr}$, for instance.
At this stage, let us take the gauge condition (\ref{g=-1}) for the Weyl transformation.
By means of the metric tensor (\ref{Metric}), the gauge condition (\ref{g=-1}) is cast to the form
\begin{eqnarray}
A (1 + B r^2) = 1.
\label{AB}
\end{eqnarray}
Using this gauge condition (\ref{AB}) and Eqs. (\ref{Metric})-(\ref{Affine}), the Ricci tensor
and the scalar curvature can be easily calculated to be
\begin{eqnarray}
R_{tt} &=& \frac{1}{2} A ( A^{\prime\prime} + \frac{n-2}{r} A^\prime ), \nonumber\\
R_{ij} &=& \left[ \frac{n-3}{r^2} ( 1 - A ) - \frac{A^\prime}{r} \right] \delta_{ij}
+ \frac{1}{r^2} \Biggl[ \frac{n-3}{r^2} ( A - 1 )
+ \frac{1}{r} \frac{A^\prime}{A} ( 1 - \frac{n}{2} + A ) - \frac{1}{2} \frac{A^{\prime\prime}}{A} \Biggr]
x^i x^j,
\nonumber\\
R &=& - A^{\prime\prime} - \frac{2(n-2)}{r} A^\prime - \frac{(n-2)(n-3)}{r^2} ( A - 1 ).
\label{Curvature}
\end{eqnarray}
These results produce the concrete expressions for the non-vanishing components of the traceless Einstein
tensor $G_{\mu\nu}^T \equiv R_{\mu\nu} - \frac{1}{n} g_{\mu\nu} R$
\begin{eqnarray}
G_{tt}^T &=& \left( \frac{1}{2} - \frac{1}{n} \right) A \left[ A^{\prime\prime} + (n-4) \frac{1}{r} A^\prime
- 2 (n-3) \frac{1}{r^2} ( A - 1 ) \right], \nonumber\\
G_{ij}^T &=& \left\{ \frac{1}{n} \delta_{ij} + \frac{1}{r^2} \frac{1}{A} \left[ -\frac{1}{2} + \frac{1}{n}( 1 - A ) \right]
x^i x^j \right\} \Biggl[ A^{\prime\prime} + (n-4) \frac{1}{r} A^\prime
\nonumber\\
&-& 2 (n-3) \frac{1}{r^2} ( A - 1 ) \Biggr].
\label{G^T}
\end{eqnarray}
As a result, Eq. (\ref{Einstein spaces}) reduces to the equation
\begin{eqnarray}
A^{\prime\prime} + (n-4) \frac{1}{r} A^\prime - 2 (n-3) \frac{1}{r^2} ( A - 1 ) = 0.
\label{A-eq}
\end{eqnarray}
This equation can be exactly solved by noticing that it is written as
\begin{eqnarray}
&{}& A^{\prime\prime} + (n-4) \frac{1}{r} A^\prime - 2 (n-3) \frac{1}{r^2} ( A - 1 )
\nonumber\\
&=& \frac{1}{r^{n-3}} \frac{d^2}{d r^2} \left[ r^{n-3} ( A - 1 ) \right]
- (n-2) \frac{1}{r^{n-2}} \frac{d}{d r} \left[ r^{n-3} ( A - 1 ) \right].
\label{A-eq 2}
\end{eqnarray}
Hence, $A(r)$ is given by
\begin{eqnarray}
A(r) = 1 - \frac{2M}{r^{n-3}} + a r^2,
\label{A-value}
\end{eqnarray}
where $M$ and $a$ are integration constants. From the boundary condition (\ref{BC}),
we have to choose $a = 0$, and we can obtain the expression for $B(r)$ in terms of the gauge
condition (\ref{AB}). Consequently, we arrive at the expressions for $A(r)$ and $B(r)$
\begin{eqnarray}
A(r) = 1 - \frac{2M}{r^{n-3}}, \quad B(r) = \frac{2M}{r^2 (r^{n-3} - 2M)}.
\label{AB-value}
\end{eqnarray}
Then, the line element is of form
\begin{eqnarray}
d s^2 = - \left( 1 - \frac{2M}{r^{n-3}} \right) d t^2 + (d x^i)^2
+ \frac{2M}{r^2 (r^{n-3} - 2M)} (x^i d x^i)^2.
\label{Schwarzschild 1}
\end{eqnarray}
Accordingly, we have succeeded in showing that the Schwarzschild metric in the Cartesian coordinate system
is a classical solution in the WTDiff gravity as in general relativity.
Here we should refer to an important remark. The Schwarzschild metric in the Cartesian coordinate system,
(\ref{Schwarzschild 1}) can be rewritten in the spherical coordinate system as
\begin{eqnarray}
d s^2 = - \left( 1 - \frac{2M}{r^{n-3}} \right) d t^2
+ \frac{1}{1 - \frac{2M}{r^{n-3}}} d r^2 + r^2 d \Omega_{n-2}^2,
\label{Schwarzschild 2}
\end{eqnarray}
where
\begin{eqnarray}
d \Omega_{n-2}^2 = d \theta_2^2 + \sin^2 \theta_2 d \theta_3^2 + \cdots
+ \prod_{i=2}^{n-2} \sin^2 \theta_i d \theta_{n-1}^2.
\label{Omega}
\end{eqnarray}
This form of the Schwarzschild metric is very familiar with physicists, but this is not a classical
solution to the equations of motion of the WTDiff gravity, (\ref{Eq of motion of WTDiff}). The reason
is that when transforming from the Cartesian coordinates to the spherical coordinates, we have
a non-vanishing Jacobian factor which is against the TDiff. To put differently, while the determinant of the
metric tensor in Eq. (\ref{Schwarzschild 1}) is $-1$, the one in Eq. (\ref{Schwarzschild 2}) is not so,
which is against the gauge condition (\ref{g=-1}).
\section{Discussions}
In this article, in order to have the WTDiff gravity, starting with the conformally invariant scalar-tensor
gravity in a general space-time dimension which is invariant under both the local Weyl transformation
and the diffeomorphisms (Diff),
we have gauge-fixed the longitudinal diffeomorphism, by which the full diffeomorphisms (Diff) are broken
to the transverse diffeomorphisms (TDiff). It is explicitly checked that not only the resultant action of
the WTDiff gravity but also its equations of motion are invariant under both the local Weyl transformation
and the TDiff.
Moreover, we have studied classical solutions of the WTDiff gravity, and found that the Schwarzschild
metric is certainly a solution when the metric is expressed in terms of the Cartesian coordinate system.
It is of interest to note that the familiar Schwarzschild metric in the spherical coordinate system is
not a classical solution in the WTDiff gravity. This dependence on the coordinate systems of the classical
solutions is a general feature in the WTDiff gravity since when fixing the Weyl symmetry by a gauge condition,
only the TDiff are remained in the WTDiff gravity.
For instance, in the WTDiff gravity, if we rewrite a flat Minkowski space-time in the spherical coordinates,
the resulting line element is not a classical solution owing to the nonvanishing Jacobian factor
even if it is a solution in the Cartesian coordinates.
As a future problem, it might be possible to show that the Reissner-Nordstrom metric, which is
a static solution to the Einstein-Maxwell field equations, is a classical solution in the WTDiff gravity
in {\it{four}} space-time dimensions. We wish to consider this problem in near future.
\begin{flushleft}
{\bf Acknowledgements}
\end{flushleft}
We wish to thank M. Maeno, K. Taniguchi and K. Uryu for discussions.
This work is supported in part by the Grant-in-Aid for Scientific
Research (C) No. 16K05327 from the Japan Ministry of Education, Culture,
Sports, Science and Technology.
|
\section*{Acknowledgements}
\noindent We express our gratitude to our colleagues in the CERN
accelerator departments for the excellent performance of the LHC. We
thank the technical and administrative staff at the LHCb
institutes. We acknowledge support from CERN and from the national
agencies: CAPES, CNPq, FAPERJ and FINEP (Brazil); NSFC (China);
CNRS/IN2P3 (France); BMBF, DFG and MPG (Germany); INFN (Italy);
FOM and NWO (The Netherlands); MNiSW and NCN (Poland); MEN/IFA (Romania);
MinES and FANO (Russia); MinECo (Spain); SNSF and SER (Switzerland);
NASU (Ukraine); STFC (United Kingdom); NSF (USA).
We acknowledge the computing resources that are provided by CERN, IN2P3 (France), KIT and DESY (Germany), INFN (Italy), SURF (The Netherlands), PIC (Spain), GridPP (United Kingdom), RRCKI and Yandex LLC (Russia), CSCS (Switzerland), IFIN-HH (Romania), CBPF (Brazil), PL-GRID (Poland) and OSC (USA). We are indebted to the communities behind the multiple open
source software packages on which we depend.
Individual groups or members have received support from AvH Foundation (Germany),
EPLANET, Marie Sk\l{}odowska-Curie Actions and ERC (European Union),
Conseil G\'{e}n\'{e}ral de Haute-Savoie, Labex ENIGMASS and OCEVU,
R\'{e}gion Auvergne (France), RFBR and Yandex LLC (Russia), GVA, XuntaGal and GENCAT (Spain), Herchel Smith Fund, The Royal Society, Royal Commission for the Exhibition of 1851 and the Leverhulme Trust (United Kingdom).
\section{Detector and simulation}
\label{sec:Detector}
The \mbox{LHCb}\xspace detector~\cite{Alves:2008zz,LHCb-DP-2014-002} is a single-arm forward
spectrometer covering the \mbox{pseudorapidity} range $2<\eta <5$,
designed for the study of particles containing {\ensuremath{\Pb}}\xspace or {\ensuremath{\Pc}}\xspace
quarks. The detector includes a high-precision tracking system
consisting of a silicon-strip vertex detector surrounding the $pp$
interaction region, a large-area silicon-strip detector located
upstream of a dipole magnet with a bending power of about
$4{\mathrm{\,Tm}}$, and three stations of silicon-strip detectors and straw
drift tubes placed downstream of the magnet.
The tracking system provides a measurement of charged particle momentum, $p$, with
a relative uncertainty that varies from 0.5\% at low momentum to 1.0\% at 200\ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace.
The minimum distance of a track to a primary vertex (PV), the impact parameter, is measured with a resolution of $(15+29/\mbox{$p_{\mathrm{ T}}$}\xspace)\ensuremath{{\,\upmu\mathrm{m}}}\xspace$,
where \mbox{$p_{\mathrm{ T}}$}\xspace (in GeV/$c$) is the component of the momentum transverse to the beam direction.
Different types of charged hadrons are distinguished using information
from two ring-imaging Cherenkov (RICH\xspace) detectors.
Photons, electrons and hadrons are identified by a calorimeter system consisting of
scintillating-pad and preshower detectors, an electromagnetic
calorimeter and a hadronic calorimeter.
Muons are identified by a
system composed of alternating layers of iron and multiwire
proportional chambers.
The trigger comprises a hardware stage and a software stage.
The hardware trigger employed in this analysis uses information
from the muon system to select single muons or muon pairs,
applying $\mbox{$p_{\mathrm{ T}}$}\xspace$ requirements.
The subsequent software trigger is composed of two stages, the first of which
performs a partial reconstruction and requires either a pair of well-reconstructed,
oppositely charged muons having an invariant mass above $2.7 \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace$, or a
single well-reconstructed muon. The second stage of the software trigger applies
a full event reconstruction, and requires at least one of the following two
conditions to be fulfilled: either two opposite-sign muons must form a
good-quality vertex that is well separated from all of the primary vertices and
must have an invariant mass within $120\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$ of the known ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace$ mass
\cite{PDG2014}, or an algorithm using a boosted decision tree must identify a
two- or three-track vertex that is well separated from all of the primary
vertices and includes a muon among the constituent tracks.
The same trigger requirements are used to select both ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\kaon^+}}\xspace$ and
${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\pion^+}}\xspace$ decays, due to the similarity in their kinematic distributions.
In the simulation, $pp$ collisions are generated using \mbox{\textsc{Pythia}}\xspace
6~\cite{Sjostrand:2006za} with a specific \mbox{LHCb}\xspace
configuration~\cite{LHCb-PROC-2010-056}, or, for the hard process
$gg\ensuremath{\rightarrow}\xspace{\ensuremath{\B_\cquark^+}}\xspace+b+{\ensuremath{\overline \cquark}}\xspace$ that is the dominant source of ${\ensuremath{\B_\cquark^+}}\xspace$ mesons, using the
dedicated generator \mbox{\textsc{Bcvegpy}}\xspace~\cite{Chang:2003cq,Chang:2005hq}. Decays of
hadronic particles are described by \mbox{\textsc{EvtGen}}\xspace~\cite{Lange:2001uf}, in which
final-state radiation is generated using \mbox{\textsc{Photos}}\xspace~\cite{Golonka:2005pn}. The
interaction of the generated particles with the detector, and its response, are
implemented using the \mbox{\textsc{Geant4}}\xspace toolkit~\cite{Allison:2006ve, *Agostinelli:2002hh}
as described in Ref.~\cite{LHCb-PROC-2011-006}.
\section{Introduction}
\label{sec:Introduction}
The {\ensuremath{\B_\cquark^+}}\xspace meson,
the lightest $\bar{b} c$ bound state, can only decay weakly.
Since it contains only heavy quarks,
its decays can be analysed
using various theoretical approaches, including
QCD-based methods~\cite{Kiselev:2000pp, Jones:1998ub, Kiselev:1999sc} and
QCD-inspired phenomenological models~\cite{Eichten:1994gt,Naimuddin:2012dy}.
A measurement of the weak decay properties of ${\ensuremath{\B_\cquark^+}}\xspace$ mesons can test these approaches and
provide insight into the dynamics of the heavy quarks in the ${\ensuremath{\B_\cquark^+}}\xspace$ meson.
The exclusive decay\footnote{The inclusion of charge-conjugate processes is implied throughout the paper.} $\ensuremath{\Bcp\to\jpsi\Kp}\xspace$ is of particular interest since it
proceeds via a ${\ensuremath{\overline \bquark}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\overline \cquark}}\xspace{\ensuremath{\Pu}}\xspace{\ensuremath{\overline \squark}}\xspace$ transition and thus is CKM-suppressed
by a factor $|V_{us}/V_{ud}|^2\sim0.05$ with respect to $\ensuremath{\Bcp\to\jpsi\pip}\xspace$,
where the dominant amplitude is a ${\ensuremath{\overline \bquark}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\overline \cquark}}\xspace{\ensuremath{\Pu}}\xspace{\ensuremath{\overline \dquark}}\xspace$ transition.
In addition to the CKM matrix elements, the ratio of branching fractions
$\ensuremath{{R}_{K/\pi}}\xspace \equiv {\ensuremath{\mathcal{B}}}\xspace(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)/{\ensuremath{\mathcal{B}}}\xspace(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$ depends on the form factors of the two decays.
Theoretical calculations of $\ensuremath{{R}_{K/\pi}}\xspace$ have been carried out
using approaches that handle the non-factorisable and non-perturbative
contributions in different ways, yielding values in the range
from $0.05$ to $0.10$~\cite{Chang:1992pt, Liu:1997hr,Anisimov:1998uk,
Colangelo:1999zn, AbdElHady:1999xh, Kiselev:2000pp, Kiselev:2002vz, Ebert:2003cn,
Ivanov:2006ni, Naimuddin:2012dy, Qiao:2012hp,Ke:2013yka}.
The decay $\ensuremath{\Bcp\to\jpsi\Kp}\xspace$ was first observed by the \mbox{LHCb}\xspace collaboration,
which reported a measurement of $\ensuremath{{R}_{K/\pi}}\xspace=0.069\pm0.019\pm0.005$~\cite{LHCb-PAPER-2013-021}.
The uncertainty on this value is too large to discriminate between the predictions quoted above.
The \ensuremath{pp}\xspace data sample used in Ref. \cite{LHCb-PAPER-2013-021}, taken at a centre-of-mass
energy of $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and corresponding to an integrated luminosity of $1\ensuremath{\mbox{\,fb}^{-1}}\xspace$,
is now reanalysed in this paper
together with an additional sample taken at a centre-of-mass energy
of $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and corresponding to an integrated luminosity of $2\ensuremath{\mbox{\,fb}^{-1}}\xspace$.
Owing to improvements in the analysis method as well as the increase in the data sample size,
the statistical uncertainty is reduced by a factor of more than two.
The systematic uncertainty is also reduced.
\section{List of all symbols}
\label{sec:listofsymbols}
\subsection{Experiments}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash lhcb} & \mbox{LHCb}\xspace & \texttt{\textbackslash atlas} & \mbox{ATLAS}\xspace & \texttt{\textbackslash cms} & \mbox{CMS}\xspace \\
\texttt{\textbackslash alice} & \mbox{ALICE}\xspace & \texttt{\textbackslash babar} & \mbox{BaBar}\xspace & \texttt{\textbackslash belle} & \mbox{Belle}\xspace \\
\texttt{\textbackslash cleo} & \mbox{CLEO}\xspace & \texttt{\textbackslash cdf} & \mbox{CDF}\xspace & \texttt{\textbackslash dzero} & \mbox{D0}\xspace \\
\texttt{\textbackslash aleph} & \mbox{ALEPH}\xspace & \texttt{\textbackslash delphi} & \mbox{DELPHI}\xspace & \texttt{\textbackslash opal} & \mbox{OPAL}\xspace \\
\texttt{\textbackslash lthree} & \mbox{L3}\xspace & \texttt{\textbackslash sld} & \mbox{SLD}\xspace & \texttt{\textbackslash cern} & \mbox{CERN}\xspace \\
\texttt{\textbackslash lhc} & \mbox{LHC}\xspace & \texttt{\textbackslash lep} & \mbox{LEP}\xspace & \texttt{\textbackslash tevatron} & Tevatron\xspace \\
\end{tabular*}
\subsubsection{LHCb sub-detectors and sub-systems}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash velo} & VELO\xspace & \texttt{\textbackslash rich} & RICH\xspace & \texttt{\textbackslash richone} & RICH1\xspace \\
\texttt{\textbackslash richtwo} & RICH2\xspace & \texttt{\textbackslash ttracker} & TT\xspace & \texttt{\textbackslash intr} & IT\xspace \\
\texttt{\textbackslash st} & ST\xspace & \texttt{\textbackslash ot} & OT\xspace & \texttt{\textbackslash spd} & SPD\xspace \\
\texttt{\textbackslash presh} & PS\xspace & \texttt{\textbackslash ecal} & ECAL\xspace & \texttt{\textbackslash hcal} & HCAL\xspace \\
\texttt{\textbackslash MagUp} & \mbox{\em Mag\kern -0.05em Up}\xspace & \texttt{\textbackslash MagDown} & \mbox{\em MagDown}\xspace & \texttt{\textbackslash ode} & ODE\xspace \\
\texttt{\textbackslash daq} & DAQ\xspace & \texttt{\textbackslash tfc} & TFC\xspace & \texttt{\textbackslash ecs} & ECS\xspace \\
\texttt{\textbackslash lone} & L0\xspace & \texttt{\textbackslash hlt} & HLT\xspace & \texttt{\textbackslash hltone} & HLT1\xspace \\
\texttt{\textbackslash hlttwo} & HLT2\xspace & \\
\end{tabular*}
\subsection{Particles}
\subsubsection{Leptons}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash electron} & {\ensuremath{\Pe}}\xspace & \texttt{\textbackslash en} & \en & \texttt{\textbackslash ep} & {\ensuremath{\Pe^+}}\xspace \\
\texttt{\textbackslash epm} & \epm & \texttt{\textbackslash epem} & {\ensuremath{\Pe^+\Pe^-}}\xspace & \texttt{\textbackslash muon} & {\ensuremath{\Pmu}}\xspace \\
\texttt{\textbackslash mup} & {\ensuremath{\Pmu^+}}\xspace & \texttt{\textbackslash mun} & \mun & \texttt{\textbackslash mumu} & {\ensuremath{\Pmu^+\Pmu^-}}\xspace \\
\texttt{\textbackslash tauon} & {\ensuremath{\Ptau}}\xspace & \texttt{\textbackslash taup} & {\ensuremath{\Ptau^+}}\xspace & \texttt{\textbackslash taum} & {\ensuremath{\Ptau^-}}\xspace \\
\texttt{\textbackslash tautau} & {\ensuremath{\Ptau^+\Ptau^-}}\xspace & \texttt{\textbackslash lepton} & {\ensuremath{\ell}}\xspace & \texttt{\textbackslash ellm} & {\ensuremath{\ell^-}}\xspace \\
\texttt{\textbackslash ellp} & {\ensuremath{\ell^+}}\xspace & \texttt{\textbackslash ellell} & \ensuremath{\ell^+ \ell^-}\xspace & \texttt{\textbackslash neu} & {\ensuremath{\Pnu}}\xspace \\
\texttt{\textbackslash neub} & {\ensuremath{\overline{\Pnu}}}\xspace & \texttt{\textbackslash neue} & {\ensuremath{\neu_e}}\xspace & \texttt{\textbackslash neueb} & {\ensuremath{\neub_e}}\xspace \\
\texttt{\textbackslash neum} & {\ensuremath{\neu_\mu}}\xspace & \texttt{\textbackslash neumb} & {\ensuremath{\neub_\mu}}\xspace & \texttt{\textbackslash neut} & {\ensuremath{\neu_\tau}}\xspace \\
\texttt{\textbackslash neutb} & {\ensuremath{\neub_\tau}}\xspace & \texttt{\textbackslash neul} & {\ensuremath{\neu_\ell}}\xspace & \texttt{\textbackslash neulb} & {\ensuremath{\neub_\ell}}\xspace \\
\end{tabular*}
\subsubsection{Gauge bosons and scalars}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash g} & {\ensuremath{\Pgamma}}\xspace & \texttt{\textbackslash H} & {\cal H} & \texttt{\textbackslash Hp} & {\ensuremath{\PH^+}}\xspace \\
\texttt{\textbackslash Hm} & {\ensuremath{\PH^-}}\xspace & \texttt{\textbackslash Hpm} & {\ensuremath{\PH^\pm}}\xspace & \texttt{\textbackslash W} & {\ensuremath{\PW}}\xspace \\
\texttt{\textbackslash Wp} & {\ensuremath{\PW^+}}\xspace & \texttt{\textbackslash Wm} & {\ensuremath{\PW^-}}\xspace & \texttt{\textbackslash Wpm} & {\ensuremath{\PW^\pm}}\xspace \\
\texttt{\textbackslash Z} & {\ensuremath{\PZ}}\xspace & \\
\end{tabular*}
\subsubsection{Quarks}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash quark} & {\ensuremath{\Pq}}\xspace & \texttt{\textbackslash quarkbar} & {\ensuremath{\overline \quark}}\xspace & \texttt{\textbackslash qqbar} & {\ensuremath{\quark\quarkbar}}\xspace \\
\texttt{\textbackslash uquark} & {\ensuremath{\Pu}}\xspace & \texttt{\textbackslash uquarkbar} & {\ensuremath{\overline \uquark}}\xspace & \texttt{\textbackslash uubar} & {\ensuremath{\uquark\uquarkbar}}\xspace \\
\texttt{\textbackslash dquark} & {\ensuremath{\Pd}}\xspace & \texttt{\textbackslash dquarkbar} & {\ensuremath{\overline \dquark}}\xspace & \texttt{\textbackslash ddbar} & {\ensuremath{\dquark\dquarkbar}}\xspace \\
\texttt{\textbackslash squark} & {\ensuremath{\Ps}}\xspace & \texttt{\textbackslash squarkbar} & {\ensuremath{\overline \squark}}\xspace & \texttt{\textbackslash ssbar} & {\ensuremath{\squark\squarkbar}}\xspace \\
\texttt{\textbackslash cquark} & {\ensuremath{\Pc}}\xspace & \texttt{\textbackslash cquarkbar} & {\ensuremath{\overline \cquark}}\xspace & \texttt{\textbackslash ccbar} & {\ensuremath{\cquark\cquarkbar}}\xspace \\
\texttt{\textbackslash bquark} & {\ensuremath{\Pb}}\xspace & \texttt{\textbackslash bquarkbar} & {\ensuremath{\overline \bquark}}\xspace & \texttt{\textbackslash bbbar} & {\ensuremath{\bquark\bquarkbar}}\xspace \\
\texttt{\textbackslash tquark} & {\ensuremath{\Pt}}\xspace & \texttt{\textbackslash tquarkbar} & {\ensuremath{\overline \tquark}}\xspace & \texttt{\textbackslash ttbar} & {\ensuremath{\tquark\tquarkbar}}\xspace \\
\end{tabular*}
\subsubsection{Light mesons}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash hadron} & {\ensuremath{\Ph}}\xspace & \texttt{\textbackslash pion} & {\ensuremath{\Ppi}}\xspace & \texttt{\textbackslash piz} & {\ensuremath{\pion^0}}\xspace \\
\texttt{\textbackslash pizs} & {\ensuremath{\pion^0\mbox\,\mathrm{s}}}\xspace & \texttt{\textbackslash pip} & {\ensuremath{\pion^+}}\xspace & \texttt{\textbackslash pim} & {\ensuremath{\pion^-}}\xspace \\
\texttt{\textbackslash pipm} & {\ensuremath{\pion^\pm}}\xspace & \texttt{\textbackslash pimp} & {\ensuremath{\pion^\mp}}\xspace & \texttt{\textbackslash rhomeson} & {\ensuremath{\Prho}}\xspace \\
\texttt{\textbackslash rhoz} & {\ensuremath{\rhomeson^0}}\xspace & \texttt{\textbackslash rhop} & {\ensuremath{\rhomeson^+}}\xspace & \texttt{\textbackslash rhom} & {\ensuremath{\rhomeson^-}}\xspace \\
\texttt{\textbackslash rhopm} & {\ensuremath{\rhomeson^\pm}}\xspace & \texttt{\textbackslash rhomp} & {\ensuremath{\rhomeson^\mp}}\xspace & \texttt{\textbackslash kaon} & {\ensuremath{\PK}}\xspace \\
\texttt{\textbackslash Kb} & {\ensuremath{\Kbar}}\xspace & \texttt{\textbackslash KorKbar} & \kern 0.18em\optbar{\kern -0.18em K}{}\xspace & \texttt{\textbackslash Kz} & {\ensuremath{\kaon^0}}\xspace \\
\texttt{\textbackslash Kzb} & {\ensuremath{\Kbar{}^0}}\xspace & \texttt{\textbackslash Kp} & {\ensuremath{\kaon^+}}\xspace & \texttt{\textbackslash Km} & {\ensuremath{\kaon^-}}\xspace \\
\texttt{\textbackslash Kpm} & {\ensuremath{\kaon^\pm}}\xspace & \texttt{\textbackslash Kmp} & {\ensuremath{\kaon^\mp}}\xspace & \texttt{\textbackslash KS} & {\ensuremath{\kaon^0_{\mathrm{ \scriptscriptstyle S}}}}\xspace \\
\texttt{\textbackslash KL} & {\ensuremath{\kaon^0_{\mathrm{ \scriptscriptstyle L}}}}\xspace & \texttt{\textbackslash Kstarz} & {\ensuremath{\kaon^{*0}}}\xspace & \texttt{\textbackslash Kstarzb} & {\ensuremath{\Kbar{}^{*0}}}\xspace \\
\texttt{\textbackslash Kstar} & {\ensuremath{\kaon^*}}\xspace & \texttt{\textbackslash Kstarb} & {\ensuremath{\Kbar{}^*}}\xspace & \texttt{\textbackslash Kstarp} & {\ensuremath{\kaon^{*+}}}\xspace \\
\texttt{\textbackslash Kstarm} & {\ensuremath{\kaon^{*-}}}\xspace & \texttt{\textbackslash Kstarpm} & {\ensuremath{\kaon^{*\pm}}}\xspace & \texttt{\textbackslash Kstarmp} & {\ensuremath{\kaon^{*\mp}}}\xspace \\
\texttt{\textbackslash etaz} & \ensuremath{\Peta}\xspace & \texttt{\textbackslash etapr} & \ensuremath{\Peta^{\prime}}\xspace & \texttt{\textbackslash phiz} & \ensuremath{\Pphi}\xspace \\
\texttt{\textbackslash omegaz} & \ensuremath{\Pomega}\xspace & \\
\end{tabular*}
\subsubsection{Heavy mesons}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash D} & {\ensuremath{\PD}}\xspace & \texttt{\textbackslash Db} & {\ensuremath{\Dbar}}\xspace & \texttt{\textbackslash DorDbar} & \kern 0.18em\optbar{\kern -0.18em D}{}\xspace \\
\texttt{\textbackslash Dz} & {\ensuremath{\D^0}}\xspace & \texttt{\textbackslash Dzb} & {\ensuremath{\Dbar{}^0}}\xspace & \texttt{\textbackslash Dp} & {\ensuremath{\D^+}}\xspace \\
\texttt{\textbackslash Dm} & {\ensuremath{\D^-}}\xspace & \texttt{\textbackslash Dpm} & {\ensuremath{\D^\pm}}\xspace & \texttt{\textbackslash Dmp} & {\ensuremath{\D^\mp}}\xspace \\
\texttt{\textbackslash Dstar} & {\ensuremath{\D^*}}\xspace & \texttt{\textbackslash Dstarb} & {\ensuremath{\Dbar{}^*}}\xspace & \texttt{\textbackslash Dstarz} & {\ensuremath{\D^{*0}}}\xspace \\
\texttt{\textbackslash Dstarzb} & {\ensuremath{\Dbar{}^{*0}}}\xspace & \texttt{\textbackslash Dstarp} & {\ensuremath{\D^{*+}}}\xspace & \texttt{\textbackslash Dstarm} & {\ensuremath{\D^{*-}}}\xspace \\
\texttt{\textbackslash Dstarpm} & {\ensuremath{\D^{*\pm}}}\xspace & \texttt{\textbackslash Dstarmp} & {\ensuremath{\D^{*\mp}}}\xspace & \texttt{\textbackslash Ds} & {\ensuremath{\D^+_\squark}}\xspace \\
\texttt{\textbackslash Dsp} & {\ensuremath{\D^+_\squark}}\xspace & \texttt{\textbackslash Dsm} & {\ensuremath{\D^-_\squark}}\xspace & \texttt{\textbackslash Dspm} & {\ensuremath{\D^{\pm}_\squark}}\xspace \\
\texttt{\textbackslash Dsmp} & {\ensuremath{\D^{\mp}_\squark}}\xspace & \texttt{\textbackslash Dss} & {\ensuremath{\D^{*+}_\squark}}\xspace & \texttt{\textbackslash Dssp} & {\ensuremath{\D^{*+}_\squark}}\xspace \\
\texttt{\textbackslash Dssm} & {\ensuremath{\D^{*-}_\squark}}\xspace & \texttt{\textbackslash Dsspm} & {\ensuremath{\D^{*\pm}_\squark}}\xspace & \texttt{\textbackslash Dssmp} & {\ensuremath{\D^{*\mp}_\squark}}\xspace \\
\texttt{\textbackslash B} & {\ensuremath{\PB}}\xspace & \texttt{\textbackslash Bbar} & {\ensuremath{\kern 0.18em\overline{\kern -0.18em \PB}{}}}\xspace & \texttt{\textbackslash Bb} & {\ensuremath{\Bbar}}\xspace \\
\texttt{\textbackslash BorBbar} & \kern 0.18em\optbar{\kern -0.18em B}{}\xspace & \texttt{\textbackslash Bz} & {\ensuremath{\B^0}}\xspace & \texttt{\textbackslash Bzb} & {\ensuremath{\Bbar{}^0}}\xspace \\
\texttt{\textbackslash Bu} & {\ensuremath{\B^+}}\xspace & \texttt{\textbackslash Bub} & {\ensuremath{\B^-}}\xspace & \texttt{\textbackslash Bp} & {\ensuremath{\Bu}}\xspace \\
\texttt{\textbackslash Bm} & {\ensuremath{\Bub}}\xspace & \texttt{\textbackslash Bpm} & {\ensuremath{\B^\pm}}\xspace & \texttt{\textbackslash Bmp} & {\ensuremath{\B^\mp}}\xspace \\
\texttt{\textbackslash Bd} & {\ensuremath{\B^0}}\xspace & \texttt{\textbackslash Bs} & {\ensuremath{\B^0_\squark}}\xspace & \texttt{\textbackslash Bsb} & {\ensuremath{\Bbar{}^0_\squark}}\xspace \\
\texttt{\textbackslash Bdb} & {\ensuremath{\Bbar{}^0}}\xspace & \texttt{\textbackslash Bc} & {\ensuremath{\B_\cquark^+}}\xspace & \texttt{\textbackslash Bcp} & {\ensuremath{\B_\cquark^+}}\xspace \\
\texttt{\textbackslash Bcm} & {\ensuremath{\B_\cquark^-}}\xspace & \texttt{\textbackslash Bcpm} & {\ensuremath{\B_\cquark^\pm}}\xspace & \\
\end{tabular*}
\subsubsection{Onia}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash jpsi} & {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace & \texttt{\textbackslash psitwos} & {\ensuremath{\Ppsi{(2S)}}}\xspace & \texttt{\textbackslash psiprpr} & {\ensuremath{\Ppsi(3770)}}\xspace \\
\texttt{\textbackslash etac} & {\ensuremath{\Peta_\cquark}}\xspace & \texttt{\textbackslash chiczero} & {\ensuremath{\Pchi_{\cquark 0}}}\xspace & \texttt{\textbackslash chicone} & {\ensuremath{\Pchi_{\cquark 1}}}\xspace \\
\texttt{\textbackslash chictwo} & {\ensuremath{\Pchi_{\cquark 2}}}\xspace & \texttt{\textbackslash OneS} & {\Y1S} & \texttt{\textbackslash TwoS} & {\Y2S} \\
\texttt{\textbackslash ThreeS} & {\Y3S} & \texttt{\textbackslash FourS} & {\Y4S} & \texttt{\textbackslash FiveS} & {\Y5S} \\
\texttt{\textbackslash chic} & {\ensuremath{\Pchi_{c}}}\xspace & \\
\end{tabular*}
\subsubsection{Baryons}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash proton} & {\ensuremath{\Pp}}\xspace & \texttt{\textbackslash antiproton} & {\ensuremath{\overline \proton}}\xspace & \texttt{\textbackslash neutron} & {\ensuremath{\Pn}}\xspace \\
\texttt{\textbackslash antineutron} & {\ensuremath{\overline \neutron}}\xspace & \texttt{\textbackslash Deltares} & {\ensuremath{\PDelta}}\xspace & \texttt{\textbackslash Deltaresbar} & {\ensuremath{\overline \Deltares}}\xspace \\
\texttt{\textbackslash Xires} & {\ensuremath{\PXi}}\xspace & \texttt{\textbackslash Xiresbar} & {\ensuremath{\overline \Xires}}\xspace & \texttt{\textbackslash Lz} & {\ensuremath{\PLambda}}\xspace \\
\texttt{\textbackslash Lbar} & {\ensuremath{\kern 0.1em\overline{\kern -0.1em\PLambda}}}\xspace & \texttt{\textbackslash LorLbar} & \kern 0.18em\optbar{\kern -0.18em \PLambda}{}\xspace & \texttt{\textbackslash Lambdares} & {\ensuremath{\PLambda}}\xspace \\
\texttt{\textbackslash Lambdaresbar} & {\ensuremath{\Lbar}}\xspace & \texttt{\textbackslash Sigmares} & {\ensuremath{\PSigma}}\xspace & \texttt{\textbackslash Sigmaresbar} & {\ensuremath{\overline \Sigmares}}\xspace \\
\texttt{\textbackslash Omegares} & {\ensuremath{\POmega}}\xspace & \texttt{\textbackslash Omegaresbar} & {\ensuremath{\overline \POmega}}\xspace & \texttt{\textbackslash Lb} & {\ensuremath{\Lz^0_\bquark}}\xspace \\
\texttt{\textbackslash Lbbar} & {\ensuremath{\Lbar{}^0_\bquark}}\xspace & \texttt{\textbackslash Lc} & {\ensuremath{\Lz^+_\cquark}}\xspace & \texttt{\textbackslash Lcbar} & {\ensuremath{\Lbar{}^-_\cquark}}\xspace \\
\texttt{\textbackslash Xib} & {\ensuremath{\Xires_\bquark}}\xspace & \texttt{\textbackslash Xibz} & {\ensuremath{\Xires^0_\bquark}}\xspace & \texttt{\textbackslash Xibm} & {\ensuremath{\Xires^-_\bquark}}\xspace \\
\texttt{\textbackslash Xibbar} & {\ensuremath{\Xiresbar{}_\bquark}}\xspace & \texttt{\textbackslash Xibbarz} & {\ensuremath{\Xiresbar{}_\bquark^0}}\xspace & \texttt{\textbackslash Xibbarp} & {\ensuremath{\Xiresbar{}_\bquark^+}}\xspace \\
\texttt{\textbackslash Xic} & {\ensuremath{\Xires_\cquark}}\xspace & \texttt{\textbackslash Xicz} & {\ensuremath{\Xires^0_\cquark}}\xspace & \texttt{\textbackslash Xicp} & {\ensuremath{\Xires^+_\cquark}}\xspace \\
\texttt{\textbackslash Xicbar} & {\ensuremath{\Xiresbar{}_\cquark}}\xspace & \texttt{\textbackslash Xicbarz} & {\ensuremath{\Xiresbar{}_\cquark^0}}\xspace & \texttt{\textbackslash Xicbarm} & {\ensuremath{\Xiresbar{}_\cquark^-}}\xspace \\
\texttt{\textbackslash Omegac} & {\ensuremath{\Omegares^0_\cquark}}\xspace & \texttt{\textbackslash Omegacbar} & {\ensuremath{\Omegaresbar{}_\cquark^0}}\xspace & \texttt{\textbackslash Omegab} & {\ensuremath{\Omegares^-_\bquark}}\xspace \\
\texttt{\textbackslash Omegabbar} & {\ensuremath{\Omegaresbar{}_\bquark^+}}\xspace & \\
\end{tabular*}
\subsection{Physics symbols}
\subsubsection{Decays}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash BF} & {\ensuremath{\mathcal{B}}}\xspace & \texttt{\textbackslash BRvis} & {\ensuremath{\BF_{\mathrm{{vis}}}}} & \texttt{\textbackslash BR} & \BF \\
\texttt{\textbackslash decay[2] \textbackslash decay\{\Pa\}\{\Pb \Pc\}} & \decay{\Pa}{\Pb \Pc} & \texttt{\textbackslash ra} & \ensuremath{\rightarrow}\xspace & \texttt{\textbackslash to} & \ensuremath{\rightarrow}\xspace \\
\end{tabular*}
\subsubsection{Lifetimes}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash tauBs} & {\ensuremath{\tau_{{\ensuremath{\B^0_\squark}}\xspace}}}\xspace & \texttt{\textbackslash tauBd} & {\ensuremath{\tau_{{\ensuremath{\B^0}}\xspace}}}\xspace & \texttt{\textbackslash tauBz} & {\ensuremath{\tau_{{\ensuremath{\B^0}}\xspace}}}\xspace \\
\texttt{\textbackslash tauBu} & {\ensuremath{\tau_{{\ensuremath{\Bu}}\xspace}}}\xspace & \texttt{\textbackslash tauDp} & {\ensuremath{\tau_{{\ensuremath{\D^+}}\xspace}}}\xspace & \texttt{\textbackslash tauDz} & {\ensuremath{\tau_{{\ensuremath{\D^0}}\xspace}}}\xspace \\
\texttt{\textbackslash tauL} & {\ensuremath{\tau_{\mathrm{ L}}}}\xspace & \texttt{\textbackslash tauH} & {\ensuremath{\tau_{\mathrm{ H}}}}\xspace & \\
\end{tabular*}
\subsubsection{Masses}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash mBd} & {\ensuremath{m_{{\ensuremath{\B^0}}\xspace}}}\xspace & \texttt{\textbackslash mBp} & {\ensuremath{m_{{\ensuremath{\Bu}}\xspace}}}\xspace & \texttt{\textbackslash mBs} & {\ensuremath{m_{{\ensuremath{\B^0_\squark}}\xspace}}}\xspace \\
\texttt{\textbackslash mBc} & {\ensuremath{m_{{\ensuremath{\B_\cquark^+}}\xspace}}}\xspace & \texttt{\textbackslash mLb} & {\ensuremath{m_{{\ensuremath{\Lz^0_\bquark}}\xspace}}}\xspace & \\
\end{tabular*}
\subsubsection{EW theory, groups}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash grpsuthree} & {\ensuremath{\mathrm{SU}(3)}}\xspace & \texttt{\textbackslash grpsutw} & {\ensuremath{\mathrm{SU}(2)}}\xspace & \texttt{\textbackslash grpuone} & {\ensuremath{\mathrm{U}(1)}}\xspace \\
\texttt{\textbackslash ssqtw} & {\ensuremath{\sin^{2}\!\theta_{\mathrm{W}}}}\xspace & \texttt{\textbackslash csqtw} & {\ensuremath{\cos^{2}\!\theta_{\mathrm{W}}}}\xspace & \texttt{\textbackslash stw} & {\ensuremath{\sin\theta_{\mathrm{W}}}}\xspace \\
\texttt{\textbackslash ctw} & {\ensuremath{\cos\theta_{\mathrm{W}}}}\xspace & \texttt{\textbackslash ssqtwef} & {\ensuremath{{\sin}^{2}\theta_{\mathrm{W}}^{\mathrm{eff}}}}\xspace & \texttt{\textbackslash csqtwef} & {\ensuremath{{\cos}^{2}\theta_{\mathrm{W}}^{\mathrm{eff}}}}\xspace \\
\texttt{\textbackslash stwef} & {\ensuremath{\sin\theta_{\mathrm{W}}^{\mathrm{eff}}}}\xspace & \texttt{\textbackslash ctwef} & {\ensuremath{\cos\theta_{\mathrm{W}}^{\mathrm{eff}}}}\xspace & \texttt{\textbackslash gv} & {\ensuremath{g_{\mbox{\tiny V}}}}\xspace \\
\texttt{\textbackslash ga} & {\ensuremath{g_{\mbox{\tiny A}}}}\xspace & \texttt{\textbackslash order} & {\ensuremath{\mathcal{O}}}\xspace & \texttt{\textbackslash ordalph} & {\ensuremath{\mathcal{O}(\alpha)}}\xspace \\
\texttt{\textbackslash ordalsq} & {\ensuremath{\mathcal{O}(\alpha^{2})}}\xspace & \texttt{\textbackslash ordalcb} & {\ensuremath{\mathcal{O}(\alpha^{3})}}\xspace & \\
\end{tabular*}
\subsubsection{QCD parameters}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash as} & {\ensuremath{\alpha_s}}\xspace & \texttt{\textbackslash MSb} & {\ensuremath{\overline{\mathrm{MS}}}}\xspace & \texttt{\textbackslash lqcd} & {\ensuremath{\Lambda_{\mathrm{QCD}}}}\xspace \\
\texttt{\textbackslash qsq} & {\ensuremath{q^2}}\xspace & \\
\end{tabular*}
\subsubsection{CKM, CP violation}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash eps} & {\ensuremath{\varepsilon}}\xspace & \texttt{\textbackslash epsK} & {\ensuremath{\varepsilon_K}}\xspace & \texttt{\textbackslash epsB} & {\ensuremath{\varepsilon_B}}\xspace \\
\texttt{\textbackslash epsp} & {\ensuremath{\varepsilon^\prime_K}}\xspace & \texttt{\textbackslash CP} & {\ensuremath{C\!P}}\xspace & \texttt{\textbackslash CPT} & {\ensuremath{C\!PT}}\xspace \\
\texttt{\textbackslash rhobar} & {\ensuremath{\overline \rho}}\xspace & \texttt{\textbackslash etabar} & {\ensuremath{\overline \eta}}\xspace & \texttt{\textbackslash Vud} & {\ensuremath{V_{\uquark\dquark}}}\xspace \\
\texttt{\textbackslash Vcd} & {\ensuremath{V_{\cquark\dquark}}}\xspace & \texttt{\textbackslash Vtd} & {\ensuremath{V_{\tquark\dquark}}}\xspace & \texttt{\textbackslash Vus} & {\ensuremath{V_{\uquark\squark}}}\xspace \\
\texttt{\textbackslash Vcs} & {\ensuremath{V_{\cquark\squark}}}\xspace & \texttt{\textbackslash Vts} & {\ensuremath{V_{\tquark\squark}}}\xspace & \texttt{\textbackslash Vub} & {\ensuremath{V_{\uquark\bquark}}}\xspace \\
\texttt{\textbackslash Vcb} & {\ensuremath{V_{\cquark\bquark}}}\xspace & \texttt{\textbackslash Vtb} & {\ensuremath{V_{\tquark\bquark}}}\xspace & \texttt{\textbackslash Vuds} & {\ensuremath{V_{\uquark\dquark}^\ast}}\xspace \\
\texttt{\textbackslash Vcds} & {\ensuremath{V_{\cquark\dquark}^\ast}}\xspace & \texttt{\textbackslash Vtds} & {\ensuremath{V_{\tquark\dquark}^\ast}}\xspace & \texttt{\textbackslash Vuss} & {\ensuremath{V_{\uquark\squark}^\ast}}\xspace \\
\texttt{\textbackslash Vcss} & {\ensuremath{V_{\cquark\squark}^\ast}}\xspace & \texttt{\textbackslash Vtss} & {\ensuremath{V_{\tquark\squark}^\ast}}\xspace & \texttt{\textbackslash Vubs} & {\ensuremath{V_{\uquark\bquark}^\ast}}\xspace \\
\texttt{\textbackslash Vcbs} & {\ensuremath{V_{\cquark\bquark}^\ast}}\xspace & \texttt{\textbackslash Vtbs} & {\ensuremath{V_{\tquark\bquark}^\ast}}\xspace & \\
\end{tabular*}
\subsubsection{Oscillations}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash dm} & {\ensuremath{\Delta m}}\xspace & \texttt{\textbackslash dms} & {\ensuremath{\Delta m_{{\ensuremath{\Ps}}\xspace}}}\xspace & \texttt{\textbackslash dmd} & {\ensuremath{\Delta m_{{\ensuremath{\Pd}}\xspace}}}\xspace \\
\texttt{\textbackslash DG} & {\ensuremath{\Delta\Gamma}}\xspace & \texttt{\textbackslash DGs} & {\ensuremath{\Delta\Gamma_{{\ensuremath{\Ps}}\xspace}}}\xspace & \texttt{\textbackslash DGd} & {\ensuremath{\Delta\Gamma_{{\ensuremath{\Pd}}\xspace}}}\xspace \\
\texttt{\textbackslash Gs} & {\ensuremath{\Gamma_{{\ensuremath{\Ps}}\xspace}}}\xspace & \texttt{\textbackslash Gd} & {\ensuremath{\Gamma_{{\ensuremath{\Pd}}\xspace}}}\xspace & \texttt{\textbackslash MBq} & {\ensuremath{M_{{\ensuremath{\PB}}\xspace_{\ensuremath{\Pq}}\xspace}}}\xspace \\
\texttt{\textbackslash DGq} & {\ensuremath{\Delta\Gamma_{{\ensuremath{\Pq}}\xspace}}}\xspace & \texttt{\textbackslash Gq} & {\ensuremath{\Gamma_{{\ensuremath{\Pq}}\xspace}}}\xspace & \texttt{\textbackslash dmq} & {\ensuremath{\Delta m_{{\ensuremath{\Pq}}\xspace}}}\xspace \\
\texttt{\textbackslash GL} & {\ensuremath{\Gamma_{\mathrm{ L}}}}\xspace & \texttt{\textbackslash GH} & {\ensuremath{\Gamma_{\mathrm{ H}}}}\xspace & \texttt{\textbackslash DGsGs} & {\ensuremath{\Delta\Gamma_{{\ensuremath{\Ps}}\xspace}/\Gamma_{{\ensuremath{\Ps}}\xspace}}}\xspace \\
\texttt{\textbackslash Delm} & {\mbox{$\Delta m $}}\xspace & \texttt{\textbackslash ACP} & {\ensuremath{{\mathcal{A}}^{{\ensuremath{C\!P}}\xspace}}}\xspace & \texttt{\textbackslash Adir} & {\ensuremath{{\mathcal{A}}^{\mathrm{ dir}}}}\xspace \\
\texttt{\textbackslash Amix} & {\ensuremath{{\mathcal{A}}^{\mathrm{ mix}}}}\xspace & \texttt{\textbackslash ADelta} & {\ensuremath{{\mathcal{A}}^\Delta}}\xspace & \texttt{\textbackslash phid} & {\ensuremath{\phi_{{\ensuremath{\Pd}}\xspace}}}\xspace \\
\texttt{\textbackslash sinphid} & {\ensuremath{\sin\!\phid}}\xspace & \texttt{\textbackslash phis} & {\ensuremath{\phi_{{\ensuremath{\Ps}}\xspace}}}\xspace & \texttt{\textbackslash betas} & {\ensuremath{\beta_{{\ensuremath{\Ps}}\xspace}}}\xspace \\
\texttt{\textbackslash sbetas} & {\ensuremath{\sigma(\beta_{{\ensuremath{\Ps}}\xspace})}}\xspace & \texttt{\textbackslash stbetas} & {\ensuremath{\sigma(2\beta_{{\ensuremath{\Ps}}\xspace})}}\xspace & \texttt{\textbackslash stphis} & {\ensuremath{\sigma(\phi_{{\ensuremath{\Ps}}\xspace})}}\xspace \\
\texttt{\textbackslash sinphis} & {\ensuremath{\sin\!\phis}}\xspace & \\
\end{tabular*}
\subsubsection{Tagging}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash edet} & {\ensuremath{\varepsilon_{\mathrm{ det}}}}\xspace & \texttt{\textbackslash erec} & {\ensuremath{\varepsilon_{\mathrm{ rec/det}}}}\xspace & \texttt{\textbackslash esel} & {\ensuremath{\varepsilon_{\mathrm{ sel/rec}}}}\xspace \\
\texttt{\textbackslash etrg} & {\ensuremath{\varepsilon_{\mathrm{ trg/sel}}}}\xspace & \texttt{\textbackslash etot} & {\ensuremath{\varepsilon_{\mathrm{ tot}}}}\xspace & \texttt{\textbackslash mistag} & \ensuremath{\omega}\xspace \\
\texttt{\textbackslash wcomb} & \ensuremath{\omega^{\mathrm{comb}}}\xspace & \texttt{\textbackslash etag} & {\ensuremath{\varepsilon_{\mathrm{tag}}}}\xspace & \texttt{\textbackslash etagcomb} & {\ensuremath{\varepsilon_{\mathrm{tag}}^{\mathrm{comb}}}}\xspace \\
\texttt{\textbackslash effeff} & \ensuremath{\varepsilon_{\mathrm{eff}}}\xspace & \texttt{\textbackslash effeffcomb} & \ensuremath{\varepsilon_{\mathrm{eff}}^{\mathrm{comb}}}\xspace & \texttt{\textbackslash efftag} & {\ensuremath{\etag(1-2\omega)^2}}\xspace \\
\texttt{\textbackslash effD} & {\ensuremath{\etag D^2}}\xspace & \texttt{\textbackslash etagprompt} & {\ensuremath{\varepsilon_{\mathrm{ tag}}^{\mathrm{Pr}}}}\xspace & \texttt{\textbackslash etagLL} & {\ensuremath{\varepsilon_{\mathrm{ tag}}^{\mathrm{LL}}}}\xspace \\
\end{tabular*}
\subsubsection{Key decay channels}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash BdToKstmm} & \decay{\Bd}{\Kstarz\mup\mun} & \texttt{\textbackslash BdbToKstmm} & \decay{\Bdb}{\Kstarzb\mup\mun} & \texttt{\textbackslash BsToJPsiPhi} & \decay{\Bs}{\jpsi\phi} \\
\texttt{\textbackslash BdToJPsiKst} & \decay{\Bd}{\jpsi\Kstarz} & \texttt{\textbackslash BdbToJPsiKst} & \decay{\Bdb}{\jpsi\Kstarzb} & \texttt{\textbackslash BsPhiGam} & \decay{\Bs}{\phi \g} \\
\texttt{\textbackslash BdKstGam} & \decay{\Bd}{\Kstarz \g} & \texttt{\textbackslash BTohh} & \decay{\B}{\Ph^+ \Ph'^-} & \texttt{\textbackslash BdTopipi} & \decay{\Bd}{\pip\pim} \\
\texttt{\textbackslash BdToKpi} & \decay{\Bd}{\Kp\pim} & \texttt{\textbackslash BsToKK} & \decay{\Bs}{\Kp\Km} & \texttt{\textbackslash BsTopiK} & \decay{\Bs}{\pip\Km} \\
\end{tabular*}
\subsubsection{Rare decays}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash BdKstee} & \decay{\Bd}{\Kstarz\epem} & \texttt{\textbackslash BdbKstee} & \decay{\Bdb}{\Kstarzb\epem} & \texttt{\textbackslash bsll} & \decay{\bquark}{\squark \ell^+ \ell^-} \\
\texttt{\textbackslash AFB} & \ensuremath{A_{\mathrm{FB}}}\xspace & \texttt{\textbackslash FL} & \ensuremath{F_{\mathrm{L}}}\xspace & \texttt{\textbackslash AT\#1 \textbackslash AT2} & \AT2 \\
\texttt{\textbackslash btosgam} & \decay{\bquark}{\squark \g} & \texttt{\textbackslash btodgam} & \decay{\bquark}{\dquark \g} & \texttt{\textbackslash Bsmm} & \decay{\Bs}{\mup\mun} \\
\texttt{\textbackslash Bdmm} & \decay{\Bd}{\mup\mun} & \texttt{\textbackslash ctl} & \ensuremath{\cos{\theta_\ell}}\xspace & \texttt{\textbackslash ctk} & \ensuremath{\cos{\theta_K}}\xspace \\
\end{tabular*}
\subsubsection{Wilson coefficients and operators}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash C\#1 \textbackslash C9} & \C9 & \texttt{\textbackslash Cp\#1 \textbackslash Cp7} & \Cp7 & \texttt{\textbackslash Ceff\#1 \textbackslash Ceff9 } & \Ceff9 \\
\texttt{\textbackslash Cpeff\#1 \textbackslash Cpeff7} & \Cpeff7 & \texttt{\textbackslash Ope\#1 \textbackslash Ope2} & \Ope2 & \texttt{\textbackslash Opep\#1 \textbackslash Opep7} & \Opep7 \\
\end{tabular*}
\subsubsection{Charm}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash xprime} & \ensuremath{x^{\prime}}\xspace & \texttt{\textbackslash yprime} & \ensuremath{y^{\prime}}\xspace & \texttt{\textbackslash ycp} & \ensuremath{y_{\CP}}\xspace \\
\texttt{\textbackslash agamma} & \ensuremath{A_{\Gamma}}\xspace & \texttt{\textbackslash dkpicf} & \decay{\Dz}{\Km\pip} & \\
\end{tabular*}
\subsubsection{QM}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash bra[1] \textbackslash bra\{a\}} & \bra{a} & \texttt{\textbackslash ket[1] \textbackslash ket\{b\}} & \ket{b} & \texttt{\textbackslash braket[2] \textbackslash braket\{a\}\{b\}} & \braket{a}{b} \\
\end{tabular*}
\subsection{Units}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash unit[1] \textbackslash unit\{kg\}} & \unit{kg} & \\
\end{tabular*}
\subsubsection{Energy and momentum}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash tev} & \ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace & \texttt{\textbackslash gev} & \ensuremath{\mathrm{\,Ge\kern -0.1em V}}\xspace & \texttt{\textbackslash mev} & \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace \\
\texttt{\textbackslash kev} & \ensuremath{\mathrm{\,ke\kern -0.1em V}}\xspace & \texttt{\textbackslash ev} & \ensuremath{\mathrm{\,e\kern -0.1em V}}\xspace & \texttt{\textbackslash gevc} & \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c}}\xspace \\
\texttt{\textbackslash mevc} & \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c}}\xspace & \texttt{\textbackslash gevcc} & \ensuremath{{\mathrm{\,Ge\kern -0.1em V\!/}c^2}}\xspace & \texttt{\textbackslash gevgevcccc} & \ensuremath{{\mathrm{\,Ge\kern -0.1em V^2\!/}c^4}}\xspace \\
\texttt{\textbackslash mevcc} & \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace & \\
\end{tabular*}
\subsubsection{Distance and area}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash km} & \ensuremath{\mathrm{ \,km}}\xspace & \texttt{\textbackslash m} & \ensuremath{\mathrm{ \,m}}\xspace & \texttt{\textbackslash ma} & \ensuremath{{\mathrm{ \,m}}^2}\xspace \\
\texttt{\textbackslash cm} & \ensuremath{\mathrm{ \,cm}}\xspace & \texttt{\textbackslash cma} & \ensuremath{{\mathrm{ \,cm}}^2}\xspace & \texttt{\textbackslash mm} & \ensuremath{\mathrm{ \,mm}}\xspace \\
\texttt{\textbackslash mma} & \ensuremath{{\mathrm{ \,mm}}^2}\xspace & \texttt{\textbackslash mum} & \ensuremath{{\,\upmu\mathrm{m}}}\xspace & \texttt{\textbackslash muma} & \ensuremath{{\,\upmu\mathrm{m}^2}}\xspace \\
\texttt{\textbackslash nm} & \ensuremath{\mathrm{ \,nm}}\xspace & \texttt{\textbackslash fm} & \ensuremath{\mathrm{ \,fm}}\xspace & \texttt{\textbackslash barn} & \ensuremath{\mathrm{ \,b}}\xspace \\
\texttt{\textbackslash mbarn} & \ensuremath{\mathrm{ \,mb}}\xspace & \texttt{\textbackslash mub} & \ensuremath{{\mathrm{ \,\upmu b}}}\xspace & \texttt{\textbackslash nb} & \ensuremath{\mathrm{ \,nb}}\xspace \\
\texttt{\textbackslash invnb} & \ensuremath{\mbox{\,nb}^{-1}}\xspace & \texttt{\textbackslash pb} & \ensuremath{\mathrm{ \,pb}}\xspace & \texttt{\textbackslash invpb} & \ensuremath{\mbox{\,pb}^{-1}}\xspace \\
\texttt{\textbackslash fb} & \ensuremath{\mbox{\,fb}}\xspace & \texttt{\textbackslash invfb} & \ensuremath{\mbox{\,fb}^{-1}}\xspace & \texttt{\textbackslash ab} & \ensuremath{\mbox{\,ab}}\xspace \\
\texttt{\textbackslash invab} & \ensuremath{\mbox{\,ab}^{-1}}\xspace & \\
\end{tabular*}
\subsubsection{Time }
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash sec} & \ensuremath{\mathrm{{\,s}}}\xspace & \texttt{\textbackslash ms} & \ensuremath{{\mathrm{ \,ms}}}\xspace & \texttt{\textbackslash mus} & \ensuremath{{\,\upmu{\mathrm{ s}}}}\xspace \\
\texttt{\textbackslash ns} & \ensuremath{{\mathrm{ \,ns}}}\xspace & \texttt{\textbackslash ps} & \ensuremath{{\mathrm{ \,ps}}}\xspace & \texttt{\textbackslash fs} & \ensuremath{\mathrm{ \,fs}}\xspace \\
\texttt{\textbackslash mhz} & \ensuremath{{\mathrm{ \,MHz}}}\xspace & \texttt{\textbackslash khz} & \ensuremath{{\mathrm{ \,kHz}}}\xspace & \texttt{\textbackslash hz} & \ensuremath{{\mathrm{ \,Hz}}}\xspace \\
\texttt{\textbackslash invps} & \ensuremath{{\mathrm{ \,ps^{-1}}}}\xspace & \texttt{\textbackslash invns} & \ensuremath{{\mathrm{ \,ns^{-1}}}}\xspace & \texttt{\textbackslash yr} & \ensuremath{\mathrm{ \,yr}}\xspace \\
\texttt{\textbackslash hr} & \ensuremath{\mathrm{ \,hr}}\xspace & \\
\end{tabular*}
\subsubsection{Temperature}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash degc} & \ensuremath{^\circ}{C}\xspace & \texttt{\textbackslash degk} & \ensuremath {\mathrm{ K}}\xspace & \\
\end{tabular*}
\subsubsection{Material lengths, radiation}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash Xrad} & \ensuremath{X_0}\xspace & \texttt{\textbackslash NIL} & \ensuremath{\lambda_{int}}\xspace & \texttt{\textbackslash mip} & MIP\xspace \\
\texttt{\textbackslash neutroneq} & \ensuremath{\mathrm{ \,n_{eq}}}\xspace & \texttt{\textbackslash neqcmcm} & \ensuremath{\mathrm{ \,n_{eq} / cm^2}}\xspace & \texttt{\textbackslash kRad} & \ensuremath{\mathrm{ \,kRad}}\xspace \\
\texttt{\textbackslash MRad} & \ensuremath{\mathrm{ \,MRad}}\xspace & \texttt{\textbackslash ci} & \ensuremath{\mathrm{ \,Ci}}\xspace & \texttt{\textbackslash mci} & \ensuremath{\mathrm{ \,mCi}}\xspace \\
\end{tabular*}
\subsubsection{Uncertainties}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash sx} & \sx & \texttt{\textbackslash sy} & \sy & \texttt{\textbackslash sz} & \sz \\
\texttt{\textbackslash stat} & \ensuremath{\mathrm{\,(stat)}}\xspace & \texttt{\textbackslash syst} & \ensuremath{\mathrm{\,(syst)}}\xspace & \\
\end{tabular*}
\subsubsection{Maths}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash order} & {\ensuremath{\mathcal{O}}}\xspace & \texttt{\textbackslash chisq} & \ensuremath{\chi^2}\xspace & \texttt{\textbackslash chisqndf} & \ensuremath{\chi^2/\mathrm{ndf}}\xspace \\
\texttt{\textbackslash chisqip} & \ensuremath{\chi^2_{\text{IP}}}\xspace & \texttt{\textbackslash chisqvs} & \ensuremath{\chi^2_{\text{VS}}}\xspace & \texttt{\textbackslash chisqvtx} & \ensuremath{\chi^2_{\text{vtx}}}\xspace \\
\texttt{\textbackslash chisqvtxndf} & \ensuremath{\chi^2_{\text{vtx}}/\mathrm{ndf}}\xspace & \texttt{\textbackslash deriv} & \ensuremath{\mathrm{d}} & \texttt{\textbackslash gsim} & \gsim \\
\texttt{\textbackslash lsim} & \lsim & \texttt{\textbackslash mean[1] \textbackslash mean\{x\}} & \mean{x} & \texttt{\textbackslash abs[1] \textbackslash abs\{x\}} & \abs{x} \\
\texttt{\textbackslash Real} & \ensuremath{\mathcal{R}e}\xspace & \texttt{\textbackslash Imag} & \ensuremath{\mathcal{I}m}\xspace & \texttt{\textbackslash PDF} & \mathcal{P} \\
\texttt{\textbackslash sPlot} & \mbox{\em sPlot}\xspace & \\
\end{tabular*}
\subsection{Kinematics}
\subsubsection{Energy, Momenta}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash Ebeam} & \ensuremath{E_{\mbox{\tiny BEAM}}}\xspace & \texttt{\textbackslash sqs} & \ensuremath{\protect\sqrt{s}}\xspace & \texttt{\textbackslash ptot} & \mbox{$p$}\xspace \\
\texttt{\textbackslash pt} & \mbox{$p_{\mathrm{ T}}$}\xspace & \texttt{\textbackslash et} & \mbox{$E_{\mathrm{ T}}$}\xspace & \texttt{\textbackslash mt} & \mbox{$M_{\mathrm{ T}}$}\xspace \\
\texttt{\textbackslash dpp} & \ensuremath{\Delta p/p}\xspace & \texttt{\textbackslash msq} & \ensuremath{m^2}\xspace & \texttt{\textbackslash dedx} & \ensuremath{\mathrm{d}\hspace{-0.1em}E/\mathrm{d}x}\xspace \\
\end{tabular*}
\subsubsection{PID}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash dllkpi} & \ensuremath{\mathrm{DLL}_{\kaon\pion}}\xspace & \texttt{\textbackslash dllppi} & \ensuremath{\mathrm{DLL}_{\proton\pion}}\xspace & \texttt{\textbackslash dllepi} & \ensuremath{\mathrm{DLL}_{\electron\pion}}\xspace \\
\texttt{\textbackslash dllmupi} & \ensuremath{\mathrm{DLL}_{\muon\pi}}\xspace & \\
\end{tabular*}
\subsubsection{Geometry}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash degrees} & \ensuremath{^{\circ}}\xspace & \texttt{\textbackslash krad} & \ensuremath{\mathrm{ \,krad}}\xspace & \texttt{\textbackslash mrad} & \ensuremath{\mathrm{ \,mrad}}\xspace \\
\texttt{\textbackslash rad} & \ensuremath{\mathrm{ \,rad}}\xspace & \\
\end{tabular*}
\subsubsection{Accelerator}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash betastar} & \ensuremath{\beta^*} & \texttt{\textbackslash lum} & \lum & \texttt{\textbackslash intlum[1] \textbackslash intlum\{2 \,\ensuremath{\mbox{\,fb}^{-1}}\xspace\}} & \intlum{2 \,\ensuremath{\mbox{\,fb}^{-1}}\xspace} \\
\end{tabular*}
\subsection{Software}
\subsubsection{Programs}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash bcvegpy} & \mbox{\textsc{Bcvegpy}}\xspace & \texttt{\textbackslash boole} & \mbox{\textsc{Boole}}\xspace & \texttt{\textbackslash brunel} & \mbox{\textsc{Brunel}}\xspace \\
\texttt{\textbackslash davinci} & \mbox{\textsc{DaVinci}}\xspace & \texttt{\textbackslash dirac} & \mbox{\textsc{Dirac}}\xspace & \texttt{\textbackslash evtgen} & \mbox{\textsc{EvtGen}}\xspace \\
\texttt{\textbackslash fewz} & \mbox{\textsc{Fewz}}\xspace & \texttt{\textbackslash fluka} & \mbox{\textsc{Fluka}}\xspace & \texttt{\textbackslash ganga} & \mbox{\textsc{Ganga}}\xspace \\
\texttt{\textbackslash gaudi} & \mbox{\textsc{Gaudi}}\xspace & \texttt{\textbackslash gauss} & \mbox{\textsc{Gauss}}\xspace & \texttt{\textbackslash geant} & \mbox{\textsc{Geant4}}\xspace \\
\texttt{\textbackslash hepmc} & \mbox{\textsc{HepMC}}\xspace & \texttt{\textbackslash herwig} & \mbox{\textsc{Herwig}}\xspace & \texttt{\textbackslash moore} & \mbox{\textsc{Moore}}\xspace \\
\texttt{\textbackslash neurobayes} & \mbox{\textsc{NeuroBayes}}\xspace & \texttt{\textbackslash photos} & \mbox{\textsc{Photos}}\xspace & \texttt{\textbackslash powheg} & \mbox{\textsc{Powheg}}\xspace \\
\texttt{\textbackslash pythia} & \mbox{\textsc{Pythia}}\xspace & \texttt{\textbackslash resbos} & \mbox{\textsc{ResBos}}\xspace & \texttt{\textbackslash roofit} & \mbox{\textsc{RooFit}}\xspace \\
\texttt{\textbackslash root} & \mbox{\textsc{Root}}\xspace & \texttt{\textbackslash spice} & \mbox{\textsc{Spice}}\xspace & \texttt{\textbackslash urania} & \mbox{\textsc{Urania}}\xspace \\
\end{tabular*}
\subsubsection{Languages}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash cpp} & \mbox{\textsc{C\raisebox{0.1em}{{\footnotesize{++}}}}}\xspace & \texttt{\textbackslash ruby} & \mbox{\textsc{Ruby}}\xspace & \texttt{\textbackslash fortran} & \mbox{\textsc{Fortran}}\xspace \\
\texttt{\textbackslash svn} & \mbox{\textsc{SVN}}\xspace & \\
\end{tabular*}
\subsubsection{Data processing}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash kbytes} & \ensuremath{{\mathrm{ \,kbytes}}}\xspace & \texttt{\textbackslash kbsps} & \ensuremath{{\mathrm{ \,kbits/s}}}\xspace & \texttt{\textbackslash kbits} & \ensuremath{{\mathrm{ \,kbits}}}\xspace \\
\texttt{\textbackslash kbsps} & \ensuremath{{\mathrm{ \,kbits/s}}}\xspace & \texttt{\textbackslash mbsps} & \ensuremath{{\mathrm{ \,Mbytes/s}}}\xspace & \texttt{\textbackslash mbytes} & \ensuremath{{\mathrm{ \,Mbytes}}}\xspace \\
\texttt{\textbackslash mbps} & \ensuremath{{\mathrm{ \,Mbyte/s}}}\xspace & \texttt{\textbackslash mbsps} & \ensuremath{{\mathrm{ \,Mbytes/s}}}\xspace & \texttt{\textbackslash gbsps} & \ensuremath{{\mathrm{ \,Gbytes/s}}}\xspace \\
\texttt{\textbackslash gbytes} & \ensuremath{{\mathrm{ \,Gbytes}}}\xspace & \texttt{\textbackslash gbsps} & \ensuremath{{\mathrm{ \,Gbytes/s}}}\xspace & \texttt{\textbackslash tbytes} & \ensuremath{{\mathrm{ \,Tbytes}}}\xspace \\
\texttt{\textbackslash tbpy} & \ensuremath{{\mathrm{ \,Tbytes/yr}}}\xspace & \texttt{\textbackslash dst} & DST\xspace & \\
\end{tabular*}
\subsection{Detector related}
\subsubsection{Detector technologies}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash nonn} & \ensuremath{\mathrm{{ \mathit{n^+}} \mbox{-} on\mbox{-}{ \mathit{n}}}}\xspace & \texttt{\textbackslash ponn} & \ensuremath{\mathrm{{ \mathit{p^+}} \mbox{-} on\mbox{-}{ \mathit{n}}}}\xspace & \texttt{\textbackslash nonp} & \ensuremath{\mathrm{{ \mathit{n^+}} \mbox{-} on\mbox{-}{ \mathit{p}}}}\xspace \\
\texttt{\textbackslash cvd} & CVD\xspace & \texttt{\textbackslash mwpc} & MWPC\xspace & \texttt{\textbackslash gem} & GEM\xspace \\
\end{tabular*}
\subsubsection{Detector components, electronics}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash tell1} & TELL1\xspace & \texttt{\textbackslash ukl1} & UKL1\xspace & \texttt{\textbackslash beetle} & Beetle\xspace \\
\texttt{\textbackslash otis} & OTIS\xspace & \texttt{\textbackslash croc} & CROC\xspace & \texttt{\textbackslash carioca} & CARIOCA\xspace \\
\texttt{\textbackslash dialog} & DIALOG\xspace & \texttt{\textbackslash sync} & SYNC\xspace & \texttt{\textbackslash cardiac} & CARDIAC\xspace \\
\texttt{\textbackslash gol} & GOL\xspace & \texttt{\textbackslash vcsel} & VCSEL\xspace & \texttt{\textbackslash ttc} & TTC\xspace \\
\texttt{\textbackslash ttcrx} & TTCrx\xspace & \texttt{\textbackslash hpd} & HPD\xspace & \texttt{\textbackslash pmt} & PMT\xspace \\
\texttt{\textbackslash specs} & SPECS\xspace & \texttt{\textbackslash elmb} & ELMB\xspace & \texttt{\textbackslash fpga} & FPGA\xspace \\
\texttt{\textbackslash plc} & PLC\xspace & \texttt{\textbackslash rasnik} & RASNIK\xspace & \texttt{\textbackslash elmb} & ELMB\xspace \\
\texttt{\textbackslash can} & CAN\xspace & \texttt{\textbackslash lvds} & LVDS\xspace & \texttt{\textbackslash ntc} & NTC\xspace \\
\texttt{\textbackslash adc} & ADC\xspace & \texttt{\textbackslash led} & LED\xspace & \texttt{\textbackslash ccd} & CCD\xspace \\
\texttt{\textbackslash hv} & HV\xspace & \texttt{\textbackslash lv} & LV\xspace & \texttt{\textbackslash pvss} & PVSS\xspace \\
\texttt{\textbackslash cmos} & CMOS\xspace & \texttt{\textbackslash fifo} & FIFO\xspace & \texttt{\textbackslash ccpc} & CCPC\xspace \\
\end{tabular*}
\subsubsection{Chemical symbols}
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash cfourften} & \ensuremath{\mathrm{ C_4 F_{10}}}\xspace & \texttt{\textbackslash cffour} & \ensuremath{\mathrm{ CF_4}}\xspace & \texttt{\textbackslash cotwo} & \cotwo \\
\texttt{\textbackslash csixffouteen} & \csixffouteen & \texttt{\textbackslash mgftwo} & \mgftwo & \texttt{\textbackslash siotwo} & \siotwo \\
\end{tabular*}
\subsection{Special Text }
\begin{tabular*}{\linewidth}{@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l@{\extracolsep{\fill}}l@{\extracolsep{0.5cm}}l}
\texttt{\textbackslash eg} & \mbox{\itshape e.g.}\xspace & \texttt{\textbackslash ie} & \mbox{\itshape i.e.}\xspace & \texttt{\textbackslash etal} & \mbox{\itshape et al.}\xspace \\
\texttt{\textbackslash etc} & \mbox{\itshape etc.}\xspace & \texttt{\textbackslash cf} & \mbox{\itshape cf.}\xspace & \texttt{\textbackslash ffp} & \mbox{\itshape ff.}\xspace \\
\texttt{\textbackslash vs} & \mbox{\itshape vs.}\xspace & \\
\end{tabular*}
\section{Results and summary}
\label{sec:Results}
Using the yield and efficiency ratios,
the ratio of branching fractions of \ensuremath{\Bcp\to\jpsi\Kp}\xspace and \ensuremath{\Bcp\to\jpsi\pip}\xspace is evaluated as
\begin{equation*}
\ensuremath{{R}_{K/\pi}}\xspace= 0.089\pm0.013\pm0.003
\end{equation*}
for the 7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace data sample
and
\begin{equation*}
\ensuremath{{R}_{K/\pi}}\xspace = 0.075\pm0.008\pm0.003
\end{equation*}
for the 8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace sample,
where the first uncertainties are statistical and the second are systematic.
The two results are combined by evaluating their weighted average.
The systematic uncertainties of both measurements are dominated by
the contribution from the non-inclusion of the partially reconstructed
background for ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\kaon^+}}\xspace$ decays, and so are assumed to be fully
correlated, while their statistical uncertainties are independent.
The combined measurement for the $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ data sample is
\begin{equation*}
\ensuremath{{R}_{K/\pi}}\xspace = 0.079\pm0.007\pm0.003\,.
\end{equation*}
This is consistent with the previous LHCb measurement $\ensuremath{{R}_{K/\pi}}\xspace = 0.069\pm0.019\pm0.005$~\cite{LHCb-PAPER-2013-021}, which was based on the 7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace data alone.
The uncertainties are significantly reduced due to both the increased sample size and the improved event selection.
The result supersedes the previous measurement \cite{LHCb-PAPER-2013-021}
and agrees with the theoretical predictions in Refs. \cite{
Kiselev:2000pp, Chang:1992pt, Liu:1997hr,
AbdElHady:1999xh, Ebert:2003cn,
Ivanov:2006ni, Naimuddin:2012dy, Qiao:2012hp, Ke:2013yka}, but disfavours that based on QCD sum rules \cite{Kiselev:2002vz}.
\section{Event selection}
Candidate ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace\,h^+$ decays,
with ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\Pmu^+\Pmu^-}}\xspace$ and $h^+$ being a $K^+$ or $\pi^+$,
are reconstructed as follows.
First a loose preselection is applied. Pairs of oppositely charged, well-reconstructed
muon tracks with $\mbox{$p_{\mathrm{ T}}$}\xspace>550\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c}}\xspace$ consistent with originating from a
common vertex are combined to form ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{\Pmu^+\Pmu^-}}\xspace$ candidates.
Hadron ($h^+$) candidates are selected from well-reconstructed tracks with $\mbox{$p_{\mathrm{ T}}$}\xspace>500\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c}}\xspace$, inconsistent with originating from any PV and with the muon hypothesis.
Candidate ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\kaon^+}}\xspace$ and ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\pion^+}}\xspace$ decays are formed from ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace h^+$ combinations
that originate from a common vertex. They must also be within $500\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$ of the known ${\ensuremath{\B_\cquark^+}}\xspace$ mass \cite{PDG2014}.
The impact parameter $\ensuremath{\chi^2}\xspace$, $\ensuremath{\chi^2}\xspace_{\rm IP}$, which is defined as the difference
in the vertex fit $\ensuremath{\chi^2}\xspace$ of the PV with and without the particle
under consideration, is required to be less than 16 for the ${\ensuremath{\B_\cquark^+}}\xspace$ candidates.
A multivariate classifier using a boosted decision tree (BDT) \cite{Breiman} is constructed to further suppress the combinatorial background.
The kinematic variables used as inputs to the BDT are chosen to discriminate between signal and background.
The twelve variables chosen are: the $\ensuremath{\chi^2}\xspace_{\rm IP}$ of the {\ensuremath{\B_\cquark^+}}\xspace, {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace, {\ensuremath{\Pmu^+}}\xspace, $\mu^-$ and $h^+$ candidates;
the \mbox{$p_{\mathrm{ T}}$}\xspace of the {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace, {\ensuremath{\Pmu^+}}\xspace, $\mu^-$ and $h^+$ candidates;
the $\ensuremath{\chi^2}\xspace$ per degree of freedom of the ${\ensuremath{\B_\cquark^+}}\xspace$ vertex fit; and the decay time and the decay length of the ${\ensuremath{\B_\cquark^+}}\xspace$ candidate.
Since the kaon-pion mass difference is small compared with the energy release of
${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace h^+$ decays, the distributions of the BDT variables
are similar for ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace K^+$ and ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace \pi^+$ decays.
The BDT is trained with simulated ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace \pi^+$ decays to represent
both the ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace K^+$ and the ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace \pi^+$ signals,
and with events from the upper mass sideband of the ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\pion^+}}\xspace$ candidates in data,
$[6444,~6528]\ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace$, to represent the combinatorial background.
For one third of the events in the training samples the centre-of-mass energy is $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$,
and for the rest it is $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ in accordance with the ratio of integrated luminosities.
Since the BDT does not use any particle identification information,
it selects both ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\kaon^+}}\xspace$ and ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\pion^+}}\xspace$ candidates.
Particle identification requirements using information from the RICH\xspace subdetectors
are then applied to the hadrons
to obtain two mutually exclusive samples of ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\kaon^+}}\xspace$ and ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\pion^+}}\xspace$ candidates.
The BDT and particle identification requirements
are optimised sequentially on the sample of ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\kaon^+}}\xspace$ candidates that
pass the loose preselection to maximise $N_K/\sqrt{N_{\rm tot}}$,
where $N_{\rm tot}$ is the total number of candidates
within $\pm 3$ times the mass resolution around the known ${\ensuremath{\B_\cquark^+}}\xspace$ mass.
Here $N_K$ refers to the ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\kaon^+}}\xspace$ signal yield and
is estimated to be $(N_{\rm tot}-N_{\rm comb})/(1+1/(r_{\rm eff}\ensuremath{{R}_{K/\pi}}\xspace))$,
where the value of $\ensuremath{{R}_{K/\pi}}\xspace$ is taken from the previous LHCb measurement~\cite{LHCb-PAPER-2013-021},
$N_{\rm comb}$ is the number of combinatorial background events in the signal region
extrapolated from the upper sideband,
and $r_{\rm eff}$ represents the ratio of the numbers of ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\kaon^+}}\xspace$ and ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace \pi^+$
events that pass the ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\kaon^+}}\xspace$ selection and fall in the signal window.
After this optimisation, the BDT rejects more than $99.8\%$ of
the combinatorial background and keeps around $70\%$ of ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace h^+$ events.
This particle identification requirement has an efficiency of about
$70\%$ for ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\kaon^+}}\xspace$ and $87\%$ for ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\pion^+}}\xspace$, while the
probabilities for a charged kaon to be misidentified as a pion and a charged
pion to be misidentified as a kaon are below $7\%$ and $1\%$, respectively.
\section{Systematic uncertainties}
Since the running conditions changed between $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$,
the systematic uncertainties on $\ensuremath{{R}_{K/\pi}}\xspace$ are determined separately for the two samples.
Table~\ref{tab:sys} summarises the relative systematic uncertainties associated with the mass fit and efficiency estimates
that affect the ratio of branching fractions. The sources of these uncertainties are discussed below.
Each of the systematic uncertainties associated with
the mass fit is studied by generating an ensemble of pseudoexperiments according to the nominal model
described above
and fitting them with an alternative model. The difference in the mean values of $\ensuremath{{R}_{K/\pi}}\xspace$ obtained
is taken as the systematic uncertainty.
Changing the signal model from
the sum of two DSCB functions to a single DSCB function leads to
relative systematic uncertainties of $0.5\%$ and $0.8\%$ for the 7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace and 8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace data, respectively.
Using a third-order polynomial in place of an exponential function for the combinatorial background
changes the mean values of $\ensuremath{{R}_{K/\pi}}\xspace$ by $1.1\%$ and $0.5\%$ for the two samples.
In the nominal fit, the partially reconstructed background
is neglected for \ensuremath{\Bcp\to\jpsi\Kp}\xspace decays for reasons of fit stability.
The associated systematic uncertainties are estimated
by including such a component in the same way as was done for \ensuremath{\Bcp\to\jpsi\pip}\xspace decays,
and are found to be $3.3\%$ and $3.2\%$ for the 7 and $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ data, respectively.
Using the sum of two DSCB functions instead of a single DSCB function for the misidentification
background events changes the mean values of $\ensuremath{{R}_{K/\pi}}\xspace$ by $0.2\%$ and $0.0\%$ for the two samples.
The selection and trigger efficiencies are calculated with simulated samples.
Systematic effects on the efficiency evaluation due to differences between data and simulation
in the distributions of variables such as muon momentum and {\ensuremath{\B_\cquark^+}}\xspace decay time are investigated.
Such effects are found to cancel in the efficiency ratio and
thus have negligible impact on $\ensuremath{{R}_{K/\pi}}\xspace$.
The kaon and pion identification efficiencies are measured as functions of momentum and pseudorapidity
with a control sample of $D^{*+} \ensuremath{\rightarrow}\xspace {\ensuremath{\D^0}}\xspace {\ensuremath{\pion^+}}\xspace$, ${\ensuremath{\D^0}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^+}}\xspace$ decays,
and represented by two-dimensional histograms.
When the histogram binning is varied, the largest changes in the efficiency ratio seen
are $0.2\%$ and $0.1\%$ for the $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and 8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace samples,
and these values are assigned as the corresponding relative systematic uncertainties.
The simulation accounts for the different interaction cross-sections
of pions and kaons with matter. However, if the amount of material in
the detector is not modelled correctly,
this would alter the efficiency ratio. A systematic uncertainty of
$0.3\%$ associated with this effect is assigned for both $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and $8 \ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ samples.
Adding all of the above contributions in quadrature, the total relative systematic uncertainties
on $\ensuremath{{R}_{K/\pi}}\xspace$ are $3.5\%$ and $3.4\%$ for the $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ results.
\begin{table}[h]
\caption{Summary of the relative systematic uncertainties on $\ensuremath{{R}_{K/\pi}}\xspace$.}
\vspace*{-0.5cm}
\begin{center}\begin{tabular}{lcc}
\hline
& $7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ & $8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ \\
\hline
Signal model & $0.5\%$ & $0.8\%$\\
Combinatorial background & $1.1\%$ & $0.5\%$\\
Partially reconstructed background & $3.3\%$ & $3.2\%$\\
Misidentification background & $0.2\%$ & $0.0\%$\\
Particle identification efficiency & $0.2\%$ & $0.1\%$\\
Detector material & $0.3\%$ & $0.3\%$\\
\hline
Total & $3.5\%$ & $3.4\%$\\
\hline
\end{tabular}\end{center}
\label{tab:sys}
\end{table}
\section{Signal yields and efficiency correction}
\label{sec:yields}
The measurement is made by evaluating
\begin{equation} \label{eq:Rk}
\ensuremath{{R}_{K/\pi}}\xspace \equiv\frac{{\ensuremath{\mathcal{B}}}\xspace(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)}{{\ensuremath{\mathcal{B}}}\xspace(\ensuremath{\Bcp\to\jpsi\pip}\xspace)}
= \frac{N(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)}{N(\ensuremath{\Bcp\to\jpsi\pip}\xspace)} \times
\frac{\epsilon(\ensuremath{\Bcp\to\jpsi\pip}\xspace)}{\epsilon(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)},
\end{equation}
where $N(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)$ and $N(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$ are the signal yields,
and $\epsilon(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)$ and $\epsilon(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$ are the total efficiencies
estimated with simulation and control samples of data.
The signal yields $N(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)$ and $N(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$
are obtained from a simultaneous unbinned maximum likelihood fit to the distribution of
{\ensuremath{\B_\cquark^+}}\xspace candidate masses in the range 6000 to 6600 \ensuremath{{\mathrm{\,Me\kern -0.1em V\!/}c^2}}\xspace.
These candidates include the part of the background training sample that passes the full selection;
the effect of doing so has been investigated and found not to lead to any systematic bias.
The fit model includes
components due to signal, combinatorial background
and misidentified decays (\ensuremath{\Bcp\to\jpsi\pip}\xspace misidentified as \ensuremath{\Bcp\to\jpsi\Kp}\xspace, or vice versa).
A partially reconstructed background component is included for \ensuremath{\Bcp\to\jpsi\pip}\xspace.
This background is mainly due to ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace \rho^+$ decays
followed by $\rho^+ \ensuremath{\rightarrow}\xspace {\ensuremath{\pion^+}}\xspace {\ensuremath{\pion^0}}\xspace$.
The data show no clear indication of partially reconstructed background
for \ensuremath{\Bcp\to\jpsi\Kp}\xspace. A systematic uncertainty is assigned due to the non-inclusion of this background component.
The signal mass distribution of ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace\,h^+$ is described by the
sum of two double-sided Crystal Ball ($F^{\rm DSCB}$) functions consisting
of a Gaussian core and power law tails on both sides,
\begin{eqnarray}
f_{\rm sig}(M_{{\ensuremath{\B_\cquark^+}}\xspace})=\alpha F^{\rm DSCB}_1(M_{{\ensuremath{\B_\cquark^+}}\xspace}) + (1-\alpha)F^{\rm DSCB}_2(M_{{\ensuremath{\B_\cquark^+}}\xspace}),
\end{eqnarray}
where $M_{{\ensuremath{\B_\cquark^+}}\xspace}$ is the invariant mass of the ${\ensuremath{\Pmu^+\Pmu^-}}\xspace\,h^+$ combination with the
mass of the ${\ensuremath{\Pmu^+\Pmu^-}}\xspace$ pair constrained to the known ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace$ mass.
In the simultaneous fit, the Gaussian mean and the core mass resolution
$\sigma_1$ of $F^{\rm DSCB}_1$ are allowed to vary, and set to be the same
for both the ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\pion^+}}\xspace$ and the ${\ensuremath{\B_\cquark^+}}\xspace\ensuremath{\rightarrow}\xspace{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace{\ensuremath{\kaon^+}}\xspace$ decays.
The tail parameters, the fraction $\alpha$ and the ratio $\sigma_2/\sigma_1$
of the core-mass resolutions of $F^{\rm DSCB}_1$ and $F^{\rm DSCB}_2$
are fixed to the values obtained in simulation.
The combinatorial background for each decay mode is modelled by an exponential distribution.
Background arising from misidentified ${\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace J/\psi h^+$ decays
is described by a DSCB function, with
shape and mass offset relative to the signal peak derived from simulation for each mode separately.
The invariant mass distribution of the partially reconstructed background is
taken to be an ARGUS function\cite{Albrecht:1990am} convolved with a Gaussian resolution function.
The mean and the width parameters of the resolution
function are set to be zero and $\sqrt{\alpha\sigma_1^2+(1-\alpha)\sigma_2^2}$.
The parameters estimated from the simultaneous fit are:
the yield $N(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$, the yield ratio $N(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)/N(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$,
the numbers of combinatorial background events for \ensuremath{\Bcp\to\jpsi\Kp}\xspace and \ensuremath{\Bcp\to\jpsi\pip}\xspace decays,
the number of misidentification background events for each of the decay modes,
the number of partially reconstructed background events for the \ensuremath{\Bcp\to\jpsi\pip}\xspace decay,
and the shape parameters describing the signal and background distributions.
The results of the separate fits to the 7 and 8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace samples are shown in Fig.~\ref{fig:fit}.
In the 7\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace sample, the yield $N({\ensuremath{\B_\cquark^+}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi\mskip 2mu}}}\xspace {\ensuremath{\pion^+}}\xspace)$ is found to be $954 \pm 36$ and the yield ratio
$N(\ensuremath{\Bcp\to\jpsi\Kp}\xspace)/N(\ensuremath{\Bcp\to\jpsi\pip}\xspace)$ is found
to be $0.069 \pm 0.010$. The corresponding values in the 8\ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace sample are $2253 \pm 53$ and $0.059 \pm 0.006$.
The ratio of branching fractions $\ensuremath{{R}_{K/\pi}}\xspace$
is obtained by correcting the yield ratio
with the relative efficiency, as shown in Eq.~\ref{eq:Rk}.
The total efficiencies include contributions from
the LHCb detector acceptance and from selection, trigger and particle identification requirements.
The selection and trigger efficiencies are calculated from simulated samples.
The simulated events are weighted to account for
differences from data in the track multiplicity distribution.
It has been checked that after this weighting, the distributions of the variables used as inputs to the BDT
are similar in data and simulation.
The particle identification efficiencies for hadrons are evaluated from simulation calibrated with
a control sample of $D^{*+} \ensuremath{\rightarrow}\xspace {\ensuremath{\D^0}}\xspace {\ensuremath{\pion^+}}\xspace$, ${\ensuremath{\D^0}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{\kaon^-}}\xspace {\ensuremath{\pion^+}}\xspace$ decays.
The efficiency ratio is determined to be
$\epsilon(\ensuremath{\Bcp\to\jpsi\pip}\xspace)/ \epsilon(\ensuremath{\Bcp\to\jpsi\Kp}\xspace) = 1.277\pm 0.007$
and $1.284 \pm 0.006$ for $7 \ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ and $8 \ifthenelse{\boolean{inbibliography}}{\ensuremath{~T\kern -0.05em eV}\xspace}{\ensuremath{\mathrm{\,Te\kern -0.1em V}}}\xspace$ data, respectively.
The efficiency difference between \ensuremath{\Bcp\to\jpsi\pip}\xspace and \ensuremath{\Bcp\to\jpsi\Kp}\xspace mainly arises from
particle identification for the hadrons.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.48\linewidth]{fig1a.pdf}
\includegraphics[width=0.48\linewidth]{fig1b.pdf}\\
\vspace*{0.5cm}
\includegraphics[width=0.48\linewidth]{fig1c.pdf}
\includegraphics[width=0.48\linewidth]{fig1d.pdf}
\vspace*{-0.5cm}
\end{center}
\caption{ Fits to the reconstructed \ensuremath{\Bcp\to\jpsi\Kp}\xspace (left) and \ensuremath{\Bcp\to\jpsi\pip}\xspace (right) mass distributions
using 7 TeV (top) and 8 TeV (bottom) data samples.
The contributions from the signal, the misidentification background,
the combinatorial background and the partially reconstructed background
are indicated in the figures.
}
\label{fig:fit}
\end{figure}
|
\section{Introduction }
White dwarfs, compact object supported by electron degeneracy pressure, have the well known maximum mass limit of 1.4 M$_\odot$ \citep{Chandrasekhar1931}. If accretion raises the mass above this Chandrasekhar limit, the white dwarf is no longer able to support itself against gravitational collapse, and the resulting rapid contraction leads to a Type-Ia supernova. The characteristic mass limit sets the standard properties of the Type-Ia supernova. Recently however a few cases have been observed of over-luminous Type Ia supernovae that require white dwarfs well above this mass limit ($\mbox{$\stackrel{>}{_{\sim}}$}$ 2 M$_\odot$) to explain their properties \citep{howel06,hicken_07,yamanaka_09, Scalzo+2010, Tanaka+2010, Silverman+2011, Taubenberger+2011}. \cite{das_m12} explored the possibility of super-Chandrasekhar mass configurations arising from quantum mechanical modification of the degenerate electron equation of state in the presence of ultra strong internal magnetic fields. The effect of this turns out to be sub-dominant to
that of the Lorentz force; the latter by itself can raise the maximum mass well above the Chandrasekhar limit \citep{Ostiker_Hartwick68, Bera+Bhattacharya2014}. The maximum mass of the magnetically supported configurations is dependent on the field geometry. Among axisymmetric structures, the maximum mass is about 1.9 M$_\odot$ for pure poloidal field \citep{Bera+Bhattacharya2014, das_m15, Franzon+Schramm2015} and more than 5 M$_\odot$ for pure toroidal field \cite{Bera+Bhattacharya2016}, with intermediate values for mixed field configurations. These limits refer to equilibrium structures without consideration of stability. In this paper, we study the stability of these equilibrium configurations.
Magnetic field is ubiquitous in the compact stars, be it neutron stars or white dwarfs. The highest magnetic field measured at the neutron star surface is $\sim 10^{15}$ G and that of a white dwarf is $\sim 10^9$ G \citep{Schmidt+2003}. While the field external to the star is primarily poloidal, \cite{Prendergast1956} suggests that in the interior both poloidal and toroidal fields must be present to ensure long term stability. The consideration of the minimum energy principle \citep{Bernstein+1958} indicates that equilibrium configurations with pure poloidal \citep{marke73} or pure toroidal \citep{tayler1973} magnetic field are unstable. For such field geometries, perturbations in the matter and the magnetic field close to the neutral line (viz. the locus of vanishing magnetic field in the stellar interior, enclosed by field lines) can generate states of lower energy than the unperturbed configuration. Perturbation of the system would therefore drive it to a new configuration by rearranging
the magnetic field and matter. This magnetic instability is intrinsic to the field geometry, even if the magnetic energy is small compared to the thermal and the gravitational energy of the configuration (see e.g. \cite{Flowers+Ruderman77}). \cite{tayler1973} and \cite{Acheson1978} show that for a pure toroidal configuration the non-axisymmetric azimuthal mode $m = 1$ is the dominant instability mode with very short instability time scale (Alfv\'en crossing time).
The non-linear evolution of the magnetic configurations with pure poloidal and pure toroidal field shows instability with growth time comparable to the Alfv\'en time of the configuration \citep{Braithwaite+Spruit2006, Braithwaite2006b, Bonanno+Uprin2013a, Bonanno+Uprin2013b, Bonanno+Uprin2013c, IbanezMejia+Braithwaite2015}. Configurations that show long term dynamical stability in numerical experiments with a stably stratified star contain comparable amounts of energy in poloidal and toroidal components \citep{Braithwaite+Nordlund2006, braithwaite09}. Numerically evolved axisymmetric or non-axisymmetric stable structures have been found for configurations with non-barotropic (stably stratified) equation of state \citep{mitchell+2015} and a helical initial field distribution with random or mixed poloidal-toroidal field \citep{Braithwaite2008}.
Numerical studies of the evolution of neutron stars with pure poloidal fields in general relativistic formalism have been carried out by \cite{Lasky+2011, Ciolfi+2011, Ciolfi+Rezzolla2012}. These studies also indicate the presence of instability near neutral line.
Magnetars, a set of neutron stars with a strong surface magnetic field ($\sim 10^{15}$ G), show repeated gamma-ray flares. Quasi-periodic oscillations (QPOs) observed in the tail of giant flares provide evidence of neutron star oscillations \citep{Isreal+2005, Strohmayer+Watts2005}. Such oscillations of strongly magnetized neutron stars have been investigated by various authors for both axisymmetric \citep{Glampedakis+2006, Lee2008, gabler+13, gabler+12} and non-axisymmetric \citep{Lander+2010, Lander+Jones2011a, Lander+Jones2011b, Asai+2015, Asai+2016} modes.
In this paper, we study the linear and non-linear evolution of perturbed variables of a magnetized equilibrium structure. The perturbation equations and the methods of evolution are described in Section~\ref{eq+method}. In Section~\ref{results} we present in brief the results obtained, which we discuss in Section~\ref{discussion}. Our conclusions are summarized in Section~\ref{conclusion}.
\section{Governing equations and Methods } \label{eq+method}
Here we model the self-gravitating, magnetized degenerate star with perfect conductivity in Newtonian gravity. The Newtonian description of gravity is adequate for this study as general relativity alters the white dwarf structure only by a few percent at maximum \citep{Bera+Bhattacharya2016}. We use a spherical polar coordinate system ($r$,~$\theta$,~$\phi$). The governing equations to describe the system are expressed as~:
\begin{align}\label{hydro_eqs}
\left(\frac{\partial\mathbf{v}}{\partial t}+(\mathbf{v}\cdot\nabla)\mathbf{v}\right) &= -\frac{1}{\rho}\mathbf{\nabla} P -\mathbf{\nabla} \Phi_g+\frac{1}{\rho}\left( \mathbf{J}\boldsymbol\times\mathbf{B}\right) \\
\frac{\partial\rho}{\partial t} &= -\nabla\cdot(\rho\mathbf{v})\\
\frac{\partial \mathbf B}{\partial t} &= \nabla\times(\mathbf{v}\times\mathbf{B})\\%-\textcolor{cyan}{\nabla\psi}\\
\mathbf{\nabla}^2\Phi_g &= 4\pi G \rho\\
P &= P(\rho
\end{align}
Here, $\mathbf{v}, P, \rho, \Phi_g, \mathbf{J}, \mathbf{B}$ are the non-rotating part of the fluid velocity, pressure, matter density, gravitational potential, current and magnetic field respectively. The magnetic field components also satisfy divergence free condition $\mathbf{\nabla\cdot B} = 0$ and $\mathbf{\nabla\times B} = \mu_0\mathbf J$, $\mu_0$ being the free space permeability. The functional form of $P(\rho)$ describes the equation of state (EoS). Here we consider Fermi degenerate EoS to model the white dwarf and polytropic EoS ($P \propto \rho^{1+\frac{1}{n}}$, $n$ : polytropic index) in some cases to verify the results.
To study the evolution of the perturbed variables over the background equilibrium structure, initially the equilibrium configurations are constructed. These equilibrium solutions without intrinsic fluid velocity are calculated from the time-independent part of the above equations.
\begin{align}\label{hydro_eui}
0 &= -\frac{1}{\rho_0}\mathbf{\nabla} P_0 -\mathbf{\nabla} {\Phi_g}_0+\frac{1}{\rho_0}\left( \mathbf{J_0}\boldsymbol\times\mathbf{B_0}\right) \\
\mathbf{\nabla}^2{\Phi_g}_0 &= 4\pi G \rho_0\\
P_0 &= P(\rho_0
\end{align}
Here, the variables with the subscript zero represent the time independent or the equilibrium part of the variable. To solve for equilibrium axisymmetric magnetic configurations we follow the self-consistent field method \citep{hachi86, tomim05, lande09, Bera+Bhattacharya2014}. From the condition of axisymmetry we are restricted to only specific field geometries: pure poloidal, pure toroidal and mixed poloidal-toroidal to a certain extent.
After obtaining an equilibrium solution we add perturbation to it and aim to study the evolution of the variables. A radial perturbation generates stable oscillations when the effective polytropic index $n<3$ and this is generally satisfied in a white dwarf except for extreme relativistic configurations. Non-radial perturbations of the spherical star are classified as polar and axial depending on the parity. On the spherical surface polar perturbation can be decomposed in terms of $\hat rY_{lm}$ and $\nabla Y_{lm}$ and the axial perturbations into $\hat r\times\nabla Y_{lm}$. The polar perturbation generates $f,p$ mode oscillations with pressure as the restoring force and in a rotating star the axial perturbation generates $r$ modes from the Coriolis force \citep{Papaloizou+Pringle1978}. In a magnetic configuration axial perturbation excites the magnetic wave modes due to the presence of Lorentz force. Being interested in the evolution of the magnetic field we provide axial perturbations to the magnetic configurations. To keep the computation less expensive we assume that Cowling approximation is valid, i.e. the
gravitational potential remains fixed to the initial condition. Cowling approximation is known to alter the frequencies of the acoustic waves (e.g., $f, g$), but the deviation is never more than 20\% in either Newtonian \citep{Cox1980} or relativistic \citep{Yoshida+Kojima1997} treatment. Modes of higher order are affected less as the effect of oscillating components get averaged out. Although there is no direct study of the effect of the Cowling approximation on $r$ modes \citep{Kastaun2008} and magnetic modes, the approximation may be considered to be reliable as long as the perturbation does not significantly modify the equilibrium structure. In the following subsections, we describe the linear and non-linear methods followed to study the evolution of the perturbations to the equilibrium configurations.
\subsection{Linear perturbation }
To study the perturbations of the star in the linear regime we evolve the perturbed variables over the background equilibrium configuration. Such studies are common in the context of stellar pulsations of both non-magnetic \citep{Passamonti+2009, Jones+2002, Lockitch+Friedman1999, Papaloizou+Pringle1978} and magnetic \citep{Lander+2010, Lander+Jones2011a, Lander+Jones2011b, Asai+2015, Asai+2016} stars. The properties of the various modes of the oscillating stars are studied either using normal mode calculations or using MHD simulations. Here we are interested in investigating the axisymmetric strongly magnetized configurations which are generally deformed to either prolate or oblate shape depending on the field structure. Hence, we followed the $\phi$-decomposed MHD methods \citep{Lander+2010, Lander+Jones2011a, Lander+Jones2011b} to utilize the axisymmetry, and to deal with non-sphericity.
\subsubsection{Perturbed equations}
The general variables can be expressed as a sum of the equilibrium part and the perturbed part, i.e, $P = P_0+\delta P$, $\rho = \rho_0+\delta\rho$, $\mathbf{B} = \mathbf{B_0}+\delta\mathbf{B}$, $\mu_0\mathbf{J_0} = \nabla\times\mathbf{B_0}$, $\Phi_g={\Phi_g}_0$. The time dependent equations with terms of linear order in the perturbed variables can be expressed as,
\begin{align}\label{hydro_pertb}
\frac{\partial\mathbf{f}}{\partial t} & = \left[-\mathbf{\nabla} \delta P + \frac{\mathbf{\nabla} P_0}{\rho_0}\delta\rho \right] \nonumber\\
& -\frac{1}{\rho_0}\left(\mathbf{J_0}\boldsymbol\times\mathbf{B_0}\right)\delta\rho + \frac{1}{\rho_0}\left(\mathbf{J_0}\boldsymbol\times\boldsymbol{\beta}\right) \nonumber\\
& + \left[\frac{1}{\rho_0}\left(\nabla\times\boldsymbol{\beta}\right)\times\mathbf{B_0} - \frac{1}{\rho_0^2}\left(\nabla\rho_0\times\boldsymbol{\beta}\right)\times\mathbf{B_0}\right]\\
\frac{\partial\delta\rho}{\partial t} & = -\nabla\cdot\mathbf{f}\\
\frac{\partial \boldsymbol\beta}{\partial t} &= \nabla\times(\mathbf{f}\times\mathbf{B_0})-\frac{\nabla\rho_0}{\rho_0}\times(\mathbf{f}\times\mathbf{B_0})\\
\delta P & = \left.\frac{\partial P}{\partial \rho}\right|_0\delta\rho \label{hydro_pertb_end
\end{align}
Here, $\mathbf{f}=\rho_0\mathbf{v}$ and $\boldsymbol\beta=\rho_0\delta\mathbf{B}$. Each of these perturbed quantities is further decomposed in azimuthal angle $\phi$ with index $m$, e.g. the perturbed pressure
\begin{align}
&\delta P (t, r, \theta, \phi) = \nonumber\\
&\sum_{m=0}^{m=+\infty}[\delta P^+(t, r, \theta) \cos m\phi + \delta P^-(t, r, \theta) \sin m\phi].
\end{align}
This reduces the three-dimensional computation to a two-dimensional grid which helps limit the computation time.
\subsubsection{Initial values and Boundary conditions}
The magnetic equilibrium structure is calculated using the self-consistent field technique \citep{hachi86, tomim05, lande09, Bera+Bhattacharya2014}. Starting from an arbitrary initial configuration, a force balanced equilibrium configuration satisfying the virial condition is achieved iteratively. The solution is obtained for polytropic or Fermi degenerate equation of state. This force balanced solution is used as the background solution for perturbation study.
These magnetic equilibrium structures are non-spherical. In the linear perturbation study, to avoid difficulty in imposing boundary conditions on the non-spherical shape we use a modified radial coordinate $x=x(r, \theta)$, fitting the surfaces of constant pressure of the unperturbed star \citep{Jones+2002}. The partial differentials are expressed as
\begin{align}
\left.\frac{\partial}{\partial r}\right|_\theta &= \left.\frac{\partial x}{\partial r}\right|_\theta \left.\frac{\partial}{\partial x}\right|_\theta,\\
\left.\frac{\partial}{\partial \theta}\right|_r &= \left.\frac{\partial x}{\partial \theta}\right|_r \left.\frac{\partial}{\partial x}\right|_\theta + \left.\frac{\partial}{\partial \theta}\right|_x.
\end{align}
Here $x$ equals $r$ at a specific $\theta$ value $\theta = \theta_x$. We consider $\theta_x = \uppi/2 ~\text{and}~ 0$ for a prolate and oblate configuration respectively. The factors containing the partial derivative of $x$ with respect to $r$ and $\theta$ can be calculated at any point using the following relations,
\begin{align}
\left.\frac{\partial x}{\partial r}\right|_\theta &= \left.\left.\frac{\partial \rho}{\partial r}\right|_\theta \right/\left.\frac{\partial \rho}{\partial x}\right|_{\theta_x},\\
\left.\frac{\partial x}{\partial \theta}\right|_r &= \left.\left.\frac{\partial \rho}{\partial \theta}\right|_r \right/\left.\frac{\partial \rho}{\partial x}\right|_{\theta_x} .
\end{align}
At the stellar surface the Lagrangian pressure perturbation is zero i.e. for a displacement vector $\boldsymbol{\xi}$,
\begin{equation}
\delta P + \boldsymbol{\xi}\cdot\nabla P_0 = 0.
\end{equation}
At the outer surface matter ($\rho_0$) and current ($\mathbf{J_0}$) densities vanish and Eq. $\ref{hydro_eui}$ provides us $\nabla P_0 = 0$. Therefore, at the outer surface $\delta P = 0$. Proper modeling of the atmosphere outside the stellar solid crust in order to study the numerical evolution of the field is difficult due to the very high value of the Alfv\'en velocity. As we are mainly interested in the evolution of the internal field, here we do not consider any evolution of the field or material outside the surface and we assume for simplicity that $\boldsymbol{\beta}=\mathbf{f}=0$ at the outer boundary. For the condition at the center we consider $\delta P(x=0) = 0$ $\boldsymbol{\beta}(x=0)=\mathbf{f}(x=0)=0$ as we plan to study the evolution of the non-axisymmetric modes (i.e. $m >0$). The perturbed variables also vanish at the pole for $m \neq 1$. For $m = 1$, $f_\theta, f_\phi,\beta_\theta ~\text{and}~ \beta_\phi$ are non-zero conserving their $\theta$-gradient.
Depending on the symmetry of the initial perturbation, perturbations of a specific class either axial or polar are excited. For the polar type of perturbation we use a spherical harmonic profile for initial density perturbations $\delta \rho \sim Y_{lm}(\theta, \phi)$. The axial type of perturbation is excited by introducing magnetic spherical harmonic perturbation $\mathbf f \sim \hat{r} \times \nabla Y_{lm}(\theta, \phi)$ in velocity.
\subsubsection{Numerical code}
The set of $\phi$ decomposed linearly perturbed equations \ref{hydro_pertb}-\ref{hydro_pertb_end} are evolved forward in time from the initial condition mentioned above by using a MacCormack predictor-corrector algorithm \citep{MacCormack1969}. To dissipate spurious higher order oscillations, resulting from finite order grid differentiation, we impose additional fourth order Kreiss-Oliger dissipation. To ensure divergence free condition of the evolved perturbed field we use hyperbolic-parabolic divergence cleaning method \citep{Dedner+2002, Lander+2010}.
To check the accuracy of the code we mainly rely on convergence tests. We provide the initial pressure perturbation as $\delta P \sim \rho \left(\frac{r}{R(\theta)}\right)^lY_{ll}(\theta, \phi)$ along with the equilibrium solution of a polytropic star of polytropic index n = 1 and study the $f-$mode oscillations for $l=m=2$. The perturbations evolve with a single peak frequency. But in the presence of rotation, it is split into two frequencies and the separation between them depends on the rotation speed. Fig.\ref{f_mode_power} show the $f-$mode oscillations of the rotating polytropic stars. These results are identical to the results reported in \citep{Jones+2002, Passamonti+2009}. Hence this verifies the performance of the non-magnetic parts of the code.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{f_modes.pdf}
\caption{A. Power spectra (in normalised unit) of $l=m=2$ $f$-mode oscillation from a polytropic star of polytropic index n = 1. B. Splitting of power spectra of a rotating star into co-rotating (lower frequency) and counter-rotating components. C. Dependence of mode frequencies on the rotation speed.}
\label{f_mode_power}
\end{figure}
Magnetic configurations suffer from various kinds of instability depending on the field geometry. Here we proceed by choosing some of the known unstable modes and find the corresponding instability time scale.
We study the evolution of the magnetic configuration for different grid resolutions and observe the appearance of instability. The initial perturbation is provided in the velocity terms $\mathbf f \sim \hat{r} \times \nabla Y_{lm}(\theta, \phi)$. Fig. \ref{instability_accuracy} suggests that the results are independent of the grid resolution and thus confirms the presence of instability.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{toroidal_Emag_vs_time.pdf}
\caption{ m = 1 instability of the magnetic degenerate star with a pure toroidal field. The perturbed magnetic energy (in normalised unit) increases with time. The behavior is independent of the grid resolution.
}
\label{instability_accuracy}
\end{figure}
\subsection{Nonlinear perturbation }
To study the evolution of the perturbation in the non-linear regime we use the MHD code $\textsc{pluto}$ which integrates a system of multidimensional conservation equations using Godunov-type shock-capturing schemes \citep{Mignone+2007}. This code is widely used to study various astrophysical problems related to accretion/protoplanetary disk, jet, outflow etc. We are interested in studying the evolution of the stellar internal field. Here also to avoid the modeling of the outer atmosphere we consider a spherical domain concentric to the stellar center and within the star. This spherical computation domain excludes the outer part near the surface with a very low density. We solve the MHD equations in $\textsc{pluto}$ using hll Riemann solver \citep{Harten+1983}. The divergence-free condition $\nabla\cdot \mathbf{B} = 0$ is enforced in the solution by coupling the induction equation to a generalized Lagrange multiplier \citep{Dedner+2002, Mignone+2010}. Here we consider the values of the parameters to be
fixed
to their initial
values at the inner radial and impose the axisymmetric condition at the axial points. At the outer boundary, we maintain the fixed gradient of the equilibrium values such that the boundary allows the flow of matter \citep{Mukherjee+2013a}. This fixed gradient condition at the outer boundary may not suitable in a situation with strong inflow/outflow but provides an effective boundary for equilibrium with small perturbations. We evolve the perturbed polytropic star adiabatically with the adiabatic index $\gamma=\frac{5}{3}$. Here we assume the white dwarf as a polytropic star with polytropic index $n=1.5$ which is the non-relativistic approximation of the degenerate EoS. Hence for the non-linear evolution, we consider low mass ($<$M$_\odot$) magnetic white dwarfs where relativistic EoS does not modify the structure significantly.
In the non-linear evolution process, we compute perturbed quantities as the difference in the value of a variable at a given time from that in the equilibrium solution. These perturbed quantities are three dimensional variables. To identify non-radial modes we choose a $r$ value where the magnitude of the perturbation is close to maximum and evaluate the $Y_{lm}$ coefficients of the $\theta-\phi$ plane data.
\section{Results } \label{results}
Here we present the results obtained from the linear and non-linear evolutions of the perturbation. We provide specific perturbations to the equilibrium solution to study the instability. The non-linear stability is studied only for cases with pure poloidal and pure toroidal field assuming polytropic EoS whereas for all other cases we study the evolution using Fermi degenerate EoS with $\mu=2$ ($\mu$~: atomic mass per electron).
\subsection{Pure toroidal}
\subsubsection{linear perturbation}
A configuration with pure toroidal field contains field only within the surface. The axial points have zero fields i.e. they form the magnetic neutral line (Fig.~\ref{lin_tor}a). Analytical perturbation study shows the presence of unavoidable non-axisymmetric (m=1) instability near the polar axis \citep{tayler1973}. To study this Tayler instability we consider the initial velocity perturbation as $\mathbf f \sim \hat{r} \times \nabla Y_{11}$. While the background field is only toroidal, the generated perturbed field contains all field components. The perturbed field energy increases with time linearly in the early phase but exponentially at later epochs (Fig.~\ref{lin_tor}c). The exponential increase in the perturbed field energy is the indication of instability. This being a linear calculation, the perturbation evolves over the background equilibrium solution and the total energy is not conserved in case of instability. Fig.~\ref{lin_tor}c shows that the appearance of the instability, i.e. the
time when the magnetic energy starts to increase exponentially, is proportional to the Alfv\'en time ($\tau_A$). The Alfv\'en crossing time is the characteristic time of a magnetic configuration and is inversely proportional to the average field strength. This is estimated as $\tau_A\approx\frac{R}{\langle c_A\rangle} = R\sqrt{\frac{\mu_0\langle\rho\rangle}{\bar B^2}}$, where $R$ is the stellar radius, $\langle c_A\rangle$ is the volume averaged Alfv\'en speed. To quantify the growth rate of the instability we assume an exponential growth near the beginning of the instability and introduce a term $\zeta$,
\begin{equation}
\zeta = \frac{\Delta(\log\delta M)}{\Delta t}.
\end{equation}
Fig.~\ref{lin_tor}d shows that the instability growth rate is proportional to the average field strength. Therefore a strongly magnetized configuration with a pure toroidal field has a very short instability time scale.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{lin_tor.pdf}
\caption{Linear perturbation study of white dwarfs with pure toroidal field: a) Toroidal field distribution of the equilibrium configuration with mass 0.88 M$_\odot$ and $\mathcal{M}/W$=0.8\%. b) The ratio of perturbed field magnitude to the equilibrium field after the onset of the instability for the $m = 1$ evolution. The instability appears near the axial region of this configuration. c) The evolution of perturbed field shows instability which appears corresponding to their Alfv\'en crossing time $\tau_A\sim 78, 39 ~\rm{and}~ 20~ s$. d) The growth rate ($\zeta$), at the beginning of the instability from the time evolution of Fig.~\ref{lin_tor}c, is almost linearly proportional to the average magnetic field strength.}
\label{lin_tor}
\end{figure}
\subsubsection{non-linear perturbation}
We study the non-linear evolution of the perturbed $\Gamma=\frac{5}{3}$ polytropic magnetic star with mass 0.78 M$_\odot$, radius 9859 km. The average field strength of this configuration is $6.1\times10^6$ T and the total magnetic energy is 0.8\% of the gravitational energy. Our computation region covers the range from 10\% to 90\% of the radius. The non-linear numerical evolution of this equilibrium solution (without any added perturbation) does not show any instability other than small fluctuations. For the perturbation study, axial perturbations as initial velocity $\mathbf v \sim \hat{r} \times \nabla Y_{22}(\theta, \phi)$ were added. The evolution of this perturbed configuration shows instability in the growth of the perturbed magnetic energy after about the Alfv\'en time $\tau_A \sim 35.7 ~s$ (Fig.~\ref{nonlin_tor}a). In the early phase, the perturbed field grows linearly before the instability. The time sequence plot of the total kinetic energy shows initial very minor change and
then a sudden increase. This sudden increase in kinetic energy begins a little before the field growth which indicates that the enhanced velocity field induces the growth of the field. After the instability, the perturbed field magnitude shows saturation but the kinetic energy and the total magnetic energy decay significantly. This decay in magnetic energy may happen due to the assumed fixed gradient condition at the outer boundary which does not consider the whole configuration. Fig.~\ref{nonlin_tor}b shows that at the early time $l=m=1$ is the only dominating mode in the $\theta$-component of the perturbed field. At a later time, the amplitude of this specific mode increases, as well as other modes appear. To check the consistency of the results we study this non-linear evolution using $64^3$ and $128^3$ grid. The properties of the instability and other characteristics of the time evolution are nearly independent of the grid resolution (Fig.~\ref{nonlin_tor}a). We also study the linear
evolution of $m=1$ mode perturbation for this polytropic configuration and observe the characteristics. Both the linear and non-linear study show consistent behavior in their time evolution. The main difference between the two is the non-appearance of the saturation in the linear case as there is no back reaction to modify the background solution.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{nonlin_tor.pdf}
\caption{Non-linear perturbation study of white dwarfs with a pure toroidal field: a) The evolution of i) perturbed magnetic energy ($\delta \mathcal M$), ii) kinetic energy (K.E.) and iii) total magnetic energy ($\mathcal M$). The results are almost independent of the grid resolution as the value of the parameters for $64^3$ and $128^3$ are identical. The linear evolution of the $m=1$ perturbation mode of this polytropic star is shown for comparison. The non-linear evolution of the equilibrium configuration without any perturbation shows the static characteristics. b) Early time (t~=~35~s) spherical harmonic mode components of the $\theta$-component of the perturbed field ($\delta B_\theta$) evaluated at $r/R=0.6$. c) Late time (t~=~83~s) mode components of $\delta B_\theta$ at the same $r$ value.}
\label{nonlin_tor}
\end{figure}
\subsection{Pure poloidal}
\subsubsection{linear perturbation}
For a pure poloidal field configuration, the magnetic neutral line forms a circle on the equatorial plan. In the case of pure toroidal field structure with the magnetic neutral line along the axis, the perturbations evolve in the $\theta-\phi$ plane and we study $\phi$-decomposed mode. The perturbations generated from magnetic instability in a pure poloidal configuration will evolve in the $r-\theta$ plane. As we use our coordinate system with the origin at the stellar center, it is difficult to specify the mode of the perturbations from the symmetry. Here we study the evolutions of the $l=m=2$ mode of the perturbations providing the initial velocity perturbation as $\mathbf f \sim \hat{r} \times \nabla Y_{22}$. The time evolution of this perturbation on the equilibrium configuration (Fig.~\ref{lin_pol}A) with pure poloidal field exhibits a significant growth of the perturbed field near the neutral line (Fig.~\ref{lin_pol}B). Fig.~\ref{lin_pol}C indicates the appearance of the instability for the perturbed
magnetic
energy after the initial linear growth. Here also the exponential growth starts in corresponding Alfv\'en time. Fig.~\ref{lin_pol}D shows that the instability growth rate varies linearly with the average field strength of the configuration.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{lin_pol.pdf}
\caption{ Linear perturbation study of white dwarfs with pure poloidal field (similar to Fig.~\ref{lin_tor}) : a) Poloidal field distribution of the equilibrium configuration with mass 0.88 M$_\odot$ and $\mathcal{M}/W$=0.8\%. b) The ratio of perturbed field magnitude to the equilibrium field after the onset of the instability for the $m = 2$ evolution. The instability appears in a region near the magnetic neutral line of the equilibrium configuration. c) The evolution of perturbed field shows instability which appears at the corresponding Alfv\'en crossing time $\tau_A\sim 69, 34 ~\rm{and}~ 17~ s$. d) The growth rate ($\zeta$), at the beginning of the instability from the time evolution of Fig.~\ref{lin_pol}c, is almost linearly proportional to the average magnetic field strength.}
\label{lin_pol}
\end{figure}
\subsubsection{non-linear perturbation}
Similar to the pure toroidal case here too we consider a $\Gamma=\frac{5}{3}$ polytropic star which has a mass 0.9 M$_\odot$, radius 9716 km and average magnetic field of strength $7\times 10^6$~T. For this configuration, the magnetic field energy is 0.9\% of the gravitational energy and the Alfv\'en crossing time $\tau_A \sim 32 ~s$. The non-linear evolution of the equilibrium structure almost remains unchanged when the variables on the spherical shell ($0.1\leq r/R\leq0.9$) are evolved using $\textsc{pluto}$. For the perturbation evolution study we add the axial velocity perturbation $\mathbf v \sim \hat{r} \times \nabla Y_{22}(\theta, \phi)$ to the equilibrium solution and let it evolve. Now, the perturbed magnetic field energy and the kinetic energy exhibit instability after about a few Alfv\'en crossing time (Fig.~\ref{nonlin_pol}a). It is also observed that after these values attain some kind of saturation the total magnetic energy decays, indicating
extreme non-linear interactions and deformation of matter flow and field geometry. The Newtonian evolution properties ( e.g. appearance of instability, significant magnetic field energy decay within a few Alfv\'en times after the beginning of instability, generation of toroidal field energy and its rise to a level comparable that of the poloidal field) match with the general relativistic evolution of the poloidal neutron star explored by \cite{Ciolfi+Rezzolla2012}. The linear study of the $m=2$ mode also shows similar instability characteristics but in this case the instability happens a little earlier in comparison to the non-linear case. This may indicate that the particular mode (i.e. $m=2$) used to study linear evolution may not be the dominant one. The mode decomposition of the $\phi$-component of the perturbed field shows initially $m=1$ or $m=2$ modes (Fig.~\ref{nonlin_pol}b) but later many other modes arise (Fig.~\ref{nonlin_pol}c).
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{nonlin_pol.pdf}
\caption{Non-linear perturbation study of white dwarfs with pure poloidal field configuration (similar to Fig.~\ref{nonlin_tor}) : a) The evolution of i) perturbed magnetic energy ($\delta \mathcal M$), ii) kinetic energy (K.E.) and iii) total magnetic energy ($\mathcal M$). The linear evolution is done for $m=2$ mode of the perturbation. b) Early time (t~=~42~s) spherical harmonic mode components of the $\phi$-component of the perturbed field ($\delta B_\phi$) evaluated at $r/R=0.6$. c) Late time (t~=~139~s) mode components of $\delta B_\phi$ at the same $r$ value.}
\label{nonlin_pol}
\end{figure}
\subsection{Effects of Mixed field, Rotation} \label{mixed+rotation}
In the last two subsections, we observed that the pure toroidal and pure poloidal field configurations are unstable as they show instability when we perturb the system. Both of these field configurations, assumed here for the perturbation study, are idealisations. We may expect more realistic magnetic configurations to contain both poloidal and toroidal components without any specific symmetry. On the other hand, the intrinsic spin of the star may play a role in aligning the magnetic field axis. Although many such complex possibilities exist, our investigation in this paper remains restricted only to axisymmetric configurations.
As mentioned above, the axisymmetric equilibrium configurations we deal with are computed using the self-consistent field method. One of the limitations of this method is that one can not obtain a configuration with comparable poloidal and toroidal field. The maximum energy of the toroidal component in the mixed field configuration is limited to less than 10\% \citep{armaza+15} of the total. Although \cite{Ciolfi+Rezzolla2013} found mixed field configurations with a significant toroidal component using a perturbative method, we do not find such solutions to represent equilibrium structures. Here we study the linear evolution of the perturbation of a mixed field configuration containing about 2\% of the magnetic energy in the toroidal form (Fig.~\ref{lin_mix}A). As the total energy in the toroidal component is not significant compared to that in the poloidal component we find the magnetic instability characteristics to be similar to the pure poloidal case (Fig.~\ref{lin_mix}C). However, in an equatorial region near the stellar surface, where the local toroidal field significantly exceeds the local poloidal component, a change in the nature of the instability is observed. A comparison of Fig.~\ref{lin_pol}B and Fig.~\ref{lin_mix}B illustrates this difference. At the boundary of the toroidal field region the poloidal current distribution displays a discontinuity. Potential numerical errors associated with this restricts our present study to configurations with relatively low toroidal field. The location of the region being very close to the surface of the star also makes it difficult to define a \textsc{pluto} computation box sufficiently in the interior. Hence for this case we restrict ourselves to linear perturbation study.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{lin_mix.pdf}
\caption{Linear instability of a mixed field configuration. a) Equilibrium field distribution is mainly dominated by the poloidal field. The toroidal field component is there within the closed field line bounded by the stellar surface. b) The distribution of the perturbed field compared to the equilibrium field strength. c) The time variation of the perturbed field magnitude shows instability similar to the pure poloidal field configuration.}
\label{lin_mix}
\end{figure}
Axisymmetric configurations with uniform rotation along the symmetry axis can be obtained by introducing the centrifugal force in the equilibrium force balance equation.The centrifugal force acts outward from the rotation axis and is proportional to the square of the angular frequency. In the presence of this force, one can get an equilibrium solution upto a frequency, known as Keplerian frequency, beyond which a bound object cannot be formed. To study the perturbation of these configurations one must consider the Coriolis force term in the perturbation equations. The presence of Coriolis force provides $r$-mode oscillations for a non-magnetic star and reduces the instability growth rates of the magnetic configurations. Fig.~\ref{tor_rot_growth_rate} shows the variation of the growth rate depending on the rotation frequency of a magnetic white dwarf of mass 0.88 M$_\odot$ and $|\frac{\mathcal M}{W}| = 0.8$\%. Here from the
linear study we observe that the instability growth rate decreases by more than 50\% as the rotation speed approaches Keplerian frequency.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{tor_rot_growth_rate.pdf}
\caption{The growth rates ($\zeta$) of the rotating 0.88 M$_\odot$ white dwarfs with pure toroidal field ($|\mathcal M/W| = 0.8$\%). The maximum rotation frequency presented here is close to its Keplerian frequency. The growth rate reduces by more than 50\% as the rotation frequency increases. }
\label{tor_rot_growth_rate}
\end{figure}
\subsection{Super-Chandrasekhar mass white dwarfs}
The presence of strong ordered magnetic field can deform the equilibrium structure. Depending on the field geometry and hence the direction of Lorentz force the equilibrium structure can be prolate or oblate. As the Lorentz force becomes significant it can support more mass relative to the non-magnetic configuration. Fig.~\ref{MR_alfven_time} shows that the mass-radius relations of the magnetic white dwarfs with equal $|\mathcal M/W|$ ratio are shifted to the higher mass for both the pure poloidal and pure toroidal field geometry. The magnetically supported super-Chandrasekhar mass white dwarfs have very short Alfv\'en crossing time scales, typically less than a second.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{M_Ralfven}
\caption{The mass-radius relation of white dwarfs with pure poloidal (upper panel) and pure toroidal field. The maximum mass for a pure poloidal field is about 1.9 M$_\odot$ and for pure toroidal field it is more than 5 M$_\odot$. The coloured squares exhibit the Alfv\'en crossing time ($\tau_A$) of these configurations.}
\label{MR_alfven_time}
\end{figure}
We study the linear evolution of perturbation in magnetized super-Chandrasekhar mass white dwarfs with central density $2\times 10^{13}~ \rm{kg/m}^3$. Fig.~\ref{lin_SuCh} shows that the equilibrium mass increases as the ratio of magnetic to the gravitational energy of the configuration increases, irrespective of the field geometry -- whether pure poloidal or pure toroidal.
The instability growth rates for these configurations vary almost linearly with the effective magnetic field strength, defined as the square-root of the ratio between magnetic to gravitational energy ($\sqrt{|\mathcal{M}/W|}$). This behaviour is akin that obtained using average field in configurations with field strengths too low to affect the stellar structure (Fig.~\ref{lin_tor}d \& Fig.~\ref{lin_pol}d) . The non-linear evolution of these perturbed highly massive magnetic configurations is expected to show instability behaviour similar to those presented above for low-mass structures. We do not attempt to study the adiabatic non-linear evolution of the perturbation as the full Fermi degenerate equation of state of super-Chandrasekhar mass white dwarfs has a varying polytropic index. The linear evolution growth rates are presented for magnetic configurations with $|\mathcal{M}/W|$ value upto to $\sim$10\%. Beyond this the value is highly dependent on the grid resolution. The obtained growth rates of these cases
from the linear study are very high ($>10~s^{-1}$), indicating the highly unstable nature of these configurations.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{lin_SuCh.pdf}
\caption{The effective magnetic field ($\sqrt{|\mathcal M/W|}$) dependency of the configuration mass, average magnetic field and the instability time scale of supar-Chandrasekhar mass white dwarfs with central density $2\times 10^{13}~ \rm{kg/m}^3$.}
\label{lin_SuCh}
\end{figure}
\section{Discussion } \label{discussion}
In this perturbation study, we consider ideal MHD with no viscosity effects. Fields with long range order as considered here cause white dwarfs to show instability with sub-second time scales. For a typical white dwarf parameter, the viscous dissipation time is about $10^{10}$~yr and so is the resistive time scale \citep{Bera+Bhattacharya2016}. Hence, the characteristics of the instability are not significantly affected by these dissipative effects in the early stage. The presence of high-speed rotation, in general, reduces the instability growth rate of magnetic structures \citep{Acheson1978, Pitts+Tayler1985, Braithwaite2006a,
Lander+Jones2011a, Lander+Jones2011b}, as shown in Section~\ref{mixed+rotation} for the specific case of toroidal fields. White dwarfs with higher central density and higher field strength also have higher Keplerian frequency \citep{Franzon+Schramm2015}, suggesting that rotation may provide a stabilizing influence. However it can not stabilize the system completely unless the toroidal field increases very slowly with the distance from the axis \citep{Bonanno+Uprin2013c}. The toroidal field increases faster near the axis for more massive magnetic white dwarfs and this ensures that even the rotating configurations remain unstable.
Another idealization considered in this work is the zero-temperature degenerate equation of state. At the very outer layers near the surface of the star, matter may however be non-degenerate and thus have a different equation of state. This is not likely to impact on the nature of the instability shown here since the driving mechanism operates in deeper layers. In fact results for a degenerate gas differ little from those for a polytropic one, suggesting near independence on the equation of state. Long-term evolution of the instability and observables such as the luminosity may, however, be affected more significantly by the state of the outermost layers. The computation of these effects is beyond the scope of this paper.
The state of the configuration after the instability is not studied here but we observe that during non-linear evolution many large-amplitude fluctuating components with significant matter flow are created as the instability grows. The ultimate destiny of the evolved state will be decided on whether it is able to settle into a configuration in dynamic equilibrium. So far such states of dynamic equilibrium have been found only in cases with magnetic energy too low to influence the stellar structure \citep{Braithwaite+Nordlund2006, braithwaite09, Ciolfi+2011}. This indicates that while low field configurations may achieve a dynamical equilibrium state, the possibility is remote for this occurring in strongly magnetized configurations where the instability is mainly driven by the electrical currents. Our numerical study suggests that the strongly magnetized configurations have an extremely complex evolution which may lead to an eventual collapse of the system as these configurations are already very close to
the maximum mass limit.
However, the formation scenarios of white dwarfs also leave little room for generating configurations supported by ultra strong magnetic fields. Strong magnetic fields in white dwarfs are thought to be generated from either i) inheritance from the progenitor star via flux-freezing \citep{Ruderman1972}, or, ii) possible dynamo action during the common envelope phase of a binary system \citep{Potter+Tout2010, Briggs+2015}. White dwarfs are formed by gravitational contraction of stars less massive than $\sim$ 8 M$_\odot$ in a process lasting several days \citep{Woosley+Weaver1986}. If an ultra strong field is inherited from flux-freezing, then the instability discussed above would prevent the ordered field structure from being retained until the formation of the white dwarf is complete. On the other hand, the dynamo mechanism during the common envelope phase is a steady process which takes about a Myr \citep{Potter+Tout2010} to produce the field at the white dwarf surface. Instabilities with short time scales
will then cause the field, if it is ultra strong, to evolve to a more stable configuration which may not possess a significant long range order. In absence of such long range order the effective Lorentz force would be insignificant and will be incapable of supporting super-Chandrasekhar mass configurations.
\section{Conclusions} \label{conclusion}
In this paper, we have studied the linear and non-linear evolution of perturbations to an axisymmetric, strongly magnetized object. Our main results are:
\begin{enumerate}
\item Axisymmetric magnetic configurations with pure poloidal or pure toroidal field suffer from magnetic instabilities with time scale comparable to the Alfv\'en crossing time ($\tau_A$) of the configuration. As the Alfv\'en crossing time is inversely proportional to the average magnetic field of the configuration, structures with ultra strong magnetic fields, with very short Alfv\'en crossing times, are strongly unstable.
\item In the case of rotating magnetic white dwarfs, the instability growth rate reduces as the rotation speed increases. However this is insufficient to fully stabilize magnetically supported configurations near their mass limits.
\item Magnetically supported super-Chandrasekhar mass white dwarfs require extremely strong magnetic fields in the interior and are hence susceptible to instabilities with a very short growth time scale (typically less than a second). Instabilities of this nature may in fact prevent the formation of such objects.
\end{enumerate}
\section{Acknowledgement}
PB thanks CSIR, India for Research Fellowship grant SPM-09/545(0221)/2015-EMR-I. We thank E. Truhlik, H. Spruit and J. P. Ostriker for the useful comments. We also thank the referee for the valuable comments that helped us to improve the paper. The most of the numerical computations were carried out using IUCAA HPC.
\def\aj{AJ}%
\def\actaa{Acta Astron.}%
\def\araa{ARA\&A}%
\def\apj{ApJ}%
\def\apjl{ApJ}%
\def\apjs{ApJS}%
\def\ao{Appl.~Opt.}%
\def\apss{Ap\&SS}%
\def\aap{A\&A
\def\aapr{A\&A~Rev.}%
\def\aaps{A\&AS}%
\def\azh{AZh}%
\def\baas{BAAS}%
\def\bac{Bull. astr. Inst. Czechosl.}%
\def\caa{Chinese Astron. Astrophys.}%
\def\cjaa{Chinese J. Astron. Astrophys.}%
\def\icarus{Icarus}%
\def\jcap{J. Cosmology Astropart. Phys.}%
\def\jrasc{JRASC}%
\def\mnras{MNRAS}%
\def\memras{MmRAS}%
\def\na{New A}%
\def\nar{New A Rev.}%
\def\pasa{PASA}%
\def\pra{Phys.~Rev.~A}%
\def\prb{Phys.~Rev.~B}%
\def\prc{Phys.~Rev.~C}%
\def\prd{Phys.~Rev.~D}%
\def\pre{Phys.~Rev.~E}%
\def\prl{Phys.~Rev.~Lett.}%
\def\pasp{PASP}%
\def\pasj{PASJ}%
\def\qjras{QJRAS
\def\rmxaa{Rev. Mexicana Astron. Astrofis.}%
\def\skytel{S\&T}%
\def\solphys{Sol.~Phys.}%
\def\sovast{Soviet~Ast.}%
\def\siamr{SIAMR}%
\def\ssr{Space~Sci.~Rev.}%
\def\zap{ZAp}%
\def\nat{Nature}%
\def\iaucirc{IAU~Circ.}%
\def\aplett{Astrophys.~Lett.}%
\def\apspr{Astrophys.~Space~Phys.~Res.}%
\def\bain{Bull.~Astron.~Inst.~Netherlands}%
\def\fcp{Fund.~Cosmic~Phys.}%
\def\gca{Geochim.~Cosmochim.~Acta}%
\def\grl{Geophys.~Res.~Lett.}%
\def\jcp{J.~Chem.~Phys.}%
\def\jgr{J.~Geophys.~Res.}%
\def\jqsrt{J.~Quant.~Spec.~Radiat.~Transf.}%
\def\memsai{Mem.~Soc.~Astron.~Italiana}%
\def\nphysa{Nucl.~Phys.~A}%
\def\physrep{Phys.~Rep.}%
\def\physscr{Phys.~Scr}%
\def\planss{Planet.~Space~Sci.}%
\def\procspie{Proc.~SPIED}%
\let\astap=\aap
\let\apjlett=\apjl
\let\apjsupp=\apjs
\let\applopt=\ao
\bibliographystyle{mnras}
|
\section{Introduction}
Allocated resources often go to waste, even when in scarce supply. It
is common in university departments, for example, to find that all
rooms are fully booked in advance, yet walking down the corridor one
sees that many rooms are in fact empty. For examples from
different domains, consider allocating spots in
a spinning class to gym members, and assigning time slots for
a public electric vehicle charging station to residents in a
neighborhood. Even a gym member who is highly uncertain about his ability to attend the class,
or a resident unsure about her need for charging, may selfishly
reserve a space just in case
this turns out to be convenient.
\paragraph{Resource Allocation under Uncertainty}
What is common in these problems is the presence of uncertainty and a subsequent, {\em ex post} decision about utilization by allocated agents.
At the time when assignments need to be made (period zero), each agent has uncertainty in her value for using an assigned resource or even whether this will be possible at all. The uncertainty is not resolved until some time in the future (period one), at which point we assume it is too late to assign an unused resource to another agent.
Our focus in this paper is on two period problems, but more general scenarios need not be limited to two periods: gym classes may take place every week, and people's uncertainty about attending may reduce gradually between the reservation and the actual time of each class.
The common practice of giving resources to agents for free (randomly,
or on a first-come-first-serve basis) wastes resources, even
though high utilization may be the first-order objective: universities
may define what they consider to be high value uses, and beyond
that want their facilities to be utilized rather than reserved and
wasted. Similarly, a gym manager may have a preference for higher
attendance in classes to retain and attract more members to
the gym; and the ongoing funding for public charging stations may rely
on them being used.
In some settings, under-utilization
can also have broader, negative effects on society. For example, less electric vehicle charging means more pollution, or to consider another domain, unused and underdeveloped wireless spectrum means a lost opportunity for broad spillover effects to the rest of the economy.\footnote{In regard to welfare considerations--- a standard second-price auction (SP) would maximize agents' expected welfare for single resource allocation. If, in addition, we model society as gaining $C>0$ from a resource being utilized, then modifying SP by adding a penalty $C$ if the resource is not used, and allowing negative bids (i.e. allowing positive transfers from the mechanism to the agents) is welfare-optimal. But this is not a satisfactory solution to our problem because it runs a deficit. In fact, with large $C$ and insisting on no-deficit, we can prove that the welfare optimal mechanisms are precisely those that are utilization maximizing (see Appendix~\ref{appx:welfare}).}
Despite appearing important to practice and simple to state, the
problem of optimal mechanism design for utilization does not appear to
have been formally defined in the literature. The na\"ive solution
of collecting bids and running the second price (SP) auction
(and more generally, the VCG mechanism) does not necessarily assign
resources to the more reliable agents, since higher bids need not
reflect higher utilization.
Moreover, the
payments are collected in period zero and thus ``sunk,'' and do not shape incentives downstream towards higher utilization. Another worry is that in many scenarios of interest
there appears to be a ``no-charge'' norm so that agents should not be charged if they act as intended; e.g., a gym member should not be charged for attending classes when she has already paid for the gym
membership, and residents should not be charged for using the community's electric vehicle charging stations.
Payments that are contingent on utilization decisions have been used in practice to disincentivize no-shows:
hospitals charge patients
penalties that are not covered by insurance for missing appointments; organizers of conferences reimburse students their
registration fee if they attend talks; and fancy restaurants charge a
fee if guests who reserved don't show up. But we are not aware of
any formal analysis of such {\em ad hoc} mechanisms, or a design approach
that takes into account the penalty that an individual participant would
be willing to face.
This is the
main conceptual contribution of our paper. Rather, existing approaches
can be viewed as simple, first-come first-served schemes,
and where it is not clear how the penalty should be set: a penalty that is too small is not effective,
whereas a penalty that is too big will drives away participation in
the scheme.
Perfect utilization could of course be achieved through always charging the assigned agent a very large penalty for no-show, or paying a very large bonus for using a resource, but this would drive away participants or run into a deficit.
\paragraph{Contingent Payment Mechanisms for Utilization Maximization}
In this paper, we formalize the utilization-optimal mechanism design problem, and define a class of {\em contingent payment mechanisms} that make use of payments that are contingent on agents' publicly observable {\em ex post} utilization decisions. In period zero, agents make reports based on their distributional information about their future values for resources (the agent types).
The mechanism assigns the resource or resources and determine the contingent payments for each assigned agent.
When period one comes, each agent privately observes
her realized value and decides whether to use an assigned resource
based on this value as well as the contingent payments. The design
objective is to maximize utilization (i.e. the expected number of
resources utilized) in a dominant strategy equilibrium. We also
insist on voluntary participation by agents and that the mechanism is no-deficit, thus precluding always charging very large penalties and paying very large rewards.
Given that types correspond to value distributions, this is a problem
of multi-dimensional mechanism design. As would be
expected from the literature, this presents technical challenges. One way to see the impact of multi-dimensional types is to consider that the payments an assigned agent faces can have two parts, consisting of one payment when she uses the resource and one when she does not.
Because of the rich types, different types have different probabilities for using the resource, and thus different probabilities of needing to pay each part of the payments, and as a consequence, different preference orderings over these payments.
The present problem is challenging, also, because we must address
information asymmetries both before and after the time of contracting, value distributions and realized values, respectively.
\paragraph{Our Contributions}
For single resource allocation, we study the natural \emph{contingent second price (CSP) mechanism}, which elicits a single bid from each agent, allocates to the highest bidder, and charges her the second highest bid only if she does not use the resource. As an indirect mechanism, CSP avoids the inconvenience of eliciting full value distributions. We show that it is a dominant strategy in CSP to bid an amount that is equal to a the maximum no-show penalty
an agent is willing to pay~(Theorem~\ref{thm:dominant_strategy}), and that the utilization achieved by CSP is always higher than that under the SP auction (Theorem~\ref{thm:CSP_beat_SP}.)
This is because CSP selects more reliable agents than SP, since a high bid indicates an agent is more reliable and less likely to pay a penalty. %
Moreover, the use of contingent payments further promotes allocated agents to use the resources--- shaping behavior of an
agent in period one and promoting an agent to use
the resource with higher probability.
In addition to the ``no-charge'' property,(agents that act as intended do not make a payment), we adopt the natural ``always allocate''
property (i.e., the resource is always allocated to some agent.)
Given this, we prove that the CSP mechanism is the only dominant
strategy, individually rational, anonymous, and deterministic
mechanism that does not run into deficit, always allocates and satisfies no-charge (Theorem~\ref{thm:csp_uniqueness}.)
Moreover, CSP is not dominated by any mechanism
(Theorem~\ref{thm:csp_not_dominated}) if we relax the last two
conditions, and is provably optimal in this class of
mechanisms for the simple model of agents types where each agent gets a fixed value from using the resource, but is only able to do so with a fixed probability smaller than one.
We also provide both theoretical results and detailed
simulation results to establish that for simple type distributions, relaxing the last two conditions does not improve utilization on average.
In Section~\ref{sec:multi_item}, we generalize the CSP mechanism to
assign multiple resources. For allocating $m$ identical
resources, we prove that the contingent ($m$+1)$^{\mathrm{st}}$ price
mechanism is the only mechanism with the set of desired criteria, and
is again optimal among a larger set of mechanisms
(Theorem~\ref{thm:ck1p_theorem}.)
Since agents' expected utilities are not quasi-linear in
contingent payments,
the generalization of CSP to allocate multiple,
heterogeneous resources makes use of
the minimum competitive equilibrium
price mechanism for non-quasi-linear assignment~\cite{demange1985strategy,
DBLP:journals/ior/AlaeiJM16}.
We present simulation results in
Section~\ref{sec:simulations} and in the appendix, comparing
utilization achieved by this generalized CSP mechanisms
with that of the VCG mechanism and other benchmarks.
We show that a significant improvement in utilization is achieved by
the generalized-CSP mechanism, and that allowing reserve prices
does not improve utilization on average.
\subsection{Related Work} \label{subsec:related_work}
Contingent payments have arisen in previous work on auction design. Prominent examples include auctioning oil drilling licenses~\cite{hendricks1988empirical}, royalties~\cite{caves2003contracts,deb2014implementation}, ad auctions~\cite{varian2007position}, and selling a firm~\cite{ekmekci2016just}. In such auctions, payments are contingent on some observable world state (e.g. amount of oil produced, a click, or the ex post cash flow)
rather than on an agent's own downstream actions.
Thus, this is a departure from our model.
Moreover, the major role of contingent payments in these applications is to improve revenue and hedge risk~\cite{skrzypacz2013auctions}, rather than providing bidders with a way to signal their own, idiosyncratic uncertainty and thus address moral hazard.
There's an extensive literature on strictly-proper
scoring-rules~\cite{gneiting2007strictly}, but this does not not appear to
helpful for eliciting the information about uncertainty because (i)
only the actions, and not realized values are observed, and thus a
scoring-rule method could not be used to elicit beliefs about value
distributions, and (ii) the utility for using an assigned resource is
entangled with the incentives to provide accurate prediction about
one's utilization action.
\if 0
Our problem is a kind of principal-agent problem~\cite{hart1986theory,holmstrom1979moral}.
Classically, this literature addresses both
problems with hidden information (e.g. seller's quality~\cite{dellarocas2003efficiency})
before the time of contracting, which are termed \emph{adverse selection}, and problems for which information asymmetry arises after
the time of contracting (e.g. shipping a low quality good,)
the problem of \emph{moral hazard}.
The distinction between the two settings is blurred in dynamic settings (see~\cite{stole2001lectures,bolton2005contract},) as is the case for our problem, where the asymmetries exist both before and after contracting.
In particular, although agents' actions are fully observable, uncertainty together with participation constraints precludes charging unbounded penalties, the standard approach when actions are observable in moral hazard problems. We are aware of
no model or methods in this literature that addresses our problem.
\fi
Our problem is a kind of principal-agent problem~\cite{hart1986theory,holmstrom1979moral}.
Classically, this literature addresses
problems with hidden information (e.g. seller's quality~\cite{dellarocas2003efficiency}) before the time of contracting--- this is
the problem of \emph{adverse selection}. In addition, this literature
addresses problems for which information asymmetry arises after
the time of contracting (e.g. shipping a low quality good)--- this
is the problem of \emph{moral hazard}.
The distinction between the two settings is blurred in dynamic settings (see~\cite{stole2001lectures,bolton2005contract},) as is the case for our problem where the asymmetries exist both before and after contracting.
Although agents' actions are fully observable, uncertainty together with participation constraints make it impossible to charge unbounded penalties, which is a standard approach when actions are observable in settings with moral hazard. We are aware of
no model or methods in the principal-agency literature that addresses our problem.
On mechanisms with actions after the time of contracting, Atakan and Ekmecki~\cite{atakan2014auctions} study auctions where the value of taking each action depends on the collective actions by all allocated agents,
but these actions are taken before rather than after observing the world state
and thus the timing of information is quite different than our model.
Courty and Li~\cite{courty2000sequential} study the problem of revenue maximization in selling airline tickets, where passengers have uncertainty about their
value for a trip at the time of booking, and decide whether to
take a trip only after realizing their actual values. Although Courty and Li~\cite{courty2000sequential} model agents' types as distributions, and the optimal mechanism in their setting can be understood as a menu of contingent contracts, the type spaces
in their model are effectively one-dimensional (because they require stochastic dominance conditions).
The closest related work is on the design of mechanisms for
incentivizing reliability in the specific setting of demand-side
response in electric power systems~\cite{Ma_ijcai16,Ma_aamas17}, where
selected agents decide whether to respond only after uncertainty in
their costs for demand response are resolved. The objective there is
to guarantee a probabilistic target on the collective actions taken by
agents, without selecting too large a number of agents or incurring
too much of a total cost. In contrast to the models in the present
paper, there is no hard feasibility constraint in the
setting of demand response--- that is, whereas only one agent can be
assigned to each one of our resources, in demand response any number
between zero and all agents can reduce demand. This makes the present problem more challenging.
\if 0
\dcp{are you comfortable with this? feel
free to rephrase if not!} \hma{I don't think this is true, though... constraint in this problem are on the feasible allocations, whereas constraints in DR are on the probabilistic outcomes. I would actually say that constraint is harder to deal with. BUT, there is no optimality result at all in the DR paper (e.g. prove that the number of agents that is being selected is small) whereas we have optimality result here. The easier constraint that we have here made it possible to actually prove something}\dcp{hmmm, do you remember the discussion
we had when writing the grant when we disagreed on this issue of
constraints? I thought you agreed with me in the end. does it help
you to remember that discussion? anyway, please try to suggest phrasing you're ok with and we can try to converge}
\hma{I'm happy with the above. I do remember the conversation about DR having no constraint on how many people we can select, however, there is constraint on the probabilistic decisions that the selected agents are going to make. These constraints can both be considered as constraints on the agent-independent contracts that we offer to the agents, and I think the hardness on satisfying the two are similar. Selecting a smaller number of agents is somewhat the goal of the DR mechanisms, thus the mechanism is able to select any number of agents doesn't really make it a lot easier. If we think about out problem, we can also say there is no constraint on the probability that the resource must be utilized, however, having no such constraint doesn't make the current problem easier, either...}
\fi
Other papers study dynamic mechanism design in assignment settings,
including models with the possibility that workers assigned to tasks
will prove to be unreliable~\cite{porter2008fault} and problems where
the goal is to maximize expected, discounted value in the presence of
uncertainty~\cite{parkes03c,cavallo10,bergemann2010dynamic}. But we
are not aware of any work in the dynamic mechanism design literature
that is applicable to our problem, which can be construed as the limit
case in which the principal has a very large value (representing
society) for agents choosing to utilize the assigned resources.
\section{Preliminaries} \label{sec:preliminaries}
We first introduce the model for the assignment of a single resource.
There is a set $N = \{1,~2,~\dots,~n\}$ of agents and two time periods. In period 0, the value of agent $i$ for using the resources is uncertain, represented by a random variable $V_i$, whose exact (potentially negative) value is not realized until period 1.
The cumulative distribution function (CDF) $F_i$ of $V_i$ is agent $i$'s private information at period 0, and corresponds to her \emph{type}. Let $F = (F_1, \dots, F_n)$ denote a type profile.
The assignment is determined in period 0, whereas the allocated agent decides on whether to use the resource at period 1, after she privately learns the realization $v_i$ of $V_i$.
Define $V_{i}^+ \triangleq \max \{V_{i}, \; 0\}$.
We make the following assumptions about $F_i$:
\begin{enumerate}[({A}1)]
\item $\E{V_{i}^+}>0$, which means $V_i$ takes positive value with non-zero probability. An agent for which this is violated would never be interested in the resource.
\item $\E{V_{i}^+} < +\infty$, which means agents do not get infinite utility from the resource.
\item $\E{V_i} < 0$, which means the hard commitment of ``always use the resource no matter what happens" is not favorable.
\end{enumerate}
(A1) and~(A2) are clearly reasonable.
(A3) may seem like a strong requirement, however, we will see later that an agent for whom (A3) is violated would be willing to commit to pay an unbounded penalty if the resource is not used, which we don't consider
to be reasonable.
The following is an example value distribution with discrete support
and models an agent who may be completely unable to use the resource.
See also Example~\ref{ex:exp_model} in Section~\ref{sec:optimality} for an example type model where values are continuously distributed.
\begin{example}[$(w_i,p_i)$ model] \label{ex:vipi}
The value for agent $i$ to use the resource is $w_i > 0$, however, she is able to do so only with probability $p_i \in (0,1)$. With probability $1-p_i$, agent $i$ is unable to show up to use the resource due to a hard constraint. The hard constraint can be modeled as $V_i$ taking value $-\infty$ with probability $1-p_i$.
See Figure~\ref{fig:pmf_vipi}. We have $\E{V_i^+} = w_i p_i > 0$ and $\E{V_i} = -\infty < 0$, thus (A1)-(A3) are satisfied. The $-\infty$ value is not critical for (A3) --- it suffices if with probability $1 - p_i$, the agent incurs a large enough cost if she forces herself to show up.
\end{example}
\begin{figure}[t!]
\spacecheat{\vspace{-1.5em}}
\centering
{
\subfloat{
\small{
\begin{tikzpicture}[scale = 0.9][font=\normalsize]
\draw (0,0) node {$V_i = \pwfun{w_i, &\text{ w.p. }\ p_i \\
-\infty, &\text{ w.p. }\ 1-p_i}$};
\draw (0,-0.5) node{{\color{white} some text}};
\end{tikzpicture}
}}
\hspace{2em}
\subfloat{
\begin{tikzpicture}[scale = 1][font = \small]
\draw[->] (0.5,0) -- (2.8,0) node[anchor=north] {$v$};
\draw[->] (1,-0.2) -- (1,0.8) ;
\draw (1,0.7) node[anchor=west] {$f_i(v)$};
\draw[dotted] (0,0) -- (0.5,0);
\draw (-0.5, 0) -- (0, 0);
\draw (2, 0) -- (2,0.6);
\draw [fill] (2, 0.6) circle [radius=0.04] node[anchor=west] {\small{$p_i$}};
\draw (-0.25, 0) -- (-0.25,0.3);
\draw [fill] (-0.25, 0.3) circle [radius=0.04]node[anchor=west] {\small{$1-p_i$}};
\draw (2, 0) node[anchor= north] {\small{$w_i$}}
(-0.25, 0) node[anchor = north] {\small{$-\infty$}};
\end{tikzpicture}
}
}
\spacecheat{\vspace{-0.5em}}
\caption{Agent value distribution in the $(w_i,~ p_i)$ type model.
\label{fig:pmf_vipi}} \spacecheat{\vspace{-1em}}
\end{figure}
\subsection{Contingent Payment Mechanisms}
At period 0, each agent makes a report $r_i$ from some set of messages $\mathcal{R}$. Let $r = (r_1, \dots, r_n)\in \mathcal{R}^n$ denote a report profile.
A mechanism is defined by $\mathcal{M} = (\mathcal{R}, x, t, t^{(0)}, t^{(1)})$.
Based on the reports, an {\em allocation rule} $x = (x_1(r), \dots, x_n(r)): \mathcal{R}^n \mapsto \{0, 1\}^n$ allocates the right to use the resource to at most one agent, which we denote as $i^\ast = i^\ast(r)$ s.t. $x_{i^\ast}(r) = 1$. Each agent is charged $t_i(r)$ in period 0.
The allocated agent $i^\ast$ is also charged $t_{i^\ast}^{(0)}(r)$ or $t_{i^\ast}^{(1)}(r)$ at the end of period 1, depending on her action of not using or using the resource, respectively.
The timeline of a contingent payment mechanism is as follows:
\noindent {\em Period 0}:
\begin{enumerate}[$\bullet$]
\item Each agent $i$ reports $r_i$ to the mechanism based on knowledge of type $F_i$.
\item The mechanism allocates the resource to agent $i^\ast$.
\item The mechanism collects $t_i(r)$ from each agent, and determines the contingent payments $t_{i^\ast}^{(0)}(r)$, $t_{i^\ast}^{(1)}(r)$ for the allocated agent.
\end{enumerate}
\noindent {\em Period 1}:
\begin{enumerate}[$\bullet$]
\item The allocated agent privately observes the realized value $v_{i^\ast}$ of $V_i^\ast$.
\item The allocated agent decides which action to take based on $v_{i^\ast}$ and $t_{i^\ast}^{(0)}(r)$, $t_{i^\ast}^{(1)}(r)$.
\item The mechanism collects a contingent payment from $i^\ast$ based on her action.
\end{enumerate}
We assume that agents are risk-neutral, expected-utility maximizers with quasi-linear utility functions. An unallocated agent's utility is $u_{i}(r) = -t_i(r)$. For the allocated agent $i^\ast$, the utility for using the resource at period one is $v_{i^\ast} - t_{i^\ast}^{(1)}(r)$ and the utility for not using the resource is $-t_{i^\ast}^{(0)}(r)$. The rational decision at period 1 is to use the resource if and only if $v_{i^\ast} - t_{i^\ast}^{(1)}(r) \geq -t_{i^\ast}^{(0)}(r)$ (breaking ties in favor of using the resource.) Thus, the expected utility to the allocated agent is
\begin{align}
u_{i^\ast}(r) = &
\E{ (V_{i^\ast} - t_{i^\ast}^{(1)}(r)) \one{V_{i^\ast} \geq (t_{i^\ast}^{(1)} (r)-t_{i^\ast}^{(0)}(r))}} - t_{i^\ast}^{(0)}(r) \Pm{V_{i^\ast} < (t_{i^\ast}^{(1)}(r) - t_{i^\ast}^{(0)}(r))} - t_{i^\ast}(r), \label{equ:exp_util_report}
\end{align}
where $\one{\cdot}$ is the indicator function.
Let $r_{-i} \triangleq (r_1, \dots, r_{i-1}, r_{i+1}, \dots, r_n)$
\begin{definition}[Dominant strategy equilibrium (DSE)] A mechanism has a {\em dominant strategy equilibrium} if for each agent $i$, all value distribution $F_i$ satisfying (A1)-(A3), there exists a report $r^\ast_i \in \mathcal{R}$ such that $u_{i}(r^\ast_i, ~ r_{-i}) \geq u_{i}(r_i, ~ r_{-i})$, $\forall r_{i} \in \mathcal{R}$, $\forall r_{-i} \in \mathcal{R}^{n-1}$.
\if 0
\begin{displaymath}
u_{i}(r^\ast_i, ~ r_{-i})\geq u_{i}(r_i, ~ r_{-i}), \quad \forall r_{i} \in \mathcal{R},~\forall r_{-i} \in \mathcal{R}^{n-1}.
\end{displaymath}
\fi
\end{definition}
A direct mechanism, for which the message space is the type space, is {\em dominant strategy incentive compatible} (DSIC) if truthful reporting of one's type is a DSE.
Let $r^\ast = (r^\ast_1, \dots, r^\ast_n)$ denote the report profile in a DSE under mechanism $\mathcal{M}$.
\begin{definition}[Individual rationality (IR)] A mechanism is {\em individually rational} if for every agent $i$, all value distribution $F_i$ satisfying (A1)-(A3), $u_i(r^\ast_i, r_{-i})\geq 0$, $\forall r_{-i}\in \mathcal{R}^{n-1}$,
\if 0
\begin{displaymath}
u_i(r^\ast_i, r_{-i})\geq 0, \quad \forall r_{-i}\in \mathcal{R}^{n-1},
\end{displaymath}
\fi
\end{definition}
In words, IR (or voluntary participation)
states that an agent's expected utility is non-negative under her dominant strategy given that she makes rational decisions in period 1 (if allocated), regardless of the reports made by the rest of the agents.
We cannot charge unallocated agents without violating IR, thus $t_i(r) \leq 0$ for $i\neq i^\ast$.
\begin{definition}[No deficit (ND)] A mechanism satisfies {\em no deficit} if, for all possible agent types $\{F_i\}_{i \in N}$ satisfying (A1)-(A3) and all report profile $r \in \mathcal{R}^n$, the expected revenue is non-negative, assuming rational decisions of agents in period 1:
\begin{displaymath}
\sum_{i\in N} t_i(r) + t_{i^\ast}^{(0)}(r) \cdot \Pm{V_{i^\ast} < (t_{i}^{(1)}(r)-t_{i}^{(0)}(r))} + t_{i^\ast}^{(1)}(r) \cdot \Pm{V_{i^\ast} \geq (t_{i}^{(1)}(r)-t_{i}^{(0)}(r))} \geq 0.
\end{displaymath}
\end{definition}
A mechanism is \emph{anonymous} if the outcome (assignment, payments) is invariant to permuting the identities of agents. \emph{Deterministic} insists that the outcome is not randomized unless there is a tie. \emph{Always-allocate} requires that the resource is always allocated as long as there is at least one agent, thus rules out the use of reserve prices. \emph{No-charge} precludes an allocated agent from making a payment to the mechanism if the resource is utilized: $t_{i^\ast} + t_{i^\ast}^{(1)} \leq 0$.
The objective is to design a mechanisms with desirable properties that
maximizes in dominant strategy the \emph{utilization} of the resource, i.e. the probability
with which the allocated agent rationally decides to use the resource:
\begin{displaymath}
ut_\mathcal{M}(F) \triangleq \Pm{V_{i^\ast} \geq (t_{i^\ast}^{(1)}(r^\ast)-t_{i^\ast}^{(0)}(r^\ast))}.
\end{displaymath}
\section{The CSP Mechanism} \label{sec:CSP}
We show in this section that under (A1)-(A3),
the CSP mechanism achieves higher utilization
than the SP auction. The uniqueness and optimality of CSP are
discussed in Section~\ref{sec:optimality}.
\begin{definition}[Contingent Second-Price Mechanism] The {\em CSP mechanism} collects the bids $b = (b_1, \dots, b_n)$ from the agents.
\begin{enumerate}[$\bullet$]
\item Allocation rule: $x_{i^\ast}(b) = 1$ for $ i^\ast \in \arg \max_{i\in N} b_i$,
breaking ties at random.
\item Payment rule: $t_{i^\ast}^{(0)} = \arg \max_{i\neq i^\ast} b_i$, $t_{i^\ast}^{(1)}(b) = 0$ and $t_i(b) = 0$ for all $i \neq i^\ast$
\end{enumerate}
\end{definition}
In words, the CSP mechanism collects a single bid $b_i$ from each
agent, allocates the resource to the highest bidder in period 0, and
charges the second highest bid as no-show penalty if the allocated
agent fails to use the resource in period 1.
\subsection{Dominant Strategy Equilibrium under CSP}
For simplicity, we consider a particular agent and
simplify notation, so that $V, ~F, ~t, ~t^{(0)}$, and $t^{(1)}$ denote
the random value, type, period 0 payment and contingent payments for
this agent. We refactor the payments, and let the ``two-part payment"
$(z,y)$ be
\begin{displaymath}
y \triangleq t + t^{(1)}, \text{ and } z \triangleq t^{(0)} - t^{(1)}.
\end{displaymath}
Here, $y$ is the ``base payment" that an agent makes if she is allocated and chooses to use the resource in period 1, and $z$ is the additional ``penalty" that she needs to make if she does not use the resource. The base payment $y$ is always zero under the CSP mechanism.
After learning the realized value $v$ in period 1, an agent gets utility $v - t - t^{(1)} = v - y$ if she uses the resource, and $- t - t^{(0)} = -z - y$ otherwise. A rational agent would decide to use the resource if and only if $v \geq -z$.
The base payment $y$ does not affect an agent's decision in period 1 on whether to use the resource, and the utilization is $\Pm{V \geq -z}$.
The expected utility~\eqref{equ:exp_util_report} for being allocated the right to use the resource as a function of the two-part payment $(z,y)$ can be rewritten as:
\begin{align}
u(z, y) &= \E{V \cdot \one{V \geq -z}} - z \cdot \Pm{V < -z} - y. \label{equ:exp_util_zy}
\end{align}
To economize on notations, we drop the second variable when $y=0$ and write
\begin{align}
u(z) \triangleq & ~u(z,0) = \E{V \cdot \one{V \geq -z}} - z \cdot \Pm{V < -z}. \label{equ:exp_util_z} \\
=& ~\E{V^+}+ \E{V \cdot \one{-z \leq V < 0}} - z \cdot \Pm{V < -z}. \label{equ:exp_util_z_parts}
\end{align}
It helps to understand the three parts of this expression: $\E{V^+}$ is the highest possible utility that an agent derives from the resource, by utilizing whenever $v \geq 0$ and paying 0 for no-show; $\E{V \cdot \one{-z \leq V \hspace{-0.1em} \leq 0}}$ is the expected loss of utility when the realized value is negative and the resource is utilized to avoid the payment; and $z \cdot \Pm{V < -z}$ is the expected amount to pay for not utilizing the resource when the realized value is very low.
\begin{figure}
\centering
\spacecheat{\vspace{-1.5em}}
\subfloat[\small{$\E{V} > 0$}]{\label{fig:utilieis_posEv}
\begin{tikzpicture}[scale = 1.1][font=\small]
\draw[->] (-0.1,0) -- (2.5,0) node[anchor=north] {$z$};
\draw[->] (0,-0.2) -- (0,0.8) node[anchor=west] {$u(z)$};
\draw[-] (0, 0.6) parabola[bend at end] (1.8, 0.2) -- (2.5, 0.2);
\draw[loosely dotted] (0,0.2) -- (2, 0.2);
\draw (0, 0.6) node[anchor=east] {\scriptsize{$\mathbb{E}[V^+]$}}
(0, 0.2) node[anchor=east] {\scriptsize{$\mathbb{E}[V]$}};
\end{tikzpicture}
}
~
\subfloat[$\E{V} < 0$]{\label{fig:utilities_negEv}
\begin{tikzpicture}[scale = 1.1][font=\small]
\draw[->] (-0.1,0) -- (2.5,0) node[anchor=north] {$z$};
\draw[->] (0,-0.2) -- (0,0.8) node[anchor=west] {$u(z)$};
\draw[-] (0, 0.6) parabola[bend at end] (2, -0.2) -- (2.3, -0.2);
\draw[loosely dotted] (0, -0.2) -- (2, -0.2);
\draw (0, 0.6) node[anchor=east] {\scriptsize{$\mathbb{E}[V^+]$}}
(0, -0.2) node[anchor=east] {\scriptsize{$\mathbb{E}[V]$}};
\draw (1.2,0) node[anchor=south] {$z^0$};
\end{tikzpicture}
}
\spacecheat{\vspace{-0.5em}}
\caption{Expected utility for being allocated the resource
as a function of penalty $z$.
\label{fig:utilities}} \spacecheat{\vspace{-0.5em}}
\end{figure}
\begin{restatable
{lemma}{lemmaexputility} \label{lem:exp_u}
Under (A2), the expected utility $u(z)$ as a function of the penalty $z$ satisfies:
\begin{enumerate}[(i)]
\setlength\itemsep{0em}
\item $u(z) = \E{V^+} - \int_0^{z} F(- v) dv$,
\item $u(0) = \E{V^+}$, $\lim_{z \rightarrow \infty} u(z) = \E{V}$, and
\item $u(z)$ is continuous and monotonically decreasing w.r.t. $z$.
\end{enumerate}
\end{restatable}
See Figure~\ref{fig:utilities}. Part (i) of Lemma~\ref{lem:exp_u} can be derived by applying integration by parts to \eqref{equ:exp_util_z}, and the proof of the rest of the lemma is straightforward. We defer the formal proofs to Appendix \ref{appx:proof_lem_exp_u}. Intuitively, the agent gets the expected positive value $\E{V^+}$ when $z =0$. As the penalty $z$ increases to infinity, the agent would end up always using the resource and never pays the penalty thus her expected utility converges to $\E{V}$.
\begin{theorem}[Dominant Strategy in CSP
\label{thm:dominant_strategy} Given assumptions~(A1)-(A3), it is a dominant strategy to bid $b^\ast_{\mathrm{CSP}} = z^0$, the unique zero-crossing of $u(z)$, in the CSP mechanism.
\end{theorem}
\begin{proof}
From parts~(ii) and~(iii) of Lemma~\ref{lem:exp_u}, we know that when $\E{V} < 0$, there is a unique zero crossing $z^0$ of $u(z)$ s.t. $u(z^0) = 0$, as shown in Figure~\ref{fig:utilities_negEv}. In this case, an agent gets non-negative expected utility iff the penalty $z$ is at most $z^0$ thus $z^0$ corresponds to the agent's maximum acceptable penalty for no-show. It is then standard for a second price mechanism, noting that the bid sets the maximum penalty the agent will face, that bidding $z^0$ is a dominant strategy.
\end{proof}
Without~(A3), we have $u(z) > 0, \; \forall z \in {\mathbb{R}}$, as shown in Figure~\ref{fig:utilieis_posEv}. This implies that the agent gains in expectation for any penalty $z$ and would always accept such a contract, which we find unreasonable (``pay \$1B if you don't show up for the spinning class.'') In this case, there is always an incentive to bid higher, and thus no
DSE exists.
\subsection{Better Utilization than Second Price Auction}
The following lemma relates the slope of an agent's utility functions with utilization.
\begin{restatable
{lemma}{lemmaexputilityutilization} \label{lem:exp_u_util}
While facing a two part payment $(z,y)$, the utilization, i.e. an agent's probability of using the resource, corresponds to the derivatives of her expected utility functions w.r.t $z$:
\begin{displaymath}
\Pm{V \geq -z } = 1 - F(-z)= 1 + \frac{d}{dz}u(z) = 1 + \frac{\partial}{\partial z}u(z,y),
\end{displaymath}
and is monotonically increasing in $z$. Moreover, $u(z)$ is convex in $z$.
\end{restatable}
The proof is straightforward given part (i) of Lemma~\ref{lem:exp_u}, the fact that $u(z,y) = u(z) -y$, and that $\frac{d}{dz}u(z) = \Pm{V \geq -z }-1$ increases in $z$. Intuitively, the agent is more likely to use the resource when the penalty is larger. Moreover, the higher the probability that the resource is to be used at penalty $z$, the less likely the agent is paying the penalty, thus $u(z)$ decreases slower as $z$ increases, which corresponds to the shallower the slope at $z$.
We now prove that in DSE, the CSP mechanism always achieves higher utilization than the SP auction, which assigns the resource to the
highest bidder and charges the second highest bid.
\begin{figure}[t!]
\centering
\spacecheat{\vspace{-0.5em}}
\begin{tikzpicture}[scale = 0.85][font=\small]
\draw[->] (-0.15,0) -- (6,0) node[anchor=north] {$z$};
\draw[->] (0,-0.3) -- (0, 1.8) node[anchor=west] {$u_i(z)$};
\draw[-] (0, 1.5) parabola[bend at end] (4,-0.6);
\draw[dashed] (0,1) parabola[bend at end] (6, -0.35);
\draw[dashdotted] (0, 0.2) -- (1.5, -0.05);
\draw
(1.9,0) node[anchor=north] {$z^0_1$}
(3.1,-0.05) node[anchor=south] {$z^0_2$}
(1.2,0) node[anchor=north] {$z^0_3$};
\draw (0, 1.6) node[anchor=east] {\small{$\E{V_1^+}$}}
(0, 0.95) node[anchor=east] {\small{$\E{V_2^+}$}}
(0, 0.2) node[anchor=east] {\small{$\E{V_3^+}$}};
\draw (4.5, 1.5) -- (5, 1.5) node[anchor=west] {$u_1(z)$};
\draw[dashed] (4.5, 1) -- (5, 1) node[anchor=west] {$u_2(z)$};
\draw[dashdotted] (4.5, 0.5) -- (5, 0.5) node[anchor=west] {$u_3(z)$};
\draw[dotted] (1.85,0) -- (1.85, 0.6);
\end{tikzpicture}
\spacecheat{\vspace{-0.8em}}
\caption{Example economy where the SP winner (agent 1) is different from the CSP winner (agent~2.)
\label{fig:switched_bids}} \spacecheat{\vspace{-0.5em}}
\end{figure}
\begin{restatable}{theorem}{thmCSPbeatSP} \label{thm:CSP_beat_SP}
For any set of agent types satisfying (A1)-(A3), the CSP mechanism has higher utilization than the SP auction under the dominant strategy equilibria.
\end{restatable}
\begin{proof}[Proof Sketch]
The full proof is provided in Appendix~\ref{appx:proof_thm_CSP_beat_SP}. Here we provide the intuition.
First, observe that the assigned agent in SP uses the resource as long as $V_{i^\ast} \geq 0$, achieving utilization $\Pm{V_{i^\ast} \geq 0}$. It is a dominant strategy for each agent to bid the her expected utility from using the resource: $b^\ast_{i,{\mathrm{SP}}} = \E{V_i \cdot \one{V_i \geq 0}} = \E{V^+_i}$. Consider the following two cases:
\textit{Case 1:} SP and CSP allocate the resource to the same agent (say agent~1), the utilization under CSP $\Pm{V_1 \geq -z}$ is always (weakly) higher than that under SP $\geq \Pm{V_1 \geq 0}$, for any penalty $z \geq 0$.
\textit{Case 2:} SP and CSP allocate the resource to agent 1 and 2 respectively. Ignoring ties, we must have $\E{V_1^+} > \E{V_2^+}$, and $z^0_1 < z^0_2$, as shown in Figure~\ref{fig:switched_bids}. This implies $u_1(0) - u_1(z^0_1) > u_2(0) - u_2(z^0_1)$. In order for this to happen, the slope of $u_2(z)$ at any penalty level $z \in [z^0_1,~ z^0_2]$ (where the CSP payment resides) must be shallower than that of $u_1(z)$ at $z=0$, due to the convexity of the expected utility functions. This translates to CSP having strictly higher utilization than SP given Lemma~\ref{lem:exp_u_util}.
\if 0
Formally, denote the probability for agent $i$ to use the resource given penalty $z$ as $ut_i(z) \triangleq \Pm{V_i \geq -z}$. Lemma~\ref{lem:exp_u_util} implies $ \frac{d}{dz} u_i(z) = ut_i(z) - 1$. $u_1(0) - u_1(z^0_1) > u_2(0) - u_2(z^0_1)$ implies $ \int_0^{z^0_1} (ut_1(v) - 1) dv < \int_0^{z^0_1} (ut_2(v) - 1) dv
\Rightarrow \int_0^{z^0_1} ut_1(v) dv < \int_0^{z^0_1} ut_2(v) dv$. Since both $ut_1(z)$ and $ut_2(z)$ are increasing in $z$, we have $z^0_1 \cdot ut_1(0) \leq \int_0^{z^0_1} ut_1(v) dv < \int_0^{z^0_1} ut_2(v) dv \leq z^0_1 \cdot ut_2(z^0_1)$, therefore $ut_1(0) < ut_2(z^0_1) \leq ut_2(z)$ for any $z \in [z^0_1, z^0_2]$.
\fi
\end{proof}
The following examples illustrate the improvement in utilization from CSP over SP, and show that SP can be arbitrarily worse than CSP.
\begin{example} [Double gain in CSP] \label{ex:3_val} Consider two agents with value distributions and expected utilities as shown in Figure~\ref{fig:example_utilities}. Compared with Agent~2, Agent~1 has higher value for the facility, but lower probability of willing to use the resource and higher probability for a hard constraint.
Under SP, $b_{{\mathrm{SP}}, 1}^\ast = 20, b_{{\mathrm{SP}}, 2}^\ast = 16$ thus Agent~1 is allocated and the utilization is $\Pm{V_1 \geq 0} = 0.2$. Whereas under CSP $b_{{\mathrm{CSP}}, 1}^\ast = z^0_1 = 30$ and $b_{{\mathrm{CSP}}, 2}^\ast = z^0_2 = 60$. Agent~2 is allocated and charged penalty $b_{{\mathrm{CSP}}, 1}^\ast = 30$ for no-show, thus the utilization is $\Pm{V_2 \geq -z^0_1} = 0.8$. This is higher than $\Pm{V_2 \geq 0} = 0.4$, the probability with which the resource is utilized, if Agent~2 is allocated the resource under SP in some other economy.
\begin{figure}[t!]
\centering{
\spacecheat{\vspace{-1em}}
\subfloat{
\small{
\begin{tikzpicture}[scale = 0.9][font=\small]
\draw (0,1.4) node {$V_1 = \pwfun{100, & \text{ w.p. } 0.2\\ -20, & \text{ w.p. } 0.4 \\ -\infty, &\text{ w.p. } 0.4},$};
\draw (0,0) node {$V_2 = \pwfun{40, & \text{ w.p. } 0.4\\ -10, & \text{ w.p. } 0.4 \\ -\infty, &\text{ w.p. } 0.2}.$};
\draw (0,-0.6) node{{\color{white} some text}};
\end{tikzpicture}
}}
\hspace{2em}
\subfloat{
\begin{tikzpicture}[scale = 0.85][font=\small]
\draw[->] (-0.2,0) -- (7,0) node[anchor=north] {$z$};
\draw[->] (0,-0.3) -- (0,2.3) node[anchor=west] {$u_i(z)$};
\draw[-] (0, 2) -- (2, 0.4) -- (4, -0.4);
\draw[dashed] (0, 1.6) -- (1, 1) -- (7, -0.2);
\draw
(2.9, -0.05) node[anchor=north] {$z^0_1 = 30$}
(6, -0) node[anchor=north] {$z^0_2 = 60$}
(1, 0) node[anchor = north] {10}
(1.9, 0) node[anchor = north] {20}
(0, 1) node[anchor = east] {10}
(0, 0.4) node[anchor = east] {4};
\draw (0, 2.0) node[anchor=east] {$20$}
(0, 1.6) node[anchor=east] {$16$};
\draw (5.5, 2) -- (6,2) node[anchor=west] {$u_1(z)$};
\draw[dashed] (5.5, 1.5) -- (6, 1.5) node[anchor=west] {$u_2(z)$};
\draw[dotted] (2,0) -- (2, 0.4) -- (0, 0.4)
(0, 1) -- (1,1) -- (1, 0);
\end{tikzpicture}}
}
\spacecheat{\vspace{-0.5em}}
\caption{Agents' value distributions and expected utilities in Example \ref{ex:3_val}.
\label{fig:example_utilities}} \spacecheat{\vspace{-0.5em}}
\end{figure}
\end{example}
\begin{example}[SP arbitrarily worse] \label{ex:SP_worse}
Under the $(w_i,p_i)$ model introduced in Example~\ref{ex:vipi}, the expected utility for agent $i$ given penalty $z$ is $u_i(z) = w_ip_i - (1-p_i)z$. It is a dominant strategy to bid $b_{i, {\mathrm{SP}}}^\ast = \E{V_i^+} = w_i p_i$ under SP and the zero-crossing $b_{i,{\mathrm{CSP}}}^\ast = z^0_i = w_ip_i/(1-p_i)$ under CSP.
Consider an economy where there are two agents with types: $p_1=\varepsilon$, $w_{1} = 1/\varepsilon$; and $p_2 = 1-\varepsilon$, $w_{2} = 1$ for some $\varepsilon>0$ very small. We can verify that agent 1 is allocated under SP since $b_{1, {\mathrm{SP}}}^\ast = 1 > b_{2,{\mathrm{SP}}}^\ast = 1-\varepsilon$ whereas agent 2 is allocated under CSP since $b_{2,{\mathrm{CSP}}}^\ast = (1-\varepsilon)/\varepsilon > 1/(1-\varepsilon) = b_{1,{\mathrm{CSP}}}^\ast$. The utilization under SP and CSP is $\varepsilon$ and $1-\varepsilon$, respectively; thus, CSP can have arbitrarily better utilization than SP by selecting a better winner.
\end{example}
We see that CSP's higher utilization comes from two
different aspects of its design:
\begin{enumerate}[$\bullet$]
\item Charging a penalty $z>0$ changes the period 1 decision of the allocated agent, promoting the resource to be used with higher probability.
\item Agents with higher probability of showing up have utility functions that decrease more slowly with penalty, thus have relatively higher zero-crossing points and bid higher in CS
\end{enumerate}
One might wonder about a comparison between slightly more general mechanisms.
Consider, for example, introducing a reserve price $R>0$ into the CSP mechanism, where the resource is only allocated when $b_{i^\ast} \geq R$ for $i^\ast \in \arg \max_{i\in N} b_i$, and the allocated agent is charged $t_{i^\ast}^{(0)} = \max(\max_{i \neq i^\ast} b_i, R)$ as penalty. We call this the \emph{CSP+R mechanism}.
Under (A1)-(A3), it remains a dominant strategy for agents to bid the zero-crossings $b_{i,\textrm{CSP+R}}^\ast = z^0_i$ under CSP+R.
Similarly, the SP auction can be modified to charge an additional fixed penalty $C>0$ (termed the \emph{SP+C mechanism}), which introduces a penalty to the assigned agent
in the event that the resource goes wasted. We can show that when the reserve penalty $R$ is set to be equal to the fixed penalty $C$, utilization under CSP+R dominates that of the SP+C mechanism profile-by-profile (see Theorem~\ref{thm:CSPC_Utilization} in Appendix~\ref{appx:proof_thm_CSPC_Utilization}.)
We discuss the effect of reserve prices in more detail in Sections~\ref{sec:optimality} and~\ref{sec:simulations}.
\section{Characterization and Optimality of CSP} \label{sec:optimality}
In this section, we
consider the full space of mechanisms, including those
with two-part payments that use both
the base payment $y$
and a no show penalty $z$.
We consider the optimal mechanism design problem under (some subset)
of the following properties:
\spacecheat{\vspace{-0.5em}}
\begin{enumerate}[{P}1.]
\begin{multicols}{2}
\setlength\itemsep{0em}
\item Dominant-strategy equilibrium
\item Individually rational
\item No deficit
\item Anonymous
\item Deterministic (unless breaking ties)
\item Always-allocates the resource
\item No charge
\end{multicols}
\end{enumerate}
\spacecheat{\vspace{-0.5em}}
We prove that when the type space is the set of all value
distributions satisfying (A1)-(A3), the CSP mechanism is unique
amongst mechanisms with properties (P1)-(P7).
For this, we make use of a key lemma (Lemma~\ref{lem:lem_P1P5_characterization}), which
characterizes the set of all possible outcomes
under mechanisms that satisfy (P1)-(P5).
Combined with (P6) and
(P7), we can show that CSP is unique.
We also prove that the utilization of the CSP mechanism is not dominated by any
mechanism in the broader class of mechanisms
that satisfy properties (P1)-(P5), and that
the CSP mechanism is optimal
for agents with the
simple $(w_i, p_i)$ types from Example \ref{ex:vipi}.
Moreover, we prove that the CSP mechanism maximizes
utilization among a broad subclass of mechanisms satisfying (P1)-(P6),
and that relaxing (P6) by adding reserve prices rarely improves
utilization.
\subsection{Utilities in The Two-Dimensional Payment Space}
For the formal analysis of uniqueness and optimality,
we appeal to the standard
revelation principle from mechanism design, and
consider the space of direct-revelation mechanisms.
\paragraph{Zero-profit Curves} Recall that in the CSP mechanism the base payment $y = t + t^{(1)}$ is always zero. To study general mechanisms, we move back to the full $(z,y)$ payment space, and work with \emph{iso-profit curves} which are sets of $(z,y)$ pairs for which $u(z,y) = c$ for some constant $c$, i.e. an agent will be indifferent to all payments $(z,y)$ that reside on the same iso-profit curve. See Figure~\ref{fig:indiffCurves}.
\begin{figure}
\spacecheat{\vspace{-1em}}
\centering
\subfloat[Iso-profit curves of an agent]{\label{fig:indiffCurves}
\begin{tikzpicture}[scale = 1.1][font = \footnotesize]
\draw[->] (-0.1,0) -- (4.4,0) node[anchor=north] {$z$};
\draw[->] (0,-0.2) -- (0,2.2) node[anchor=east] {$y$};
\draw[-, name path = uA] (-0.1, 1.2) parabola[bend at end] (4,-0.2);
\draw[dashdotted] (-0.1, 2) parabola[bend at end] (4, 0.5);
\draw[dashed] (-0.1, 0.4) to [out = -40, in = 160] (0.7, -0.05);
\draw
(2.3, 0) node[anchor=north] {{ $b_{{\mathrm{CSP}}}^\ast = z^0$}}
(0, 1.4) node[anchor=east] {$b_{{\mathrm{SP}}}^\ast$};
\draw[dotted] (1.0,-0.1) -- (1.0, 0.9);
\draw (1.0, 0) node[anchor=north]{{\scriptsize $C$}};
\draw[dotted] (-0.1, 0.55) -- (1.45, 0.55);
\draw (0, 0.65) node[anchor = east]{$b_{{\mathrm{SP+C}}}^\ast$};
\draw[densely dotted, name path = gcsp](-0.1, -0.1)--(1.2, 1.7);
\path [name intersections={of=uA and gcsp, by=intgcap}];
\draw (0.6, 1.9) node[anchor = west] {$y = \frac{\gamma}{1-\gamma}z$};
\filldraw [black] (intgcap) circle (1pt);
\draw (intgcap) node[anchor = south]{$A$};
\draw[dashdotted] (2.3, 2) -- (2.7, 2) node[anchor=west] {{\scriptsize $u(z,y) = -10$}};
\draw (2.3, 1.55) -- (2.7, 1.55) node[anchor=west] { {\scriptsize $u(z,y) = 0$}};
\draw [dashed](2.3, 1.1) -- (2.7, 1.1) node[anchor=west] {{\scriptsize $u(z,y) = 10$}};
\end{tikzpicture}
}
\hspace{2em}
\subfloat[Examples of IR and ND ranges]{\label{fig:irbb}
\begin{tikzpicture}[scale = 1.2][font=\footnotesize]
\fill [pattern =north west lines, pattern color = black!40, opacity=0.5](-3, 0.68)--(-0.8, -0.2)--(-0.8, -0.6)--(-3, -0.6)--(-3, 0);
\fill [pattern=crosshatch dots, pattern color = black!60, opacity=0.5] (-3, 0.08)--(-1.3, -0.6)--(-0.8, -0.6)--(-0.8, 0.8)--(-3, 0.8) -- (-3, 0.08);
\draw[->] (-3, 0) -- (-0.6,0) node[anchor=north] {$z$};
\draw[->] (-2.8,-0.8) -- (-2.8,1.1) node[anchor=west] {$y$};
\draw[-] (-3, 0.68) -- (-0.8, -0.2);
\draw[dashed] (-3, 0.08) -- (-1.3, -0.6);
\draw (-1.2, 0) node[anchor=south] {$z^0$};
\draw (-1.8, -1.4) node[anchor=south] {(i) $(w_i,p_i)$};
\fill [pattern =north west lines, pattern color = black!40, opacity=0.5](-0.1, 0.7) parabola[bend at end] (2, -0.4)--(2, -0.6)--(-0.1, -0.6)--(-0.1, 0);
\fill [pattern=crosshatch dots, pattern color = black!60, opacity=0.5](-0.1, 0.1) parabola[bend at end] (0.6, -0.4) to [out = 0, in=-170] (2, -0.1)--(2, 0.8)--(-0.1, 0.8) -- (-0.1, 0.1);
\draw[->] (-0.1,0) -- (2.2,0) node[anchor=north] {$z$};
\draw[->] (0,-0.8) -- (0,1.1) node[anchor=west] {$y$};
\draw[-] (-0.1, 0.7) parabola[bend at end] (2, -0.4);
\draw[dashed] (-0.1, 0.1) parabola[bend at end] (0.6, -0.4);
\draw[dashed] (0.6, -0.4) to [out = 0, in=-170] (2, -0.1);
\draw (0.85,0) node[anchor=south] {$z^0$};
\filldraw [black] (1.32, -0.28) circle (1pt);
\draw (1.32, -0.28) node[anchor = north]{{\scriptsize $B$}};
\draw (1, -1.4) node[anchor=south] {(ii) Exponential};
\draw (2.5, 0.6) -- (2.8, 0.6) node[anchor=west] { {\scriptsize $u(z,y) = 0$}};
\draw [dashed](2.5, 0.2) -- (2.8, 0.2) node[anchor=west] { {\scriptsize $rev(z,y) = 0$}};
\fill [pattern =north west lines, pattern color = black!40, opacity=0.5](2.5, -0.1) -- (2.8, -0.1) -- (2.8, -0.25) -- (2.5, -0.25);
\draw (2.8,-0.2) node[anchor=west] { {\scriptsize IR range}};
\fill [pattern=crosshatch dots, pattern color = black!60, opacity=0.5](2.5, -0.5) -- (2.8, -0.5) -- (2.8, -0.65) -- (2.5, -0.65);
\draw (2.8, -0.6) node[anchor=west] { {\scriptsize ND range}};
\end{tikzpicture}
}
\spacecheat{\vspace{-0.5em}}
\caption{Illustration of iso-profit curves, IR and ND ranges in the 2-D payment space. \label{fig:iso_irbb}} \spacecheat{\vspace{-0.0em}}
\end{figure}
Because $u(z, y) = u(z) - y$ (see \eqref{equ:exp_util_zy} and \eqref{equ:exp_util_z}), the \emph{zero-profit curve} (i.e. where $c=0$, the solid line depicted in Figure~\ref{fig:indiffCurves}) is characterized by $\{(z,y)\ |\ y = u(z) \}$. This has the same shape as the expected utility function $u(z)$ in Figure~\ref{fig:utilities} and is continuously decreasing and convex according to Lemmas~\ref{lem:exp_u} and \ref{lem:exp_u_util}.
Observing $\frac{\partial}{\partial y}u(z,y) = -1$, we know that other iso-profit curves are vertical shifts of the zero-profit curve, and recall from Lemma~\ref{lem:exp_u_util} that the utilization for an agent facing payments $(z, y)$ directly relates utilization to the slope of the zero-profit curve at point $(z,y)$: $\Pm{V \geq -z} = 1 + \frac{d}{dz}u(z) = 1+\frac{\partial}{\partial z}u(z,y)$.
\paragraph{IR and ND Ranges} If an allocated agent is charged payment $(z,y)$ above her zero-profit curve, her expected utility would be negative since $u(z,y)$ decreases in $z$ and $y$. We call the area weakly below an agent's zero-profit curve the \emph{IR-range} for the agent. For any two-part payment $(z,y)$, the expected revenue (e.g. payment from the agent to the mechanism) is $rev(z, y) = y + z \cdot \Pm{V < -z}$.
\if 0
\begin{align*}
rev(z, y) = y + z \cdot \Pm{V < -z}.
\end{align*}
\fi
We call the set of payments $(z, y)$ for which $rev(z, y)$ is non-negative as the \emph{ND-range} (no-deficit range) for this agent. The ND range depends an agent's own type, meaning that fixing two part payment $(z,y)$, whether the mechanism runs into deficit depends on the type of the assigned agent.
\begin{example} \label{ex:irbb}
Consider the $(w_i, p_i)$ model (Example~\ref{ex:vipi}.) The zero-profit curve is characterized by $\{ (y,z) \in {\mathbb{R}}^2 | y = u(z) = w_i p_i - (1-p_i)z \}$. The expected revenue that the mechanism collects is $rev(z,y) = y + (1-p_i)z$, thus the ND range is lower-bounded by $y = -(1-p_i)z$. See Figure~\ref{fig:irbb}(i).
\end{example}
\begin{example}[Exponential model]
\label{ex:exp_model}
Under the exponential model, the utility for agent $i$ to use the resource is a fixed value $w_i > 0$ minus a random opportunity cost, which is exponentially distributed with parameter $\lambda_i>0$, as illustrated in Figure~\ref{fig:pdf_exp}. The expected value $\E{V_i} = w_i - 1/\lambda_i$ where $1/\lambda_i$ is the expected value of the opportunity cost. With $w_i < 1/\lambda_i$, assumptions (A1)-(A3) are satisfied.
For an agent with an exponential value distribution, the zero-profit curve, IR and ND ranges are as shown in Figure~\ref{fig:irbb}(ii) (see derivations in Appendix~\ref{appx:bids}.)
\begin{figure}[t!]
\spacecheat{\vspace{-1em}}
\centering
{
\subfloat{
\small{
\begin{tikzpicture}[scale = 0.9][font=\normalsize]
\draw (0,0) node {$ f_i(v) = \pwfun{ \lambda_i e^{\lambda_i (v - w_i)}, & v \leq w_i \\
0, & v > w_i}$};
\draw (0,-0.6) node{{\color{white} some text}};
\end{tikzpicture}
}}
\hspace{2em}
\subfloat{
\begin{tikzpicture}[scale = 1.3][font=\small]
\draw[->] (-1.8,0.3) -- (2.2,0.3) node[anchor=north] {$v$};
\draw[->] (0,0.1) -- (0,1.1);
\draw (0,1) node[anchor=west] {$f_i(v)$};
\draw[scale=1,domain=-1.8:1.5,smooth,variable=x] plot (x,{0.5*exp(0.8*(x-1.5))+0.3});
\draw[-] (1.5, 0.3) -- (1.5,0.8);
\draw (1.5, 0.3) node[anchor=north] {$w_i$};
\draw[dotted] (0, 0.8) -- (1.5, 0.8);
\draw (0, 0.8) node[anchor=east] {$\lambda_i$};
\end{tikzpicture}
}
}
\spacecheat{\vspace{-0.7em}}
\caption{Agent value distribution n the exponential type model.
\label{fig:pdf_exp}} \spacecheat{\vspace{-0.5em}}
\end{figure}
\end{example}
\paragraph{The First Best}
Under any IR and ND mechanism, the payment facing the assigned agent must reside in the intersection of the IR and ND ranges of the assigned agent. Given an agent's type, we can compute the highest utilization induced by any two-part payment in this intersection, which we call the \emph{first-best} utilization of this agent.
For the exponential type, the first-best is achieved by charging the agent $(z_B, y_B)$ at point $B$ in Figure~\ref{fig:irbb}(ii), which is the the two-part payment with the highest penalty component in the IR and ND range. For the $(w_i, p_i)$ type, though there is no upper bound on the highest IR and ND penalty, the utilization is always $p_i$, and is not affected by $z$.
\paragraph{DSE Bids}
The mechanisms that we have discussed so far restrict the payments to an ordered, one-dimensional subspace of the $(z,y)$ payment space: CSP considers $\{(z,y)|y=0\}$ where all agents prefer a smaller $z$, and SP restricts $\{(z,y)|z = 0\}$ where all agents prefer a smaller $y$.
The crossing point of the an agent's zero-profit curve and the one-dimensional payment space determines the agent's maximum willingness to pay under each mechanism, thus translates into the agent's DSE bids in these second price mechanisms.
Under CSP, the DSE bid $b_{i,{\mathrm{CSP}}}^\ast = z^0_i$ corresponds to the crossing point of $y=0$ and the agent's zero-profit curve, and under SP, the DSE bids $b_{i,{\mathrm{SP}}}^\ast = \E{V_i^+}$ correspond to the crossing points of the zero-profit curve with $z=0$. See Figure~\ref{fig:indiffCurves}.
To see another example, consider a generalization of CSP
that collects the second highest bid as the no-show penalty
and also charges a $\gamma$-fraction of the second highest bid as the base payment.
\begin{definition}[The $\gamma$-CSP Mechanism] The \emph{$\gamma$-CSP mechanism}, parametrized by $\gamma \in [0, 1]$, collects a single bid from each agent.
\begin{enumerate}[$\bullet$]
\item Allocation rule: $x_{i^\ast}(b) = 1$ for $ i^\ast \in \arg \max_{i\in N} b_i$, breaking ties at random.
\item Payment rule: $t_{i^\ast}^{(0)} = \arg \max_{i\neq i^\ast} b_i$, $t_{i^\ast}^{(1)}(b) = \gamma \cdot t_{i^\ast}^{(0)}$ and $t_i(b) = 0$ for all $i \neq i^\ast$
\end{enumerate}
\end{definition}
Payments reside on $\{ (z,y) | y = \gamma/(1-\gamma)z\}$,
and the SP ($\gamma = 1$) and CSP ($\gamma=0$) are special cases. We prove that it is a dominant strategy to bid the sum of the two coordinates of the crossing point between the zero-profit curve and $y= \frac{\gamma}{1-\gamma}z$ (i.e. $b_i^\ast = z_A + y_A$ as in Figure~\ref{fig:indiffCurves}; see Theorem~\ref{thm:dse_gammaCSP} in Appendix~\ref{appx:proof_thm_dse_gammaCSP}.)
We return to this mechanism below.
\subsection{Uniqueness and Optimality of CSP} \label{sec:uniq_opt}
Define the {\em frontier} of an economy $N$ to be the upper-envelope of the zero-profit curves of all agents, i.e., $\{(z,y)\ |\ y = \bar{u}_N(z) \}$ where $\bar{u}_N(z) \triangleq \max_{i \in N} u_i(z)$. This characterizes the maximum willingness to pay by all agents in $N$. As the upper envelope of a finite set of decreasing convex functions, $\bar{u}_N(z)$ is also decreasing and convex, and has a unique zero-crossing, which we denote as $z^0_N$.
Define the frontier of the sub-economy without agent $i$ as $\bar{u}_{N\backslash \{i\}}(z) \triangleq \max_{j\neq i}u_j(z)$, and the \emph{$m^{\text{th}}$ frontier} of the economy as the $m^{\mathrm{th}}$ upper envelope of $\{u_i(z)\}_{i\in N}$. See Figure~\ref{fig:P1P5_characterization}.
\begin{figure}
\centering
\spacecheat{\vspace{-0.5em}}
\begin{tikzpicture}[scale = 1][font=\small]
\draw[-, name path = uA] (-0.2, 3.3) to[out=-65, in = 150] (2.5, -0.2);
\draw[dashed, name path = uB] (-0.2, 2.2) to[out=-35, in = 175] (6,-0.1);
\draw[dashdotted,name path = uC] (-0.2, 1.2) parabola[bend at end] (5,-0.1);
\draw[dotted] (-0.2, 0.5) parabola[bend at end] (1.5,-0.1);
\path [name intersections={of=uA and uB, by=intAB}];
\path [name intersections={of=uA and uC, by=intAC}];
\draw[blue!80, line width= 1mm, opacity=0.5] (-0.2, 3.3) to[out=-65, in=-240] (intAB);
\draw[blue!80, line width=1mm, opacity=0.5] (intAB) to[out=-32, in = 175] (6,-0.1);
\draw[yellow!80, line width= 1mm, opacity=0.5] (-0.2, 2.2) to[out=-35, in=-213] (intAB);
\draw[yellow!80, line width=1mm, opacity=0.5] (intAB) to[out=-60, in = -220] (intAC) parabola[bend at end] (5,-0.1);
\fill [pattern = north east lines, pattern color = black!40, opacity=0.7] (-0.2, 2.2)--(-0.2, 3.3)--(intAB)--(-0.2, 2.2);
\fill [pattern = north east lines, pattern color = black!40, opacity=0.7] (intAB) to[out=-32, in = 172] (5.2, 0) --(3.55, 0) --(intAC)--(intAB);
\draw[->] (-0.3,0) -- (6.5,0) node[anchor=north] {$z$};
\draw[->] (0,-0.3) -- (0, 3.5) node[anchor=west] {$y$};
\draw (5.3, 0) node[anchor = north]{{\small $z^0_N$}};
\filldraw [black] (5.2, 0) circle (1pt);
\draw (5.3, 0) node[anchor = south]{{\small $A$}};
\draw[loosely dotted](3.55, -0.1)--(3.55, 0.6);
\filldraw [black] (3.55, 0) circle (1pt);
\draw (3.55, 0) node[anchor = north]{{\small $B$}};
\if 0
\draw[loosely dotted](2.7, -0.1)--(2.7, 1);
\filldraw [black] (2.7, 0.17) circle (1pt);
\draw (2.7, 0) node[anchor = north]{{\small $C$}};
\fi
\draw[densely dotted, name path = gcsp](-0.1, -0.07)--(2, 1.5);
\path [name intersections={of=uA and gcsp, by=intgcap}];
\filldraw [black] (intgcap) circle (1pt);
\draw (intgcap) node[anchor = east]{{\small $D$}};
\if 0
\filldraw [black] (intAB) circle (1pt);
\draw (intAB) node[anchor = west]{ {\small $E$}};
\draw (1.7, 0.6) node[anchor = west]{ {\small $G$}};
\fi
\filldraw [black] (0, 2.07) circle (1pt);
\draw (0, 1.9) node[anchor = east]{ {\small $F$}};
\draw [blue!80, line width=1mm, opacity=0.5] (7, 3.4) -- (7.5, 3.4);
\draw (7.5, 3.4) node[anchor=west] {\small{The Frontier}};
\draw [yellow!80, line width=1mm, opacity=0.5] (7, 2.9) -- (7.5, 2.9);
\draw (7.5, 2.9) node[anchor=west] {\small{The 2nd Frontier}};
\fill [pattern = north east lines, pattern color = black!40, opacity=0.5](7, 2.33) -- (7.5, 2.33) -- (7.5, 2.53) -- (7, 2.53);
\draw (7.5, 2.4) node[anchor=west] { {\small Possible Payments}};
\draw [-] (7, 1.9) -- (7.5, 1.9) node[anchor=west] {\footnotesize{$u_1(z,y) = 0$}};
\draw [dashed] (7, 1.4) -- (7.5, 1.4) node[anchor=west] {\small{$u_2(z,y) = 0$}};
\draw [dashdotted] (7, 0.9) -- (7.5, 0.9) node[anchor=west] {\small{$u_3(z,y) = 0$}};
\draw [dotted] (7, 0.4) -- (7.5, 0.4) node[anchor=west] {\small{$u_4(z,y) = 0$}};
\draw [densely dotted] (7, -0.2) -- (7.5, -0.2) node[anchor=west] {\small{$y=\frac{\gamma}{1-\gamma}z$}};
\end{tikzpicture}
\spacecheat{\vspace{-0.9em}}
\caption{Characterization of possible outcomes for mechanisms under (P1)-(P5).
\label{fig:P1P5_characterization}} \spacecheat{\vspace{-0.5em}}
\end{figure}
The following key lemma characterizes possible outcomes for mechanisms satisfying (P1)-(P5).
\begin{restatable
{lemma}{lemCharacterization}\label{lem:lem_P1P5_characterization} Assume the type space is the set of all value distributions satisfying (A1)-(A3). Fix any mechanism that satisfies (P1)-(P5), the allocated agent $i^\ast$ and the two-part payment $(z^\ast, y^\ast)$ the allocated agent $i^\ast$ is charged:
\begin{enumerate}[(i)]
\item $(z^\ast, y^\ast)$ resides weakly below $u_{i^\ast}(z)$;
\item $u_{i^\ast}(z)$ must be a part of the frontier $u_N(z)$ of all agents;
\item $(z^\ast, y^\ast)$ resides weakly above the frontier of the rest of the economy $u_{N \backslash \{i^\ast\}}(z)$;
\item The allocated agent faces a non-negative base payment $y^\ast \geq 0$;
\item The utilization corresponds to the slope of the frontier at $z^\ast$: $ut_\mathcal{M}(F) = \frac{d}{dz}\bar{u}_N(z)|_{z = z^\ast} +1$, which increases in $z^\ast$ and attains its maximum at $z^\ast = z^0_N$.
\end{enumerate}
\end{restatable}
\begin{proof}[Proof Sketch]
We defer the full proof to Appendix~\ref{appx:proof_P1P5} and provide intuitions here. (i) is implied by IR. (ii) and (iii) require that the allocated agent $i^\ast$ to be on the frontier of all agents, and that $(z^\ast, y^\ast)$ resides in between the first and second frontiers, as illustrated in Figure~\ref{fig:P1P5_characterization}.
Otherwise, there exists $i\neq i^\ast$ that strictly prefers the winner's outcome of ``getting allocated and pay $(z^\ast, y^\ast)$" to her outcome ``not getting allocated", which enables us to construct an economy where DSIC is violated.
(iv) implies $i^\ast$ cannot be paid a bonus for showing up. Assume otherwise, that $i^\ast$ with type $u_{i^\ast}$ is allocated and charged payment $(z^\ast, y^\ast)$ with $y^\ast < 0$. We can construct a type $u_{i^\ast}'$ of $(w_i, p_i)$ model s.t. $(z^\ast, y^\ast)$ is inside $u_{i^\ast}'$'s IR range but outside of her ND range. Replace $u_{i^\ast}$ with $u_{i^\ast}'$ in the original economy, we can show that either ND is violated, or the agent with type $u_{i^\ast}'$ has a useful deviation.
For part (v), note that (i)-(iii) imply $u_{i^\ast}(z) = \bar{u}_N(z)$ must hold for a small neighborhood around $z^\ast$, therefore the utilization, which corresponds to the slope of $u_{i^\ast}(z)$ at $z^\ast$ also corresponds to the slope of the frontier: $\Pm{V_{i^\ast} \geq z^\ast} = \frac{d}{dz}u_{i^\ast}(z) + 1|_{z = z^\ast} = \frac{d}{dz}\bar{u}_N(z)|_{z = z^\ast} +1$. Since $u_{N}(z)$ is convex, the utilization $\frac{d}{dz}\bar{u}_N(z)|_{z = z^\ast} +1$ increases $z^\ast$. The highest possible utilization under (P1)-(P5) is achieved when the penalty is $z^\ast = z^0_N$, the highest penalty in the range of possible payments determined by (i)-(iv). In this case, the agent with the highest zero-crossing must be allocated.
\end{proof}
With this characterization, we prove the uniqueness of CSP given (P1)-(P7).
\begin{restatable}{theorem}{thmCSPUniqueness} \label{thm:csp_uniqueness}
Assume the type space is the set of all value distributions satisfying (A1)-(A3). Then CSP is unique: it is the only mechanism with properties (P1)-(P7).
\end{restatable}
We defer the full proof to Appendix~\ref{appx:optimality} due to the space limit. For some intuition, given (P7) and
Lemma~\ref{lem:lem_P1P5_characterization}, the payment facing the
allocated agent must be consisted of a penalty in between the second
and highest zero-crossings, and no base payment. In this case, the
agent with the highest zero-crossing $z^0_i$ (i.e. the CSP winner)
must be allocated since IR is violated otherwise. (P6) further pins
down the penalty to be exactly the second highest zero-crossing (i.e.
the CSP payment.) If not, either the allocated agent has a useful
deviation, or in some other economy the resource would not be
allocated. We also prove that CSP optimal for $(w_i,p_i)$ model given (P1)-(P5).
\begin{theorem} \label{thm:csp_opt_wipi} Assume the type space is the set of all $(w_i, p_i)$ value distributions. CSP achieves highest possible utilization among mechanisms satisfying (P1)-(P5), type profile by type profile.
\end{theorem}
This follows because, by
part~(v) of Lemma~\ref{lem:lem_P1P5_characterization}, we know that
the only way to beat the CSP mechanism
is to charge the CSP winner a penalty higher
than the second highest zero-crossing (this would incentivize higher
utilization). But penalties do not improve the
utilization in the $(w_i, p_i)$ model because if the agent cannot
use the resource, she cannot whatever the penalty.
Another implication of Lemma~\ref{lem:lem_P1P5_characterization}
is that CSP is not dominated given (P1)-(P5).
\begin{restatable
{theorem}{thmCSPnotDominated} \label{thm:csp_not_dominated} Assume the type space is the set of all value distributions satisfying
(A1)-(A3). No mechanism under (P1)-(P5) achieves weakly higher utilization than CSP for all economies, and a strictly higher utilization than CSP for at least one economy.
\end{restatable}
See Appendix~\ref{appx:optimality} for the proof. Intuitively, if a
mechanism charges the allocated agent a higher penalty than the second
highest zero-crossing in some economy, in some other economy it must
leave the resource unallocated or charge a smaller penalty, resulting
in lower utilization than CSP.
\subsection{Tightness of the Result} \label{sec:tightness}
CSP is unique given (P1)-(P7), and is undominated given (P1)-(P5). In regard to tightness of the results, note that relaxing any one of (P1)-(P3) respectively, we achieve higher utilization than CSP for all economies by (P1) implementing the full information first best, (P2) charging the allocated agent infinite no-show penalty, (P3) paying the allocated agent infinite bonus for showing up. Relaxing (P4) by always allocating to the same agent or (P5) allocating uniformly at random may achieve higher utilization than CSP for some economies, but results in lower utilization on average.
\dcp{reshef says: it may not be obvious why P4 and P5 are bad mechanisms. Perhaps compare with a more familiar scenario, e.g. combinatorial auction, where allocating at random (or to a dictator) clearly beats some mechanisms, yet there are more visible disadvantages.}
\hma{Was actually able to find one line of space and explained this a little bit. Actually, relaxing (P1)-(P3) we get something that dominates CSP. For the last two, isn't it obvious that the average utilization would be really bad?}
Relaxing either (P6) or (P7), CSP is no longer unique or optimal profile by profile. On the other hand, the performance of CSP should be expected to not only be undominated but good on average.
Relaxing (P6) and imposing (P7), we may set reserve penalties which improves utilization for some economies.
Simulation results, however, show that reserve penalties decrease average utilization for simple type distributions (see Section~\ref{sec:simulations}.)
The reason why reserve prices improve revenue in second price auctions but not utilization is that utilization given penalty $z$ consists of two parts:
\begin{align}
ut_{i^\ast}(z) = \Pm{V_{i^\ast} \geq z} = \Pm{V_{i^\ast} \geq 0} + \Pm{-z \leq V_{i^\ast} < 0}. \label{equ:utilization_res}
\end{align}
If an item is not sold because of a small reserve price, the loss of revenue is small and bounded by the reserve. In contrast, if a resource is not allocated, the loss of utilization may not be small, due to the loss of the first part of \eqref{equ:utilization_res}: $\Pm{V_{i^\ast} \geq 0}$, the probability with which the agent uses the resource if she is given the resource for free.
In fact, we prove that for economies with a single agent and simple type distributions, the optimal reserve penalty is zero (see Appendix~\ref{appx:reserve}.)
Imposing (P7) ``no charge" may seem without loss--- if an allocated agent $i^\ast$ is charged a positive base payment $y^\ast \hspace{-0.2em}> \hspace{-0.2em}0$, reducing $y^\ast$ to zero, increasing $z^\ast$ while keeping $u_{i^\ast}(z^\ast, y^\ast)$ the same, utilization is (weakly) improved without decreasing the utility of agent~$i^\ast$. This intuition seems to be supported by the result that the $\gamma$-CSP mechanism family, which violates~(P7), is still dominated by CSP.
\begin{restatable}{corollary}{gammaCSP_Utilization}
\label{thm:gammaCSP_Utilization}
For any set of agent types satisfying (A1)-(A3), CSP achieves weakly higher utilization than the $\gamma$-CSP mechanism for any $\gamma \in (0,1]$.
\end{restatable}
However, CSP does not dominate all mechanisms that satisfy (P1)-(P6) profile-by-profile. This is because with the multi-dimensional types and the $(z,y)$ payment space, different types disagree on which two-part payment is more favorable due to different probabilities of paying the penalty.
This disagreement enables mechanisms to select the $(z,y)$ contract as a function of an agent's {\em own} report without violating DSIC, and to sometimes charge a higher penalty than the CSP payment.
We have experimented with such a class of mechanisms,
but have not found them to achieve higher
average utilization than CSP in simulation (see Appendix~\ref{appx:agent_dependent_mechs})
\if 0
(P7) restricts payments in a 1D subspace thus agent-independence is necessary for DSIC. More generally, for any mechanism that satisfies agent-independence, payments must be on the second frontier of the economy thus the penalty cannot exceed the CSP payment and CSP is optimal (see Theorem~\ref{thm:CSP_Utilization} in Appendix~\ref{appx:optimality}.) This generalizes Theorem~\ref{thm:CSP_beat_SP}, and also implies that CSP achieves highest utilization among all $\gamma$-CSP mechanisms, since the $\gamma$-CSP mechanisms are agent-independent.
\fi
\section{Assignment of Multiple Resources} \label{sec:multi_item}
In this section, we generalize the model to allow for assigning
multiple resources but where each agent remains interested in
receiving an assignment of at most one resource (i.e., the unit-demand
model). We propose two DSIC mechanisms, the contingent
($m$+1)$^\mathrm{th}$ price mechanism for assigning $m$ identical
resources (e.g., spots in a spinning class,) and the generalized CSP
mechanism for assigning $m$ heterogeneous resources (e.g., time slots
of a charging station.)
\subsection{Multiple Identical Resources}
Consider first the scenario of assigning $m$ identical resources. An
agent's type is still $F_i$, describing her value distribution for
using one unit.
The CSP mechanism can be generalized as a ($m$+1)$^\mathrm{th}$ price mechanism, which collects a single bid from each agent, allocates to the $m$ highest bidders, and charge the ($m$+1)$^\mathrm{th}$ highest bid as the no-show penalty:
\begin{definition}[Contingent ($m$+1)$^\mathrm{th}$ Price Mechanism]
Let $b = (b_1, \dots, b_n)$ be a bid profile and w.l.o.g. reorder the agents as $b_1 \geq b_2 \geq \dots \geq b_n$, breaking ties at random.
\begin{enumerate}[$\bullet$]
\item Allocation rule: $x_i(b) = 1$ for $i \leq m$ and $x_i(b) = 0$, otherwise.
\item Payment rule: $t_{i}^{(0)}(b) = b_{m+1}$ and $t_{i}^{(1)}(b) = 0$ for $i \leq m$; $t_i(b) = 0$ for all $i$.
\end{enumerate}
\end{definition}
It remains a dominant strategy for an agent to bid her zero-crossing
$z^0_i$ under the contingent ($m$+1)$^\mathrm{th}$ price mechanism.
The CSP+R, $\gamma$-CSP and SP mechanisms can also be generalized
similarly, and the dominant strategies remain the same as in the
allocation of a single resource.
\begin{theorem} \label{thm:ck1p_theorem}
Assume the type space is the set of all value distributions satisfying (A1)-(A3). For allocating $m$ identical resources, the contingent ($m$+1)$^\mathrm{th}$ price mechanism is the only mechanism that satisfy (P1)-(P7). Moreover, this mechanism is optimal for the $(w_i,p_i)$ type model under (P1)-(P5), and dominates the generalization of the $\gamma$-CSP mechanisms for all $\gamma \in (0,1]$ in utilization.
\end{theorem}
\if 0
\dcp{i don't think we know what 'uniform' means, and i tried to hide the 'agent-independent' phrasing from above because it may cause confusion. if you need to state it this way then explain and define eg in words before the thm} \hma{ Ahh uniform means all assigned agents must be charged the same prices. How about we just replace the last piece of the theorem with ``is optimal under (P1)-(P5) for $(w_i, p_i)$ and dominates the $\gamma$ versions of the mechanism"}\dcp{yes, much better!} \hma{rephrased as above!}\dcp{great}
\hma{Also simplified the following proof intuition}\dcp{ok}
\fi
In particular, the contingent ($m$+1)$^\mathrm{th}$ price mechanism dominates the ($m$+1)$^\mathrm{th}$ price auction (i.e., the generalization of SP.) Similar to Lemma~\ref{lem:lem_P1P5_characterization}, we can show that the allocated agents must be on the top $m$ upper envelopes of the economy, the payment must be weakly above the ($m$+1)$^\mathrm{th}$ envelope, and the summation of the slopes of the top $m$ upper envelopes, which corresponds to the total utilization, increases monotonically as the penalty increases. The rest of the proof follows the same arguments as those of Theorem~\ref{thm:csp_uniqueness} and Corollary~\ref{thm:gammaCSP_Utilization}.
\if 0
With (P6) and (P7), payment must be the ($m$+1)$^\mathrm{th}$ zero-crossing, and thus the contingent ($m$+1)$^\mathrm{th}$ price mechanism is unique. When payments are agent-independent, the penalty cannot exceed the crossing point of the ($m$+1)$^\mathrm{th}$ frontier with the horizontal axis, so higher utilization is not possible.
\dcp{careful, agent-ind is used again here} \hma{will rephrase}
\fi
\subsection{Multiple Heterogeneous Resources}
Consider now the problem of assigning $m$ different resources.
Let $N = \{1, 2, \dots, n \}$ be the set of agents and $M = \{a, b, \dots, m\}$ be the set of resources. For each $a \in M$, the value for each agent $i$ to use resource $a$ is a random variable $V_{i,a}$ with CDF $F_{i,a}$. We assume $\{V_{i,a}\}_{i\in N, a \in M}$ are independent, and that for each $i \in N$, $a \in M$, $F_{i,a}$ satisfies (A1)-(A3). $\{ F_{i,a} \}_{a \in M}$ corresponds to agent $i$'s type.
This is the utilization-maximization version of the classical unit-demand assignment problem. The SP mechanism can be generalized as the VCG mechanism~\cite{vickrey1961counterspeculation,clarke1971,groves1973}, which collects a single bid from each agent for each alternative $b_{i,a}$, assigns the resources to maximizes the summation of total bids, and charges each assigned agent in period 0 the negative externality (in terms of bids) that she imposes on the rest of the agents.
Since expected utility \eqref{equ:exp_util_report} is quasi-linear in the period 0 payment $t_i$, bidding one's true values $b_{i,a}^\ast = \E{V_{i,a}^+}$.
What is interesting is that the na\"ive generalization of CSP, which
collects bids and charges the allocated agents the VCG payments as
penalties, fails to be incentive compatible because $u_{i,a}(z)$ as in
\eqref{equ:exp_util_z} is not quasi-linear in the penalty $z$. In this
way, we see a surprising connection with the literature on DSIC
assignment under non quasi-linear utilities.
\if 0
\dcp{take a bit of time to explain why and/or point to an example} \hma{will do}\hma{Hmm I'm a bit confused about what you have in mind for me to explain? Also found the following a bit confusing and rephrased. Original:
We generalize the CSP mechanism using contingent payments that are defined as with the
minimum competitive equilibrium
mechanism~\cite{demange1985strategy,DBLP:journals/ior/AlaeiJM16}. We
adopt $\{z_a\}_{a \in M}$ to define a set of penalties and assign
resources to agents in an agent-maximizing manner given these
penalties. We have a competitive equilibrium if, in addition, no
resource is assigned more than once and unassigned resources have
penalty zero.}\dcp{i'm happy with where we are!}
\fi
We generalize the CSP mechanism by setting the contingent payments for each resource to be the minimum competitive equilibrium (CE)
prices~\cite{demange1985strategy,DBLP:journals/ior/AlaeiJM16}.
A set of penalties $\{z_a\}_{a \in M}$ is a set of CE prices if when resources are assigned to agents in an agent-maximizing manner given these penalties, no resource is assigned more than once, and unassigned resources have penalty zero.
\begin{definition}[Generalized CSP Mechanism (GCSP)]
The GCSP mechanism collects from agents value distributions $\{F_{i,a}\}_{i \in N, a \in M}$, computes the expected utilities as functions of no-show penalties $\{ u_{i,a}(z)\}_{i \in N, a \in M}$ according to \eqref{equ:exp_util_z} and the minimum competitive equilibrium penalties $\{z_a\}_{a \in M}$.
\begin{enumerate}[$\bullet$]
\item Allocation rule: assign to each agent her most preferred resource given $\{z_a\}_{a \in M}$: $x_i = a_i^\ast = \arg \max_{a \in M} u_{i,a}(z_a)$ if $\max_{a \in M} u_{i,a}(z_a) > 0$, breaking ties to clear the market.
\item Payment rule: charge each allocated agent $t_i^{(0)} = z_{a_i^\ast}$ as no-show penalty if the agent is allocated some resource $a_i^\ast \in M$. All other payments are zero.
\end{enumerate}
\end{definition}
Demange and Gale~\cite{demange1985strategy} prove that for the unit-demand multi-item
assignment problem where agents' utilities are non-quasi-linear but
decrease continuously in payments, as is the case for $\{
u_{i,a}(z)\}_{i \in N, a \in M}$, the minimum CE prices exist.
Moreover, the mechanism which sets the minimum CE prices and assigns the items in an agent-maximizing manner, is truthful, and
clears the market. Thus we conclude that the generalized CSP
mechanism is well defined, and that truthful reporting is a DSE.
For the computation of the CE
prices, Alaei et. al.~\cite{DBLP:journals/ior/AlaeiJM16} provide a recursive
algorithm that computes prices based on outcomes in economies with
less agents and less alternatives.
\if 0
\hma{Got rid of the description of the algorithm all together. Should we include this in the appendix?}
We use the algorithm for non-quasi-linear assignment (Alaei et al.~\cite{alaei2010walrasian}) to generalize the CSP mechanism to this setting. The generalized CSP mechanism computes agent-independent prices recursively, and allocates in the agent-maximizing manner.
\begin{algorithm*}[t]
\caption{Alaei's mechanism $\mathrm{ALG}_1$} \label{alg:Alaei}
\KwIn{Utility functions $u = \{(u_{i,j}(z) \}_{i\in N,j\in M}$}
\KwOut{Assignment $x_{i,j} \in \{0,1\}$, payments $z_i$ and resulting expected utilities $\pi_i$}
\tcp{\small{If there is only one agent, the agent gets her favorite resource free of charge}}
\uIf{$|N| = 1$ } {
$x_{1,j^\ast} \leftarrow 1$ for $j^\ast = \arg \max u_{1,j}(0)$; $x_{1,j} \leftarrow 0$ for $j \neq j^\ast$; $z_1 \leftarrow 0$; $\pi_1 \leftarrow u_{1,j^\ast}(0)$.
}
\Else{
\tcp{\small{If there is only one resource to allocate, compute the CSP outcome}}
\uIf{$|M| = 1$}{
$x_{i^\ast,1} \leftarrow 1$ for $i^\ast = \arg \max z^0_{i,1}$ where $z^0_{i,1} = u_{i,1}^{-1}(0)$;
$z_{i^\ast} \leftarrow \max_{i\neq i^\ast} z^0_{i,1}$;
$\pi_{i^\ast} \leftarrow u_{i^\ast, 1}(z_{i^\ast})$;
$x_{i,1} \leftarrow 0$, $z_i \leftarrow 0$ and $\pi_i \leftarrow 0$ for all $i\neq i^\ast$.
}
\tcp{\small{Allocating $|M|>1$ resources to $|N| > 1$ agents}}
\Else{
\For{$i\in N$, $j \in M$}{
\tcp{\small{Compute the outcome in the economy without agent $i$ or resource $j$}}
$(x_{(-i,-j)}, z_{(-i,-j)}, \pi_{(-i,-j)}) \leftarrow \mathrm{ALG}_1(u_{-i,-j})$;
\lFor{$i' \neq i$}{
$q_{(-i,-j),i'} \leftarrow u_{i', j}^{-1}(\pi_{(-i,-j),i'})$
}
$z_{i,j} \leftarrow \max \{ 0, \max_{i'\neq i} q_{(-i,-j),i'} \} $
}
\tcp{\small{Compute the agent-maximizing allocation and determine the payments}}
\For{$i \in N$}{
\If{ $\max_{i\in M} u_{i,j}(z_{i,j}) > 0$ } {
$x_{i,j^\ast} \leftarrow 1$ for $j = \arg \max u_{i,j}(z_{i,j})$; $x_{i,j} \leftarrow 0$ for $j\neq j^\ast$;
$z_{i} \leftarrow z_{i,j^\ast}$; $\pi_i \leftarrow u_{i,j^\ast}(z_i)$;
}
\lElse{$x_{i,j} \leftarrow 0,~\forall j$; $z_i \leftarrow 0$; $\pi_i \leftarrow 0$}
}
}
}
\Return $x$, $z$, $\pi$.
\end{algorithm*}
Intuitively, Algorithm~\ref{alg:Alaei} computes for each agent $i$ and each resource $j$ the critical payment $z_{i,j}$, at which the rest of the agents $-i$ are indifferent between getting $j$ at this payment $z_{i,j}$ or getting the outcome in the economy without agent $i$ or resource $j$.
By construction, the critical penalties $z_{i,j}$ are independent to agent $i$'s own reports, and the assignment rule is agent maximizing thus the mechanism Algorithm~\ref{alg:Alaei} is DSIC, thus so is GCSP. We can verify that GCSP satisfies (P2)-(P7).
We can compare this mechanism with the appropriate generalization of the SP mechanism, in this case the Vickrey-Clarke-Groves (VCG) mechanism~\cite{vickrey1961counterspeculation,clarke1971,groves1973}. In particular, we use VCG with a fixed penalty (termed VCG+C.) This collects a single bid from each agent for each alternative $b_{i,j}$, computes the assignment that maximized the summation of total bids, and charges each allocated agent at period 0 the negative externality (in terms of bids) that she imposes on the rest of the agents. It is a dominant strategy to bid true values in VCG mechanisms, and with the same analysis as that of the SP+C mechanism, we can show that it is a dominant strategy for each agent to bid $b_{i,j}^\ast = \E{V_{i,j} \cdot \one{V_{i,j} \geq -C}} - C \cdot \Pm{V_{i^\ast} < -C}$ on each alternative $j$ if this is non-negative, but we omit the formal definition and the proof.
\fi
For assigning $m$ identical resources, the minimum CE penalties are
exactly the $m+1^{\mathrm{th}}$ highest zero-crossing among all
agents, thus GCSP coincides with the contingent $m+1^{\mathrm{th}}$
price mechanisms and generalizes both that and the CSP
mechanism.
Unlike the assignment of identical resources, the utilization under GCSP need not dominate that of the VCG mechanism. We show this through a pair of examples. Still, simulation results in Section~\ref{sec:simulations} show that the GCSP mechanism almost always achieves higher utilization than VCG.
\if 0
\savespace{, except for some extreme economies where there is little competition}.\dcp{agree ok to cut this, and even if
we don't need to for space}
\fi
\begin{example}[GCSP beats VCG] \label{ex:Mk} Consider an economy two resources $a$, $b$ and two agents 1 and 2 and the following types according to the $(w_i, p_i)$ type model:
\begin{enumerate}[$\bullet$]
\item Agent 1: $w_{1,a} = 200$, $p_{1,a} = 0.2$; $w_{1,b} = 20$, $p_{1,b} = 0.8$.
\item Agent 2: $w_{2,a} = 50$, $p_{2,a} = 0.8$; $w_{2,b} = 80$, $p_{2,b} = 0.4$.
\end{enumerate}
The expected utilities are as shown in Figure~\ref{fig:example_CVCG_better}. Under VCG, agents bid $b_{i,j}^\ast = w_{i,j}p_{i,j}$. Agent 1 is allocated resource $a$ and agent 2 is allocated resource $b$, and the total utilization is $p_{1,a} + p_{2,b} = 0.6$.
\if 0
\hma{Keeping the detailed computation in the .tex in case this is helpful later}
We now compute the outcome under generalized CSP. The expected utility functions are of the form $u_{i,j}(z) = v_{i,j}p_{i,j} - (1-p_{i,j})z$ for all $i,j$. We feed $u$ into Algorithm~\ref{alg:Alaei}, which computes agent-independent prices $z_{i,j}$ for all $i,j$ then do agent-maximizing allocation. Recall that $\pi_{(-i,-j),i'}$ denotes the utility of $i'$ in the economy without $i,j$. First, for agent 1 and resource $a$, we need to compute the outcome the outcome in the economy without 1 or $a$ (denoted $E_{(-1,-a)}$.) In $E_{(-1,-a)}$, there is only one agent in the economy, thus agent 2 gets resource $b$ for free and her utility is $\pi_{(-1,-a),2} = w_{2,b}p_{2,b} = 32$.
The payment at which agent 2 is indifferent between getting $a$ and getting $\pi_{(-1,-a),2}$ is therefore $q_{(-1,-a),2} = u_{2,a}^{-1}(\pi_{(-1,-a),2}) = 40$, thus agent 1 face a payment $z_{1,a} = \max(q_{(-1,-a),2}, ~0) = 40$ on resource $a$.
Similarly, we can compute $z_{1,b} = 0$, $z_{2,a} = 30$ and $z_{2,b} = 0$. Under these payments, $u_{1,a}(z_{1,a}) = 8 < u_{1,b}(z_{1,b}) = 16$ thus agent 1 is allocated resource $b$. $u_{2,a}(z_{2,a}) = 34 > u_{2,b}(z_{2,b}) = 32$, thus agent 2 is allocated resource $a$.
\fi
The lowest CE penalties can be computed as $z_a = 30$ and $z_b = 0$. Under these penalties, $u_{1,a}(z_a) = w_{1,a}p_{1,a} - (1-p_{1,a}) z_a = 16$, $u_{1,b}(z_b) = w_{1,b}p_{1,b} - (1-p_{1,b}) z_a = 16$ thus agent 1 is indifferent between the two resources. For agent 2, $u_{2,a}(z_a) = w_{2,a}p_{2,a} - (1-p_{2,a}) z_a = 40 - 0.2 \cdot 30 = 34$, $u_{2,b}(z_b) = w_{2,b}p_{2,b} - (1-p_{2,b}) z_b = 32$, thus agent 2 strictly prefers resource $a$. Under the GCSP mechanism, agent 1 is allocated resource $b$, while agent $b$ is allocated resource $a$. The total utilization under GCSP is therefore $p_{1,b} + p_{2,a} = 1.6$, higher than that of VCG.
\end{example}
\begin{figure}[t!]
\spacecheat{\vspace{-1em}}
\centering
\subfloat[Example~\ref{ex:Mk}]{
\includegraphics[scale=0.75]{figures/example_CVCG_better-eps-converted-to.pdf}
}
\hspace{-1em}
\subfloat[Example~\ref{ex:VCG_better}]{\label{fig:example_CVCG_worse}
\includegraphics[scale=0.75]{figures/example_CVCG_worse-eps-converted-to.pdf}
}
\spacecheat{\vspace{-0.8em}}
\caption{An example of utility curves for the heterogeneous assignment problem. \label{fig:example_CVCG_better} \spacecheat{\vspace{-1em}}
}
\end{figure}
We can see from Example~\ref{ex:Mk} that comparing with the VCG mechanism, which favors agents with high vertical intercept, agents with high zero-crossings and therefore shallower expected utility functions are more likely to get allocated in GCSP. This can lead in turn to higher utilization. For agents with continuous types (e.g. the exponential type model,) utilization is further improved due to the penalty charged by the mechanism.
Still, VCG may provide higher utilization in somewhat extreme examples where the number of agents is small and there is little competition.
\begin{example}[VCG beats GSCP]\label{ex:VCG_better}
Consider an economy with the following two agents 1 and 2, two items $a$ and $b$, the $w_i p_i$ type model, and agent types:
\begin{itemize}
\item Agent~1: $v_{1,a} = 200$, $p_{1,a} = 0.2$; $v_{1,b} = 550$, $p_{1,b} = 0.1$.
\item Agent~2: $v_{2,a} = 37.5$, $p_{2,a} = 0.8$; $v_{2,b} = 66.67$, $p_{2,b} = 0.6$.
\end{itemize}
The expected utility curves are as shown in Figure~\ref{fig:example_CVCG_worse}. Under VCG, agent 1 gets resource $b$ and pays 10, agent 2 gets resource $a$ and pays 0. Total utilization is $0.1+0.8 = 0.9$. Under GCSP, we can compute following the same steps as in Example~\ref{ex:Mk} that agent 1 gets resource $a$ and pays 0, agent 2 gets resource $b$ and pays 16.67 for no show penalty.
The total utilization is $0.2 + 0.6 = 0.8$, which is lower than that of VCG.
\end{example}
\section{Simulation Results} \label{sec:simulations}
In this section, we compare resource utilization for different
mechanisms under the natural exponential type model (Example~\ref{ex:exp_model},) where agents' uncertain values are parametrized by a fixed value and the expected opportunity cost. A type distribution specifies the distributions of these parameters.
The simulation results are very robust
to different type models and different distributions on
types.\footnote{See Appendices~\ref{appx:bids} and~\ref{appx:additional_simulation} for
derivation of bids in different models, and for a large sweep of
simulation results.}
\subsection{Single Resource}
For allocating a single resource, an agent's value for using the resource is $V_i = w_i - O_i$ where $w_i>0$ is the fixed value and $O_i \sim Exp(\lambda_i)$ is the opportunity cost. $\E{V_i} = w_i - \lambda_i^{-1}$ where $\lambda_i^{-1}$ is the expected value of the opportunity cost.
We consider the type distribution where the value and the expected opportunity cost $\lambda_i^{-1}$ are uniformly distributed: $\lambda_i^{-1} \sim \mathrm{U}[0,~L]$ and $w_i \sim \mathrm{U}[0,~\lambda_i^{-1}]$. The results are robust to the range $L$, and $w_i < \lambda_i^{-1}$ guarantees (A1)-(A3) are satisfied.
\paragraph{SP, CSP and Benchmarks}
We set $L = 10$, vary the number of agents from 2 to 15, and
consider the average utilization over 10,000 randomly generated profiles under SP, CSP and other benchmarks. See Figure~\ref{fig:singleItem_exp_SP_CSP_FB}. The {\em First Best} is the highest achievable utilization subject to the IR and ND as discussed in Section~\ref{sec:optimality}. {\em P1-P5 UB} is the upper bound on utilization under (P1)-(P5) as discussed in Lemma~\ref{lem:lem_P1P5_characterization}.
{\em Random} is the average utilization achieved by randomly allocating the resource to one of the agents without charging any payment. The CSP mechanism achieves a large improvement compared with the SP auction and Random, and performs relatively well when compared with the P1-P5 UB.
With three agents, Figure~\ref{fig:singleItem_exp_SP_CSP_scatter} compares the utilization under CSP and SP profile-by-profile. The utilization of CSP dominates that of SP, as proved in Theorem~\ref{thm:CSP_beat_SP}.
\begin{figure}
\spacecheat{\vspace{-1em}}
\centering
\subfloat[SP, CSP and benchmarks]{\label{fig:singleItem_exp_SP_CSP_FB}
\includegraphics[scale=0.78]{figures/singleItem_exp_SP_CSP_FB-eps-converted-to.pdf}
}
\hspace{-0.8em}
\subfloat[]{\label{fig:singleItem_exp_SP_CSP_scatter}
\includegraphics[scale=0.78]{figures/singleItem_exp_SP_CSP_scatter-eps-converted-to.pdf}
}
\subfloat[Effect of reserve prices]{\label{fig:singleItem_exp_SPC_CSPR}
\includegraphics[scale=0.78]{figures/singleItem_exp_SPC_CSPR-eps-converted-to.pdf}
}
\spacecheat{\vspace{-0.8em}}
\caption{Utilization of a single resource under the exponential model
\label{fig:utilization_exp_threeItem}
} \spacecheat{\vspace{-1em}}
\end{figure}
\paragraph{Fixed Penalty and Reserve Prices}
\if 0
\todo{ Under SP+C, an assigned agent uses the resource only when $V_i \geq -C$ thus the DSE bids $b_{i,{\mathrm{SP+C}}}^\ast = \E{V_{i} \cdot \one{V_{i} \geq -C}} - C \cdot \Pm{V_{i} < -C}$.
}
\fi
Here, we fix the number of agents to be $n=10$. Figure~\ref{fig:singleItem_exp_SPC_CSPR} examines utilization under CSP with reserve price (CSP+R) and under SP with fixed penalty (SP+C).
We can see that while $C>0$ may improve the utilization of SP, it is not enough to achieve higher utilization than CSP. Fixing $R=C$, the utilization of CSP+R not only achieves a higher average utilization than SP+C but dominates that of SP+C profile-by-profile (as proved in Theorem~\ref{thm:CSPC_Utilization}, not shown in the figure.)
We see that CSP+R is best with zero reserve penalty. This means
that the improvement in utilization due to an additional penalty is smaller than the loss of utilization due to the resource becoming unallocated (see discussions in Section~\ref{sec:tightness}.)
\subsection{Multiple Resources}
In this section, we compare the performance of different mechanisms when the resources are not identical. For each agent $i$ and alternative $a$, we model $V_{i,a}$ as
being an exponential type with parameters $w_{i,a}$ and $\lambda_{i,a}$, and we assume the opportunity costs $\lambda_{i,a}^{-1}$ are i.i.d. $\lambda_{i,a}^{-1} \sim \mathrm{U}[0, ~L]$ and that given $\lambda_{i,a}$, the value $w_{i,a} \sim \mathrm{U}[0,~\lambda^{-1}_{i,a}]$.
We fix the number of resources to be $m = 3$, and the maximum opportunity cost $L = 10$. The results are robust to different $m$ and $L$. The average utilization over 10,000 randomly generated profiles is shown in Figure~\ref{fig:3Item_exp_SPCSP}, as we vary the number of agents from 2 to 15. The {\em First Best} benchmark computes for each economy the assignment that maximizes total utilization under the full information first best, and \emph{FCFS} (first come first serve) is equivalent to the random serial dictatorship, where agents arrive in random order and gets assigned for free her favorite resource that is still available.
\begin{figure}
\spacecheat{\vspace{-1em}}
\centering
\subfloat[]{\label{fig:3Item_exp_SPCSP}
\includegraphics[scale=0.78]{figures/3Item_exp_SPCSP_alt-eps-converted-to.pdf}
}
\hspace{-0.8em}
\subfloat[]{\label{fig:3Item_exp_fractions}
\includegraphics[scale=0.78]{figures/3Item_exp_fractions_alt-eps-converted-to.pdf}
}
\subfloat[]{\label{fig:3items_exp_CSPR_SPC}
\includegraphics[scale=0.78]{figures/3items_exp_CSPR_SPC_alt-eps-converted-to.pdf}
}
\spacecheat{\vspace{-0.5em}}
\caption{Utilization for assigning three resources under the exponential type model. \label{fig:utilization_exp_multItem}} \spacecheat{\vspace{-1em}}
\end{figure}
As the number of agents increases, the utilization of GCSP increases significantly because competition results in agents with higher probability of utilization being selected and increases the penalties which further boost utilization.
Unlike in the single resource case, the utilization under GCSP does not dominate that of VCG profile-by-profile. The fraction of profiles for which GCSP has strictly higher utilization, and for which VCG has strictly higher utilization are shown in Figure~\ref{fig:3Item_exp_fractions}. When the number of agents is very small, for about 1\% of the profiles, VCG has strictly higher utilization than GCSP. With three or more agents, GCSP is almost always better.
As with the SP auction, the VCG mechanism for assigning multiple
resources can also be generalized to include a fixed penalty $C$
(termed VCG+C). The effect of reserve prices $R$ for GCSP (the
GCSP+R) and fixed penalty $C$ for VCG+C is shown in
Figure~\ref{fig:3items_exp_CSPR_SPC}, where the number of agents $n$
is fixed to be 10 and the reserve $R$ and penalty $C$ vary from 0 to
10. There is a very small improvement in utilization for the GCSP+R
mechanism comparing with no reserve when $R$ is close to zero,
otherwise the patterns are similar as in the case of single resource
assignment.
\section{Conclusion}
We introduced the problem of maximizing resource utilization in a setting where
a plan needs to be made but there is uncertainty about the actual values of
agents that will be realized for resources (and indeed, these may be negative,
reflecting that they have better things to do!).
For allocating a single resource, we proved that the natural contingent second-price
mechanism is unique under a set of desired criteria, and maximizes utilization across a larger class of mechanisms. We also generalized the mechanism to assign multiple resources,
connecting this problem with that of assignment under non quasi-linear utility,
and giving theoretical results for the case of multiple, identical items,
and strong support from simulation for the case of multiple, heterogeneous items.
Directions for future work include weakening the assumptions on the type space for the optimality of CSP, developing theoretical results for assigning multiple, heterogeneous resources, and designing contingent mechanisms to incentivize utilization or reliability in repeated assignments of resources or tasks (e.g. scheduling serving times at a childcare co-op operated by the parents, rebalancing in bike-sharing systems by the customers.)
More generally, we see a rich agenda in the design of what we think about as ``coordination
mechanisms''--- these are mechanisms that need to form a plan and construct payment
schedules to promote coordinated outcomes by a system of agents.
Exploring the effect of present-bias~\cite{Frederick2002} in the coordination of future events and designing commitment devices~\cite{bryan2010commitment} through contingent-payment mechanisms is another interesting direction.
\if 0
\section*{Acknowledgements}
The authors thank Rakesh Vohra, Ilya Segal, Scott Kominers, Di Pei, Thibaut Horel and Debmalya Mandal for helpful discussions.
\fi
\bibliographystyle{plain}
|
\section{Correspondence between \texorpdfstring{$B\tau_E$ and $Q/Q_{\rm gB}$}{energy confinement time and heat flux} scalings}
\label{scaling_correspondence}
In the experimental scaling (\ref{MASTscaling}),
$\tau_E$ is the energy confinement time, i.e., the time it takes the power loss $P \sim QA$ to carry away the thermal energy $W \sim nTV$, where $A$ and $V$ scale with flux surface area and volume respectively, $Q$ is the energy flux, $n$ is density, and $T$ is temperature. Therefore,
\begin{equation}
\tau_E \sim \frac{W}{P} \sim \frac{nT a}{Q},
\label{tauE}
\end{equation}
where $a \sim V/A$ is the macroscopic length scale.
The appropriate normalising quantity for $Q$ is the ``gyro-Bohm'' heat flux $Q_{\rm gB} = nTv_{te}\rho_\ast^2$,
where $v_{te}=(2T/m)^{1/2}$ is the thermal speed,
$\rho_\ast = \rho_e/a$, $\rho_e = v_{te}/\Omega_e$ is the gyroradius and $\Omega_e$ is the
cyclotron frequency (since we are interested in electron transport, we are using electron
quantities, but the argument for ions is exactly the same).
In the gyrokinetic ordering, the heat flux is
$Q = \mathcal O(\rho_\ast^2)$ automatically (see, e.g., \cite{Abel13}),
so the normalised heat flux $Q/Q_{\rm gB}$ must be order unity and independent of $\rho_\ast$.
This normalised heat flux is
\begin{equation}
\frac{Q}{Q_{\rm gB}} = \frac{Q}{n T v_{te} \rho_\ast^2} = \frac{Q}{n T a \Omega_e\rho_\ast^3}
\sim \frac{1}{\Omega_e\tau_E \rho_\ast^3},
\label{QQgb}
\end{equation}
where we have used (\ref{tauE}) to obtain the last expression.
At constant $\rho_\ast$, (\ref{QQgb}) implies
\begin{equation}
\frac{Q}{Q_{\rm gB}} \propto \frac{1}{B \tau_E},
\end{equation}
so we are probing the same dependence on collisionality by measuring $Q/Q_{\rm gB}$ in
gyrokinetic simulations as the experiments do by measuring $B\tau_E$ with $\rho_\ast$ fixed.
Let us now explain what is meant when scaling (\ref{MASTscaling}) is claimed to hold
experimentally \cite{Valovic}. The starting point is to assume that
the following parametrisation of the energy confinement time holds:
\begin{equation}
\Omega_e\tau_E = K\nu_\ast^{x_\nu}\rho_\ast^{x_\rho}q^{x_q} \beta^{x_\beta}\kappa^{x_\kappa},
\label{Ansatz}
\end{equation}
where $K$ is a constant, $\kappa$ is elongation \cite{Miller98},
$\beta$ is plasma beta, $q$ is the safety
factor and
\begin{equation}
\nu_\ast = \frac{\nu}{\varepsilon^{3/2}v_{te}/qR}
\label{nustar}
\end{equation}
is the effective electron collision rate $\nu/\varepsilon$ divided by the bounce
frequency $\varepsilon^{1/2}v_{te}/qR$, where $\varepsilon = a/R$ is the inverse aspect ratio,
and $R$ the major radius. This is the dimensionless quantity characterising collisional detrapping.
The experiments looking for the ``collisionality scaling''
strive to hold all these parameters constant apart from $\nu_\ast$ and measure
the scaling exponent $x_\nu$ of $B\tau_E$ with respect to $\nu_\ast$.
Since in our numerical study, we do not vary the equilibrium geometry,
the same exponent $x_\nu$ will apply to the
normalised collisionality in our gyrokinetic simulations,
which is simply $\nu a/v_{te}$.
Note finally that, in view of (\ref{QQgb}), if $Q/Q_{\rm gB}$ is independent of $\rho_\ast$
(as it is hard-wired to be within the gyrokinetic ordering), it must be the case
in (\ref{Ansatz}) that $x_\rho = -3$, which is indeed consistent with
the experimental findings \cite{Valovic}.
\section{Plasma parameters}
\label{plasma_parameters}
The reference, or ``nominal'', plasma parameters for our simulations represent MAST shot 8500 at $t = 0.289$ s.
They are specified in Table \ref{parameters}. We use a
Miller \cite{Miller98} representation of the flux-surface geometry, which is assumed to be axisymmetric.
\begin{table}[htbp]
\center
\begin{tabular}{lcr}
\hline
minor radius & $r/a$ & 0.65\phantom{0} \\
major radius & $R/a$ & 1.46\phantom{0} \\
safety factor & $q$ & 1.9\phantom{00} \\
magnetic shear & $\hat{s}$ & 1.8\phantom{00} \\
Shafranov shift & $R^\prime/a$ & $-0.25$\phantom{0} \\
pressure gradient & $p^\prime\beta/p$ & $-0.12$\phantom{0}\\
elongation & $\kappa$ & 1.57\phantom{0} \\
elongation gradient & $\kappa^\prime$ & 0.40\phantom{0} \\
triangularity & $\delta$ & 0.22\phantom{0} \\
triangularity gradient & $\delta^\prime$ & 0.16\phantom{0} \\
\hline
density gradient & $a/L_n$ & $-1.2\phantom{00}$ \\
electron temperature gradient & $a/L_T$ & 3.42\phantom{0} \\
effective ion charge & $Z_{\mathrm{eff}}$ & 1.15\phantom{0} \\
species temperature ratio & $T_e/T_i = 1/\tau$ & 1.06\phantom{0} \\
electron collisionality & $\nu_{\mathrm{nom}} a/v_{te}$ & 0.02\phantom{0}\\
flow shear & $\gamma_E a/v_{te}$ & $-0.003$\\
\hline
\end{tabular}
\caption{Nominal dimensionless local equilibrium parameters based on MAST shot 8500 at $t = 0.289$ s.
Here $a = 0.55$\,m is the half diameter of the last closed flux surface at the elevation of the magnetic axis and $r$ is the same for the particular flux surface we have chosen.
The prime symbol denotes the derivative with respect to normalised minor radius: $^\prime\equiv a(d/dr)$. Scale lengths are defined by $a/L_n\equiv-n^\prime/n$ (density) and $a/L_T\equiv-T^\prime/T$ (temperature).
}
\label{parameters}
\end{table}
Note that the reversed density gradient is not rare in MAST \cite{McConeThesis} and was found to be linearly stabilizing \cite{Garzotti}.
In our simulations, we varied the collisionality $\nu$ relative to its nominal value $\nu_{\mathrm{nom}}$
given in Table \ref{parameters}.
We carried out such a collisionality scan for two values of $a/L_T$, the nominal one given
in Table \ref{parameters} and a slightly smaller (slightly more marginal) $a/L_T = 3.3$.
All other parameters were held fixed.
A small amount of flow shear $\gamma_E$, given in Table~\ref{parameters},
was included in the nonlinear simulations. As in \cite{Roach09}, this was found to assist
convergence and is of the same order as the experimental level in MAST (see \ref{shear}).
The sign of $\gamma_E$ is unimportant because the equilibrium is up-down symmetric \cite{Parra11}.
\section{Numerical details}
\label{numerical_parameters}
\subsection{GS2 spatial coordinates}
\label{coords}
GS2 solves the gyrokinetic equation (\ref{GKE}) in a flux tube that winds its way
toroidally and poloidally around the toroidal flux surface designated by a label $r$
(see Table \ref{parameters}), which is defined as its half diameter at the elevation
of the magnetic axis of the tokamak. The position along the flux tube
is determined by the poloidal angle $\theta$, which can be thought of as
the coordinate parallel to the magnetic field. The flux-tube domain that we use
in GS2 extends over $-\pi\le\theta\le\pi$, where
$\theta = 0$ is the outboard midplane and $\theta = \pm\pi$ the inboard midplane.
If we demand that
\begin{equation}
\mathbf{B} = \nabla \alpha \times \nabla \psi,
\label{Clebsch}
\end{equation}
where $\psi$ is the poloidal magnetic flux and $\alpha$ is chosen to satisfy (\ref{Clebsch}),
each field line is labelled by the pair of Clebsch coordinates $(\psi,\alpha)$.
Then $(\psi, \alpha, \theta)$ can be used as the three independent (but not orthogonal!)
curvilinear coordinates. Designating the central field line of a flux tube by
$(\psi_0,\alpha_0)$, the GS2 coordinates $(x,y)$ transverse to the magnetic-field direction are
related to $(\psi,\alpha)$ in the following manner \cite{BeerCowley,EdmundThesis}:
\begin{eqnarray}
\label{defx}
x &= \frac{q(\psi_0)}{B_0 r(\psi_0)} (\psi-\psi_0),\\
y &= \frac{1}{B_0}\left.\frac{d\psi}{d r}\right|_{r=r(\psi_0)} (\alpha - \alpha_0),
\label{defy}
\end{eqnarray}
where the safety factor $q(\psi)$ and the radial flux-surface label $r(\psi)$ are both
flux functions and $B_0$ is a reference (constant) magnetic field defined to be
the toroidal magnetic field at the mid-radius of the flux surface containing our flux tube.
The $x$ coordinate is ``radial'' in the sense that it is transverse
to the flux surface and increases away from the magnetic axis.
The second coordinate $y$ is often called ``binormal'' because it
effectively measures distances perpendicular to the field line
but within the flux surface (at constant $\psi$).
In the outboard midplane, $(x,y)$ form an approximately orthogonal grid
in the cross-section of the flux tube
perpendicular to the magnetic field, with $x$ and $y$ approximately
measuring true distances from the centre of the flux tube.
Following the field line around the flux surface, the flux tube twists
due to magnetic shear, with $\nabla x$, $\nabla y$ and $\nabla \theta$
generally not orthogonal to each other.
GS2 is pseudospectral with respect to $(x, y)$ but not $\theta$; that is, except for the
evaluation of the nonlinear term, its data grid has the coordinates $(k_x, k_y) = \mathbf{k}_\perp$ and $\theta$. For the evaluation of the nonlinear term, the Fast Fourier Transform is used
to go from $(k_x, k_y)$ to $(x,y)$
(and then back again after the evaluation). In the context of the GS2 grids, the terms ``parallel'' and ``perpendicular'' refer to $\theta$ and to the other two coordinates (either $(x, y)$ or $(k_x, k_y)$ as appropriate),
respectively.
\subsection{Numerical grids, resolution and boundary conditions}
\label{numerics}
In our largest nonlinear simulations,
the spatial grid has 48 cells in $\theta$ and $108 \times 144$ points in $x \times y$, the range of positive perpendicular wavenumbers being
$(k_x\rho_{e},k_y\rho_{e})\in[0.012,0.40]\times[0.012,0.57]$.
We refer to these as ``large-box simulations'' in the text.
The simulations referred to as ``small-box simulations'' are the same except for the perpendicular grid parameters: the box size in real space is halved, but extended to somewhat higher perpendicular wavenumbers (i.e., higher spatial resolution): $80 \times 108$ in $x \times y$, spanning the spectral range
$(k_x\rho_{e},k_y\rho_{e})\in[0.023,0.60]\times[0.024,0.84]$.
The parameter regimes in which these two classes of simulations are useful
are discussed in \ref{convergence_study}.
The wavenumber ranges are intended to capture electron physics;
the Boltzmann ion response ensures that we need not worry about where they lie in relation to ion scales.
Dealiasing is used for the nonlinear term: zero-padding with higher-wavenumber modes is introduced before the Fourier step into real space; any nonlinear interactions that wrap around unphysically in reciprocal space will be nullified when these modes are thrown away after the inverse Fourier step.
The number of $x \times y$ points indicated here (applicable only to nonlinear runs) is the same as the total number of modes including the dealiasing modes, but the ranges of wavenumbers given above do not include the dealiasing modes. All data shown here, including plots in real space, derive solely from the evolved, non-dealiasing modes: the additional dealiasing modes are used only transiently during the nonlinear part of each time step, and are not included in the code's output.
Twist-and-shift parallel boundary conditions \cite{BeerThesis, BeerCowley} link together some of the
$k_x$ modes at low $k_y$, whereas the
$k_y = 0$ modes are periodic in $\theta$. At the remaining (open) domain ends, the parallel boundary condition is $h = 0$ for incoming particles. The boundary conditions are automatically periodic in the perpendicular
directions $(x,y)$ by virtue of the spectral representation.
The velocity-space grid has 18 energies ${\cal E} = v^2/2$, and, for passing particles, 16 grid-points in $\lambda = \mu/\cal E$ for each sign of $v_\parallel = \pm(2 {\cal E} - 2\mu B)^{1/2}$, where $\mu= v_\perp^2/2B$ is the magnetic moment.
Trapped particles are accommodated by additional $\lambda$ grid points, one for particles bouncing at each $\theta$.
Linear simulations (\ref{linear_simulations}) use the same setup as nonlinear ones, but evolve each perpendicular wavenumber independently.
For the ``simplified'' simulations in section \ref{zeffscansection}, in which only the electron-ion collisionality was varied, the grid sizes were: 24 cells in $\theta$; $40 \times 54$ points in $x \times y$, spanning the spectral range
$(k_x\rho_{e},k_y\rho_{e})\in[0.023,0.30]\times[0.024,0.41]$;
12 energies $\cal E$; and 8 $\lambda$'s for passing particles in each direction. This version of the code applied the zero incoming boundary condition to
$g = \langle\delta f\rangle = h + e\langle\phi\rangle F/T$,
rather than to $h$.\footnote{
The option to use $h$ was added to the code during the course of the present work. The difference
between the two options vanishes as $\phi\rightarrow0$ at $\theta\rightarrow\pm\infty$, but it has been found in linear simulations \cite{ColinGreg} that the use of $h$ leads to faster convergence as the actual, finite
$\theta$ domain size is increased.}
\section{Linear growth rates}
\label{linear_simulations}
\begin{figure}[t]
\centering
\includegraphics[scale=0.75]{linear_enorm_nobox}
\caption{Linear growth rates versus $k_y$, at $k_x = 0$, with zero flow shear $\gamma_E = 0$,
for various collisionalities as indicated.
All other parameters are the nominal ones given in Table \ref{parameters} (including $a/L_T=3.42$).
}
\label{linear}
\end{figure}
Figure \ref{linear} shows the growth rate of the most unstable linear mode at $k_x = 0$
for each $k_y$ in the numerical domain (which was described in \ref{numerics}).
These linear simulations have zero flow shear, $\gamma_E=0$.
At the given $a/L_T$, the effect of decreasing collisionality is to decrease slightly the maximum growth rate, and to extend to somewhat lower $k_y$ the unstable wavenumber interval. At
$k_y\rho_{te}\approx 0.2$
where the saturated heat flux is maximal (see Figure \ref{kyspectrum}), the dependence of the linear spectrum on collisionality is weak.
Note that this scan is consistent with the previous observation \cite{Roach09} that
collisions lead to a stability gap between ETG- and ITG-driven modes (the range of $k_y$'s unstable to ITG being lower than the wave numbers shown here).
Figure \ref{critical} shows the dependence of the maximum growth rate on the temperature gradient
(still with $\gamma_E=0$). We see that the linear critical gradient varies only between
$2.3$ and $2.5$ throughout this whole range spanning a factor of 50 in collisionality.
This is in agreement with the earlier findings \cite{Jenko2001} that collisions have very little
effect on the linear critical gradient.
\begin{figure}[b]
\centering
\includegraphics[scale=0.75]{gamma_max_by_nu4_colour2_enorm_nobox}
\caption{
Linear growth rates versus $a/L_T$, with zero flow shear $\gamma_E = 0$,
for various collisionalities as indicated,
showing that the linear critical gradient (where the growth rate crosses zero)
is insensitive to collisionality.
}
\label{critical}
\end{figure}
Note that while these linear simulations were done at zero flow shear, our nonlinear runs
had $\gamma_E = -0.003\,v_{te}/a$, a value representing the measured rotational shear in the MAST shot on which
the rest of our equilibrium parameters were based. The ETG instability in the
presence of flow shear can become subcritical in the sense that growth is transient
and all modes eventually decay (cf.\ \cite{Newton10,Highcock10,Barnes11,Highcock11,Sch12,vanWyk16}).
Given a finite initial perturbation, such transient growth still leads to a saturated nonlinear state
(because long enough transient growth can be as good as a formally unstable mode as far as
maintaining turbulence is concerned \cite{Highcock10,Barnes11,Highcock11,vanWyk16}),
although we have found the minimum value of $a/L_T$ required for this to be higher
than the critical gradient implied by Figure \ref{critical}: it is typically between
$3$ and $3.3$. The mapping out of a ``zero-turbulence manifold''
(cf.\ \cite{Highcock12,Ghim14}) for the ETG problem has been left outside the scope
of this study --- and would be a formidable computational challenge, requiring very many
very-long-time simulations. It appears plausible that the key parameters in such a study
would be the electron temperature gradient $a/L_T$ and collisionality $\nu_{ei}a/v_{te}$, rather
than the ion temperature gradient and the perpendicular flow shear
$\gamma_E$, which has a more profound effect on the turbulence threshold in the ITG problem \cite{Highcock12,Ghim14,vanWyk16}. The magnetic geometry factor $qR/r\approx B/B_p$, which appears in the scaling (\ref{Q_scaling}), may play a similar role --- improving confinement as this quantity is reduced --- in both cases, although for rather different reasons.
\section{Issues of numerical convergence in nonlinear simulations}
\label{convergence_study}
We discuss here the selection of runs used to construct Figure \ref{collscan}, and the associated numerical convergence issues. We also include additional evolution plots, complementing the two that were shown in Figure \ref{overlap}. This appendix concludes with a short discussion on the role of flow shear.
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.75,clip]{qescan_uncvgd_full_enorm_gB_noleg}
\label{qe_uncvgd}
}
\subfigure[]{
\includegraphics[scale=0.75,clip]{zvscan_uncvgd_full_enorm_nobox}
\label{zv_uncvgd}
}
\caption{Variation of \subref{qe_uncvgd} the time-averaged normalised electron heat flux $Q/Q_{\rm gB}$ and \subref{zv_uncvgd} the rms zonal velocity $k_\mathrm{Z} \phi_\mathrm{Z}$, defined by (\ref{kzphiz}), versus normalised electron collisionality $\nu a/v_{te}$ for our entire simulation
series using both small and large boxes, and two temperature gradients, including unresolved runs (hollow symbols; infilled symbols are the resolved points shown in Figure \ref{collscan}).}
\label{qescan_uncvgd}
\end{figure}
\begin{figure}[p]
\subfigure[]{
\includegraphics[scale=0.75]{qe_evolution_l2_enorm_gB}
\label{qe_evolution}
}
\subfigure[]{
\includegraphics[scale=0.75]{zp_evolution_l2_enorm}
\label{zp_evolution}
}
\subfigure[]{
\includegraphics[scale=0.75]{zv_evolution_l2_enorm}
\label{zv_evolution}
}
\subfigure[]{
\includegraphics[scale=0.75]{zs_evolution_l2_enorm}
\label{zs_evolution}
}
\subfigure[]{
\includegraphics[scale=0.75]{kz_evolution_l2_enorm}
\label{kz_evolution}
}
\caption{Evolution in time of \subref{qe_evolution} the normalised turbulent electron heat flux $Q/Q_{\rm gB}$, \subref{zp_evolution} the square of the zonal electrostatic potential, $\phi_{\mathrm Z}^2 = \sum_{k_x} |\phi_{k_x,0}|^2$, \subref{zv_evolution} the square of the zonal velocity, $(k_{\mathrm Z}\phi_{\mathrm Z})^2 = \sum_{k_x} k_x^2|\phi_{k_x,0}|^2$, \subref{zs_evolution} the square of the zonal shear, $\sum_{k_x} k_x^4|\phi_{k_x,0}|^2$, and \subref{kz_evolution} the rms zonal wavenumber $k_{\mathrm Z}$, for large-box
simulations at the nominal $a/L_T = 3.42$ and at various collisionalities (colours: same as Figure \ref{linear}, plus yellow for $\nu = 2\,\nu_{\mathrm{nom}}$).
}
\label{evolution}
\end{figure}
\subsection{Convergence of the heat flux}
Figure \ref{collscan} was based on a set of simulations picked from a larger parameter scan that we have carried out. The choice of simulations was based on whether we deemed them to be numerically converged. Here we explain how this was decided.
Figure \ref{qescan_uncvgd} shows the results of both small- and large-box simulations
(see \ref{numerics} for the explanation of what this means), over a range of collisionalities and for two values of the temperature gradient. Runs
excluded from Figure \ref{collscan} are shown as hollow points. They have been excluded because one
or more important physical quantity measured in these runs was found to be dominated by either the box scale (the lowest $k$) or by grid scales
(the highest $k$'s), and, therefore, the convergence of the simulation with respect to box size and/or spatial resolution could not be relied upon.
Two particular pathologies have been identified, and are also visible in Figure \ref{evolution}, which shows the time evolution of the full set of large-box simulations at the nominal temperature gradient ($a/L_T=3.42$).
(i) As the collisionality is reduced, it can be seen from the growth of the zonal velocity (Figure \ref{evolution}\subref{zv_evolution}) that the increasing zonal $\phi^2$ (Figure \ref{evolution}\subref{zp_evolution}) more than compensates
initially for the gradually falling zonal wavenumber (Figure \ref{evolution}\subref{kz_evolution}), until the final level of zonal velocity is reached. However, the zonal shear (Figure \ref{evolution}\subref{zs_evolution}) remains relatively constant in time.
For the highest collisionality (yellow), the heat flux
never collapses, and we believe that in this case it is because the required zonal wavenumber
is too close to the simulation box scale; see Figure \ref{evolution}\subref{kz_evolution}.
Such uncollapsed runs are visible as the high points in Figure \ref{qescan_uncvgd}\subref{qe_uncvgd}.
(ii) At the low end of the
collisionality scale in Figure \ref{evolution} (blue, red and green), it can be seen that there is no significant difference in the heat fluxes as the collisionality is varied,
leading to the plateau of hollow red squares in Figure \ref{qescan_uncvgd}\subref{qe_uncvgd} (and a similar plateau for the
other temperature gradient, blue squares). These runs are also excluded because of box-scale effects. In these
cases, somewhat counterintuitively, the small-box runs, which, however, extend to higher $k$'s, appear to be resolved and do not exhibit the plateau. Figure \ref{uncvgd_spectra} shows an example (the lowest-temperature-gradient and lowest-collisionality case), in which the perpendicular temperature is dominated by the box scale for the large-box run, which is excluded.
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.75]{zonal_spectra_uncvgd_enorm_cut}
\label{uncvgd_zonal_spectrum}
}
\subfigure[]{
\includegraphics[scale=0.75]{zonal_tperp_spectra_uncvgd_enorm_cut}
\label{uncvgd_zonal_tperp_spectrum}
}
\caption{Spectrum of \subref{uncvgd_zonal_spectrum} the zonal potential, \subref{uncvgd_zonal_tperp_spectrum} the zonal perpendicular temperature perturbation, for $\nu = 0.1 \,\nu_{\mathrm{nom}}$, $a/L_T = 3.3$, for a small-box simulation (green) and a large-box simulation (yellow), at the final simulation time in each case.
}
\label{uncvgd_spectra}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.9]{zp_overlap_enorm_cut_nobox}
\label{zp_overlap}
}
\subfigure[]{
\includegraphics[scale=0.9]{zs_overlap_enorm_cut_nobox}
\label{zs_overlap}
}
\subfigure[]{
\includegraphics[scale=0.9]{kz_overlap_enorm_cut_nobox}
\label{kz_overlap}
}
\caption{Evolution in time of \subref{zp_overlap} the square of the zonal electrostatic potential, $\phi_{\mathrm Z}^2 = \sum_{k_x}|\phi_{k_x,0}|^2$, \subref{zs_overlap} the square of the zonal shear,
$\sum_{k_x} k_x^4|\phi_{k_x,0}|^2$, and \subref{kz_overlap} the rms zonal wavenumber $k_{\mathrm Z}$, for $\nu = 0.2 \,\nu_{\mathrm{nom}}$, $a/L_T = 3.3$ (green: small-box simulations; yellow: large-box simulations). This complements Figure \ref{overlap}.}
\label{overlap2}
\end{figure}
Eliminating both pathologies gives a fairly clean collisionality scaling when the remaining resolved runs are combined.
Given that the zonal damping, according to the scaling demonstrated in section \ref{zonal_damping_section}, is lowest for
the box-scale mode, it is not surprising that there are convergence problems at this end of the spectrum. Nonetheless, all the simulations that we consider resolved have saturated with the nonlinear state dominated by zonal modes with wavenumbers above the box-scale mode.
Thus, to summarise, at higher collisionalities, we need larger boxes but can sometimes get away with less spatial
resolution, whereas at low collisionalities, we need
higher spatial resolution, but can sometimes get away with smaller boxes.
For completeness, we include as Figure \ref{overlap2} the time traces of various zonal quantities for the
simulations used
in Figure \ref{overlap}. When these two figures are combined, the complete set analogous to Figure \ref{evolution} is obtained. Note that these demonstrate some lack of convergence in the ``quasi-saturated'' state, but the long-time state is resolved.
Spatial cross-sections of the electrostatic potential for these two simulations are shown in Figure \ref{contours} (large-box) and Figure \ref{contours2} (small-box). The morphology of the quasi-saturated state is
qualitatively distinguishable between the two simulations. This difference is perhaps associated with the clear difference in the zonal shear shown in in Figure \ref{overlap2}\subref{zs_overlap}.
The small-box simulation has higher spatial resolution, i.e., a higher maximum $k$, and the higher power of $k_x$ in the zonal shear (compared to the zonal velocity) emphasises the high end of the spectrum. The long-time saturated states, on the other hand, appear qualitatively the same. Note that in the large-box simulation in Figure \ref{satslice}, there are seven periods of the dominant zonal wavelength across the box, whereas in the doubled small box in Figure \ref{contours2}\subref{satslice2x2} there can only be an even number, which is six (two times the three across each smaller periodic box). This level of discrepancy between the zonal wavenumbers must be expected given the finite domain size.
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.65,clip]{overlap_qsat_morexy_enorm}
\label{quasisatslice2x2}
}
\subfigure[]{
\includegraphics[scale=0.65,clip]{overlap_sat_morexy_enorm}
\label{satslice2x2}
}
\caption{Electrostatic potential $\phi$ at the outboard midplane, for $\nu = 0.2 \,\nu_{\mathrm{nom}}$, $a/L_T = 3.3$:
\subref{quasisatslice2x2} quasi-saturated state at $t = 1204.2 \,a/v_{te}$, \subref{satslice2x2} long-time saturated state at $t = 7841.5 \,a/v_{te}$, for small-box simulations showing $2\times 2$ copies of each periodic domain. The large-box version of this simulation is shown in Figure \ref{contours}.}
\label{contours2}
\end{figure}
\subsection{Background flow shear}
\label{shear}
Various authors have reported runaway growth in ETG turbulence simulations.
To aid saturation, Guttenfelder and Candy~\cite{GuttenfelderCandy}, following earlier work \cite{CandyPPCF2007},
added background flow shear to their simulations.
Roach et al.~\cite{Roach09}, extending the simulations of Joiner et al.~\cite{Joiner},
observed runaway growth that was healed by adding either flow shear or sufficient collisions
(attributed by them to the stabilisation of weak trapped-electron modes close to the box scale
at $k_y\rho_e\approx 0.02$).
In our simulations, we have included an experimentally realistic level of flow shear,
$\gamma_E= -0.003\,v_{te}/a$, and we do not find runaway behaviour.
Note that in all of the cases reported above, the normalised perturbed zonal shear (Figure \ref{evolution}\subref{zs_evolution} and Figure \ref{overlap2}\subref{zs_overlap}) is an order of magnitude larger than the background flow shear. In the ``simplified'' simulations of section \ref{zeffscansection}, zero flow shear was used, and it is reasonable to suppose that the fixed electron-electron collisions helped regularise these simulations over a wider range of collisionalities than was accessible when varying electron-ion
and electron-electron collisions together.
The simplified simulations support our view that the scaling of the heat flux with the collision rate that we have found does not depend directly on flow shear.
Of course, physically, in a real tokamak, flow shear is still needed to suppress the ion-scale turbulence, which we did not model.
\section{Long-time linear damping of the electron zonal modes}
\label{Felix}
In this Appendix, we derive the scaling (\ref{gammaZ}) of the zonal damping rate.
\subsection{Zonal evolution equation}
Consider the linearised form of the zonal gyrokinetic equation (\ref{GKE_Z}) (i.e.,
ignore its right-hand side). It is convenient to Fourier transform this equation
in the field-perpendicular coordinate $x$ (there is no dependence on $y$
by definition of the zonal modes), while leaving the field-parallel derivatives
in position space (the zonal modes here are just ones with $k_y=0$, but are allowed
to have parallel variation). This gives us
\begin{equation}
\frac{\partial}{\partial t}\left[h + J_0(k_\perp\rho)\frac{e\phi}{T} F\right]
+ v_\parallel\mathbf{b}\cdot\nabla h + ik_x v_{Bx} h =
\langle C[h e^{-ik_x\rho_x}]e^{ik_x\rho_x}\rangle,
\label{hzonal}
\end{equation}
where $J_0(k_\perp\rho)$ with $\rho = v_\perp/\Omega_e$ is the Fourier-space form of the
gyroaveraging operator, $k_\perp = k_x|\nabla x|$ (here the curvilinear coordinate
$x$ is defined by (\ref{defx}) and so $\nabla x = (q/B_0 r)\nabla\psi$),
$\rho_x = (\nabla x)\cdot(\mathbf{b}\times\mathbf{v}_\perp)/\Omega_e$, the angle brackets
$\langle\dots\rangle$ denote the averaging of the linearised collision operator over the gyroangle \cite{SCHEKOCHIHIN_APJS2009,Abel08}, and
\begin{equation}
v_{Bx} = \mathbf{v}_{B}\cdot\nabla x = \frac{q}{B_0 r} \mathbf{v}_{B}\cdot\nabla \psi,
\label{vbx_def}
\end{equation}
where $\mathbf{v}_{B}$ is given by (\ref{vB}).
The quasineutrality equation (\ref{QN2}) becomes
\begin{equation}
\frac{e\phi}{T}\left(1 + \frac{1}{\tau}\right) = - \frac{1}{n}\int d^3\mathbf{v} J_0(k_\perp\rho) h.
\label{phizonal}
\end{equation}
\subsection{Long-wavelength ordering and time scales}
Assuming that the zonal modes have long wavelengths, $k_x\rho_e\ll1$,
we seek the solution to equations (\ref{hzonal}) and (\ref{phizonal}) in the form
of an asymptotic expansion
\begin{equation}
h = h^{(0)} + h^{(1)} + h^{(2)} + \dots,
\end{equation}
where $h^{(n)}\sim(k_x\rho_e)^n h^{(0)}$.
In this ordering, $J_0(k_\perp\rho)\approx 1$ to lowest order and
clearly, from (\ref{phizonal}), to lowest order, $e\phi/T \sim h^{(0)}/F$.
We introduce a formal ordering of time scales in (\ref{hzonal}):
\begin{equation}
\fl
v_\parallel\mathbf{b}\cdot\nabla \sim \frac{v_{te}}{a} \sim \nu,\quad
k_x v_{Bx} \sim \frac{k_x}{\Omega_e}\frac{v_{te}^2}{a} = k_x\rho_e \frac{v_{te}}{a}
\sim (k_x \rho_e) \nu, \quad
\frac{\partial}{\partial t} \sim \gamma_\mathrm{Z} \sim (k_x\rho_e)^2\nu.
\label{ordering}
\end{equation}
This ordering anticipates the kind of solution we are expecting.
Generally speaking, any initial zonal perturbation will evolve on three successive
time scales: first, it will be damped collisionlessly by the streaming and
magnetic-drift terms on the time scale $\sim a/v_{te}$, leaving a finite
residual perturbation \cite{RH}; then this residual, which is non-Maxwellian,
will be damped by collisions on the time scale $\sim 1/\nu$
\cite{HR,Kim,Xiao,Xiao07}, leaving a residual perturbed Maxwellian
(this will be our $h^{(0)}$); and finally, this perturbed Maxwellian
will be damped diffusively at the rate $\gamma_\mathrm{Z}$, as we are
about to show.
Thus, the zonal perturbation whose evolution we are going to calculate
will be close to a steady state, with its evolution essentially governed
by the (neoclassical) collisional transport theory.
\subsection{Zeroth order}
\label{order0}
With the ordering (\ref{ordering}), at zeroth order, (\ref{hzonal}) becomes
\begin{equation}
v_\parallel \mathbf{b}\cdot\nabla h^{(0)} = C[h^{(0)}].
\label{zerothorder}
\end{equation}
If we multiply this equation by $h^{(0)}/F$ and flux-surface average,
the left-hand side vanishes (see, e.g., \cite{Abel13}, \S 6.1)
and we get
\begin{equation}
-\overline{\int d^3\mathbf{v} \frac{h^{(0)} C_{ee}[h^{0}]}{F}}
+ \overline{\int d^3\mathbf{v} \frac{\nu_{ei}v_{te}^3}{v^3 F}\frac{1-\xi^2}{2}
\left(\frac{\partial h^{(0)}}{\partial\xi}\right)^2} = 0,
\label{vanishing}
\end{equation}
where we have denoted the flux-surface average by an overbar, used the collision
operator (\ref{model_e}) taken to lowest order in $k_\perp\rho_e$ and
integrated by parts in the term involving electron-ion collisions
(cf.\ \cite{SCHEKOCHIHIN_APJS2009}, \S 4.2). Both terms in (\ref{vanishing})
are non-negative definite and so must vanish individually.
The vanishing of the first of these implies that $h^{(0)}$ must be
a perturbed Maxwellian, the vanishing of the second that this perturbed Maxwellian
has no mean parallel flow. Therefore,
\begin{equation}
h^{(0)} = \left[-\frac{e\phi}{T}\left(1 + \frac{1}{\tau}\right)
+ \left(\frac{v^2}{v_{te}^2}-\frac{3}{2}\right)\frac{\delta T}{T}\right] F,
\label{h0}
\end{equation}
where the density perturbation associated with $h^{(0)}$ has been fixed
by the requirement that, to lowest order, $h^{(0)}$ and $\phi$ must satisfy (\ref{phizonal}).
Note that we omit the superscript $(0)$ on $\phi$ and $\delta T$ because we will not need to calculate
these fields explicitly to any higher order.
Finally, substituting (\ref{h0}) back into (\ref{zerothorder}) and using the fact
that, $h^{(0)}$ being a flow-less perturbed Maxwellian, $C[h^{(0)}]=0$, we find that both
$\phi$ and $\delta T$ must be flux functions to zeroth order:
\begin{equation}
\mathbf{b}\cdot \nabla\frac{e\phi}{T} = \mathbf{b}\cdot \nabla \frac{\delta T}{T} = 0.
\end{equation}
The zeroth-order zonal solution (\ref{h0}) is precisely the promised
quasi-steady state to which any initial zonal perturbations will relax after initial collisionless
and collisional transients.
\subsection{First order}
\label{order1}
At first order in $k_x\rho_e$, equation (\ref{hzonal}) is
\begin{equation}
v_\parallel\mathbf{b}\cdot\nabla h^{(1)} + ik_x v_{Bx} h^{(0)} = C[h^{(1)}].
\label{firstorder}
\end{equation}
Note that any gyroaveraging or FLR effects in the collision operator will only appear
in the second order.
In dealing with (\ref{firstorder}), it is convenient to take advantage of the fact that in
tokamaks~\cite{HintonWong},
\begin{equation}
\mathbf{v}_B\cdot\nabla\psi = - v_\parallel\mathbf{b}\cdot\nabla\left(\frac{v_\parallel I}{\Omega_e}\right),
\label{drifttrick}
\end{equation}
where $I(\psi) = B_T R$ is a flux function, $B_T$ is the toroidal magnetic field and $R$ the major
radius coordinate. Note that $\nabla$ on the right-hand side of (\ref{drifttrick}) is taken at constant $\mathcal{E}$ and $\mu$, which, owing to the variation of $B$, is not at constant $v_\parallel$. Using (\ref{vbx_def}) and recalling that the safety factor $q(\psi)$ and the
radial flux-surface label $r(\psi)$ are both flux functions and the reference magnetic
field $B_0$ is a constant, we have
\begin{equation}
v_{Bx} = - v_\parallel\mathbf{b}\cdot\nabla\left(\frac{v_\parallel}{\Omega_e}\frac{q I}{B_0 r}\right)
= - v_\parallel\mathbf{b}\cdot\nabla\left(\rho_{pe} \frac{v_\parallel}{v_{te}}\right),
\label{vbx}
\end{equation}
where $\rho_{pe} = (qI/B_0 r)\rho_e$ is, by definition, the ``poloidal gyroradius'',
so called because $qI/B_0 r \sim B/B_p$,
where $B_p$ is the poloidal magnetic field.
Note that $\Omega_e$ and, therefore, $\rho_e = v_{te}/\Omega_e$, is calculated using
the total, space-dependent field $B$, and so $\rho_{pe}$ is {\em not} a flux function.
Since $h^{(0)}$ is a flux function ($\mathbf{b}\cdot\nabla h^{(0)}=0$),
we may then rewrite (\ref{firstorder}) as follows
\begin{equation}
v_\parallel\mathbf{b}\cdot\nabla\left( h^{(1)} - ik_x\rho_{pe}\frac{v_\parallel}{v_{te}} h^{(0)}\right)
= C[h^{(1)}].
\label{h1}
\end{equation}
The solution of this equation is a standard problem in neoclassical theory
\cite{HintonWong,Helander}, but we shall not require its explicit form
in order to work out the scalings that we seek. Note from (\ref{h1}) that
the first-order zonal perturbation has a parallel electron flow
(and, therefore, current) of order
\begin{equation}
\frac{u_\parallel}{v_{te}} \sim k_x\rho_{pe} \frac{\delta p}{p},
\end{equation}
where $\delta p/p$ is the zonal pressure perturbation associated with $h^{(0)}$.
The perpendicular $\mathbf{E}\times\mathbf{B}$ zonal flow is, of course, just
\begin{equation}
\frac{v_{Ey}}{v_{te}} \sim k_x\rho_e \frac{e\phi}{T}.
\end{equation}
It is the resistive damping of zonal flows (currents) that will lead to the damping
of the zonal modes.
\subsection{Second order}
In order to calculate this damping, we need evolution equations for the zeroth-order
fields $\phi$ and $\delta T$. The time derivatives enter at the second order in (\ref{hzonal}).
Using (\ref{h0}) and (\ref{vbx}), we have at this order:
\begin{eqnarray}
\nonumber
& \frac{\partial}{\partial t}\left[-\frac{1}{\tau} \frac{e\phi}{T}
+ \left(\frac{v^2}{v_{te}^2}-\frac{3}{2}\right)\frac{\delta T}{T}\right] F
+ v_\parallel\mathbf{b}\cdot\nabla h^{(2)} - h^{(1)}
v_\parallel\mathbf{b}\cdot\nabla\left(ik_x\rho_{pe}\frac{v_\parallel}{v_{te}}\right) \\
& = k_x^2 \langle C[h^{(0)}\rho_x]\rho_x\rangle +
\left\langle C\left[h^{(2)} - \frac{1}{2}k_x^2\rho_x^2h^{(0)}\right] \right\rangle\vphantom{\biggl(},
\end{eqnarray}
where the leading-order FLR parts of the collision operator acting on $h^{(0)}$
have appeared. We now transit-average this equation, integrate
the magnetic-drift term by parts and use (\ref{h1}) to express
$v_\parallel\mathbf{b}\cdot\nabla h^{(1)}$.
The result, with the transit-average denoted by overbar, is
\begin{equation}
\fl
\frac{\partial}{\partial t}\left[-\frac{1}{\tau} \frac{e\phi}{T}
+ \left(\frac{v^2}{v_{te}^2}-\frac{3}{2}\right)\frac{\delta T}{T}\right] F
= \overline{- i k_x \rho_{pe} \frac{v_\parallel}{v_{te}} C[h^{(1)}]
+ k_x^2 \langle C[h^{(0)}\rho_x]\rho_x\rangle + \langle C[\dots] \rangle}.
\label{secondorder}
\end{equation}
To separate the evolution of $\phi$ and $\delta T$, we take the density
and energy moments of (\ref{secondorder}).
Integrating it over velocities, and noting that,
by conservation of particles, $\int d^3\mathbf{v}\langle C[\dots]\rangle = 0$,
we get\footnote{The last term in (\ref{slow_phi}) is related to the spatial FLR diffusion term
that we wrote out explicitly in the gyroaveraged collision operator (\ref{model_e}).}
\begin{equation}
-\frac{n}{\tau} \frac{\partial}{\partial t} \frac{e\phi}{T}
= \overline{-i k_x \rho_{pe} \int d^3\mathbf{v} \frac{v_\parallel}{v_{te}} C_{ei}[h^{(1)}]
+ k_x^2\rho_e^2\int d^3\mathbf{v} \frac{C_{ei}[h^{(0)}\rho_x]\rho_x}{\rho_{e}^2}}.
\label{slow_phi}
\end{equation}
Only the electron-ion collision terms have survived because, to lowest order,
conservation of momentum by the electron-electron collisions implies
\begin{equation}
\int d^3\mathbf{v} v_\parallel C_{ee}[h^{(1)}] = 0, \quad
\int d^3\mathbf{v} C_{ee}[h^{(0)}\rho_x] \rho_x = 0.
\end{equation}
In the same vein, multiplying (\ref{secondorder}) by $v^2/v_{te}^2$ and integrating over velocities,
we find, using energy conservation by collisions, $\int d^3\mathbf{v} v^2\langle C[\dots]\rangle = 0$,
\begin{equation}
\fl
\frac{3n}{2} \frac{\partial}{\partial t}\frac{\delta T}{T}
= \overline{-i k_x \rho_{pe} \int d^3\mathbf{v} \frac{v_\parallel}{v_{te}}
\left(\frac{v^2}{v_{te}^2}-\frac{3}{2}\right) C[h^{(1)}]
+ k_x^2\rho_e^2\int d^3\mathbf{v} \frac{C[h^{(0)}\rho_x]\rho_x}{\rho_e^2}
\left(\frac{v^2}{v_{te}^2}-\frac{3}{2}\right)}.
\label{slow_dT}
\end{equation}
The difference between this equation and (\ref{slow_phi}) is that the
contributions from electron-electron collisions do not vanish
(because same-species collisions can support non-zero heat fluxes).
\subsection{Damping rate}
We do not need to solve the neoclassical equation (\ref{h1}) for $h^{(1)}$ explicitly
to see that, this equation being linear, $h^{(1)}$ will be a linear combination of
$e\phi/T$ and $\delta T/T$ with velocity-dependent coefficients, all of which
are proportional to $k_x\rho_{pe}$; any homogeneous part of $h^{(1)}$ satisfies
(\ref{zerothorder}) and so can be absorbed into $h^{(0)}$.
Therefore, we may schematically represent (\ref{slow_phi}) and (\ref{slow_dT})
as the following matrix equation at this order:\footnote[1]{Note that in the absence of magnetic drifts, in a straight field,
$h^{(1)}=0$ and we must replace $\rho_{pe}$ with $\rho_e$ everywhere, with the
collisional evolution of the zonal modes now controlled by the FLR spatial-diffusion terms
in the gyrokinetic collision operator.}
\begin{eqnarray}
\label{phi_schematic}
\frac{\partial}{\partial t} \frac{e\phi}{T} &=
\nu_{ei} k_x^2\rho_{pe}^2\left( a_{11} \frac{e\phi}{T} + a_{12} \frac{\delta T}{T} \right)\\
\frac{\partial}{\partial t} \frac{\delta T}{T} &=
\nu_{ei} k_x^2\rho_{pe}^2\left( a_{21} \frac{e\phi}{T} + a_{22} \frac{\delta T}{T} \right)
+ \nu_{ee} k_x^2\rho_{pe}^2\left( b_{21} \frac{e\phi}{T} + b_{22} \frac{\delta T}{T} \right)
\label{dT_schematic}
\end{eqnarray}
where $a_{ij}$ and $b_{ij}$ are dimensionless coefficients that depend only
on equilibrium parameters. Clearly, if $\nu_{ei}=0$, the matrix is defective (has a zero row corresponding to
the upper equation (\ref{phi_schematic})) and the damping rate of both the
potential and temperature perturbations of the slowest damped eigenmode vanishes. Therefore,
\begin{equation}
\gamma_\mathrm{Z} \propto \nu_{ei} k_x^2 \rho_{pe}^2.
\end{equation}
The order-unity numerical prefator in the exact expression for $\gamma_\mathrm{Z}$
depends on the ratio $\nu_{ee}/\nu_{ei}$, which itself is order unity.
Thus, we have shown that the long-time damping rate of the zonal modes always
satisfies~(\ref{gammaZ}), a scaling that is indeed well reproduced in our
numerical simulations (Figure \ref{zonal_damping}).
\subsection{Zonal damping by same-species collisions}
\label{intraspecies}
The result (\ref{gammaZ}) survives even if, artifically, the ratio $\nu_{ee}/\nu_{ei}$
is made large --- as was done in the ``simplified'' simulations of section \ref{zeffscansection}.
In this case, (\ref{dT_schematic}) simply implies
that
\begin{equation}
\frac{\delta T}{T} \approx -\frac{b_{21}}{b_{22}}\frac{e\phi}{T},
\label{hardcouple}
\end{equation}
and, according to (\ref{phi_schematic}),
the zonal perturbation is still damped at the
rate $\gamma_\mathrm{Z} \sim \nu_{ei} k_x^2 \rho_{pe}^2$.
If we consider an even more artificial situation in which $\nu_{ei}$ is made to vanish
completely, we must go to next order when calculating the density moment (\ref{slow_phi})
of the zonal kinetic equation (\ref{hzonal}) to allow higher-order finite-drift-orbit-width
FLR terms to come in. These terms will be small by an extra
factor of $(k_x\rho_{pe})^2$ and so technically we would have to reorder the time
derivative two powers of $k_x\rho_e$ higher and then work through two
extra orders in our expansion.\footnote{This was indeed confirmed
independently by such a calculation \cite{Jack_note}.} It is clear though that the result of this
amounts to replacing (\ref{phi_schematic}) and (\ref{dT_schematic}) with
\begin{eqnarray}
\label{phi_schematic2}
\frac{\partial}{\partial t} \frac{e\phi}{T} &=
\nu_{ee} k_x^4\rho_{pe}^4\left( b_{11} \frac{e\phi}{T} + b_{12} \frac{\delta T}{T} \right),\\
\frac{\partial}{\partial t} \frac{\delta T}{T} &=
\nu_{ee} k_x^2\rho_{pe}^2\left( b_{21} \frac{e\phi}{T} + b_{22} \frac{\delta T}{T} \right),
\label{dT_schematic2}
\end{eqnarray}
using (\ref{dT_schematic2}) to infer again the coupling (\ref{hardcouple})
between the zonal potential and temperature perturbations, and finally arriving at
a zonal damping rate
\begin{equation}
\gamma_\mathrm{Z} \sim \nu_{ee} k_x^4\rho_{pe}^4.
\end{equation}
This is the scaling that is indeed obtained numerically when $\nu_{ei}$ is switched off
(black crosses in Figure~\ref{zonal_damping}).
These considerations might appear moot, as $\nu_{ei}=0$ is unphysical and achievable only in numerical experiments.
They do, however, help us gain some insight into the way in which our theoretical aruments
would be modified if we applied them to zonal dynamics in ITG rather than ETG turbulence,
as suggested in section \ref{ITG}.
\subsection{Case of ion zonal modes}
\label{ionzfs}
The only differences between the linearised gyrokinetic equation for the ion zonal modes
and (\ref{hzonal}) are a minus sign in front of the $e\phi/T$ term, accounting for the ion charge,
and the absence of interspecies collisions:
\begin{equation}
\frac{\partial}{\partial t}\left[h_i - J_0(k_\perp\rho)\frac{e\phi}{T_i} F\right]
+ v_\parallel\mathbf{b}\cdot\nabla h_i + ik_x v_{Bx} h_i =
\langle C_{ii}[h_i e^{-ik_x\rho_x}]e^{ik_x\rho_x}\rangle.
\label{hizonal}
\end{equation}
Another very important difference between the ion and electron cases
is the Boltzmann-electron closure (\ref{ai2}),
which removes the flux-surface-averaged part of the zonal potential from the electron-density
response. Combining (\ref{ai2}) with the expression (\ref{dnn}) for the ion density
perturbation in terms of $h_i$, we get the following version of the quasineutrality
equation:
\begin{equation}
\frac{e\phi}{T_i} + \frac{e(\phi - \overline{\phi})}{T_e}
= \frac{1}{n}\int d^3\mathbf{v} J_0(k_\perp\rho) h_i.
\label{QNizonal}
\end{equation}
That the difference between (\ref{QNizonal}) and (\ref{phizonal}) is significant
becomes obvious if we take the density moment of (\ref{hizonal}), i.e.,
integrate it over velocities keeping $\mathbf{r}$ constant (equivalently,
multiply by $J_0(k_\perp\rho)$ and integrate over $\mathbf{v}$), and then
flux-surface average. To lowest order in $k_\perp\rho_i$,
the term in the square brackets becomes, using (\ref{QNizonal}),
\begin{equation}
\overline{\frac{1}{n}\int d^3\mathbf{v} J_0(k_\perp\rho) \left[h_i - J_0(k_\perp\rho)\frac{e\phi}{T_i} F\right]}
\approx \frac{1}{2}k_\perp^2\rho_i^2 \frac{e\overline{\phi}}{T_i}.
\label{ddti}
\end{equation}
This has two extra powers of $k_x$ compared to the analogous term for electron zonal
flows (see the left-hand side of (\ref{slow_phi})).
By a calculation analogous to \ref{order0}, the zeroth-order solution $h_i^{(0)}$
of (\ref{hizonal}) in the long-time limit must again be a quasi-steady
(i.e., slow-evolving), constant on flux surfaces, flow-less perturbed Maxwellian.
The rest of the calculation
proceeds similarly to the electron case with the exception that only momentum-conserving
ion-ion collisions are present and so the velocity integral of the collision terms
in the density equation will be smaller by an extra factor of $k_x^2\rho_{pi}^2$,
as explained in \ref{intraspecies}.
In view of (\ref{ddti}), the analog of (\ref{phi_schematic2}) for ion zonal modes will then be
\begin{equation}
\frac{\partial}{\partial t}
\frac{1}{2}k_x^2\rho_{pi}^2 \frac{e\phi}{T_i} = \nu_{ii} k_x^4\rho_{pi}^4
\left(c_{11} \frac{e\phi}{T_i} + c_{12} \frac{\delta T_i}{T_i}\right),
\end{equation}
where we have absorbed all factors accounting for differences between
$k_\perp$ and $k_x$ and between $\rho_i$ and $\rho_{pi}$ into the
dimensionless coefficients $c_{11}$ and $c_{12}$.
The extra factors of $k_x^2\rho_{pi}^2$ on the left-hand side
(due to the special role of flux-surface-averaged potential in the Boltzmann-electron closure)
and on the right-hand side (due to momentum conservation by the ion-ion collision operator)
cancel, leaving $e\phi/T_i$ to evolve on the diffusive time scale $\sim 1/\nu_{ii} k_x^2\rho_{pi}^2$.
The evolution equation for $\delta T_i/T_i$ does not have
these extra factors on either side and so feature the same time scale. Thus, the ion zonal
modes, similarly to the electron ones (although for a different reason),
will be damped at a rate
\begin{equation}
\gamma_{\mathrm{Z} i} \sim \nu_{ii} k_x^2 \rho_{pi}^2.
\end{equation}
We remind the reader that the above calculation, like the whole of \ref{Felix}, is concerned solely with the linear physics of the zonal modes considered on their own. Its significance nonlinearly will depend upon the interactions between these modes and the nonzonal modes. In the main part of the paper, we report nonlinear results for ETG turbulence only, but we do discuss ITG turbulence further, in this wider nonlinear context, in the concluding section \ref{ITG}.
\section{Introduction}
\label{etg-introduction}
Experiments on MAST in which heat transport is dominated by the electron channel find that the thermal energy confinement time $\tau_E$ varies with the normalised electron collisionality $\nu_{\ast}$ according to the scaling \cite{Valovic}
\begin{equation}
B\tau_E \propto\nu_{\ast}^{-0.82\pm 0.1},
\label{MASTscaling}
\end{equation}
where $B$ is the magnetic field.
This scaling is favourable towards improved confinement in the hotter, lower collisionality plasmas
anticipated in future devices.
In this paper, we investigate how the electron heat flux $Q$ varies with electron collisionality in
simulations of electron-scale plasma turbulence in MAST, using the local gyrokinetic flux-tube code GS2 \cite{KOTSCHENREUTHER1995, GS2}.
At constant geometry and $\rho_\ast = \rho_e/a$ ($\rho_e$ is the electron Larmor radius, $a$ is the equilibrium length scale),
it can be shown (see \ref{scaling_correspondence}) that
\begin{equation}
B\tau_E \propto \left(\frac{Q}{Q_{\rm gB}}\right)^{-1},
\end{equation}
where $Q_{\rm gB} = n T v_{te} \rho_\ast^2$ is the electron gyroBohm heat flux.
It therefore ought to be possible to recover the scaling (\ref{MASTscaling}) from a local calculation of the electron heat flux.
We wish to discover whether this experimental MAST scaling may be understood in terms of electron
temperature gradient (ETG) driven turbulent transport.
With this goal in mind, we carry out local gyrokinetic simulations restricted to electrostatic
perturbations at electron-gyroradius scales, ignoring both electromagnetic (e.g., ``microtearing''
\cite{Doerk11,Gutten11})
modes and ion-gyroradius-scale effects. The restriction to electrostatic perturbations is a matter of considering as simple a model as possible for a plasma where beta is low and magnetic perturbations are relatively small. The neglect of the ion-gyroradius-scale turbulence is justifiable on the grounds that,
in typical MAST plasmas, ion turbulence
is considerably suppressed by radial shear in the background flow, and ion transport is close to the neoclassical
level \cite{Field04, Roach05, FieldIAEA2010, FieldNF2011}. This is fortunate, as spanning both ion- and electron-gyroradius scales requires prohibitively large computational resources.
We further limit
ourselves here to a Boltzmann-ion model, at electron scales, treating only the electrons kinetically. Whilst greatly
simplified, the equations we solve (which are described in more detail in section \ref{setup}) reproduce the experimental scaling of electron heat flux with collisionality.
\begin{figure}[b]
\centering
\subfigure[]{
\includegraphics[scale=0.65,clip]{overlap_qsat_l2_enorm2}
\label{quasisatslice}
}
\subfigure[]{
\includegraphics[scale=0.65,clip]{overlap_sat_l2_enorm2}
\label{satslice}
}
\caption{Non-dimensionalised electrostatic potential $e\phi/T\rho_\ast$ (where $\rho_\ast=\rho_e/a$) at the outboard midplane, for $\nu = 0.2 \,\nu_{\mathrm{nom}}$ (here $\nu_\mathrm{nom}$ is the ``nominal'', i.e., experimental value of collisionality), $a/L_T = 3.3$: \subref{quasisatslice} quasi-saturated state at $t = 1200.3\,a/v_{te}$, \subref{satslice} saturated state at $t = 7835.8\,a/v_{te}$, for large-box simulations. See \ref{plasma_parameters} and \ref{numerical_parameters} for the meaning of symbols.}
\label{contours}
\end{figure}
The reference simulation parameters are based on experiment, which is close to the threshold for the onset of turbulence. In this region of parameter space, we find that the saturated turbulent heat flux varies with collisionality in a manner consistent with the experimental scaling. This scaling is only revealed, however, if the simulation times are sufficiently long to reach a true steady state, which requires them to be much longer than the electron-collision time scale. At earlier times, there is a transient ``quasi-saturated'' state with higher heat flux, in which the zonal modes (which do not themselves contribute to the heat flux) are small but slowly growing. When they have grown to a sufficient level, the nonzonal modes and the heat flux are significantly suppressed.
Figure \ref{contours} illustrates these two regimes by showing the electrostatic potential (which is proportional to the density perturbation) in an outboard-midplane cross-section of a flux tube in MAST both in the earlier quasi-saturated state and the later long-time saturated state, based on one of the simulations reported below. In the quasi-saturated state, the zonal modes do not appear to play a special role, and radially extended ``streamers'' can be seen, as is usually expected for ETG turbulence \cite{Dorland00, Jenko00}. In contrast, in the long-time saturated state, a strong zonal component comes to dominate, structuring the turbulence into
``vortex streets'' and dramatically weakening radial transport.
In this final saturated state, the nonlinear drive of the zonal modes is balanced by their weak collisional damping, dominated by electron-ion collisions. Scans in collisionality reveal that the saturated heat flux increases with increasing collisionality, in rough proportionality: $Q/Q_{\rm gB}\propto\nu_{\ast }$.
A brief outline of the rest of the paper is as follows. In section \ref{setup}, we describe the equations that are solved and the simulation set-up.
In section \ref{results}, we present our main results, including the long-time evolution of the turbulence, the dependence of the saturated heat flux on collisionality, and the structure of the
saturated turbulent state.
We also sketch a simple theoretical argument that explains the collisionality scaling
of the heat flux (section \ref{twiddle}). In section \ref{discussion}, a summary of our findings
is given, our results are put in the context of some earlier work, and a discussion is given
of the apparent differences and possible similarities between the ETG and ITG turbulent states
in light of the conclusions of the present study.
\section{Governing equations and numerical set-up}
\label{setup}
Our study is based on the MAST H-mode shot 8500, which had 2 MW of NBI heating, and for which data are available
from the International Tokamak Profile Database \cite{Roach08}. This shot was analysed by Field et al. \cite{Field04}, and a linear gyrokinetic study was performed by Roach et al. \cite{Roach05}. In the present work, we consider a single flux surface, for which the detailed plasma parameters are given in \ref{plasma_parameters} (they are referred to as ``nominal'' parameters); these were kept fixed throughout our study, except for varying collisionality and electron temperature gradient where indicated.
In \ref{numerical_parameters}, we provide the technical details about our numerical simulations: the coordinate system used, numerical grids, resolution and boundary conditions.
\subsection{Gyrokinetic equation}
We use the GS2 continuum gyrokinetic code \cite{GS2} to obtain the perturbed distribution function and electrostatic field in local flux-tube geometry.
The electron distribution function is written
\begin{equation}
f = F + \delta f
\end{equation}
(we omit species subscripts when this will cause no confusion), where
\begin{equation}
F = n\left(\frac{m}{2\pi T}\right)^{3/2}\exp\left(-\frac{mv^2}{2T}\right)
\end{equation}
is the equilibrium Maxwellian background distribution, $n$, $m$, $T$ and $\mathbf v$ are the electron density, mass, temperature and velocity, respectively, and the perturbed distribution function is split into a Boltzmann response associated with the perturbed electrostatic potential $\phi$ and a gyrotropic non-Boltzmann, generally non-Maxwellian part:
\begin{equation}
\delta f = \frac{e\phi(\mathbf{r})}{T} F+h(\mathbf{R},v_\perp,v_\parallel),
\label{dfe}
\end{equation}
where $e$ is the absolute value of the electron charge.
Note that $\phi$ is a function of the particle position $\mathbf r$, whereas $h$ is
a function of the gyrocentre position $\mathbf R=\mathbf r - \mathbf b\times\mathbf v/\Omega_e$,
where $\mathbf{b}$ is the unit vector along the magnetic field and $\Omega_e=-eB/mc$ is the electron
cyclotron frequency (its sign is the negative sign of the electron charge). The gyrocentre distribution $h$ is otherwise independent of the gyroangle,
being a function of the parallel $v_\parallel = {\mathbf v}\cdot\mathbf b$ and
perpendicular $v_\perp = (v^2 - v_\parallel^2)^{1/2}$ velocities.
Equivalently, its velocity-space variables can be (and are) chosen to be
the energy ${\cal E} = v^2/2$ and magnetic moment $\mu = v_\perp^2/2B$,
so $v_\parallel = \pm(2{\cal E} - 2\mu B)^{1/2}$. In this representation,
the gyrocentre distribution $h$ satisfies the electrostatic gyrokinetic equation (GKE)
\cite{Frieman82} (see review \cite{Abel13})
\begin{equation}
\frac{\partial}{\partial t}\left(h + \frac{e\langle\phi\rangle}{T}F\right)
+\left(v_\parallel\mathbf{b}+\mathbf{v}_B\right)\cdot\nabla h
+\langle\mathbf{v}_E\rangle\cdot\nabla h +\langle\mathbf{v}_E\rangle\cdot\nabla F=
\langle C[h]\rangle,
\label{GKE}
\end{equation}
where $\langle\dots\rangle$ denotes gyroaveraging at constant gyrocentre position $\mathbf R$,
\begin{equation}
\mathbf{v}_B= \frac{\mathbf{b}}{\Omega_e}\times\left(v_\parallel^2 \mathbf{b}\cdot\nabla\mathbf{b}
+ \frac{v_\perp^2}{2}\frac{\nabla B}{B}\right)
\label{vB}
\end{equation}
is the magnetic drift velocity, and
\begin{equation}
\mathbf{v}_E = \frac{c}{B}\mathbf{b}\times \nabla\phi
\label{ExB}
\end{equation}
is the $\mathbf{E\times B}$ drift velocity. The energy is injected into the system
via the last term on the left-hand side of (\ref{GKE}), which contains the radial gradients
of the equilibrium distribution:
\begin{equation}
\langle\mathbf{v}_E\rangle\cdot\nabla F =
\langle v_{Ex}\rangle\frac{\partial F}{\partial x}
= \langle v_{Ex}\rangle
\left[\frac{1}{L_n} + \left(\frac{v^2}{v_{te}^2} - \frac{3}{2}\right)\frac{1}{L_T}\right] F,
\end{equation}
where $v_{te}=(2T/m)^{1/2}$ is the electron thermal speed. We have defined
$L_n= -d\ln n/dx$ and $L_T = -d\ln T/dx$, the
gradient scale lengths of the equlibrium electron density and temperature profiles,
respectively. The natural normalisation of these lengths
is the tokamak minor radius $a$. In the above definitions,
$x$ is the spatial coordinate transverse to the flux surface
and $y$ is a coordinate within the flux surface transverse to the
magnetic field (at the outboard midplane, it is approximately the poloidal arc length) --- these curvilinear, non-orthogonal coordinates,
as used by GS2, are explained in \ref{coords}.
Note that, in these coordinates,
\begin{equation}
v_{Ex} = \mathbf{v}_{E}\cdot\nabla x = \frac{c q}{B_0^2 r}\frac{d\psi}{d r}
\frac{\partial \phi}{\partial y}
\label{vEx}
\end{equation}
(see \ref{coords} for the definition of all symbols).
\subsection{Adiabatic ions}
In a simple (two-species, hydrogenic, i.e., with ion charge $=+e$) plasma, the quasineutrality condition for the perturbations
is
\begin{equation}
\delta n_e = \delta n_i,
\label{QN}
\end{equation}
where
$\delta n = \int d^3\mathbf{v}\delta f$ is the density perturbation for each species.
In terms of $h$, this density perturbation is
\begin{equation}
\frac{\delta n}{n} = -\frac{Z e\phi}{T}+\frac{1}{n}\int d^3\mathbf{v}\langle h\rangle_{\mathbf{r}},
\label{dnn}
\end{equation}
where $\langle\cdot\rangle_{\mathbf{r}}$ denotes the gyroaveraging operator at constant
particle position $\mathbf r$ (needed here because the velocity integral must be performed
at fixed particle position $\mathbf r$, while $h$ is a function of $\mathbf R$),
and $Z=1$ for the ions and $-1$ for the electrons.
We have assumed a decomposition of the perturbed ion distribution function
analogous to (\ref{dfe}).
If we consider only perturbations at scales
perpendicular to the magnetic field that are much smaller than the ion gyroradius,
$k_\perp \rho_i \gg 1$, the non-Boltzmann part of the ion density response
in (\ref{dnn}) can be neglected because $\langle h_i\rangle_{\mathbf{r}}$
is suppressed by the averaging over large ion Larmor orbits.
In formal terms, this approximation is the lowest-order
expansion in the electron-ion mass ratio. Therefore, in the present study, we will
assume that the ion distribution is entirely described by the
Boltzmann (sometimes called adiabatic) response:
\begin{equation}
\frac{\delta n_i}{n_i} = -\frac{e\phi}{T_i}.
\label{ai}
\end{equation}
Combining this with the quasineutrality (\ref{QN}) and the full gyrokinetic
electron density response given by (\ref{dnn}) with $Z=-1$,
we get the following equation for $\phi$:
\begin{equation}
\frac{e\phi}{T_e}\left(1+\frac{1}{\tau}\right) =-\frac{1}{n}\int d^3\mathbf{v}\langle h\rangle_{\mathbf{r}},
\label{QN2}
\end{equation}
where $h$ satisfies (\ref{GKE}) and $\tau=T_i/T_e$.
Equations (\ref{GKE}) and (\ref{QN2}) constitute a closed system, which is solved numerically
by the version of the GS2 code that we use here.
\subsection{Collisions}
\label{colls}
The electron collision operator used in GS2, appearing on the right-hand side of (\ref{GKE}), consists of the electron-ion pitch-angle scattering operator (to lowest order in the mass-ratio expansion) and a particle-, momentum- and energy-conserving simplified model of the electron-electron collision operator \cite{Roach05,Abel08,Barnes09}: in $\mathbf{k}_\perp$ space, it is
\begin{equation}
\fl
\langle C[h]\rangle_{\mathbf{k}_\perp} =
\langle C_{ee}[h]\rangle_{\mathbf{k}_\perp} + \nu_{ei}\frac{v_{te}^3}{v^3}
\left[
\frac{\partial}{\partial\xi}\frac{(1-\xi^2)}{2}\frac{\partial h_{\mathbf{k}_\perp}}{\partial\xi}
- \frac{(1 + \xi^2)}{4} \frac{v^2}{v_{te}^2}
k_\perp^2 \rho_e^2 h_{\mathbf{k}_\perp}
\right],
\label{model_e}
\end{equation}
where $\xi = v_\parallel/v$, the electron-electron operator $\langle C_{ee}[h]\rangle$
is given in \cite{Abel08} and the numerical implementation of the whole operator is detailed in \cite{Barnes09}.\footnote{As originally written, GS2 was not configured to simulate kinetic electrons
with adiabatic (Boltzmann) ions and also to include electron-ion collisions. This is because GS2 always formally requires at least one kinetic ion species, and so single-species KE+AI simulations were actually performed as single-species kinetic-ions simulations with the sign of the charge in the Boltzmann response appropriately reversed. The equations solved are then correct for simulations with kinetic electrons and adiabatic ions without collisions or with same-species collisions only. For the present work, we added the capability to include the electron-ion collision term in such simulations; previously it was included only for electrons in multi-species simulations with both kinetic ions and kinetic electrons.}
The electron-ion collision rate is $\nu_{ei} = Z_{\mathrm{eff}}\nu_{ee}$, where
$\nu_{ee} \equiv \nu = \sqrt 2 \pi n e^4 T^{-3/2} m^{-1/2} \ln\Lambda$ is the electron-electron collision rate and $\ln\Lambda$ is the Coulomb logarithm. In a plasma with multiple ion species,
$Z_{\mathrm{eff}} = \sum_i n_i Z_i^2/\sum_i n_i Z_i$
arises from summing the individual electron-ion collision operators
over all ion species ($Z_ie$ is that species' charge); in a hydrogenic plasma, $Z_\mathrm{eff}=1$. Note, however, that GS2 treats $Z_{\mathrm{eff}}$ as an independent parameter in equation (\ref{model_e}), formally allowing one to vary $\nu_{ei}$
relative to $\nu_{ee}$ without affecting the quasineutrality equation (\ref{QN}).
This is normally done to include the effect of electron collisions with impurity ion species that are not being modelled kinetically, but are present in real experiments. In this work, it will also be useful to us in section \ref{zeffscansection} as a method of varying this ratio artificially.
Electron-ion collisions relax the parallel electron flow towards the stationary ion background, resistively dissipating the current associated with an electron flow below the ion gyroscale.\footnote{The electron flow is also a current because, in the limit of $k_\perp\rho_i\gg1$, ion gyromotion averages out over the electron scales, leaving, to lowest order, no ion-flow response in either parallel or perpendicular direction; cf.\ the adiabatic-ion approximation (\ref{ai}), which arises from the averaging out of the gyrokinetic ion-density response.} As a result, the electron-ion collision operator, and hence the electron collision operator overall, does not conserve (electron) momentum.\footnote{As we see in section \ref{results},
electron-ion collisions reduce the final saturation amplitude of electron-scale zonal flows, and this gives rise to a favourable scaling of electron heat flux with collisionality.}
The last term in the bracket in (\ref{model_e}) is a spatial diffusion --- it is a finite--Larmor-radius (FLR) effect
arising from the gyroaveraging of the collision operator \cite{Abel08}.
\section{Results}\label{results}
\subsection{Time evolution: from \texorpdfstring{``quasi-saturation''}{"quasi-saturation"} to long-time steady state}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.9]{qe_overlap_enorm_cut_gB_nobox}
\label{qe_overlap}
}
\subfigure[]{
\includegraphics[scale=0.9]{zv_overlap_enorm_cut_nobox}
\label{zv_overlap}
}
\caption{Evolution in time of \subref{qe_overlap} the turbulent electron heat flux in electron gyroBohm units, and \subref{zv_overlap} the square of the zonal velocity, $(k_{\mathrm Z}\phi_{\mathrm Z})^2$, for two simulations with electron collisionality $\nu = 0.2 \,\nu_{\mathrm{nom}}$, electron temperature gradient $a/L_T = 3.3$ (green: small-box simulation; yellow: large-box simulation; see \ref{numerics} for details).}
\label{overlap}
\end{figure}
Figure \ref{qe_overlap}
shows the turbulent electron heat flux as a function of time for two simulations that have identical
plasma parameters and differ only in numerical grid parameters (their coincidence
is evidence of numerical convergence; for more details of the numerics and of the
various issues of convergence, see \ref{convergence_study}, which also contains plots for other
values of $\nu$ and $a/L_T$ and of other zonal quantities). The heat flux $Q$
is calculated from the solution of equations (\ref{GKE}) and (\ref{QN2}) according to
\begin{equation}
Q = \overline{\int d^3\mathbf{v}\frac{mv^2}{2} \delta f v_{Ex}}\,,
\label{Qdef}
\end{equation}
where the overbar indicates a flux-surface average. We will normalise $Q$ to the gyroBohm value $Q_{\rm gB} = n T v_{te} \rho_\ast^2$, where $\rho_\ast = \rho_e/a$.
Besides its physical meaning as the heat flux,
$Q$ is also a good proxy for the turbulent fluctuation level of the non-zonal
modes ($k_y\neq 0$; the $k_y=0$ components of $\phi$ or $h$ do not contribute to $Q$).
After a short exponential transient during the linear growth phase, the system reaches a
``quasi-saturated'' turbulent state, which, however, is not the
final steady state. The final saturated state is reached much later, after a slow
decline in heat flux accompanied by slow growth of the zonal ($k_y=0$) component of the turbulence: Figure \ref{zv_overlap} shows the evolution of the zonal velocity squared,
\begin{equation}
(k_{\mathrm Z}\phi_{\mathrm Z})^2 = \sum_{k_x} k_x^2|\phi_{k_x,0}|^2
\label{kzphiz}
\end{equation}
(this equation also introduces the definition of the characteristic zonal scale $k_\mathrm{Z}$).
Snapshots of $\phi$, shown in Figure \ref{contours} and corresponding to early and late times in
one of these simulations, are a vivid illustration of the different structure of the ``quasi-saturated'' and the
final steady state: the former resembles the streamer-dominated state usually associated with ETG turbulence
\cite{Dorland00, Jenko00}, whereas the latter is a zonal-mode-dominated state, which we will now proceed
to analyse. Note that it is this long-time saturated state that matters for determining the level
of transport because, even though the time for it to emerge is long by the standards of
electron-gyroscale dynamics, it is still much shorter than the transport time scale in a tokamak.\footnote{Using MAST data from \cite{Valovic}, $\tau_E$ is of order 10 ms, whereas $a/v_{te}$ is of order 10 ns, a factor of $10^6$ smaller. The longest runs in this paper evolve for times only of order $10^4\,a/v_{te}$, so there is still good scale separation. }
This is because the self-consistent combination of $\delta f$ local (gradient-driven) gyrokinetics
at the gyroradius scale, and radial transport evolution at the system scale, implies a scale separation
in both space and time \cite{BarnesTrinity, Abel13}.
\subsection{Collisionality scaling: numerical results and theory}
\label{twiddle}
The two clearest results from the final saturated state of these simulations are summarised in Figure \ref{collscan}, which is based on a parameter scan in electron collisionality for two values of the
electron temperature gradient, at two different numerical resolutions.
The saturated heat flux increases roughly proportionally to the collisionality, whereas the zonal velocity is essentially independent of it. The heat-flux scaling is in remarkable agreement
with the experimental scaling (\ref{MASTscaling}).\footnote{To facilitate a quantitative comparison of our results with \cite{Valovic} -- whilst noting that we consider a particular flux surface whereas the authors of \cite{Valovic} considered the plasma globally -- we convert our nominal $\nu = 0.02\,v_{te}/a$ (Table \ref{parameters}) using the expression (\ref{nustar}) to get a nominal $\nu_\ast\approx 0.1$. The collisionality range shown in our Figure \ref{collscan} is thus approximately the same as the range of $\nu_\ast$ accessed experimentally (approximately 0.03--0.1 in Figure 3 of \cite{Valovic}) for which the scaling (\ref{MASTscaling}) was obtained.}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.75,clip]{qescan_mult_powerlines_enorm_gB_nobox2}
\label{qe_scaling}
}
\subfigure[]{
\includegraphics[scale=0.75,clip]{zvscan_mult_enorm_nobox2}
\label{zv_scaling}
}
\caption{Variation of \subref{qe_scaling} the time-averaged normalised electron heat flux $Q/Q_{\rm gB}$, and \subref{zv_scaling} the rms zonal velocity $k_\mathrm{Z} \phi_\mathrm{Z}$, defined by (\ref{kzphiz}), versus normalised electron collisionality $\nu a/v_{te}$, at the nominal (experimental) value of the temperature gradient $a/L_T = 3.42$, and at $a/L_T = 3.3$. Symbol shapes indicate the simulation series (``small box'' or ``large box''), as explained in \ref{convergence_study}. The dot-dashed line shows the experimental power-law scaling, and
the dotted line shows the theoretical linear scaling, $Q/Q_{\rm gB}\propto\nu_{\ast}$, equation (\ref{Q_scaling}).}
\label{collscan}
\end{figure}
Let us outline a simple explanation of these results, which we will then follow up with
a series of numerical experiments designed to test its plausibility.
Let us split the gyrokinetic equation (\ref{GKE}) explicitly into equations
governing the evolution of the
nonzonal and zonal components of the distribution function and the associated
electrostatic potential (cf.\ \cite{Sugama2009}),
\begin{equation}
h = h_{\mathrm{NZ}}+h_{\mathrm{Z}},
\quad
\phi = \phi_{\mathrm{NZ}}+\phi_{\mathrm{Z}},
\end{equation}
where the subscripts NZ, Z denote nonzonal ($k_y\neq0$) and zonal ($k_y=0$) modes, respectively:
\begin{eqnarray}
\fl
\frac{\partial}{\partial t}\left(h_{\mathrm{NZ}} + \frac{e \langle\phi_{\mathrm{NZ}}\rangle}{T} F\right)
+\left(v_\parallel\mathbf{b}+\mathbf{v}_{B}\right)\cdot\nabla h_{\mathrm{NZ}}
- \langle C[h_{\mathrm{NZ}}]\rangle
\nonumber\\
\fl
\qquad +
\underbrace{\langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla h_{\mathrm{Z}}
+\langle\mathbf{v}_{E,{\mathrm{Z}}}\rangle\cdot\nabla h_{\mathrm{NZ}}}_\mathrm{Z-NZ~interaction~(I)}
+\langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla h_{\mathrm{NZ}}
-\overline{\langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla h_{\mathrm{NZ}}}
\nonumber\\
\fl
\qquad = \underbrace{- \langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla F}_{\mathrm{energy~injection~(II)}},
\label{GKE_NZ}
\\
\fl
\frac{\partial}{\partial t}\left(h_{\mathrm{Z}} + \frac{e \langle\phi_{\mathrm{Z}}\rangle}{T} F\right)
+\left(v_\parallel\mathbf{b}+\mathbf{v}_{B}\right)\cdot\nabla h_{\mathrm{Z}}
- \underbrace{\langle C[h_{\mathrm{Z}}]\rangle}_\mathrm{damping~(III)}
= \underbrace{- \overline{\langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla h_{\mathrm{NZ}}}}_\mathrm{energy~injection~(IV)},
\label{GKE_Z}
\end{eqnarray}
where the overline denotes spatial averaging over $y$, i.e., the $k_y=0$ component.
Equation (\ref{GKE_Z}) is the $y$ average of the gyrokinetic equation (\ref{GKE});
then (\ref{GKE_NZ}) is the result of subtracting (\ref{GKE_Z}) from (\ref{GKE}).
We conjecture that the dominant balance governing the saturated state of the
nonzonal modes is between the zonal-nonzonal interaction terms (I) and the
the linear drive (energy-injection) term (II) in (\ref{GKE_NZ}).\footnote[1]{We are thus treating the collision term in the nonzonal equation as subdominant, or at least as
unimportant to this aspect of the dynamics. Numerically we find that it cannot be neglected as it regularises the fine velocity-space
structure arising due to the phase-mixing of $h_{\mathrm{NZ}}$ \cite{Sch16}.
The parallel streaming and magnetic-drift terms likely play a part
in determining the spatial structure of the turbulence \cite{Barnes11cb,Ghim13,Sch16}, but we shall see that we do not need to determine $k_y$, $k_{\mathrm{Z}}$ or $k_\parallel$.
In \cite{Sugama2009}, a split between nonzonal and zonal components is performed for integrated entropy balance equations (equations (67) and (68) of \cite{Sugama2009}) in which these other terms do not appear.
One could base a similar argument to the
one presented here on these equations.
}
We estimate these terms as
\begin{equation}
\mathrm{(I)}\sim
\langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla h_{\mathrm{Z}}
\sim \frac{c}{B} k_y\phi_\mathrm{NZ} k_\mathrm{Z} h_\mathrm{Z},
\label{ZNZint}
\end{equation}
and
\begin{equation}
\mathrm{(II)}\sim
\langle\mathbf{v}_{E,{\mathrm{NZ}}}\rangle\cdot\nabla F
\sim \frac{c}{B} k_y \phi_\mathrm{NZ} \frac{F}{L_T}
\label{NZdrive}
\end{equation}
where $k_{\mathrm{Z}}$ is the typical zonal wavenumber and $k_y$ the typical nonzonal wavenumber
(we have used $v_{Ex}\sim (c/B) k_y\phi$; see (\ref{vEx})).
The second zonal-nonzonal interaction term,
$\langle\mathbf{v}_{E,{\mathrm{Z}}}\rangle\cdot\nabla h_{\mathrm{NZ}}$,
is of the same order as (\ref{ZNZint}) if we assume that
\begin{equation}
\frac{h}{F} \sim \frac{e\phi}{T}
\label{h_phi_ordering}
\end{equation}
for both zonal and nonzonal modes. Balancing (\ref{ZNZint}) and (\ref{NZdrive}),
we find, after cancellation of $k_y \phi_{\mathrm{NZ}}$, that
\begin{equation}
k_{\mathrm{Z}} h_{\mathrm{Z}} \sim \frac{F}{L_T}.
\label{constraint_from_nz}
\end{equation}
This is a form of mixing-length hypothesis \cite{Wesson_mixing_length},
suggesting that the perturbed zonal gradients $\sim k_{\mathrm{Z}} h_{\mathrm{Z}}$ compensate
the background equilibrium gradients associated with $F$.\footnote{They need not necessarily flatten the background gradient completely or everywhere. From Figure \ref{zv_overlap}, the normalised zonal velocity $k_{\mathrm{Z}}\rho_e e\phi_{\mathrm{Z}}/T\rho_\ast \approx 0.3$ (we also find $k_{\mathrm{Z}}\rho_e\delta T_{\mathrm{Z}}/T\rho_\ast\approx 0.3$), which should be compared with the background gradient $a/L_T = 3.3$. For the (random-noise) initial conditions used in our simulations, the minimum $a/L_T$ required to sustain turbulence is between $3.0$ and $3.3$, so the perturbation levels correspond roughly to the distance away from this nonlinear critical gradient (of which we have not made a precise determination because the system appears to be strictly subcritical in the presence of even small flow shear; the linear critical gradient {\em without} flow shear is about $2.4$, as shown in \ref{linear_simulations}). Furthermore, the perturbed gradient is not constant (since $k_{\mathrm{Z}}\neq 0$); the actual magnitude of the gradient at any particular $x$ can exceed 0.3.
}
It follows from (\ref{h_phi_ordering}) and (\ref{constraint_from_nz}) that
\begin{equation}
k_{\mathrm{Z}} \frac{e\phi_{\mathrm{Z}}}{T} \sim \frac{1}{L_T}.
\label{zv_equation}
\end{equation}
Thus, the gradients of the zonal modes (e.g., the zonal velocity or the zonal temperature
gradient) are independent of collisionality. This independence is indeed seen in Figure \ref{zv_scaling}.
The only nonlinearity present in the zonal equation (\ref{GKE_Z}) is the nonzonal-nonzonal interaction
term (IV); zonal modes are not directly driven by background gradients
because $\nabla F$ is in the $x$ direction.
One can think of (\ref{GKE_Z}) as a kind of Langevin equation for zonal modes, which are
excited by coupling between nonzonal modes and damped by collisions (at long times,
the only damping mechanism is collisional; see \ref{Felix}).\footnote{How precisely
the zonal modes are generated is a matter for future research, but see section \ref{additional}.}
Therefore, the dominant balance in (\ref{GKE_Z}) is between the nonlinear energy injection (IV)
and collisional damping (III):
\begin{equation}
\frac{c}{B} k_{\mathrm{Z}} k_y \phi_{\mathrm{NZ}} h_{\mathrm{NZ}} \sim \gamma_{\mathrm{Z}} h_{\mathrm{Z}},
\label{constraint_from_z}
\end{equation}
where $\gamma_{\mathrm{Z}}$ is the collisional damping rate of the zonal modes.
Note that in conjecturing such a balance, we are assuming the zonal modes
are not subject to a strong nonlinear instability that would break them down
back into nonzonal perturbations, thus resulting in a purely collisionless
saturated state. This possibility (which, for example, appears to be realised
for zonal flows in certain regimes of ITG turbulence, where it is known as the
tertiary instability \cite{Rogers00}) would amount to a dominant balance
in (\ref{GKE_Z}) between the energy-injection and energy-removal
effects within the nonlinear term (IV). We are expressly assuming that this is
not the dominant balance in our near-marginal ETG turbulence.
Combining equations (\ref{zv_equation}), (\ref{constraint_from_z}), and (\ref{h_phi_ordering}) again,
we get
\begin{equation}
\frac{h_{\mathrm{NZ}}^2}{h_{\mathrm{Z}}^2} \sim \frac{\phi_{\mathrm{NZ}}^2}{\phi_{\mathrm{Z}}^2}
\sim \frac{\gamma_{\mathrm{Z}} eBL_T}{c k_y T} \sim \frac{1}{k_y\rho_e}\frac{\gamma_{\mathrm{Z}}}{v_{te}/L_T}.
\label{hratio_scaling_equation}
\end{equation}
Therefore, estimating the heat flux (\ref{Qdef}), we find
\begin{equation}
\frac{Q}{Q_{\rm gB}}
\sim
\frac{n \delta T_\mathrm{NZ} v_{Ex}}{n T v_{te} \rho_\ast^2}
\sim
\frac{\delta T_{\mathrm{NZ}}}{T}\frac{c k_y \phi_{\mathrm{NZ}}}{B v_{te}\rho_\ast^2}
\sim
k_y\rho_e\left(\frac{e\phi_{\mathrm{NZ}}}{T \rho_\ast}\right)^2,
\label{Q_form}
\end{equation}
where we have used (\ref{h_phi_ordering}) to estimate ${\delta T}/{T}\sim {e \phi}/{T}$.
Using (\ref{hratio_scaling_equation}) to relate $\phi_{\mathrm{NZ}}^2$ to $\phi_{\mathrm{Z}}^2$
and (\ref{zv_equation}) to estimate $\phi_{\mathrm{Z}}^2$, we get
\begin{equation}
\frac{Q}{Q_{\rm gB}}
\sim
\frac{\gamma_{\mathrm{Z}}}{v_{te}/L_T}\left(\frac{e\phi_{\mathrm{Z}}}{T \rho_\ast}\right)^2
\sim
\frac{\gamma_{\mathrm{Z}}}{k_{\mathrm{Z}}^2\rho_e^2(v_{te}/a)}\frac{a}{L_T}.
\label{conductivity}
\end{equation}
Thus, we are able to estimate the electron heat conductivity entirely in terms of the linear damping rate and characteristic scale of the zonal flows.
It is possible to show analytically (\ref{Felix}) and confirm numerically (section \ref{zonal_damping_section}) that in the long term, zonal modes are damped by collisional (Ohmic) resistivity:
\begin{equation}
\gamma_{\mathrm{Z}}\sim\nu_{ei}k_{\mathrm{Z}}^2\rho_{pe}^2,
\label{gammaZ}
\end{equation}
where $\nu_{ei}$ is the electron-ion collision rate and $\rho_{pe} = \rho_e B/B_p$ is the
``poloidal Larmor radius'' of the electrons ($B_p$ is the poloidal magnetic field; see
\ref{order1} for a more precise definition of $\rho_{pe}$).
Using (\ref{gammaZ}) in (\ref{conductivity}), we finally obtain
\begin{equation}
\frac{Q}{Q_{\rm gB}}\sim
\frac{\nu_{ei}}{v_{te}/a}\left(\frac{B}{B_p}\right)^2 \frac{a}{L_T}\propto\nu_{ei}.
\label{Q_scaling}
\end{equation}
Thus, a relatively simple theoretical argument
has produced a linear scaling of the heat flux with collisionality. Note that all dependence on $k_\mathrm{Z}$
or any other wave numbers has cancelled in the final expression (\ref{Q_scaling}),
and so, in order to obtain the heat flux, we need not know
the spatial scales of either zonal or nonzonal modes. Considering the simplicity of the argument, and the level of agreement between it, our numerical results (Figure \ref{collscan}), and the experimental MAST scaling (\ref{MASTscaling}), we find it quite compelling.
\subsection{Damping of zonal modes}
\label{zonal_damping_section}
In the theoretical argument of section \ref{twiddle},
a crucial step was to use the expression (\ref{gammaZ}) for the
collisional damping of the zonal modes, which allowed
us to estimate the heat flux according to (\ref{conductivity}) and avoid having to
theorise about the characteristic scale of the zonal modes (a nontrivial question, with, as our simulations indicate, possibly
a nonuniversal answer).
In \ref{Felix},
the damping rate (\ref{gammaZ}) is derived analytically.
Physically, the situation can be summarised as follows.
\begin{figure}[t]
\centering
\includegraphics[scale=0.8]{damping_newaxeslabels_powerlines3_nobox}
\caption{Zonal damping rate normalised to collisionality, $\gamma_{\mathrm{Z}}/\nu$, versus $k_x\rho_e$, spanning the range of collisionalities shown in Figure \ref{collscan} (solid colors). The final states of various
saturated nonlinear simulations were used as initial conditions, with the nonlinearity switched off. Also shown (black crosses) are the corresponding damping rates for a simulation at $\nu = \nu_{\mathrm{nom}}$ in which electron-ion
collisions were turned off (formally by setting $Z_\mathrm{eff}=0$; see section \ref{colls}); and a simulation at $\nu = \nu_{\mathrm{nom}}$ in which electron-ion collisions were retained but
magnetic drifts were turned off (blue open circles). The solid line is the slope
$\propto k_x^2$, corresponding to the scaling (\ref{gammaZ});
the dashed line is $\propto k_x^4$, corresponding to the scaling expected
when $\nu_{ei}=0$ (see \ref{intraspecies}).}
\label{zonal_damping}
\end{figure}
Consider a zonal perturbation with some perpendicular wave number $k_x$
satisfying (\ref{GKE_Z}) with zero right-hand side --- a linear
equation. In the absence of collisions, this perturbation will decay quickly
(on the time scale $\sim a/v_{te}$), but not
to zero, leaving a finite residual zonal field \cite{RH}. With collisions present,
after a period of a few collision times, which is still much shorter than the damping
time, $\nu^{-1}\ll\gamma_\mathrm{Z}^{-1}\sim (\nu k_x^2\rho_e^2)^{-1}$ in the
long-wavelength limit $k_x\rho_e\ll1$, it is possible to show that, to lowest
order in $k_x\rho_e$, the remaining perturbation is a perturbed Maxwellian with
a density (or, equivalently, $\phi$) and a temperature perturbation, both constant
on each flux surface. These
perturbations then decay diffusively due to perpendicular particle diffusion (equivalently, resistivity)
arising from the collision operator. We already saw in section \ref{colls} that
the gyrokinetic collision operator (\ref{model_e})
contains FLR terms that have the form of a spatial diffusion.
These terms correspond to the displacement of gyrocentres by distances $\sim\rho_e$
due to collisions during Larmor rotation. Solving the ``zonal transport'' problem
more carefully, one can show that collisions also displace the centres of banana
(and corresponding passing) orbits by distances of order the poloidal gyroradius
$\rho_{pe} = (B/B_p)\rho_e$, which is larger. This leads to the damping rate (\ref{gammaZ}).
It is proprtional to $\nu_{ei}$ (rather than $\nu_{ee}$, on which it depends weakly)
because it is essentially the Ohmic resistive
damping of electron currents (both parallel and perpendicular; see \ref{order1}), and
it is due to electron-ion friction (cf.~\cite{Helander}).
The calculation of \ref{Felix}, where this is demonstrated more carefully, can
be checked in our numerical simulations, to ascertain that it is indeed this effect
that is responsible for the zonal damping. Figure \ref{zonal_damping} shows the
zonal damping rate normalised by the collision frequency, $\gamma_\mathrm{Z}/\nu$,
for a number of simulations in which the nonlinearity in equation (\ref{GKE_Z}) was
turned off (the right-hand side set to zero) and the zonal field allowed to decay linearly.
We see that, for a range of collisionalities $\nu$ and in a broad range of wave numbers
$k_x\rho_e$, the scaling (\ref{gammaZ}), $\gamma_\mathrm{Z}/\nu \sim k_x^2\rho_{pe}^2$,
is followed quite well. As a further successful test, we find that if we turn off
magnetic drifts ($\mathbf{v}_{B}\cdot\nabla h$ in equation (\ref{GKE_Z})), thus
removing the banana orbits, the zonal damping rates drop by close to an order of
magnitude (blue open circles in Figure \ref{zonal_damping}). This is roughly consistent
with a reduction of $\gamma_\mathrm{Z}$ by a factor of $(B/B_p)^2\approx (qR/r)^2\approx 18$
in our geometry,
to $\gamma_\mathrm{Z}\sim \nu k_x^2\rho_e^2$,
with the dominant diffusion due in this case to finite Larmor orbits, as explained above.
Finally, we test the theoretical expectation that the dominant contribution to the
damping of the electron zonal flows, which are also currents, is proportional
specifically to the electron-ion collision frequency $\nu_{ei}$. Figure \ref{zonal_damping}
(black crosses) shows that when the electron-ion collisions are turned off,
$\nu_{ei}=0$, leaving only the (momentum-conserving) electron-electron collisions $\nu_{ee}$,
the damping rates drop dramatically and scale as $\gamma_\mathrm{Z} \sim \nu_{ee} k_x^4\rho_{pe}^4$,
as indeed expected theoretically (see \ref{intraspecies}).
Thus, the scaling (\ref{gammaZ}) and the theory that leads to it (\ref{Felix}) appear to be sound
and successfully reproduced in our simulations.\footnote{It is perhaps worth pointing
out that such an agreement is only possible in simulations that use a
sufficiently realistic electron gyrokinetic collision operator (see section \ref{colls}),
an indispensable property being momentum conservation by the electron-electron
collisions and a correct capturing of Ohmic resistivity by the electron-ion ones.
See, however, section \ref{zeffscansection} for certain simplifications that are allowed.}
\subsection{Numerical tests of the role of collisions and zonal modes}
\label{modified_dynamics}
With the theoretical argument presented in section \ref{twiddle} in mind, let us now build up the evidence that the long-time steady state of the ETG turbulence in our simulations is controlled by zonal-nonzonal interactions and by the electron-ion collisional damping of the zonal modes. First, we remark that whilst
linear simulations (see \ref{linear_simulations})
indicate that the linear instability growth rates may be comparable to the nominal (experimental)
collisionality,
the growth rates depend only weakly on collisionality. It does not seem plausible that such insensitive linear physics can explain the strong collisionality dependence of the nonlinearly saturated state. Clues
to the actual (nonlinear) origin of this dependence can be obtained via nonlinear simulations with
modified dynamics, as described below.
\begin{figure}[t]
\centering
\includegraphics[scale=0.9]{qe_overlap_perts4_enorm2_cut_gB_nobox}
\caption{The large-box simulation shown in Figure \ref{overlap} (yellow) was restarted
at $t = 8397.7\, a/v_{te}$
without zonal interactions in the nonzonal evolution equation (cyan). The heat flux
in this modified simulation returns to a level that is close to the high early ``quasi-saturated''
level. For direct comparison, the same modified simulation was also rerun from initial noise,
giving the same heat flux level (purple).}
\label{perts}
\end{figure}
First let us show that the zonal component regulates the amplitude of the rest of the turbulence, which
determines the heat flux. The cyan and purple curves in Figure \ref{perts} show the time evolution of the heat flux in simulations with identical parameters to one of the simulations in Figure \ref{overlap},
but with the nonlinear term artificially adjusted
in such a way that the zonal modes no longer affect the evolution of the
nonzonal modes: namely, the zonal components have been zeroed out in the
calculation of the nonlinear term, so that in equation (\ref{GKE}),
$\langle\mathbf{v}_{E}\rangle\cdot\nabla h$ is replaced by
$\langle\mathbf{v}_{E,\mathrm{NZ}}\rangle\cdot\nabla h_\mathrm{NZ}$.
The zonal modes are still allowed to be nonlinearly driven by the nonzonal modes,
but the zonal modes do not then feed back on the nonzonal modes; the
nonzonal evolution is entirely independent of the zonal evolution.
This eliminates the nonlinear terms used to obtain the dominant balance (\ref{constraint_from_nz}).
The heat-flux collapse occuring in the full simulation is prevented by this change, confirming that the collapse is indeed mediated by the effect of the zonal modes (which was turned off) on the nonzonal modes (which carry the heat flux).
The ``quasi-saturated'' streamer-dominated state is thus just the saturated state that would have emerged had the zonal flows been prohibited or suppressed.
\begin{figure}[t]
\centering
\includegraphics[scale=0.9]{qe_double_perts_enorm_gB}
\caption{The two simulations corresponding to the red squares in Figure \ref{collscan} are shown: the large-box simulations with $a/L_T = 3.42$ and collisionalities $\nu = \nu_{\mathrm{nom}}/2$ (cyan) and $\nu = \nu_{\mathrm{nom}}/3$ (red).
The higher-collisionality case is restarted
at $t = 10003.6\, a/v_{te}$
with certain terms in the collision operator reduced
to the lower collisionality value, whilst others are retained at the higher value: (a) only electron-electron
collisions reduced (blue); (b) only electron-ion collisions reduced (purple); (c) only nonzonal collisions reduced (green); (d) only zonal collisions reduced (yellow).
}
\label{perts2}
\end{figure}
Collisions damp the zonal modes, and in this context are important precisely \emph{because} they are small, as this means that the finite zonal fields that emerge from any fast collisionless evolution \cite{RH} are damped
very weakly and so can grow to dynamically significant amplitudes and regulate the turbulence.
Let us show that it is the electron-ion collisions that affect the zonal modes in the crucial way;
these are momentum
non-conserving (for electrons), as they act to relax the electron flow and thereby
dissipate the associated current --- in other words, they give rise to Ohmic resistivity
(see discussion in section \ref{colls}).
In Figure \ref{perts2}, we show time evolution of the heat flux corresponding to two different collisionalities in the same series of simulations (the two shown as red squares in Figure \ref{collscan} --- the large-box simulations with $a/L_T = 3.42$). The heat flux in the higher-collisionality simulation is shown by the cyan curve, the heat flux in the lower-collisionality one by the red curve. If we rerun the higher-collisionality simulation with $\nu_{ei}$ unchanged but $\nu_{ee}$ reduced to the lower value (blue curve), or with the zonal collisionality
unchanged but the nonzonal collisionality reduced to the lower value (green curve), there is no
significant change in the saturated heat flux.
By contrast, if we reduce only
$\nu_{ei}$, leaving $\nu_{ee}$ unchanged (purple curve), or if we reduce only the zonal collisionality,
leaving the nonzonal collisionality unchanged (yellow curve), the heat flux drops to a value
consistent with the lower-collisionality case (red).
Thus it is the electron-ion collisions on the zonal component that primarily
determine the heat-flux collisionality dependence.
\subsection{Simplified simulations for extended collisionality range}
\label{zeffscansection}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.75,clip]{qescan_allnom_powerlines2_enorm_gB_nobox}
\label{qe_scaling_zeff}
}
\subfigure[]{
\includegraphics[scale=0.75,clip]{zvscan_allnom_enorm_nobox}
\label{zv_scaling_zeff}
}
\caption{Variation of \subref{qe_scaling_zeff} the time-averaged electron heat flux $Q/Q_{\rm gB}$,
\subref{zv_scaling_zeff} the rms zonal velocity $k_\mathrm{Z} \phi_\mathrm{Z}$,
adding to the points from Figure \ref{collscan} (now circles) further points (red crosses) obtained by varying only the electron-ion collisionality $\nu_{ei}$, but keeping the electron-electron collisionality at the nominal value, $\nu_{ee} = \nu_\mathrm{nom}$, all at the nominal temperature gradient $a/L_T = 3.42$.
The dot-dashed line shows the experimental scaling (\ref{MASTscaling}) and
the dotted line shows the theoretical linear scaling, $Q/Q_{\rm gB}\propto\nu_{\ast}$, equation (\ref{Q_scaling}).}
\label{zeffscan}
\end{figure}
We have argued that it is the effect of electron-ion collisions on the zonal modes that sets the collisionality dependence of the saturated heat flux. Collisions in the nonzonal gyrokinetic equation (\ref{GKE_NZ}) are regularising, in that collisions are needed to dissipate the fine structure that the distribution function
develops in velocity space. This means that, while collisions cannot be dropped for the nonzonal modes,
the particular value of $\nu$ is unimportant or, at most, has a weak effect --- for example
on linear growth rates (see \ref{linear_simulations}).
This is similar to the well-known situation in fluid turbulence, where a small viscosity is needed to provide dissipation but the saturated state is independent of the exact value of this viscosity. This understanding of the underlying physics opens up an opportunity to probe the collisionality dependence of electron transport without paying the
high price of increased velocity-space resolution that reducing collisionality would exact. The
strategy is to vary only the collisionality affecting the zonal modes or only $\nu_{ei}$
(although in the latter case, the subdominant part of the zonal damping rate
$\sim \nu_{ee} k_x^4\rho_e^4$ --- see section \ref{zonal_damping_section} and \ref{intraspecies} --- will
eventually take over).\footnote{We further stress that one should be careful in the interpretation
of such simplified simulations outside the range of collisionalities used for the full simulations.
The simplified simulations could become unrepresentative of the original system if the fixed
electron-electron collisions suppress new modes that would otherwise have emerged and
dominated the dynamics at either low or high collisionality.
We have not investigated whether this is the case for the simulations reported here.}
Figure \ref{zeffscan} shows the collisionality dependence of the electron heat flux over a wider range of values of $\nu_{ei}$ than in Figure \ref{collscan}, accessed using the latter strategy: varying $\nu_{ei}$ only while
keeping $\nu_{ee}$ at its nominal value. We see that the result of this extended collisionality scan is to
confirm the general plausibility of our picture of the ETG transport: $Q/Q_{\rm gB}$ stays approximately proportional to $\nu_{ei}$ across a wider range of its values, and the independence of the
zonal flow velocity of $\nu_{ei}$ also approximately persists over this wider range.
Note that these simulations were performed with
reduced perpendicular spatial resolution and without flow shear;
see \ref{numerics} and \ref{convergence_study} for details.
\subsection{Spatial structure of the saturated state}
\label{additional}
Finally, let us provide some details about the structure of the zonal and nonzonal components of the
turbulence in its final saturated state.
The zonal component of the saturated turbulent field is long-lived, with a spatial structure that can persist, once established, for as long as the total simulation time. Figure \ref{zp_hov} is a Hovm\"oller (space-time) diagram of the zonal potential for the same case as shown in Figure \ref{overlap}.
The zonal pattern barely changes during the time window that we have used
for time-averaging the heat flux and other quantities in the saturated state of this
simulation.\footnote{In this context, we gain some insight into the mechanism
of generation of the zonal modes by returning to the numerical experiment shown in Figure \ref{perts}
(section \ref{modified_dynamics}). We find that disconnecting the
zonal feedback on the nonzonal modes but keeping the nonzonal-nonzonal interactions
in the evolution equation for the zonal modes
still leads to growth of the zonal modes in the ``quasi-saturated'' state, but they are quite different from
the ones in simulations where their feedback is preserved: namely,
they are more incoherent in time and they also grow more vigorously (see Figure \ref{zp_hov_start_comparison}).
Thus, in the fully coupled system, even when the zonal modes are small,
their excitation by the nonzonal modes still depends on the small modifications
(subdominant as far as the heat flux is concerned)
that they produce in the latter --- and thus the mechanism of this excitation
is not as straightforward as just stochastic coupling of nonzonal modes into $k_y=0$.}
\begin{figure}[t]
\centering
\includegraphics[scale=0.65]{zp_hov_enorm_set_full}
\caption{Zonal electrostatic potential $\phi_{\mathrm{Z}}$ in a flux-tube cross-section in the outboard midplane as a function of the radial spatial coordinate $x$ and time, for the same ``large-box'' simulation as shown in Figure \ref{overlap} ($\nu = 0.2 \,\nu_{\mathrm{nom}}$, $a/L_T = 3.3$). The black line shows the start of the time-averaging window used to characterise the saturated state.}
\label{zp_hov}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.65]{zp_hov_enorm_set_start}
\label{zp_hov_start}
}
\subfigure[]{
\includegraphics[scale=0.65]{zp_hov_enorm_set_start_zap}
\label{zp_hov_start_zap}
}
\caption{
\subref{zp_hov_start} The same as Figure \ref{zp_hov},
but for the initial period $t\leqslant 2000 \,a/v_{te}$;
\subref{zp_hov_start_zap} the same for
the case in which zonal feedback onto the nonzonal modes has been disconnected (the simulation
for which the heat flux is shown as the purple curve in Figure~\ref{perts}).}
\label{zp_hov_start_comparison}
\end{figure}
The spatial spectrum of the zonal potential is almost monochromatic in this particular case, as shown in Figure \ref{zonal_spectrum}. In general, we find that the zonal spectra typically have only a small number of sharp peaks. A single extra harmonic is just visible in Figure \ref{zonal_spectrum}.
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.75]{overlap_zonal_spectra_enorm_cut}
\label{zonal_spectrum}
}
\subfigure[]{
\includegraphics[scale=0.75]{overlap_zonal_tperp_spectra_enorm_cut}
\label{zonal_tperp_spectrum}
}
\subfigure[]{
\includegraphics[scale=0.75]{overlap_ky_spectra_enorm_gB_loglog}
\label{kyspectrum}
}
\caption{For the same case as Figure \ref{overlap}: \subref{zonal_spectrum} the spectrum of the zonal potential;
\subref{zonal_tperp_spectrum} the spectrum of the zonal perpendicular temperature perturbation;
\subref{kyspectrum} the heat-flux spectrum;
all are
averaged over the time window of the saturated state, which is shorter for the large-box simulation (yellow) than for the small-box simulation (green).
}
\label{spectra}
\end{figure}
By contrast with the monochromatic zonal spectrum, the nonzonal spectrum is broadband when averaged over time. Figure \ref{kyspectrum} shows
the heat-flux spectrum against $k_y$, and Figure \ref{meanspectrum} shows the nonzonal $\phi$ spectrum for the large-box simulation of Figure \ref{overlap} against both $k_x$ and $k_y$.
Again by contrast with the static zonal component, the nonzonal component is
rapidly fluctuating. Figures \ref{flicker}\subref{slice1spectrum}-\subref{slice3spectrum} show the same 2D spectrum of $\phi$ at
particular instants of time. Thus, the nonzonal spectrum has a ``flickering'' appearance:
at any given time, a small number of modes are much more intense than the
others, but the dominant modes change over time, giving rise to a smooth time-averaged spectrum. This high ``$\mathbf k$-space intermittency'' is perhaps natural in the near-marginal saturated state at experimentally relevant parameters, with only a small number of modes excited at any given time.
The details may be dependent on the simulation grid in $k_x$ and $k_y$, and this would be an interesting topic for future investigations.
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.75]{2dspectrum_mean_enorm}
\label{meanspectrum}
}
\subfigure[]{
\includegraphics[scale=0.75]{2dspectrum_slice1_enorm}
\label{slice1spectrum}
}
\subfigure[]{
\includegraphics[scale=0.75]{2dspectrum_slice2_enorm}
\label{slice2spectrum}
}
\subfigure[]{
\includegraphics[scale=0.75]{2dspectrum_slice3_enorm}
\label{slice3spectrum}
}
\caption{2D spectra of the nonzonal electrostatic potential for the same large-box simulation as in Figure \ref{overlap}: \subref{meanspectrum} the spectrum averaged over the saturated time-window $t \geqslant 7502.3 \, a/v_{te}$; \subref{slice1spectrum} instantaneous spectrum at the same time as Figure \ref{satslice}; \subref{slice2spectrum} \& \subref{slice3spectrum} instantaneous spectra at later times during the saturated state. The grid-cell size in these plots is 0.012 in both directions: the minimum positive $k_x\rho_e$ and $k_y\rho_e$ (see \ref{numerics}).}
\label{flicker}
\end{figure}
\section{Summary and discussion}
\label{discussion}
The prevailing view of the structure of ETG turbulence in tokamaks and
the associated levels of transport has its origin in the first gyrokinetic simulations,
which did not include collisions \cite{Dorland00, Jenko00} (and did not correspond to near-threshold conditions in a spherical tokamak).
Whereas the ITG turbulent state has long been believed to be zonal-flow dominated
\cite{Diamond05}, the ETG fluctuations were characterised by long radial eddies
(``streamers'') that enhanced the transport to a level comparable with ITG turbulence,
overcoming the reduction by a factor of $(m_e/m_i)^{1/2}$ expected from the relationship
between the electron and ion gyroscales (the scales at which the two types of fluctuations
were driven).
The present study differs from the more traditional approach to modelling ETG turbulence
in three respects: collisions are included; simulations are run for a
much longer time; and the equilibrium parameters correspond to the experimental situation in a real device, namely MAST, and therefore place the system close to a marginal state with respect to the
ETG drive (note also that MAST is a spherical tokamak, so has a somewhat different
magnetic geometry compared to the more prevalent large fusion devices such as
TFTR, JET or ITER).
As a result, we find that, in application to the physical regime that we have considered,
the standard picture of ETG transport is in need of revision.
The high-transport, streamer-dominated nonlinear state does indeed emerge, and is not strongly
dependent on the collisionality of the plasma, but it persists only
transiently, over relatively short simulation times (short compared to the energy
confinement time but still long compared to a typical eddy turnover time,
and long compared to typical simulation times used for ETG turbulence in the past).
It turns out that this state
is not entirely steady --- while the heat flux might appear to be statistically stationary,
there is a slow growth of the zonal component of the fluctuations,
which eventually (after $t v_{te}/a \sim$ a few
thousand) reaches dynamical strength compared to the transport-setting nonzonal modes
and proceeds to change the character of the turbulence. A new, long-time,
zonal-dominated saturated state emerges, whose structure is more reminiscent
of what is traditionally expected of ITG, rather than ETG, turbulence
(see Figure \ref{contours} and further discussion in section \ref{ITG}).
We emphasize that it is the {\em final} saturation level of the heat flux in
gyrokinetic flux-tube simulations, averaged over the turbulent fluctuation scales
in length and time, that is physically relevant for transport calculations.
We have found that the turbulent heat flux supported by the new long-time nonlinear
state can be much lower than in the
``quasi-saturated'' streamer-dominated state if the collisionality of the
plasma is low (see Figures \ref{qe_overlap}, \ref{qe_scaling} and \ref{qe_scaling_zeff}).
The (roughly linear) collisionality dependence of the heat flux
found in our simulations turns out to be in remarkably good agreement
with the experimental scaling (\ref{MASTscaling}) \cite{Valovic}.\footnote[1]{Earlier
simulations of ETG in MAST \cite{Joiner,Roach09} did report heat fluxes roughly consistent
with the estimated experimental level of electron heat transport in the cases
that were simulated.
Whilst we obtain a difference in heat flux between the ``quasi-saturated'' state and the final state across the entire
range of collisionality shown in Figure \ref{collscan}, including at the nominal collisionality,
the most dramatic differences occur when the collisionality is well below its nominal value (see Figure \ref{evolution}\subref{qe_evolution}).}
We have proposed a phenomenological argument (section \ref{twiddle})
whereby this collisionality scaling can be understood if one assumes that the saturation
of the nonzonal modes is governed by the zonal gradients coming into approximate
balance with the equilibrium gradients (to be more precise,
the zonal-nonzonal interactions balancing the linear drive), while the saturation of the zonal modes
is set by a balance between their nonlinear excitation by the nonzonal interactions and
their damping by Ohmic resistivity. The latter is operative because electron-scale flows
are also currents --- and so it is the electron-ion collisions that
play the defining role in setting the zonal damping rate. We have supported our
view by a series
of numerical experiments that confirmed the crucial role of the zonal modes
in enabling the emergence of the new saturated state (section \ref{modified_dynamics})
and the crucial role of the electron-ion collisions on the zonal modes
in setting its collisionality dependence (sections \ref{modified_dynamics} and \ref{zeffscansection}).
The Ohmic damping of the zonal flows is an analytical result (\ref{Felix}), but
we have also systematically confirmed that it is captured correctly in our simulations
(section \ref{zonal_damping_section}). We have also documented some
key qualitative features of the long-time saturated state: the long-time
coherence and approximate monochromaticity of the zonal modes (cf.\ \cite{Parker_Krommes_2013,Parker_Krommes_2014}) and, in contrast, the dominant individual nonzonal modes ``flickering'' with time
in and out of existence to give rise to a broad-band time-averaged spectrum
(section \ref{additional}).
\subsection{Previous work on ETG and the collisionality scaling}
Long-time changes in the saturated state of gyrokinetic simulations have
previously been reported in other numerical studies.
Mantica et al.~\cite{ManticaEPS2011} reported a change
in the zonal-nonzonal balance at long times in an ITG simulation.
Guttenfelder and Candy~\cite{GuttenfelderCandy}, in their collisionless ETG simulations
with adiabatic ions using the GYRO code, found a long-time reduction in transport
associated with an increased level of zonal perturbations, which occurred at a low but
not at a higher level of flow shear.
We have not investigated here the dependence on flow shear, except via the simplified
simulations with zero flow shear reported in section \ref{zeffscansection}, which appear to obey
the same scalings (Figure \ref{zeffscan}).
It is not impossible that the transport scaling
found in the present paper exists within a window in which the flow shear is not too large
(although it should perhaps still be large enough to suppress ion-scale transport
in order for a heat-flux calculation restricted to electron scales to make sense).
Perhaps most relevantly for
comparisons with our work, Nakata et al.~\cite{Nakata2010}, using
a slab ETG model, adiabatic ions, a Lorentz collision operator (i.e.,
the pitch-angle scattering operator in (\ref{model_e})), and comparing two different
sets of parameters, reported that transport at long times in their linearly more unstable case
could be suppressed below the level of their linearly more stable case, owing to the formation
of zonal flows that collimated the turbulence into ``vortex streets'' and acted as a barrier
to radial transport (see also earlier studies of ETG zonal flows \cite{Idomura05,Idomura06}).
This suppression appears to be consistent with our findings.
Recent direct measurements using the Doppler back-scattering system installed on MAST \cite{Hillesheim} are consistent with small-scale turbulence due to ETG.
Gurchenko and Gusakov \cite{GurchenkoGusakov} established an experimental correlation in several devices between electron-scale turbulence and anomalous electron transport, but not a causation pathway directly attributing the transport to ETG. The transport scaling that we obtain here derives from, and may, therefore, be an experimental signature for, not only the presence of electron-scale turbulence but also the saturation mechanism involving a dominant interaction with weakly damped zonal modes. In a regime in which this were not the saturation mechanism, the scaling of transport with collisionality may well be much weaker or absent.
An alternative explanation for the experimental collisionality scaling of heat
transport in spherical tokamaks that has been previously suggested relies on the transport
associated with microtearing turbulence \cite{Valovic,Guttenfelder2012,Guttenfelder2013}.
Since, in the present work, we have limited ourselves to electrostatic perturbations,
the microtearing instability is excluded and and so it is clear that the collisionality
scaling that we have found does not require it.
Note that the arguments we apply to ETG-driven turbulence are fairly generic and
may well apply in some form to other instabilities. We have, of course, also not excluded
the possibility that other modes may produce a similar dependence for different reasons.
The microtearing contribution in particular to the overall electron transport in real
fusion devices remains a live and pressing research subject
(made challenging, however, by the particular difficulty
of obtaining well-resolved simulations of electromagnetic turbulence in tokamaks).
It is an interesting question how general our picture
of ETG turbulent transport might prove to be. NSTX \cite{Kaye} and MAST \cite{Valovic} both exhibit the strong scaling of confinement time with collisionality considered here, but in conventional tokamaks (as opposed to these two spherical ones), the scaling exponent between $B\tau_E$ and $\nu_\ast$ has been found to be much weaker or essentially zero \cite{ITERPB1999, Cordey2005NF, Bourdelle2011}.
Clearly, identifying the reason or reasons for this difference is an important issue for further study.
We expect ETG to be more important in STs for the overall energy confinement, because of the large flow shear available to suppress ion-scale modes. Considering just ETG-dominated electron heat transport, however, one possibility is that, since in conventional tokamaks $B/B_p$ is large compared to STs, giving larger zonal damping, the heat flux given by the scaling (\ref{Q_scaling}) may be as large as the heat flux in our ``quasi-saturated''
state, in which case the ``quasi-saturated'' state may be the actual saturated state, not regulated by the mechanism discussed here, and with only weak or no collisionality dependence.\footnote{More generally, something similar may happen in ranges of collisionality, background gradients or other aspects of the magnetic geometry that differ from those considered here. However, even in conventional tokamaks, the collisionality scaling from the argument that we have presented may still be recovered at sufficiently low values of collisionality. Testing this hypothesis in simulations of conventional tokamaks would be an interesting avenue to explore in the future.}
\subsection{ETG vs.\ ITG turbulence near and far from threshold}
\label{ITG}
The idea that the ETG turbulent state is dominated by zonal modes leads one naturally
to the question of whether the saturated states of ETG and ITG turbulence are
essentially similar, at least qualitatively.
Our ETG turbulence model, equations (\ref{GKE}) and (\ref{QN2}), describes
kinetic electrons with Boltzmann (adiabatic) ions.
The simplest possible model for ITG turbulence would involve
kinetic ions with Boltzmann electrons --- the latter physically justified
by fast streaming of electrons along field lines. The main mathematical difference
between these two models is that the Boltzmann electron response must be restricted
to perturbations that have variation along magnetic-field lines. Namely,
the density perturbation is, in contrast to the ETG case (\ref{ai}),
\begin{equation}
\frac{\delta n}{n} = \frac{e(\phi-\bar\phi)}{T_e},
\label{ai2}
\end{equation}
where $\bar\phi$ is the flux-surface average \cite{Dorland93,Hammett93,AbelCowley}.
It is this latter feature that is normally believed to be responsible for the
difference between the zonal-flow-dominated ITG state and the streamer-dominated ETG state:
radial variations in the zonal $\phi$ do not affect $\delta n$ as
electrons cannot respond to radially varying zonal modes at ion scales;
these can, therefore, grow large enough to break up the primary ITG-driven streamers
via a secondary instability \cite{Cowley91,Rogers00}
and isotropise the turbulence, whereas for ETG turbulence,
the latter effect was believed to be too weak to destroy the streamers \cite{Dorland00,Jenko00}
(although the physics that determined the radial scale
of the streamers perhaps remained unclear \cite{Nevins06} --- possibly again a (weaker) secondary
instability \cite{Cowley91,Jenko02}).
What we have found in our simulations is that, whereas there is indeed no
trace of a fast onset of a secondary instability similar to one that breaks
up the ITG streamers (and is indeed seen in most ITG simulations),
the ETG streamers' lease on life granted by a stronger density response
to zonal perturbations is nevertheless only temporary:
the zonal component does find a way to grow slowly,
until it is large enough to take control over the nonzonal fluctuations.
In addition to the ITG instability giving rise to streamer-like modes and
the secondary instability breaking them up into zonal flows, the third pillar
of the standard picture of ITG saturation is a ``tertiary instability''
whereby the zonal modes that have grown to a certain critical amplitude
break up, returning energy to nonzonal perturbations \cite{Rogers00}. This mechanism
of regulating the zonal component of the turbulence does not rely on collisional damping
and thus should give rise to heat fluxes that are mostly independent
of collisionality --- as indeed it appears to do, at least when the ITG
turbulence is simulated in regimes that are far from the instability threshold \cite{Barnes11cb}.
However, it has long been known that at temperature gradients that are close
to the linear threshold (and thus, arguably, most relevant experimentally),
the tertiary instability is ineffective, and a state of strongly suppressed
transport ensues, called the Dimits upshift of the critical temperature
gradient \cite{Dimits00,Rogers00}. Clearly, in the case of ETG turbulence,
should a tertiary instability of the zonal modes appear, their collisional
damping would cease to matter and consequently the collisionality dependence
of the heat flux should flatten off (cf.\ \cite{Ricci06}). It is then tempting to suppose that
our ETG state, in which the zonal modes are stable, collisionally damped
and the heat flux scales with collisionality, is simply the ETG version
of the Dimits-shift regime. Ascertaining whether this is indeed the case
requires a larger parameter scan (in particular, in the temperature gradients)
than has been undertaken here and has to be left for future work.
By analogy with the ETG case, another tempting conjecture would be that the
Dimits regime for the ITG turbulence should also have an (ion) collisionality
scaling, by way of an argument similar to one presented
in section \ref{twiddle}.\footnote{In porting this argument to the ion scales,
one must not forget that an important difference between the ETG and ITG cases
is the specially pivotal role played by electron-ion collisions for the ETG zonal modes.
For ions, ion-electron collisions are always
subdominant in the mass-ratio expansion because the ions carry nearly all of the momentum,
and collisions with the lighter electrons have a negligible impact on the ion flow.
By contrast, the relative flow of electrons (which constitutes a current) is strongly
affected by collisions with the heavier, sluggish ions. Therefore, for electrons, collisions are
not momentum-conserving, whereas for ions, to lowest approximation they are.
How this difference plays out in the calculation of the rate of damping of ion
zonal flows is outlined in \ref{ionzfs}. In the end, it turns out, however,
that the damping rate scales in the same way as for the electron case,
$\gamma_{\mathrm{Z}i}\sim \nu_{ii} k_x^2 \rho_{pi}^2$, although for a
different reason.}
This in fact is not a new idea \cite{Lin,Lin00,Diamond05,Ricci06}, although the situation
remains murky as including more realism (in particular, kinetic electron response)
in gyrokinetic simulations of ITG turbulent transport can
nullify or even reverse its collisionality dependence \cite{MikkelsenDorland}.
Since the collisionality scaling of the ETG transport that we report
at least has some experimental support \cite{Valovic},
we feel more secure about the relevance of our conclusions --- although
``stress-testing'' them by relaxing various assumptions (e.g., perhaps most urgently,
allowing kinetic ions \cite{Nevins06}) and by extending the equilibrium parameter
ranges is an essential direction for future research.
Thus, the relationship between ETG and ITG is complicated and indeed within each
of these two regimes, there are likely to be several sub-regimes depending
on the part of the parameter space that one is interested in --- in particular, how
far from the threshold the turbulence is. This said, our results suggest that there
are perhaps more similarities between ETG and ITG physics than previously thought.
Seeking a unified and universal description of various types of drift-wave turbulence
may therefore be a worthwhile and reasonable aspiration.
In conclusion, let us summarise again what has been achieved here.
We have found that in the saturated state of ETG turbulence at driving gradients
close to their experimental values in MAST, the electron heat transport decreases with
collisionality, in agreement with experimental evidence \cite{Valovic}.
This behaviour points toward improved confinement in future devices.
We have explained it based on a simple theoretical picture (backed up by
numerical tests) of an ETG
turbulent state dominated by the interaction between nonzonal and zonal modes,
with the collisionality dependence of the heat flux originating from the
collisionality dependence of the resistive damping of the zonal modes.
On a practical note, ETG simulations with generic initial conditions
must be run to very long times to capture the (crucial) effect of zonal modes
on the saturated state.
\section*{References}
|
\section*{Introduction}
The social nature of human beings has attracted the interest of researchers since the beginning of science. Being part of a social entity and actively participating in it is a part of everyone of us. This natural tendency explained for the first time by Social Darwinism represents one of the environmental pressures that has probably affected most of the human evolution. Cooperation, competition and other social skills (e.g., social problem solving, social self- efficacy) as well as dynamics (e.g., individualism and pluralism, in-group favoritism) have determined human evolution and are still distinguishing our society more than other environmental features. Within such complex dynamics, social and cultural groups have developed different strategies in order to maintain an efficient way to cooperate within their community while simultaneously adopting an effective strategy to compete with the other groups within their ecological niche.
Nowadays the revolution in information and communication technologies (ICT) is changing the old social equilibrium of the world by providing a brand new way to interact with others, new social structures and possibilities, as well as new "ecological niches" to be explored and exploited.
In the 1990s, Dunbar’s number theory set the cognitive limit of the number of people with whom one can maintain stable social relationships to $150$ \cite{dunbar1992neocortex, dunbar2009social}.
It is clear that with the massive use of social networks and online communication tools the number of social relations we deal with daily is increasing, which suggests the need to re-think the concepts of ecological niches as well as the concepts of community, participation and privacy.
Similarly, even the ecological system theory \cite{bronfenbrenner1992ecological} needs to be updated by considering the possible relations that can be developed through new scenarios such as the world wide web.
Such a brand new way to interact is in facts affecting our lives as well as the world organization, and presumably scientists should consider this as the most important environmental factor that will affect the human evolution in the future.
For example recent and important historical events (i.e. Arab Spring, Occupy Wall Street Movements, anti-coup movements in Turkey) took place thanks to the possibility provided by social networks to gather an extraordinary amount of people in such a brief amount of time: before the ICT revolution it would have been at least harder to make those movements happen. Sure enough social networks, and new media in general, are having a key role in driving debates and public opinion in consequence of socially relevant events and can be seen as mediators of the consequences of these.
At the same time recent historical events are shaping our way to interact with ICT tools so that similar cultures are approaching social networks and new media in a different way, depending on the degree of affordances and concerns related to their use.
In order to properly approach this environmental change, we need to consider the cultural differences characterizing the interaction between people within this brand new world (i.e., perception of privacy, virtual sense of community and web based participation) and distinguishing the new equilibria that interconnects human society through the world wide web \cite{prensky2001digital}.
Understanding the human factors involved in the use of social networks and online applications, such as mobile applications or crowdsourcing platforms, is now one the most challenging objectives for several disciplines \cite{arnaboldi2013egocentric, passarella2012ego}. In fact by considering the massive and daily use of web tools in our lives, we need to understand how several well-known factors and socio-psychological dynamics in literature can be translated into virtual contexts. The opportunities in terms of communication and the sharing of knowledge that the "virtual world" is providing us are astonishing. However in order to gain an actual comprehension of the usage of such tools, a deeper understanding of the complex dynamics behind those processes needs to be achieved. While a proper and comprehensive model for the human interaction with technologies is still under development, it can be useful to approach this research by using well known techniques adopted in the past in the "real world" as the cross-cultural analysis of features involved in the behaviours of interest.
Cross-cultural differences involved in the perception of behaviors, attitudes or emotions have frequently been a precious tool in order to understand peculiarities or commonalities in several areas of research . The enrichment provided by different cultures can, in fact, highlight peculiarities related to the context or the history of a population that may help researchers in modeling complex behaviors. Investigating the connections between well known constructs and taking into account the differences due to cultural specificity or recent historical events is necessary for researchers to understand new important features, excluding some others related to a specific context, but showing the important factors in common with the human race.
The main objective of our study is to measure the relationship between Sense of Community, Participation and Privacy compared to Culture and Sex, taking into account two culture (Italian and Turkish) similar for many features, but surely different for recent historical events. After a brief theoretical introduction to the constructs (Sense of Community, Participation and Privacy), we will describe our study in which 180 digital native participants (83 Italians and 97 Turkish People) have been asked to fill in an online questionnaire.
\subsection{Sense of Community}
Sense of community (SoC) is a construct largely used in community psychology, where it has been used to describe the complex relationship that an individual has while feeling as an active part of his community. Since the term “community” has a multi-level nature, which according to Brodsky and Marx (2002) can be referred to neighborhood, block groups, housing complexes, schools and cities, it has been studied in a large variety of contexts \cite{brodsky2002expanding, brodsky2011layers}. Given that several definitions have been taken into account, the variety of features that this construct can express for different individuals in different settings is phenomenal. Sarason defined it as the feeling that community members have about each other \cite{sarason1974psychological, blanchard2008testing}, McMillan and Chavis (1986) claimed that SoC is a concept related to membership, emotional safety, identity, belonging and attachment to a group \cite{mcmillan1986sense}. The perception of similarity to others and willingness to communicate with people are also key features in defining SoC \cite{castellini2011sense}. According to these authors, when we feel part of a community we are more willing to share responsibility, to improve face to face relationships and to participate with others in social activities. On the other hand, several social scientists have shown that the increased complexity and change in technologies have affected the meaning and importance of communities \cite{brodsky2002expanding}. Relevant to this new approach to SoC, in which the contexts are evolving from a physical reality into a virtual scenario, all research is aimed at identifying a definition of the Sense of Virtual Community \cite{blanchard2002sense, rovai2005feelings, francescato2006evaluation}. The Sense of Virtual Communities has been found to be overlapping in the SoC in many features: for example participants in some studies by Francescato (2006; 2007) have shown similar levels in collaborative learning, developed self-efficacy, developed problem solving and perception of social capital when comparing face to face and online groups \cite{francescato2006evaluation, francescato2007developing}.
\subsection{Participation}
Due to its multi-disciplinary nature, participation has been investigated through approaches coming from several academic disciplines (e.g. Sociology, Anthropology Psychology) even though a proper descriptive and comprehensive model is still lacking at the moment. Given that, it can be analyzed starting from different levels: macro-social/institutional level, micro-social levels, psychological level. According to Cicognani and Zani (2011), when discussing participation on the macro-social/institutional level we refer to the electoral system, civic education, culture, religion, social-economic development and the history of the country in question \cite{cicognani2011civic}. On the micro-social level, contexts such as family, friends, school, and voluntary associations are taken into account . On the psychological level, Wilkenfeld et al. (2010) highlight the importance of participation, on both civic and political issues as well as on the cognitive and social development of adolescents \cite{wilkenfeld2010relation}. A key feature in promoting participation is the presence of a collaborative group sustaining the action. Sense of collective identity stimulates engagement in collective actions \cite{klandermans2002methods} as well as identification with the group \cite{cicognani2011civic}. To this effect, SoC \cite{mcmillan1986sense} has been proven to have a positive, bidirectional, association with participation \cite{simon1998collective}. One key feature related to participation is also the construct of empowerment \cite{rappaport1991healing} and as found in Zimmermann (2000) the action-oriented nature of empowerment finds a natural consequence in both individual and social participation \cite{zimmerman2000empowerment}. Combining the contribution from psychological literature and the recent findings coming from the renewed interest in this construct, especially from Information and Communication Technologies studies, several features have been taken into account to find the proper incentives promoting participation (inhibiting competition). For instance Gachter investigated the role of rewards and punishment, finding out that punishments can increase cooperation even in the long term, because the gains from cooperation itself give enough gains to became stable even if the punishment is removed \cite{gachter2008long}. In a more recent review on this topic, Gachter shows that even the size of the incentives has a crucial role in these dynamics. In a contest of rewards, a very large incentive can enhance a stronger cooperation, while with a higher punishment a low threat can be enough to reach the same result \cite {rand2009positive} \cite {sutter2010choosing} \cite {gachter2012social}
\subsection{Privacy}
With the increasing use of social networks (e.g., Facebook, Twitter, Google+), and collaborative web platforms (i.e., crowdsourcing tools) people find the opportunity to share ideas and feelings with others on internet everyday . While exploring these brand new and innovative media, we have learned that we also need to take into account the potential dangers that can occur while exchanging personal data online. Participating in social media is highly related to the feeling of safety that people experience while sharing their contents on the web. According to Song, Hao and Daqing (2013), people that have a higher trust in social networks are more likely to participate actively than those who have less \cite{song2013empirical}. When people feel a risk about sharing personal data on social networks they do not contribute to them. As mentioned above in SoC, behavior in real physical contexts and web scenarios seem to be overlapping in some of its features. In fact, privacy is a very important dimension of human life, affecting personal and social lives. Since it involves the control of the amount of contact with others and the perceived safety of these interactions, Solove (2008) has found that when people do not meet their privacy needs, it may result in antisocial and stressful behaviors \cite{solove2008end}.
Altman has defined privacy as “the selective control of access to information to the self or to one’s group” \cite{altman1975environment}. According to Pedersen (1997) six types of privacy can be identified: solitude, isolation, anonymity, reserve, intimacy with friends and family \cite{pedersen1997psychological}. Privacy attitudes have been measured in a sample of $210$ men and $165$ women from high schools and colleges in Turkey \cite{rustemli1993privacy}. Results show that women have higher mean scores for measures of intimacy with friends and lower mean scores for isolation and reserve than men. There were no mean differences for solitude, intimacy with family, and anonymity. Cross-cultural differences in the perception of privacy regulation have been investigated by Kaya and Weber in 2003, involving American and Turkish students \cite{kaya2003cross}. The result showed that American students desired more privacy in their residence hall rooms than Turkish students. Regardless of culture, males reported a greater desire for privacy than females. The relationship between privacy and security was investigated in a study regarding the educational use of cloud service such as social networks, Google drive and Dropbox \cite{arpaci2015effects}. Here, security was taken into account as the degree to which students believed that cloud-services were secure platforms for storing and sharing sensitive personal data. The results showed that the perception of a low level of security may have affected students’ attitudes towards using such services. In other words, students with low tolerance for technological risks may defer their use of these services and privacy concerns may then impede attitude towards educational use of cloud services. Several scales have also been recently developed in order to investigate privacy.
The Privacy Behavior Scale is a six-item, $5$ point Likert scale, assessing the need for privacy related to the use of Internet \cite{buchanan2007development}. The Privacy Attitude Scale \cite{buchanan2007development} measures privacy concerns for a number of Internet security topics, such as the use of e-mail with $16$ items on a $5$ point Likert scale. The Privacy Concerns Scale \cite{dinev2004internet} is a three-item Likert scale assessing concerns towards personal information provided on the Internet. The Identity Information Disclosure Scale \cite{stutzman2006evaluation} is specifically addressed to concerns regarding the disclosure of identity information in the the usage of social networks.
\subsection*{Aims and hypotheses of the present study}
The main objective of our study is to measure the relationship between the constructs of Sense of Community, Participation and Privacy compared to Culture and Sex. We examined the score on the scale for Sense of Community, Participation and Privacy in respect to Sex and Culture. We wanted to assess the effect of Culture and Sex on the perceived Sense of Community, Sense of Confidence/Concern toward the representation of the construct of Participation and Privacy (in order to consider cultural values referred to this construct, we provided an Italian/Turkish definition of the construct). In order to appreciate more how cultural differences are involved in representing the constructs of Participation and Privacy, we provided a semantic differential for both Italian and Turkish sub-samples. This measure has also been analyzed in respect to Sex. At the end of our analysis, we present the correlation structure between all the measures adopted for both the entire sample as well as the single sub-samples.
\subsection{Methods}
\subsection*{Participants}
A total of $180$ participants were recruited for the online questionnaires of our study. We divided the sample on the basis of their origin (i.e. Italy/Turkey).
The Italian sub-sample ($29$ M; $54$F) shows a mean age of $24.69$ S.D. $3.9$ (Male= $24.24$ S.D. $4.46$; Female = $24.93$ S.D. $3.57$) and the Turkish sub-sample ($49$ M; $48$ F) shows a mean age of $22.4$ S.D. $2.67$ ( Male = $22.96$ S.D. $2.68$; Female = $21.83$ S.D. $2.55$). Most of the participants were students ($84.\%$) recruited from The University of Florence (Italian sub-sample) and the Cukurova University (Turkish sub-sample).
\subsection*{Procedures}
The main objective of this study, as mentioned above, is to measure the relationship between Sense of Community, Participation and Privacy, in two different cultures (Italian and Turkish). In order to assess this relationship in respect to Culture and Sex, a set of questionnaires was prepared.
The scale we used to asses Sense of Community was the union of two different scales: The Classroom and Community Inventory (CSCI) as found in Rovai, Wighting \& Lucking (2004), and the Sense of Community in School as found in Vieno, Santinello, Pastore \& Perkins (2007), for a total of $16$ items on a $5$ point Liker scale, from $1$= "strongly" disagree to $5$="Strongly agree" \cite{rovai2004classroom, vieno2007social}. Since this questionnaires are aimed to assess Sense of Community in a scholastic environment, we took in consideration only participants that at the moment of the administration were university students or university workers (e.g., teachers, researchers, PhD students).
In order to obtain a measurement for both the Participation and Privacy constructs, two different questionnaires were developed. Both questionnaires were structured following this scheme: definition, semantic differential, and scale.
Definition: for both dimensions a general definition of the construct has been provided using those found in popular dictionaries (Treccani dictionary for Italian, “Wikipedia” and “Türk Dil Kurumu “ for Turkish) for both Italian and Turkish sub-samples.
Semantic differential: the Semantic differential technique was used in order to appreciate more the cultural differences regarding both constructs \cite{osgood1964semantic}. According to the Osgood’s semantic differential (1964) a list of $10$ items made of two bipolar adjectives was provided (e.g., warm-cold; useful-useless; safe-dangerous). Participants were asked to choose where his/her position lie in the continuum (a Likert Scale from $1$ to $10$) from the positive to the negative side (adjective) of each of the $10$ items.
Scale: for both Participation and Privacy an ad-hoc questionnaire was developed. Both scales presented $10$ items assessing participants’ perception of Participation/Privacy in relation to different contexts (real life, online and social networks). Items were alternated into two sub-sections: $5$ items assessed the confidence towards these constructs, $5$ assessed concerns towards them. Answers on a $5$-point Likert scale ($1$="Absolutely no", $2$="A few”, $3$="Moderately", $4$="A lot”, $5$="Very much”) were coded and summed up in order to obtain a general score for both Concerns/Confidence sub-scales.
\subsection*{Data analysis}
The data analysis activity was structured in two main phases. In the first one, we calculated the descriptive statistics, assessing the pre-conditions required by the subsequent inferential analysis. In particular, we checked the Gaussian distribution of the continuous variables, i.e. skewness and kurtosis $\in (-1;+1)$, and the sufficient balancing and size of the sub-samples of interest (i.e., sex and nationality). Then in the second phase we conducted the inferential analyses required by our main hypothesis.
In particular, within the context of this research the Pearson's \textit{r} correlation between the variables such as participation, privacy and sense of community were analyzed. Furthermore, an ANOVA analysis was conducted to find out if these three variables differentiate in terms of nationality and sex factors. SPSS.20 was used for these analyses.
\section*{Results}
In this section the main results of our research are presented. Descriptive statistics are presented first, where we show the results of the socio-demographic and operative variables (Participation, Privacy, Sense of Community) of our sample. In the following section, we show results of the inferential analysis (i.e. correlation structures, ANOVA analysis, effects of socio-demographics on a semantic differential scale).
\subsection*{Descriptive statistics}
\subsubsection*{Sociodemographic variables}
The task has been administered to Italian and Turkish university students. The Italian sub-samples shows a mean age of $24.68$ and is composed by $54$ females and $29$ males, the Turkish people (mean age= $22.40$) by $48$ females and $49$ males.
\begin{table}[h!]
\caption{Sociodemographic dimensions for the Italian and Turkish sub-samples. The standard deviation of each variable ($\sigma$) is reported between brackets}
\begin{center}
\small
\begin{tabular}{lccc}
\label{tab:descriptive_1}
Subsample&n&Age\\
\hline
Italian&83&24.68 (3.9)\\
Turkish&97&22.40 (2.7)\\
\hline
It Females &54&24.92 (3.6)\\
It Males &29&24.24 (4.5)\\
\hline
Tk Females &48&21.83 (2.6)\\
Tk Males &49&22.96 (2.7)\\
\hline
\end{tabular}
\normalsize
\end{center}
\end{table}
Most of the participants were students ($84\%$), single ($75\%$) and had an high degree of education (Bachelor degree= $71\%$). The Turkish sub-sample seem to be more uniform regarding civil status and education.
\begin{table}[h!]
\caption{Summary of descriptive statistics about the discrete sociodemographic variables of the sample}
\label{tab:Descriptive_0}
\begin{tabular}{lcccccc} \toprule
Variable & Total & \multicolumn{2}{l}{Italian} & \multicolumn{2}{l}{Turkish} \\
&& \multicolumn{2}{l}{Subsamples} & \multicolumn{2}{l}{Subsamples}\\
\cmidrule(r){3-6}
&& Female & Male & Female & Male \\ \midrule
\textbf{Occupation} & 180 &&&& \\
Student & 152 & 43&20&44&45 \\
Employed & 20 & 4&8&4&4 \\
Unemployed & 8 & 7&1&0&0 \\
\hline
\textbf{Civil status} & 180 &&&& \\
Single & 136 & 24&19&47&46 \\
Engaged & 38 & 29&6&1&2 \\
Married & 6 & 1&4&0&1 \\
\hline
\textbf{Education} & 180 &&&& \\
High Sc. & 24 & 12&11&0&1 \\
Bachelor & 128 & 27&11&44&46 \\
Mast. Sc & 28 & 15&7&4&2 \\
\hline\\
\end{tabular}
\end{table}
\subsubsection*{Operative variables}
Our scales about Participation and Privacy seem to be able to discern between the confidence and the concerns the participant has about these two dimensions (Table \ref{tab:Descriptive_2}). In the Italian sample, we found general lower values compared to the Turkish sub-sample, for both Confidence and Concerns. Average scores suggest that Italian females are more confident ($16.4$) than males ($15.4$) and less concerned (female=$12.9$; male= $12.1$) about participation. On the contrary they are more concerned ($14.9$) than males ($13.9$) and less confident (female=$12.1$; male=$13.3$) about privacy.
In the Turkish sample we see the same trend, but with higher values. Males and females have similar values about Participation Confidence (18.9 and 18.3), generally higher than the Concerns (female=$14.6$; males=$15.1$). Like in the Italian sample, Turkish participants have higher values in Privacy Concerns ($18.1$ for female and $16.9$ for male) than in Privacy Confidence ($13.6$ and $15.7$) and females show higher values than males on concerns for both measures.
Italian females ($55.17$) show higher values on the sense of community scale than Italian males ($51.38$). The Turkish sub-sample has generally lower values than the Italians, with similar scores for females ($50.23$) and males ($50.44$).
\begin{table*}[htbp!]
\centering
\captionsetup{justification=centering}
\small
\caption{Summary of descriptive statistics of the operative variables about participation and privacy}
\label{tab:Descriptive_2}
\begin{tabular}{lccccc} \toprule
Measure & \multicolumn{2}{l}{Italian} & \multicolumn{2}{l}{Turkish} \\
& \multicolumn{2}{l}{Subsamples} & \multicolumn{2}{l}{Subsamples}\\
\cmidrule(r){2-5}
& Female & Male & Female & Male \\ \midrule
\textbf{Participation} &&&& \\
Confidence & 16.4(3.3)&15.4(4.3)&18.3(3.9)&18.9(4.0) \\
Concerns & 12.9(3.6)&12.1(3.9)&14.6(2.6)&15.1(3.4) \\
\hline
\textbf{Privacy} &&&& \\
Confidence & 12.1(1.8)&13.3(3.0)&13.6(3.2)&15.7(3.0) \\
Concerns & 14.9(2.9)&13.9(2.8)&18.1(4.2)&16.9(4.5) \\
\hline
\textbf{Sense of } & &&&&\\
\textbf{Community} & 55.17(7.4)&51.38(12.3)&50.23(11.6)&50.44(11.1)\\
\hline
\end{tabular}
\end{table*}
According to the Semantic differential, the most interesting results about Participation is the elevated contrast between Italians and Turkish People.
As you can see in Figure \ref{fig:Participation_differential} Italians assessed almost all the adjectives at one of the extremes of the continuum; both males and females, showed very high scores for "Useless-Useful”, "Bad-Good”, "Dirty-Clean” and "Personal-Social”; for almost all other differentials the scores are very low, except for "Easy-Di cult”. Turkish samples present a totally different trend, where all values are low, only "Personal-Social”, "Easy-Di cult” and "Concrete-Abstract” have slightly higher values. For both Italians and Turkish people, the differences between males and females are minimal.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Participation_OsgoodSemantic_figure.jpg}
\caption{\label{fig:Participation_differential} Semantic differential for all the sub-samples with respect to the concept of "Participation".}
\end{center}
\end{figure}
In the Privacy Semantic Differential \ref{fig:Privacy_differential} the Turkish sample shows a trend similar to the Participation one with mid-low values in almost all items, except for "Personal-Social”, "Easy-Difficult" and "Concrete-Abstract”. The Italian sample has some similarities, but we can see lower values for "Important-Unimportant”, whereas we found mid values in the "Easy-Difficult”, "Concrete-Abstract” and "Warm-Cold” items. Exactly like the Participation Semantic differential and even for privacy, the differences between males and females are very low in both samples.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.45\textwidth]{Privacy_OsgoodSemantic_figure.jpg}
\caption{\label{fig:Privacy_differential} Semantic differential for all the sub-samples with respect to the concept of "Privacy".}
\end{center}
\end{figure}
Finally, the sense of community scale reports higher values for Italians in general. In particular, females report higher values in the Italian sub-sample ($55.17$ vs. $51.38$), and lower but similar values in the Turkish sub-sample ($50.23$ vs. $50.44$), as indicated in Table \ref{tab:Descriptive_2}.
\subsection*{Inferential statistics}
The inferential analysis was structured in three different phases investigating the correlation structure characterizing the observable quantities taken into account in this study (i.e., age, education, sense of community, participation and privacy). The second phase ran three ANOVA analyses in order to investigate the univariate and combined effects of nationality and sex on the three fundamental variables defining the constructs under scrutiny (i.e., sense of community, participation and privacy). Finally, the last phase was dedicated to assessing the effects of the socio-demographic factors of interest (i.e., nationality and sex) on the differential semantic dimensions describing the participation and privacy perception.\\
\subsubsection*{Univariate analysis: correlation between the order parameters}
\begin{table*}[htbp!]
\centering
\captionsetup{justification=centering}
\caption{Correlations between the dependent measures}
\begin{adjustbox}{angle=90}
\begin{tabular}{lccccccc} \toprule
\label{tab:Correlation_General}
Measure & Age & Education & Sense &\multicolumn{2}{l}{Participation} & \multicolumn{2}{l}{Privacy} \\
&&& of Community &\multicolumn{2}{l}{Submeasures} & \multicolumn{2}{l}{Submeasures} \\
\cmidrule(r){5-8}
&&&& Confidence & Concerns & Confidence & Concerns \\ \midrule
\textbf{Sense of Community} &&&& \\
It. Female & ns&$.28^*$&1&ns&ns&ns&ns \\
It. Male & ns&$-.45^*$&1&$.69^{**}$&$.54^{**}$&ns&$.58^{**}$ \\
Tk. Female & ns&ns&1&ns&ns&ns&ns \\
Tk. Male & ns&ns&1&$.39^{**}$&$-.33^*$&ns&ns \\
\hline
\textbf{Par. Confidence} &&&& \\
It. Female & ns&ns&ns&1&$-.53^{**}$&$.28^*$&ns \\
It. Male & ns&ns&$.69^{**}$&1&ns&ns&$.40^{*}$ \\
Tk. Female & ns&ns&ns&1&ns&$.61^{**}$&$.35^*$ \\
Tk. Male & ns&ns&$.39^{**}$&1&ns&$.40^{**}$&ns \\
\hline
\textbf{Par. Concerns} &&&& \\
It. Female & ns&ns&ns&$-.53^{**}$&1&ns&$.34^*$ \\
It. Male & ns&ns&$.54^{**}$&ns&1&ns&$.47^{**}$ \\
Tk. Female & ns&ns&ns&ns&1&ns&$.29^*$ \\
Tk. Male & ns&ns&$-.33^{*}$&ns&1&ns&$.40^{**}$ \\
\hline
\textbf{Pri. Confidence} &&&& \\
It. Female & ns&$-.31^{*}$&ns&$.28^{*}$&ns&1&ns \\
It. Male & ns&ns&ns&ns&ns&1&ns \\
Tk. Female & $.29^{*}$&ns&ns&$.61^{**}$&ns&1&ns \\
Tk. Male & ns&ns&ns&$.40^{**}$&ns&1&ns \\
\hline
\textbf{Pri. Concerns} &&&& \\
It. Female & ns&$-.40^{**}$&ns&ns&$.34^{*}$&ns&1 \\
It. Male & ns&ns&$.58^{**}$&$.40^{*}$&$.47^{**}$&ns&1 \\
Tk. Female & ns&ns&ns&$.35^{*}$&$.29^{*}$&ns&1 \\
Tk. Male & ns&ns&ns&ns&$.40^{**}$&ns&1 \\
\hline
\multicolumn{8}{l}{Pearson \textit{r} correlation, *: \textit{p} < .05, **: \textit{p} < .01, \textit{ns}: not significant.}
\end{tabular}
\end{adjustbox}
\end{table*}
The Pearson's \textit{r.} correlation statistic was adopted to estimate the relations between the fundamental variables of interest with respect to the $4$ sub population considered by our study \ref{tab:Correlation_General}.
The age of the participants appears to have no significant effects. The only exception is related to the Turkish female sample where a weak relation emerged between age and privacy confidence.
Education appears to be related to the sense of community only in the Italian sample. In particular, it showed quite an interesting reversed effect regarding sex for Italian males, where we obtained a negative correlation explaining about $20\%$ of the variance which suggested that males with a higher education are characterized by a lower sense of community. However, the Italian females reported a positive correlation between the two measures explaining the $8\%$ of the variance. For what concerns the privacy construct, education appears to affect only the Italian female perception. In particular, we always obtained a negative correlation accounting for $9\%$ and $16\%$ of the variance respectively for privacy confidence and privacy concerns.
The sense of community correlation structure shows interesting features as well. First of all no significant correlation appears to relate the sense of community with privacy and participation for the females disregarding nationality. On the contrary, the male sub-samples both reports positive correlations between sense of community and participation confidence with higher values for Italians ($\textit{r.} = 0.69$), and lower values for Turkish students($\textit{r.} = 0.39$). While an opposite behavior appears to characterize the relation with the participation concerns, with values respectively of $\textit{r.}=0.54$ for Italians, and of $\textit{r.}=-0.33$ for Turkish males. Only the Italian males show a significant positive correlation between sense of community and privacy concerns ($\textit{r.}=0.58$). In order to evaluate the non-trivial relationships existing between concerns and confidence for both the participation and privacy perceptions, we explicitly took into account all the correlation structures describing such sub-measures separately.
The participation confidence appears to be significantly and negatively related to the participation concerns only within the Italian females sub-sample, explaining $28\%$ of the variance. For what concerns the relationship between participation and privacy confidence, all the sub-samples report positive correlations with the only exception of the Italian males, suggesting how an increment in the first produces a coherent change in the second. In particular, the Turkish females show the higher correlation ($\textit{r.}=0.61$), and the Turkish males the second one ($\textit{r.}=0.40$), with the Italian females that report the weakest relation ($\textit{r.}=0.28$). Finally the participation confidence is even related with the privacy concerns for Italian males ($\textit{r.}=0.40$), and Turkish females ($\textit{r.}=0.35$).
In brief, the participation concerns are reported to correlate with the privacy concerns, and not with the privacy confidence for all the investigated sub-samples. All the correlations here are between $11\%$ and the $22\%$.\\
\subsubsection*{Sense of community: nationality and sex effects}
The effects of nationality and sex, as well as their interaction on the sense of community were evaluated by means of ANOVA analysis. As reported in table \ref{tab:Senseofcommunity_1}, no differences emerge regarding the sex, nor the interaction between nationality and sex. The only factor affecting such a dimension appears to be the nationality, even if the effect is very moderate explaining just $3.5\%$ of the variance, with the Italian sub-sample reporting a higher value.
\begin{table}[h!]
\centering
\caption{In table the best ANOVA model for the dependent variable Sense of Community, considering the factors sex and nationality, is reported.}
\label{tab:Senseofcommunity_1}
\begin{tabular}{lccc} \toprule
\multicolumn{3}{l}{\textbf{Sense of community}} \\
Factor & \textbf{F} &\textit{Sign}& $\eta^2$\\
\hline
Nationality & 6.4 & \textit{p.} < 0.05 & 0.035 (3.5 \%)\\
Sex & 1.2 & \textit{ns} & 0.007 (0.7\%)\\
Nat*Gen & 1.5 & \textit{ns} & 0.009 (0.9\%)\\
\hline
\hline
\end{tabular}
\end{table}
\subsubsection*{Participation dimensions: nationality and sex effects}
Confidence and concerns sub-measures concerning participation were separately analyzed by ANOVA (table \ref{tab:Participation_1}), in order to estimate the role of nationality and sex on the two dimensions. For both categories, sex and interaction vs sex and nationality do not have statistically significant effects, while again nationality appears to be the only factor impacting on such dimensions. In both the cases, the cultural differences seem to account for approximately $10\%$ of the variance. In particular, both the participation confidence as well as the participation concerns appear to be significantly higher within the Turkish sample.
\begin{table}[h!]
\centering
\caption{In table the best ANOVA model for the variables connected to the participation dimension (i.e., confidence \& concerns) are presented, considering the factors sex and nationality}
\label{tab:Participation_1}
\begin{tabular}{lccc} \toprule
\multicolumn{3}{l}{\textbf{Participation Confidence}} \\
Factor & \textbf{F} &\textit{Sign}& $\eta^2$\\
\hline
Nationality & 20.7 & \textit{p.} < 0.01 & 0.105 (10.5 \%)\\
Sex & 0.1 & ns & 0.001 (0.1\%)\\
Nat*Gen & 1.8 & ns & 0.010 (1\%)\\
\hline
\multicolumn{3}{l}{\textbf{Participation Concerns}} \\
Factor & \textbf{F}& $\eta^2$\\
\hline
Nationality & 20.1 & \textit{p.} < 0.01 & 0.102 (10.2 \%)\\
Sex & 0.1 & ns & 0.001 (0.1\%)\\
Nat*Gen & 0.6 & ns & 0.009 (0.9\%)\\
\hline
\hline
\end{tabular}
\end{table}
\subsubsection*{Privacy dimension: nationality and sex effects}
In table \ref{tab:Privacy_1} the best predicting model linking the nationality and sex factors to the sub-scales related to the privacy dimension (i.e., confidence and concerns) is reported. In these cases, we obtained a significant effect both on nationality as well as on sex, and again no effects were revealed about the interaction between the factors of interest. In particular, nationality explains $11.4\%$ for the privacy confidence, and the $14.1\%$ of the privacy concerns. In both cases, again the Turkish sample is characterized by higher values. In regards to sex, females reported higher values for privacy concerns, and lower values for privacy confidence, respectively explaining $8.3\%$ and $2.1\%$ of the total variance.
\begin{table}[h!]
\centering
\caption{In table the best ANOVA model for the variables connected to the privacy dimension (i.e., confidence \& concerns) are presented, considering the factors sex and nationality}
\label{tab:Privacy_1}
\begin{tabular}{lccc} \toprule
\multicolumn{3}{l}{\textbf{Privacy Confidence}} \\
Factor & \textbf{F} &\textit{Sign}& $\eta^2$\\
\hline
Nationality & 22.6 & \textit{p.} < 0.01 & 0.114 (11.4 \%)\\
Sex & 15.8 & \textit{p.} < 0.05 & 0.083 (8.3\%)\\
Nat*Gen & 1.3 & ns & 0.007 (0.7\%)\\
\hline
\multicolumn{3}{l}{\textbf{Privacy Concerns}} \\
Factor & \textbf{F}& $\eta^2$\\
\hline
Nationality & 28.9 & \textit{p.} < 0.01 & 0.141 (14.1 \%)\\
Sex & 3.7 & \textit{p.} < 0.05 & 0.021 (2.1\%)\\
Nat*Gen & 0.1 & ns & 0.001 (0.1\%)\\
\hline
\hline
\end{tabular}
\end{table}
\subsubsection*{Semantic differential about Participation and Privacy}
Very interesting results emerged from the semantic differential analysis regarding the nationality and sex as well as their combined effect. Table \ref{tab:Semantic_Differentials_models} reports only the significant effects, that nevertheless appear to be present in both cases (i.e., semantic differential related to privacy and participation separately) for $8$ dimensions out of the original $10$. Some dimensions are characterized by moderate effects, with an explained variance ranging between $2.6\%$ and $9.6\%$. It is worth noting some dimensions appear as strongly affected by the nationality factor. For instance the concept of participation appears to elicit impressive differences, with "bad-good" accounting for $61.2\%$ of the total variance, "dirty-clean" for $50.2\%$, and "unusefull-usefull" for $46.5\%$. Even the privacy concept appears differently represented by the two cultural samples, with "personal-social" dimension explaining $45.8\%$, "unuseful-useful" $45.2\%$, "dirty-clean" $44.1\%$, and "bad-good" $40.4\%$. In particular, as reported in figures \ref{fig:Participation_differential} and \ref{fig:Privacy_differential}, the Italian sample seems to perceive the participation concept as more "clean", "good" and "useful" compared to the Turkish sample. The same effects are reported for the privacy concept with the only exception of the dimension "personal-social", in which the Turkish sample perceives "privacy" as more social than the Italians.
\begin{table}[h!]
\centering
\small
\caption{In table the best ANOVA models for the variables connected to the semantic differentials about both Participation and Privacy are presented, considering the factors sex and nationality. Statistical significance (\textit{p.}) is reported in table as follows: $^*: <0.05$ and $^**:< 0.01$}
\label{tab:Semantic_Differentials_models}
\begin{tabular}{lccc} \toprule
\multicolumn{4}{l}{\textbf{Participation}} \\
Dimension & Factor & \textbf{F} & $\eta^2$\\
\hline
Warm-Cold & Nationality & 4.7* & 0.026 (2.6 \%)\\
& Nat*Gen & 5.5* & 0.031 (3.1\%)\\
\hline
Dirty-Clean & Nationality & 177.6** & 0.502 (50.2 \%)\\
\hline
Safe-Dangerous & Nat*Gen & 7.3** & 0.040 (4.0\%)\\
\hline
Bad-Good & Nationality & 277.8** & 0.612 (61.2 \%)\\
& Sex & 12.3** & 0.065 (6.5\%)\\
\hline
Unuseful-Useful & Nationality & 153.0** & 0.465 (46.5 \%)\\
& Sex & 15.2* & 0.080 (8.0\%)\\
\hline
Imp.-Unimportant & Nationality & 7.4** & 0.041 (4.1 \%)\\
& Nat*Gen & 11.2** & 0.060 (6.0\%)\\
\hline
Personal-Social & Nationality & 18.8** & 0.096 (9.6 \%)\\
& Sex & 8.4** & 0.045 (4.5\%)\\
\hline
Concrete-Abstract & Nationality & 63.9** & 0.266 (26.6 \%)\\
\hline
\hline
\multicolumn{3}{l}{\textbf{Privacy}} \\
Dimension & Factor & \textbf{F} & $\eta^2$\\
\hline
Warm-Cold & Nationality & 12.6** & 0.067 (6.7 \%)\\
& Sex & 4.5* & 0.025 (2.5\%)\\
\hline
Dirty-Clean & Nationality & 138.6** & 0.441 (44.1 \%)\\
\hline
Bad-Good & Nationality & 199.4** & 0.404 (40.4 \%)\\
\hline
Fun-Boring & Nationality & 9.1** & 0.049 (4.9 \%)\\
\hline
Unuseful-Useful & Nationality & 145.2** & 0.452 (45.2 \%)\\
& Sex & 4.2* & 0.023 (2.3\%)\\
& Nat*Gen & 5.2* & 0.028 (2.8\%)\\
\hline
Imp.-Unimportant & Sex & 4.1* & 0.023 (2.3 \%)\\
\hline
Personal-Social & Nationality & 148.6** & 0.458 (45.8 \%)\\
& Nat*Gen & 4.9* & 0.027 (2.7\%)\\
\hline
Concrete-Abstract & Nationality & 5.5* & 0.030 (3.0 \%)\\
\hline
\hline
\end{tabular}
\end{table}
\section*{Discussion}
The psychological constructs of Participation and Privacy are currently among the most interesting psychological concepts to investigate. The growing multi-disciplinary literature about them, ranging from engineering to complex systems and social sciences, witnesses the impact of ICT on modern human society, and in general on every new kind of interactions among people and groups \cite{castellano2009statistical,Vilone201684,cecconi2016new}. Many societal dynamics, from the microscopic interaction within families to the macroscopic cultural and sociological movements involving people from all over the world, are based on the same human tendencies of interaction with others by means of the adoption of cultural and personal dependent strategies of coping. From the "Arab spring" in 2010, to the "Occupy Wall Street movements" in 2011, and the recent protests in Turkey (i.e., from Taksim square protests to the recent coup), internet has been considered by many commentators as one of the fundamental ingredients. Of course the connection between people, especially within the younger generations, has provided new incentives and mainstreams that have nurtured the protest movements. On the other hand, such events could be considered as a direct consequence of the changes produced by internet on its users. Actually, from this point of view it seems very hard to clarify the role of the world wide web on such cultural and societal changes, while it seems more reasonable to take into account both the dynamics described above. Human beings change in order to satisfy environmental requirements (i.e., increasing their adaptation), to negotiate social constructs, norms and behaviors useful to communicate, interact and solve problems with others as well as represent elements affecting and actively changing the environment they belong to.
For instance, the spreading of social networks in the last few years required non-digital natives to make a great effort to adapt to such a brand new social order, while it profoundly shaped the social skills and attitudes of digital natives. Such a double process (i.e., to adapt old strategies to new tasks, and to adopt new concepts and habits) is not just a question of age (i.e., digital natives vs non-digital natives), but it represents a double process acting in different degrees within every participant, and in different ways (i.e., with different speed and effectiveness) along the different cultures, and socio-demographic clusters. In general, it is expected that the dynamics of such changes will define new interesting scenarios and phenomena, and of course both the spreading of new concepts and social negotiations of old habits and norms will continue to determine societal as well as personal changes.
The hypotheses of our study moved from the basic assumption that the "fluid" human perception of socially-negotiated concepts like Participation and Privacy, can differ a lot between two different cultural groups, depending on structural differences of the cultures and on different recent events affecting the university students participating in the survey.
The Turkish and the Italian cultures have ancient common roots, and could be both considered as European cultures even if with a very different recent history. Finally, the different role of sex within the two cultural groups was considered and evaluated in the multivariate analysis.
The two sub-samples appeared as comparable regarding the size, average age and sex balance. With the age that actually presented very little dispersion around the average values, and that can be considered as a constant describing a very specific social cluster (i.e., those of the university students from a medium size city). Consequently, it is not surprising that no significant correlations emerged concerning age. The only correlation with the age variable is reported in the Turkish female sub-sample as affecting the Privacy Confidence. The older the participant is the greater is her confidence and trust in the concept of Privacy (Tab. \ref{tab:Correlation_General}).
The education variable represents the number of formal years of study accomplished by the participants, and the different distribution of the two sub-samples is due only to the different organization of the two educational systems. For this reason only the Italian sample presents significant correlations between the education variable and the operative dimensions. The opposite relationship between sex and sense of community is very interesting. In the Italian male sub-sample a negative correlation is revealed, while in the female one a significant positive correlation is present. Moreover in the Italian sample, the males are characterized by a negative correlation between sense of community and education, while the Italian females show the opposite trend. The Privacy sub-dimensions, confidence and concerns, show a negative correlation in age only for Italian females, and consequently appear to reduce both their confidence and concerns about Privacy with the increase of education.
The sense of community dimension shows very interesting relation patterns within our sample, reporting significant correlations only for the male sub-samples, and not always in the same direction. The Participation concept is perceived as more positive and trustworthy as the sense of community increases, the same effect acts on Italian male sub-samples even for the concerns, while it presents an opposite effect on the Turkish male sub-sample (i.e., the concerns about participation decrease with the increase of sense of community). Finally the Italian sub-sample presents a positive correlation between the sense of community and Privacy concerns scores (Tab \ref{tab:Correlation_General}).
In general, it is interesting to observe how the results of the correlation matrix between Participation and Privacy measurements are quite empty. In particular, larger correlations were expected between the confidence and concerns sub-measures. Such data indicates the multi-structured and complex architecture of the concepts under scrutiny and how in different cultures their interplay could not be easy to reveal or predict.
In order to evaluate the interplay between sex and culture on the operative variables of our study (i.e., sense of community, participation and privacy concerns and confidence), we conducted a series of separate ANOVA in order to test the effects of the factors as well as of their interaction.
The sense of community does not appear to differ much regarding the nationality ($3,5\%$), and the sex does not play a significant role in the interaction between sex and nationality. In other words, the two cultures appear to be comparable in terms of sense of community.
On the contrary, more relevant differences are detected regarding the Participation and Privacy measures. The ANOVA investigating the sex and nationality effects on the sub-scores of the Participation scale (Tab. \ref{tab:Participation_1}) reports that around the $10\%$ of the variance is explained by the nationality factor. In particular, both the confidence and the concerns about participation are significantly higher in the Turkish sample. A greater portion of the variance is explained by the ANOVA investigating such effects on the Privacy sub-scores. Concerning this construct respectively $11,4\%$ and the $14,1\%$ are associated with the nationality, again indicating the Turkish sub-sample as more confident and concerned about Privacy (Tab. \ref{tab:Privacy_1}). Finally an effect of the sex is revealed as explaining respectively the $8,3\%$ and the $2,1\%$ of the variance of Privacy confidence and concerns sub-scores, with the Females reporting always a lower score.
The most fascinating results are those produced by the semantic differential measures. Even in this case a series of ANOVA were conducted analyzing every dimension presented by the semantic differential, considering the sex and the nationality as independent factors of variance. The Fisher’s Fs and the Etas squared reported in table 8 delineates impressive differences in the perception of such concepts always mainly related to the nationality (i.e., the culture).
Both the perception of the Participation and Privacy appear to be affected by nationality and sex; in particular in 8 cases out of 16 the explained variance (i.e., the entity of the difference explained by the factors) is under $10\%$. Surprisingly the other 8 cases report an explained variance ranging between the $61,2\%$ and the $26\%$.
For what concerns Participation, it seems that Italians perceive it as a more "clean", "good", "useful" and "social" than Turkish, with no significant difference to sex. Among the samples, there are no significant differences on the perceived importance, warmness, easiness and perceived "fun" related to the experience of participating. The fact that Turkish people perceive it as more "dirty", "concrete", "bad", "useless" could be related to the recent experiences of the Taksim square protests in which really violent clashes happened.
As regards to the concept of Privacy, it seems that the Italians perceive it as a more "useful", "good", "clean" and "personal" concept than the Turkish students. In order to appreciate these results with a cultural interpretation, we wanted to highlight the key role that the recent Italian mass media and political debates have given to the concept of privacy. In fact, the Italian political agenda in the recent years has been highly focused on a privacy regulation, especially regarding important legal topics. This concern has led to a more developed concept of privacy, highly related to the concept of "human rights that need to be respected and protected" as found in the definition provided by the main Italian dictionary (Treccani dictionary). This concept of privacy as a right is not stressed at the same level for example in the current English definition, which stresses more the perspective of an "ability of an individual or group to seclude themselves" (Wikipedia), demonstrating how this concept is still in development and culturally shaped and variable. The fact that the Turkish students perceive the construct of Privacy as more "useless", "social", "dirty" and "bad" could be given to the fact that this concept is quite new to them. In fact, the word "privacy" is not in the main Turkish dictionaries at the moment and even the literal translation of this term in this language ("gizlilik") needed a well-structured explanation to make it comprehensible to the Turkish sub-sample.
We believe that a better understanding of the dynamics underneath the use of ICT tools can provide us the necessary knowledge in order to shape these environments in a more efficient way (i.e. taking into account cultural, sex and historical differences), strengthening positive features and mitigating the negative ones.
\section*{Acknowledgments}
This work was funded by the European Commission under the FP7-ICT-2013-10 call, proposal No. 611299, project SciCafe 2.0, and by H2020 FETPROACT-GSS CIMPLEX Grant No. 641191.
\bibliographystyle{apacite}
|
\section{Introduction}
It has recently been noted that the classical dynamics of black holes
simplifies in the limit of a large number of dimensions. The key observation
- first made by Emparan, Suzuki, Tanabe and collaborators in \cite{Emparan:2013moa,Emparan:2013xia,Emparan:2013oza,Emparan:2014cia,Emparan:2014jca,Emparan:2014aba,Emparan:2015rva} - is that black holes
at large $D$ have two effective length scales. The first of these, $r_0$,
is the size of the black holes. The second
is the thickness of the black hole's gravitational tail, i.e. the
distance beyond the black hole event horizon after which the gravitational
potential rapidly decays to zero. In four dimensions
the black hole size and thickness are comparable. In the large $D$ limit,
however, the thickness of the gravitational tail turns out to
scale like $r_0/D$ \cite{Emparan:2013moa} and so is much smaller than the the black hole
size.
This observation suggests the possibility of an effective `dimensional
reduction' of black hole dynamics to the membrane region; a slab of spacetime
of thickness $1/D$ centered around the codimension one event horizon.
In work done over the last year, this expectation has been borne out in
various contexts. In this paper we will focus on black holes propagating
in an otherwise unperturbed flat space. Assuming that $r_0$ (see above)
and the length scale of variation along the horizon are both of order
unity, the dimensional reduction described above was worked out to leading
nontrivial order in the $1/D$ expansion for the most general nonlinear
dynamical context in \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk};
the special case of stationary solutions and their small fluctuations has
also been studied at higher orders in the $1/D$ expansion in
\cite{Emparan:2015hwa,Suzuki:2015iha,Tanabe:2015isb, Tanabe:2016opw}.
In addition the dimensional reduction of small horizon ripples at
length scale $1/\sqrt{D}$ about particular solutions (black strings or black
branes in flat, $AdS$ or $dS$ space) has been studied in
\cite{Emparan:2015gva,Suzuki:2015axa,Tanabe:2015hda,Emparan:2016sjk,Tanabe:2016pjr}. Further developments were presented in
\cite{Sadhu:2016ynd,Herzog:2016hob,Rozali:2016yhw,Chen:2015fuf,Giribet:2013wia,Prester:2013gxa,Chen:2016fuy}.
In this paper we further develop the general nonlinear dynamical
construction of \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}.
In particular we demonstrate that the reduction of black hole dynamics to
membrane dynamics, worked out to leading nontrivial order in the $1/D$
expansion in \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}, can be
systematically generalized to every order in $1/D$. As an application
of this systematic framework we explicitly work out the first subleading
corrections to the membrane equations of motion in the $1/D$ expansion, and
also determine the spacetimes dual to any particular membrane solution
at next subleading order in the $1/D$ expansion. In this introduction
we first review the leading order construction presented in
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} and then present our
explicit higher order results.
\subsection{Review of earlier work}
Consider a class of $D$ dimensional metrics of the form
\begin{equation}\label{ansatz0}
g_{MN}= \eta_{MN}+ \frac{ (n_M-u_M)(n_N-u_N) }{\psi^{D-3}}
\end{equation}
The metrics \eqref{ansatz0} are parameterized by a smooth $D$ dimensional function $\psi$ and a smooth
oneform field $u_M$. $n_M$ in \eqref{ansatz0} is the normal field to
surfaces of constant $\psi$, (i.e. $n_M= \frac{\partial_M \psi} {\sqrt{\partial_P \psi \partial_Q\psi \eta^{PQ}}}$). The oneform field $u_M$ is assumed to be unit normalized (i.e. $u_N u_M \eta^{MN}=-1$) and tangent to surfaces of
constant $\psi$ (i.e. $u_M n_N \eta^{MN}=0$).
In order to gain intuition for spacetimes of the form \eqref{ansatz0} it is useful to first consider
a special case. Working with coordinates in which the metric on Minkowski space takes the form
$$ds^2=-dt^2 + dr^2 + r^2 d \Omega_{D-2}^2, $$
the choice $u=-dt$ and $ \psi= \frac{r}{r_0}$
turns \eqref{ansatz0} into the metric of a Schwarzschild black hole of radius $r_0$
in the so called Kerr Schild coordinates.
Note $\psi =1$ is the event horizon of the Schwarzschild black hole. More generally the surface
$\psi=1$ is easily verified to be a null submanifold of \eqref{ansatz0} for every choice of
$\psi$ and $u$. This null manifold coincides with the event horizon of the \eqref{ansatz0} provided that
$\psi$ and $u$ are chosen such that the metric \eqref{ansatz0} settles down into a collection of stationary
black holes at late times. Following
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} we refer to the submanifold $\psi=1$ as
the membrane world volume. \footnote{Through this paper we assume that $\psi$ in \eqref{ansatz0} is chosen to
ensure that the membrane surface is a smooth codimension one surface that is timelike when viewed
as a submanifold of flat space (we have emphasized above that this surface is a null submanifold of
the metric \eqref{ansatz0}). We also assume that
$\psi$ is chosen to ensure that $\frac{1}{\psi^{D-3}}$ decays at spatial
infinity.}
Note that as $\psi$
increases past unity $\frac{1}{\psi^{D-3}}$ decays to zero very rapidly.
This decay is exponential in $D$ once $\psi-1 \gg \frac{1}{D}$. It follows
that \eqref{ansatz0} represents a class of asymptotically flat spacetimes
with the following property; the spacetime outside the event horizon deviates
significantly from flat space only in a slab of thickness $\frac{1}{D}$
around the event horizon. We will refer to this as the membrane region.
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} set out to characterize
solutions of the vacuum Einstein equations, $R_{MN}=0$, that reduce to
metrics of the form \eqref{ansatz0} in the large $D$ limit, with corrections
in a power series in $\frac{1}{D}$. As we have reviewed above, when $\psi-1 \gg \frac{1}{D}$ the spacetimes \eqref{ansatz0} reduce to flat space.
Deviations from flatness are nonperturbatively small
in the $\frac{1}{D}$ expansion. Thus Einstein's equations are automatically
solved at all order in $1/D$ outside the membrane region. In order to
obtain a true solution of Einstein's equations, the solution
\eqref{ansatz0} needs to be corrected order by order in the $\frac{1}{D}$
expansion only in the membrane region.
Consider a region of size $\frac{1}{D}$ centered around any point $x_0$ on the
event horizon of \eqref{ansatz0}. It may be shown that the metric of this ball is closely
approximated by the metric in an equivalent small region centered around
the appropriate event horizon point of {\it some} boosted Schwarzschild
black hole provided that
\begin{equation}\label{ansatzt}
\nabla^2 \left( \frac{1}{\psi^{D-3}} \right)=0, ~~~\nabla.u=0,
\end{equation}
(the contraction of all indices is achieved by use of the metric
$\eta_{MN}$ in the equations above) \footnote{ When an expression like $\nabla^2$ acts
on $\frac{1}{\psi^{D-3}}$ we get two distinct terms of order $D^2$ in two ways.
The first term is $\propto (D-3)(D-2) \frac{(\nabla \psi)^2}{\psi^{D-1}}$.
The second term is $\propto (D-3) \frac{\nabla^2 \psi}{\psi^{D-2}}$. Though
the second term has one less explicit factor of $D$ than the first,
it actually contributes at the same order in the $1/D$ expansion - i.e.
at leading order - because of the contraction of indices in $\nabla^2$.
This is the reason that \eqref{ansatz0} solves the leading order equations
only if $\nabla^2 \psi$ takes the same value as it does in a Schwarzschild
black hole, leading to the first requirement listed in \eqref{ansatzt}.
In a similar manner worldvolume
derivatives of the horizon shape and velocity field - which are of order
unity - compete with derivatives acting on $\frac{1}{\psi^{D-3}}$ only if their
order is enhanced by the contraction of a worldvolume index.
The only first derivative expression involving the black hole velocity that
has such a contraction is $\nabla.u$. It follows that \eqref{ansatz0}
satisfies the leading order equations only if $\nabla.u$ takes the same value
as it does on a Schwarzschild black hole. This leads to the second of
\eqref{ansatzt}.}
. These equations need only be satisfied
at leading order in $D$ and can be violated at subleading orders.
As Schwarzschild black holes are exact solutions to
Einstein's equations, it follows
as a consequence that the spacetimes \eqref{ansatz0} {\it almost}
solve Einstein's equations in the membrane region, provided that \eqref{ansatzt}
is satisfied at every point on the membrane.
The statement that Einstein's equations are `almost' solved in the membrane
region has the following precise meaning. When evaluated in the membrane
region the four derivative scalar $R_{AB} R^{AB}$ is in general of order $D^4$.
This estimate follows immediately from the fact that the metric
varies on a length scale of order $1/D$ in the membrane region. Once we impose
\eqref{ansatzt}, on the other hand, $R_{AB} R^{AB}$ turns out to be of order $D^2$, i.e. In a coordinate system in which all components of the metric
are of order unity, $R_{AB}$ is of order $D$; one order lower than the
generic order suggested by a dimensional estimate.
In other words \eqref{ansatzt} ensures that Einstein's equations are obeyed to
leading order - but are generically violated at first subleading order.
Consequently the metrics
\eqref{ansatz0} - with the conditions \eqref{ansatzt} imposed at leading order- are plausible
starting points for the construction of true
solutions of Einstein's equations in a power series in $\frac{1}{D}$.
The authors of \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} were able to
carry out this perturbative expansion to first subleading order in $\frac{1}{D}$
(see below for a review). Interestingly they discovered that
arbitrary metrics of the form \eqref{ansatz0} could {\it not} be corrected
to yield regular solutions to Einstein's equations at next order in
$\frac{1}{D}$. It turns out to be possible to correct
\eqref{ansatz0} at first order in $1/D$ only when the fields $\psi$
and $u$ obey an integrability constraint - a membrane equation of motion -
that we will describe in considerable detail below. Whenever this condition
is obeyed, a regular correction (of order $1/D$) to the metric \eqref{ansatz0}
was found in \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}.
The corrected metric obeys $R_{AB}={\cal O}(1)$
\footnote{More precisely, $R_{AB}={\cal O}(1)$ in coordinates in which
all metric components are of order unity. More generally,
$R_{AB} R^{AB}$ is of order unity. } ;
i.e. once the corrections are taken into account,
Einstein's equations are solved at leading {\it and first subleading order}
in $\frac{1}{D}$.
We now turn to a description of the integrability constraints mentioned
in the previous paragraph. Consider the surface
$\psi=1$, viewed
as a submanifold of flat space with metric $\eta_{MN}$; we refer to this
submanifold as the membrane. Let $K_{MN}$ represent
the extrinsic curvature of this (generically timelike) submanifold.
Recall also that the velocity oneform field $u_M$ on the membrane surface
is tangent to the membrane and so may be regarded as a oneform field in the
membrane world volume. The authors of \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} found that the metric \eqref{ansatz0}
could be corrected to a regular \footnote{By a regular solution we mean a
solution with a smooth event horizon that is regular everywhere outside
the event horizon.} solution of Einsteins equations
at first order if and only if the following constraints are obeyed
\begin{equation} \label{VE1copy}
\left(\frac{\nabla^2 u_{A}}{\mathcal{K}}- \frac{\nabla_A\mathcal{K}}{\mathcal{K}}+ u_{C}K^C_A-u.\nabla u_A \right)\mathcal{P}^A_B=0
\end{equation}
where $\mathcal{P}^A_B=\delta^A_B + u^A u_B$ is the projector orthogonal
to the velocity vector on the membrane world volume, and all covariant
derivatives are taken with respect to the induced metric on the
membrane. The quantity $\mathcal{K}$ is the trace of the extrinsic curvature of
the membrane worldvolume.
The integrability conditions \eqref{VE1copy} have an interesting interpretation.
They may be thought of as a set of $D-2$ equations for $D-2$ variables
(one of these variables is the shape of the membrane, and the other $D-3$
variables are the components of the unit normalized, divergence free
velocity field). In other words the equations \eqref{VE1copy} define an
initial value problem for membrane dynamics. As every configuration that
obeys \eqref{VE1copy} gives rise to a metric that obeys Einstein's equations to
the appropriate order in $1/D$, it follows that solutions of the membrane
equations \eqref{VE1copy} are in one to one correspondence with asymptotically
flat dynamical black hole configurations that solve Einstein's equations
to first subleading order in $1/D$.
\subsection{The membrane paradigm at higher orders in $1/D$}
In this paper we demonstrate that first order perturbative procedure outlined above
extends systematically to arbitrary orders in the expansion in $\frac{1}{D}$. We will now
very briefly outline our inductive argument. We assume that the perturbative procedure
has been implemented upto $n^{th}$ order, i.e. that corrections to the metric
\eqref{ansatz0} have been determined upto $n^{th}$ order in the $1/D$ expansion
in such a manner that $R_{MN}$ evaluated on the corrected solution is of order
$D^{1-n}$. We then add further corrections of order $1/D^{n+1}$ to the metric
(see \eqref{pertthy} and \eqref{metcorr}). At order $D^{n-1}$ we demonstrate that the
Einstein constraint equations are independent of
the new unknown correction functions when evaluated on the event horizon
$\psi=1$. These equations determine the correction
to the membrane equations (and the divergence condition on the velocity) at
order $1/D^{n+1}$. Moving away from the horizon we argue that the order $D^{1-n}$ part of
$R_{MN}$ takes the form listed in table \ref{Rbasisex}. Setting the expressions
in this table yields a set of inhomogeneous linear differential equations that can be used
to determine order $1/D^{n+1}$ corrections to the metric. Explicit expressions
for the sources in these differential equations can only be obtained
by grinding through the perturbative procedure, but we use
a contracted Bianchi identity to demonstrate that the sources that occur in these equations
are not all independent, but obey certain relations (see \eqref{bianchisrc})
at every order of perturbation theory.
Using these relations we are able to integrate the inhomogeneous differential
equations for any source functions and obtain an explicit and unique expressions
for the metric corrections at order $1/D^{n+1}$ (see Section \ref{pfo}) that
are manifestly regular and obey all required boundary conditions.
As an illustration of the general method outlined above we explicitly
implement the perturbative procedure to second subleading order in
$\frac{1}{D}$. We find that the modified membrane equations take the form
\begin{eqnarray}\label{VE2copy}
\nonumber&&\Bigg[\frac{\nabla^{2}u_{A}}{\mathcal{K}}-\frac{\nabla_{A} \mathcal{K}}{\mathcal{K}}+u^{B} K_{BA}-u\cdot\nabla u_{A}\Bigg]{\cal P}^{A}_{C} \\\nonumber&+& \Bigg[\left(-\frac{u^{C}K_{CB}K^{B}_{A}}{\mathcal{K}}\right)+\left(\frac{\nabla^2\nabla^2 u_{A}}{\mathcal{K}^3}-\frac{u \cdot \nabla \mathcal{K}\nabla_{A}\mathcal{K}}{\mathcal{K}^3}-\frac{\nabla^{B}\mathcal{K}\nabla_{B}u_{A}}{\mathcal{K}^2}-2\frac{K^{CD}\nabla_{C}\nabla_{D}u_{A}}{\mathcal{K}^2}\right) \\ \nonumber &+&\left(-\frac{\nabla_{A}\nabla^2 \mathcal{K}}{\mathcal{K}^3} +\frac{\nabla_{A}\left(K_{BC}K^{BC}\mathcal{K}\right)}{\mathcal{K}^3}\right) + 3\frac{(u\cdot K\cdot u)(u\cdot \nabla u_{A})}{\mathcal{K}}-3\frac{(u\cdot K\cdot u)(u^{B} K_{BA})}{\mathcal{K}}\\ &-&6\frac{(u \cdot \nabla \mathcal{K})(u\cdot \nabla u_{A})}{\mathcal{K}^2}+6\frac{(u \cdot \nabla \mathcal{K})(u^{B}K_{BA})}{\mathcal{K}^2} + \frac{3}{(D-3)} u \cdot \nabla u_{A}-\frac{3}{(D-3)} u^{B}K_{BA} \Bigg]{\cal P}^{A}_{C}=0\nonumber\\
\end{eqnarray}
while the divergence free condition on the velocity field is modified, at second subleading order, to the equation
\begin{equation}\label{Divu}
\nabla\cdot u=\frac{1}{2\mathcal{K}}\left(\nabla_{(A}u_{B)}\nabla_{(C}u_{D)}{\cal P}^{BC}{\cal P}^{AD}\right)
\end{equation}
Note that the first line in \eqref{VE2copy} is simply a
rewriting of \eqref{VE1copy}; the 2nd-4th lines of this equations represent
corrections to \eqref{VE1copy}. There is a well defined sense (see below) in
which each of these correction terms is of order $\frac{1}{D}$ relative
to the leading order terms in the first line. It follows that the
equations \eqref{VE2copy} represent small corrections to the leading order
equations \eqref{VE1copy}. The first order corrected membrane equation
of motion \eqref{VE2copy} and \eqref{Divu} are the main result of this paper.
We then present explicit expressions for the second order sources
for all the inhomogeneous differential equations (see table \ref{Rbsrc2}).
Plugging these sources into the general equations for the metric corrections
at any order we obtain explicit results for the second order correction to
the spacetime metric dual to any particular solution of the membrane equations
of motion.
The second order corrected membrane equations \eqref{VE2copy} admit a simple solution; a spherical
membrane at rest. This solution is dual to the Schwarzschild black hole.
As a check of our second order corrections to the membrane equations we use
\eqref{VE2copy} to compute the spectrum of small fluctuations
about this simple solutions. This spectrum is easy to obtain, and turns
out to be in perfect agreement with the second order corrected spectrum
of quasinormal modes obtained by Emparan Suzuki and Tanabe in
\cite{Emparan:2014aba}, providing confidence in the correctness of
\eqref{VE2copy}.
\section{Perturbation theory: general structure} \label{pertgen}
\subsection{A more detailed description of the starting ansatz} \label{md}
As we have explained in the introduction, the starting point of our perturbative construction
of large $D$ solutions to Einstein's equations is the metric \eqref{ansatz0}. In the introduction
we noted that the metrics \eqref{ansatz0} are parameterized by the $D$ dimensional function $\psi$
and the oneform field $u$. We assume these fields have a good large $D$ limit,
i.e. that the length scale of variation in $\psi$ and $u$ is of order unity.
Following
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}, however, consider
two different functions $\psi$ with the same membrane surface (i.e. with coincident zero sets for $\psi -1$). These two functions define metrics
\eqref{ansatz0} that coincide (outside the event horizon) at leading order
in $1/D$ but differ at subleading orders in $1/D$.
Similarly $u$ functions that agree on the membrane
but differ off it lead to metrics \eqref{ansatz0} that differ only at subleading order in $1/D$.
Any two metrics \eqref{ansatz0} that differ only at subleading orders in $1/D$ constitute equivalent
starting points for the perturbative construction of solutions in the following sense: the end
result of perturbation theory starting from the two different starting points will be the same.
In order to construct all distinct final metrics we need only consider one member of each `equivalence
class' of metrics \eqref{ansatz0}. As explained above the equivalence classes are labeled by
the zero set of the function $\psi -1$ (the membrane world volume) and the value of the velocity
field on the membrane world volume. In order to pick a representative from each equivalence class that
we can use to set up our perturbation theory we invent an arbitrary way of constructing the full
function $\psi$ from its zero set, and the full velocity field $u$ from its values on the membrane.
Following \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} we refer to the (essentially arbitrary)
rule for achieving this construction as a subsidiary condition on the functions $\psi$ and $u$.
For technical reasons, in this paper we utilize the subsidiary conditions of \cite{Bhattacharyya:2015dva} rather than that of \cite{Bhattacharyya:2015fdk}. We now describe these conditions in detail.
Consider a given timelike membrane submanifold in flat space. At each point on the manifold consider a
geodesic that shoots outwards from the manifold along its normal vector. The
resultant collection of curves \footnote{These `curves' are actually straight
lines as they are all geodesics in flat space. We use the term `curve' to bring
to mind the obvious generalization of this construction when the membrane is
embedded in a curved spacetime.}
is a
spacefilling congruence of spacelike geodesics; caustics of this congruence, if any, only occur
at distances of order unity (rather than $1/D$) away from the membrane.
\footnote{The quantity $\frac{D}{{\mathcal K}}$ gives a rough estimate for the
distance away from the membrane at which the geodesics caustic. Below we
explain that ${\mathcal K}$ is of order $D$ so that this caustic length
scale is of order unity. }
We define the scalar function
$B$ in the neighborhood of the membrane as follows; $B$ at any point is defined to be the signed proper distance, along the geodesic that passes through it, to the membrane. This distance is defined to be positive
outside the membrane and negative inside the membrane. Note that $B$ vanishes on the membrane. We define
\begin{equation}\label{ndef}
n_M=\nabla_M B
\end{equation}
It follows from our construction above that
\begin{equation} \label{nnorm}
n.n=1
\end{equation}
$n_A$ is the normal oneform to surfaces of constant
$B$. We use the symbol $K_{MN}$ denote the extrinsic curvature of surfaces of constant $B$. Note
of course that $n^AK_{AB}=0$. We also define ${\mathcal K}=K_A^A$. We then proceed to define the
function $\psi$ as
\begin{equation}\label{psidef}
\psi= 1 + \frac{{\mathcal K} B}{D-3}
\end{equation}
In a similar manner we use the velocity function on the membrane to define a velocity oneform
field in spacetime simply by parallel transport along our congruence of geodesics. It follows from
our definitions above that
\begin{equation}\label{auxeq} \begin{split}
&n.\nabla n_A=0 \\
& n.\nabla u_A=0
\end{split}
\end{equation}
The first line of \eqref{auxeq} follows upon differentiating 0
\eqref{nnorm}, using \eqref{ndef} and interchanging derivatives.
This equation is in fact simply the geodesic equations
for the congruence of geodesics that defines $B$.
The equation on the second line of
\eqref{auxeq} follows from the fact that $u$ is defined off the membrane
by parallel transport. It follows from \eqref{auxeq} that
\begin{equation}\label{kab}
K_{AB}= (\eta_A^C-n_A n^C)\left( \nabla_C n_D \right)
\left( \eta^D_B -n^D n_B \right)
= \left( \nabla_A - n_A (n. \nabla) \right) n_B
= \nabla_A n_B= \nabla_A \nabla_B B
\end{equation}
Note that our definition of $n_A$ in this section, and the rest of this paper,
differs slightly from the definition given
in the introduction. The two definitions agree at leading order (which was all that was
required in the discussion around \eqref{ansatz0} ) but differ at subleading orders in $1/D$. The
vector $n_A$ defined in this section - rather than the normal vector defined
in the introduction - will be used through the rest of this paper.
Using \eqref{psidef} it is easily verified that on the submanifold $B=0$
\begin{equation}\label{psicond} \begin{split}
&\psi \nabla^2 \psi= \frac{{\mathcal K}^2}{D-3} + 2 \frac{n.\nabla \mathcal{K}}{D-3}\\
&(D-2) \nabla \psi. \nabla \psi =\frac{D-2}{D-3}\frac{{\mathcal K}^2}{D-3}
\end{split}
\end{equation}
As we explain below, in the large $D$ limit taken in this paper $2 \frac{n.\nabla \mathcal{K}}{D-3}$
is of order unity while $\frac{{\mathcal K}^2}{D-3}$ is order $D$. It follows that to leading order
in $D$
$$ (D-2) \nabla \psi. \nabla \psi= \psi \nabla^2 \psi, ~~~i.e.
\nabla^2 \left( \frac{1}{\psi^{D-3}} \right) =0$$
In other words our construction satisfies the first equation of \eqref{ansatzt}. We satisfy the second
equation in \eqref{ansatzt} by construction; we simply choose our $u$ oneform on the membrane such that its divergence vanishes at leading order in $D$.
The divergence of $u$ will turn out not to vanish at a subleading order.
\subsection{Coordinate Choice for the correction metric}
In this paper we search for solutions of Einstein's equations in a power series expansion in
$\frac{1}{D}$
\begin{equation} \label{pertthy}
\begin{split}
G_{MN}&= \eta_{MN} + h_{MN},\\
h_{MN}&= \sum_{n=0}^\infty \frac{h_{MN}^{(n)}}{(D-3)^n}\,,\\
\text{with, }h_{MN}^{(0)}&= \frac{O_M O_N}{\psi^{D-3}} ,\\
\end{split}
\end{equation}
Here
\begin{equation}\label{Odef}
O_M=n_M-u_M
\end{equation}
We fix coordinate redefinition ambiguities by demanding
\begin{equation}\label{corrcon}
h_{MN}O^N=0,
\end{equation}
Consider any point in the metric \eqref{ansatz0}. The tangent space built about
this point has two special vectors; the vector $n$ and the vector $u$. All the
other $D-2$ directions orthogonal to $n$ and $u$ are equivalent and can be
rotated into each other. It is thus useful to parameterize the most general
fluctuation field $h_{MN}$ (subject to the gauge condition
\eqref{corrcon}) in the form
\begin{equation}\label{metcorr}
\begin{split}
h_{MN}^{(n)} &= H^{(S,n)} O_M O_N +O_{(M} H_{N)}^{(V,n)} +H^{(T,n)}_{MN} + \frac{1}{D-3} H^{(Tr,n)} {\cal P}_{MN},\\
\text{where,}&\\
{\cal P}_{MN}= &\eta_{MN}-O_M n_N -O_N n_M + O_M O_N,\\
O^{N} H_{N}^{(V,n)}&=0,~~~n^{N} H_{N}^{(V,n)}=0, ~~~O^{M} H^{(T,n)}_{MN} =0,~~~n^{M} H^{(T,n)}_{MN} =0, ~~~{\cal P}^{MN}H^{(T,n)}_{MN}=0,
\end{split}
\end{equation}
The superscripts $S$, $V$ and $T$ stand for scalar, vector and tensor
respectively, and denote the transformation properties of the relevant
symbol under the $SO(D-2)$ rotations in tangent space that leave $n$ and
$u$ fixed. The superscript $Tr$ stands for trace, and labels a second scalar.
\subsection{Orders of $D$}
As we have explained above, in this paper we solve Einstein's equations
in a systematic expansion in $\frac{1}{D}$. In order for this process to
be well defined, we need to be able to unambiguously estimate the scaling
with $D$ of various terms that appear in the metric and in the membrane
equation of motion. Such an estimation is only unambiguous within
subclasses of solutions, as we will now explain with an example.
Consider a membrane whose world volume is a $D-2$ sphere (of radius $R$)
times time. The trace of extrinsic curvature, ${\mathcal K}$,
of this surface is easily shown to be $\frac{D-2}{R}$ and so is of order
$D$ (assuming $R$ is of order unity). On the other hand
the surface $S^p \times R^{D-2-p}$ times time has
${\mathcal K}=\frac{p}{R}$. If $p$ and $R$ are both held fixed as $D$ is taken
to infinity, ${\mathcal K}$ is of order unity for this surface. It follows
that ${\mathcal K}$ cannot unambiguously be assigned a scaling with $D$
without making further assumptions. The same holds true of various other
quantities (e.g. $\nabla^2 u_M$) that enter the metric and equation of motion.
In this paper we follow \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}
and estimate the $D$ scalings of all terms as follows. We assume that
\begin{itemize}
\item Our starting ansatz is constructed by sewing together bits of the
event horizon of black holes of radii $R$ and timelike velocity $u^M$ where
$R$ and $u^M$ are everywhere finite and of order unity.
\item Our starting configuration (and so our full solution) preserves
an $SO(D-p-2)$ rotational invariance with $p$ held fixed as $D$ is taken
to infinity
\end{itemize}
As explained in \cite{Bhattacharyya:2015fdk}, these assumptions unambiguously
specify the scaling with $D$ of all quantities of interest (in particular
they force ${\mathcal K}$ to be of order $D$).
We emphasize that in this paper we use the assumptions listed above only
to estimate the scalings of $D$ of various quantities. When the assumptions
listed in the previous paragraph are obeyed, the membrane equations
and metrics listed in this paper certainly apply. However the formulae
of this paper apply more generally to any spacetime whose variables
scale with $D$ in the same manner in which they would if the assumptions
above were obeyed - a much larger class of configurations.
\subsection{All orders definition of the membrane surface and velocity}
\label{sv}
As explained in subsection \ref{md}, the metric \eqref{ansatz0} - the starting
point of our perturbative expansion - is completely determined by the
shape of a membrane and a velocity field on the membrane. To what precision
can this procedure be reversed? In other words if we are given a solution
to Einstein's equations of the appropriate kind, how precisely can we
read off the corresponding `shape' and `velocity' of the membrane?
We could attempt to identify the membrane shape and velocity field by
simply expanding the exact solution in powers of
$1/D$ and focusing attention on the leading order term. By comparing with \eqref{ansatz0} we could then read off
the membrane shape and velocity field. While this procedure is simple,
a moment's thought will convince the reader that it is ambiguous at
all orders in $1/D$ save the leading order. \footnote{For instance, the
velocity redefinition $u^\mu \rightarrow u^\mu + \delta u^\mu/D$ does not
change the metric at leading order in $1/D$.} In other words the requirement
that our solution reduce to \eqref{ansatz0} defines the membrane shape and
velocity only at leading order, leaving the subleading corrections to
these quantities ambiguous. In this subsection we will fix this ambiguity
by adopting a more precise definition of the shape and velocity field. This
definition agrees with that of \eqref{ansatz0} at leading order, but
is precise at all orders. We use this precise definition
in the computations presented in the rest of this paper.
We define the membrane shape to be the location of the event horizon of our
spacetime, and will choose higher order corrections to the metric
\eqref{ansatz0} to ensure that this event horizon coincides with the
surface $\psi=1$.
Turning to the velocity field, let $G^{AB}$ denote the full spacetime
inverse metric. Let $n_A$ be the oneform normal to the event horizon.
We define the velocity field on the membrane by the requirement that
\begin{equation}\label{vfm}
u^A= G^{AB}n_B
\end{equation}
on the event horizon (i.e. at $\psi=1$). In other words the velocity field
is a tangent vector to the generators of
the event horizon. It is easily verified that \eqref{vfm} is
a true equation for the starting point of perturbation theory
\eqref{ansatz0}. We will choose corrections to the perturbative ansatz
to ensure that \eqref{vfm} holds at all orders in $1/D$.
The requirement \eqref{vfm} together with the requirement that $\psi=1$ is the
exact event horizon of our spacetime are easily seen to be
satisfied provided that
\begin{equation}\label{velhor}
\begin{split}
H^{(S)}(\psi=1)&=0\\
H^{(V)}_M(\psi=1)&=0\\
\end{split}
\end{equation}
The first condition ensures that $G^{MN}\partial_M \psi \partial_N \psi =0$,
i.e. $d\psi$ is null at $\psi=1$ while the second
condition then ensures that the full spacetime metric on the event horizon
takes the form
$$ \eta_{MN} + O_M O_N+ H^{(T)}_{MN} + \frac{1}{D-3} H^{Tr} {{\cal P}}_{MN}$$
Let us write this metric in a the local basis of oneforms
$(n, u, Y_{a})$ where $Y_a$ is any $D-2$ dimensional basis of oneforms
chosen orthogonal to $n$ and $u$. In this basis the metric takes a block
diagonal form with a $2 \times 2$ block (with basis $n$ and $u$) and a
$D-2 \times D-2$ block (with basis $Y_a$). It follows that the inverse metric
also has this block diagonal structure. Note that the $2 \times 2$
block is universal, i.e. it is the same at every order in perturbation theory.
This block is the only one that contributes in \eqref{vfm}. As
\eqref{vfm} holds at leading order, it follows
that the conditions \eqref{velhor} ensure that \eqref{vfm} holds at every
order in perturbation theory.
Recall that according to \eqref{ansatzt} the velocity field used in
\eqref{ansatz0} is divergence free at leading order in $\frac{1}{D}$.
As we will see below, the divergence of the velocity field defined in this
subsection will not, in general, vanish at subleading orders in $1/D$.
\subsection{Structure of the equations of perturbation theory}
Our perturbative procedure proceeds as follows. We assume that our
solution takes the form \eqref{pertthy} together with
\eqref{corrcon} and \eqref{metcorr}. The Ricci tensor of this metric -
evaluated in a slab of spacetime of thickness $1/D$ around $\psi=1$ -
takes the schematic form
\begin{equation}\label{rt}
R_{MN}= \sum_n D^{2-n} R_{MN}^{n}
\end{equation}
Let us imagine that we have implemented our perturbative procedure to order
$n-1$, i.e. that we have determined $h^{(m)}_{MN}$ for $m=1 \ldots n-1$ in a
manner that ensures that $R_{MN}^{(m)}=0$ for $m= 0 \ldots n-1$.
In order to go to one higher order in perturbation theory we must solve
for $h_{MN}^{(n)}$ to ensure that $R_{MN}^n$ also vanishes.
Schematically
$$ R_{MN}^{(n)}= C_{MN}^{PQ} h^{(n)}_{PQ} + {\mathcal S}^{(n)}_{MN}$$
where $C_{MN}^{PQ}$ is a linear differential operator with derivatives only in
the $\psi$ direction and ${\mathcal S}^{(n)}_{MN}$ is a source function.
As $h^{(n)}_{PQ}$ is already of order $n$, the differential operator
$C_{MN}^{PQ}$ is built entirely out of the zero order background metric
\eqref{ansatz0}, and so is the same at every order. On the other
hand the source function ${\mathcal S}^{(n)}_{MN}$ is proportional to expressions
of $n^{th}$ order in $1/D$ built out of derivatives of the membrane velocity
and shape function, and is different at every order.
At every point of the event horizon of the ansatz metric \eqref{ansatz0}
there are two distinguished vectors; $n^A$ and $u^A$. Let
$${\cal P}_{AB}= \eta_{AB} -n_A n_B + u_A u _B$$
denote the projector orthogonal to these two vectors (all dot products
taken in flat space). Instead of dealing directly with the components of
$R_{MN}$ we find it more convenient to use a basis adopted to $u^A$ and $n^A$ listed in table \ref{Rbasis}.
\begin{center}
\begin{table}[h]
\caption{Basis of components of $R_{MN}$}\label{Rbasis}
\begin{tabular}{ |c|c|c| }
\hline
Scalar sector & Vector sector & Tensor sector \\
\hline
$R^{S_1} = O^M R_{MN} O^N $& $ R^{V_1}_{L}= O^{M}R_{MN}{\cal P}^{N}_{L} $ & $R^{T}_{AB} = {\cal P}^M_A R_{MN} P^N_B - \frac{{\cal P}_{AB}}{D-2}{\cal P}^{MN}R_{MN} $ \\
$R^{S_2} = O^M R_{MN} u^N $& $R^{V_2}_{L}= u^{M}R_{MN}{\cal P}^{N}_{L}$& $~$ \\
$ R^{S_3} = u^M R_{MN} u^N $ & $~$ & $~$ \\
$R^{S_4}= R_{MN}{\cal P}^{MN} $& $~$ & \\
\hline
\end{tabular}
\end{table}
\end{center}
\noindent
By explicit computation (plugging \eqref{pertthy} into the formula for
the Ricci tensor) we find that the linear combinations listed in
Table \ref{Rbasis} of the curvature components $R^n_{MN}$ (see \eqref{rt})
are given by the expressions listed in Table \ref{Rbasisex}.
In table \ref{Rbasisex}, fluctuation fields $H^S$, $H^{Tr}$ $H^V_A$ and $H^T_{MN}$ are taken
to be of $n^{th}$ order and all source functions
(e.g. ${\mathcal S}^{S_1}$) also understood to be $n^{th}$ order sources.
All appearances of $\nabla.u$
\footnote{$\nabla.u$ is the divergence of the velocity field thought of
as a vector field in $R^{D-1,1}$. On the surface $\psi=1$, however,
$\nabla.u$ coincides with the membrane worldvolume divergence of
velocity field (this follows upon using the second of \eqref{auxeq}).}
in the table \ref{Rbasisex} should also be understood
as follows. Naively $\nabla.u$ is of order $D$. For that reason we expand
\begin{equation}\label{nu}
\nabla.u=(D-3) \left( \sum_{n=0}^\infty \frac{(\nabla.u)_n}{(D-3)^{n}} \right)
\end{equation}
Every appearance of $\nabla.u$ in table \ref{Rbasisex} should actually
be replaced by $(\nabla.u)_n$. We have already seen in the introduction
that $(\nabla.u)_0=0$. We will see below that $(\nabla.u)_1$ also vanishes,
but that $(\nabla.u)_2$ is nonzero.
\begin{center}
\begin{table}[t]
\caption{Expressions for basis of $R_{MN}$}\label{Rbasisex}
\resizebox{\columnwidth}{!}{
\begin{tabular}{ |c| }
\hline
Scalar sector \\
\hline
$R^{S_1} = \left( \frac{-\mathcal{K}^2}{2(D-3)^2} \right) \frac{d^2 H^{(Tr)}}{dR^2}+{\mathcal{S}}^{S_1}(R) $\\
$R^{S_2} = \left( \frac{\mathcal{K}^2}{2(D-3)^2} \right) e^{-R}\frac{d}{dR}\left(e^R\frac{d}{dR}H^{(S)}\right) - \frac{\mathcal{K}^2}{4(D-3)^2}e^{-R}\frac{d}{dR}H^{(Tr)} + \frac{\mathcal{K}}{2(D-3)}\nabla^MH^{(V)}_M $\\$+ {\mathcal{S}}^{S_2}(R) + \frac{\mathcal{K}}{2(D-3)}e^{-R}~\nabla.u $\\
$R^{S_3} = \left( \frac{\mathcal{K}^2}{2(D-3)^2} \right)e^{-2R}(1-e^R)\frac{d}{dR}(e^R\frac{dH^{(S)}}{dR}) $\\$ - \left( \frac{\mathcal{K}^2}{4(D-3)^2} \right)e^{-2R}(1-e^R)\frac{dH^{(Tr)}}{dR} - \frac{\mathcal{K}}{2(D-3)}e^{-R}\nabla^MH^{(V)}_M + {\mathcal{S}}^{S_3}(R) + \frac{\mathcal{K}}{2(D-3)}~e^{-2R}~\nabla.u$ \\
$R^{S_4} = \left( \frac{\mathcal{K}^2}{(D-3)^2}\right) e^{-R}\frac{d}{dR}(e^R H^{(S)}) + \left( \frac{\mathcal{K}^2}{2(D-3)^2} \right)e^{-2R}(1-e^R)\frac{d}{dR} \left( e^R \frac{d}{dR}H^{(Tr)} \right) $\\$ - \left( \frac{\mathcal{K}^2}{2(D-3)^2} \right) \frac{dH^{(Tr)}}{dR} + \frac{\mathcal{K}}{D-3}\nabla^M H^{(V)}_M + \frac{2\mathcal{K}}{D-3}\frac{d}{dR}\nabla^M H^{(V)}_M + \nabla^M \nabla^N H^{(T)}_{MN} + {\mathcal{S}}^{S_4}(R)-~\frac{\mathcal{K}}{(D-3)} e^{-R}\nabla.u$ \\
\hline
Vector sector \\
\hline
$R_M^{V_1} =\left( \frac{\mathcal{K}^2}{2(D-3)^2} \right) e^{-R} \frac{d}{dR}(e^{R} \frac{d}{dR} H^{(V)}_{M}) + \frac{1}{2}\frac{\mathcal{K}}{(D-3)} \frac{d}{dR} \left( \nabla^N H^{(T)}_{NM} \right)+{\mathcal{S}}^{V_1}_{M}(R) $ \\
$R_M^{V_2} = \left( \frac{\mathcal{K}^2}{2(D-3)^2} \right) e^{-2R}(1- e^{R}) \frac{d}{dR}(e^{R} \frac{d}{dR} H^{(V)}_{M}) + {\mathcal{S}}^{V_2}_{M}(R)$\\
\hline
Tensor sector \\
\hline
$R^{T}_{AB} = \left( \frac{-\mathcal{K}^2}{2(D-3)^2} \right) e^{-R} \frac{d}{dR}\left(\left(e^R-1\right) \frac{dH^{(T)}_{AB}}{dR}\right) + {\mathcal{S}}^{T}_{AB}(R) $ \\
\hline
\end{tabular}}
\end{table}
\end{center}
\noindent
In order to obtain Table \ref{Rbasisex} we have worked in the neighbourhood
of the surface
$\psi=1$ and the variable $R$ is defined by $R=(D-3)(\psi-1)$.
\footnote{We will explain below that the sources listed in Table \ref{Rbasisex}
are not completely independent, but are constrained by the well known relation
\begin{equation}\label{bi}
\nabla^M \left( R_{MN} - \frac{\tilde{R}}{2}G_{MN} \right) = 0
\end{equation} .}
\subsection{The Einstein Constraint Equations}
In the process of solving for the fluctuation fields $h_{MN}^{(n)}$ we
will find the Einstein constraint equations (relevant to the
foliation of our spacetime in slices of constant $\psi$) particularly useful.
We will now provide a careful definition of these equations.
Let us define
\begin{equation} \label{eeqs}
E_{MN}\equiv R_{MN}-\tilde{R} \frac{G_{MN}}{2}
\end{equation}
where $\tilde{R}$ is the Ricci scalar.
The constraint equations are defined by the relations
\begin{equation}\label{EEconstr}
E^{(ec)}_M = E_{MN}G^{NL}n_L
\end{equation}
We have a total of $D$ constraint equations. These equations decompose into two scalars and one
vector under local $SO(D-2)$ rotations.
Let us imagine we have solved for our membrane metric at $(n-1)^{th}$ order in perturbation
theory, and are now attempting to solve for the metric correction at $n^{th}$ order.
If, in this process, we evaluate the constraint equation \eqref{EEconstr} and retain
terms only up to $n^{th}$ order then we need use $G^{NL}$ on the RHS of \eqref{EEconstr}
only at zero order (i.e. from the metric \eqref{ansatz0}), because $E_{MN}$ is already of
$n^{th}$ order. It follows that the $n^{th}$ order scalar and vector constraint equations
are simply linear combinations of the $n^{th}$ order scalars and vectors listed in
table \ref{Rbasis}. We will now determine the relevant linear combinations. In order to
to this we first determine the $n^{th}$ order Ricci scalar ${\tilde R}$ as a linear combination
of the scalars in table \ref{Rbasis}.
\begin{equation}\label{rs}
\tilde{R} = R_{AB}G^{AB}\\ =\left( R^{AB}P_{AB} + O.R.O(1 - e^{-R}) + 2O.R.u \right)
= (R^{S_4}+(1-e^{-R})R^{S_1}+2R^{S_2})
\end{equation}
Using this equation we find
\begin{equation}\label{EEconflat}
\begin{split}
E^{(ec)}_M &= \left( R_{MN}- \frac{\tilde{R}}{2}G_{MN} \right) G^{NL}n_L \\
&= R_{MN}O^N(1-e^{-R}) + R_{MN}u^N -\frac{1}{2}\tilde{R}~ n_M \\
\end{split} \end{equation}
By dotting \eqref{EEconflat} with $n$ and $u$ or by projecting it orthogonal to these vectors
we finally obtain the $n^{th}$ order constraint equations written as linear combinations
of the scalars and vectors in table \ref{Rbasis}.
\begin{equation}\label{DCrel}
\begin{split}
&E^{S_1}=E^{(ec)}_M u^M=(1-e^{-R})R^{S_2}+R^{S_3} \\
&E^{S_2}=E^{(ec)}_M O^M=\frac{1}{2}\left((1-e^{-R})R^{S_1}-R^{S_4}\right) \\
&E^{V}_L=E^{(ec)}_N{\cal P}^N_L=(1-e^{-R})R^{V1}_L+R^{V_2}_L
\end{split}
\end{equation}
The explicit form of the $n^{th}$ order constraint equations is listed in table \ref{Ebasis}
below
\begin{center}
\begin{table}[h]
\caption{Listing of constraint equations}\label{Ebasis}
\resizebox{\columnwidth}{!}{
\begin{tabular}{ |c| }
\hline
Vector constraint \\
\hline
$E^{V}_M = E^{(ec)}_N{\cal P}^N_M =(1-e^{-R})R^{V_1}_M+R^{V_2}_M $\\$ = \frac{1}{2}\frac{\mathcal{K}}{(D-3)}(1-e^{-R}) \frac{d}{dR} \left( \nabla^A H^{(T)}_{AM} \right) + {\mathcal V}_M^{V}(R) $\\
\hline
Scalar constraint 1 \\
\hline
$E^{S_1} = E^{(ec)}_M u^M =(1-e^{-R})R^{S_2}+R^{S_3} $\\$ =\frac{\mathcal{K}}{2(D-3)}(1-e^R) \frac{d}{dR}\left( \nabla^M H^{(V)}_M \right) - \frac{\mathcal{K}}{2(D-3)}e^{-R} \nabla^M H^{(V)}_M + {\mathcal V}^{S1}(R) + \frac{\mathcal{K}}{2(D-3)}~e^{-R}~\nabla.u $\\
\hline
Scalar constraint 2 \\
\hline
$E^{S_2} = E^{(ec)}_M O^M =\frac{1}{2}\left((1-e^{-R})R^{S_1}-R^{S_4}\right)= -\frac{\mathcal{K}}{2(D-3)} \frac{d}{dR}\left( \nabla^M H^{(V)}_M \right)-\frac{\mathcal{K}}{(D-3)} \nabla^M H^{(V)}_M $ \\ $+ \frac{\mathcal{K}^2}{4(D-3)^2}(2-e^{-R})\frac{d}{dR}H^{(Tr)} - \frac{\mathcal{K}^2}{2(D-3)^2}\left( \frac{d}{dR}H^{(S)}+H^{(S)} \right) - \frac{1}{2} \nabla_M\nabla_N H^{(T)}_{MN} + {\mathcal V}^{S2}(R) + \frac{\mathcal{K}}{2(D-3)} e^{-R}~\nabla.u $\\
\hline
\end{tabular}}
\end{table}
\end{center}
\noindent
As in table \ref{Rbasis}, all fluctuation fields in table \ref{Ebasis} should be taken to be
of $n^{th}$ order. The source functions in table \ref{Ebasis} are also of $n^{th}$ order
and are given in terms of the sources in table \ref{Rbasis} and the as yet unknown quantity
$\nabla.u$ by
\begin{equation}\label{constsc}
\begin{split}
&{\mathcal V}^{S_1}(R) =(1-e^{-R}){\mathcal S}^{S_2}(R)+{\mathcal S}^{S_3}(R) \\
&{\mathcal V}^{S_2}(R) =\frac{1}{2}\left[(1-e^{-R}){\mathcal S}^{S_1}(R)- {\mathcal S}^{S_4}(R)\right] \\
&{\mathcal V}^{V}_L(R)=(1-e^{-R}){\mathcal S}^{V_1}_L(R)+{\mathcal S}^{V_2}_L(R)
\end{split}
\end{equation}
Now it is well known that the Einstein tensor obeys the identity
\begin{equation}\label{bianchi}
\nabla_M E^{MN}=0
\end{equation}
It is also well known (and easy to see) that this identity ensures that the `normal' derivative
of the constraint equations is a linear combination of the `in plane' derivatives of Einstein's
equations. \footnote{This is the fact that ensures that if all Einstein constraint equations are solved on one
`time' slice then they are automatically solved on the next `time' slice. In other
words, in order to solve Einstein's equations you need only solve the constraint equations on one time
slice provided you solve the other equations - lets call them the dynamical equations - everywhere.}
Within the perturbation theory of interest to this paper the equation \eqref{bianchi} may be evaluated and projected onto its scalar and vector sectors and shown to be equivalent to the following relations
\begin{equation}\label{bianchirel}
\begin{split}
&\frac{d}{dR}E^{V}_M + E^{V}_M + \frac{(D-3)}{\mathcal{K}} \nabla^N R^T_{NM}=0 \\
&\frac{d}{dR}E^{S_1} + E^{S_1} + \frac{(D-3)}{\mathcal{K}} \nabla^N R^{V_2}_{N}=0 \\ &\frac{d}{dR}E^{S_2} + E^{S_2} + \left( \frac{1}{2}R^{S_1} + R^{S_2} + \frac{1}{2}R^{S_4} \right) + \frac{(D-3)}{\mathcal{K}} \nabla^N R^{V_1}_{N}=0
\end{split}
\end{equation}
Using \eqref{DCrel} the RHS of these relations may be recast in the equivalent form
\begin{equation}\label{biarel}
\begin{split}
&\frac{d}{dR}E^{V}_M + (1-e^{-R})R_M^{V_1} + R_M^{V_2} + \frac{(D-3)}{\mathcal{K}} \nabla^N R^T_{NM}=0 \\ &\frac{d}{dR}E^{S_1} + (1-e^{-R})R^{S_2} + R^{S_3} + \frac{(D-3)}{\mathcal{K}} \nabla^N R^{V_2}_{N}=0 \\ &\frac{d}{dR}E^{S_2} + \frac{1}{2}e^{-R}R^{S_1} + (1-e^{-R})R^{S_1} + R^{S_2} + \frac{(D-3)}{\mathcal{K}} \nabla^N R^{V_1}_{N}=0
\end{split}
\end{equation}
In either form these equations express the $R$ derivatives of the Einstein constraint equations
\eqref{DCrel} in terms of linear combinations of the Einstein equations.
Using the explicit expressions in tables \ref{Rbasisex} and \ref{Ebasis}, it is possible to verify
that the equations \eqref{bianchirel} are indeed obeyed, provided that the scalar and vector sources in
table \ref{Rbasisex} and \ref{Ebasis} are not all independent but are constrained by the following relations
\begin{equation}\label{bianchisrc}
\begin{split}
&\frac{d}{dR}{\mathcal V}^{V}_M + {\mathcal V}^{V}_M + \frac{(D-3)}{\mathcal{K}} \nabla^N {\mathcal S}^T_{NM}=0 \\
&\frac{d}{dR}{\mathcal V}^{S_1}+ {\mathcal V}^{S_1} + \frac{(D-3)}{\mathcal{K}} \nabla^N {\mathcal S}^{V_2}_{N}=0 \\ &\frac{d}{dR}{\mathcal V}^{S_2} + {\mathcal V}^{S_2} + \left[ \frac{1}{2}{\mathcal S}^{S_1} + \left( {\mathcal S}^{S_2} +\frac{\mathcal{K}}{2(D-3)}e^{-R}~\nabla.u \right) + \frac{1}{2}\left({\mathcal S}^{S_4}-\frac{\mathcal{K}}{(D-3)}e^{-R}\nabla.u\right) \right] \\&+ \frac{(D-3)}{\mathcal{K}} \nabla^N {\mathcal S}^{V_1}_{N}=0
\end{split}
\end{equation}
Note that we have two relations between the four scalar sources and one relation between the
two vector sources in table \ref{Rbasisex}. Note that the relations also involve the as yet unknown
quantity $\nabla.u$. Later in this paper we will explicitly verify that the sources that appear in the first and second
order calculation obey the relations \eqref{bianchisrc}. However we would like to emphasize here
that these relations are necessarily obeyed at every order in perturbation theory.
\subsection{Choice of basis for the constraint and dynamical equations}
Because we have the linear relationship between constraint and dynamical equations we use the following basis for solving the scalar, vector and tensor fluctuations
\begin{equation}\label{basisRE}
\begin{split}
&\text{Tensor:}~~~~ R^{T}_{AB} \\
&\text{Vector:}~~~~ R^{V_2}_M,~~E^V_M \\
&\text{Scalar:}~~~~ R^{S_1},~~R^{S_2},~~E^{S_1},~~E^{S_2}
\end{split}
\end{equation}
From now on we write every expression in this basis. The expressions that we get from Bianchi identities i.e. equations \eqref{bianchirel},\eqref{biarel} can be converted to the basis \eqref{basisRE} as
\begin{equation}\label{BiBa}
\begin{split}
&\frac{d}{dR}E^{V}_M + E^{V}_M + \frac{(D-3)}{\mathcal{K}} \nabla^N R^T_{NM}=0 \\
&\frac{d}{dR}E^{S_1} + E^{S_1} + \frac{(D-3)}{\mathcal{K}} \nabla^N R^{V_2}_{N}=0 \\
&\frac{d}{dR}E^{S_2} + (1-\frac{1}{2}e^{-R})R^{S_1} + R^{S_2} + \frac{1}{1-e^{-R}} \frac{(D-3)}{\mathcal{K}}\nabla^M \left( E_M^V- R^{V_2}_{M} \right) =0
\end{split}
\end{equation}
The corresponding relationship between the sources is given by
\begin{equation}\label{BiBaSrc}
\begin{split}
&\frac{d}{dR}{\mathcal V}^{V}_M + {\mathcal V}^{V}_M + \frac{(D-3)}{\mathcal{K}} \nabla^N {\mathcal S}^T_{NM}=0 \\
&\frac{d}{dR}{\mathcal V}^{S_1}+{\mathcal V}^{S_1} + \frac{(D-3)}{\mathcal{K}} \nabla^N {\mathcal S}^{V_2}_{N}=0 \\ &\frac{d}{dR}{\mathcal V}^{S_2} + (1-\frac{1}{2}e^{-R}){\mathcal S}^{S_1} + {\mathcal S}^{S_2}+ \frac{1}{1-e^{-R}}\frac{(D-3)}{\mathcal{K}} \nabla^N \left({\mathcal V}^{V}_{N}-{\mathcal S}^{V_2}_{N}\right)=0
\end{split}
\end{equation}
\section{Perturbation theory at first order} \label{pfo}
In this section we will explicitly solve for the first order correction metric $h^{(1)}_{MN}$.
However we will perform our analysis in a manner that makes the generalization to higher
orders obvious.
\subsection{Listing first order source functions}
As we have explained in the previous section, the components of $R^1_{MN}$ are given in terms of
$h^{(1)}_{MN}$ by the expressions in Table \ref{Rbasisex} with particular values for the source
functions in that table. By explicit calculation at first order we find that these source functions
are given by the values listed in the table \ref{Rbsrc}.
\begin{center}
\begin{table}[h!]
\caption{Sources of $R_{MN}$ equations at 1st order}\label{Rbsrc}
\resizebox{\columnwidth}{!}{
\begin{tabular}{ |c| }
\hline
Scalar sector \\
\hline
${\mathcal{S}}^{S_1}(R) = 0 $\\
${\mathcal{S}}^{S_2}(R) = \frac{\mathcal{K}}{2(D-3)}e^{-R}u.K.u-\frac{e^{-R}(-1+R)}{2}\frac{u.\nabla\mathcal{K}}{(D-3)} - \frac{\mathcal{K}^2}{2(D-3)^2}e^{-R}(-3+2R)$\\
${\mathcal{S}}^{S_3}(R) = \frac{1}{2\mathcal{K}(D-3)}Re^{-R}\nabla^2\mathcal{K}-\frac{e^{-2R}(-2+2e^R+R)}{2}\frac{u.\nabla \mathcal{K}}{(D-3)} + \frac{\mathcal{K}^2}{2(D-3)^2}e^{-2 R} \left(3 e^R (R-1)-2 R+3\right) $ \\
${\mathcal{S}}^{S_4}(R) = e^{-R}(-1+R)\frac{u.\nabla\mathcal{K}}{(D-3)} + \frac{\mathcal{K}^2}{(D-3)^2}e^{-R}(-1+2R) $ \\
\hline
Vector sector \\
\hline
$ {\mathcal{S}}^{V_1}_{A}(R) = \frac{\mathcal{K}}{2(D-3)} e^{-R} \left( u^M K_{MN} - u^{M}\nabla_{M} u_{N} \right) {\cal P}^N_A $ \\
$ {\mathcal{S}}^{V_2}_{A}(R) = \frac{\mathcal{K}}{2(D-3)} e^{-2R} \left( u^M K_{MN} - u^{M}\nabla_{M} u_{N} \right) {\cal P}^N_A + \frac{e^{-R}}{2}\left( \frac{\nabla^2 u_A}{(D-3)}-\frac{\nabla_A \mathcal{K}}{(D-3)} \right) $\\
\hline
Tensor sector \\
\hline
$ {\mathcal{S}}^{T}_{AB}(R) = 0 $ \\
\hline
\end{tabular}}
\end{table}
\end{center}
\noindent
Moreover the constraint equations take the form listed in Table \ref{Ebasis} with first order source functions listed in Table \ref{Esrc}.
We list the corresponding sources to the constraint equations at 1st order in table \ref{Esrc}. We have verified that our explicit expressions for the sources obey the constraints \eqref{bianchisrc}.
We now proceed to solve the metric corrections at 1st order i.e. $h^{(1)}_{MN}$. We impose the conditions \eqref{velhor} as discussed in section \ref{sv}.
\begin{center}
\begin{table}[t]
\caption{Sources to constraint equations at 1st order}\label{Esrc}
\begin{tabular}{ |c| }
\hline
Vector constraint source\\
\hline
$ {\mathcal V}_M^{V}(R) = \frac{e^{-R}}{2}\left( \frac{\nabla^2 u_M}{(D-3)}-\frac{\nabla_M \mathcal{K}}{(D-3)} + \frac{\mathcal{K}}{(D-3)}(u^AK_{AM}-u.\nabla u_M) \right) $\\
\hline
Scalar constraint 1 source\\
\hline
$ {\mathcal V}^{S_1}(R) = \frac{1}{2\mathcal{K}(D-3)}Re^{-R}\nabla^2\mathcal{K} - \frac{-e^{-2R}+e^{-R}(1+R)}{2}\frac{u.\nabla\mathcal{K}}{(D-3)} $\\$+ \frac{\mathcal{K}}{2(D-3)}e^{-R}(1-e^{-R})u.K.u + Re^{-R}\frac{\mathcal{K}^2}{2(D-3)^2} $\\
\hline
Scalar constraint 2 source\\
\hline
$ {\mathcal V}^{S_2}(R) = \frac{e^{-R}}{2} \left( \frac{\mathcal{K}^2}{(D-3)^2}(1-2R) + \frac{u.\nabla \mathcal{K}}{(D-3)} (1-R) \right) $\\
\hline
\end{tabular}
\end{table}
\end{center}
\noindent
\subsection{Tensor sector}
In this sector we have a single equation for the single variable $H^{(T)}_{MN}$.
This equation is obtained by equating the last line of Table
\ref{Rbasisex} to zero and takes the form
\begin{equation} \label{teneq}
\begin{split}
R^{T}_{AB} = e^{-R} \frac{d}{dR}\left(\left(e^R-1\right) \frac{dH^{(T)}_{AB}}{dR}\right) \left( \frac{-\mathcal{K}^2}{2(D-3)^2} \right) + {\mathcal{S}}^{T}_{AB}(R)=0
\end{split}
\end{equation}
where ${\mathcal{S}}^{T}_{AB}(R)$ is the source for the tensor sector. At first
order it turns out that ${\mathcal{S}}^{T}_{AB}(R)=0$ (see Table \ref{Esrc}).
In order to facilitate generalizations to higher orders however,
in this subsection we will solve \eqref{teneq} for an arbitrary source
function, and substitute ${\mathcal{S}}^{T}_{AB}(R)=0$ only at the end of the
calculation.
Integrating \eqref{teneq} once we find
\begin{equation}\label{fi}
\begin{split}
\frac{d}{dR}( H^{(T)}_{AB})= \left( \frac{-2(D-3)^2}{\mathcal{K}^2} \right) \frac{-1}{e^R-1}\int_0^R e^{x}{\mathcal{S}}^{T}_{AB}(x) dx
\end{split}
\end{equation}
The condition that $H^{(T)}_{AB}$ (and so RHS
of \eqref{fi}) is regular at $R=0$ fixes
the lower limit of the integral in \eqref{fi}.
Integrating a second time we find
\begin{equation}\label{si}
\begin{split}
H^{(T)}_{AB} &=\left( \frac{-2(D-3)^2}{\mathcal{K}^2} \right) \int_R^{\infty}\frac{dy}{e^y-1}\int_0^y e^{x}{\mathcal{S}}^{T}_{AB}(x) dx \\
&= \left( \frac{2(D-3)^2}{\mathcal{K}^2} \right) \bigg[ \log(1-e^{-R})\int_0^R e^x S^{T}_{AB}(x)dx + \int_R^{\infty} \log(1-e^{-x})e^x S^{T}_{AB}(x) \bigg]
\end{split}
\end{equation}
where the upper limit in the outer integral in \eqref{si} is fixed by the requirement that
$H^{(T)}_{AB}$ decay at large $R$.
In summary, the tensor fluctuation $H^{(T)}_{AB}$ is given at any order, in
terms of the tensor source function ${\mathcal{S}}^{T}_{AB}(x)$ at that order, by the
expression \eqref{si}. Note that $H^{(T)}_{AB}$ is uniquely determined
by its source function; requirements of regularity at $R=0$
and decay at infinity unambiguously fix all integration constants in
\eqref{teneq}.
As we have mentioned above, at first order ${\mathcal{S}}^{T,1}_{AB}(R)=0$.
It follows from \eqref{si} that the first order tensor fluctuation $H^{(T)}_{AB}$ also vanishes and so
\begin{equation} \label{foten}
H^{(T,1)}_{AB}=0
\end{equation}
\subsection{Vector Sector}
\subsubsection{Constraint Equation and the Membrane Equation of Motion}
In the vector sector we have two equations for the single
variable $H^{(V)}_M$. The two equations may be chosen to be the vector
constraint equation $E^V_M$ (see the first line of Table \ref{Ebasis}) and
the equation $R_L^{V_2}=0$ (see Table \ref{Rbasisex}).
One cannot, of course, solve two equations for a single variable unless
one linear combination of the two equations is an identity. Indeed the first equation of
\eqref{BiBa}
\begin{equation}\label{vecident}
\frac{d}{dR}E^{V}_M + E^{V}_M + \frac{(D-3)}{\mathcal{K}} \nabla^N R^T_{NM}=0
\end{equation}
asserts that the vector constraint equation is automatically solved at
all values of $R$ if its solved at one value of $R$ (we use here that
we have already solved the tensor equation so that $R^T_{AB}=0$).
We will find it convenient to solve the vector constraint equation at
$R=0$. From Table \ref{Ebasis} we see that
$$E^V_M=\frac{1}{2}\frac{\mathcal{K}}{(D-3)}(1-e^{-R}) \frac{d}{dR} \left( \frac{\nabla^M H^{(T)}_{MN}}{(D-3)} \right) + {\mathcal V}_M^{V}(R)$$
At $R=0$
$$E^V_M= {\mathcal V}_M^{V}(0)$$
It follows that the constraint equation is solved at $R=0$ if and only if
$ {\mathcal V}_M^{V}(0)$ vanishes (here we use the fact that
$H^{(T)}_{MN}$ is regular at $R=0$; see the previous subsection) .
This requirement is a statement of the membrane equations of motion.
We would like to reemphasize that the membrane equations of motion at $n^{th}$ order
are obtained simply by evaluating the $n^{th}$ order vector constraint equation at
$R=0$. At $R=0$ this equation is independent of all the unknown $n^{th}$ order
fluctuation fields. As a consequence the membrane equations of motion may be obtained
at $n^{th}$ order {\it before} solving for the fluctuation fields at $n^{th}$ order,
as in studies of the fluid gravity correspondence.
The analysis presented in this subsection so far has been valid at
every order in perturbation theory. Specializing now to the first order, we
read off the value of ${\mathcal V}_M^{V}(0)$ from Table \ref{Esrc}. Equating
this expression to zero we find the first order membrane equation of motion
\begin{equation} \label{VE1}
\left(\frac{\nabla^2 u_{A}}{\mathcal{K}}- \frac{\nabla_A\mathcal{K}}{\mathcal{K}}+ u_{C}K^C_A-u.\nabla u_A \right)\mathcal{P}^A_B=0
\end{equation}
While all fields in \eqref{VE1} live in the full bulk spacetime $R^{D-1,1}$,
and all derivatives in that equation are bulk spacetime derivatives, the
equation \eqref{VE1} itself holds only on the membrane surface $\psi=1$.
Using the subsidiary conditions \eqref{auxeq} it is possible to rewrite
\eqref{VE1} as an equation restricted to the membrane. As demonstrated
in \cite{Bhattacharyya:2015fdk} the equation of motion of motion turns
out to take exactly the same form as \eqref{VE1} in this language. In other
words \eqref{VE1} also holds true if we think of $K_{MN}$ and $u_M$ as
membrane world volume fields, and regard every derivative in that equation
as a covariant derivative on the membrane world volume.
\subsubsection{Solving for the vector fluctuation}
As we have explained in the previous subsubsection, the
constraint vector equation is automatically solved at every
$R$ provided the membrane equation is obeyed. Assuming this is
the case, we have already solved one of the two vector equations.
In order to solve for the unknown function, $H^{(V)}_M$, in the vector sector,
we now turn to the second vector equation $R_L^{V_2}=0$. This equation
takes the form
\begin{equation} \label{dynveceq}
\left( \frac{-\mathcal{K}^2}{2(D-3)^2} \right) e^{-2R}(-1+ e^{R}) \frac{d}{dR}(e^{R} \frac{d}{dR} H^{(V)}_{M}) + {\mathcal{S}}^{V_2}_{M}(R) = 0
\end{equation}
As in the previous subsection we will proceed to solve \eqref{dynveceq}
for an arbitrary source function, plugging in the first order result for
the source
\begin{equation}\label{fos}
{\mathcal{S}}^{V_2,1}_{A}(R) = -\frac{\mathcal{K}}{2(D-3)} e^{-2R}(-1+ e^{R})
\left( u^M K_{MN} - u^{M}\nabla_{M} u_{N} \right) {\cal P}^N_A
\end{equation}
only at the end of the computation.
Notice that the LHS of \eqref{dynveceq} vanishes at $R=0$. It follows
that \eqref{dynveceq} admits regular solutions if and only if
${\mathcal S}^{V_2}_M(R)$ also vanishes at $R=0$. It would naively seem that
this requirement imposes a new constraint on membrane data, independent
of \eqref{VE1}. \footnote{Had this step of the programme imposed a new
constraint, we would have obtained a new membrane equation - and so obtained
more membrane equations than membrane variables, leading to an inconsistent
dynamical system.} However it turns out that the vanishing of
${\mathcal S}^{V_2}_M(R)$ is automatic; indeed it follows from \eqref{DCrel} that
$R^{V_2}_M$ is simply identical to the vector constraint equation
$E^{V}_M$ at $R=0$. It follows as a consequence that ${\mathcal S}^{V_2}_M(R)$
is proportional to the LHS of
\eqref{VE1} at $R=0$. \footnote{ To see this we note that \eqref{dynveceq}
reduces to ${\mathcal S}^{V_2}_M(R)$ at $R=0$ while
$E^{V}_M$ reduces to the LHS of \eqref{VE1} at $R=0$.}.
Using the fact that ${\mathcal{S}}^{V_2,1}_{M}(0)$ vanishes, we integrate
\eqref{dynveceq} once to find
\begin{equation}
\begin{split}
& e^{R} \frac{d}{dR} H^{(V)}_{M} = \left( \frac{-2(D-3)^2}{\mathcal{K}^2} \right)
\left[ \int_0^R \left( \frac{-e^{y}}{1-e^{-y}} \right){\mathcal{S}}^{V_2}_{M}(y)dy
+ C_M^{V_2} \right]
\end{split}
\end{equation}
where $C_M^{V_2}$ is an as yet undetermined integration constant. Integrating
a second time we find
\begin{equation}\label{vecsolf}
\begin{split}
& H^{(V)}_{M} = \left( \frac{2(D-3)^2}{\mathcal{K}^2} \right) \int_R^{\infty} e^{-x} \left[ \int_0^x \left( \frac{-e^{y}}{1-e^{-y}} \right){\mathcal{S}}^{V_2}_{M}(y)dy \right]dx
- C_M^{V_2} e^{-R}
\end{split}
\end{equation}
The upper limit on the the outer integral of \eqref{vecsolf} has been
determined from the requirement that $H^{(V)}_{M}$ vanishes at large $R$.
The expression for $H^{V}_M$ may be simplified by integrating by parts; we
find
\begin{equation}\label{vecsolsu}
H^{(V)}_{M}(R) = \left( \frac{2(D-3)^2}{\mathcal{K}^2} \right)
\left( e^{-R} \int_0^R \left( \frac{-e^{x}}{1-e^{-x}} \right){\mathcal{S}}^{V_2}_{M}(x)dx
- \int_R^\infty \frac{ {\mathcal{S}}^{V_2}_{M}(x)}{1-e^{-x}} \right)
- C_M^{V_2} e^{-R}
\end{equation}
In particular that
\begin{equation}\label{vecsolc}
H^{(V)}_{M}(0) = -\left( \frac{2(D-3)^2}{\mathcal{K}^2} \right)
\int_0^\infty \frac{ {\mathcal{S}}^{V_2}_{M}(x)}{1-e^{-x}}
- C_M^{V_2}
\end{equation}
It follows (see \eqref{velhor}) that
\begin{equation}\label{vecint}
C_M^{V_2} = -\left( \frac{2(D-3)^2}{\mathcal{K}^2} \right)
\int_0^\infty \frac{ {\mathcal{S}}^{V_2}_{M}(x)}{1-e^{-x}}
\end{equation}
so that
\begin{equation}\label{vecsols}
H^{(V)}_{M}(R) = \left( \frac{2(D-3)^2}{\mathcal{K}^2} \right)
\left( e^{-R} \int_0^R \left( \frac{-e^{x}}{1-e^{-x}} \right){\mathcal{S}}^{V_2}_{M}(x)dx
- \int_R^\infty \frac{ {\mathcal{S}}^{V_2}_{M}(x)}{1-e^{-x}}
+e^{-R} \int_0^\infty \frac{ {\mathcal{S}}^{V_2}_{M}(x)}{1-e^{-x}}
\right)
\end{equation}
The expression \eqref{vecsols} is our final expression for $H^{(V)}_{M}(R)$
at any order in perturbation theory in terms of the source function at that
order. Note that $H^{(V)}_{M}(R)$ is uniquely determined in terms of its
source function; the integration constants in \eqref{dynveceq} are
uniquely determined by the requirement that $H^{(V)}_{M}(R)$ vanish at infinity
and that \eqref{velhor} is obeyed at $R=0$.
Plugging the first order expression for the source \eqref{fos} into
\eqref{vecsols}, at first order we find
\begin{equation}
H^{(V,1)}_{M} = \frac{(D-3)}{\mathcal{K}}Re^{-R}\left( u^{A} K_{AN}-u^{A}\nabla_A u_N \right) P^N_M
\end{equation}
\subsection{Scalar sector}
In the scalar sector we have four equations for the two variables $H^{(Tr)}$ and
$H^{(S)}$. As a basis for the four equations we find it convenient to use the
two scalar constraint equations $E^{S_1}$ and $E^{S_2}$ (see Table \ref{Ebasis})
together with the two additional equations $R^{S_1}=0$ and $R^{S_2}=0$
(see Table \ref{Rbasis}).
\subsubsection{Constraint Equations and $\nabla.u$}
As in the previous subsection it is consistent to have four equations
for two variables only if two of the four equations are identities. The last two
equations in \eqref{BiBa}
\begin{equation}\label{scaconst}
\begin{split}
&\frac{d}{dR}E^{S_1} + E^{S_1} + \frac{(D-3)}{\mathcal{K}} \nabla^N R^{V_2}_{N}=0 \\
&\frac{d}{dR}E^{S_2} + (1-\frac{1}{2}e^{-R})R^{S_1} + R^{S_2} + \frac{(D-3)}{\mathcal{K}}\frac{1}{1-e^{-R}} \nabla^M \left( E_M^V- R^{V_2}_{M} \right) =0
\end{split}
\end{equation}
assert that this is indeed the case. As we have already solved the vector sector at $n^{th}$ order
$R^{V_2}_{N}$ vanishes. It follows that the first equation in \eqref{scaconst} asserts that if
$E^{S_1}$ is solved at any $R$ it is automatically solved at every $R$.
When evaluated at $R=0$ this equation reduces to the condition
\begin{equation}\label{condo}
{\cal V}^{S_1}(0)+ \frac{\mathcal{K}}{2(D-3)}~\nabla.u =0
\end{equation}
Recall that at leading order $\nabla.u=0$. \eqref{condo} determines the
correction to this statement at subleading orders.
As in the previous subsection we emphasize that the expression for $\nabla.u$ at
$n^{th}$ order is determined simply by evaluating the $n^{th}$ order
constraint equation $E^{S_1}$ at $R=0$. In order to obtain this correction we
do not need to solve for any of the $n^{th}$ order fluctuation fields,
all of which drop out in $E^{S_1}$ evaluated at $R=0$.
The analysis of this subsection has, so far, been valid at every order in perturbation theory.
Specializing to first order it is easily verified from Table \ref{Esrc} that ${\mathcal V}^{S_1}(0)=0$.
It follows that the zero order relation $\nabla.u=0$ is uncorrected at
first order~~(since $(\nabla . u)_{0}=\mathcal{V}^{S_1}(0)=0 $). As we will see in the next section, the situation is
different at second order.
The constraint equation $E^{S_2}$ plays a distinct logical role from $E^{S_1}$ in our perturbative
programme. Once the tensor and vector equations had been solved, \eqref{scaconst} assured us
that $E^{S_1}(R)$ obeys a homogeneous differential equation in $R$ (see
\eqref{bianchirel} which makes no reference
to any of the other equations in the scalar sector. On the other hand the differential
equation obeyed by $E^{S_2}$ involves the other scalar equations (see the last equation in
\eqref{biarel}). The most useful way to view the last equation in \eqref{biarel} is as
follows. It might, a priori, have seemed that we have 4 equations in the scalar sector.
We have already dealt with $E^{S_1}$ above leaving behind a three dimensional space of
equations. A useful basis for this space is given by $E^{S_2}$, $R^{S_1}$ and $R^{S_2}$.
The last equation in \eqref{biarel} allows us to eliminate $R^{S_2}$ from this basis.
In order to complete solving in the scalar sector we need only solve the equations
$E^{S_2}$, $R^{S_1}$. In other words the constraint
equation $E^{S_2}$ does not constrain data: instead it may be used to solve for the
scalar fluctuation. We turn to this task in the next subsubsection.
\subsubsection{Solving for the scalar fluctuations}
The equation $R^{S_1}$
\begin{equation}
R^{S_1}=\left( \frac{-\mathcal{K}^2}{2(D-3)^2} \right) \frac{d^2 H^{(Tr)}}{dR^2}+{\mathcal{S}}^{S_1}(R)=0
\end{equation}
is easily solved. Integrating the above equation once we get
\begin{equation}
\frac{d H^{(Tr)}}{dR} = \left(\frac{-2(D-3)^2}{\mathcal{K}^2}\right) \int_R^{\infty} dx~{\mathcal{S}}^{S_1}(x)
\end{equation}
Where we have fixed the boundary condition from the requirement that $H^{(Tr)}$ and so its
derivative $\frac{d H^{(Tr)}}{dR}=0$ vanish at large $R$.
Integrating this equation once again we have
\begin{equation}\label{htreq}
\begin{split}
H^{(Tr)} &= \left(\frac{2(D-3)^2}{\mathcal{K}^2}\right) \int_R^{\infty} dy \int_y^{\infty} dx~ {\mathcal{S}}^{S_1}(x) \\
&= \left(\frac{2(D-3)^2}{\mathcal{K}^2}\right) \left[ -R \int_R^{\infty} dx~ {\mathcal{S}}^{S_1}(x) + \int_R^{\infty}dx~ x~{\mathcal{S}}^{S_1}(x) \right]
\end{split}
\end{equation}
where, once again we have fixed the integration constant from the requirement that
$H^{(Tr)}=0$ at large $R$.
Specializing now to first order we note ${\mathcal{S}}^{S_1,1}=0$ so that
\begin{equation}
H^{(Tr,1)}=0
\end{equation}
The equation $E^{S_2}$ takes the form
\begin{equation}\label{sctt}
\begin{split}
&\frac{d}{dR} (H^{(S)}e^{R}) =\frac{2(D-3)^2}{\mathcal{K}^2}e^{R}\mathcal{S}_{S}(R) ~~~~\text{where,} \\
&\mathcal{S}_{S}(R)= -\frac{\mathcal K}{2 (D-3)} \frac{d}{dR}\left( \nabla^M H^{(V)}_M \right)-\frac{\mathcal{K}}{(D-3)} \nabla^M H^{(V)}_M \\
&+\frac{\mathcal{K}^2}{4(D-3)^2}(2-e^{-R})\frac{d}{dR}H^{(Tr)} - \frac{1}{2} \nabla^M\nabla^N H^{(T)}_{MN} + {\mathcal V}^{S_2}(R) + \frac{\mathcal{K}}{2(D-3)}e^{-R}~\nabla.u
\end{split}
\end{equation}
Plugging in the already obtained
expressions of $H^{(V)}_M ~~,H^{(T)}_{MN}~, ~H^{(Tr)} $ (see \eqref{vecsols},\eqref{htreq} and \eqref{si}) and using \eqref{BiBaSrc}, the source function
${\mathcal S}_S(R)$ can be rewritten as a linear functional of the elementary sources
${\mathcal S}^{S_1}$, ${\mathcal S}^{S_2}$ and ${\mathcal V}^{S_1}$
\footnote{It turns out that all dependence on the fourth independent scalar
source, ${\mathcal V}^{S_2}$ cancels.}. Upon simplifying (by integrating
by parts on several occasions) we find
\begin{equation}\label{compss}
\begin{split}
\mathcal{S}_{S}(R)&= \int_{R}^{\infty}{\mathcal{S}}^{S_2}(x)dx +\frac{1}{2}\int_{R}^{\infty}(2-e^{-x}){\mathcal{S}}^{S_1}(x)dx -\frac{1}{2}(2-e^{-R})\int_{R}^{\infty}{\mathcal{S}}^{S_1}(x)dx\\&-\left(1-e^{-R}\right)\int_{R}^{\infty}\left(\frac{e^{x}\left({{\mathcal V}^{S_1}}^{'}(x)+{\mathcal V}^{S_1}(x)\right)}{(e^{x}-1)}dx\right)dy-\mathcal{V}^{S_{1}}(R)+e^{-R}\mathcal{V}^{S_1}(0)\\&+\log (1-e^{-R})\left({{\mathcal V}^{S_1}}^{'}(0)+{\mathcal V}^{S_1}(0)\right) + (\nabla \cdot u)\frac{\mathcal{K} e^{-R}}{2(D-3)}
\end{split}
\end{equation}
We note that ${\mathcal S}_S$ is analytic at $R=0$ if and only if
\begin{equation} \label{sccond}
{{\mathcal V}^{S_1}}^{'}(0)+{\mathcal V}^{S_1}(0)=0
\end{equation}
This condition is, in fact, automatic. It follows from the second of
\eqref{BiBaSrc} that the LHS of \eqref{sccond} is proportional to
$\nabla^N{\mathcal S}_N^{V_2}(0)$. We have already argued, however, that
${\mathcal S}_N^{V_2}$ vanishes at $R=0$. Since this condition holds at
every point on the membrane, it follows also that $\nabla^N{\mathcal S}_N^{V_2}(0)
=0$
establishing \eqref{sccond}.
\footnote{In studies of the fluid gravity correspondence a derivative
of the equation of the $n^{th}$ order equation contributes to sources
only at $(n+1)^{th}$ order in the derivative expansion. In the large
$D$ expansion of this paper, however, the suppression in order resulting
from using an extra derivative can be compensated for by an enhancement
in order resulting from the contraction of a spacetime index. Consequently
the equation of motion and its contracted derivatives are of the same
order in the large $D$ expansion.}
Plugging \eqref{compss} into \eqref{sctt}, integrating (and simplifying
using integration by parts) we find
\begin{equation}\label{hsans}
\begin{split}
H_{S}(R)&=\frac{2(D-3)^2}{\mathcal{K}^2}e^{-R}\Bigg(\frac{(\mathcal{K}(\nabla \cdot u))R}{2(D-3)} + e^{R}\int_{R}^{\infty}{\mathcal{S}}^{S_2}(x)dx -\int_{0}^{\infty} {\mathcal{S}}^{S_2}(x)dx+\int_{0}^{R}e^{x}{\mathcal{S}}^{S_2}(x)dx \\&+ \frac{e^R}{2}\int_{R}^{\infty}(2-e^{-x}){\mathcal{S}}^{S_1}(x)dx+\frac{1}{2}\int_{0}^{R}e^x(2-e^{-x}){\mathcal{S}}^{S_1}(x)dx-\frac{1}{2}\int_{0}^{\infty}(2-e^{-x}){\mathcal{S}}^{S_1}(x)dx\\&-\frac{1}{2}(2 e^R-R)\int_{R}^{\infty}{\mathcal{S}}^{S_1}(x)dx+\int_{0}^{\infty}{\mathcal{S}}^{S_1}(x)dx-\frac{1}{2}\int_{0}^{R}(2e^y-y){\mathcal{S}}^{S_1}(x)dx\\&-\int_{0}^{R}\left(e^{y}-1\right)\int_{y}^{\infty}\left(\frac{e^{x}\left({{\mathcal V}^{S_1}}^{'}(x)+{\mathcal V}^{S_1}(x)\right)}{(e^{x}-1)}dx\right)dy-\int_{0}^{R}e^x{\mathcal V}^{S_1}(x)dx+R \mathcal{V}^{S_1}(0)\Bigg)
\end{split}
\end{equation}
Explicitly at first order
\begin{equation}\label{hsfirst}
H^{(S,1)} = \frac{D-3}{\mathcal{K}}R e^{-R}\left( R \left( -\frac{\mathcal{K}}{D-3}-\frac{u\cdot\nabla \mathcal{K}}{\mathcal{K}}+\frac{u\cdot K \cdot u}{2} \right) +\left( \frac{\mathcal{K}}{D-3}+u\cdot K \cdot u \right) \right)
\end{equation}
\subsection{Final Result for the first order metric}
After integrating the ordinary differential equations corresponding to Einstein's equations and imposing the condition that the metric is regular at the horizon, matches flat space at the end of the membrane region and \eqref{velhor}, we get the following solutions for the various components of the metric correction.
\begin{equation}
\begin{split}
& H^{(T,1)}_{MN} = 0 \\
& H^{(Tr,1)} = 0 \\
& H^{(V,1)}_{M} = \frac{(D-3)}{\mathcal{K}}Re^{-R}\left( u^{A} K_{AL}-u^{A}\nabla_A u_L \right){\cal P}^L_M\\
& H^{(S,1)} = \frac{D-3}{\mathcal{K}}R e^{-R}\left( R \left( -\frac{\mathcal{K}}{D-3}-\frac{u\cdot\nabla \mathcal{K}}{\mathcal{K}}+\frac{u\cdot K \cdot u}{2} \right) +\left( \frac{\mathcal{K}}{D-3}+u\cdot K \cdot u \right) \right)
\end{split}
\end{equation}
Thus we can write the 1st order corrected metric as
\begin{equation}\label{met1}
\begin{split}
&g_{MN} = \eta_{MN} +\frac{O_M O_N}{\psi^{D-3}}\\ &+ \frac{1}{D-3}\bigg[ \frac{D-3}{\mathcal{K}}R e^{-R}\left( R \left( -\frac{\mathcal{K}}{D-3}-\frac{u\cdot\nabla \mathcal{K}}{\mathcal{K}}+\frac{u\cdot K \cdot u}{2} \right) +\left( \frac{\mathcal{K}}{D-3}+u\cdot K \cdot u \right) \right)O_MO_N\\ &+ \frac{(D-3)}{\mathcal{K}}Re^{-R} \left( u^{A} K_{AL}-u^{A}\nabla_A u_L \right) P^L_{(M} O_{N)} \bigg]
\end{split}
\end{equation}
\section{2nd order solution}
The metric \eqref{met1} solves Einstein equation to first subleading order. In this section we implement the perturbative procedure
to one higher order. In other words we determine the correction $H^{(2)}_{MN}$ in a way that
ensures that $R_{AB}$ evaluated on the corrected metric is of order $1/D$ (more precisely that
$R_{AB}R^{AB}$ is of order $1/D^2$).
The procedure we follow is exactly that of the previous section: in fact second order corrections to the
metric are given directly by the formulae of the previous subsection with one modification: we need to use the second order rather than first order source functions. In other words the computation at second order
boils down entirely to determining the second order sources.
In order to determine the sources at second order we plug the first order corrected metric \eqref{met1}
together with an as yet undetermined second order correction $h^2_{MN}$
into Einstein's equations. We use the fact that the shape and velocity functions in the first order
corrected metric obey the equation of motion
\begin{equation}\label{meom}
\left(\frac{\nabla^2 u_{A}}{\mathcal{K}}- \frac{\nabla_A\mathcal{K}}{\mathcal{K}}+ u_{C}K^C_A-u.\nabla u_A \right)\mathcal{P}^A_B + \frac{1}{D}\mathcal{E}_A {\cal P}^A_B =0
\end{equation}
where $\mathcal{E}_B$ is an as yet undetermined `2nd order' correction to the equations of motion.
As in the previous subsection we solve the equations in the neighbourhood of a particular point on the
event horizon. In our analysis, however, we use the fact that the membrane equations of motion
\eqref{meom} are obeyed not just at the particular point we are expanding about but everywhere on the
membrane. In other words we use the fact that the derivative of \eqref{meom} vanishes at the point of
interest. Finally we also use the fact that $\nabla.u$ is an as yet undetermined quantity of order
$1/D$.
We find by explicit computation that the curvature components listed table \ref{Rbasis} do indeed take the
form listed in table \ref{Rbasisex},\ref{Ebasis} once all metric fluctuation fields in that table are identified with second order
fluctuations. Our explicit computations also yield explicit expressions for all the second order source functions. We present an explicit listing of these source functions in Tables \ref{Rbsrc2} and \ref{Esrc2}
in the Appendix.
In the rest of this section we obtain the second order correction to the metric by inserting the
second order sources listed above into the general integral formulae of the previous section and
performing all integrals.
\subsection{Constraints on membrane data}
\subsubsection{Correction to the membrane equations from the vector
sector}
As in the previous subsection \eqref{vecident} guarantees that the
vector constraint equation $E^V_M=0$
is solved at any $R$ if the equation is obeyed at $R=0$. As in the previous subsection the constraint equation at $R=0$ is independent of the second order fluctuation fields. From table \ref{Esrc2} we see that this constraint
equation at $R=0$ determines $-\frac{1}{D}\mathcal{E}_A {\cal P}^A_B$ - the second order correction
to the membrane equation of motion - in terms of appropriate expressions involving the
membrane extrinsic curvature and velocity fields. Adding these correction terms to the
first order membrane equation \eqref{VE1copy} we recover the second order corrected membrane
equation
\begin{eqnarray}\label{veq1geom}
\nonumber&&\Bigg[\frac{\nabla^{2}u}{\mathcal{K}}-\frac{\nabla \mathcal{K}}{\mathcal{K}}+u\cdot K-(u\cdot\nabla )u \bigg]\cdot {\cal P} +\bigg[\frac{\nabla^2\nabla^2 u}{\mathcal{K}^3}-\frac{\nabla (\nabla^2 \mathcal{K})}{\mathcal{K}^3} \\\nonumber&+&3\frac{(u\cdot K\cdot u)(u\cdot \nabla u)}{\mathcal{K}}-3\frac{(u\cdot K\cdot u)(u\cdot \nabla n)}{\mathcal{K}}-6\frac{(u \cdot (\nabla^2 n))(u\cdot \nabla u)}{\mathcal{K}^2}\\&+&6\frac{(u \cdot (\nabla^2 n))(u\cdot \nabla n)}{\mathcal{K}^2}+\frac{3}{D-3} u \cdot \nabla u-\frac{3}{D-3}u \cdot \nabla n\Bigg]\cdot {\cal P}=0
\end{eqnarray}
where
\begin{equation}
{\cal P}^{AB}= \eta^{AB}- n^A n^B +u^A u^B
\end{equation}
The 1st square bracket in \eqref{veq1geom} is simply the 1st order equation of motion while the 2nd square bracket represents subleading corrections. \footnote{Note that we can write the equation \eqref{veq1geom} in a nicer looking form by using the subsidiary conditions \eqref{auxeq}, divergence of first order membrane equation of motion \eqref{VE1copy} and divergence of velocity condition \eqref{Divu}. The form is
\begin{equation}
\left( \frac{\nabla^2O}{\nabla.O}+O.\nabla O \right)\cdot {\cal P} + \left( \frac{\nabla^2\nabla^2 O}{(\nabla.O)^3} + 3\frac{\nabla^2(\nabla.O)}{(\nabla.O)^3}O.\nabla O \right)\cdot {\cal P} = 0
\end{equation}
}
We would like, however, to emphasize an important technical
point. All the fields in \eqref{veq1geom} are assumed to live in all of the embedding flat spacetime;
they are extended off the surface of the membrane by the subsidiary conditions listed earlier in this paper.
While all covariant derivatives listed in \eqref{veq1geom} are evaluated on the surface of the membrane,
they act on fields defined in all of spacetime.
As the membrane equations of motion are intrinsic to the membrane, it is clearly unnatural
to write them in terms of spacetime derivatives of an essentially arbitrary extension of
membrane fields into the embedding spacetime. The equation of motion \eqref{veq1geom} can be
rewritten so that all fields in that equation are purely membrane world volume fields, and every
derivative in the equation is a covariant derivative on the membrane world volume. We now
explain how this is done.
The relationship between the bulk covariant derivatives
of tensors (e.g. $u_M$) and membrane worldvolume derivatives of the same quantities is quite
straightforward when no more than one derivative acts on the same object. The spacetime covariant
derivative is obtained from the corresponding bulk quantity by projecting every index (not just the
derivative indices) onto the membrane world volume. However this relationship is more complicated
when we have two or more derivatives acting on the same object; the reason for the additional complication
is that the formula for multiple worldvolume covariant derivatives involves inserting projectors at each
step (when you define the first derivative in terms of bulk derivatives, then again when you define
the second derivative in terms of bulk derivatives etc); when such expressions are opened out, outer
derivatives act on projectors used to define the inner derivatives. Tracing through
the required algebra we find that the corrected second order membrane equation of motion,
written in terms of fields and covariant derivatives that live purely on the membrane
world volume, takes the form
\begin{eqnarray}\label{veq1geomwv}
\nonumber&&\Bigg[\frac{\nabla^{2}u_{A}}{\mathcal{K}}-\frac{\nabla_{A} \mathcal{K}}{\mathcal{K}}+u^{B} K_{BA}-u\cdot\nabla u_{A}\Bigg]{\cal P}^{A}_{C} \\\nonumber&+& \Bigg[\left(-\frac{u^{C}K_{CB}K^{B}_{A}}{\mathcal{K}}\right)+\left(\frac{\nabla^2\nabla^2 u_{A}}{\mathcal{K}^3}-\frac{u \cdot \nabla \mathcal{K}\nabla_{A}\mathcal{K}}{\mathcal{K}^3}-\frac{\nabla^{B}\mathcal{K}\nabla_{B}u_{A}}{\mathcal{K}^2}-2\frac{K^{CD}\nabla_{C}\nabla_{D}u_{A}}{\mathcal{K}^2}\right) \\ \nonumber &+&\left(-\frac{\nabla_{A}\nabla^2 \mathcal{K}}{\mathcal{K}^3} +\frac{\nabla_{A}\left(K_{BC}K^{BC}\mathcal{K}\right)}{\mathcal{K}^3}\right) + 3\frac{(u\cdot K\cdot u)(u\cdot \nabla u_{A})}{\mathcal{K}}-3\frac{(u\cdot K\cdot u)(u^{B} K_{BA})}{\mathcal{K}}\\ &-&6\frac{(u \cdot \nabla \mathcal{K})(u\cdot \nabla u_{A})}{\mathcal{K}^2}+6\frac{(u \cdot \nabla \mathcal{K})(u^{B}K_{BA})}{\mathcal{K}^2} + \frac{3}{D-3} u \cdot \nabla u_{A}-\frac{3}{D-3} u^{B}K_{BA} \Bigg]{\cal P}^{A}_{C}=0\nonumber\\
\end{eqnarray}
The projector ${\cal P}^{AB}$ used in this equation
\begin{equation}
{\cal P}^{AB}= \tilde{g}^{AB} + u^A u^B
\end{equation}
where $\tilde{g}^{AB}$ is the induced metric on the world volume of the membrane.
The equation \eqref{veq1geomwv} can be slightly simplified as follows.
Let us first note that \eqref{veq1geomwv} takes the schematic form
\begin{equation}\label{seomo}
F^A + \frac{S^A}{ \mathcal{K}}=0
\end{equation}
where $F^A$ is the first order contribution to the equation of motion (the first line
of \eqref{veq1geomwv}) while $\frac{S^A}{\mathcal{K}}$ is the second order contribution (the second-fourth lines of \eqref{veq1geomwv}). $F^A$ and $S^A$ are each vector fields of order unity.
Let us now consider the modified equation of motion
\begin{equation}\label{seomt}
F^A + \frac{S^A}{ \mathcal{K}} + \nabla. F \frac{\zeta^A}{\mathcal{K}^2}=0
\end{equation}
where $\zeta^A$ is any vector field of order unity. As $\nabla.F$ is naively
of order $D$, the difference between the equations \eqref{seomt} and \eqref{seomo}
is naively of order $\frac{1}{D}$ suggesting that \eqref{seomo} and \eqref{seomt}
differ at first subleading order. This is not the case. By taking a divergence of
either \eqref{seomo} or \eqref{seomt}, the reader can easily convince herself that,
onshell, $\nabla.F$ is of order unity (rather than the naive estimate of order $D$).
If follows that \eqref{seomt} and \eqref{seomo} actually differ only at second
subleading order ( $\frac{1}{D^2}$ ) and are equivalent at first subleading order.
We are thus allowed to simplify \eqref{veq1geomwv} by adding any expression of the form
$\nabla. F \frac{\zeta^A}{\mathcal{K}^2}$ to it.
Now it was demonstrated in \cite{Bhattacharyya:2015fdk} that
\begin{equation}\label{SE1}
\frac{\nabla.F}{\mathcal{K}}=\frac{\nabla^2\mathcal{K}}{\mathcal{K}^2} - 2~\frac{u.\nabla\mathcal{K}}{\mathcal{K}} + u.K.u
\end{equation}
Using this relation and making the the choice
\begin{equation}\label{zcho}
\zeta^A= -3 \left( (u.\nabla u)_A - u_BK^B_A \right)
\end{equation}
we find that \eqref{veq1geomwv} is equivalent to \eqref{seomt} whose explicit form
is
\begin{eqnarray}\label{VE2alt}
\nonumber&&\Bigg[\frac{\nabla^{2}u_{A}}{\mathcal{K}}-\frac{\nabla_{A} \mathcal{K}}{\mathcal{K}}+u^{B} K_{BA}-u\cdot\nabla u_{A}\Bigg]{\cal P}^{A}_{C} \\\nonumber&+& \Bigg[\left(-\frac{u^{C}K_{CB}K^{B}_{A}}{\mathcal{K}}\right)+\left(\frac{\nabla^2\nabla^2 u_{A}}{\mathcal{K}^3}-\frac{u \cdot \nabla \mathcal{K}\nabla_{A}\mathcal{K}}{\mathcal{K}^3}-\frac{\nabla^{B}\mathcal{K}\nabla_{B}u_{A}}{\mathcal{K}^2}-2\frac{K^{CD}\nabla_{C}\nabla_{D}u_{A}}{\mathcal{K}^2}\right) \\ \nonumber &+&\left(-\frac{\nabla_{A}\nabla^2 \mathcal{K}}{\mathcal{K}^3} +\frac{\nabla_{A}\left(K_{BC}K^{BC}\mathcal{K}\right)}{\mathcal{K}^3}\right) - 3\frac{\nabla^2\mathcal{K}~u\cdot \nabla u_{A}}{\mathcal{K}^3}+3\frac{\nabla^2\mathcal{K}~u^{B} K_{BA}}{\mathcal{K}^3}\\ &+& \frac{3}{D-3} u \cdot \nabla u_{A}-\frac{3}{D-3} u^{B}K_{BA} \Bigg]{\cal P}^{A}_{C}=0\nonumber\\
\end{eqnarray}
\subsubsection{Divergence of velocity from a scalar constraint}
As we have explained in the previous section, the Einstein constraint equation
$E^{S_1}$ is satisfied at all $R$ if it is satisfied at $R=0$. As explained
in the previous subsection, the equation at $R=0$ simply asserts that
$$\nabla.u_{2}=- \frac{2(D-3)}{{\mathcal K}}{\mathcal V}^{S_1}(0)$$
Reading off the value of ${\mathcal V}^{S_1}(0)$ from the table \ref{Esrc2} we find
\begin{equation}
\begin{split}
\nabla\cdot u=\frac{(\nabla.u)_2}{D-3}=\frac{1}{2\mathcal{K}}\left(\nabla_{(A}u_{B)}\nabla_{(C}u_{D)}{\cal P}^{BC}{\cal P}^{AD}\right)
\end{split}
\end{equation}
\subsection{Second order corrections to the metric}\label{2ndmet}
\subsubsection{Tensor Sector}
The metric correction in the tensor sector is given by \eqref{si}
\begin{equation}
\begin{split}
H^{(T)}_{AB} &=\left( \frac{-2(D-3)^2}{\mathcal{K}^2} \right) \int_R^{\infty}\frac{dy}{e^y-1}\int_0^y e^{x}{\mathcal{S}}^{T}_{AB}(x) dx \\
&= \left( \frac{2(D-3)^2}{\mathcal{K}^2} \right) \bigg[ \log(1-e^{-R})\int_0^R e^x S^{T}_{AB}(x)dx + \int_R^{\infty} \log(1-e^{-x})e^x S^{T}_{AB}(x) \bigg]
\end{split}
\end{equation}
where ${\mathcal{S}}^{T}_{AB}$ is the second order source listed in table \ref{Rbsrc2}.
All the integrals that appear in the final answer can easily be performed
analytically, but the final results (given in terms of polylogs)
are not very illuminating; we prefer to leave our final result
in terms of an explicit integral.
\subsubsection{Vector Sector}
The solution for $H^{(V)}_M(R)$ at second order is given by \eqref{vecsols}
\begin{equation} \label{vfin}
H^{(V)}_{M}(R) = \left( \frac{2(D-3)^2}{\mathcal{K}^2} \right)
\left( e^{-R} \int_0^R \left( \frac{-e^{x}}{1-e^{-x}} \right){\mathcal{S}}^{V2}_{M}(x)dx
- \int_R^\infty \frac{ {\mathcal{S}}^{V2}_{M}(x)}{1-e^{-x}}
+ e^{-R} \int_0^\infty \frac{ {\mathcal{S}}^{V2}_{M}(x)}{1-e^{-x}}
\right)
\end{equation}
with all sources read off at 2nd order from table \ref{Rbsrc2}.
As in the tensor sector, all integrals that appear in \eqref{vfin}
can be explicitly performed in terms of polylogs, but we find the
expression \eqref{vfin} in terms of explicit integrals more illuminating.
\subsubsection{Scalar Sector}
Equation $R^{S_1}$ is decoupled equation for $H^{(Tr)}$. The integrated form is given by \eqref{htreq} which we write again
\begin{equation}
\begin{split}
H^{(Tr)} &= \left(\frac{2(D-3)^2}{\mathcal{K}^2}\right) \int_R^{\infty} dy \int_y^{\infty} dx~ {\mathcal{S}}^{S_1}(x) \\
&= \left(\frac{2(D-3)^2}{\mathcal{K}^2}\right) \left[ -R \int_R^{\infty} dx~ {\mathcal{S}}^{S_1}(x) + \int_R^{\infty}dx~ x~{\mathcal{S}}^{S_1}(x) \right]
\end{split}
\end{equation}
The source ${\mathcal{S}}^{S_1}$ for 2nd order is given in table \ref{Rbsrc2}. Substituting this we get the final form of the metric correction
\begin{equation}\label{tracesolution}
H^{(Tr,2)}=-\left(\frac{2(D-3)^2}{\mathcal{K}^2}\right)e^{-R}(1+R) \left( \left( u \cdot K-u\cdot \nabla u \right)\cdot {\cal P} \cdot \left( u\cdot K-u\cdot \nabla u \right) \right)
\end{equation}
In a similar manner the fluctuation ${H}^{S}$ can is given by
\eqref{hsans} upon plugging in the explicit values of the second order sources
from Tables \ref{Rbsrc2},\ref{Esrc2}.
\section{The spectrum of small fluctuations around a spherical membrane}
The simplest solution of the second order membrane equations of motion is a static spherical membrane
dual to a Schwarzschild Black hole.
In this section we compute the spectrum of small fluctuations about this solution. Our answers agree perfectly with earlier results for the spectrum of light quasinormal modes obtained by direct gravitational
analysis, in \cite{Emparan:2014aba}. We regard this detailed agreement as a nontrivial consistency check of the
second order membrane equations of motion derived in this paper.
The computation presented in this section is a straightforward extension of the first order
computation presented in section 5 of \cite{Bhattacharyya:2015fdk}. We have kept the discussion
of this section brief. We refer the reader to section 5 of \cite{Bhattacharyya:2015fdk} for a fuller discussion
of the logic behind our computation.
We work in standard spherical polar coordinates (see Eq 5.1 of \cite{Bhattacharyya:2015fdk}). The static spherical
membrane is given by
\begin{equation} \label{Schwcon}
r=1, ~~~u=-dt,
\end{equation}
We study the small fluctuations
\begin{equation}\label{SchwPert}
\begin{split}
r &= 1 + \epsilon~\delta r(t,\theta),\\
u &= -dt + \epsilon~\delta u_{\mu}(t,\theta)dx^{\mu}.
\end{split}
\end{equation}
about this solution and work to linear order in $\epsilon$. As explained in \cite{Bhattacharyya:2015fdk}, to linear order
the metric on membrane worldvolume is given by
\begin{equation}\label{SchwLin}
ds^2= -dt^2 + \left(1+2 \epsilon \delta r \right)d\Omega_{D-2}^2~~.
\end{equation}
As in \cite{Bhattacharyya:2015fdk} we find it convenient to work with covariant derivatives with respect to the unperturbed
spherical metric
\begin{equation}\label{IndFlat}
ds^2=-dt^2+ d\Omega_{D-2}^2~~,
\end{equation}
The derivatives appearing from now on are all with respect to metric \eqref{IndFlat}.
We use the following notation for the laplacian with respect to this fixed metric
$$\overline{\nabla}^2=\nabla_\mu\nabla^\mu=-\partial_t^2+\nabla_a\nabla^a=-\partial_t^2+\nabla^2$$
\subsection{The divergence condition}
The RHS of \eqref{Divu} is quadratic in $\epsilon$, and so vanishes upon linearizing in $\epsilon$.
At linear order, therefore, \eqref{Divu} reduces to $\nabla.u=0$ (where the divergence is taken along the dynamical
membrane world volume). As explained in \cite{Bhattacharyya:2015fdk}, this equation can be rewritten as
\begin{equation}\label{velco1}
\nabla_\mu \delta u^\mu = - (D-2) \partial_t \delta r,
\end{equation}
where, the covariant derivatives \eqref{velco1} are now taken w.r.t. the fixed metric \eqref{IndFlat}.
$u^0$ deviates from unity only at quadratic order in $\epsilon$. For the linearized considerations of this
section, therefore, the LHS of \eqref{velco1} is simply the spatial divergence of the velocity
\begin{equation}\label{velco2}
\nabla_a \delta u^a = - (D-2) \partial_t \delta r.
\end{equation}
As in \cite{Bhattacharyya:2015fdk}, \eqref{velco2} may be solved by separating $u$ into its
gradient and curl parts, i.e. by setting
\begin{equation}\label{veldec1}
\delta u_a = \nabla_a \Phi + \delta v_a,
\end{equation}
with
\begin{equation}\label{veldec2}
\nabla\cdot \delta v=0.
\end{equation}
It follows from \eqref{velco2} that
\begin{equation} \label{phicon}
\nabla^2 \Phi= -(D-2) \partial_t \delta r.
\end{equation}
\subsection{Linearized equation of motion}
In order to obtain the linearized membrane equations of motion we use Eq 5.7 of
\cite{Bhattacharyya:2015fdk} together with
\begin{eqnarray} \nonumber
\frac{u^{E}K_{EB}K_{a}^{B}}{\mathcal{K}}&=&-\epsilon\frac{(\nabla_{a}\partial_{t}\delta r-\delta u_{a})}{D-2}\\\nonumber
\frac{\nabla ^2\nabla ^2 u_{a}}{\mathcal{K}^3}&=&\epsilon\frac{\bar{\nabla} ^2 \bar{\nabla} ^2 \delta u_{a}+\bar{\nabla }^2 \nabla_{a}\partial_{t}\delta r}{(D-2)^3} \\\nonumber
\frac{K^{CD}\nabla_C\nabla_D u_{a}}{\mathcal{K}^2}&=&\epsilon\frac{\bar{\nabla} ^2 \delta u_{a}- \nabla_{a}\partial_{t}\delta r}{(D-3)(D-2)}\\\nonumber
\frac{\nabla_{a}\nabla ^2 \mathcal{K}}{\mathcal{K}^3} &=&-\epsilon\frac{\nabla_{a} \bar{\nabla}^2(\bar{\nabla}^2 \delta r+\delta r(D-2))}{(D-2)^3}\\\nonumber
\frac{\nabla_{a}(K^{BC}K_{BC}\mathcal{K})}{\mathcal{K}^3}&=&\epsilon\frac{3\nabla_{a}(-\bar{\nabla}^2\delta r-\delta r(D-2))}{(D-3)(D-2)}\\\nonumber
\end{eqnarray}
(the equations above are accurate only to linear order in $\epsilon$ and all covariant
derivatives are taken with respect to \eqref{IndFlat}). The linearized membrane equation
is given by
\begin{equation}\label{LinEE}
\begin{split}
&\left[ \left(1+\frac{\overline{\nabla}^2}{D-2}\right)\delta u_a + \nabla_a\left(1+\frac{\overline{\nabla}^2}{D-2}\right)\delta r-\partial_t\nabla_a\delta r \left( 1-\frac{1}{D-2} \right) - \partial_t\delta u_a \right] \\
& +\bigg[ \frac{\nabla_a\partial_t \delta r-\delta u_a}{D-2}+\frac{\overline{\nabla}^2\overline{\nabla}^2 \delta u_a+\overline{\nabla}^2 \nabla_a \partial_t \delta r}{(D-2)^3}+ 2\frac{-\overline{\nabla}^2 \delta u_a+\nabla_a \partial_t \delta r}{(D-3)(D-2)}+\frac{\nabla_a \overline{\nabla}^2(\overline{\nabla}^2 \delta r+(D-2)\delta r)}{(D-2)^3} \\&+ 3\frac{\nabla_a(-\overline{\nabla}^2 \delta r-(D-2)\delta r)}{(D-3)(D-2)} + 3\frac{\partial_t \delta u_a}{(D-3)} + 3\frac{\partial_t \nabla_a \delta r-\delta u_a}{(D-3)} \bigg] =0.
\end{split}
\end{equation}
(\eqref{LinEE} generalizes equation (5.9) of \cite{Bhattacharyya:2015fdk}).
Substituting \eqref{veldec1} into \eqref{LinEE} we find the generalized version of of
(5.15) of \cite{Bhattacharyya:2015fdk},
\begin{equation} \label{VElin}
\begin{split}
&\left(\frac{\nabla^2}{D-2}+ 1 - \partial_t +\frac{\overline{\nabla}^2 \overline{\nabla}^2 }{(D-2)^3}- \frac{2 (\nabla^2)}{(D-2)^2}+\frac{3 \partial_t}{(D-3)}-\frac{3 }{(D-3)}\right)\delta v_a=\\
& -\Bigg(\frac{\partial_t \nabla_a }{D-2}+ \frac{\nabla_a \nabla^2 }{D-2} + \nabla_a -\nabla_a \partial_t + \frac{2 \nabla_a \partial_t }{(D-2)^2}- \frac{\nabla_a \overline{\nabla}^2(\nabla^2 + (D-2))}{(D-2)^3}\\& - \frac{9 \nabla_a ( (D-2)^2 -(D-2)(9\nabla^2 - \partial_t^2 ))}{3(D-2)^3}+\frac{3 \partial_t \nabla_a }{(D-3)}\Bigg)\delta r\\
& -\Bigg(\frac{\nabla^2}{D-2}+ 1 - \partial_t +\frac{\overline{\nabla}^2 \overline{\nabla}^2 }{(D-2)^3}- \frac{2 (\nabla^2)}{(D-2)^2}+\frac{3 \partial_t}{(D-3)}-\frac{3}{(D-3)}\Bigg) \nabla_a \Phi
\end{split}
\end{equation}
\subsection{Scalar quasinormal modes}
Using \eqref{velco2} and \eqref{phicon} we take the divergence of \eqref{VElin} to obtain
\begin{equation}\label{DivSca}
\begin{split}
&-(\overline{\nabla}^2+D-3)\partial_{t}\delta r+\frac{\partial_{t}\nabla^2 \delta r}{D-2}+\frac{\nabla^2\bar{\nabla}^2\delta r}{(D-2)}+\nabla^2 \delta r-\partial_{t}\nabla^2 \delta r-(D-2)\partial_{t} \delta r+(D-2)\partial_{t}^2 \delta r \\&+ \frac{\nabla^2\partial_{t} \delta r+ (D-2)\partial_{t} \delta r}{D-2} - \frac{ (\overline{\nabla}^2+D-3)^2(D-2)\partial_{t}\delta r+(\overline{\nabla}^2+D-3)\nabla^2\partial_{t}\delta r}{(D-2)^3} \\&+ 2\frac{(\overline{\nabla}^2+D-3)(D-2)\partial_t\delta r+\nabla^2\partial_{t}\delta r}{(D-2)^2}+\frac{\nabla^2\bar{\nabla}^2(\bar{\nabla}^2 \delta r +\delta r(D-2))}{(D-2)^3}\\ &-\frac{\nabla^2(3\nabla^2 \delta r-\partial_t^2\delta r+3\delta r(D-2))}{(D-2)^2}-3\frac{D-2}{(D-3)}\partial_{t}^2 \delta r +\frac{3}{(D-3)}(\partial_{t} \nabla^2 \delta r +(D-2)\partial_{t} \delta r)=0
\end{split}
\end{equation}
As in \cite{Bhattacharyya:2015fdk} we expand
\begin{equation}\label{ScaFlu}
\delta r= \sum_{l,m}a_{lm} Y_{lm} e^{-i \omega^r_l t}~~.
\end{equation}
where the spherical harmonics $Y_{lm}$ obey
\begin{equation} \label{ScalSph}
-\nabla_{S^{D-2}}^2 Y_{lm} = l(D+l-3) Y_{lm}.
\end{equation}
Inserting \eqref{ScaFlu} into \eqref{DivSca} we obtain
\begin{equation}\label{ScaQNM}
\omega_l^r=\pm \sqrt{l-1}-i(l-1)+\frac{1}{D}\left(\pm \sqrt{l-1}\left(\frac{3l}{2}-2\right)-i(l-1)(l-2)\right)
\end{equation}
The result \eqref{ScaQNM} is in perfect agreement with the result obtained by EST in Equations (5.30) and (5.31) of \cite{Emparan:2014aba}.
As explained in \cite{Bhattacharyya:2015fdk}, the modes with $l=0$ and $l=1$ are special. At $l=0$ the formula
\eqref{ScaQNM} yields $\omega=0,2i-\frac{4i}{D}$. The second solution is, however, spurious
(see \cite{Bhattacharyya:2015fdk}). The first solution is the zero mode corresponding to rescaling the black hole; this
is an exact zero mode at all orders in $1/D$.
At $l=1$ \eqref{ScaQNM} yields the frequencies $\omega=0,0$. As explained in \cite{Bhattacharyya:2015fdk} these two modes
correspond to translations and boosts of the membrane.
\subsection{Vector quasinormal modes}
We expand the velocity fluctuations in a basis of vector spherical harmonic
\begin{equation}\label{VecFlu}
\delta v_a = \sum_{l,m} b_{lm} Y_a^{lm} e^{-i \omega^v_l t}
\end{equation}
Where, $l=1,2,3,...$. The vector spherical harmonics satisfy the property
\begin{equation} \label{VecSph}
\nabla^2 V=- [(D+l-3)l-1]V
\end{equation}
Plugging \eqref{VecFlu} into \eqref{VElin}, using \eqref{VecSph} and equating the coefficients of
independent vector spherical harmonics (see \cite{Bhattacharyya:2015fdk} for more discussion) we obtain
\begin{equation}\label{VecQNM}
\omega^v_l = -i(l-1)-\frac{i}{D} (l-1)^2.
\end{equation}
\eqref{VecQNM} is in perfect agreement with the formula (5.22) of \cite{Emparan:2014aba} derived earlier by EST by purely
gravitational analysis. Note that the mode with $l=1$ has vanishing frequency. As explained in
\cite{Bhattacharyya:2015fdk} $l=1$ is the exact zero mode corresponding to setting the black hole spinning.
\section{Discussion}
In this paper we have worked out the duality between the dynamics of black holes in a large number of dimensions
and the motion of a non gravitational membrane in flat space to second subleading order in $1/D$. Our work
generalizes the analysis of \cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}. The
concrete new results of this paper are
\begin{itemize}
\item The second order corrected membrane equations of motion listed in \eqref{VE2copy}.
\item The formula \eqref{Divu} for the divergence of the velocity field (which vanished at first order).
\item The explicit form of the second order corrected metric dual to any given membrane motion
(see subsection \ref{2ndmet}
\end{itemize}
In addition to obtaining the new results listed above we have also achieved
an improved understanding of the structure of the perturbative expansion in
$1/D$. We have demonstrated that the perturbative programme,
implemented to first nontrivial order in
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk}, can systematically be
extended to every order in the $1/D$ expansion. In particular we have shown
that the algebraically nontrivial `integrability' properties that
allowed for the existence of a first order solution in
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} are actually automatic at
all orders as as a consequence of the well known equation \eqref{bianchi}.
We have also explained that the membrane equations may directly be obtained
by evaluating the Einstein constraint equation on the event horizon.
In particular the membrane equations at $(n+1)^{th}$ order in $1/D$ are
obtained by evaluating the constraint equations on $n^{th}$ order metric,
without needing to solve for the $(n+1)^{th}$ order metric. We have also
explained that the assumption of $SO(D-p-2)$ isometry, made in
\cite{Bhattacharyya:2015fdk}, is not necessary; the membrane equations
can be derived under much more general conditions
The fact our membrane equations arise from the Einstein constraint equations at the event horizon is
strongly reminiscent of the `traditional' membrane paradigm of
black hole physics. It would be very interesting to better understand the relationship between the the
large $D$ membrane and the traditional membrane paradigm.
\cite{Price:1986yy,Price:1988sci,PhysRevD.18.3598}.
As black holes are thermodynamical objects, the black hole membrane studied in
\cite{Bhattacharyya:2015dva,Bhattacharyya:2015fdk} and this paper should carry an entropy current. At leading order
in $1/D$ it turns out (see \cite{bh}) that this
entropy current is given simply by a constant times $u^M$. The divergence of this entropy current is thus proportional to
$\nabla.u$. It follows that the RHS of the formula \eqref{Divu} gives an expression
for the rate of entropy production on the membrane. It would be interesting to further investigate this observation
and its consequences.
On a related note, it would be interesting to derive the most general stationary solution of the second order
corrected equations of motion derived in this paper and compare our results with those of \cite{Suzuki:2015iha}.
In this paper we have focused our attention on black holes propagating in an otherwise perfectly flat spacetime.
It would be interesting to generalize our study to the motion of black holes propagating in any vacuum solution
of Einstein's equations, e.g. a gravity wave. Such a generalization would allow us, for instance, to
study the absorption of gravity waves by black holes at large $D$. At first order in the derivative
expansion we expect the generalized effective membrane equation to be given simply by covariantizing first order flat
space equations of motion. At second order, however, the equations of motion could receive genuinely new
contributions from the background Riemann tensor of the space in which the black hole propagates
\footnote{Something similar happens in the study of forced fluids in the fluid gravity correspondence \cite{Bhattacharyya:2008ji}}.
It would be interesting to work this out in detail.
Finally, it would be interesting to put the membrane equations derived in this paper to practical use to allow us
to learn new things about black holes. One possible direction would be to test out how well the large D expansion
does in astrophysical contexts (i.e. when $D=4$). Another direction would be to use the formalism developed herein to
address interesting unanswered structural questions about gravity, e.g. questions about the second law of
thermodynamics in higher derivative gravity. We leave such investigations for the future.
\vskip .8cm
\section*{Acknowledgments}
We would like to thank K. Inbasekar , S. Thakur and M. Mandlik for many useful
discussions during the progress in the project. We would like especially to
thank S. Bhattacharyya for several very useful discussions and explanations.
S. Mazumdar would like to acknowledge the hospitality of the University of
Barcelona , ETH Zurich and IISER Mohali while this work was in progress.
Y.D. and S. Mazumdar would also like to thank ICTP,Trieste for hospitality
while this work was in progress. Y.D. ,A.D., S. Mazumdar and A.S would
also like to thank the organizers of The Fourth Indo-Israel Meeting, Goa
for hospitality while this work was in progress . S.M. would like to
thank IAS Princeton for hospitality while this work was in progress. The work
of all authors was supported by the Infosys Endowment for the study of the
Quantum Structure of Spacetime, as well as an Indo Israel (UGC/ISF) grant.
Finally we would all like to acknowledge our debt to the people of India
for their steady and generous support to research in the basic sciences.
|
\section{Introduction}
Quantum technology relies on the utilization of resources, like quantum coherence and entanglement, which improve considerably quantum information processing protocols relative classical ones $\cite{Perrin, Nielsen, Benenti}$.
However, the quantum entanglement is so fragile and undergoes either an asymptotic decay or a sudden death $\cite{Hornberger, Merali}$.
This is due to decoherence, whereby the unavoidable interaction between any real quantum system with its surrounding environment alters the quantum system and consequently disentanglement occurs. Therefore, it is very important to investigate the physical systems and the physical effects that
may lead to effective long-time entanglement preservation.
Although there are so-called decoherence-free states which the initial entanglement remains invariant in time, however, there is only a certain kind of entangled state which can be
decoupled from the influence of the environment in this way $\cite{Maniscalco, Li, An1}$.
So far, a lot of researches have been devoted to entanglement manipulation and protection.
For instance, the quantum Zeno effect (QZE) is a promising way to avoid the decaying behaviour of the entanglement in dissipative systems. This effect which refers to the inhibition of the temporal evolution of a quantum system by repeated projective measurements during a defined period of time, has been discussed in $\cite{Maniscalco, Mundarain, Rossi, Hou}$.
But, this scheme is relatively difficult since one has to perform a series of measurements to the system during the course of the evolution.
The other scheme for protecting the entanglement is detuning modulation $\cite{An1, Xiao}$, where the quantum entanglement is not preserved in the long-time limit.
Another proposed scheme, without the disadvantages of the
aforementioned schemes, focuses on the long time limit protection of quantum
entanglement through the additional qubits. The protection of entanglement between two qubits via the additional qubits was first observed in $\cite{An2}$, where the bipartite entanglement is sustained by the addition of a third qubit. Improving the preservation of entanglement by much more additional qubits introduced
in $\cite{An3, Flores}$, where all of the entangled qubits and the additional ones were contained in a common environment.
Distantly non-interacting two-level quantum systems, each of which interacts with an
environment independently, are preferable elements for a quantum hardware in order to
accomplish the individual control required for quantum information processing $\cite{Ladd, Xiang}$.
Therefore, finding strategies to protect quantum resources, such as coherence and entanglement, is an essential task in these configurations $\cite{Franco}$.
In this paper, we consider a system of $N$ non-interacting qubits immersed in a common zero-temperature thermal Lorentzian reservoir and show how quantum coherence of a given single-qubit in this system can be preserved by increasing the number of additional $N-1$ qubits. At the long time limit ($t\rightarrow\infty$), we show that, apart from Markovian and non-Markovian dynamics, the coherence measure for the single-qubit approaches asymptotically to a non-zero steady value depending on the number of additional qubits and the initial state coherence of the single-qubit.
Also, by considering $N\rightarrow\infty$, our calculations show that the coherence measure reaches to its initial value.
In the next step, we extend this scheme to protect bipartite and tripartite entanglements
among two and three distant qubits, each of which contained in an independent reservoir.
It is observed that adding other qubits to any of the reservoirs, leads to protecting entanglement from sudden death. At the limit $t\rightarrow\infty$, we obtain simple analytical expressions for each of the bipartite concurrence $\cite{Wootterrs}$ and tripartite lower bound of concurrence (LBC) $\cite{gao}$
in terms of the number of additional qubits contained in each of the reservoirs and
the respective initial state entanglement. Moreover, increasing the number of additional qubits in
each of the reservoirs completely preserves the initial state entanglement from the dissipative processes
of the reservoirs.
As illustrated in the text, this mechanism of entanglement preservation works even for different Markovian and non-Markovian dynamics occurred individually in each of the related subsystems. After this Introduction, in Sec. II we formulate our scheme for preserving the quantum coherence of a single-qubit. Sec. III is devoted for extending the scheme proposed in Sec II for protection of bipartite and tripartite entanglements. And finally, the paper is ended with a brief conclusion in Sec. IV.
\section{Quantum coherence preservation}
In this section, we consider a system consisting of $N$ independent qubits (two-level atoms) immersed in a common zero-temperature thermal reservoir,
as depicted in Fig. 1(a). The Hamiltonian $\hat{H}$ of the system contains two parts ($\hbar=1$)
\begin{eqnarray}
\hat{H}=\hat{H}_{0}+\hat{H}_{I},
\end{eqnarray}
with $\hat{H}_{0}$ the free Hamiltonian and $\hat{H}_{I}$ describes the interaction Hamiltonian
\begin{eqnarray}
\hat{H}_{0}=\Omega \sum_{j=1}^{N} {\hat{\sigma}_{j}^{+} \hat{\sigma}_{j}^{-}}+\sum_{k} \omega_{k} \hat{b_{k}}^{\dagger} \hat{b_{k}},
\end{eqnarray}
\begin{eqnarray}
\hat{H}_{I}= \sum_{j=1}^{N} \sum_{k} \beta_{j}(g_{k} {\hat{\sigma}_{j}^{+}} \hat{b_{k}}+g_{k}^{*} {\hat{\sigma}_{j}^{-}} \hat{b_{k}}^{\dagger}),
\end{eqnarray}
where $\hat{\sigma}_{j}^{+}(\hat{\sigma}_{j}^{-})$ is the raising (lowering) operator of the $j^{th}$ qubit with transition frequency $\Omega$
and $\hat{b_{k}}$ ($\hat{b_{k}^{\dagger}}$) is the annihilation (creation) operator of the $k^{th}$ field mode with frequency $\omega_{k}$.
The strength of coupling between the $j^{th}$ qubit and the $k^{th}$ field mode is described by $g_{k}$ and
the dimensionless real constants $\beta_{j}$ are introduced to individualize the qubits.
It is convenient to work in the interaction picture where the state $|\psi(t)\rangle$ of the system obeys the Schrodinger equation as
\begin{eqnarray}
i \frac{d}{dt}|\psi(t)\rangle=\hat{H}_{I}(t) |\psi(t)\rangle,
\end{eqnarray}
and the Hamiltonian in this picture is given by
\begin{eqnarray}
\hat{H}_{I}(t)=e^{i \hat{H}_{0}t} \hat{H}_{I} e^{-i \hat{H}_{0}t}=\sum_{j=1}^{N} \sum_{k} \beta_{j}(g_{k} {\hat{\sigma}_{j}^{+}} \hat{b_{k}} e^{i(\Omega-\omega_{k})t}+g_{k}^{*} {\hat{\sigma}_{j}^{-}} \hat{b_{k}}^{\dagger} e^{-i(\Omega-\omega_{k})t}).
\end{eqnarray}
Since the total Hamiltonian commutes with the number of excitations, i.e. $\big[(\sum_{j=1}^{N} {\hat{\sigma}_{j}^{+} \hat{\sigma}_{j}^{-}}+\sum_{k} \hat{b_{k}}^{\dagger} \hat{b_{k}}),H\big]=0$,
therefore, any initial state of the form
\begin{eqnarray}
|\psi(0)\rangle=C_{0}(0)|0\rangle_{S} |0\rangle_{E}+\sum_{j=1}^{N}C_{j}(0)|j\rangle_{S} |0\rangle_{E},
\end{eqnarray}
evolves after time $t$ into the following state
\begin{eqnarray}
\begin{array}{c}
|\psi(t)\rangle=C_{0}(t)|0\rangle_{S} |0\rangle_{E}+\sum_{j=1}^{N}C_{j}(t)|j\rangle_{S} |0\rangle_{E}+\sum_{k} D_{k}(t) |0\rangle_{S} |1_{k}\rangle_{E},
\end{array}
\end{eqnarray}
where $|j\rangle_{S}=|g\rangle^{\bigotimes N}_{j^{th}\equiv e}$, which means that all of the qubits are in the ground state $|g\rangle$ except the $j^{th}$ qubit,
which is in the excited state $|e\rangle$ and $|0\rangle_{S}=|g\rangle^{\bigotimes N}=|g,g,...,g\rangle$.
Also, we denote $|0\rangle_{E}$ being the vacuum state of the reservoir and $|1_{k}\rangle_{E}$ the state of it with only one excitation in the $k^{th}$ field mode.
Substituting Eqs. (5) and (7) into Eq. (4), gives the following set of $N+1$ differential equations for the probability amplitudes as
\begin{eqnarray}
\dot{C}_{j}(t)=-i \beta_{j} \sum_{k} g_{k} D_{k}(t) e^{i(\Omega-\omega_{k})t},\\
\dot{D}_{k}(t)=-i \sum_{j=1}^{N} \beta_{j} g_{k}^{*} C_{j}(t) e^{-i(\Omega-\omega_{k})t},
\end{eqnarray}
where $j=1,2,...,N$. It is clear that $\dot{C}_{0}(t)=0$, then $C_{0}(t)=C_{0}(0)=C_{0}$. Integrating Eq. (9) and substituting it into the Eq. (8), gives
the following set of $N$ integro-differential equations
\begin{eqnarray}
\frac{dC_{j}(t)}{dt}=-\int_{0}^{t} f(t-t')\beta_{j} \sum_{l=1}^{N} \beta_{l} C_{l}(t') dt',
\end{eqnarray}
and the correlation function $f(t-t')$ is related to the spectral density $J(\omega)$ of the reservoir by
\begin{eqnarray}
f(t-t')=\int_{-\infty}^{\infty} d\omega J(\omega) e^{i(\Omega-\omega)(t-t')}.
\end{eqnarray}
The exact form of $C_{j}(t)$ thus depends on the particular choice for the spectral density of the reservoir. We take the Lorentzian spectral density for the reservoir as
\begin{eqnarray}
J(\omega)=\frac{1}{2\pi} \frac{\gamma_{0} \lambda^{2}}{(\omega-\Omega+\Delta)^2+\lambda^{2}},
\end{eqnarray}
where $\Delta$ is the detuning between the transition frequency of the qubits $\Omega$ and the central frequency of the reservoir.
The parameter $\lambda$ defines the spectral width of the coupling and the parameter $\gamma$ is the coupling constant.
By using the spectral density $J(\omega)$ given by Eq. (12), the exact solutions of the probability amplitudes $C_{j}(t)$ can be obtained (see appendix A) as
\begin{eqnarray}
\begin{array}{c}
C_{j}(t)=e^{-\Lambda t/2}\bigg(\mathrm{cosh}{(\frac{Dt}{2})}+\frac{\Lambda}{D} \mathrm{sinh}{(\frac{Dt}{2}})\bigg) C_{j}(0)+\\\\
\frac{\sum_{l\neq j}^{N} {\beta_{l}}^2 C_{j}(0)-\beta_{l} \beta_{j} C_{l}(0)}{\sum_{l=1}^{N} {\beta_{l}}^2}\bigg(1-e^{-\Lambda t/2}\Big(\mathrm{cosh}{(\frac{Dt}{2})}+\frac{\Lambda}{D} \mathrm{sinh}{(\frac{Dt}{2})}\Big)\bigg),
\end{array}
\end{eqnarray}
where $\Lambda=\lambda-i\Delta$ and $D=\sqrt{\Lambda^{2}-2\gamma_{0} \lambda \sum_{l=1}^{N} {\beta_{l}}^2}$.
Here, there are two regimes for the system environment coupling $\cite{An1,An2,Flores}$: weak coupling regime ($\lambda>2\gamma_{0} \sum_{l=1}^{N} {\beta_{l}}^2$)
and strong coupling regime ($\lambda<2\gamma_{0} \sum_{l=1}^{N} {\beta_{l}}^2$).
In the weak coupling regime the behaviour of the system is Markovian and irreversible decay occurs. However, in
the strong coupling regime the non-Markovian dynamics occurs accompanied by an oscillatory reversible decay.
After tracing out the zero-temperature thermal reservoir and the qubits except $j^{th}$ one, the reduced density
matrix of the $j^{th}$ qubit in the basis $\{|e\rangle, |g\rangle \}$, at time $t$, becomes
\begin{eqnarray}
\rho_{j}(t)=\left(
\begin{array}{cc}
|C_{j}(t)|^2 & C_{0}^{*} C_{j}(t) \\\\
C_{0} C_{j}^{*}(t) & 1-|C_{j}(t)|^2 \\
\end{array}
\right).
\end{eqnarray}
Furthermore, if we let $\beta_{l}=1$ for $l=1,2,...,N$ and $C_{l}(0)=0$ with $l\neq j$, Eq. (13) reduces to
\begin{eqnarray}
C_{j}(t)=G(t)C_{j}(0),
\end{eqnarray}
where
\begin{eqnarray}
G(t)=\frac{N-1}{N}+\frac{e^{-\Lambda t/2}}{N} \Big(\mathrm{cosh}{(\frac{Dt}{2})}+\frac{\Lambda}{D} \mathrm{sinh}{(\frac{Dt}{2})}\Big),
\end{eqnarray}
and $D=\sqrt{\Lambda^{2}-2\gamma_{0} \lambda N}$. The reduced density matrix of the $j^{th}$ qubit in Eq. (14), can be rewritten as
\begin{eqnarray}
\rho_{j}(t)=\left(
\begin{array}{cc}
|G(t)|^{2}|C_{j}(0)|^2 & C_{0}^{*}G(t)C_{j}(0) \\\\
C_{0}G(t)^{\ast}C_{j}^{*}(0) & 1-|G(t)|^{2}|C_{j}(0)|^2 \\
\end{array}
\right).
\end{eqnarray}
The $j^{th}$ qubit dynamics thus depends only on the function ${G(t)}$ ($0 < |G(t)| \leq 1$), which in turns depends on the spectral density and the number of qubits. Also, as discussed in $\cite{Breuer}$, by differentiating of Eq. (17) with respect to time and comparing it with
an exact master equation, decay rate of the $j^{th}$ qubit can be obtained as
\begin{eqnarray}
\Gamma(t)=-2 Re\Big\{\frac{\dot{C}_{j}(t)}{C_{j}(t)}\Big\}=Re\bigg\{\frac{2\gamma_{0} \lambda e^{-\Lambda t/2} sinh{(\frac{Dt}{2})}}{D\bigg(\frac{N-1}{N}+\frac{e^{-\Lambda t/2}}{N} \Big(\mathrm{cosh}{(\frac{Dt}{2})}+\frac{\Lambda}{D} \mathrm{sinh}{(\frac{Dt}{2}})\Big)\bigg)}\bigg\}.
\end{eqnarray}
Baumgratz $et.al$ introduced an intuitive measure of quantum coherence based on the off-diagonal elements of density matrix for the desired quantum state $\cite{Baumgratz}$, as
\begin{eqnarray}
\xi(\rho(t))=\sum_{m,n(m\neq n)} |\rho_{m,n}(t)|,
\end{eqnarray}
where $\rho_{m,n}(t) (m\neq n)$ are the off-diagonal elements of the system density matrix. Indeed, it has been
recently shown that $\xi(\rho(t))$ satisfies the physical requirements which make it as a proper coherence measure.
Suppose the $j^{th}$ qubit is initially prepared in the state $\alpha_{1} |g\rangle+\alpha_{2} |e\rangle$ ($|\alpha_{1}|^2+|\alpha_{2}|^2=1$), the other qubits are prepared in the state $|g\rangle$ and the reservoir also is in its respective vacuum state, so in this way, $\xi(\rho_{j}(0))=2|C_{0} C_{j}(0)|=2|\alpha_{1} \alpha_{2}|$. Therefore, at time $t>0$, the coherence of the $j^{th}$ qubit becomes $\xi(\rho_{j}(t))=2|C_{0} C_{j}(t)|=2|G(t)||C_{0} C_{j}(0)|$.
The dynamical behavior of the coherence in terms of the dimensionless time $\gamma_{0} t$
has been shown in Fig. 2, where we have assumed $C_{0}=C_{j}(0)=1/\sqrt{2}$.
In Fig. 2 (a, b), the coherence measure of the qubit exhibits a Markovian dynamics and monotonically approaches to zero in the absence of additional qubits ($N=1$).
However, it is readily observed that it can be greatly inhibited by increasing the number of additional qubits ($N=2,3,6$).
Moreover, in the non-Markovian regime and in the absence of additional qubits ($N=1$), the measure of coherence oscillatory undergoes to sudden death and in the presence of additional qubits ($N=2,3,6$), undergoes to a non-zero steady value, as depicted in Fig. 2(c, d).
The oscillations of coherence in the non-Markovian regime constitute a sufficient condition to signify the presence of memory effects in the system dynamics,
being due to information backflow from the environment to the system.
We point out that the existence of detuning ($\Delta$) tend to slow down the decay process of the coherence in the non-Markovian regime.
It should be noted that, in the long time limit, the measure of coherence reduces to the following expression
\begin{eqnarray}
\xi(\rho_{j}(t))=2|C_{0} C_{j}(t)|=2|\alpha_{1} \alpha_{2}|(\frac{N-1}{N}).
\end{eqnarray}
Indeed, the single-qubit coherency can be determined only by knowing the number of additional $N-1$ qubits in the reservoir and the initial state of the qubit.
Also, as $N\rightarrow \infty$, $\xi(\rho_{j}(t))$ reaches to $2|\alpha_{1} \alpha_{2}|$ (i.e. the initial coherency of the $j^{th}$ qubit).
On the other hand, these observations are confirmed by the behavior of the related decay rate (18) in terms of the scaled parameter $\gamma_{0} t$, as shown in Fig. 3.
Before extending the obtained results for protecting entanglement in the next sections, let us make a discussion in order to clarify the connection between the proposed scheme in this paper and the decoherence-free subspace method $\cite{lid1, lid2, lid3}$. According to these references, consider that the related Hilbert space of the qubit system regarded as our open system is as $\mathcal{H}_{S}=\mathcal{H}_{D-F}\oplus \mathcal{H}_{N}$ where $\mathcal{H}_{D-F}$ is the decoherence-free subspace and $\mathcal{H}_{N}$ denotes the noisy one. When we have only the considered single-qubit, the initial state of this qubit completely belongs to the noisy subspace $\mathcal{H}_{N}$. Therefore, at long time limit and according to Eq. (20), the qubit ultimately loses its coherency as shown in Fig. 2. In fact, in this stage, there is no decoherence-free subspace $\mathcal{H}_{D-F}$. As the non-interacting additional qubits enter to the reservoir some decoherence-free or subradiant states are created and we have a $\mathcal{H}_{D-F}$ subspace where the initial state of the single-qubit has a non-zero support in that subspace. So, in this way, it is observed a non-zero steady value for the coherence measure (see Eq. (20) and Fig. 2). Consequently, when the number of additional qubits becomes very large (infinity), the $\mathcal{H}_{D-F}$ subspace will be sufficiently large such that the initial state of the respective single-qubit completely belongs to this subspace so its coherence remains unchanged. It is concluded, in this regard, that the decoherence-free subspace can be effectively manipulated through the additional qubits. This argument is also valid for entanglement protection which will be discussed in the next sections.
Since the appearance of decoherence-free subspace $\mathcal{H}_{D-F}$ depends only on the presence of additional qubits so it is expected that it can not be depend on the structure of the reservoir as confirmed in $\cite{bh}$. Therefore, the preservation of quantum coherence and entanglement, in long time limit, can not be dependent on the spectral shape of the reservoir such as Lorentzian, Ohmic, sub-Ohmic or super-Ohmic cases.
\section{Quantum entanglement preservation}
\subsection{EPR-type entanglement}
To achieve to the scheme of entanglement preservation, we consider a composite system consisting of two subsystems $A$ and $B$ contained in two independent Lorenzian reservoirs.
Each of the subsystems contains $N_{A}$ and $N_{B}$ qubits, respectively.
Let's consider the $j^{th}$ qubit of subsystem $A$ and the $l^{th}$ qubit of subsystem $B$ prepared initially in an EPR-type entangled state as follows
\begin{eqnarray}
|\phi(0)\rangle_{j,l}=C_{j}^{A}(0)|e,g\rangle+C_{l}^{B}(0)|g,e\rangle,
\end{eqnarray}
as depicted in Fig. 1 (b).
According to the Ref. $\cite{Bellomo}$, the complete dynamics of the above two-qubit system can be obtained easily (see appendix B). Therefore, in the standard computational basis such as $\{|1\rangle \equiv|e,e\rangle, |2\rangle \equiv |e,g\rangle, |3\rangle \equiv |g,e\rangle, |4\rangle \equiv |g,g\rangle\}$, the explicit forms of the matrix elements of the density operator at time $t$ becomes
\begin{eqnarray}
\begin{array}{c}
\rho_{22}(t)=|G^{A}(t)|^{2}|C_{j}^{A}(0)|^2,\\\\
\rho_{33}(t)=|G^{B}(t)|^{2}|C_{l}^{B}(0)|^2,\\\\
\rho_{44}(t)=1-|G^{A}(t)|^{2}|C_{j}^{A}(0)|^2-|G^{B}(t)|^{2}|C_{l}^{B}(0)|^2,\\\\
\rho_{23}(t)=\rho_{32}^{*}(t)=G^{A}(t) {G^{B}}^{*}(t) C_{j}^{A}(0) {C_{l}^{B}}^{*}(0),\\\\
\rho_{11}(t)=\rho_{12}(t)=\rho_{13}(t)=\rho_{14}(t)=\rho_{24}(t)=\rho_{34}(t)=0.
\end{array}
\end{eqnarray}
where $G^{A}(t)$, $G^{B}(t)$ are determined in similar way as case of the Eq. (16).
The parameter $\lambda_{A} (\lambda_{B})$ which appears in the $G^{A}(t)$ ($G^{B}(t)$) defines the spectral width for the coupling of the qubits at subsystem $A$ (subsystem $B$) to its respective reservoir. Also, we assume that the coupling constants of the qubits to their respective reservoirs are equal, i.e. $\gamma_{A}=\gamma_{B}=\gamma_{0}$.
To quantify the amount of entanglement for the state (22), we use concurrence as a measure of two-qubit entanglement. The following analytical form for the concurrence is obtained
\begin{eqnarray}
C(\rho_{j,l}(t))=2 |G^{A}(t) G^{B}(t)| |C_{j}^{A}(0) C_{l}^{B}(0)|.
\end{eqnarray}
The time dependency of the concurrence has been shown in Fig. 4 in the Markovian and Non-Markovian regimes.
Notice that in the case of $N_{A}=N_{B}=1$ (in the absence of additional qubits in the subsystems A and B), the concurrence ultimately decays to zero but there is a steady non-zero value for it by increasing $N_{A}$ and $N_{B}$ in both of the Markovian and non-Markovian regimes.
An interesting result for the concurrence occurs at the asymptotical limit $t\rightarrow \infty$, as follows
\begin{eqnarray}
C(\rho_{j,l})=2 \frac{(N_{A}-1)(N_{B}-1)}{N_{A}N_{B}} |C_{j}^{A}(0) C_{l}^{B}(0)|.
\end{eqnarray}
Indeed, in the long time limit, the concurrence can be determined only by knowing the number of qubits in the reservoirs and the initial state of the two-qubit system.
It is clear that when $N_{A}=1$ and $N_{B}>1$ or $N_{B}=1$ and $N_{A}>1$, then the concurrence decays to zero in the long time limit.
On the other hand,
as $N_{A}$, $N_{B}\rightarrow \infty$, the concurrence approaches to its initial value, i.e. 2$|C_{j}^{A}(0) C_{l}^{B}(0)|$. Also, if we let, for instance, $N_{B}\rightarrow \infty$ and rename $N_{A}\equiv N$ then the Eq. (24) becomes
\begin{eqnarray}
C(\rho_{j,l})=2 \frac{N-1}{N} |C_{j}^{A}(0) C_{l}^{B}(0)|.
\end{eqnarray}
The concurrence evolution in (23) can be evaluated for different non-Markovian behaviors of the related subsystems as shown in Fig. 5 with $N_{A}=N_{B}=6$. At the long time limit, it is interesting to note that, apart from different non-Markovian nature of the subsystems, the asymptotic concurrences are completely coincided to each others.
\subsection{W-type entanglement}
In this section, we extend the results of the previous section to three subsystems $A$, $B$ and $C$ contained in three independent Lorenzian reservoirs (see Fig. 1(c)).
Each subsystem has $N_{A}$, $N_{B}$ and $N_{C}$ qubits respectively. A W-type entangled state of $j^{th}$ qubit of subsystem $A$, $l^{th}$ qubit of subsystem $B$ and $m^{th}$ qubit of subsystem $C$, at time $t=0$, can be written as follows
\begin{eqnarray}
|W\rangle_{ABC}=C_{j}^{A}(0)|e,g,g\rangle+C_{l}^{B}(0)|g,e,g\rangle+C_{m}^{C}(0)|g,g,e\rangle.
\end{eqnarray}
As shown in appendix B, in the standard computational basis as $\{|1\rangle \equiv|e,e,e\rangle, |2\rangle \\ \equiv|e,e,g\rangle, |3\rangle \equiv|e,g,e\rangle,|4\rangle \equiv|e,g,g\rangle, |5\rangle \equiv|g,e,e\rangle, |6\rangle \equiv|g,e,g\rangle, |7\rangle \equiv|g,g,e\rangle, |8\rangle \equiv|g,g,g\rangle\}$, the matrix elements of the three-qubit density operator at time $t$ are
\begin{eqnarray}
\begin{array}{c}
\rho_{44}(t)=|G^{A}(t)|^{2} |C_{j}^{A}(0)|^2,\\\\
\rho_{66}(t)=|G^{B}(t)|^{2} |C_{l}^{B}(0)|^2,\\\\
\rho_{77}(t)=|G^{C}(t)|^{2} |C_{m}^{C}(0)|^2,\\\\
\rho_{88}(t)=1-|G^{A}(t)|^{2}|C_{j}^{A}(0)|^2-|G^{B}(t)|^{2}|C_{l}^{B}(0)|^2-|G^{C}(t)|^{2}|C_{m}^{C}(0)|^2,\\\\
\rho_{46}(t)=\rho_{64}^{*}(t)=G^{A}(t) {G^{B}}^{*}(t) C_{j}^{A}(0) {C_{l}^{B}}^{*}(0),\\\\
\rho_{47}(t)=\rho_{74}^{*}(t)=G^{A}(t) {G^{C}}^{*}(t) C_{j}^{A}(0) {C_{m}^{C}}^{*}(0),\\\\
\rho_{67}(t)=\rho_{76}^{*}(t)=G^{B}(t) {G^{C}}^{*}(t) C_{l}^{B}(0) {C_{m}^{C}}^{*}(0),
\end{array}
\end{eqnarray}
and zero for the remained ones and $G^{A}(t), G^{B}(t)$ and $G^{C}(t)$ are similar to the case of Eq. (16). The parameter $\lambda_{A}$ ($\lambda_{B}$ and $\lambda_{C}$) is the spectral width of the coupling of $N_{A}$ ($N_{B}$ and $N_{C}$) qubits to its respective reservoir. Also, we assume that the coupling constants of the subsystems to their respective reservoirs are equal $\gamma_{A}=\gamma_{B}=\gamma_{C}=\gamma_{0}$.
To assess to the degree of tripartite entanglement analytically, the so-called lower bound of concurrence ($LBC$) is used in this way. Any
separable states have a vanishing $LBC$ but its inverse is not true. However, a positive
$LBC$ indicates the detection of entanglement with certainty yet. Thus, using of $LBC$ for evaluating the entanglement dynamics is acceptable. According to Ref. $\cite{Fei}$, the $LBC$ for a three-qubit state $\rho$ is defined as
\begin{eqnarray}
LBC(\rho)=\sqrt{\frac{1}{3}\sum_{r=1}^{6} \Big\{\big[C_{r}^{(12|3)}(\rho)\big]^2+\big[C_{r}^{(23|1)}(\rho)\big]^2+\big[C_{r}^{(31|2)}(\rho)\big]^2\Big\}},
\end{eqnarray}
where
\begin{eqnarray}
C_{r}^{(uv|w)}(\rho)=\mathrm{max}\Big\{0,\sqrt{\lambda_{r, s}^{(uv|w)}}-\sum_{t>s} \sqrt{\lambda_{r, t}^{(uv|w)}}\Big\}.
\end{eqnarray}
In Eq. (29), $\lambda_{r, t}^{(uv|w)}$ are the eigenvalues of the density matrix $\rho(L_{r}^{uv}\otimes \sigma_{y}^{w})\rho^*(L_{r}^{uv}\otimes \sigma_{y}^{w})$ in decreasing order where $L_{r}^{uv}(r=1, 2, ..., 6)$ are the six generators of the $SO(4)$ group acting on the qubits $u$ and $v$, and $\sigma_{y}^{w}$ is the y-component Pauli matrix acting on the qubit $w$.
For the reduced three-qubit density matrix of our concern, $LBC$ can be obtained as
\begin{eqnarray}
\begin{array}{c}
LBC(\rho_{j, l, m})=\sqrt{\frac{8}{3}}\big(|G^{A}(t) G^{B}(t)|^2|C_{j}^{A}(0) C_{l}^{B}(0)|^2+|G^{A}(t) G^{C}(t)|^2|C_{j}^{A}(0) C_{m}^{C}(0)|^2+ \\\\
|G^{B}(t) G^{C}(t)|^2|C_{l}^{B}(0) C_{m}^{C}(0)|^2\big)^{\frac{1}{2}}.
\end{array}
\end{eqnarray}
Fig. 6, shows the $LBC$ for the three-qubit in Markovian and non-Markovian regimes.
Notice that similar to the previous section, for $N_{A}=N_{B}=N_{C}=1$ (in the absence of additional qubits in the subsystems A, B and C), the $LBC$ eventually
decays to zero but there is a steady value for $LBC$ retained by increasing $N_{A}$, $N_{B}$ and $N_{C}$ in both Markovian and non-Markovian regimes.
In the limit $t\rightarrow \infty$, the $LBC$ reduces to
\begin{eqnarray}
\begin{array}{c}
LBC(\rho_{j, l, m})=\sqrt{\frac{8}{3}}\big(\frac{(N_{A}-1)^{2}(N_{B}-1)^{2}}{N_{A}^{2}N_{B}^{2}}|C_{j}^{A}(0) C_{l}^{B}(0)|^2+
\frac{(N_{A}-1)^{2}(N_{C}-1)^{2}}{N_{A}^{2}N_{C}^{2}}|C_{j}^{A}(0) C_{m}^{C}(0)|^2+\\\\
\frac{(N_{B}-1)^{2}(N_{C}-1)^{2}}{N_{B}^{2}N_{C}^{2}}|C_{l}^{B}(0) C_{m}^{C}(0)|^2\big)^{\frac{1}{2}}.
\end{array}
\end{eqnarray}
Consider, for example, $N_{C}=1$ then the $LBC$ in (31) becomes
\begin{eqnarray}
LBC(\rho_{j, l, m})=\sqrt{\frac{8}{3}}\frac{(N_{A}-1)(N_{B}-1)}{N_{A}N_{B}}|C_{j}^{A}(0) C_{l}^{B}(0)|,
\end{eqnarray}
and in the same way, if $N_{B}=1$ and $N_{C}=1$, then $LBC$ is equal to zero.
Ultimately, in the limits $N_{A},N_{B},N_{C}\rightarrow \infty$, the $LBC$ reaches
\begin{eqnarray}
LBC(\rho_{j, l, m})=\sqrt{\frac{8}{3}}\big(|C_{j}^{A}(0) C_{l}^{B}(0)|^2+|C_{j}^{A}(0) C_{m}^{C}(0)|^2+|C_{l}^{B}(0) C_{m}^{C}(0)|^2\big)^{\frac{1}{2}},
\end{eqnarray}
and also for the $LBC$ in (32), as $N_{A}, N_{B}\rightarrow \infty$, we have
\begin{eqnarray}
LBC(\rho_{j, l, m})=\sqrt{\frac{8}{3}}|C_{j}^{A}(0) C_{l}^{B}(0)|.
\end{eqnarray}
The $LBC$ evolution in (30) can be evaluated under different non-Markovian behaviors of the related subsystems as shown
in Fig. 7 with $N_{A}=N_{B}=N_{C}=6$. Obviously, at long time limit, it is interesting to note that, apart from
different non-Markovian behaviors of the subsystems, the asymptotic $LBC$s are completely coincided to each others again.
\section{Conclusion}
We investigated the preservation of quantum coherence of a single-qubit interacting with a zero-temperature thermal reservoir through the addition of non-interacting qubits into the reservoir.
Also, we discussed the extension of this scheme for entanglement protection of two
and three distant non-interacting qubits, each of which individually has been contained in an independent reservoir.
At limit $t\rightarrow\infty$, explicit dependence of the coherence measure, bipartite and tripartite concurrences on the number of additional qubits were derived.
It was pointed out that, by increasing the number of additional qubits in each
reservoir, the initial coherence and the respective entanglements are completely protected in both
Markovian and non-Markovian regimes.
Interestingly it was shown that, for preserving of initial state entanglement, the dynamics of each subsystem is not necessarily similar to dynamics of the other subsystem (subsystems) from Markovian and non-Markovian point of views. On the other hand, from the experimental point of view, the proposed scheme can be realized using lossy (imperfect) cavities. As illustrated, for preserving the entanglement, it is not important for the cavities to have the same spectral density. It should be noted that the scheme can be extended for protecting higher order multipartite entanglement of qubits distantly contained in the Lorentzian reservoirs such as cavities. Finally, the proposed scheme of this paper can be extended for investigating the possibility of protection of coherence and entanglement against the temperature of the reservoir through the additional qubits which can be regarded as the subject of future research.
\newpage
\vspace{1cm} \setcounter{section}{0}
\setcounter{equation}{0}
\renewcommand{\theequation}{A.\arabic{equation}}
{\Large{Appendix A:}}\\
\textbf{Details of derivation of Eq. (13)}:
Taking the Laplace transform from both sides of Eq. (10) gives the following set of equations
\begin{eqnarray}
p C_{j}(p)-C_{j}(0)=-\beta_{j} \mathcal{L}\{f(t)\} \sum_{l=1}^{N} \beta_{l} C_{l}(p),
\end{eqnarray}
where $j=1,2,...,N$. Here we use the notation $F(p)=\mathcal{L} \{F(t)\}=\int_{0}^{\infty} F(t) e^{-pt} dt$.
By dividing Eq. (A.1) to $\beta_{j}$, we observe that the right hand sides of $N$ equations are equal so the following relation between the coefficients is obtained
\begin{eqnarray}
\frac{p C_{1}(p)-C_{1}(0)}{\beta_{1}}=\frac{p C_{2}(p)-C_{2}(0)}{\beta_{2}}=...=\frac{p C_{j}(p)-C_{j}(0)}{\beta_{j}}=...=\frac{p C_{N}(p)-C_{N}(0)}{\beta_{N}}.
\end{eqnarray}
By writing the coefficients $C_{l}(p)$ ($l\neq j$) in terms of $C_{j}(p)$ and inserting them into the Eq. (A.1), we have
\begin{eqnarray}
p C_{j}(p)-C_{j}(0)=-\mathcal{L}\{f(t)\} \bigg{(}\beta_{j}^2 C_{j}(p)+\frac{1}{p}\sum_{l\neq j}^{N} \Big{[} \beta_{l}^2(pC_{j}(p)-C_{j}(0))+\beta_{j}\beta_{l}C_{l}(0) \Big{]} \bigg{)},
\end{eqnarray}
For a Lorentzian spectral density given by Eq. (12), the correlation function $f(t)$ can be calculated by using Eq. (11), therefore we have
\begin{eqnarray}
f(t)=\frac{\gamma_{0} \lambda}{2}e^{-(\lambda-i\Delta)t},
\end{eqnarray}
and its Laplace transform is written as
\begin{eqnarray}
\mathcal{L}\{f(t)\}=\frac{\gamma_{0} \lambda}{2(p+\lambda -i\Delta)},
\end{eqnarray}
After substituting Eq. (A.5) into Eq. (A.3), the coefficients $C_{j}(p)$ can be obtained as
\begin{eqnarray}
C_{j}(p)=\frac{2(p+\lambda -i\Delta)}{2(p+\lambda -i\Delta)+ \sum_{l=1}^{N} \beta_{l}^2}C_{j}(0)+\frac{\gamma_{0} \lambda \big{(}\sum_{l\neq j}^{N} \big{[} \beta_{l}^2 C_{j}(0)-\beta_{j}\beta_{l} C_{j}(0)\big{]} \big{)}}{p \Big{(} 2(p+\lambda -i\Delta)+ \sum_{l=1}^{N} \beta_{l}^2\Big{)}},
\end{eqnarray}
Finally, the inverse Laplace transform of $C_{j}(p)$ gives Eq. (13).
\vspace{1cm} \setcounter{section}{0}
\setcounter{equation}{0}
\renewcommand{\theequation}{B.\arabic{equation}}
{\Large{Appendix B:}}\\
\textbf{Dynamics of two and three independent qubits}:
As discussed in Ref. $\cite{Bellomo}$, a single-qubit dynamics has the form
\begin{eqnarray}
\rho_{mm'}(t)=\sum_{nn'} \chi^{nn'}_{mm'}(t) \rho_{nn'}(0),
\end{eqnarray}
where $\rho_{mm'}(t)=\langle m| \rho(t) |m'\rangle$ with $m,n=e,g$.
The matrix whose entries are the values $\chi^{nn'}_{mm'}(t)$ is said to form a matrix representation of $\chi(t)$. By using Eq. (17),
the matrix representation of $\chi(t)$ in the standard computational basis $\{|e,e\rangle, |e,g\rangle, |g,e\rangle, |g,g\rangle\}$ can be obtained as
\begin{eqnarray}
\chi=\left(
\begin{array}{cccc}
|G(t)|^2 & 0 & 0 & 0 \\
0 & G(t) & 0 & 0 \\
0 & 0 & G(t)^{*} & 0 \\
1-|G(t)|^2 & 0 & 0 & 1
\end{array}
\right),
\end{eqnarray}
where $\chi^{nn'}_{mm'}(t)=\langle m,m'| \chi(t) |n,n'\rangle$.
We now consider a system consisting of two independent qubits, each locally interacting with its own reservoir.
The complete dynamics of the two-qubit system can be obtained by knowing the single-qubit dynamics which has been obtained in Eq. (17).
Under these conditions, given the dynamics of each qubit as $\rho^{A}_{m_{1}m'_{1}}(t)=\sum_{n_{1}n'_{1}} A^{n_{1}n'_{1}}_{m_{1}m'_{1}}(t) \rho_{n_{1}n'_{1}}(0)$ and
$\rho^{B}_{m_{2}m'_{2}}(t)=\sum_{n_{2}n'_{2}} B^{n_{2}n'_{2}}_{m_{2}m'_{2}}(t) \rho_{n_{2}n'_{2}}(0)$,
the dynamics of the two-qubit system is expressed by
\begin{eqnarray}
\begin{array}{c}
\rho^{AB}_{m_{1}m'_{1},m_{2}m'_{2}}(t)=\sum_{n_{1}n'_{1}} \sum_{n_{2}n'_{2}} A^{n_{1}n'_{1}}_{m_{1}m'_{1}}(t) B^{n_{2}n'_{2}}_{m_{2}m'_{2}}(t) \rho^{AB}_{n_{1}n'_{1},n_{2}n'_{2}}(0)=\\\\
\sum_{n_{1}n'_{1}} \sum_{n_{2}n'_{2}} E^{n_{1}n'_{1},n_{2}n'_{2}}_{m_{1}m'_{1},m_{2}m'_{2}}(t) \rho^{AB}_{n_{1}n'_{1},n_{2}n'_{2}}(0),
\end{array}
\end{eqnarray}
where $\rho^{AB}_{m_{1}m'_{1},m_{2}m'_{2}}(t)=\langle m_{1},m_{2}| \rho^{AB}(t) |m'_{1},m'_{2}\rangle$ and $E^{n_{1}n'_{1},n_{2}n'_{2}}_{m_{1}m'_{1},m_{2}m'_{2}}(t)=
\langle m_{1}m'_{1},m_{2}m'_{2}| A\otimes B |n_{1}n'_{1},n_{2}n'_{2}\rangle$ with $m_{1},n_{1},m_{2},n_{2}=e,g$. Since the qubits are in general in different environments so that their evolution is characterized by the different functions $A(t)$ and $B(t)$ as
\begin{eqnarray}
\begin{array}{c}
A(t)=\left(
\begin{array}{cccc}
|G^{A}(t)|^2 & 0 & 0 & 0 \\
0 & G^{A}(t) & 0 & 0 \\
0 & 0 & {G^{A}}^{*}(t) & 0 \\
1-|G^{A}(t)|^2 & 0 & 0 & 1
\end{array}
\right),\\\\
B(t)=\left(
\begin{array}{cccc}
|G^{B}(t)|^2 & 0 & 0 & 0 \\
0 & G^{B}(t) & 0 & 0 \\
0 & 0 & {G^{B}}^{*}(t) & 0 \\
1-|G^{B}(t)|^2 & 0 & 0 & 1
\end{array}
\right),
\end{array}
\end{eqnarray}
where $G^{A}(t)$ and $G^{B}(t)$ can be considered as
\begin{eqnarray}
G^{j}(t)=\frac{N_{j}-1}{N_{j}}+\frac{e^{-\Lambda_{j} t/2}}{N_{j}} \big(cosh{(\frac{D_{j}t}{2})}+\frac{\Lambda_{j}}{D_{j}} sinh{(\frac{D_{j}t}{2})}\big),
\end{eqnarray}
and $D_{j}=\sqrt{\Lambda_{j}^{2}-2\gamma_{j} \lambda_{j} N_{j}}$ with $j=A, B$.
In the following, we extend this procedure to explore the dynamics of three independent qubits, each locally interacting with its own reservoir.
By considering the dynamics of the third qubit as $\rho^{C}_{m_{3}m'_{3}}(t)=\sum_{n_{3}n'_{3}} C^{n_{3}n'_{3}}_{m_{3}m'_{3}}(t) \rho_{n_{1}n'_{1}}(0)$, the dynamics of the three-qubit system is simply given by
\begin{eqnarray}
\begin{array}{c}
\rho^{ABC}_{m_{1}m'_{1},m_{2}m'_{2},m_{3}m'_{3}}(t)=\sum_{n_{1}n'_{1}} \sum_{n_{2}n'_{2}} \sum_{n_{3}n'_{3}} A^{n_{1}n'_{1}}_{m_{1}m'_{1}}(t) B^{n_{2}n'_{2}}_{m_{2}m'_{2}}(t) C^{n_{3}n'_{3}}_{m_{3}m'_{3}}(t) \rho^{ABC}_{n_{1}n'_{1},n_{2}n'_{2},n_{3}n'_{3}}(0)=\\\\
\sum_{n_{1}n'_{1}} \sum_{n_{2}n'_{2}} \sum_{n_{3}n'_{3}} F^{n_{1}n'_{1},n_{2}n'_{2},n_{3}n'_{3}}_{m_{1}m'_{1},m_{2}m'_{2},m_{3}m'_{3}}(t) \rho^{ABC}_{n_{1}n'_{1},n_{2}n'_{2},n_{3}n'_{3}}(0),
\end{array}
\end{eqnarray}
where $\rho^{ABC}_{m_{1}m'_{1},m_{2}m'_{2},m_{3}m'_{3}}(t)=\langle m_{1},m_{2},m_{3}| \rho^{ABC}(t) |m'_{1},m'_{2},m'_{3}\rangle$ and also $F^{n_{1}n'_{1},n_{2}n'_{2},n_{3}n'_{3}}_{m_{1}m'_{1},m_{2}m'_{2},m_{3}m'_{3}}(t)=
\langle m_{1}m'_{1},m_{2}m'_{2},m_{3}m'_{3}| A \otimes B \otimes C |n_{1}n'_{1},n_{2}n'_{2},n_{3}n'_{3}\rangle$ with $m_{1},n_{1},m_{2},n_{2},m_{3},n_{3}=e,g$.
By knowing the matrix representations of $A(t)$ and $B(t)$ in Eq. (B.4), the function $C(t)$ is obtained as
\begin{eqnarray}
\begin{array}{c}
C(t)=\left(
\begin{array}{cccc}
|G^{C}(t)|^2 & 0 & 0 & 0 \\
0 & G^{C}(t) & 0 & 0 \\
0 & 0 & {G^{C}}^{*}(t) & 0 \\
1-|G^{C}(t)|^2 & 0 & 0 & 1
\end{array}
\right),
\end{array}
\end{eqnarray}
where $G^{A}(t)$, $G^{B}(t)$ and $G^{C}(t)$ are as
\begin{eqnarray}
G^{j}(t)=\frac{N_{j}-1}{N_{j}}+\frac{e^{-\Lambda_{j} t/2}}{N_{j}} \big(cosh{(\frac{D_{j}t}{2})}+\frac{\Lambda_{j}}{D_{j}} sinh{(\frac{D_{j}t}{2})}\big),
\end{eqnarray}
and $D_{j}=\sqrt{\Lambda_{j}^{2}-2\gamma_{j} \lambda_{j} N_{j}}$ with $j=A, B, C$.
\newpage
|
\section{Introduction}
Counting walks in a graph is a basic and interesting problem.
Let $G=(V,E)$ be a locally finite graph with adjacency matrix $A$.
Then the matrix element $(A^m)_{xy}$ counts the number of $m$-step walks
connecting $x$ and $y$.
For $x=y=o\in V$ this number is expressible in the integral:
\[
(A^m)_{oo}=\int_{\mathbb{R}} x^m \mu(dx),
\qquad m=0,1,2,\dots,
\]
where $\mu$ is a probability distribution on $\mathbb{R}=(-\infty,+\infty)$,
called the \textit{spectral distribution} of $A$ at a vertex $o$.
Thus the number of walks may be studied from
an analytic or probabilistic point of view.
During the last fifteen years the quantum probability has been employed
for the asymptotic analysis of graph spectra
as well as the study of product structures in connection with several
notions of independence,
see e.g., a monograph \cite{Hora-Obata}.
This paper focuses on
the notion of \textit{Kronecker product} of graphs $G_1\times_K G_2$,
which is also called the \textit{direct product}
and is one of the most basic graph products,
see the comprehensive monographs
\cite{Hammack-Imrich-Klavzar2011}, \cite{Imrich-Klavzar2000}.
It is well known that the spectral distribution of
the Cartesian product of two graphs $G_1\times_C G_2$
is obtained by the usual convolution of probability distributions
defined by
\[
\int_{\mathbb{R}}h(x)\mu_1*\mu_2(dx)
=\int_{\mathbb{R}}\int_{\mathbb{R}}h(x+y)\mu_1(dx)\mu_2(dy),
\qquad h\in C_{\mathrm{bdd}}(\mathbb{R}).
\]
The convolution $\mu_1*\mu_2$ is known to be the distribution of
the sum of two independent random variables $X_1+X_2$.
Quantum probability allows us to discuss variations of
independence of non-commutative variables.
The comb product of graphs is related to
the monotone convolution,
the star product to the Boolean convolution,
and the free product to the free convolution,
see e.g., \cite{Hora-Obata}, for further relevant results see
\cite{Accardi-Lenczewski-Salapata2007}.
The \textit{Mellin convolution}
in the original sense is the convolution product
on the locally compact abelian group $(\mathbb{R}_{>0},\cdot)$
with the Haar measure $dx/x$ defined by
\begin{equation}\label{eq-Mellin-Convolution}
f\star g(x)
=\int_0^\infty f(y)g\Big(\frac{x}{y}\Big) \frac{dy}{y}
=\int_0^\infty f\Big(\frac{x}{y}\Big)g(y) \frac{dy}{y}
\end{equation}
for $f,g \in L^1((0,\infty),dx/x)$,
see e.g., \cite{Marichev}.
Extending the above definition naturally to
symmetric probability distributions on $\mathbb{R}$,
we define the \textit{Mellin convolution} $\mu_1*_M\mu_2$
of two symmetric distributions $\mu_1$ and $\mu_2$ by
\[
\int_{\mathbb{R}}h(x)\mu_1*_M\mu_2(dx)
=\int_{\mathbb{R}}\int_{\mathbb{R}}h(xy)\mu_1(dx)\mu_2(dy),
\qquad h\in C_{\mathrm{bdd}}(\mathbb{R}).
\]
Recall that a measure $\mu$ on $\mathbb{R}$ is called symmetric
if $\mu(-dx)=\mu(dx)$.
The Kronecker product of graphs becomes
a new member of the corresponding list of ``product structures"
of graphs and ``convolution products"
of probability distributions on $\mathbb{R}$.
This paper is organized as follows.
In Section \ref{Sec:Counting walks in a graph} we assemble
basic notations and notions
for counting walks in terms of the spectral distribution.
In Section \ref{Sec:Kronecker product of graphs} we introduce
the concept of Kronecker product of graphs and
show some elementary properties with illustrations.
The main result is stated in Theorem \ref{main theorem}.
In Section \ref{Sec:Subgraphs of 2-dimensional lattice as Kronecker products}
two-dimensional integer lattices restricted to certain domains
which admit the Kronecker product structure.
We derive formulas for counting walks and show that
the density functions of the spectral distributions
are expressible in terms of elliptic integrals.
Finally in Section \ref{sec:Examples in higher dimension} we
discuss towards higher dimensional extension,
where we find unexpectedly that
the restricted integer lattice $\{x\ge y\ge z\}$
and the mixed product $(\mathbb{Z}_+\times_K \mathbb{Z}_+)\times_C \mathbb{Z}_+$
are not isomorphic but have a common spectral distribution
at the origin $(0,0,0)$.
\section{Counting walks in a graph}
\label{Sec:Counting walks in a graph}
A graph $G=(V,E)$ is a pair,
where $V$ is a non-empty set and $E$ a subset of two-point subsets of $V$,
i.e., $E\subset\{\{x,y\}\,;\, x,y\in V, x\neq y\}$.
We deal with both finite and infinite graphs.
If $\{x,y\}\in E$, we say that $x$ and $y$ are adjacent and
write $x\sim y$.
The degree of $x\in V$ is defined to be the number of
vertices that are adjacent to $x$, and is denoted by $\deg x=\deg_G x$.
A graph under consideration in this paper is always assumed to be
locally finite, i.e., $\deg x<\infty$ for all vertices $x\in V$.
For $m=1,2,\dots$ an $m$-step walk from a vertex $x\in V$ to another $y\in V$
is an (ordered) sequence of vertices $x_0,x_1,\dots,x_m$ such that
\[
x=x_0\sim x_1\sim x_2\sim \dotsb\sim x_{m-1}\sim x_m=y.
\]
The number of such walks is interesting to study.
The adjacency matrix of a graph $G=(V,E)$ is a matrix
$A$ indexed by $V\times V$ whose entries are defined by
\[
(A)_{xy}=
\begin{cases}
1, &\text{if $x\sim y$}; \\
0, &\text{otherwise}.
\end{cases}
\]
By local finiteness the powers of $A$ are well-defined and
the matrix entry $(A^m)_{xy}$ counts the number of
$m$-step walks connecting $x$ and $y$.
It is convenient to introduce the Hilbert space $\ell^2(V)$ of
$\mathbb{C}$-valued square-summable functions on $V$ with the inner product
\[
\langle f, g\rangle=\sum_{x\in V}\overline{f(x)}\,g(x),
\qquad
f,g\in \ell^2(V).
\]
Let $\{\delta_x\,;\,x\in V\}$ be the canonical orthonormal basis of
$\ell^2(V)$.
Then we have
\[
(A^m)_{xy}=\langle \delta_x,A^m\delta_y\rangle,
\quad m=0,1,2,\dots.
\]
We are particularly interested in counting the number of walks
from a vertex $o\in V$ to itself, which is denoted by
\[
W_m(o;G)=(A^m)_{oo}\,,
\qquad m=0,1,2,\dots.
\]
We tacitly understand that $W_0(o;G)=1$.
\begin{theorem}\label{2thm:spectral distribution}
Let $G=(V,E)$ be a graph with a distinguished vertex $o\in V$.
Then there exists a probability distribution $\mu$ on $\mathbb{R}$ such that
\[
W_m(o;G)
=(A^m)_{oo}
=\langle \delta_o,A^m\delta_o\rangle
=M_m(\mu),
\qquad m=0,1,2,\dots,
\]
where
\[
M_m(\mu)=\int_{\mathbb{R}} x^m \mu(dx)
\]
is the $m$-th moment of $\mu$.
\end{theorem}
The proof is by the Hamburger theorem, see e.g., \cite{Hora-Obata}.
The probability distribution in Theorem \ref{2thm:spectral distribution}
is called the \textit{spectral distribution} of $A$ in the vector state
at $o\in V$.
The spectral distribution is not uniquely determined in general
due to the indeterminate moment problem,
however, it is unique if the degrees of vertices are uniformly bounded,
i.e., if $\sup\{\deg (x)\,;\,x\in V\}<\infty$.
If $W_{2m+1}(o;G)=(A^{2m+1})_{oo}=0$ for all $m=0,1,2,\dots$,
the spectral distribution may be assumed to be symmetric.
\section{Kronecker product of graphs}
\label{Sec:Kronecker product of graphs}
\subsection{Definition and elementary properties}
Let $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$ be two
(finite or infinite) graphs with adjacency matrices $A^{(1)}$ and $A^{(2)}$,
respectively.
Let $V=V_1\times V_2$ be the Cartesian product set and
define a matrix $A$ indexed by $V\times V$ by
\[
(A)_{(x,y),(x^\prime,y^\prime)}
=A^{(1)}_{xx^\prime} A^{(2)}_{yy^\prime},
\qquad
(x,y),(x^\prime,y^\prime)\in V.
\]
Since $A$ is a symmetric matrix whose
diagonal entries are all zero
and off-diagonal ones take values in $\{0,1\}$,
there exists a graph $G$ on $V=V_1\times V_2$ whose adjacency matrix is $A$,
or equivalently, whose edge set is given by
\[
E=\{\{(x,y),(x^\prime,y^\prime)\}\,;\,
(A)_{(x,y),(x^\prime,y^\prime)}=1\}.
\]
The above graph $G$ is called the \textit{Kronecker product} of $G_1$ and $G_2$,
and is denoted by
\[
G=G_1\times_K G_2\,.
\]
In other words,
the Kronecker product of $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$
is a graph on $V_1\times V_2$ with adjacency relation
$(x,y)\sim_K(x^\prime,y^\prime)\Longleftrightarrow x\sim x^\prime$
and $y\sim y^\prime$.
\begin{remark}
The term \textit{Kronecker product} appears in
\cite{Brouwer-Haemers2010} for instance,
while there are many synonyms.
The \textit{direct product} is another common term
used in \cite{Godsil1993}, \cite{Hammack-Imrich-Klavzar2011},
\cite{Imrich-Klavzar2000} and so forth.
In this paper we prefer to the former
in order to avoid confusion with some terms in quantum probability.
\end{remark}
Through the canonical unitary isomorphism
$\ell^2(V_1\times V_2)\cong \ell^2(V_1)\otimes \ell^2(V_2)$
given by $\delta_{(x,y)}\leftrightarrow \delta_x\otimes\delta_y$,
the adjacency matrix $A$ of $G_1\times_K G_2$ is written as
\begin{equation}\label{3eqn:def od A for Kronecker product}
A=A^{(1)}\otimes A^{(2)}.
\end{equation}
In fact, by definition we have
\[
(A)_{(x,y),(x^\prime,y^\prime)}
=\langle\delta_{(x,y)}, A\delta_{(x^\prime,y^\prime)}\rangle
=\langle\delta_x\otimes \delta_y, A(\delta_{x^\prime}\otimes\delta_{y^\prime})\rangle
\]
and
\[
A^{(1)}_{xx^\prime} A^{(2)}_{yy^\prime}
=\langle\delta_x, A^{(1)}\delta_{x^\prime}\rangle
\langle\delta_y, A^{(2)}\delta_{y^\prime}\rangle
=\langle\delta_x\otimes \delta_y,
(A^{(1)}\otimes A^{(2)})(\delta_{x^\prime}\otimes\delta_{y^\prime})\rangle,
\]
from which \eqref{3eqn:def od A for Kronecker product} follows.
We collect some elementary properties, of which the proofs are
straightforward.
For further relevant results, see
the comprehensive monographs \cite{Hammack-Imrich-Klavzar2011},
\cite{Imrich-Klavzar2000}.
\begin{proposition}
For any graphs $G_1,G_2,G_3$ we have
\begin{gather*}
G_1\times_K G_2 \cong G_2\times_K G_1\,,
\\
(G_1\times_K G_2)\times_K G_3
\cong
G_1\times_K (G_2\times_K G_3).
\end{gather*}
\end{proposition}
\begin{proposition}
Let $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$ be two connected graphs
with $|V_1|\ge2$ and $|V_2|\ge2$.
Then the Kronecker product $G_1\times_K G_2$ has at most two
connected components.
\end{proposition}
\begin{proposition}
Let $P_1$ be the graph consisting of a single vertex.
Then for any graph $G=(V,E)$
the Kronecker product $P_1\times_K G$ is a graph on $V$ with no edges,
i.e., an empty graph on $V$.
\end{proposition}
The \textit{Cartesian product}
of two graphs $G_1$ and $G_2$,
denoted by $G_1\times_C G_2$,
is a graph on $V_1\times V_2$ with adjacency matrix defined by
\[
(A)_{(x,y),(x^\prime,y^\prime)}
=A^{(1)}_{xx^\prime}\delta_{yy^\prime}
+\delta_{xx^\prime}A^{(2)}_{yy^\prime},
\]
or equivalently under the isomorphism
$\ell^2(V_1\times V_2)\cong \ell^2(V_1)\otimes \ell^2(V_2)$,
\[
A=A^{(1)}\otimes I^{(2)}+I^{(1)}\otimes A^{(2)},
\]
where $I^{(i)}$ is the identity matrix indexed by
$V_i\times V_i$ for $i=1,2$.
The distance-2 graph of $G_1\times_C G_2$ is a graph on $V_1\times V_2$
with adjacency relation:
\begin{align}
(x,y)\sim(x^\prime,y^\prime)
&\Longleftrightarrow
\mathrm{dis}_{G_1\times_C G_2}((x,y),(x^\prime,y^\prime))=2
\label{3eqn:in proof 3.4}\\
&\Longleftrightarrow
\mathrm{dis}_{G_1}(x,x^\prime)+\mathrm{dis}_{G_2}(y,y^\prime)=2.
\nonumber
\end{align}
It is then easy to see that
the Kronecker product $G_1\times_K G_2$ is a subgraph of
the distance-2 graph of $G_1\times_C G_2$.
However, $G_1\times_K G_2$ is not necessarily
an induced subgraph of the distance-2 graph of $G_1\times_C G_2$.
\subsection{Counting walks}
The Kronecker product of graphs has a significant property
from the viewpoint of counting walks.
\begin{theorem}\label{thm:number of walks of Kronecker product}
Let $G_1\times_K G_2$ be the Kronecker product of
two graphs $G_1=(V_1,E_1)$ and $G_2=(V_2,E_2)$.
For $o_1\in V_1$ and $o_2\in V_2$ we have
\[
W_m((o_1,o_2);G_1\times_K G_2)=W_m(o_1;G_1)W_m(o_2;G_2),
\qquad m=0,1,2,\dots.
\]
\end{theorem}
\begin{proof}
Let $A^{(1)}$ and $A^{(2)}$ denote the adjacency matrices of $G_1$ and $G_2$,
respectively.
Let $A$ be the adjacency matrix of the Kronecker product $G_1\times_K G_2$.
Using the natural isomorphism $\ell^2(V_1\times V_2)\cong
\ell^2(V_1)\otimes\ell^2(V_2)$ and
$A=A^{(1)}\otimes A^{(2)}$ as in \eqref{3eqn:def od A for Kronecker product}
we calculate as follows:
\begin{align*}
W_m((o_1,o_2);G_1\times_K G_2)
&=\langle \delta_{(o_1,o_2)},A^m\delta_{(o_1,o_2)}\rangle \\
&=\langle \delta_{o_1}\otimes\delta_{o_2},
(A^{(1)}\otimes A^{(2)})^m\delta_{o_1}\otimes\delta_{o_2}\rangle \\
&=\langle \delta_{o_1}, (A^{(1)})^m\delta_{o_1}\rangle
\langle \delta_{o_2}, (A^{(2)})^m\delta_{o_2}\rangle \\
&=W_m(o_1;G_1)W_m(o_2;G_2),
\end{align*}
which completes the proof.
\end{proof}
\subsection{Mellin convolution of
symmetric probability distribution on $\mathbb{R}$}
We focus on symmetric probability distributions $\mu$ on $\mathbb{R}$
having finite moments of all orders.
Since $M_{2m+1}(\mu)=0$ holds for all $m=0,1,2,\dots$,
we are mostly interested in the even moments.
For such probability distributions $\mu$ and $\nu$,
there exists a probability distribution, denoted by $\mu*_M\nu$,
uniquely specified by
\[
\int_{\mathbb{R}}h(x)\mu*_M\nu(dx)
=\int_{\mathbb{R}}\int_{\mathbb{R}}h(xy)\mu(dx)\nu(dy),
\qquad h\in C_{\mathrm{bdd}}(\mathbb{R}).
\]
We call $\mu*_M\nu$ the \textit{Mellin convolution}.
It is easily seen that $\mu*_M\nu$ is symmetric
and has finite moments of all orders.
In fact,
\begin{proposition}\label{prop:moments Mellin convolution}
$M_m(\mu*_M \nu)=M_m(\mu)M_m(\nu)$ for all $m=0,1,2,\dots$.
\end{proposition}
Combining Theorems
\ref{2thm:spectral distribution},
\ref{thm:number of walks of Kronecker product}
and Proposition \ref{prop:moments Mellin convolution},
we come to the following fundamental result.
\begin{theorem}\label{main theorem}
For $i=1,2$ let $G_i=(V_i,E_i)$ be a graph with a distinguished vertex $o_i$.
Let $\mu_i$ be the spectral distribution of the adjacency matrix $A^{(i)}$ of
$G_i$ in the vector state at $o_i$.
Assume that $\mu_i$ is symmetric,
or equivalently that $W_{2m+1}(G_i, o_i)=0$
for all $m=0,1,2,\dots$ and $i=1,2$.
Then we have
\[
W_m((o_1,o_2);G_1\times_K G_2)
=M_m(\mu_1*_M\mu_2),
\qquad m=0,1,2,\dots.
\]
In other words, the spectral distribution of
the Kronecker product $G_1\times_K G_2$ in the vector state at
$(o_1,o_2)$ is the Mellin convolution of $\mu_1$ and $\mu_2$.
\end{theorem}
The Mellin convolution is originally introduced
on the basis of the locally compact abelian group $\mathbb{R}_{>0}=(0,\infty)$,
see Introduction. In this connection we should note the following
\begin{proposition}\label{Prop:density of Mellin convolution}
Let $f(x)$ and $g(x)$ be symmetric density functions on $\mathbb{R}$
and consider the probability distributions
$\mu(dx)=f(x)dx$ and $\nu(dx)=g(x)dx$.
Then $\mu*_M\nu$ admits a symmetric density function $2f\star g(x)$,
where $f\star g$ is the (original) Mellin convolution
defined in \eqref{eq-Mellin-Convolution}.
\end{proposition}
\begin{proof}
By definition, for a symmetric function
$h\in C_{\mathrm{bdd}}(\mathbb{R})$ we have
\begin{align*}
\int_{\mathbb{R}}h(x)\mu*_M\nu(dx)
&=\int_{\mathbb{R}}\int_{\mathbb{R}}h(xy)\mu(dx)\nu(dy) \\
&=4\int_0^{\infty}\int_0^{\infty}h(xy)f(x)g(y)dxdy \\
&=4\int_0^{\infty} g(y)dy \int_0^{\infty}h(x)f\Big(\frac{x}{y}\Big)
\frac{dx}{y} \\
&=2\int_{\mathbb{R}} h(x) dx
\int_0^{\infty}f\Big(\frac{x}{y}\Big)g(y)\frac{dy}{y}\,.
\end{align*}
Hence, $2f\star g(x)$ is the density function of $\mu*_M\nu$.
\end{proof}
For the readers' convenience we make comparison with the Cartesian product.
The classical convolution of two probability distributions
$\mu$ and $\nu$ is
a probability distribution, denoted by $\mu*\nu$,
uniquely specified by
\[
\int_{\mathbb{R}}h(x)\mu*\nu(dx)
=\int_{\mathbb{R}}\int_{\mathbb{R}}h(x+y)\mu(dx)\nu(dy),
\qquad h\in C_{\mathrm{bdd}}(\mathbb{R}).
\]
By applying the binomial expansion we get the following.
\begin{proposition}\label{prop:M_m(G_1XCG_2)}
For $i=1,2$ let $G_i=(V_i,E_i)$ be a graph with a distinguished vertex $o_i$.
Let $\mu_i$ be the spectral distribution of the adjacency matrix $A^{(i)}$ of
$G_i$ in the vector state at $o_i$.
Then we have
\begin{align*}
W_m((o_1,o_2);G_1\times_C G_2)
& =\sum_{k=0}^m \binom{m}{k} W_k(o_1;G_1)W_{m-k}(o_2;G_2)\\
& =M_m(\mu_1*\mu_2),
\qquad m=0,1,2,\dots,
\end{align*}
where $\mu_1*\mu_2$ is the (classical) convolution.
In other words, the spectral distribution of
the Cartesian product $G_1\times_C G_2$ in the vector state at
$(o_1,o_2)$ is the convolution of $\mu_1$ and $\mu_2$.
\end{proposition}
\section{Subgraphs of 2-dimensional lattice as Kronecker products}
\label{Sec:Subgraphs of 2-dimensional lattice as Kronecker products}
\subsection{The Kronecker product $\mathbb{Z}\times_K \mathbb{Z}$}\enspace
In order to avoid confusion we use the symbol $\mathbb{Z}^2$ just for the
Cartesian product set.
The Kronecker product $\mathbb{Z}\times_K \mathbb{Z}$
is by definition a graph on
$\mathbb{Z}^2=\{(u,v)\,;\, u,v\in \mathbb{Z}\}$ with adjacency relation:
\begin{equation}\label{eqn:Mellin adjacency (1)}
(u,v)\sim_K(u^\prime, v^\prime)
\quad\Longleftrightarrow\quad
u^\prime=u\pm 1
\quad\text{and}\quad
v^\prime=v\pm 1.
\end{equation}
While, the so-called 2-dimensional integer lattice
is a graph on $\mathbb{Z}^2$ with adjacency relation:
\[
(x,y)\sim(x^\prime, y^\prime)
\quad\Longleftrightarrow\quad
\begin{cases}
x^\prime=x\pm 1, \\
y^\prime=y,
\end{cases}
\text{or}\quad
\begin{cases}
x^\prime=x, \\
y^\prime=y\pm 1.
\end{cases}
\]
We see immediately from definition that
$\mathbb{Z}\times_K \mathbb{Z}$ has two connected components,
each of which is isomorphic to
the 2-dimensional integer lattice $\mathbb{Z}\times_C \mathbb{Z}$.
Denoting by $(\mathbb{Z}\times_K \mathbb{Z})^o$
the connected component of $\mathbb{Z}\times_K \mathbb{Z}$
containing $o=(0,0)$,
we claim the following
\begin{theorem}\label{thm:ZXMZ}
$(\mathbb{Z}\times_K \mathbb{Z})^o
\cong\mathbb{Z}\times_C \mathbb{Z}$,
where the isomorphism preserves the origin.
\end{theorem}
Here we prepare a general result.
\begin{proposition}\label{4prop:induced subgraph}
For $i=1,2$ let $G_i=(V_i,E_i)$ be a graph
and $H_i=(W_i,F_i)$ an induced subgraph of $G_i$.
Then $H_1\times_K H_2$ is an induced subgraph of $G_1\times_K G_2$.
\end{proposition}
\begin{proof}
By definition the vertex set of $H_1\times_K H_2$ is $W_1\times W_2$.
For two verices $(x,y), (x^\prime,y^\prime)\in W_1\times W_2$
we have $(x,y)\sim(x^\prime,y^\prime)$
in $H_1\times_K H_2$ if and only if
$x\sim x^\prime$ in $H_1$ and
$y\sim y^\prime$ in $H_2$ by definition.
Since $H_1$ and $H_2$ are respectively induced subgraphs of $G_1$ and $G_2$,
the last condition is equivalent to that
$x\sim x^\prime$ in $G_1$ and
$y\sim y^\prime$ in $G_2$,
hence to that
$(x,y)\sim (x^\prime,y^\prime)$ in $G_1\times_K G_2$.
Consequently,
$H_1\times_K H_2$ is an induced subgraph of
$G_1\times_K G_2$ spanned by $W_1\times W_2$.
\end{proof}
\subsection{Subgraphs of 2-dimensional integer lattice}
For a subset $D\subset \mathbb{Z}^2$ let
$L[D]$ denote the lattice restricted to $D$,
i.e., the induced subgraph of $\mathbb{Z}\times_C\mathbb{Z}$
spanned by the vertices in $D$.
We are particularly interested in
restricted lattices which admit Kronecker product structure.
Theorem \ref{thm:ZXMZ} says that
$\mathbb{Z}\times_C \mathbb{Z}=L[\mathbb{Z}^2]$ itself
is isomorphic to the Kronecker product $(\mathbb{Z}\times_K \mathbb{Z})^o$.
\begin{theorem}\label{thm:ZXMP5 and Z_+XMZ}
For $n\ge2$ we have
\[
L\{(x,y)\in\mathbb{Z}^2\,;\, x\ge y \ge x-(n-1)\}
\cong (P_n\times_K \mathbb{Z})^o,
\]
where the right-hand side stands for
the connected component of $P_n\times_K \mathbb{Z}$
containing $o=(0,0)$,
$P_n$ being the path on $\{0,1,\dots,n-1\}$.
Similarly,
\[
L\{(x,y)\in\mathbb{Z}^2\,;\, x\ge y \}
\cong (\mathbb{Z}_+\times_K \mathbb{Z})^o.
\]
\end{theorem}
\begin{proof}
The path $P_n$ is naturally
regarded as an induced subgraph of $\mathbb{Z}$
spanned by $\{0,1,2,n-1\}$.
It then follows from Proposition \ref{4prop:induced subgraph} that
$P_n\times_K\mathbb{Z}$ is an induced subgraph
of $\mathbb{Z}\times_K \mathbb{Z}$.
Therefore, $(P_n\times_K \mathbb{Z})^o$ is
an induced subgraph of $(\mathbb{Z}\times_K \mathbb{Z})^o$.
Then, in view of Figure \ref{fig:restricted lattices},
we see that $(P_n\times_K \mathbb{Z})^o$ is
isomorphic to the induced subgraph of $\mathbb{Z}\times_C \mathbb{Z}$
spanned by $D=\{(x,y)\,;\, x\ge y \ge x-(n-1)\}$.
The second assertion is proved similarly.
\end{proof}
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=150pt,keepaspectratio,clip]{07_Z_+XMZ.pdf}
\qquad
\includegraphics[width=150pt,keepaspectratio,clip]{07_Z_+XMZ_+.pdf}
\caption{$(\mathbb{Z}_+\times_K \mathbb{Z})^o\cong L(x\ge y)$
and $(\mathbb{Z}_+\times_K \mathbb{Z}_+)^o\cong L(-x\le y\le x)$}
\label{fig:restricted lattices}
\end{center}
\end{figure}
\begin{theorem}
\label{thm:PmXMPn}
For $k \ge2$ and $l\ge2$ we have
\[
L\left\{(x,y)\in\mathbb{Z}^2\,;\,
\begin{array}{l}
0\le x+y\le k-1, \\
0\le x-y\le l-1
\end{array}
\right\}
\cong (P_k\times_K P_l)^o.
\]
Moreover,
\[
L\{(x,y)\in\mathbb{Z}^2\,;\, x\ge y\ge -x\}
\cong (\mathbb{Z}_+\times_K \mathbb{Z}_+)^o.
\]
\end{theorem}
The proof is similar as above, see also Figure \ref{fig:restricted lattices}.
\subsection{Counting walks}
The number of walks on one-dimensional integer lattice $\mathbb{Z}$
from the origin $0$ to itself is well known.
We have
\begin{equation}\label{4eqn:walks in Z}
W_{2m}(0;\mathbb{Z})=\binom{2m}{m},
\quad
W_{2m+1}(0;\mathbb{Z})=0,
\quad m=0,1,2,\dots.
\end{equation}
A similar result for $\mathbb{Z}_+=\{0,1,2,\dots\}$ is also well known.
We have
\begin{equation}\label{4eqn:walks in Z+}
W_{2m}(0;\mathbb{Z}_+)=C_m=\frac{1}{m+1}\binom{2m}{m},
\quad
W_{2m+1}(0;\mathbb{Z}_+)=0,
\quad m=0,1,2,\dots,
\end{equation}
where $C_m$ is the renowned Catalan number.
We start with typical restricted lattices.
\begin{example}\label{thm:xgey}
(1) For $L=L\{(x,y)\in\mathbb{Z}^2\,;\, x\ge y \}$ we have
\[
W_{2m}((0,0);L)=C_m\binom{2m}{m}=\frac{1}{m+1}\binom{2m}{m}^2,
\quad m=0,1,2,\dots,
\]
and $W_{2m+1}((0,0);L)=0$.
Indeed, by Theorem \ref{thm:ZXMP5 and Z_+XMZ} we have
$L\cong (\mathbb{Z}_+\times_K \mathbb{Z})^o$,
where the origin $(0,0)$ in $L$
corresponds to $o=(0,0) \in\mathbb{Z}_+\times_K \mathbb{Z}$.
Hence
\[
W_{m}((0,0);L)
=W_m((0,0);\mathbb{Z}_+\times_K \mathbb{Z})
=W_m(0;\mathbb{Z}_+)W_m(0;\mathbb{Z}),
\]
where Theorem \ref{thm:number of walks of Kronecker product} is applied.
Then the result follows from
\eqref{4eqn:walks in Z} and \eqref{4eqn:walks in Z+}.
(2) For $L=L\{(x,y)\in\mathbb{Z}^2\,;\, x\ge y\ge -x\}$ we have
\[
W_{2m}((0,0);L)=C_m^2=\frac{1}{(m+1)^2}\binom{2m}{m}^2,
\quad m=0,1,2,\dots,
\]
and $W_{2m+1}((0,0);L)=0$.
Indeed, we get the result from Theorem \ref{thm:PmXMPn}
along with a similar argument as in the previous example.
(3) For $L[\mathbb{Z}^2]=\mathbb{Z}\times_C \mathbb{Z}$ we have
\begin{equation}\label{04eqn:counting walks in Z^2}
W_{2m}((0,0);\mathbb{Z}\times_C\mathbb{Z})
=\binom{2m}{m}^2,
\qquad m=0,1,2,\dots.
\end{equation}
Indeed, from Theorem \ref{thm:ZXMZ} we see that
$L[\mathbb{Z}^2]=\mathbb{Z}\times_C \mathbb{Z}
\cong (\mathbb{Z}\times_K \mathbb{Z})^o$.
Then we obtain
\begin{align*}
W_{2m}((0,0);\mathbb{Z}\times_C\mathbb{Z})
&=W_{2m}((0,0);\mathbb{Z}\times_K\mathbb{Z}) \\
&=W_{2m}(0;\mathbb{Z})W_{2m}(0;\mathbb{Z})
=\binom{2m}{m}^2,
\end{align*}
as desired.
Formula \eqref{04eqn:counting walks in Z^2} is derived in a different way.
Applying Proposition \ref{prop:M_m(G_1XCG_2)} to
the Cartesian product $\mathbb{Z}\times_C \mathbb{Z}$, we obtain
\begin{align*}
W_{2m}((0,0);\mathbb{Z}\times_C\mathbb{Z})
&=\sum_{k=0}^m \binom{2m}{2k} W_{2k}(0;\mathbb{Z}) W_{2m-2k}(0;\mathbb{Z})
\label{4eqn:in ZXZ(2)} \\
&=\sum_{k=0}^m \binom{2m}{2k} \binom{2k}{k}
\binom{2m-2k}{m-k},
\nonumber
\end{align*}
where $W_{2m+1}(0;\mathbb{Z})=0$ is taken into account.
By comparing with \eqref{04eqn:counting walks in Z^2}
we get the following interesting relation:
\[
\sum_{k=0}^m \binom{2m}{2k} \binom{2k}{k}
\binom{2m-2k}{m-k}
=\binom{2m}{m}^2.
\]
Of course, one may calculate the left-hand side directly
by using the Vandermonde convolution formula for binomial coefficients
to get the right-hand side.
\end{example}
Finally we record the case where $D\subset \mathbb{Z}^2$ is bounded
in one or two directions,
see Theorems \ref{thm:ZXMP5 and Z_+XMZ} and \ref{thm:PmXMPn}.
\begin{example}
(1) For $L=L\{(x,y)\in\mathbb{Z}^2\,;\, x\ge y \ge x-(n-1)\}$ with
$n\ge2$ we have
\[
W_{2m}((0,0);L)
=W_{2m}(0;P_n)W_{2m}(0;\mathbb{Z})
=\binom{2m}{m} W_{2m}(0;P_n),
\qquad
m=0,1,2,\dots.
\]
(2) For $L=L\{(x,y)\in\mathbb{Z}^2\,;\,
0\le x+y\le k-1, \,\, 0\le x-y\le l-1\}$ with $k\ge2$ and $l\ge2$,
we have
\[
W_{2m}((0,0);L)=W_{2m}(0;P_k)W_{2m}(0;P_l),
\qquad
m=0,1,2,\dots.
\]
\end{example}
\begin{remark}\label{rem-moments-P_n}
A closed formula for $W_m(0;P_n)$ may be written down.
Set
\[
\lambda_k=2\cos\frac{k\pi}{n+1}\,,
\qquad k=1,2,\dots,n,
\]
which are, in fact, obtained from zeroes of the Chebyshev polynomials
of the second kind.
We know that $\{\lambda_1,\dots,\lambda_n\}$ constitute
the spectrum of $P_n$ (\cite[Section 1.4.4]{Brouwer-Haemers2010}).
Then there exist real constants $a_1,\dots,a_n$ such that
\begin{equation}\label{04eqn:in remark 4.12}
W_m(0;P_n)=\sum_{k=1}^n a_k \lambda_k^m,
\qquad m=0,1,2,\dots.
\end{equation}
Then, \eqref{04eqn:in remark 4.12} gives rise to
a linear system $\bm{b}=\Lambda \bm{a}$.
For $m\le 2n$ we have
\[
W_m(0;P_n)=W_m(0;\mathbb{Z}_+)
=\begin{cases}
C_{m/2}, & \text{if $m$ is even}, \\
0, & \text{otherwise},
\end{cases}
\]
and the Vandermonde matrix $\Lambda$ is
easily inverted, we obtain $a_1,\dots,a_n$ uniquely from
$\bm{a}=\Lambda^{-1}\bm{b}$.
Here is a concrete example:
\[
W_{2m}(0;P_4)=\frac{5-\sqrt5}{10}\bigg(\frac{3+\sqrt5}{2}\bigg)^m
+\frac{5+\sqrt5}{10}\bigg(\frac{3-\sqrt5}{2}\bigg)^m
\]
for $m=0,1,2,\dots$,
and, of course, $W_{2m+1}(0;P_4)=0$.
\end{remark}
\subsection{Spectral distributions}
We will describe spectral distributions
corresponding to graphs with Kronecker product structures.
We begin with their building blocks, namely,
spectral distributions associated to $\mathbb{Z}$, $\mathbb{Z}_+$ and $P_n$.
The \textit{arcsine distribution} with mean 0 and variance 2
is defined by the density function:
\begin{equation}\label{eqn:arcsine}
\alpha(x)=\frac{1}{\pi\sqrt{4-x^2}}\,1_{(-2,2)}(x),
\qquad x\in\mathbb{R}.
\end{equation}
The \textit{semicircle distribution} with mean 0 and variance 1
is defined by the density function:
\begin{equation}\label{eqn:semicircle}
w(x)=\frac{1}{2\pi}\sqrt{4-x^2}\,1_{[-2,2]}(x),
\qquad x\in\mathbb{R}.
\end{equation}
By elementary calculus we have
\begin{align}
M_{2m}(\alpha)
&=\int_{\mathbb{R}} x^{2m} \alpha(x)\,dx
=\binom{2m}{m} = W_{2m}(0;\mathbb{Z}),
\label{eqn:W_{2m}(0;Z)}\\
M_{2m}(w)
&=\int_{\mathbb{R}} x^{2m} w(x)\,dx
=C_m=\frac{1}{m+1}\binom{2m}{m}
=W_{2m}(0;\mathbb{Z}_+),
\label{eqn:W_{2m}(0;Z+)}
\end{align}
for $m=0,1,2,\dots$.
We see from Remark \ref{rem-moments-P_n}
that the spectral distribution $\pi_n$ associated to $P_n$ is given by
\[
\pi_n = \sum_{k=1}^n a_k \delta_{\lambda_k},
\]
where $\delta_x$ is the Dirac measure on the point $x\in \mathbb{R}$.
Now we move to the 2-dimensional cases associated to Cartesian and Kronecker products.
\begin{example}
For the Cartesian product $\mathbb{Z}\times_C\mathbb{Z}$
we have
\begin{align*}
W_{m}((0,0);\mathbb{Z}\times_C\mathbb{Z})
&=\sum_{k=0}^m \binom{m}{k} W_{k}(0;\mathbb{Z})W_{m-k}(0;\mathbb{Z}) \\
&=\sum_{k=0}^m \binom{m}{k} M_{k}(\alpha) M_{m-k}(\alpha)
=M_{m}(\alpha*\alpha).
\end{align*}
While, for the Kronecker product we have
\begin{align*}
W_{m}((0,0);\mathbb{Z}\times_K\mathbb{Z})
&=W_{m}(0;\mathbb{Z})W_m(0;\mathbb{Z}) \\
&=M_{m}(\alpha) M_m(\alpha)
=M_{m}(\alpha *_M \alpha).
\end{align*}
Since $\mathbb{Z}\times_C\mathbb{Z}\cong
(\mathbb{Z}\times_K\mathbb{Z})^o$, we have
\begin{equation}\label{04eqn:M_m coincide}
M_{m}(\alpha*\alpha)=M_{m}(\alpha*_M\alpha),
\qquad m=0,1,2,\dots.
\end{equation}
Since $\alpha*\alpha$ (as well as $\alpha*_M \alpha$) has a compact support,
\eqref{04eqn:M_m coincide} is sufficient to claim
that $\alpha*\alpha=\alpha*_M\alpha$.
By similar argument we obtain
the spectral distributions for some restricted lattices.
The following table summarizes the results.
\smallskip
\begin{center}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{|c|c|c|}\hline
Domain $D$ & $W_{2m}(L[D],O)$ & spectral distribution \\ \hline
$\mathbb{Z}$ & $\binom{2m}{m}$ & $\alpha$ \\
$\mathbb{Z}_+$ & $C_m$ & $w$ \\ \hline
$\mathbb{Z}^2$ & $\binom{2m}{m}^2$ & $\alpha*\alpha=\alpha*_M\alpha$ \\
$\{x\ge y\}$
& $C_m \binom{2m}{m}$ & $w*_M \alpha$ \\
$\{x\ge y\ge-x\}$
& $C_m^2$ & $w*_M w$ \\
$\{x\ge0, \, y\ge0\}$
& (A) & $w* w$ \\
$\{x\ge y \ge x-(n-1)\}$
& (B) & $\pi_n*_M\alpha$ \\
$\left\{\begin{array}{l}
0\le x+y\le k-1, \\
0\le x-y\le l-1
\end{array}
\right\}$ & (C) & $\pi_k*_M \pi_l$
\\
\hline
\end{tabular}
\vspace*{10pt}
\end{center}
Concise formulas for (A)--(C) are not known, but we have
\begin{gather*}
\mathrm{(A)}=\sum_{k=0}^m \binom{2m}{2k} C_kC_{m-k},
\qquad
\mathrm{(B)}=W_{2m}(0;P_n)\binom{2m}{m}, \\
\mathrm{(C)}=W_{2m}(0;P_k)W_{2m}(0;P_l).
\end{gather*}
\end{example}
\subsection{Calculating density functions}
In this section we investigate closed forms of density functions of
the spectral distributions $\alpha*\alpha$, $w*_M\alpha$ and $w*_M w$.
\begin{example}
(1) It follows from Proposition \ref{Prop:density of Mellin convolution} that
the density function of $w*_M \alpha$ is given by $2w \star \alpha$. Since both $w(x)$ and $\alpha(x)$ are supported by
the interval $[-2,2]$, we see easily that $w\star \alpha(x)=0$ for $x>4$.
Then, in terms of the explicit forms \eqref{eqn:arcsine}
and \eqref{eqn:semicircle}, we have:
\begin{align}\label{eqn:w star alpha}
w\star \alpha(x)
&=\int_0^{\infty} w(y)\alpha\Big(\frac{x}{y}\Big) \frac{dy}{y}\\
&=\frac{1}{2\pi^2} \int_{x/2}^2 \sqrt{4-y^2}\,
\frac{1}{\sqrt{4-(x/y)^2}}\, \frac{dy}{y}
\nonumber\\
&=\frac{1}{2\pi^2} \int_{x/2}^2
\sqrt{\frac{4-y^2}{4y^2-x^2}}\,dy,
\qquad 0\le x\le 4.
\nonumber
\end{align}
Here we need elliptic integrals and some relevant formulas \cite{Jeffrey}.
The complete elliptic integrals of the first and second kinds
are defined respectively by
\begin{align*}
K(k)&=\int_0^{\pi/2}\frac{d\theta}{\sqrt{1-k^2\sin^2 \theta}}
=\int_0^1 \frac{dx}{\sqrt{(1-x^2)(1-k^2x^2)}}, \\
E(k)&=\int_0^{\pi/2}\sqrt{1-k^2\sin^2 \theta}\, d\theta
=\int_0^1 \sqrt{\frac{1-k^2x^2}{1-x^2}}\,dx,
\end{align*}
where $k^2<1$.
Using the formula:
\[
\int_b^a \sqrt{\frac{a^2-t^2}{t^2-b^2}}\, dt
=a(K(k)-E(k)),
\quad
0<b<a,
\quad
k=\frac{\sqrt{a^2-b^2}}{a}\,,
\]
\eqref{eqn:w star alpha} becomes
\[
w\star \alpha(x)
=\frac{1}{2\pi^2}\{ K(\xi(x))-E(\xi(x)) \},
\]
where
\[
\xi(x)=\sqrt{1-\frac{x^2}{16}}\,.
\]
Consequently, the density function of $w*_M \alpha$ is given by
\[
\frac{1}{\pi^2}\{ K(\xi(x))-E(\xi(x)) \}1_{[-4,4]}(x),
\quad x\in\mathbb{R}.
\]
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=200pt,keepaspectratio,clip]{wMa.pdf}
\caption{The density function of $w*_M\alpha$}
\label{fig:wMa}
\end{center}
\end{figure}
(2) Similarly, the density function of $\alpha*_M \alpha=\alpha*\alpha$
is given by
\[
\frac{1}{2\pi^2}\,
K(\xi(x))1_{[-4,4]}(x),
\qquad x\in\mathbb{R},
\]
and the density function of $w*_M w$ by
\[
\frac{2}{\pi^2}\left\{
\left(1+\frac{x^2}{16}\right)K(\xi(x))-2E(\xi(x))\right\}
1_{[-4,4]}(x),
\qquad x\in\mathbb{R}.
\]
\end{example}
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=200pt,keepaspectratio,clip]{aMa.pdf}
\caption{The density function of $\alpha*_M\alpha$}
\label{fig:aMa}
\end{center}
\end{figure}
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=200pt,keepaspectratio,clip]{wMw.pdf}
\caption{The density function of $w*_Mw$}
\label{fig:wMw}
\end{center}
\end{figure}
\section{Examples in higher dimension}
\label{sec:Examples in higher dimension}
In this section we focus on some higher dimensional examples.
We begin with all possible combinations of products on $\mathbb{Z}^3$,
namely $\mathbb{Z}\times_K \mathbb{Z}\times_K\mathbb{Z}$ ,
$(\mathbb{Z}\times_K \mathbb{Z})\times_C\mathbb{Z}$,
$(\mathbb{Z}\times_C \mathbb{Z})\times_K\mathbb{Z}$
and $\mathbb{Z}\times_C \mathbb{Z}\times_C\mathbb{Z}$.
\begin{example}\label{5ex:1}
(1) The Kronecker product $\mathbb{Z}\times_K \mathbb{Z}\times_K\mathbb{Z}$
has 4 connected components, which are mutually isomorphic.
We have
\[
W_{2m}((0,0,0);\mathbb{Z}\times_K \mathbb{Z}\times_K\mathbb{Z})
=\binom{2m}{m}^3,
\qquad m=0,1,2,\dots.
\]
The connected component containing $O(0,0,0)$, as is illustrated in
Figure \ref{fig:ZXMZXMZ}, is the body-centerd cubic lattice
or a kind of octahedral honeycomb.
For $(\mathbb{Z}\times_C \mathbb{Z})\times_K \mathbb{Z})^o$ we have
\[
((\mathbb{Z}\times_C \mathbb{Z})\times_K \mathbb{Z})^o
\cong
((\mathbb{Z}\times_K \mathbb{Z})^o\times_K \mathbb{Z})^o
\cong
(\mathbb{Z}\times_K \mathbb{Z}\times_K \mathbb{Z})^o.
\]
Hence counting walks in
$(\mathbb{Z}\times_C \mathbb{Z})\times_K \mathbb{Z}$ is
reduced to the previous one.
\begin{figure}[hbt]
\begin{center}
\includegraphics[width=320pt,keepaspectratio,clip]{08_ZXMZXMZ.pdf}
\caption{$(\mathbb{Z}\times_K \mathbb{Z}\times_K\mathbb{Z})^o$}
\label{fig:ZXMZXMZ}
\end{center}
\end{figure}
(2) For other combinations of products of $\mathbb{Z}$ we see that
\[
((\mathbb{Z}\times_K \mathbb{Z})\times_C \mathbb{Z})^o
\cong (\mathbb{Z}\times_K \mathbb{Z})^o\times_C \mathbb{Z}
\cong (\mathbb{Z}\times_C \mathbb{Z})\times_C \mathbb{Z},
\]
which is the usual 3-dimensional integer lattice.
Hence
\begin{align*}
&W_{2m}((0,0,0);(\mathbb{Z}\times_C \mathbb{Z})\times_C\mathbb{Z})
=W_{2m}((0,0,0);(\mathbb{Z}\times_K \mathbb{Z})\times_C\mathbb{Z}) \\
&\qquad
=\sum_{k=0}^m \binom{2m}{2k}\binom{2k}{k}^2\binom{2m-2k}{m-k}
=\sum_{k=0}^m \frac{(2m)!(2k)!}{(m-k)!^2 k!^4}.
\end{align*}
Of course the above result is well known,
and our contribution here would be the derivation using the Kronecker product.
\end{example}
The last example is a very interesting case of products on $\mathbb{Z}^3_+$, which is related to a restricted lattice in $\mathbb{Z}^3$.
\begin{example}
The graph $(\mathbb{Z}_+\times_K \mathbb{Z}_+)\times_C \mathbb{Z}_+$
has two connected components
and we consider the connected component
$((\mathbb{Z}_+\times_K \mathbb{Z}_+)\times_C \mathbb{Z}_+)^o$
containing $O=(0,0,0)$.
Then we have
\begin{align}
&W_{2m}((0,0,0);(\mathbb{Z}_+\times_K \mathbb{Z}_+)\times_C \mathbb{Z}_+)
\label{5eqn:ex 5.2} \\
&\qquad=\sum_{k=0}^m\binom{2m}{2k}
W_{2k}((0,0);\mathbb{Z}_+\times_K \mathbb{Z}_+)
W_{2m-2k}(0;\mathbb{Z}_+)
\nonumber \\
&\qquad=\sum_{k=0}^m\binom{2m}{2k} C_k^2 C_{m-k}
\nonumber \\
&\qquad=\sum_{k=0}^m\frac{(2m)!(2k)!}{(m-k)!(m-k+1)! k!^2(k+1)!^2}\,.
\nonumber
\end{align}
It is remarkable that
the last summation has been already obtained in \cite{Wimp-Zeilberger1989}
as the number of walks in the 3-dimensional
restricted lattice $L\{x\ge y\ge z\} = \{ (x,y,z) \in \mathbb{Z}^3 : x\ge y\ge z \}$, namely,
\[
W_{2m}((0,0,0);(\mathbb{Z}_+\times_K \mathbb{Z}_+)\times_C \mathbb{Z}_+)
=W_{2m}((0,0,0);L\{x\ge y\ge z\}),
\]
for all $m=0,1,2,\dots$.
It is, however, noted that
$((\mathbb{Z}_+\times_K \mathbb{Z}_+)\times_C \mathbb{Z}_+)^o$
and $L\{x\ge y\ge z\}$ are not isomorphic.
For example, in the former graph
there is a unique vertex with degree 2 (that is, $O=(0,0,0)$),
while there are many vertices with degree 2 in the latter.
\end{example}
A similar phenomenon is observed also in the two-dimensional case.
\begin{example}
It follows by the usual reflection argument that
\[
W_{2m}(1;\mathbb{Z}_+)
=\binom{2m}{m}-\binom{2m}{m+2}
=C_{m+1}\,.
\]
On the other hand, it is known \cite{Guy-Krattenthaler-Sagan1992} that
\[
W_{2m}((0,0);\mathbb{Z}\times_C\mathbb{Z}_+)
=\binom{2m}{m}\binom{2m+2}{m}
-\binom{2m+2}{m+1}\binom{2m}{m-1}
=C_mC_{m+1}\,.
\]
Therefore,
\[
W_{2m}((0,0);\mathbb{Z}\times_C\mathbb{Z}_+)
=W_{2m}((0,1);\mathbb{Z}_+\times_K\mathbb{Z}_+),
\]
though two graphs $\mathbb{Z}\times_C\mathbb{Z}_+$
and $\mathbb{Z}_+\times_K\mathbb{Z}_+$ are not isomorphic.
\end{example}
|
\section{Introduction}
\label{intro}
Full exploitation of a radio telescope requires detailed and accurate knowledge of its radiation properties. Modern tools of antenna engineering, electromagnetic simulators, offer the ability to analyze telescope performance. To what extent can they deliver a true picture of telescope characteristics? Can we use them to calculate telescope behavior to useful accuracy? In this paper we investigate these questions by analyzing the properties of a large radio telescope and comparing the computations with measurements. The telescope that we study is the John A. Galt Telescope at the Dominion Radio Astrophysical Observatory.
We focus on two problems particularly relevant to mapping the extended emission from the Milky Way, which is partially linearly polarized at decimeter wavelengths. First, we need to know the gain (or, equivalently, the aperture efficiency) of the telescope so that we can report our observations in units of brightness temperature. Second, we need to know the polarization behavior of the telescope so that we can correct observations for instrumental polarization.
The John A. Galt Telescope (we will refer to it as the Galt Telescope) is a 26 m axially symmetric paraboloidal reflector. The Galt Telescope has recently been used to map the polarized emission from the entire northern sky over the frequency range 1280 to 1750~MHz as part of the Global Magneto-Ionic Medium Survey (GMIMS -- \citealp{woll09}). The results reported here have been used in processing the data from that survey.
The challenge in astronomical polarimetry is the accurate measurement of a small polarized signal embedded in a (usually) much larger randomly polarized signal.{\footnote{A randomly polarized signal is one whose time average has no net polarization} Instrumental polarization has a deleterious effect on polarimetry because it converts an unpolarized signal into an apparently polarized one. Radiation from the ground is unavoidable in most reflector telescopes, and this emission is similarly converted into an unwanted polarized signal, usually a substantial one. We examine the properties of ground emission and the telescope response to it.
To attain our two goals we compute the total radiation pattern of the telescope, including its polarization response. We describe telescope response in terms of Stokes parameters, widely used in astronomy to characterize the polarization state of a signal \citep{tinb96,wils14} but less frequently used in antenna engineering. Stokes parameter $I$ is proportional to the total intensity of the signal, parameters $Q$ and $U$ together describe the state of linear polarization and parameter $V$ describes the state of circular polarization. For a linearly polarized signal the polarized intensity is ${\rm PI}=\sqrt{{Q^2}+{U^2}}$ and the polarization angle is ${\alpha}={0.5~{{\rm{tan}}^{-1}({U/Q})}}$. Positive $V$ corresponds to right-hand circular polarization (RHCP) and negative $V$ to left-hand circular polarization (LHCP) \citep{ieee79}.
The terminology of antenna engineering is rife with terms that were developed while considering the antenna as a transmitter, for example the use of {\it{feed}} to describe the antenna placed at the focus of a reflector, or {\it{spillover}} to describe radiation from the feed that enters the far field without encountering the reflector. We calculate the radiation properties of the Galt Telescope as a transmitter, and can confidently use the results to understand its behavior as a receiver because reciprocity informs us that the behavior of an antenna as a receiver is completely described by its properties as a transmitter. Our calculations assume that the antenna is transmitting a signal at a single frequency, but we apply our results to receiving the wideband noise signals of radio astronomy.
In describing radiation patterns we use the terms $E$ plane and $H$ plane. The $E$ plane is the plane which contains the axis of a linearly polarized feed (or antenna) and the electric vector of the excitation. The $H$ plane is orthogonal to the $E$ plane, and also contains the antenna axis.
\section{Calculating the Radiation Pattern}
\label{calc}
\subsection{Calculating the Radiation Properties of the Feed}
\label{feed}
The feed is based on the design of \citet{wohl72}, scaled to a center frequency of 1576~MHz. The radiation properties of the feed were calculated using the CST software package \citep{cst14}; representative radiation patterns are shown in Figure~\ref{primary}. At the nominal center frequency the $E$- and $H$-plane patterns have nearly equal width and closely match the patterns measured by \citet{wohl72} on a 2800 MHz version of the feed. At frequencies below the design frequency of 1576~MHz the $E$-plane pattern is narrower than the $H$-plane pattern, and above that frequency the reverse is true.
\begin{figure}[ht!]
\centerline{\includegraphics[width=0.48\textwidth]{feed_patterns_rev.pdf}}
\caption{Radiation patterns of the feed as a function of the polar angle $\rho$ at the indicated frequencies,
calculated using CST. Solid curves: $E$-plane patterns. Dashed curves: $H$-plane
patterns. Dot-dash curves: cross-polar patterns in the $45^{\circ}$ plane.
The vertical lines mark the outer edge of the reflector at
${\rho}={80^{\circ}}$.}
\label{primary}
\end{figure}
\subsection{Calculating the Radiation Properties of the Telescope}
\label{grasp}
Our computations used the software package GRASP-10, version 10.1.0 \citep{gras12}, which employs physical optics (PO) and the physical theory of diffraction (PTD) to calculate the far field radiated from the telescope. The calculation includes the effects of the feed, feed-support struts, and reflector, taking account of aperture blockage. CST provides the incident field from the feed, including its full polarization characteristics, as input to the calculation. GRASP-10 uses this input to calculate surface currents on the basis of PO, modified by PTD which models the current near the rim of the reflector. The radiation pattern of the telescope is then calculated using those currents. Therefore, GRASP-10 accurately models shadowing and the effects commonly attributed to diffraction.
The reflector has a focal-length-to-diameter ratio of 0.298, and subtends $160.1^{\circ}$ at the feed. The feed struts make an angle of $34^{\circ}$ with the telescope axis. Because of limitations of GRASP-10 the feed struts in the model do not contact the reflector surface or the feed (the feed is not a physical structure in GRASP-10, but merely a radiation source with known properties). Furthermore, the struts are modelled as metal cylinders, when in fact they are made of fiberglass, have a cigar shape, and two of them have metal-sheathed cables running along them.\footnote{After version 10.3.0 GRASP is able to model dielectric feed support struts.} We will discuss the impact of these approximations in Section~\ref{disc}.
Each GRASP-10 calculation follows this procedure: (a) scattered fields from the struts, induced by the spherical wave emanating from the feed, are calculated, (b) the reflected (plane) wave from the paraboloidal surface is calculated, taking into account both the spherical wave from the feed and the fields scattered from the struts, (c) the fields scattered by the struts are calculated, as the struts are illuminated by the plane wave from the reflector, and (d) fields scattered from the struts and the fields of the plane wave from the reflector are added. Note that the struts scatter radiation {\it{twice}}. The work was extended by including the effects of departures of the reflector surface from a perfect paraboloid (surface roughness), and of leakage of radiation through the mesh surface of the reflector. Computations were made at 50 MHz intervals from 1250 to 1750~MHz.
The output of GRASP-10 contained the real and imaginary components of the electric far field, $E$. The Stokes parameters were calculated as
\begin{subequations}\label{stokes}
\begin{align}
I &= {[{\rm Re}(E_x)]^2} + {[{\rm Im}(E_x)]^2} + {[{\rm Re}(E_y)]^2} + {[{\rm Im}(E_y)]^2} \thinspace,\\
Q &= {[{\rm Re}(E_x)]^2} + {[{\rm Im}(E_x)]^2} - {[{\rm Re}(E_y)]^2} - {[{\rm Im}(E_y)]^2} \thinspace,\\
U &= 2\thinspace{\rm Re}(E_x){\thinspace}{\rm Re}(E_y) + 2\thinspace{\rm Im}(E_x){\thinspace}{\rm Im}(E_y) \thinspace,\\
V &= 2\thinspace {\rm Im}(E_x){\thinspace}{\rm Re}(E_y) - 2\thinspace{\rm Re}(E_x){\thinspace}{\rm Im}(E_y) \thinspace,
\end{align}
\end{subequations}
where $x$ and $y$ are the co- and cross-polar directions.
We need to calculate the response of the telescope to an incoming unpolarized signal. Since we treat the antenna as a transmitter, we need to simulate an antenna that transmits an unpolarized signal.{\footnote{An unpolarized transmitter is a theoretical concept: all real antennas are polarized, and therefore all transmitted signals are polarized}} Random polarization is simulated by rotating the feed to four positions in $45^{\circ}$ steps and averaging the resulting $I$, $Q$, $U$, and $V$ patterns: this technique was
developed by \citet{ng05}. Thus
\begin{equation}
{Q_{\rm random}}=( {{Q_{45^{\circ}}}+{Q_{90^{\circ}}}+{Q_{135^{\circ}}}+{Q_{180^{\circ}}}} )/4 \thinspace,
\end{equation}
and similarly for the other Stokes parameters.
\section{Results - Radiation Patterns}
\label{radpa}
\begin{figure*}[ht]
\centerline{
\includegraphics[bb = 195 150 598 596,width=0.27\textwidth,clip]
{I_fd_noleakage.pdf}
\includegraphics[bb = 195 150 598 596,width=0.27\textwidth,clip]
{I_bd_noleakage.pdf}
\includegraphics[bb = 195 150 598 596,width=0.27\textwidth,clip]
{I_bd_meshleakage.pdf}}
\centerline{
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{Q_fd_noleakage.pdf}
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{Q_bd_noleakage.pdf}
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{Q_bd_meshleakage.pdf}}
\centerline{
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{U_fd_noleakage.pdf}
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{U_bd_noleakage.pdf}
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{U_bd_meshleakage.pdf}}
\centerline{
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{V_fd_noleakage.pdf}
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{V_bd_noleakage.pdf}
\includegraphics[bb = 195 150 598 470,width=0.27\textwidth,clip]
{V_bd_meshleakage.pdf}}
\caption{Computed radiation patterns at 1400~MHz for Stokes $I$, $Q/I$,
$U/I$, and $V/I$ shown for front and back hemispheres in
stereographic projection in the coordinate system ${\theta},{\phi}$,
where $\theta$ is the angle from boresight and $\phi$ is the
circumferential angle. Units of $I$ are decibels relative to
isotropic. Units of $Q/I$, $U/I$ and $V/I$ are fractional intensity. The left
two columns show the front and back hemispheres of the radiation pattern
from a reflector with surface errors (see text for details) but no
leakage through the mesh. The right column shows the back hemisphere of
the radiation patterns when leakage is added. The three lines superimposed
on the forward $I$ pattern show the orientation of the feed-support struts;
this figure is not an accurate picture of the telescope.}
\label{rp}
\end{figure*}
\begin{figure}
\centerline{\includegraphics[width=0.48\textwidth]
{average_rad_pattern.pdf}}
\caption{Radial average of the computed radiation pattern at 1400~MHz as a function of the polar angle $\theta$, shown with and without radiation leaking through the mesh. The peak at $\theta\approx68^\circ$ arises from the scatter cones, and the peak at $\theta\approx95^\circ$ from spillover. Mesh leakage affects the pattern only for $\theta > 100^\circ$.}
\label{radial}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=0.3\textwidth,clip]{sph_blockage_shadow_rmk_clear.pdf}}
\caption{Face-on view of the telescope. The parts of the reflector
shadowed by the feed-support struts are hatched. The pale shaded
structure behind the reflector is the support tower.}
\label{shadows}
\end{figure}
Figure~\ref{rp} shows the results of GRASP-10 calculations using the technique described above. The pattern denoted $I$ is the familiar radiation pattern of antenna engineering. The three other patterns are generated by computing the radiated field in terms of $Q$, $U$ and $V$ and dividing each by $I$. The $Q/I$, $U/I$, and $V/I$ patterns indicate the conversion of unpolarized emission into apparently polarized signal. They indicate a potential error term that must be corrected.\footnote{While this is the dominant problem in
polarimetry of the linearly polarized Galactic emission, we note that a similar technique could be used to compute the conversion of any Stokes parameter into any other, for example $Q$ or $U$ into $V$ (the conversion of linearly polarized
emission into apparently circularly polarized signal). If a linearly polarized radiation pattern is calculated, the transfer of signal into $V$ can easily be determined. Such a calculation would be useful for observations of Zeeman
splitting of the H\thinspace\scriptsize{I} line.} They show the response of a polarization receiver attached to the telescope to a randomly polarized sky. Such a sky should produce a response in $I$ only, and the responses in $Q/I$, $U/I$ and
$V/I$ constitute the instrumental polarization whose effects we must correct for if we are to obtain a true representation of the sky.
The calculations shown in the left two columns of Figure~\ref{rp} include the effects of the three feed-support struts and include departure of the surface from a perfect paraboloid, but they do not include leakage through the mesh that forms the reflector surface. The $I$ radiation pattern is dominated by the main beam, the ring of {\it{spillover}} emission, and the {\it{scatter cones}}. Spillover is direct feed radiation that is not intercepted by the reflector. It begins at the reflector rim (${\theta}={80^{\circ}}$) in the front hemisphere and continues to dominate in the back hemisphere, with some redistribution from diffraction at the reflector rim. The scatter cones \citep{land91} are the three circles that intersect on the main beam and dominate sidelobes in the forward hemisphere. They are generated by interaction of the plane wave rising from the paraboloidal surface of the reflector with the feed-support struts. The plane wave induces currents in each strut, and the strut becomes a travelling-wave antenna, radiating into a cone whose opening angle ($68^{\circ}$) is twice the angle between the strut and the telescope axis. The dominance of the spillover and the scatter cones can be appreciated from the plot in Figure~\ref{radial} which shows the radial average of the radiation pattern.
Shadowing by the struts is strong in this telescope (see Figure~\ref{shadows}). This translates into the $I$ pattern, producing a variation with $\phi$ up to $\pm$5 dB in the range ${60^{\circ}}<{\theta}<{100^{\circ}}$.
Both the spillover lobes and the scatter cones are strong features of the $Q/I$, $U/I$, and $V/I$ images. In other words these regions of the radiation pattern are strongly polarized. This is expected: the rim of the reflector and the feed-support struts are linear structures (certainly they are linear on the scale of the wavelength, ${\lambda}\approx{20\thinspace{\rm{cm}}}$); each will have markedly different response to radiation parallel to the linear structure and radiation orthogonal to it. It is also clear from Figure~\ref{rp} that the struts have a strong interaction with the spillover lobes. While the principal source of the spillover lobes is direct radiation from the feed, the spherical wave from the feed has a strong interaction with the struts.
The calculated $I$ response shows that the radiation scattered by the struts is quite concentrated into the scatter cones. Sidelobes which follow the form of the scatter cones are visible in other parts of the radiation pattern, but at a low level. However, in $Q/I$, $U/I$, and $V/I$ the effects of the scatter cones heavily modulate the radiation pattern over almost the entire front hemisphere. Figure~\ref{rp} tells us that understanding instrumental polarization is a matter of understanding how the spillover lobes and the scatter cones interact with the telescope environment.
All the patterns, $Q/I$, $U/I$, and $V/I$, reach values close to 100\%, telling us that there are regions of the radiation pattern that are almost completely polarized. The $Q/I$ and $U/I$ patterns reach these values at ${\theta}\approx{45^{\circ}}$ within the scatter cones and at certain azimuths in the spillover region. Peak values of $V/I$ are found at large $\theta$, but the dominant effects will come from the spillover regions.
We expect the $Q/I$ and $U/I$ patterns to resemble each other but with a $45^{\circ}$ rotation between them. This is evident in the rear hemisphere, but in the front hemisphere the influence of the $120^{\circ}$ azimuthal spacing of the struts has an equal effect.
Computation of the radiation pattern from a perfect reflector serves to illustrate the major influences at work, but the real antenna has a surface that deviates from the perfect paraboloidal shape. The surface accuracy was established by \citet{higg00} as ${\delta}={0.45}\thinspace{\rm{cm}}$\thinspace{rms} from measurements of the aperture efficiency at different wavelengths. We included surface roughness of this magnitude in the GRASP-10 calculations. Surface roughness decreases the gain, $G$, of the antenna, and we compared the gain reduction factor, $k_G$, derived from our GRASP-10 results with that expected from the formula of \citet{ruze66}, ${k_G}={e^{{-(4{\pi}{\delta}/{\lambda})}^2}}$. The results of GRASP-10 calculations agree with the predictions of the Ruze formula within 3\% at all frequencies. The gain reduction, ${\Delta}G$, from surface roughness at 1400~MHz is 6.4\%.
The mesh surface is also partly transparent at ${\lambda}\approx{20\thinspace{\rm{cm}}}$, and allows some feed radiation to pass directly into the rear hemisphere (and, equivalently, some ground radiation to reach the feed through the mesh). We calculated the transparency of the mesh from its dimensions using the empirical formula given by \cite{mumf61}. The power leaking through the mesh ranges from 0.3\% of the incident power at 1250~MHz to 0.7\% at 1750~MHz. We calculated the transmission coeffient, $\mu$, and used it to compute a radiation pattern incorporating the leakage radiation with the GRASP-10 output, combined vectorially at the appropriate levels. The leakage signal is feed radiation, attenuated, with its polarization state unaltered. The leakage through the mesh swamps the diffracted fields in the rear hemisphere, and the fractional polarization becomes quite low when leakage is included. Leakage through gaps in the reflector surface was not considered.
Radiation patterns at 1400~MHz showing all contributing factors are presented in Figure~\ref{rp}, and a radial average of the Stokes $I$ pattern is shown in
Figure~\ref{radial}. In all subsequent discussion we use these radiation patterns and the similar patterns at other frequencies.
\section{Results --- Aperture and Beam Efficiencies}
\subsection{Aperture Efficiency and Beam Efficiency from GRASP-10}
\label{gain}
Figure~\ref{eta-beta} shows the calculated aperture efficiency, $\eta_A$, and beam efficiency, $\eta_B$, of the telescope. Fitted curves are shown for both parameters, confined to the frequency range of the GMIMS observations, 1280 to 1750~MHz.
To calculate $\eta_B$ we used the following procedure. The half-power beamwidth, $\Theta(\nu)$, of the telescope at frequency $\nu$ varies from $\sim 40'$ at 1270~MHz to $\sim 30'$ at 1740~MHz. We used measured values of $\Theta(\nu)$ and assumed equal $E$- and $H$- plane widths at all $\nu$ (a very good approximation). At each frequency the solid angle of a Gaussian beam was calculated as ${{\Omega}_B({\nu})}={1.13\thinspace{\Theta(\nu)}^2}$. Total antenna solid angle, $\Omega(\nu)$, was derived from the GRASP-10 calculations, and main beam efficiency was calculated as ${\Omega_B}(\nu)/{\Omega}(\nu)$.
Much more complicated determinations of the solid angle of the beam have been used in the past. Some authors have integrated the main beam to the first null in the response. We tried this, identifying the first null in the GRASP-10 patterns, but there were fluctuations in the radius to the first null from one frequency to the next, and these fluctuations led to rapid and unrealistic variation in $\eta_B$ with frequency, so we rejected this option. Other authors have used the ``full-beam brightness temperature'', integrating a measured beam out to some radius (e.g. \citealp{reic82}). This seemed to us an arbitrary procedure, and we could think of no logical way to either choose the radius of the full beam, or alter that radius with frequency. With the wide bandwidth of the GMIMS dataset there are investigations (e.g. spectral index determination) which will compare the data at two frequencies within the dataset. We considered it important to keep consistent processing from one frequency to the next.
It might also be argued that the antenna beams are not Gaussian, and that a better approximation to the beamshape might be found from some other mathematical function. While that is probably true, we note that the GMIMS data will be subject to further processing which will assume that the observing beams {\it{are}} Gaussian. For example, in preparation for Rotation Measure Synthesis \citep{bren05} all data are convolved to a common beamwidth, that at the lowest frequency of the observations.
The small dip in value of $\eta_A$ and $\eta_B$ at 1550~MHz appears to arise from a property of the feed. At that frequency the feed radiation pattern has a distinctly flat top (see Figure~\ref{primary}). This dip in aperture efficiency cannot be seen in our measurements (Figure~\ref{eta-meas}), and we have not investigated it any further. The fitted curves in Figure~\ref{eta-beta} are fits to all data points, including the value at 1550~MHz, but only a simple quadratic function has been fitted and this cannot include the dip.
Values of $\eta_A$ and $\eta_B$ were needed for reduction of the GMIMS data. The survey was calibrated (in janskys) with daily observations of strong sources whose flux densities were well known. Values of $\eta_A$ were used to convert observational data from janskys to antenna temperature. We used the curve fitted to the GRASP-10 values of $\eta_A$ shown in Figure~\ref{eta-beta}, adjusted to fit our measurements of $\eta_A$ (see Section~\ref{meas} and Figure~\ref{eta-meas}). Finally, the data were converted to main-beam brightness temperature using the values of $\eta_B$.
\begin{figure}
\centerline{\includegraphics[width=0.48\textwidth]{ap_beam_eff_rev.pdf}}
\caption{Aperture efficiency, $\eta_A$, calculated with GRASP-10 and main beam
efficiency, $\eta_B$, calculated as described in the text. Surface roughness
and leakage have been included in the calculations. Curves have been fitted
to the data over the frequency range of the GMIMS data (the points at 1250
MHz were excluded from the fits).}
\label{eta-beta}
\end{figure}
\subsection{Aperture Efficiency from Geometrical Optics and Ray Tracing}
We used a second approach to the determination of aperture efficiency, using much simpler techniques in order to provide a check of the GRASP-10 calculations. For this discussion it is convenient to subdivide the aperture efficiency into a number of efficiencies, each of which can be separately evaluated. Thus
\begin{equation}
{\eta_A}={{\eta_i}\,{\cdot}\,{\eta_s}\,{\cdot}\,{\eta_e}
\,{\cdot}\,{\eta_l}\,{\cdot}\,{\eta_b}} \thinspace.
\label{subdiv}
\end{equation}
We now discuss the individual terms on the right-hand side of
equation~\ref{subdiv} (using the language of transmitting and receiving
interchangeably).
\begin{itemize}
\item{Two efficiencies--- $\eta_i$, the illumination efficiency, and $\eta_s$, the spillover efficiency--- reflect the compromise inherent in the design of the feed between the efficient use of the reflector and loss of power into the spillover region. $\eta_i$ is the integrated power across the aperture relative to the power in a uniformly illuminated aperture. $\eta_s$ is the fraction of power that is intercepted by the reflector; the fraction $1-\eta_s$ goes beyond the edge of the reflector and forms the spillover lobes. High levels of $\eta_i$ usually mean low levels of $\eta_s$ and vice versa.}
\item{$\eta_e$ is the surface efficiency. Small surface deflections produce phase errors that reduce antenna gain by $1-\eta_e$.}
\item{$\eta_l$ is the leakage efficiency. This reflects loss of a fraction of $1-\eta_l$ through the mesh surface of the reflector or through gaps between reflector panels.}
\item{$\eta_b$ is the blockage efficiency. The focus equipment and the struts that support it obviously block the aperture, so ${\eta_b}<{1}$. Two effects must be considered: the struts block the incoming plane wave, but they also shadow part of the reflector surface so that the feed cannot receive signal from there. These two effects are known as plane-wave blockage and spherical-wave blockage respectively and we separately evaluate efficiencies ${\eta}_{\rm pw}$ and ${\eta}_{\rm sph}$.}
\end{itemize}
We calculated $\eta_i$, $\eta_s$, $\eta_{\rm pw}$, and $\eta_{\rm sph}$ based on geometrical optics and ray tracing. The calculations were made using a program written by L.A. Higgs \citep{hk00}. The inputs to the program were the physical dimensions of the antenna and the radiation patterns of the feed. We used the patterns calculated by CST (see Section~\ref{feed}), except that the $E$- and $H$-plane patterns were averaged to provide a circularly symmetric illumination of the reflector (the actual patterns come close to this).
Assuming an unblocked aperture, the program calculates the field on the aperture, taking into account the shape of the reflector. The level of feed radiation drops off towards the edge of the reflector (see Figure~\ref{primary}) but there is an additional ``free-space attenuation'' \citep{baar07} which further reduces illumination at the edge of the aperture. The radiation pattern of the feed is measured (or calculated) on a spherical surface centered on the feed, but the outer edges of the aperture plane are further from the focus than the center. This adds an extra taper
\begin{equation}
{\zeta} = {1 + \Biggl({\frac{rD}{4f}}\Biggr)^2} \thinspace,
\end{equation}
where $f$ is the focal length, $D$ is the antenna diameter, and $0<r<1$ is radial distance in the aperture \citep{baar07}. The free-space attenuation is stronger for deeper reflectors; for the Galt Telescope, where ${f/D}=0.298$, it amounts to an additional taper of 4.6~dB at the edge of the aperture beyond that provided by the feed radiation pattern.
The illumination efficiency, $\eta_i$, was calculated from the numerical data for the field distribution on the aperture using the equation
\begin{equation}
{\eta_i} =
{\frac{\biggl(\int{F{\thinspace}dA}\biggr)^2}{{A_p}\int{{F^2}{\thinspace}dA}}} \thinspace,
\label{etai}
\end{equation}
where $F$ is the power distribution in the aperture, including the free-space attenuation, and the integrations extend over the aperture whose physical area is $A_p$ \citep{coll85}.
Calculation of $\eta_s$ is simply an integration of the power radiated by the feed that misses the reflector.
Plane-wave blockage is derived by calculating the geometrical shadow of the central focus equipment and the feed-support struts on the aperture under plane-wave illumination. The tapered illumination of the aperture is taken into account.
The program derives spherical wave blockage by calculating the geometric shadows of the struts on the aperture, taking into account their cigar shape. We use the
relationship from \citet{lamb86} to calculate the blockage
\begin{equation}
{B}= {\frac{ \int_{\rm blockage}{E{\thinspace}dA}}{\int_{\rm aperture}{E{\thinspace}dA}}} \thinspace,
\label{sph-block}
\end{equation}
where $E$ is the amplitude of the aperture field and $dA$ is an area element.
Then
\begin{equation}
{{\eta}_{\rm sph}}={(1-B)^2} \thinspace.
\end{equation}
Effects of surface roughness and of leakage through the mesh, ${\eta}_e$ and ${\eta}_l$, were evaluated as described in Section~\ref{radpa}.
The results of these calculations are summarized in Table~1. Results are tabulated to four significant figures; four-digit precision is not justified by the accuracy of the input data but is preserved so that the gradual trends with frequency remain clear to the reader.
\begin{table*}
\begin{center}
\small
\begin{tabular}{|c|c|c|c|c|c|c||c|c|}
\hline
frequ- & illumin- & spill- & surf- & leak- & plane & spherical & ray & GRASP \\
ency & ation & over & ace & age & wave & wave & tracing & efficiency \\
(MHz) & ${\eta}_I$ & ${\eta}_s$ & ${\eta}_e$ & ${\eta}_l$ & ${\eta}_{\rm pw}$ &
${\eta}_{\rm sph}$ & efficiency & \\
\hline
\hline
1250 & 0.6807 & 0.9323 & 0.9460 & 0.9967 & 0.9554 & 0.8032 & 0.4592 & 0.5044 \\
\hline
1300 & 0.6978 & 0.9533 & 0.9417 & 0.9964 & 0.9598 & 0.8701 & 0.5213 & 0.5372 \\
\hline
1350 & 0.7191 & 0.9602 & 0.9373 & 0.9961 & 0.9628 & 0.8673 & 0.5383 & 0.5602 \\
\hline
1400 & 0.7290 & 0.9602 & 0.9327 & 0.9958 & 0.9641 & 0.8660 & 0.5428 & 0.5670 \\
\hline
1450 & 0.7338 & 0.9591 & 0.9280 & 0.9955 & 0.9646 & 0.8655 & 0.5428 & 0.5745 \\
\hline
1500 & 0.7359 & 0.9573 & 0.9237 & 0.9952 & 0.9650 & 0.8651 & 0.5407 & 0.5787 \\
\hline
1550 & 0.7215 & 0.9551 & 0.9182 & 0.9949 & 0.9664 & 0.8668 & 0.5273 & 0.5696 \\
\hline
1600 & 0.7295 & 0.9681 & 0.9131 & 0.9946 & 0.9669 & 0.8656 & 0.5368 & 0.5779 \\
\hline
1650 & 0.7252 & 0.9677 & 0.9078 & 0.9942 & 0.9640 & 0.8622 & 0.5289 & 0.5774 \\
\hline
1700 & 0.7128 & 0.9694 & 0.9024 & 0.9939 & 0.9619 & 0.8679 & 0.5174 & 0.5698 \\
\hline
1750 & 0.6970 & 0.9717 & 0.8969 & 0.9935 & 0.9592 & 0.8701 & 0.5037 & 0.5595 \\
\hline
\hline
\end{tabular}
\end{center}
\begin{center}
{\bf{Table 1:}} Contributions to aperture efficiency, and a comparison of ray tracing and GRASP-10 calculations.
\end{center}
\end{table*}
Values of $\eta_A$ obtained by the two methods are compared in Figure~\ref{eta-comp}. The ray-tracing results are below the GRASP-10 results by 1\% to 7\%, a fair agreement considering the simplicity of the ray-tracing approach. We conclude that the ray-tracing technique can be used to determine aperture efficiency of a reflector antenna with useful accuracy. Both GRASP-10 and ray-tracing reveal a dip in $\eta_A$ at $\sim$1550~MHz. Since both techniques yield a similar result, we can trace the cause to the feed radiation pattern that is calculated by CST. We see from Table 1 that the illumination efficiency, which depends entirely on the feed characteristics, is lower at 1550 MHz than at adjacent frequencies. However, as remarked above, our measurements do not show this dip. We have not investigated it further and used the fitted curves shown in Figure 5 to process the GMIMS data.
\begin{figure}
\centerline{\includegraphics[width=0.48\textwidth]{ap_eff_comp_rev.pdf}}
\caption{Comparison of aperture efficiency, $\eta_A$, calculated with
GRASP-10 and with ray tracing as a function of frequency.}
\label{eta-comp}
\end{figure}
Table 1 shows that the dominant effects in determining the aperture efficiency are the illumination efficiency and the spherical wave blockage. Together these two factors restrict aperture efficiency to be below about 0.65; other effects are relatively minor. For the Galt Telescope spherical wave blockage is especially important because the feed-support struts are at an angle of only $34^{\circ}$ from the telescope axis while the reflector edge is ${\sim}80^{\circ}$ from the axis; the struts shadow a substantial part of the reflector from the spherical wave emanating from the feed, as is evident from Figure~\ref{shadows}.
\begin{figure}
\centerline{\includegraphics[width=0.44\textwidth]{eta_A_measured_rev.pdf}}
\caption{Measured aperture efficiency, from observations of Cyg~A, in RHCP
(lower curve) and LHCP (upper curve, displaced upwards by 0.2 for clarity).
Effects of interfering signals are evident. The smooth curve is fitted to the
GRASP-10 calculations, adjusted as described in the text.}
\label{eta-meas}
\end{figure}
\section{Measurement of Aperture Efficiency}
\label{meas}
The GMIMS observations were calibrated with four strong sources, Cyg A, Tau A, Vir A, and Cas A. A self-consistent set of flux densities and spectral indices was generated by observations of these sources at the start of the survey \citep{woll10a}. Aperture efficiency of the telescope was subsequently measured from observations of Cyg A, assuming a flux density at 1.4~GHz of 1589~Jy and a spectral index of $-$1.07.
The receiver made all measurements of antenna temperature relative to a switched noise signal injected at 25 Hz into both LHCP and RHCP receiver channels. In a first measurement the amplitude of the noise signal in kelvins was established relative to coaxial resistive terminations at known temperatures, one at liquid nitrogen temperature, one at ambient temperature, and one in a temperature-controlled oven at ${\sim}100 ^{\circ}$C. This experiment did not include the feed and its associated waveguide components, so a second measurement was made which did include those components in the signal path. The high reference temperature was provided by a box filled with absorbing foam placed in front of and around the feed (all at ambient temperature). It was not possible to immerse this large box in liquid nitrogen, so the cold temperature was provided by driving the telescope to the zenith and using the cold sky together with informed estimates of the various contributions to the zenith antenna temperature.
The two measurements gave closely consistent results for the equivalent temperature of the injected noise signal over most of the band, with the exception of two frequency ranges where there were moderate mismatch problems in the feed. The second measurement, by including the feed in the signal path, overcame these difficulties, but had the slight disadvantage that the results were quite sensitive to the adopted value for the antenna temperature at the zenith. Our estimates of antenna temperature at the zenith allowed for ground emission received through the spillover sidelobes (see Section~\ref{spill}) and through the partly transparent mesh reflecting surface, atmospheric emission \citep{gibb86} and the brightness temperature of the sky (including Galactic and extragalactic contributions and the cosmic microwave background).
Figure~\ref{eta-meas} shows the aperture efficiency derived by observing Cyg~A. Results for the two hands of circular polarization, essentially two independent measurements, are shown individually. The smooth curve in Figure~\ref{eta-meas} is based on the fit to the GRASP-10 values of $\eta_A$ shown in Figure~\ref{eta-beta}. The curve has been adjusted downwards by 0.29~dB (6.5\%) preserving its shape, to fit well with the Cyg~A determination. The observations provide strong confirmation that GRASP-10 has correctly computed the {\it{shape}} of the variation of $\eta_A$ with frequency, even if the absolute value requires some adjustment. The values of $\eta_A$ using ray tracing are a good fit to the measurements without any adjustment. We discuss the discrepancy between experimental and GRASP-10 values in Section~\ref{disc}.
The good agreement between the measured value of $\eta_A$ and calculations using GRASP-10 and using ray-tracing gives us strong confidence in the results. We adopt the adjusted GRASP-10 result, the smooth curve in Figure~\ref{eta-meas}, in processing the data from the GMIMS survey. We estimate the error in $\eta_A$ to be ${\pm}3\%$. The relative error in values of $\eta_A$ between any two frequencies in the band is less than this, about 1\%.
A strong ripple in $\eta_A$ is very obvious in Figure~\ref{eta-meas}. This is almost certainly an interaction of the feed with the reflector: the incident signal is not completely absorbed by the feed, and some is scattered from the feed and, after reflection from the vertex of the reflector, a delayed signal interferes with the prompt signal to form the ripple. The frequency separation of maxima is exactly what can be predicted from the focal length of the telescope.
The measured beamwidth of the telescope also shows a ripple with frequency (see Figure~1 of \citealp{woll10a}). The aperture efficiency is high at frequencies where the beamwidth is low and vice-versa, as expected. A change of $a\%$ in beamwidth will result in a change of $2a\%$ in gain and aperture efficiency, and this is (roughly) borne out by the measurements. Similar effects are reported by \citet{popp08}.
\citet{silv49} gives an expression for the reflection coefficient of the reflector-feed combination
\begin{equation}
{|\Gamma_r|}={\frac{{G_f}\thinspace{\lambda}}{4{\pi}f}} \thinspace,
\label{gamma_refl}
\end{equation}
where $G_f$ is the gain of the feed and $f$ is the focal length of the reflector. \citet{morr78} shows that the peak-to-peak magnitude of the ripple in the power seen by the receiver will be
\begin{equation}
{\frac{{\Delta}P}{P}}={4{\gamma}{\Gamma_r}} \thinspace,
\label{ripple}
\end{equation}
where $\gamma$, the scattering factor, describes scattering by the feed. \citet{morr78} evaluates $\gamma$ for a ${\rm{TE_{11}}}$ waveguide feed with a flange at its aperture. This is close to the feed on the Galt Telescope, but not exactly the situation. We have some difficulty deciding on the value of $\gamma$ because there are other things near the focus of the telescope that will aggravate the scattering, for example the feed-support structure and the receiver box. For a point source in the main beam (relevant to the data in Figure~\ref{eta-meas}) we take ${\gamma}\approx{1}$.
Equations \ref{gamma_refl} and \ref{ripple} predict that the ripple will be strongest at the low frequency end of the band, and the measurements in Figure~\ref{eta-meas} show this. Equation \ref{gamma_refl} is dominated by the wavelength, longest at the low end of the band, and feed gain is higher there than at band center. Both decrease to higher frequencies. However, in detail the match of measurement and prediction is not exact. At 1300~MHz equation~\ref{ripple} predicts ${\frac{{\Delta}P}{P}}={0.07}$; the measured value is 0.11. At 1500~MHz the predicted value is 0.05 and the measured value 0.02.
\citet{morr78} also points out that bandpass ripple will arise from extended emission in the main beam and from spillover radiation. The factor $\gamma$ in those two cases is different, and differs from the point-source value. We cannot confirm this from measurements, but note that amplitude ripples with the same characteristic frequency structure are certainly present in the GMIMS data, particularly at the low-frequency end of the passband.
\section{Spillover and Ground Radiation}
\label{spill}
Emission from the ground is a major factor in very sensitive radio astronomy systems because it contributes to system noise, and is often the largest contributor after the receiver itself. The principal route by which ground emission reaches the receiver is by radiating directly into the feed, over the edge of the reflector, through the spillover sidelobes. The levels and forms of the spillover sidelobes are largely determined by the design of the feed: spillover radiation is feed radiation. The edge of the reflector is at an angle ${\sim}80^{\circ}$ from the feed boresight, and the spillover region extends from here past $90^{\circ}$ to an angle where the feed no longer has significant response. One can hardly expect that radiation at these angles, far from the axis of a circular waveguide feed, will have good polarization characteristics. Spillover levels will also be influenced by diffraction at the edge of the reflector. Again, diffraction from an edge is a very polarization-sensitive process \citep{bach86}. Ground radiation is scattered from the feed-support struts into the aperture and enters the receiver \citep{ande91,land91}. The struts are long, relatively thin structures, and one would expect their scattering properties to be polarization dependent. For all these reasons, the paths through which ground emission enters the reflector are all likely to have poor polarization properties, which means that they are highly likely to convert the ground emission into spurious polarized signal.
Any discussion of the interaction of the radiation pattern with the ground is complicated by the fact that the ground radiation is itself partly polarized. We first discuss the properties of the ground as a polarized emitter and then consider the error introduced by ignoring the polarization and assuming that the ground signal is unpolarized.
\subsection{The Polarization of Ground Emission}
\label{ground-pol}
At frequencies around 1.5~GHz the ground is not a perfect absorber, it is a lossy dielectric, and its reflection and emission properties are polarization dependent. This is a well known effect at meter wavelengths where horizontal polarization is usually used for communications applications because such signals reflect strongly from the ground at grazing incidence and propagate reliably over large distances. In this discussion `horizontal' and `vertical' have particular meanings. The horizontally polarized component is normal to the propagation direction and parallel to the ground surface; the vertically polarized component is normal to the propagation direction and in the plane that is perpendicular to the ground. Note that these definitions are unique to each sidelobe of the telescope because they involve the look direction of that sidelobe.
We adopt a value of ground permittivity, ${\epsilon_r}={7}$, appropriate for 1.5~GHz \citep{hall79,itu00}. We know that the ground is an emitter at our frequencies, so soil conductivity cannot be zero, but it is small under the dry soil conditions around the Observatory, typically ${\kappa}={0.001}\thinspace{\rm{S}}\thinspace{\rm{m}}^{-1}$. Initially we set soil conductivity to zero and assume that permeability is~1.0. The reflection coefficients $r_v$ and $r_h$ for vertical and horizontal polarizations are then
\begin{subequations}
\begin{align}
{r_v}&= \frac{\epsilon_r {\rm cos}{\theta_i}-\sqrt{\epsilon_r-{\rm sin}^2\theta_i} }
{\epsilon_r{\rm{cos}}\theta_i+\sqrt{\epsilon_r -{\rm{sin}}^2\theta_i}} \thinspace,\\
{r_h}&= \frac{{\rm{cos}}\theta_i-\sqrt{{\epsilon_r}-{\rm{sin}}^2\theta_i}}
{{\rm{cos}}\theta_i+\sqrt{{\epsilon_r}-{\rm{sin}}^2\theta_i}} \thinspace,
\end{align}
\end{subequations}
where ${\theta}_i$ is the angle of incidence. $r_v$ and $r_h$ are ratios of reflected to incident field amplitudes. The reflectivities ${R_v}={r_v}^2$ and ${R_h}={r_h}^2$, which we will use later, are the ratios of reflected to incident powers. For small values of conductivity, to good approximation, the brightness temperatures of the ground in the two polarizations are
\begin{subequations}
\begin{align}
{T_v}&={(1-{{r_v}^2})T_p}\label{T_v} \thinspace,\\
{T_h}&={(1-{{r_h}^2})T_p}\label{T_h} \thinspace,
\end{align}
\end{subequations}
where $T_p$ is the physical temperature of the ground, which we take to be 300~K. We plot these values in Figure~\ref{groundtemp} (together with other
quantities that will be discussed below).
There are several reasons to expect that the vertical and horizontal components of the ground emission are correlated. First, consider the emission process. A thermal signal is generated in lossy material below the ground surface as an unpolarized wave. This propagates towards the surface and it is the direction-dependent transmissivity/reflectivity of the ground/air interface that leads to the apparent polarization of the ground signal. The ground emission signal will be partly polarized, with polarized brightness $T_p$ and unpolarized brightness $T_u$, leading to fractional polarization
\begin{equation}\label{f_p}
{f_p}={\frac{T_p}{{T_p}+{T_u}}}={\frac{{T_v}-{T_h}}{{T_v}+{T_h}}} \thinspace.
\end{equation}
Equation \ref{f_p} shows that ${f_p}=0$ at vertical incidence and ${f_p}=1$ at grazing incidence. Ground emission from directly below the telescope is unpolarized, ground emission from the horizon is totally polarized, and at intermediate zenith angles is partially polarized. $T_v$ is always greater than $T_h$, so the ground emission is always partially vertically polarized. The apparent brightness temperature of the ground then has a dependence on angle of incidence that is intermediate between the two values suggested by equations \ref{T_v} and \ref{T_h}.
Furthermore, the ground surface is not smooth, and on rough ground the distinction between vertical and horizontal polarization is blurred, since the antenna sidelobe will intersect a rough ground surface at a range of incidence angles, not at a single angle. The effects are treated quantitatively by \cite{kerr90}. Using data from satellite remote sensing of the Earth's surface, these authors define a polarization coupling factor, $q$, as
\begin{equation}
{q}={0.35(1-{e^{{-0.6{\sigma_G}^2}{\nu}}})} \thinspace,
\label{coupling}
\end{equation}
where $\sigma_G$ is the rms deviation of the ground (cm) and $\nu$ is the frequency (GHz). The reflectivities, $R_{\rm rv}$ and $R_{\rm rh}$, of the rough ground for vertical and horizontal polarization are
\begin{subequations}
\begin{align}
{R_{\rm rv}}&={q{R_h}+(1-q){R_v}{e}^{(-h{\rm{cos}}^2{\theta_i})}}\label{R_rv},\\
{R_{\rm rh}}&={q{R_v}+(1-q){R_h}{e}^{(-h{\rm{cos}}^2{\theta_i})}}\label{R_rh}.
\end{align}
\end{subequations}
Here ${h}={(2k{\sigma_G})^2}$ is a roughness factor, where $k={2{\pi}/\lambda}$ is the wave number. Equation~\ref{coupling} and equations \ref{R_rv} and \ref{R_rh} indicate that surface roughness leads quickly to coupling between vertical and horizontal polarizations. As the ground roughness increases, its reflectivity tends towards zero: it becomes a better absorber (and so a stronger emitter) and the effective ground temperature tends towards the physical ground temperature, about 300~K. These effects are illustrated in Figure~\ref{groundtemp}. This figure will apply to all frequencies for which our original assumption, ${{\epsilon}_r}=7$, remains valid; no significant change is expected across the band 1280 to 1750~MHz. A detailed discussion of Figure~\ref{groundtemp} can be found in Section~\ref{ground-obs}.
\begin{figure}
\centerline{\includegraphics[width=0.48\textwidth]{ground_temp.pdf}}
\caption{Ground temperature as a function of angle of incidence in vertical
and horizontal polarizations, shown for smooth ground, and rough ground with
rms deviation 0.5~cm and 2~cm. The star symbols show measured points (see Section~\ref{ground-obs}).
Note the measured point at ${\theta_i}={0^{\circ}}$. Values of ground
temperature are valid for the entire frequency range 1250 to 1750~MHz.}
\label{groundtemp}
\end{figure}
\subsection{Prediction of Spurious Polarization from Ground Emission}
\label{ground-predict}
We have computed the expected contribution of ground emission to $I$, $Q$, and $U$ based on the equations above. We adopted a greatly simplified profile for the ground, illustrated in Figure~\ref{groundprofile}, which approximates the topography around the telescope. We apply this radial profile to all azimuths, considering the ground to be symmetrical around the location of the telescope. The zone around the telescope to a viewing angle\footnote{The viewing angle is the angle, measured from the nadir, at which the emission enters the telescope.} of $69^{\circ}$ is flat with roughness ${\sigma_G}={0.5{\thinspace}{\rm{cm}}}$ (it consists of manicured lawns and paved surfaces). From ${\theta_i}={69^{\circ}}$ to ${\theta_i}={78^{\circ}}$ the ground is still flat, but rougher, with ${\sigma_G}={2{\thinspace}{\rm{cm}}}$. From ${\theta_i}={78^{\circ}}$ the ground begins to slope upwards at an angle of $10^{\circ}$ to approximate the hillsides, extending upwards at this angle to the mountaintops, at an elevation $10^{\circ}$ above the horizon. The roughness on the hillsides is also ${\sigma_G}={2{\thinspace}{\rm{cm}}}$.
\begin{figure}
\centerline{\includegraphics[width=0.48\textwidth]{model_setup.pdf}}
\caption{The brightness temperature of the ground used in computing the telescope response to ground radiation.
This is based on the hatched ground profile as shown at the bottom of the plot. The ground is flat to a viewing angle of $78^{\circ}$. The hill rises from there to a viewing angle of $100^{\circ}$. In the paved area ${{\sigma}_G}={0.5}~{\rm{cm}}$. In the unpaved area and on the hill
${{\sigma}_G}={2}~{\rm{cm}}$. The correspondence between viewing angle and distance from the telescope is shown by the labels at the top of the figure.}
\label{groundprofile}
\end{figure}
We have calculated the signals expected in $I$, $Q$ and $U$ based on the ground model in Figure~\ref{groundprofile} together with the equations of Section~\ref{ground-pol}. At the same time we have calculated the signals expected if the ground is an unpolarized emitter with an effective temperature of 240~K with no variation with angle of incidence. The results are shown in Figure~\ref{ground1} together with measured values (see Section~\ref{ground-obs} for a description of the measurement method and for a justification of the use of 240~K as the effective temperature of the ground; see Section~\ref{compare} for detailed discussion). $I$, $Q$, and $U$, arising from ground emission, are shown as a function of zenith angle as the telescope moves along the meridian from zenith angle $-40^{\circ}$ (north of the zenith, corresponding approximately to declination $90^{\circ}$), to zenith angle $70^{\circ}$ (close to the southern limit of telescope operation at declination $-30^{\circ}$). This is the observing track used for the GMIMS observations.
\subsection{Ground Contribution Determined from Observations}
\label{ground-obs}
Also shown in Figure~\ref{groundtemp} are measured ground temperatures at 1.4~GHz from \citet{ande91}, made at 1420~MHz with an antenna with a beam of ${\sim}1.5^{\circ}$. Angles of incidence were derived from a topographical map. The measurement at ${\theta_i}={0^{\circ}}$ was made with a handheld horn. These measurements imply that the effective ground temperature is $\sim$240~K, somewhere between the temperatures expected for vertical and horizontal polarization (as predicted in Section~\ref{ground-pol}).
We interpret Figure~\ref{groundtemp} in the following way. The measurement at vertical incidence, ${\theta_i}={0^{\circ}}$, is believable. It was made with a handheld
horn antenna pointed vertically downward at the ground. It is credible that ${\sigma_G}\lessapprox{0.5\thinspace{\rm{cm}}}$ in the small patch of ground under the horn. However, this measurement is mostly irrelevant, because the spillover lobes do not "see" the ground directly below the telescope. Emission from this part of the ground does leak through the mesh into the receiver, but this accounts for only $\sim$10\% of ground emission. The spillover lobes mostly intersect more distant ground, far from the pavement and manicured lawns that surround the telescope, and the ground is rougher out there. ${\sigma_G}\approx{2\thinspace{\rm{cm}}}$ is quite a good approximation to ground in that zone, although the oblique angles of incidence need to be considered. The rougher ground pushes the curves for vertical and horizontal polarization closer together. Furthermore, lines of sight never reach grazing incidence because the Observatory is in a bowl-shaped valley, and ${\theta_i}$ probably never exceeds ${80^{\circ}}$; the region relevant to spillover is ${30^{\circ}}\leq{\theta_i}\leq{80^{\circ}}$. The measurements lead us to believe that an average value of brightness temperature, the mean of vertical and horizontal, is quite appropriate, and the measured data support this view. Our estimate is consistent with the work of \citet{kalb10}, who quote a value for the albedo of the ground at 1420~MHz of 0.2, yielding a ground brightness temperature of $\sim$240~K.
\begin{figure}
\centerline{
\includegraphics[width=0.42\textwidth]
{ground_profile_stacked.pdf}}
\caption{Computed contribution at 1400~MHz of ground radiation to Stokes
parameters $I$, $Q$, and $U$ as a function of elevation. The solid lines show
computed contributions if the ground emission is partially polarized, as
discussed in Section~\ref{ground-pol}. The dashed lines show calculations
assuming an unpolarized ground at an effective temperature of 240~K. The dotted
curves are derived from measurements (see Section~\ref{ground-obs}).}
\label{ground1}
\end{figure}
The GMIMS project, described in Section~\ref{intro}, provides us with measurements that we can compare with the results of our simulations. The observations were made using a meridian-nodding mode, where the Galt Telescope moved vertically at a rate of approximately $1^\circ$ per minute, covering the declination range $-30^{\circ}$ to $+87^{\circ}$. Each such telescope track is referred to as a ``scan''. The equivalent zenith angle range for these ``scans'' is $79.5^{\circ}$ to the South, through zenith, to $40.5^{\circ}$ to the north. Earth rotation during a declination scan caused each observing path to be a diagonal line in the equatorial coordinate system. Successive tracks created individual diagonal lines until the entire sky was mapped with crossing tracks, with ${\sim}5{\times}10^5$ crossing points. After calibration against sources of known flux density the many crossing scans were reconciled using a ``basketweaving'' routine which iteratively deduced the best-fit zero level for each scan.
The output from this process was an image of the polarized signal from the sky plus the emission received from the ground and from the atmosphere. These contaminating signals were isolated by averaging across right ascension, for each declination, all signals recorded for the interval between right ascensions $04^{\rm h}$ and $08^{\rm h}$, an area of low sky brightness. It is safe to assume that the polarized signals are relatively low and the polarization angle will take almost all possible values, so the sky signal will average to zero over this large area, leaving only the contaminating signals. This process yielded curves of the spurious $Q$ and $U$ contributions produced by ground emission plus atmospheric emission as a function of elevation angle.
The total intensity (Stokes $I$) emission from the ground and atmosphere was estimated by stacking all signals in the same range of right ascension and taking the lower envelope of the measured signal. The known atmospheric contribution (see Section~\ref{atmos}) was subtracted to leave the ground contribution profile as a function of declination (elevation).
\subsection{Comparison of Calculated and Measured Ground Contributions}
\label{compare}
Figure~\ref{ground1} demonstrates that the spurious $Q$ and $U$ signals that arise from ground emission can reach magnitudes of 0.2 to 0.3~K. These values are significant, considering that the brightest polarized emission at 1.5~GHz is only $\sim$0.5~K. The spurious signals are lowest when the telescope is operating near the zenith; there is a high degree of symmetry in the radiation patterns, especially in $Q/I$ and $U/I$ (see Figure~\ref{rp}) and positive and negative responses fall on the ground in roughly equal measure. Examination of Figure~\ref{ground1} shows that there is relatively little difference between the contributions calculated on the basis of a partially polarized ground and on the basis of an unpolarized ground with an effective temperature of 240~K. The greatest effect of the dielectric properties of the ground is to reduce its effective temperature below its physical temperature, without significant change to the polarization properties of the ground radiation in the spurious $Q$ and $U$ signals. The measured ground contribution in $I$ is much lower than the calculated contribution. This is the result of the basketweaving process, which removes any uniform background level. This effect will have to be corrected in the GMIMS dataset, but is not a concern here.
The ground $I$ signal at the zenith is a local maximum, surrounded by minima $\sim$1~K lower at ${\pm}20^{\circ}$ on either side. This is easily understood in an intuitive way. At the zenith, a ring of spillover sidelobes lies uniformly on the ground. As the telescope is tipped, one side of the ring lifts off the ground while the contribution from the opposite side is relatively unchanged, and the ground contribution drops. Interestingly, this zenith `bump' was reproduced in the model of the unpolarized ground, but the bump did not appear with the polarized ground until the model of the polarized ground (Figure~\ref{groundprofile}) included a realistic characterization of the mountainsides. The most rapid changes in $Q$ and $U$ occur in the range of zenith angle between ${\sim}12^{\circ}$ and ${\sim}25^{\circ}$ (on both sides of the zenith). This is the result of interaction of the scatter cones with the emission from the surrounding mountains. The scatter cone maxima are at $68^{\circ}$ from the telescope boresight, and will intersect the top of the mountains (at elevation $10^{\circ}$) at zenith angle ${\sim}12^{\circ}$.
\begin{figure*}
\centerline{
\includegraphics[width=0.72\textwidth,clip] {Stokes_I_Ground_v3.pdf}}
\centerline{
\includegraphics[width=0.72\textwidth,clip] {Stokes_Q_Ground_v3.pdf}}
\centerline{
\includegraphics[width=0.72\textwidth,clip] {Stokes_U_Ground_v3.pdf}}
\caption{Illustrating the interaction of the antenna radiation pattern with ground and atmospheric emission. The left panels show $I$, $Q/I$, and $U/I$ in a horizontal coordinate system with the telescope directed at $40^{\circ}$ from the zenith (declination $9.5^{\circ}$) towards the south. The upper part of the radiation patterns sees the sky; the lower part sees the ground. The two dashed lines indicate, from upper to lower, the top and the bottom of the hill. The unit of $I$ is dB relative to isotropic and the color scale of the $Q/I$ and $U/I$ plots spans $-1$ to $+1$ as in Figure~\ref{rp}. The right panels show the brightness temperature of the surroundings as a function of zenith angle as in Figure~\ref{groundprofile}.
\label{ground2}}
\end{figure*}
As an illustration of the interaction of the radiation pattern with the surroundings of the antenna (ground and atmosphere) we have prepared Figure~\ref{ground2}. The radiation patterns (in $I$, $Q/I$, and $U/I$) are shown for the telescope directed at zenith angle $40^{\circ}$. This is the zenith angle where the $Q/I$ and $U/I$ signals are quite strong (Figure~\ref{ground1}). Looking at the $I$ pattern in Figure~\ref{ground2} we see that the spillover lobes will have a major influence on the ground contribution, simply because of their large solid angles. The spillover lobes in $Q/I$ and $U/I$ are broken up into alternating positive and negative responses, so the ground contributions tend to average out. The larger fraction of the ground contribution comes from the large lobes, uniform in sign, positive in the case of $Q/I$ and negative in $U/I$, that are associated with the scatter cones.
How successful have we been in reproducing the measured ground contributions to $Q$ and $U$? Figure~\ref{ground1} shows that we have the sense of the variation with zenith angle approximately right, but that there are still significant differences, especially in $Q$, and especially when pointing to the south (zenith angles $0^{\circ}$ to $80^{\circ}$). Neither the unpolarized ground model nor the polarized model comes close to the measurement. Either our model of the ground is wrong, or our model of the strut that generates the scatter cone in this direction is wrong. This is the top strut (Figure~\ref{shadows}) the one that has no cables. Experiments showed that the ground model did affect the ground profiles in Figure~\ref{ground1}. A model based on the actual topography around the telescope might be useful, but there is a building close to the telescope, just north of it, and it is not clear what impact this has. Such an investigation is beyond the scope of this paper. If we want to rely on calculation of the spurious $Q$ and $U$ ground
contributions to make corrections to observations we would need to reach an accuracy of about $20\thinspace{\rm{mK}}$. We are at least a factor of 5 from achieving this goal.
\section{Atmospheric Emission}
\label{atmos}
Atmospheric emission contributes to system noise, especially at short wavelengths, entering the telescope through the main beam and the sidelobes. Emission from the atmosphere is unpolarized but can be converted into apparently polarized radiation by the instrumental polarization. We have calculated the contribution to polarized emission using the same routines as we applied to the problem of ground emission.
For this calculation the sky was represented as a thermal emitter whose temperature was dependent on direction. We used the equations given by \citet{gibb86} to calculate the atmospheric attenuation in dB/km. This was converted to a noise temperature using a scale height of 6 km and an average temperature of 260~K \citep{alln89}. The temperature at the zenith at 1.4 GHz is 2.0~K, and the variation with zenith angle, $z$, is taken as sec$\thinspace(z)$, the variation expected of a plane parallel slab of absorbing atmosphere.
The results are shown in Figure~\ref{atmo}. The variation of $Q$ and $U$ as a function of $z$ is similar in form to the variation of ground radiation, but inverted and much smaller in amplitude (the inversion can be seen clearly by comparing Figure~\ref{ground1} and Figure~\ref{atmo}). We can easily understand these effects. The effective temperature of the atmosphere is much lower than that of the ground, and the atmospheric contribution is correspondingly smaller (see below for a more quantitative discussion). The resemblance of form indicates that, just as the scatter cones dominate the appearance of the ground emission profile, so they dominate the atmospheric emission profile. The inversion of the shape between the ground and atmosphere profiles can be understood this way: as $z$ increases through $12^{\circ}$ the scatter cones lift off the ground, and must then encounter the brightest parts of the atmosphere; as the ground ontribution drops the atmospheric contribution rises.
At first sight the spurious polarization arising from atmospheric emission seems negligibly small, but it is worthwhile to go deeper into the data. If the ground temperature is 240~K and the atmospheric contribution is 2~K (the zenith value at 1.4~GHz) then we would expect the atmospheric contribution to be 1/120 of the ground contribution. The actual ratio of the two contributions is about 40 (from a comparison of the change in the $Q$ contribution at zenith angles ${\pm}20^{\circ}$ in Figure~\ref{ground1} and Figure~\ref{atmo}), which implies that atmospheric emission must be about 6~K, three times the zenith value. In fact sec$(z)$ reaches a value of 3 at ${z}{\approx}{70^{\circ}}$, so there is a ring of ``warmer'' atmosphere around the telescope at large zenith angles. If the scatter cones are interacting with atmospheric emission this far from the zenith (just above the surrounding hills) then our simulation results can be understood. Spurious polarization arising from atmospheric emission is unlikely to be a significant factor at $\sim$1.5~GHz, but it could be a factor at higher frequencies where astronomical polarized signal levels are lower and atmospheric emission becomes more intense.
\begin{figure}
\includegraphics[width=0.45\textwidth]
{Atm_profile_stacked.pdf}
\caption{Computed contribution at 1400~MHz of atmospheric emission for Stokes
parameters $I$, $Q$, $U$, and $V$ as a function of elevation.}
\label{atmo}
\end{figure}
\section{Concluding Discussion}
\label{disc}
We have used GRASP-10 to calculate the properties of a large radio telescope, and have compared the results of those calculations with measurements wherever possible. The GRASP-10 evaluation of aperture efficiency of the Galt Telescope has successfully predicted the frequency dependence of that parameter, but the measured values are approximately 6.5\% lower than the calculated value. This difference amounts to an error of 0.3~dB in the calculation of the gain, which is $\sim{50}$~dB. While this is good accuracy for an engineering tool, radio astronomy hopes to do better.
We also used simpler tools, ray tracing and geometrical optics, to examine aperture efficiency, and demonstrated that they too can provide useful accuracy and can provide useful insights into performance. We decomposed aperture efficiency into constituent efficiencies amenable to calculation on the basis of geometrical optics assumptions. The dominant effects in determining aperture efficiency are the illumination efficiency and blockage by the feed-support struts of the spherical wave from the feed: together they limit overall aperture efficiency to a maximum value of about 0.65. The ray tracing code was of particular value in estimating spherical wave blockage because it was designed to handle the cigar-shape of the support legs. Spherical wave blockage is severe because the axes of the feed-support struts are only $34^{\circ}$ from the telescope axis, and a substantial part of the aperture is shadowed by the
struts.
The principal weakness of GRASP-10 in this application is its inability to model the struts precisely. GRASP-10 models the struts as straight metal cylinders, while in fact they are cigar-shaped fiber glass and two of them carry metal sheathed cables. We modelled the struts as metal cylinders of diameter 25.4~cm. If we had increased this by 3\%, a mere 8~mm, the gain would have decreased by about twice this, 6\%, bringing the calculation close to the measured result.
We have investigated the properties of a radio telescope over a very wide band, 1250 to 1750 MHz. The work that we describe here has established the absolute calibration of a wideband dataset. Previous work at these frequencies has been able to depend on absolutely calibrated horn measurements (e.g. the survey of \citealp{reic82} traced its calibration to \citealp{webs74}) but such measurements are confined to the narrow frequency bands allocated to radio astronomy. We have established an absolute scale of brightness temperature for the GMIMS survey. The overall error in this scale is 3\% but the relative error between any two frequencies in the band is less than this, about 1\%.
We have used GRASP-10 to investigate instrumental polarization in the far sidelobes. Signals from the ground (and to a lesser extent from the atmosphere) that are inherently unpolarized or slightly polarized can be converted to apparently strongly polarized signals. Our analysis has shown that the principal routes of entry of these spurious signals are the spillover lobes and the scatter cones that are generated by the feed-support struts. The most rapid changes in spurious polarization occur when the scatter cone sidelobes pass from the sky to the ground, or vice versa. For maximum stability of instrumental polarization, observations should be planned to avoid telescope pointings where the scatter cones come close to the horizon. Our predictions based on GRASP-10 agree in intensity with the measured results, but there are differences in detail. Again, the less than perfect representation in GRASP-10 of the feed-support struts is a problem. We also have the fact that our ground model is very simple; it does not take into account surrounding buildings, some of which are quite close to the telescope.
The ground is not a perfect absorber; it is a lossy dielectric. As a consequence its reflection and emission properties are polarization dependent. We have presented an analysis based on empirical data on ground polarization from studies using satellite-borne microwave radiometers. We conclude that the inherent polarization of ground emission is low at the angles that matter in the spillover lobes of the telescope. In fact our prediction of spurious polarization based on a polarized ground is not very different from our
prediction assuming an unpolarized ground, and in a comparison with the measured data it is hard to say which route gives the better prediction. We conclude that the best estimate of ground brightness temperature at $\sim$1.5~GHz is about 240~K independent of angle of incidence. \citet{kalb10} derive a very similar value at 1.42~GHz.
This conclusion does not conflict with observations that show that the Moon and planets are polarized at these frequencies (e.g. \citealp{zhan12,heil63}). A telescope observing the Moon or the planets samples only one angle of incidence using its main beam. In our case, however, the sidelobes of the Galt Telescope are sampling many different angles of incidence through many sidelobes. All these contributions are averaged and the ground appears to be almost unpolarized as seen by the sidelobes.
\acknowledgments
The Dominion Radio Astrophysical Observatory is a National Facility operated by the National Research Council Canada. The Global Magneto-Ionic Medium Survey is a Canadian project with international partners. The participation of KAD and MW in this research was supported by the Natural Sciences and Engineering Research Council (NSERC). Without the skilled work of Rob Messing and Ev Sheehan in maintaining the Galt Telescope this research would not have been possible. We thank Lloyd Higgs for his outstanding software, including the
ray-tracing program.
\bibliographystyle{aa}
|
\section{Introduction}
For discrete-alphabet signals, the Shannon entropy function $H(X)$ and the entropy rate $\bar{H}({\mathbb{\mathbf{X}}})=\lim_{n\to\infty}H(X_n|X^{n-1})$ measure the complexity of a random variable $X$ and a stationary stochastic process ${\mathbb{\mathbf{X}}}=\{X_i\}$, respectively. Both of these measures are closely connected to the minimum number of bits per symbol required for representing stochastic sources \cite{cover} and can also be thought of as measures of signal structure. However, when we shift from discrete alphabet to continuous alphabet, both the entropy and the entropy rate become infinite. Instead, for analog signals, the notion of information dimension (ID) introduced by R\'enyi \cite{Renyi:59} provides a framework that can be used to quantify signal structure.
To illustrate what is meant for an analog process to be structured, consider a stationary memoryless (i.e., independent and identically distributed or i.i.d.) process ${\mathbb{\mathbf{X}}}=\{X_i\}_{i=0}^{\infty}$ such that $X_i\sim (1-p)\delta_0+p f_c$, where $f_c$ denotes the probability density function (pdf) of an absolutely continuous distribution and $\delta_0$ denotes the Dirac measure with an atom at $0$. In other words, for each $i$, with probability $p\in[0,1]$, $X_i$ is exactly equal to zero; otherwise, it is drawn from $f_c$. By the strong law of large numbers, for large values of blocklength $n$, with probability approaching one, a block $X^n$ generated by this source contains around $n(1-p)$ entries equal to zero, and the rest of the entries are real numbers in the domain of $f_c$. To describe $X^n$ with a certain precision, for zero entries, it suffices to describe their locations. The number of bits required for this description does not depend on the reconstruction quality. However, for the remaining approximately $np$ elements of $X^n$, it is known from rate-distortion theory that the required number of bits grows with the desired reconstruction quality. This intuitively suggests that $p$, which controls the number of non-zero elements in $X^n$, is a fundamental quantity related to the complexity and structure of $X^n$. This intuition is accurately captured by the ID of this source which can be shown to be equal to $p$ \cite{Renyi:59}. In fact, $\delta_0$ can be changed to any discrete probability distribution with finite entropy and the result will not change since the R\'enyi ID of a discrete source is 0.
A further significance of the ID as a measure of structure is its relationship to the problem of compressed sensing. Consider the problem of recovering a signal $X_o^n$ from under-determined measurements $Y^m=AX_o^n$, where $m<n$. It is known that if the input signal $X_o^n$ is sparse, or in general ``structured'', it can be accurately recovered from the measurements, even if $m$ is far fewer than $n$ \cite{Donoho:06,CandesT:06,RichModelbasedCS, ChRePaWi10, VeMaBl02, DoKaMe06}. For stationary memoryless processes, under some mild conditions on the distribution, the R\'enyi ID of the first order marginal distribution of the source characterizes the fundamental limits of compressed sensing, \ie the minimum number of measurements required for asymptotically almost lossless recovery \cite{WuV:10}. The notion of the R\'enyi ID is extended to stationary processes in \cite{JalaliP:14-arxiv}, where it is proved that there is a direct relationship between the ID of a stationary process and the number of random linear measurements required for its universal recovery.
While the aforementioned results give an operational meaning to the ID of a signal, evaluating the ID of a stationary process is in general difficult. Kawabata and Dembo defined the rate-distortion dimension (RDD) of i.i.d. random variables (or vectors) based on the rate-distortion trade-off in the asymptoticly low distortion regime \cite{KawabataD:94}. They proved that for a random variable, its (upper and lower) RDD is equal to its (upper and lower) ID.
The main contribution of this paper is to extend the notion of RDD to analog stationary processes, and to prove that, under some regularity conditions, the RDD of a stationary process is equal to its ID, defined in \cite{JalaliP:14-arxiv}. This provides an extension of the result of Kawabata and Dembo to stochastic processes, and thereby provides a computationally feasible way of finding the ID of a stochastic process. In order to illustrate this, we compute the RDD of piecewise-constant stochastic processes, which are widely used to model many natural signals such as images. We derive upper and lower bounds on the rate-distortion functions of such signals, and use these bounds to evaluate the RDD and hence, the ID of such processes. Furthermore, our results in \cite{RJEP:16-arxiv} suggest that the RDD of a stochastic process is closely related to the fundamental limits of compressed sensing for the process, and hence RDD and ID can be thought of as measures of structure/complexity for arbitrary stationary stochastic processes.
The organization of the paper is as follows. Section \ref{sec:RDD} defines and examines the properties of ID and RDD. Section \ref{sec:RDD-ID} contains our main result which establishes a connection between the ID and the RDD of stochastic processes. Upper and lower bounds on the rate-distortion region of the piecewise constant source modeled by a first-order Markov process are provided in Section \ref{sec:pwc_source} to illustrate the relationship between RDD and ID. Section \ref{sec:conc} concludes the paper.
\subsection{Notation}
Capital letters like $X$ and $Y$ represent random variables.
For $x\in\mathds{R}$, $\lceil x \rceil$ ($\lfloor x \rfloor$) represents the smallest (largest) integer larger (smaller) than $x$. For $b\in\mathds{N}^+$, $[x]_b$ denotes the $b$-bit approximation of $x$, \ie for $x=\lfloor x \rfloor+\sum_{i=1}^{\infty}(x)_i2^{-i}$, $(x)_i\in\{0,1\}$, $[x]_b=\lfloor x \rfloor+\sum_{i=1}^{b}(x)_i2^{-i}.$ Also, let $\langle x\rangle_b$ be defined as $\langle x\rangle_b={\lfloor bx\rfloor \over b}.$ For $x^n\in\mathds{R}^n$, $[x^n]_b$ and $\langle x^n\rangle_b$ are defined as $([x_1]_b,\ldots,[x_n]_b)$ and $(\langle x_1\rangle_b,\ldots,\langle x_n\rangle_b)$, respectively.
Throughout the paper, $\log$ refers to the logarithm in base 2 .
\section{Background}\label{sec:RDD}
In this section, we provide formal definitions of ID and RDD and an overview of the literature.
\begin{definition}[R\'enyi information dimension \cite{Renyi:59}]\label{eq:def1}
The R\'enyi upper and lower IDs of an analog random variable $X$ are defined as
\[
\bar{d}(X)=\limsup_{b\to\infty} {H(\langle X\rangle_b)\over \log b},
\]
and
$
\underline{d}(X)=\liminf_{b\to\infty} {H(\langle X\rangle_b)\over \log b},
$
respectively. If the two limits coincide, the R\'enyi ID of $X$ is defined as $d(X)=\bar{d}(X)=\underline{d}(X)$.
\end{definition}
Definition \ref{eq:def1} can also be applied to analog vectors. For instance, for a random vector $X^n$, $\bar{d}(X^n)=\limsup_{b\to\infty} {H(\langle X^n\rangle_b)\over \log b}.$
While the above definition of the R\'enyi ID is in terms of the entropy of the $b$-level quantized version of $X$ normalized by the number of bits required for binary representation of it, $\log b$, as proved in Proposition 2 of \cite{WuV:10}, it can equivalently be defined in terms of the entropy of the $b$-bit quantized version of $X$, $[X]_b$, normalized by $b$, \ie
\[
\bar{d}(X)=\limsup_{b\to\infty} {H([X]_b)\over b},
\]
and $\underline{d}(X)=\liminf_{b\to\infty} {H([X]_b)\over b}.$
The notion of R\'enyi ID for random variables or vectors was extended in \cite{JalaliP:14-arxiv} to define the ID of analog stationary processes.
\begin{definition}[ID of a stationary process \cite{JalaliP:14-arxiv}]
The $k$-th order upper and lower IDs of stationary process ${{\mathbb{\mathbf{X}}}}=\{X_i\}_{i=-\infty}^{\infty}$ are defined as
\[
\bar{d}_k({{\mathbb{\mathbf{X}}}})=\limsup_{b\to \infty} {1\over b}H([X_{k+1}]_b|[X^k]_b),
\]
and $ \underline{d}_k({{\mathbb{\mathbf{X}}}})=\liminf_{b\to \infty} {1\over b}H([X_{k+1}]_b|[X^k]_b),$
respectively. The upper and lower ID of the process ${{\mathbb{\mathbf{X}}}}$ are defined as
\[
\bar{d}_o({{\mathbb{\mathbf{X}}}})=\lim_{k\to\infty}\bar{d}_k({{\mathbb{\mathbf{X}}}})
\]
and $\underline{ d}_o({{\mathbb{\mathbf{X}}}})=\lim_{k\to\infty}\underline{d}_k({\mathbb{\mathbf{X}}}),$
respectively, when the limits exist. If $\bar{d}_o({{\mathbb{\mathbf{X}}}})=\underline{d}_o({{\mathbb{\mathbf{X}}}})$, the ID of process $\mathbb{\mathbf X}$, ${d}_o({{\mathbb{\mathbf{X}}}})$, is defined as ${d}_o({{\mathbb{\mathbf{X}}}})=\bar{d}_o({{\mathbb{\mathbf{X}}}})=\underline{d}_o({{\mathbb{\mathbf{X}}}})$.
\end{definition}
As proved in \cite{JalaliP:14-arxiv}, both $\bar{d}_k({\mathbb{\mathbf{X}}})$ and $\underline{d}_k({\mathbb{\mathbf{X}}})$ are both non-negative decreasing sequences in $k$. Hence, if they are also bounded, which is the case for instance for bounded sources, their limits as $k\to\infty$ also exist.
For a stationary memoryless process ${{\mathbb{\mathbf{X}}}}=\{X_i\}_{i=-\infty}^{\infty}$, this definition coincides with that of R\'enyi\rq{}s ID of the first-order marginal distribution of the process ${\mathbb{\mathbf{X}}}$. That is $\bar{d}_o({{\mathbb{\mathbf{X}}}})=\bar{d}(X_1)$ and $\underline{d}_o({{\mathbb{\mathbf{X}}}})=\underline{d}(X_1)$. For sources with memory, taking the limit as the memory parameter $k$ grows to infinity allows $d_o({\mathbb{\mathbf{X}}})$ to capture the overall structure that is present in an analog stationary process. It can be proved that $d_o({\mathbb{\mathbf{X}}})\leq 1$, for all bounded stationary processes, and if the stationary process ${\mathbb{\mathbf{X}}}$ is structured, $d_o({\mathbb{\mathbf{X}}})$ is strictly smaller than one \cite{JalaliP:14-arxiv}.
Under some mild conditions on the distribution, \cite{WuV:10} proves that the R\'enyi ID of the first-order marginal distribution of a stationary memoryless process characterizes the fundamental limits of its compressed sensing. In other words, given a stationary memoryless process ${{\mathbb{\mathbf{X}}}}$, asymptotically, as the blocklength $n$ grows to infinity, the minimum number of linear projections ($m$) normalized by the blocklength ($n$) that is required for recovering source $X^n$ is shown to be equal to $d(X_1)$. In \cite{JalaliP:14-arxiv}, it is shown that, asymptotically, slightly more than $n\bar{d}_o({{\mathbb{\mathbf{X}}}})$ random linear projections suffice for \emph{universal} recovery of $X^n$ generated by any stationary process that satisfies some mixing conditions. These results provide an operational interpretation of the ID of a random process.
The rate-distortion function of a stationary source measures the minimum number of bits per source symbol required for achieving a given reconstruction quality. In some cases, as the reconstruction becomes finer, the behavior of the rate-distortion function is connected to the level of structuredness of the source process and also to its ID mentioned earlier. In the rest of this section, we review the known results on these connections.
Consider a metric space $(\mathds{R}^k,\rho)$, and random vector $X^k$. The standard rate-distortion function \cite{cover} of vector $X^k$ under distortion measure
$d(x^k,{\hat{x}}^k)=\rho(x^k,{\hat{x}}^k)^r$, where $r>0$, is defined as
\[
R_r(X^k,D)=\inf_{ \mathrm{E}[d(X^k,{\hat{X}}^k)]\leq D}I(X^k;{\hat{X}}^k).
\]
\begin{definition}[Rate-distortion dimension (RDD) of a random vector \cite{KawabataD:94}]
The upper and lower RDDs of $X^k$ are defined as
\[
\overline{\dim}_R(X^k)=r\limsup_{D\to0}{R_r(X^k,D)\over \log{1\over D}},
\]
and
$\underline{\dim}_R(X^k)=r\liminf_{D\to0}{R_r(X^k,D)\over \log{1\over D}},
$
respectively. If $\overline{\dim}_R(X^k)=\underline{\dim}_R(X^k)$, the RDD of $X^n$ is defined as ${\dim}_R(X^k)=r\lim_{D\to0}{R_r(X^k,D)\over \log{1\over D}}$.
\end{definition}
The following theorem from \cite{KawabataD:94} establishes the connection between the R\'enyi ID of a random vector $X^k$ and its RDD, for a general distribution on $X^k$.
\begin{theorem}[Proposition 3.3 in \cite{KawabataD:94}] \label{thm:prop3-3}
Consider the metric space $(\mathds{R}^k,\rho)$, such that there exists $0<a_1\leq a_2<\infty$ for which $a_1\max_{i=1}^k|x_i-{\hat{x}}_i|\leq \rho(x^k,{\hat{x}}^k)\leq a_2\max_{i=1}^k|x_i-{\hat{x}}_i|,$
for all $x^k,{\hat{x}}^k\in\mathds{R}^k$. Then, for any distribution of $X^k$,
\[
\overline{\dim}_R(X^k)=\bar{d}(X^k),
\]
and
$\underline{\dim}_R(X^k)=\underline{d}(X^k),
$
where $\overline{\dim}_R(X^k)$, and $\underline{\dim}_R(X^k)$ denote the upper and lower RDD of $X^k$ under fidelity constraint $d(x^k,{\hat{x}}^k)=\rho(x^k,{\hat{x}}^k)^r$.
\end{theorem}
\section{Equivalence of RDD and ID for Analog Processes} \label{sec:RDD-ID}
This section provides the main result of this paper which extends the notion of RDD to stationary processes and establishes its connection of the ID of the process.
Consider an analog stationary process ${{\mathbb{\mathbf{X}}}}=\{X_i\}_{i=-\infty}^{\infty}$. The rate-distortion function $R({{\mathbb{\mathbf{X}}}},D)$ of the source ${{\mathbb{\mathbf{X}}}}$ under squared error distortion can be characterized as \cite{book:Berger,Gallager}
\[
R({{\mathbb{\mathbf{X}}}},D)=\lim_{m\to\infty} R^{(m)}({{\mathbb{\mathbf{X}}}},D),
\]
where
\[
R^{(m)}({{\mathbb{\mathbf{X}}}},D)=\inf_{\mathrm{E}[d_m(X^m,{\hat{X}}^m)]\leq D}{1\over m} I(X^m;{\hat{X}}^m)
\]
and
\begin{align}
d_m(x^m,{\hat{x}}^m)={1\over m}\|x^m-{\hat{x}}^m\|_2^2. \label{sq-err-distortion}
\end{align}
Note that with this distortion metric, we have $r=2$ and $R^{(m)}({{\mathbb{\mathbf{X}}}},D)= {1\over m}R_2({X}^m,D)$. It can also be shown that $\inf_{m} R^{(m)}({{\mathbb{\mathbf{X}}}},D)=R({{\mathbb{\mathbf{X}}}},D)$ \cite{Gallager}.
\begin{definition}[RDD of a stationary process]
The upper and lower RDDs of a stationary process ${\mathbb{\mathbf{X}}}$ is defined as
\[
\overline{\dim}_R({{\mathbb{\mathbf{X}}}})=2\limsup_{D\to0}{R({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}
\]
and
$\underline{\dim}_R({{\mathbb{\mathbf{X}}}})=2\liminf_{D\to0}{R({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}.$
If $\overline{\dim}_{R}({\mathbb{\mathbf{X}}})=\underline{\dim}_{R}({\mathbb{\mathbf{X}}})$, then the RDD of ${\mathbb{\mathbf{X}}}$ is defined as $\dim_{R}({\mathbb{\mathbf{X}}})=\overline{\dim}_{R}({\mathbb{\mathbf{X}}})=\underline{\dim}_{R}({\mathbb{\mathbf{X}}})$.
\end{definition}
The following theorem extends the equivalence of R\'enyi ID and RDD established in Theorem \ref{thm:prop3-3} for i.i.d. random vectors to stationary processes.
\begin{theorem}\label{thm:ID-eq-RDD}
For a stationary process ${\mathbb{\mathbf{X}}}=\{X_i\}_{i=-\infty}^{\infty}$, assume that $\lim_{D\to 0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}$ exists for all $m$. Then,
\[
{\dim}_{R}({\mathbb{\mathbf{X}}}) = \bar{d}_o({{\mathbb{\mathbf{X}}}}).
\]
\end{theorem}
The main ingredients of the proof of Theorem \ref{thm:ID-eq-RDD} are the following two lemmas.
\begin{lemma}\label{lemma:connect-UID-URDD}
For any stationary process ${\mathbb{\mathbf{X}}}$, we have
\[
\overline{\dim}_{R}({\mathbb{\mathbf{X}}}) \leq \bar{d}_o({{\mathbb{\mathbf{X}}}})\leq \inf_{m} 2\Big(\limsup_{D\to0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}\Big).
\]
\end{lemma}
\begin{lemma}\label{lemma:uniform_conv}
Assume that $\lim_{D\to 0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}$ exists for all $m$, and also there exists $\sigma_{\max}^2>0$, such that ${R^{(m)}({{\mathbb{\mathbf{X}}}},D)}$ uniformly converges to ${R({{\mathbb{\mathbf{X}}}},D)}$, for $D\in(0,\sigma_{\max}^2)$, as $m$ grows to infinity. Then, ${\dim}_{R}({\mathbb{\mathbf{X}}}) = \bar{d}_o({{\mathbb{\mathbf{X}}}}).$
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lemma:connect-UID-URDD}]
Given $k$, define distance measure $\rho_k$ such that for $x^k,{\hat{x}}^k\in\mathds{R}^k$, $\rho_k(x^k,{\hat{x}}^k)\triangleq \sqrt{kd_k(x^k,{\hat{x}}^k)}$ where $d_k(\cdot,\cdot)$ is defined in (\ref{sq-err-distortion}). Note that $(\mathds{R}^k,\rho_k)$ is a metric space. Furthermore, since $\max_{i=1}^k|x_i-{\hat{x}}_i|\leq \rho_k(x^k,{\hat{x}}^k)\leq \sqrt{k} \max_{i=1}^k|x_i-{\hat{x}}_i|$, from Theorem \ref{thm:prop3-3},
\[
2\limsup_{D\to0}{kR^{(k)}({{\mathbb{\mathbf{X}}}},{D\over k})\over \log{1\over D}}=\bar{d}(X^k).
\]
By a change of variable, $2\limsup_{D\to0}{kR^{(k)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}+\log{1\over k}}=\bar{d}(X^k),$
or
\[
2\limsup_{D\to0}{R^{(k)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}={1\over k}\bar{d}(X^k).
\]
Taking the limit of both sides as $k$ grows to infinity, and employing Lemma 2 from \cite{JalaliP:14-arxiv}, which shows that the upper ID of a process ${\mathbb{\mathbf{X}}}$ can alternatively be represented as
\[
\bar{d}_o({{\mathbb{\mathbf{X}}}})=\lim_{k\to\infty}{1\over k}\bigg(\limsup_{b\to\infty} {H([X^k]_b)\over b}\bigg),
\]
yields
\begin{align}
\lim_{k\to\infty}\bigg(2\limsup_{D\to0}{R^{(k)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}\bigg)&=\lim_{k\to\infty}{1\over k}\bar{d}(X^k)\nonumber\\
&=\bar{d}_o({{\mathbb{\mathbf{X}}}}).\label{eq:application-lemma-1}
\end{align}
Since $R^{(k)}({{\mathbb{\mathbf{X}}}},D)\geq \inf_mR^{(m)}({{\mathbb{\mathbf{X}}}},D)$, from \eqref{eq:application-lemma-1},
\begin{align*}
\bar{d}_o({{\mathbb{\mathbf{X}}}})&\geq \lim_{k\to\infty}\bigg(2\limsup_{D\to0}{\inf_m R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}\bigg)\nonumber\\
&\stackrel{(a)}{=} \lim_{k\to\infty}\bigg(2\limsup_{D\to0}{R({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}\bigg)=\overline{\dim}_{R}({\mathbb{\mathbf{X}}}),
\end{align*}
where (a) follows from the fact that $R({\mathbb{\mathbf{X}}},D)=\inf_{m} R^{(m)}({\mathbb{\mathbf{X}}},D)$ \cite{Gallager}. This proves the lower bound in the desired result.
To prove the upper bound, fix a positive integer $m\in\mathds{N}$. Any integer $k$ can be written as $k=sm+r$, where $r\in\{0,\ldots,m-1\}$. Since $kR^{(k)}({{\mathbb{\mathbf{X}}}},D)$ is a sub-additive sequence \cite{Gallager}, and $k=m+\ldots+m+r$, $kR^{(k)}({{\mathbb{\mathbf{X}}}},D)\leq sm R^{(m)}({{\mathbb{\mathbf{X}}}},D)+rR^{(r)}({{\mathbb{\mathbf{X}}}},D),$ it follows that
or
\begin{align}
R^{(k)}({{\mathbb{\mathbf{X}}}},D)\leq {sm\over k} R^{(m)}({{\mathbb{\mathbf{X}}}},D)+{r\over k}R^{(r)}({{\mathbb{\mathbf{X}}}},D).\label{eq:sub-additive}
\end{align}
Combining \eqref{eq:application-lemma-1} and \eqref{eq:sub-additive}, it follows that
\begin{align}
\bar{d}_o({{\mathbb{\mathbf{X}}}})\leq &\; 2\lim_{k\to\infty}\bigg(\limsup_{D\to0} {sm\over k} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}\bigg)\nonumber\\
&+2\lim_{k\to\infty}\bigg( \limsup_{D\to0}{r\over k}{R^{(r)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}\bigg)\nonumber\\
= &\; 2\lim_{k\to\infty}\Big({sm\over k}\Big)\bigg(\limsup_{D\to0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}\bigg)\nonumber\\
&+2\lim_{k\to\infty}\Big({r\over k}\Big)\bigg( \limsup_{D\to0}{R^{(r)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}\bigg)\nonumber\\
= &\; 2\bigg(\limsup_{D\to0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}\bigg).\label{eq:ub_wo_inf}
\end{align}
Since $m$ is selected arbitrarily, we can take the infimum of the right hand side of \eqref{eq:ub_wo_inf} and derive the desired result.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lemma:uniform_conv}]
By the lemma's assumption, $\overline{\dim}_{R}({\mathbb{\mathbf{X}}}) ={\dim}_{R}({\mathbb{\mathbf{X}}}) $; therefore, from Lemma \ref{lemma:connect-UID-URDD}, \begin{align}\label{eq:b1}
{\dim}_{R}({\mathbb{\mathbf{X}}}) \leq \bar{d}_o({{\mathbb{\mathbf{X}}}})\leq 2\Big(\lim_{D\to0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}\Big),
\end{align}
for all $m$. Given the uniform convergence assumption, for any $\epsilon>0$, there exists $m_{\epsilon}\in\mathds{N}$, such that for all $m>m_{\epsilon}$,
\begin{align}\label{eq:cond1}
\left|{R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}-{R({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}\right|<\epsilon,
\end{align}
for all $D\in(0,\sigma^2_{\max})$.
On the other hand, for any $\epsilon'>0$ and $m$, there exists $\delta_{\epsilon',m}>0$, such that for all $D\in(0,\delta_{\epsilon',m})$,
\begin{align}\label{eq:cond2}
\lim_{D\to0} { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}}\leq { R^{(m)}({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}} +\epsilon'.
\end{align}
Also, for any $\epsilon''>0$, there exists $\delta_{\epsilon''}>0$, such that for all $D\in(0,\delta_{\epsilon''})$,
\begin{align}\label{eq:cond3}
{ R({{\mathbb{\mathbf{X}}}},D)\over \log {1\over D}} \leq \frac{1}{2}\left({\dim}_{R}({\mathbb{\mathbf{X}}}) +\epsilon''\right).
\end{align}
Therefore, for any $\epsilon,\epsilon'$ and $\epsilon''$, choosing $m>m_{\epsilon}$, and $D\in(0,\min(\delta_{\epsilon',m},\delta_{\epsilon''}))$, and combining \eqref{eq:cond1}, \eqref{eq:cond2} and \eqref{eq:cond3} yields
\begin{align}\label{eq:b2}
\bar{d}_o({{\mathbb{\mathbf{X}}}})\leq {\dim}_{R}({\mathbb{\mathbf{X}}}) +\epsilon+\epsilon'+\epsilon''.
\end{align}
Since $\epsilon,\epsilon'$ and $\epsilon''$ are selected arbitrarily, combining \eqref{eq:b1} and \eqref{eq:b2} proves that ${\dim}_{R}({\mathbb{\mathbf{X}}}) = \bar{d}_o({{\mathbb{\mathbf{X}}}})$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:ID-eq-RDD}]
It is shown in \cite{WynerZ:71} that for any stationary process ${\mathbb{\mathbf{X}}}$
\begin{align}
|R^{(m)}({\mathbb{\mathbf{X}}},D)-R({\mathbb{\mathbf{X}}},D)|\leq{1\over m}I(X^m;X^0_{-\infty}). \label{eq:bd-Wyner-Ziv}
\end{align}
Note that while some of the results in \cite{WynerZ:71} hold only for sources that are either absolutely continuous or discrete, as shown in \cite{RJEP:16-arxiv}, this bound holds for sources with general distributions. Since the right hand side of \eqref{eq:bd-Wyner-Ziv} does not depend on $D$, it shows that $R^{(m)}({\mathbb{\mathbf{X}}},D)$ uniformly converges to $R({\mathbb{\mathbf{X}}},D)$ for all $D>0$. On the other hand, for any $0<\sigma_{\max}<1$, and any $D\in(0,\sigma_{\max}^2)$, $0<1/\log{1\over D}<1/\log{1\over \sigma_{\max}^2}$. Therefore, ${R^{(m)}({\mathbb{\mathbf{X}}},D)\over \log{1\over D}}$ uniformly converges to ${R({\mathbb{\mathbf{X}}},D)\over \log{1\over D}}$, for $D\in(0,\sigma_{\max}^2)$, and by Lemma \ref{lemma:uniform_conv}, ${\dim}_{R}({\mathbb{\mathbf{X}}}) = \bar{d}_o({{\mathbb{\mathbf{X}}}}).$
\end{proof}
For an i.i.d. source ${\mathbb{\mathbf{X}}}$, under some mild conditions, ${d}_o({\mathbb{\mathbf{X}}})$ characterizes the fundamental limits of compressed sensing \cite{WuV:10}. In other words, asymptotically, almost lossless recovery of $X^n$ generated by the source ${\mathbb{\mathbf{X}}}$ from measurements $Y^m=AX^n$ is feasible, if and only if the normalized number of measurements ($m/n$) is larger than $d_o({\mathbb{\mathbf{X}}})$. If the rate-distortion function of the source satisfies the condition of Theorem \ref{thm:ID-eq-RDD}, then ${\dim}_R({{\mathbb{\mathbf{X}}}})=\bar{d}_o({{\mathbb{\mathbf{X}}}})$, which implies that the RDD of an i.i.d.~process can also be used to characterize its compressed sensing fundamental limits.
On the other hand, compression-based compressed sensing of stochastic processes is studied in \cite{RJEP:16-arxiv}. It is shown in \cite{RJEP:16-arxiv} that there exists a compression-based recovery algorithm that achieves almost lossless recovery by using slightly more than $n\overline{\dim}_R({{\mathbb{\mathbf{X}}}})$ random linear measurements. This implies that $\overline{\dim}_{R}({\mathbb{\mathbf{X}}})$ is achievable for general sources. (Note that, by Lemma \ref{lemma:connect-UID-URDD}, in general $\overline{\dim}_{R}({\mathbb{\mathbf{X}}})\leq\bar{d}_o({{\mathbb{\mathbf{X}}}})$.)
\begin{remark}
\label{rem:id}
Theorem \ref{thm:ID-eq-RDD}, by proving the equivalence of ID and RDD, provides a potentially easier path to computing the ID of stochastic processes. Note that while to directly compute the ID of a process one needs to take the limit over the quantized approximations and then over the memory length, to be able to calculate the RDD of a process, the exact characterization of the rate-distortion function is not required. In fact, it is easy to see that it would be enough to have upper and lower bounds on the rate-distortion function of the source, $R({\mathbb{\mathbf{X}}},D)$, that are within a reasonable gap. More precisely, as long as the gap between the bounds grows as $o(\log{1 \over D})$, they can be used to evaluate the RDD. Moreover, since the RDD depends only on the low-distortion behavior of the rate-distortion function, studying its asymptotic small distortion performance is sufficient for computing the RDD, and as by Theorem \ref{thm:ID-eq-RDD}, ID of a source, without knowing the rate-distortion function explicitly. For instance, \cite{Gyorgy99} studies the asymptotic behavior of the rate-distortion function of some stochastic sources and employs those results to evaluate the RDD of some i.i.d.~processes.
\end{remark}
The next section illustrates computation of RDD and its relation to ID for a piecewise constant process.
\section{RDD of a Piecewise-Constant Process} \label{sec:pwc_source}
In general, deriving the rate-distortion function of sources with memory is very challenging. For instance, even for the binary symmetric Markov chain, the rate-distortion function is not known, except in a low-distortion region \cite{Gray:71}, and we have to resort to upper and lower bounds \cite{Berger:77,JalaliW:07}.
In this section we consider a piecewise constant signal modeled by a first order Markov process ${{\mathbb{\mathbf{X}}}}=\{X_i\}_{i=1}^\infty$, such that conditioned on $X_{i-1}=x_{i-1}$, $X_i$ is distributed according to $(1-p)\delta_{x_{i-1}}+pf_c$, where $f_c$ denotes the pdf of an absolutely continuous distribution with bounded support, defined over an interval $(l,u)$. In other words, at each time $i$, the process either makes a jump and takes a value drawn from distribution $f_c$, or it stays at $X_{i-1}$. The decision is made based on the outcome of an i.i.d. $\operatorname{Bern} \left({p}\right)$ random process independent of all past values of ${\mathbb{\mathbf{X}}}$. While the output of this source is not sparse, it is clearly a structured process. The following theorem provides upper and lower bounds on $R(\mathbf{X},D)$ of the piecewise-constant source. While there is a gap between the bounds on $R(\mathbf{X},D)$, since the gap does not depend on $D$, as shown in the following corollary, they can be used to evaluate RDD of the source exactly.
\begin{theorem}\label{thm:r-d-bound}
Consider a first-order stationary Markov process ${{\mathbb{\mathbf{X}}}}=\{X_i\}_{i=0}^\infty$, such that conditioned on $X_{i-1}=x_{i-1}$, $X_i$ is distributed according to $(1-p)\delta_{x_{i-1}}+pf_c$, where $f_c$ denotes the pdf of an absolutely continuous distribution with bounded support, $(l,u)$.
If $d_{\max}\triangleq \sup_{x,{\hat{x}}\in(l,u)}d(x,{\hat{x}})<\infty$, then
\[
pR_{f_c}(D)\leq R(\mathbf{X},D)\leq H(p)+pR_{f_c}(D),
\]
where $R_{f_c}(D)$ and $H(p)$ denote the rate distortion function of an i.i.d.~process distributed according to pdf $f_c$, and the binary entropy function ($-p\log_2 p-(1-p)\log_2(1-p)$), respectively.
\end{theorem}
\begin{proof}
To prove the upper bound (achievability), we consider a code that describes the positions of the jumps losslessly at rate $H(p)$. Since the source is piecewise constant, after describing the positions of the jumps, the encoder removes the repeated values and applies a lossy compression code of blocklength close to $np$. Therefore, to describe the values at distortion $D$ the encoder roughly needs to spend $npR_{f_c}(D)$ bits. For the lower bound (converse), we consider a genie-aided decoder that has access to the positions of the jumps. Then intuitively, to describe the values at distortion $D$, it still needs a rate of at least $pR_{f_c}(D)$. The proof presented in \cite{RJEP:16-arxiv} makes these steps formal by properly analyzing the reduced block length which is a random number.
\end{proof}
\begin{corollary}\label{thm:RDD-piecewise-constant}
For the piecewise constant source in Theorem \ref{thm:r-d-bound}, we have
$
{\dim}_R({{\mathbb{\mathbf{X}}}})=\bar{d}_o({\bf X})=p.
$ In other words, the RDD is equal to $p$ which is in turn equal to the ID of this source.
\end{corollary}
\begin{proof}
Given the bound on the rate-distortion process derived in Theorem \ref{thm:r-d-bound}, it is easy to directly derive the RDD of such a source. More precisely, given the upper bound, it follows that
\begin{align*}
\overline{\dim}_R({{\mathbb{\mathbf{X}}}})=2\limsup_{D\to0}{R({{\mathbb{\mathbf{X}}}},D)\over \log{1\over D}}
= p(\limsup_{D\to0} {R_{f_c}(D) \over \log{1\over D}})
=p,
\end{align*}
where the last step follows from \cite{Renyi:59} and \cite{KawabataD:94}. Similarly, given the lower bound, we have $\underline{\dim}_R({{\mathbb{\mathbf{X}}}})\geq p$. Therefore, $p\leq\underline{\dim}_R({{\mathbb{\mathbf{X}}}})\leq\overline{\dim}_R({{\mathbb{\mathbf{X}}}})\leq p$. In other words, for this source RDD exists and is equal to ${\dim}_R({{\mathbb{\mathbf{X}}}})=p$. Hence, the condition of Theorem \ref{thm:ID-eq-RDD} holds and we have $ {\dim}_R({{\mathbb{\mathbf{X}}}})=\bar{d}_o({\bf X}).$
This agrees with the ID of this source found in Theorem 2 in \cite{JalaliP:14-arxiv}, $\bar{d}_o({\bf X})=\underline{d}_o({\bf X})=p.$
\end{proof}
Corollary \ref{thm:RDD-piecewise-constant} states that the RDD of the piecewise constant source described in Theorem \ref{thm:r-d-bound} is equal to $p$, which is also the ID of this process \cite{JalaliP:14-arxiv}. While \cite{JalaliP:14-arxiv} directly computes the ID of such processes, Theorem \ref{thm:ID-eq-RDD} provides an easier alternate method for computing the ID as suggested in Remark~\ref{rem:id}.
\section{Conclusions}\label{sec:conc}
In this paper we have defined the RDD of stationary processes, as a generalization of the RDD of stochastic vectors introduced in \cite{KawabataD:94}. We have proved that under some mild conditions the RDD of a stationary process is equal to its ID introduced in \cite{JalaliP:14-arxiv}.
This gives an operational method to evaluate the ID of a stationary process, which was previously shown to be related to the fundamental limits of compressed sensing \cite{WuV:10,JalaliP:14-arxiv,RJEP:16-arxiv}.
\bibliographystyle{unsrt}
|
\section{introduction}
\label{sec:intro}
The transport of energy through flaring coronal loops is well studied, both observationally and theoretically, but not yet fully understood. The release of energy from magnetic reconnection events drives the acceleration of particles, generation of waves, and {\it in situ} heating of the coronal plasma, although it is not clear how energy is partitioned between the mechanisms. Further complicating the problem is that the partition of energy amongst the loops that comprise the arcade that forms along the flare ribbon has not been determined to date.
Flare energy release undoubtedly occurs across many magnetic threads, as has been known for a long time ({\it e.g.} \citealt{svestka1982}). \citet{aschwanden2001} presented an analysis of a large flare occurring across more than 100 loops, to infer cooling times across a well-observed arcade. \textit{Yohkoh}\ observations pointed to a temperature gradient in the arcade, where the outermost loops are the hottest \citep{tsuneta1996}. Tracing the motion of hard X-ray (HXR) sources, \citet{grigis2005} showed that as a disturbance propagates along the arcade, it triggers reconnection and particle acceleration in successive loops as it proceeds, thus heating the loops sequentially.
Multi-threaded models have been employed by a number of authors to study solar flares. \citet{hori1997,hori1998} adopted a multi-stranded model to explain the observation of stationary \ion{Ca}{19} emission during the impulsive phase of many flares, when single loop models consistently predicted strong blue-shifts. Similarly, \citet{reeves2002} developed a multi-threaded model to show that \textit{TRACE}\ and \textit{Yohkoh}\ light-curves were more readily explained by an arcade rather than a single loop. \citet{warren2005} derived an algorithm to compute energy inputs for successive threads comprising a flare, by calculating the discrepancy between the observed and calculated GOES flux. They showed that the absence of strongly blue-shifted \ion{Ca}{19} emission in \textit{Yohkoh}\ observations is because that emission is masked by previously heated threads. \citet{warren2006} studied the duration of heating on successive threads, concluding that short heating time scales lead to significantly higher temperatures, inconsistent with \textit{Yohkoh}\ observations. \citet{falewicz2015} compared one and two-dimensional models of a flare to find that the observed dynamics were better reproduced by their 2D model, which approximated a multi-stranded model. On the other hand, \citet{doschek2015b} found that while a single loop model can reproduce high temperature evaporation flows, there were numerous discrepancies between the observed and modeled cooler, red-shifted lines. Recently, \citet{qiu2016}, using the 0D model EBTEL \citep{klimchuk2008}, studied the cooling phase of flares with a multi-threaded model, and only found consistency with EUV emission if there is prolonged gradual phase heating occurring on many threads.
In the first paper (\citealt{warren2016}, hereafter Paper \rom{1}), we studied extensively a small flare that was seen with \textit{Interface Region Imaging Spectrograph}\ (\textit{IRIS}, \citealt{depontieu2014}), \textit{The Reuven Ramaty High Energy Solar Spectroscopic Imager}\ (\textit{RHESSI}, \citealt{lin2002}), \textit{Atmospheric Imaging Assembly}\ (\textit{AIA}, \citealt{lemen2012}), and \textit{Extreme Ultraviolet Imaging Spectrometer}\ and \textit{X-Ray Telescope}\ aboard \textit{Hinode}\ (\textit{EIS}\ and \textit{XRT}, respectively, \citealt{culhane2007} and \citealt{golub2007}). The combination of instruments allows coverage across a wide temperature range, from the chromosphere through the transition region (TR) and upper corona, to temperatures exceeding 10 MK. The unique perspective allowed us to measure temperatures, emission measures (EMs), non-thermal electron beam parameters, energy input, and individual TR brightenings at high cadence and spatial resolution.
In Paper \rom{1}, we presented observations of \ion{Si}{4} and \ion{C}{2} as seen by \textit{IRIS}, both of which brightened during the rise phase along with the HXR emission measured with \textit{RHESSI}. The two lines were red-shifted during that time period, and remained red-shifted even after the impulsive phase, gradually decreasing in magnitude over time-scales exceeding 20 minutes at some positions. Similar trends in \ion{Si}{4} and other cool lines were reported by other authors in larger flares seen with \textit{IRIS}, {\it e.g.} \citet{sadykov2015,brannon2015, polito2016}.
In this paper, we seek to explain the persistent red-shifts by developing a model which requires a partition of energy amongst the magnetic strands comprising the flare. In Section \ref{sec:modeling}, we describe the hydrodynamic code used to model this flare. We then split up the results in Section \ref{sec:results} into two parts: a simple model (both single loop and multi-threaded loop) and a multi-threaded Monte Carlo simulation. We finally discuss the implications and conclusions of this work in Section \ref{sec:implications}.
\section{Modeling}
\label{sec:modeling}
We have run hydrodynamic simulations with the HYDrodynamics and RADiation code (HYDRAD; \citealt{bradshaw2003,bradshaw2013}) in order to study potential heating mechanisms. The code solves the one-dimensional hydrodynamic equations, conservation of mass, momentum, and energy, applicable to a two-fluid plasma confined to a magnetic flux tube rooted beneath the surface and extending into the corona. The energy equations include terms for thermal conduction, enthalpy flux, small-scale electric fields, viscosity, gravity, inter-species collisions, and radiation. A key strength of the HYDRAD code is its speed and portability. In this paper, we present a total of 45 simulations, each for 1,000 seconds of simulation time with highly resolved grids and a wide parameter space. All of the simulations were performed on a desktop computer, with up to 8 running simultaneously.
Because the corona is a low $\beta$ plasma \citep{reidy1968,dulk1978,gary2001}, cross-field conduction is negligible \citep{goedbloed2004}, and the fields are frozen-in \citep{alfven1943}, each coronal loop may be treated as an isolated structure, with no interaction between adjacent loops. We assume each thread is semi-circular in shape, oriented vertically from the solar surface, with constant cross-section from foot-point to apex.
We treat radiative losses in the corona and TR with a full calculation of losses with CHIANTI v.8 \citep{dere1997,delzanna2015}, via the equation ({\it e.g.} \citealt{mason1994, bradshaw2013b}):
\begin{align}
E_{R}(X) &= n^{2} \Lambda \nonumber \\
&= n^2 \Bigg(0.83 \times Ab(X) \times \sum_{i=1}^{i=Z+1} \epsilon_{i}\ X_{i} \Bigg)
\label{eqn:radlosses}
\end{align}
\noindent where $n$ is the number density, $Ab(X)$ the abundance of element $X$ relative to hydrogen, $\epsilon_{i}$ the emissivity of all lines from ion $i$ of element $X$, and $X_{i}$ the population fraction of ion $i$. We solve a continuity equation for non-equilibrium ionization states of hydrogen, carbon, oxygen, silicon, and iron in the work here \citep{bradshaw2003}, as non-equilibrium ionization is expected to be significant during impulsive bursts of heating. Further, because hydrogen may not be ionized in the chromosphere, collisions and thermal conduction due to neutrals are included in the code \citep{orrall1961}. We treat optically thick radiation in the chromosphere with the recipes of \citet{carlsson2012}. We adopt the photospheric abundances of \citet{asplund2005} as some recent evidence indicates that flares are photospheric in composition \citep{warren2014}, although there is also evidence to the contrary \citep{dennis2015,doschek2015}.
We have assumed that the loop is heated by an electron beam under the collisional thick-target model \citep{brown1971,hudson1972}, where accelerated electrons stream through the corona, depositing their energy through collisions with chromospheric plasma. We have treated heating by an electron beam using the model of \citet{emslie1978}, with the details of the implementation described in \citet{reep2013}. We assume an electron beam distribution of the form:
\begin{equation}
\mathfrak{F}(E_{0}, t) = \frac{F_{0}(t)}{E_{c}^{2}}\ (\delta - 2) \times
\begin{cases}
0 & \text{if } E_{0} < E_{c} \\
\Big(\frac{E_{0}}{E_{c}}\Big)^{- \delta} & \text{if } E_{0} \geq E_{c}
\end{cases}
\label{eqn:sharpdist}
\end{equation}
\noindent where $F_{0}(t)$ is the beam energy flux (erg\,sec$^{-1}$\,cm$^{-2}$), $E_{c}$ is the low energy cut-off (keV), $\delta$ is the spectral index, and $E_{0}$ is the initial kinetic energy of a given electron (keV). This distribution, referred to as a sharp cut-off, is commonly assumed and allows for easy comparison to measured \textit{RHESSI}\ data. We use the actual \textit{RHESSI}\ parameters in the model when available, although in Section \ref{subsec:montecarlo} we treat the energy input for a given loop as randomly selected on a power-law distribution, as described in that section.
With \textit{RHESSI}, although the spectra are not spatially resolved, we measured the electron beam parameters along with temperature and emission measure as a function of time early in the event, shown in Figure \ref{fig:rhessi_params}. The points in color refer to measurements from a single detector (specifically, detectors 1, 3, 6, 8, and 9), while the black points refer to the mean of all the detectors. The power carried by the electron beam gradually increases, though it stays around $1 - 3 \times 10^{27}$\,erg\,s$^{-1}$. A cross-sectional area $A$ must be measured (or assumed) to determine the energy flux $F_{0}(t) = P(t) / A(t)$. The spectral index $\delta$ gradually increases, but is approximately $6$ at most times (slightly lower than the median value for a microflare, \citealt{hannah2008}). The cut-off energy $E_{c}$ is approximately $11$\,keV for the entire duration (compared to a median of $12$\,keV in microflares, \citealt{hannah2008}).
\begin{figure*}
\centering
\includegraphics[width=7.in]{f01.eps}
\caption{The electron beam parameters, as well as temperature and emission measure, as measured with \textit{RHESSI}\ over the course of the event. The points in color refer to measurements with a single detector, while the black points denote the average of all detectors. }
\label{fig:rhessi_params}
\end{figure*}
From these simulations, we forward model spectral lines as might be seen by \textit{IRIS}, following the methodology of \citet{bradshaw2011}, using the \textit{IRIS}\ response functions calculated with SolarSoft. In this work we focus on the \ion{Si}{4} 1402.770 \AA\ and \ion{C}{2} 1334.535 \AA\ lines, which are useful diagnostics of heat transport to the lower atmosphere ({\it e.g.} \citealt{testa2014}), and strongly correlated with energy input \citep{cheng1981}.
\section{Results}
\label{sec:results}
In Paper \rom{1}, we presented detailed observations of SOL2014-11-19T14:25UT, a small flare which was observed simultaneously by \textit{IRIS}, \textit{RHESSI}, and \textit{Hinode}. The \ion{Si}{4} emission along the slit shows distinct brightenings (typically lasting less than 60\,s) between 14:15 and 14:25, at which times the line profile is red-shifted to $\approx 30$\,km\,s$^{-1}$. After these brightenings, however, the red-shifts persist, remaining greater than $20$\,km s$^{-1}$ for more than 30 minutes, gradually declining. Emission from \ion{C}{2} is similar (see Paper \rom{1}). In this section, we focus on reproducing long-lasted red-shifts, while simultaneously remaining within the constraints given by the other instruments (temperature, density, EM, power).
\subsection{Simple Model}
\label{subsec:simple}
We first attempted to model this event with a single loop, on which seven heating bursts occurred (based on the number of observed \ion{Si}{4} brightenings during the peak intensity). We also attempted to model this event using a multi-threaded loop composed of seven strands, on each of which one heating burst occurred. In both cases, we use the observationally measured cut-off energy $E_{c} \approx 11$\,keV and spectral index $\delta \approx 6$, with a maximum beam flux of $5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$ (above the explosive threshold for this cut-off, \citealt{reep2015}). Figure \ref{fig:simulated_Si4} shows synthesized \ion{Si}{4} emission for the two cases. On the left hand side, the plots show a single loop modeled with seven distinct bursts of heating, while the right hand shows a simulation with seven threads each with one heating burst. The top plots show the \ion{Si}{4} intensity as a function of time for 15 minutes while the bottom plots show the Doppler shifts of the line (fit with a single Gaussian).
\begin{figure*}
\begin{minipage}[t]{0.5\linewidth}
\centering
\includegraphics[width=3.3in]{f02.eps}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width=3.3in]{f03.eps}
\end{minipage}
\caption{Synthesized \ion{Si}{4} emission relative to \textit{IRIS}\ for the two simulations: seven bursts of heating on a single loop (left) and seven threads with one heating burst each (right). The top plots show the synthesized intensity as a function of time while the bottom plots show the Doppler shift as a function of time.}
\label{fig:simulated_Si4}
\end{figure*}
There is an obvious discrepancy between these simulations and the observations. First, the intensity in the single thread case is more or less constant after the initial heating event, and there are distinct intensity bursts at each heating burst in the multi-threaded case. In the observations, however, the \ion{Si}{4} intensity sharply brightens during the rise phase, with a few local maxima, and then very gradually decays. Second, in the monolithic loop simulation, the Doppler velocity shows only a single strong red-shift at the time of the first heating burst, while the multi-threaded simulation shows seven strong red-shifts, corresponding to a heating event on each thread (which may be more akin to the red-shifts seen by \citealt{brosius2015}). The observations, however, show a nearly constant red-shift of $> 20$\,km\,s$^{-1}$, with weak spikes in the velocity when the intensity spikes, which neither simulation reproduces.
The quick decay of red-shifts in the simulations is straight-forward to explain. The heating burst quickly heats the chromosphere to coronal temperatures, which then drives energy down through the transition region, forcing a flow of material downward. As the material proceeds to greater depths, the local density increases, causing a loss of momentum and a quick dissipation of the bulk flows. Further, in the monolithic simulation, once evaporation brings material into the corona, later heating bursts cannot cause a strong velocity flow due to the increased inertia. In the multi-threaded simulation, each loop shows a strong red-shift at the time of a heating burst, as material is pushed downward. In both cases, the down-flows quickly dissipate in the span of $\approx$20--30\,s. \citet{fisher1989} analytically derived this quick decay of condensation flows, showing that, at most, they last for 45--60\,s, which is independent of the heating duration (although that does not imply that emission in the same spectral line lasts that long, so it is effectively an upper limit).
The observed values are irreconcilable with these simulations. A persistent down-flow of $> 20$\,km\,s$^{-1}$ cannot be explained by constant heating, or with any number of heating bursts on a single loop. A multi-threaded model could be consistent if the number of threads were significantly higher ($> 50$) than what we have assumed, as we will show in the next section.
\subsection{Monte Carlo Model}
\label{subsec:montecarlo}
The inadequacy of a simple model highlights the complexity of the event. Neither a single loop model nor a simple multi-threaded model with a small number of threads produces emission consistent with the observed values. Thanks to the wealth of observations available during this event, we can constrain the parameters of the model more thoroughly and produce a more realistic model. The first major consideration is that there must have been a very large number of strands comprising the emission {\it within a single \textit{IRIS}\ pixel}.
The \textit{RHESSI}\ observations indicate the presence of an electron beam with mean values of low energy cut-off $E_{c} \approx 11$\,keV, spectral index $\delta \approx -6$, and power input $P \approx 10^{27}$\,erg\,s$^{-1}$, which we take as inputs to the model. It is clear both observationally and from the modeling that a single loop model is insufficient, however, and in a multi-threaded model the power must be divided among many threads comprising an arcade. \textit{RHESSI}\ does not have the spatial resolution nor the cadence necessary to determine that distribution. Since the HXR emission is cospatial to and strongly correlated with the intensity of TR lines (e.g. Paper \rom{1}, \citealt{cheng1981,poland1982,woodgate1983,poland1984,simoes2015b}), we use the intensity distribution of \ion{Si}{4} emission measured with \textit{IRIS}\ to estimate this distribution of energy carried by the electron beam to various threads.
In Paper \rom{1}, we presented a histogram of \textit{IRIS}\ 1400 \AA\ intensities at various times, that were fit to a power law. The slope $\alpha$ was determined to be $\approx -1.6$ at most times, ranging from about $-1.0$ to $-2.5$. We therefore assume that the intensity distribution is correlated with the energy flux carried by the electron beam on each thread, thus composing a power-law of beam fluxes. Although the correlation between TR lines and HXR emission is well established at the large scale, both spatially and temporally, future HXR instrumentation ({\it e.g.} \textit{FOXSI}, \citealt{krucker2014}) should verify that this remains true at smaller scales.
We have run 37 simulations with HYDRAD, adopting electron beam heating with $E_{c} = 11$\,keV, $\delta = -6$ (the average values measured with \textit{RHESSI}), and beam flux ranging from $10^{8} - 10^{11}$\,erg\,s$^{-1}$\,cm$^{-2}$. We assume that the heating lasts for 10 seconds on each thread, with a flat temporal envelope. We then calculate the \ion{Si}{4} and \ion{C}{2} emission from each simulation along the loop at all times, using the \textit{IRIS}\ response function.
From this set of simulations, on a power-law distribution with slope $\alpha(t)$, varying in time as measured with \textit{IRIS}, we then randomly sample $N$ total threads to comprise the emission (threads can be used more than once). We allow these threads to occur randomly at an average rate of 1 per $r$ unit time ({\it i.e.} on a Poisson distribution with average waiting time $r$), and assume $N \times r \gtrsim 600$\,s in order to last through the period under consideration (approximately 14:14-14:24 UT). We then sum the emission from all $N$ threads to calculate a light-curve and line profile at all times as if they were all contained in the same \textit{IRIS}\ pixel, and then fit the profiles to calculate the Doppler velocity as a function of time. We present the intensities and velocities at both 1 and 8 second integration times in order to compare directly with the observations (8 second integration) and with what might be seen by a faster instrument. Since the cross-sectional area of an individual strand is unknown and necessary to calculate line intensities, we assume the pixel area is divided evenly among the $N$ strands, and additionally assume a filling factor of 1.
\subsubsection{Synthesized light-curves and Doppler shifts}
There are many variables at play that cannot be directly measured, despite the abundance of instruments that observed the event: the number of threads comprising the emission $N$, the rate of heating bursts onto these threads $r$, the minimum and maximum sizes of energy release, $F_{\text{min}}$ and $F_{\text{max}}$ (if any). We discuss ways to constrain these variables in this section.
First, consider the maximum size of the heating rate. If that maximum is too small, the lines are generally blue-shifted, as in Figure \ref{fig:low_energy}. We have limited the power-law to the range of energy flux $F_{0} = 10^8 - 10^9$\,erg\,s$^{-1}$\,cm$^{-2}$. We show the emission calculated with $N = 120$ total threads and $r = 5$ seconds average waiting time. The top plot shows the intensity of \ion{Si}{4} as a function of time, where the black dotted line shows a 1-second integration time and the red 8-second. The bottom plot shows the calculated Doppler shifts, based on fitting a single Gaussian to the profiles, where we have defined red-shifted flows to be positive. The line is, on average, weakly blue-shifted during the heating period ({\it cf.} \citealt{testa2014}), which was not observed during the event (see Paper \rom{1}). It is clear that there must have been events with energy flux such that $F_{\text{max}} > 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$.
\begin{figure}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width=3.3in]{f04.eps}
\end{minipage}
\caption{Synthesized \ion{Si}{4} profiles for a set of randomly selected weak bursts, ranging in energy from $10^8 - 10^9$\,erg\,s$^{-1}$\,cm$^{-2}$. The top plot shows the calculated intensity, and the bottom the Doppler shift. Red-shifts are defined as positive velocities. The line was calculated with integration times of one (black dotted) and eight seconds (red dashed), compared to the eight second integration time used with \textit{IRIS}\ in Paper \rom{1}. Note the absence of significant red-shifts.}
\label{fig:low_energy}
\end{figure}
Figure \ref{fig:all_energies} shows the calculated light-curves in the case where the power-law extends from $10^{8} - 10^{11}$\,erg\,s$^{-1}$\,cm$^{-2}$ for $N = 120$ threads and $r = 5$\,s per thread. At energy fluxes less than $10^{8}$\,erg\,s$^{-1}$\,cm$^{-2}$, there is little to no heating so that only minimal emission would result, while energy fluxes greater than $10^{11}$\,erg\,s$^{-1}$\,cm$^{-2}$ are so rare that we can discount them in general. In this case, the stronger events quickly produce a red-shift in both lines as the chromosphere quickly heats to TR temperatures. The stronger events also produce considerably higher intensities in both spectral lines, so that the emission is weighted heavily by them, despite being less common overall. The net result is that the intensities are markedly more bursty in nature due to large events rising above the background of weak events, and that the calculated Doppler shift is weakly red-shifted in both lines when the intensity is high. However, the velocities are not persistent over the period of heating, as in the observations. There are two possible explanations: that the assumed minimum energy release is too low ($F_{\text{min}} > 10^{8}$\,erg\,s$^{-1}$\,cm$^{-2}$) or that the assumed spacing between events is too large ($r < 5$\,s per thread). We examine both of these possibilities.
\begin{figure}
\begin{minipage}[b]{0.5\linewidth}
\centering
\includegraphics[width=3.3in]{f05.eps}
\end{minipage}
\caption{Synthesized \ion{Si}{4} profiles for a set of randomly selected bursts, ranging in energy from $10^{8} - 10^{11}$\,erg\,s$^{-1}$\,cm$^{-2}$, otherwise as before. The red-shifts are bursty and short-lived, in stark contrast to the observations.}
\label{fig:all_energies}
\end{figure}
We now consider the minimum energy flux. In Figure \ref{fig:minimum}, we show synthesized \ion{Si}{4} emission and Doppler shifts for the cases where the minimum energy flux is taken to be $10^{9}$, $3 \times 10^{9}$, and $5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, with $N = 120$ and $r = 5$\,s, which also can be compared with the previous figure. For $F_{\text{min}} < 5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, there are sharp drops in the red-shift, which were not observed, while above that value, the red-shifts remain at around 30\,km\,s$^{-1}$ for the duration of the heating events. This minimum beam flux corresponds roughly to the transition between gentle and explosive evaporation flows for the cut-off used here (see Section 5 of \citealt{reep2015}), suggesting that the majority of the threads were heated explosively.
\begin{figure*}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width=2.2in]{f06.eps}
\end{minipage}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width=2.2in]{f07.eps}
\end{minipage}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width=2.2in]{f08.eps}
\end{minipage}
\caption{Synthesized \ion{Si}{4} profiles for a set of randomly selected bursts, with minimum energy fluxes $10^{9}$, $3 \times 10^{9}$, and $5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, respectively. The red-shifts become more persistent with larger minimum energy fluxes, and the sharp drops to zero disappear. Interestingly, an instrument with a higher cadence might be able to detect drops in the velocity. }
\label{fig:minimum}
\end{figure*}
One might ask whether the persistent red-shifts might also be explained by simply increasing the rapidity with which new threads are energized. In Figure \ref{fig:r}, we show three plots, two where $F_{\text{min}} = 10^{8}$\,erg\,s$^{-1}$\,cm$^{-2}$, but now $r$ has been decreased to 1 and 3\,s per thread, and one where $F_{\text{min}} = 5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, with $r = 10$\,s per thread. In all three of these cases, we see that the red-shifts have sharp drops toward 0, either due to the weighting of the weaker heating events or the quick decay of red-shifts on the individual threads. However, we can deduce that: 1. $r < 10$\,s per thread, that is, new threads are energized at a rate faster than one per 10 seconds on average, 2. $N \times r > 600$\,s, so $N > 60$ threads must have been energized during the HXR burst, and 3. the rate of energization $r$ alone cannot explain persistent red-shifts, there must also be a minimum energy flux in general. There is one caveat to mention, though, in that a constant low-energy cut-off was assumed on each thread, which cannot be verified with \textit{RHESSI}, although the parameter strongly affects flow speeds \citep{reep2015}.
\begin{figure*}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width=2.2in]{f09.eps}
\end{minipage}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width=2.2in]{f10.eps}
\end{minipage}
\begin{minipage}[b]{0.33\linewidth}
\centering
\includegraphics[width=2.2in]{f11.eps}
\end{minipage}
\caption{Synthesized \ion{Si}{4} profiles, from left to right, with minimum energy fluxes $10^{8}$, $10^{8}$, and $5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, respectively, with $r$ values of 1, 3, and 10\,s per thread, respectively. In all three cases, there are sharp drops in the red-shifts, indicating that the rate of new threads or minimum energy alone cannot explain the observed persistent red-shifts, but that both are required.}
\label{fig:r}
\end{figure*}
\subsubsection{Checks on the model}
The synthesized emission appears consistent with the observed values, given the constraints discussed above. However, an important check on the model is to synthesize other observables to find if it is well bounded by them. Can the model simultaneously reproduce the \textit{IRIS}\ velocities, the EM distribution found with \textit{EIS}, \textit{XRT}, and \textit{AIA}, and the density constraint found with \textit{EIS}? How do the line profiles compare to those found in Paper \rom{1}?
Consider the case where $r = 5$\,s per thread, $N = 120$ threads, $F_{\text{min}} = 5 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, and $F_{\text{max}} = 10^{11}$\,erg\,s$^{-1}$\,cm$^{-2}$, which nicely reproduces the persistent red-shifts in both \ion{Si}{4} and \ion{C}{2}. Figure \ref{fig:other_params} shows the line profiles and light-curves for \ion{Si}{4} and \ion{C}{2}, the emission measure per area (defined as $n^{2} \Delta s$, where we have integrated over 1 arcsec near the apices of the loops), and finally the density sensitive line ratio \ion{Fe}{14} $264.8/274.2$\,\AA\ as might be seen by \textit{EIS}\ near the foot-point (the red data point denotes the observed ratio). The observed line profiles in Paper \rom{1} can be compared with those calculated here. Importantly, at most times, both lines show a bright red-shifted component, with a much weaker stationary component ($\lesssim 10\%$). This compares favorably with the observed \ion{Si}{4} profiles, although the stationary component of \ion{C}{2} is smaller than that observed (Figure 9 of Paper \rom{1}). The \ion{C}{2} line likely requires a full radiative transfer calculation to reproduce accurately ({\it e.g.} \citealt{lin2015}), as it forms in the upper chromosphere or base of the TR, whereas the contribution function in CHIANTI cuts off below $10^{4}$\,K. The forward model only considers thermal broadening in determining line widths, and in general both spectral lines are thinner than the observed ones, suggesting that another broadening mechanism contributes to the observed values.
\begin{figure*}
\centerline{\includegraphics[width=0.85\linewidth]{f12.eps}}
\centerline{\includegraphics[width=0.5\linewidth]{f13.eps}%
\includegraphics[width=0.5\linewidth]{f14.eps}}
\vspace{-0.2in}
\centerline{\includegraphics[width=0.5\linewidth]{f15.eps}%
\includegraphics[width=0.5\linewidth]{f16.eps}}
\caption{$r = 5$\,s per thread, $N = 120$ threads. At top, the line profiles for \ion{Si}{4} and
\ion{C}{2} at 8-second integration, which can be compared with the observed values in Paper
\rom{1}. At center, the light-curves, as before. The bottom left plot shows the ratio of the
\ion{Fe}{14} $264.8$ and $274.2$ lines as might be seen by \textit{EIS}, where the red data point
indicates the observed value. Finally, the bottom right plot shows the emission measure per
area ($= n^{2} \Delta s$) integrated over 1\,arcsec near the apices of the loops, at a few
selected times (multiplied by factors of 10 for clarity). }
\label{fig:other_params}
\end{figure*}
The density measured from the \ion{Fe}{14} ratio is slightly lower than that observed, but within the error bars. No significant trend was found between the number of threads $N$ and the calculated ratio, although higher minimum flux values increase the ratio, in general. The line ratio oscillates more for smaller number of threads, although the cadence of the \textit{EIS}\ raster is too slow to draw any conclusions from this.
The emission measure distribution (EMD, e.g. \citealt{graham2013,simoes2015b}) reveals a good deal about the dynamics. The maximum temperature in the loops tends to decrease with time - peaking at about 15, 10, 8, 6, and 5 MK at the times shown. A line was fit to the cool-ward side of the EM from its peak value down to log $T = 6.2$, as done in Paper \rom{1}, with temperature bins of 0.1 d $\log{T}$. The slopes of those lines steepen during the heating period gradually, since the hottest material dominates the emission during this time, and then become shallower as the loops cool after the heating ceases, as might be expected. At 500 seconds, the calculated slope is about $5.4$ which is intermediate to the observed values measured with \textit{AIA}\ and \textit{XRT}\, $6.4 \pm 0.9$ at UT14:22:49, and that found with \textit{EIS}, $4.5 \pm 1.1$ (see Figure 13 of Paper \rom{1}).
In Figure \ref{fig:em}, we show the EMD calculated for $r = 1, 3, 10$\,s per thread, which can be compared with the previous EMD plot. There is no clear trend in the slopes for varying values of $r$, despite what may be expected, which is likely due to the randomized energy and timing of individual threads. However, for smaller values of $r$, where the number of threads $N$ is much larger, there is more emission at lower temperatures than in the case of {\it e.g.} $r = 10$\,s per thread, because there are many threads cooling at a given time. In general, however, the slopes and peak temperatures are consistent with the observed values found in Paper \rom{1}.
\begin{figure}
\centerline{\includegraphics[width=\linewidth]{f17.eps}}
\centerline{\includegraphics[width=\linewidth]{f18.eps}}
\centerline{\includegraphics[width=\linewidth]{f19.eps}}
\caption{Emission measure plots for $r = 1, 3, 10$\,s per thread, with $F_{\text{min}} = 5 \times
10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, and $F_{\text{max}} = 10^{11}$\,erg\,s$^{-1}$\,cm$^{-2}$.}
\label{fig:em}
\end{figure}
In general, we find that the model simultaneously reproduces the persistent red-shifts and intensities along with the EMD, maximum temperature, and density measured from \ion{Fe}{14}, given the numerous constraints mentioned above. Most importantly:
\begin{enumerate}
\item{The energy partition between threads is described by a power-law with slope $\approx -1.6$ as determined observationally}
\item{The average time between threads $r \lesssim 10$\,s per thread}
\item{The number of threads $N > 60$ within a single \textit{IRIS}\ pixel}
\item{A faster rate $r$ of new threads cannot explain the persistent red-shifts by itself. There also must be a minimum heating flux on the majority of threads.}
\item{The results here suggest that for this event the minimum flux $F_{\text{min}} \gtrsim 3 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$, or more generally that the majority of threads are heated explosively. }
\end{enumerate}
\section{Implications \& Conclusions}
\label{sec:implications}
In Paper \rom{1}, we presented observations of the flare SOL2014-11-19T14:25UT, which was observed with many different instruments, covering a wide range of energies and temperatures. Observed red-shifts in the \ion{Si}{4} 1402.770 \AA\ and \ion{C}{2} 1334.535 \AA\ lines persisted for longer than 30 minutes, which is difficult to reconcile with simple theoretical models. Specifically, \citet{fisher1989} showed that condensation flows persist for about $45$ seconds, regardless of the strength or duration of heating.
Alternatively, \citet{brosius2003} suggested that a ``warm rain'' scenario can produce long-lasting red-shifts in TR lines (see also \citealt{tian2015}). During the impulsive phase of a large M-flare, \citet{brosius2003} found \ion{O}{3}, \ion{O}{5}, \ion{Mg}{10}, and \ion{Fe}{19} were all initially blue-shifted. The two oxygen lines gradually transitioned into down-flows that lasted for half an hour, while the \ion{Mg}{10} line was found to be composed of a strong stationary component and a weaker red wing, and \ion{Fe}{19} remained stationary thereafter. Those red-wing components, termed ``warm rain'', were interpreted as signatures of the cooling and draining of a loop, and lasted for half an hour or so after the flare's onset.
However, the event studied here differs in a few important respects. First, there were no signatures of blue-shifts in the TR lines during the impulsive phase. \ion{Si}{4} was fully red-shifted for the duration of the event (Figure 9 of Paper \rom{1}), in contrast to the behavior of the oxygen lines reported in \citet{brosius2003}. Second, there is insufficient time for the loops to drain. The red-shifts begin simultaneously with the HXR burst, suggesting that they are signatures of chromospheric condensation as the energy is deposited by electron beams. After heating ceases on a given coronal loop, there is a long time period during which the coronal density does not drain significantly, and energy losses are first dominated by thermal conduction, then by radiation, and only then by an enthalpy flux (see the thorough treatment by \citealt{bradshaw2010}). The time scales for coronal loops to cool and drain were derived analytically and checked numerically by \citet{cargill1995,bradshaw2005,bradshaw2010,cargill2013}, and typically are on the order of 45 minutes to an hour. Finally, the cooling between successive \textit{AIA}\ channels often seen in flares ({\it e.g.} \citealt{petkaki2012}) was seen in the coronal section of the loops, with a cooling time-scale of about 40 minutes, suggesting they did not drain significantly for nearly as long.
In this paper, by adopting a multi-threaded model, we have shown that these observations are consistent with a power-law distribution of heating occurring on a very large number of threads. The following important conclusions can be drawn from the work here.
\begin{enumerate}
\item \textbf{Multi-stranded heating.}
The single loop model is woefully inadequate to explain the intensities or Doppler shifts observed in this event, regardless of the number of heating events on the loop or duration of heating. A simple multi-stranded model of 7 loops similarly fails, as the observed Doppler shifts are essentially continuous, not single discrete events. However, a multi-stranded loop model as presented in Section \ref{subsec:montecarlo} captures many of the observed properties of the \textit{IRIS}\ emission, while being within the bounds of the observed density, temperature, and emission measure. Compare the conclusions of many prior multi-threaded studies, {\it e.g.} \citet{hori1998,reeves2002,warren2006}, etc.
\item \textbf{Energy partition among strands.}
We measured the distribution of \textit{IRIS}\ SJI intensities to be well described by a power-law, with slope $\alpha = -1.6$ at most times (see Paper \rom{1} for details). Since the intensities of many TR lines are proportional, both spatially and temporally, to the HXR intensities (Paper \rom{1}, \citealt{cheng1981,poland1984,simoes2015b}), and since the energy flux of the electrons is proportional to the non-thermal HXR intensity \citep{brown1971,holman2011}, we take this distribution as a proxy for the partition of energy among the threads. This distribution over a large number of threads produces \textit{IRIS}\ \ion{Si}{4} and \ion{C}{2} intensities and Doppler shifts that are consistent with values measured in Paper \rom{1}. In future work, we will examine this distribution for more events of varying \textit{GOES}\ class to determine statistical trends and properties.
\item \textbf{Resolving loop structures.}
It does not seem possible to explain the observed red-shifts with a single loop model, or with a small number of strands. Further, the background level of emission does not strongly show Doppler shifts (blue or red), and the shifts correspond to brightenings above background emission so that the red-shifts must be a signature of the flare itself. As the red-shifts were measured in single pixels, then, we conclude that there is loop structure not being resolved at the sub-pixel level of \textit{IRIS}. In order to maintain a red-shift in these lines without sharp drops in the speed, threads must be energized at a rate $r < 10$\,s per thread, giving a lower limit on the number of threads $N > 60$ {\it rooted within a single \textit{IRIS}\ pixel} for the duration of the HXR burst. For the duration of the entire event, this number must be appropriately increased. In comparison, \citet{simoes2015} estimated a rate of $r = 3$\,s per thread (total of 120 threads during the impulsive phase) over the entire reconnection region of a small C2.6 flare. Their analysis was based on the released non-thermal energy, and constitutes a lower bound. \\
What is the size, then, of an individual strand? If the IRIS pixel were divided evenly between strands, then the diameter is on the order of $\frac{1}{100}$\,arcsec or less, significantly smaller than previous suggestions. This may provide evidence for the fractal model of reconnection in flares \citep{shibata2001,singh2015,shibata2016}, where the current sheet becomes exceedingly thin due to the secondary tearing instability \citep{zweibel1989}.
\item \textbf{Beam energy flux constraints.}
\textit{RHESSI}\ measures the power contained in the electron beam integrated over the entire foot-point, which can then be divided by an area to give an estimate of the beam energy flux. However, since it is integrated over the entire foot-point, that does not specify what the flux was on the many threads comprising that area. Combined with the power-law distribution, we have constrained the maximum and minimum values of the flux. The cut-off energy during this event was measured at 11-13 keV for the duration of the HXR burst (Figure \ref{fig:rhessi_params}). At that cut-off, the threshold between gentle and explosive evaporation is $\approx 3 \times 10^{9}$\,erg\,s$^{-1}$\,cm$^{-2}$ \citep{reep2015}. For lower beam fluxes, the \ion{Si}{4} line is in fact blue-shifted (compare \citealt{testa2014}), which was never observed during this event. We therefore can reasonably conclude that the maximum beam flux must be greater than this value. What's more, since small events are far more likely on a power-law distribution of energies, they strongly weight the emission and often cause sharp drops in the measured Doppler shift, so that it seems likely that the majority of the threads were heated explosively.
\end{enumerate}
This work has given a great deal of insight into the dynamics of this small flare. We have reasonably found a lower limit to the number of magnetic field threads, and have found the partition of energy among them, which allows us to build a realistic multi-threaded model. This model is well constrained by the abundance of observations from many different instruments, and can be applied to flares which do not have coverage as good as this one. There are still many areas of this work that can be improved to remove assumptions and generalize the model, however, such as determining how the electron beam parameters vary from thread to thread or finding an upper limit to the number of threads. It is also often true that \ion{Si}{4} has a stronger stationary component in other flares than was seen in this one ({\it e.g.} \citealt{tian2014}), so that further work may be required to determine whence the difference arises.
We speculate that spectral lines seen in larger flares such as \ion{Fe}{21} 1354.08\,\AA\ may further improve our understanding of energy deposition between threads. The results of \citet{fisher1989} make it clear that the duration of condensation flows are insensitive to the heating strength and duration. However, evaporation flows are not limited in the same manner, and in fact there are indications that the flows last as long as the heating does ({\it e.g.} the flows in Figures 4 and 5 of \citealt{reep2015} or the \ion{Fe}{21} shifts in \citealt{polito2016}). Unfortunately, for this event, \ion{Fe}{21} was not observed, and so no hard conclusions can yet be drawn regarding the heating durations.
\acknowledgments The authors thank Lucas Tarr, George Doschek, and the anonymous referee for discussions which improved the modeling and understanding of this event. This research was performed while JWR held an NRC Research Associateship award at the US Naval Research Laboratory with the support of NASA. The research leading to these results has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 606862 (F-CHROMA) (PJAS). IRIS is a NASA Small Explorer developed and operated by LMSAL with mission operations executed at NASA Ames Research center and major contributions to downlink communications funded by the Norwegian Space Center through an ESA PRODEX contract. CHIANTI is a collaborative project involving George Mason University, the University of Michigan (USA) and the University of Cambridge (UK).
|
\section{Few more uncertainty relations} Here, we provide few more uncertainty relations particularly, using triangle inequality. If one applies the triangle inequality for the two vectors $\vec{u}$ and $\vec{v}$ as defined earlier, another tighter uncertainty relation for the sum of standard deviations for the two incompatible observables can be obtained as
\end{document} |
\section{Introduction}
\label{sec:introduction}
Band structure has become a critical component of condensed matter physics because it provides information on a crystalline material's electronic, magnetic, and optical properties \cite{bk:ashcroft-1976}. Materials that are mostly, but not perfectly, crystalline are also of interest, but band structure no longer provides a complete characterization of these materials due to disorder (\textit{e.g.},\ an impurity or surface). Complex band structure (CBS) \cite{goodwin-205-1939, kohn-809-1959, heine-300-1963, heine-1-1964, krieger-776-1967, prodan-035128-2006} generalizes this conventional band structure to account for disorder by considering wavevectors with complex components. Although CBS has been derived and developed in numerous contexts (\textit{vide infra}), the key physical and mathematical structures that underly CBS have gone unrecognized. Consequently, CBS's generality has not been fully realized and some CBS results are known only to the specific communities for which they were derived. In this work we expose the key assumptions behind CBS and then use them to show that CBS completely describes a material's static and dynamic electronic properties. We also interpret CBS as the minimal amount of information to accomplish this. In a broader sense, our analysis examines the information content of a material's Hamiltonian so that this information can be exploited to characterize the material's properties. With this interpretation of CBS, we proceed to discuss CBS's applications and unify its many results.
We accomplish this as follows. The remainder of this section provides additional background on CBS, including some of its history and applications. We proceed to introduce a concrete example in section \ref{sec:example}, which illustrates the fundamental properties of CBS. Our primary contribution comes in section \ref{sec:unified}, where we develop our interpretation of CBS by considering the information content of a material's Hamiltonian. This analysis exposes the underlying physical and mathematical structures that imply the existence of CBS. It is also performed at the level of operators, rather than that of matrices and basis sets, thereby eliminating possible ambiguities about basis-set effects. Next, sections \ref{sec:cbs-wf}--\ref{sec:extended-coupling} review existing results of CBS using a consistent nomenclature and notation. Finally, section \ref{sec:conclusions} summarizes our results and speculates on future directions.
\subsection{Background}
\label{sec:introduction:background}
At its core, conventional band structure is rooted in the translational symmetry of a crystalline material, where periodicity of the potential allows invocation of Bloch's theorem. We thus consider wavefunctions of the form
\begin{equation}
\psi_{n, \vec{k}} (\vec{r}) = e^{i \vec{k}\cdot \vec{r}} u_{n, \vec{k}}( \vec{r} ),
\label{eq:bloch-theorem}
\end{equation}
where $\vec{k}$ is a wavevector (usually in the first Brillouin zone), $n$ is the band index, and $u$ is a function with periodicity of the lattice. When used in the Schr\"odinger equation, $\vec{k}$ can be treated as a parameter (\textit{i.e.},\ each $\vec{k}$ produces an independent system), and the eigenstates for a given $\vec{k}$ are calculated using standard techniques. The ensuing dispersion relation for the $n$\textsuperscript{th} band, $E_n(\vec{k})$, details the material's band structure.
Despite the numerous applications of this conventional band structure, many systems of interest lack perfect translational symmetry. Surfaces and interfaces, for instance, break translational symmetry in (at least) one direction. These systems still possess a ``repeat unit'' similar to the periodic unit cell, and Eq.\ \eqref{eq:bloch-theorem} may remain applicable if solutions with complex $\vec{k}$ are also admitted \cite{kohn-809-1959, heine-300-1963, heine-1-1964, krieger-776-1967}.
Imaginary components of $\vec{k}$, when present, indicate evanescence; that is, the state grows or decays in magnitude from one repeat unit to the next. These states have sometimes been called ``generalized Bloch functions''. Conversely, bulk states that propagate throughout the material have real $\vec{k}$. For physicality, we restrict our attention to complex wavevectors that describe states with real energies \cite{heine-300-1963, heine-1-1964, krieger-776-1967}, and the set of such $\vec{k}$ constitutes the material's complex band structure (CBS). In other words, CBS extends the dispersion relation to complex $\vec{k}$, such that it describes both the propagating bulk states from conventional band structure and also any evanescent states that are forbidden by translational symmetry \cite{heine-1-1964}.
Given CBS's broad applicability, it has found use in myriad contexts: %
Surfaces and interfaces \cite{maue-717-1935, goodwin-205-1939, shockley-317-1939, heine-1-1964, feuchtwang-731-1967, marcus-925-1968, garcia-moliner-1789-1969, garcia-moliner-1797-1969, pendry-59-1970, kalkstein-85-1971, jepsen-416-1971, appelbaum-2166-1972, cottey-2583-1972, appelbaum-4973-1974, rehr-448-1974, pendry-87-1975, appelbaum-479-1976, bross-173-1977, allen-917-1979, allen-1454-1979, dy-4237-1979, brasher-4868-1980, lee-4988-1981, lee-4997-1981, lee-355-1981, dy-633-1982, chang-605-1982, chang-3975-1982, kambe-443-1982, schulman-2346-1983, tersoff-465-1984, tersoff-275-1986, wachutka-8512-1986, inglesfield-57-1987, chen-923-1989, garcia-moliner-1405-1990, stampfl-8461-1992, garcia-moliner-332-1994, hummel-1620-1998, mahboob-201307r-2004, demkov-195306-2005, schleife-012014-2009, james-155439-2010, bravi-155445-2014, bk:inglesfield-2015}, %
the construction of Wannier functions \cite{kohn-809-1959, des-cloizeaux-a685-1964, des-cloizeaux-a698-1964, krieger-776-1967, kohn-2485-1973, rehr-1981-1974, rehr-448-1974, he-5341-2001}, %
impurities \cite{schmidt-425-1957, kohn-2485-1973, rehr-1981-1974, inkson-369-1980, blow-359-1980, burt-1825-1980, bylander-4157-1980, blow-3711-1982, blow-5267-1983, kostyrko-3241-1999, hjort-5245-2000, kostyrko-2458-2000, dwivedi-134304-2016}, %
high-energy electron diffraction \cite{buxton-3941-1977}, %
superlattices \cite{inglesfield-162-1971, schulman-4149-1981, krishnamurthy-1027-1985, mailhiot-8360-1986, ram-mohan-6151-1988, chen-923-1989, ghahramani-1102-1989, garcia-moliner-1405-1990, smith-173-1990, ferreira-8198-1998, kostyrko-2458-2000, reynoso-035301-2014}, %
heterostructures \cite{kurtin-756-1970, kurtin-3368-1971, osbourn-2124-1979, schulman-2346-1983, tersoff-4874-1984, marsh-285-1986, ko-9945-1988, ting-3583-1992, boykin-7670-1996, monch-5076-1996, ting-985-1999, demkov-195306-2005, monch-2-2014}, %
quantum wells \cite{brand-607-1987, chen-923-1989, garcia-moliner-1405-1990, boykin-8107-1996, boykin-7670-1996, ogawa-1527-1998}, %
magnetic systems \cite{maclaren-5470-1999, sanvito-11936-1999, mavropoulos-1088-2000, butler-054416-2001, dederichs-108-2002, velev-216601-2005, eames-252511-2006, lukashev-224414-2012, zhang-222401-2012}, %
electron transport \cite{ando-8017-1991, bowen-2754-1995, boykin-7670-1996, kemp-8349-1996, choi-2267-1999, kostyrko-3241-1999, maclaren-5470-1999, kostyrko-2458-2000, wortmann-165103-2002, tomfohr-245105-2002, tomfohr-59-2002, tomfohr-235105-2002, kostyrko-4393-2002, krstic-205319-2002, picaud-3731-2003, tomfohr-1542-2004, fagas-268-2004, khomyakov-195402-2004, pecchia-1497-2004, wang-016401-2004, pomorski-115408-2004, inglesfield-155120-2005, khomyakov-035450-2005, li-194113-2006, lee-215204-2007, zhang-035108-2007, rungger-035407-2008, sorensen-155301-2008, prodan-035124-2009, sorensen-205322-2009, varga-085102-2009, james-155439-2010, vergniory-544-2010, guan-1296-2011, jiang-057202-2012}, %
nanomaterials \cite{choi-2267-1999, kostyrko-3241-1999, ferreira-16040-2000, kostyrko-2458-2000, pomorski-115408-2004, xia-1597-2004, hod-114704-2006, hod-233401-2007, reuter-085412-2011, reuter-084707-2013, reynoso-035301-2014, szczesniak-355301-2016}, %
topological materials \cite{hatsugai-11851-1993, hatsugai-3697-1993, mong-125109-2011, avila-137-2012, dang-155307-2014, reynoso-035301-2014, tauber-115008-2015, dwivedi-134304-2016}, %
solar cells \cite{sulzer-3074-2016}, %
quantum size effects \cite{kalkstein-85-1971, cottey-1734-1971, cottey-2591-1972, cottey-2446-1973, brasher-4868-1980, inglesfield-57-1987, hod-114704-2006, hod-233401-2007, reuter-034703-2010, reuter-085412-2011, reuter-084707-2013}, %
and many others. However, these numerous applications often independently redevelop the main concepts and results of CBS, frequently with different notations or nomenclature. As a result, there are several formulations of CBS that appear unrelated at a quick inspection. Concrete examples of these formulations include %
the transfer or companion matrices \cite{schmidt-425-1957, cottey-1235-1971, wood-1400-1973, lee-4988-1981, lee-4997-1981, lee-355-1981, chang-3975-1982, schulman-2346-1983, strandberg-60-1983, biczo-51-1985, biczo-1992-1985, gies-267-1987, ram-mohan-6151-1988, ko-9945-1988, ghahramani-1102-1989, stampfl-8461-1992, bowen-2754-1995, boykin-8107-1996, ting-985-1999, wortmann-165103-2002, zhang-035108-2007, avila-137-2012, sanchez-soto-191-2012, reynoso-035301-2014, dwivedi-134304-2016},
the propagation matrix \cite{marcus-925-1968, jepsen-416-1971, bross-173-1977}, %
bivariational methods \cite{wachutka-3083-1982}, %
wavefunction matching techniques \cite{goodwin-205-1939, shockley-317-1939, marsh-285-1986, brand-607-1987, wachutka-8512-1986, inglesfield-57-1987, hummel-1620-1998, prodan-035128-2006}, %
and Green function-based approaches \cite{allen-917-1979, dy-4237-1979, mostoller-552-1979, brasher-4868-1980, kambe-443-1982, garcia-moliner-1405-1990, garcia-moliner-332-1994, umerski-5266-1997, reuter-085412-2011}. %
Note that these terms are very broad; \textit{e.g.},\ several different ``transfer matrices'' have been reported. To exacerbate these inconsistencies, many of the CBS derivations are unnecessarily tied to the choice of basis set and/or are laden with application-specific details. Consequently, the handful of studies that compare different CBS formulations \cite{hoffstein-99-1970, chang-3975-1982, wachutka-3083-1982, biczo-51-1985, wachutka-8512-1986, gies-267-1987, chen-923-1989, schulman-6282-1992, garcia-moliner-1405-1990, hummel-1620-1998, wortmann-165103-2002, pecchia-1497-2004, khomyakov-035450-2005} laid crucial groundwork but have yet to illuminate the full generality of CBS.
In this work we present (section \ref{sec:unified}) a unified perspective of CBS that delineates the key underpinnings of CBS that are often unstated or unclear. From these we develop an interpretation for CBS that relates to the amount of information in the material's Hamiltonian. We then use this realization to review previous studies (sections \ref{sec:cbs-wf}--\ref{sec:extended-coupling}) and demonstrate the commonalities among the various CBS formulations.
\section{Fundamentals and a Motivating Example}
\label{sec:example}
Similar to conventional band structure, the dispersion relation, $E(\vec{k})$, is a central quantity in CBS. The difference is that we now consider evanescent states in the material by letting the components of $\vec{k}$ be complex. As in conventional band structure, each $\vec{k}$ produces an eigenvalue problem that can be solved for $E$; in this way we can regard $E(\vec{k})$ as a multi-valued complex function \cite{bk:needham-1997}. Even though there will be multiple (and generally complex) $E$ associated with each complex $\vec{k}$, we are only interested in $\vec{k}$ that produce real $E$. Unless explicitly noted, we assume $E$ is real in what follows. This section uses complex analysis \cite{bk:needham-1997} to establish general properties of $E(\vec{k})$; methods for calculating $E(\vec{k})$ will be discussed in later sections. Without loss of generality [see point \ref{properties:general:reciprocal} below], we restrict $\mathrm{Re}(\vec{k})$ to the first Brillouin zone.
As a concrete example, we begin by introducing a model system that exemplifies the properties of $E(\vec{k})$ we seek to discuss. This one-dimensional, two-band, tight-binding model is schematically depicted in Figure \ref{fig:symmetric}(a), with a layer shaded in gray. Each layer of the material has two orbitals, one with energy $\alpha$ (red circle) and the other with energy $-\alpha$ (white circle). Orbitals within the same layer are coupled by $\beta_1$ (dashed lines), and an orbital couples to the opposite-energy orbital in the two adjacent layers with $\beta_2$ (solid lines). The Hamiltonian is
\begin{equation}
\mathbf{H} = \sum_n \left[ \alpha \mathbf{a}_{n,+}^\dagger \mathbf{a}_{n,+} - \alpha \mathbf{a}_{n,-}^\dagger \mathbf{a}_{n,-} + \left( \beta_1 \mathbf{a}_{n,\pm}^\dagger \mathbf{a}_{n,\mp} + \beta_2 \mathbf{a}_{n+1,\pm}^\dagger \mathbf{a}_{n,\mp} + \mathrm{H.c.}\ \right) \right],
\label{eq:model-h}
\end{equation}
where $\mathbf{a}_{n,\pm}^{(\dagger)}$ is the annihilation (creation) operator for an electron in layer $n$'s orbital with energy $\pm\alpha$, and ``$\mathrm{H.c.}$''\ denotes Hermitian conjugate. Figure \ref{fig:symmetric}(b,c) displays $E(\vec{k})$, that is, the CBS, for this model system with $\alpha=0.8$ eV, $\beta_1=-0.4$ eV, and $\beta_2=-1.3$ eV; the lattice constant is taken to be $a=1$ (in arbitrary length units). These parameters are arbitrarily chosen to provide a representative example that clearly displays the key features of CBS. Examples of more realistic materials can be found in the literature \cite{chang-605-1982, dy-633-1982}.
\begin{figure*}
\resizebox{6.5in}{!}{\includegraphics{fig1-symmetric}}
\caption{\label{fig:symmetric}Complex band structure, $E(\vec{k})$, for the one-dimensional, tight-binding model discussed in the text ($\alpha=0.8$ eV, $\beta_1=-0.4$ eV, $\beta_2=-1.3$ eV, and $a$ is the lattice constant). (a) A schematic of the system. The gray shading displays our choice of layer. (b) A three-dimensional view of the dispersion relation, illuminating Heine's ``lines of real $E$'' \cite{heine-300-1963, heine-1-1964} in complex $\vec{k}$ space. The two lines (red and blue) are also projected onto the base of the plot---the complex $\vec{k}$ plane---for additional perspective. The letters A--H trace out one of the ``lines of real energy'' and are references for the main discussion. (c) A standard, two-dimensional view of the dispersion relation. On the right is the conventional band structure, where $\vec{k}$ is real and in the first Brillouin zone. If a complex $\vec{k}$ produces a real energy $E$, the imaginary part of $\vec{k}$ (in magnitude) is plotted on the left. Black dots in (b) and (c) denote branch points of $E(\vec{k})$.}
\end{figure*}
Development of CBS began by considering the fundamental properties of $E(\vec{k})$ when $\vec{k}$ is only complex in one dimension \cite{kohn-809-1959, heine-300-1963, heine-1-1964, krieger-776-1967, prodan-035128-2006}. Extensions to higher-dimensional cases exist \cite{goodwin-205-1939, blount-305-1962, heine-300-1963, des-cloizeaux-a685-1964, des-cloizeaux-a698-1964, avron-85-1978, inkson-369-1980, avila-137-2012, szczesniak-355301-2016} but are uncommon and relatively undeveloped. Accordingly, we restrict our discussion to one-dimensional cases. This may correspond to an effectively one-dimensional material (\textit{e.g.},\ a nanotube or nanoribbon) or to a three-dimensional material with a surface, where translational symmetry is only broken in one direction. In the latter case, components of $\vec{k}$ parallel to the surface remain good quantum numbers; they are real and can be treated parametrically \cite{heine-300-1963, jones-443-1966, krieger-776-1967, kalkstein-85-1971}. Because our discussion thus focuses on one-dimensional $\vec{k}$, the vector notation will be suppressed.
$E(k)$ is generally complex for an arbitrary, complex $k$, and we are physically interested in real $E$. This is trivially satisfied when $k$ is real (\textit{i.e.},\ conventional band structure); the complex $k$ that lead to real $E$ constitute the CBS. Heine \cite{heine-300-1963, heine-1-1964} and Krieger \cite{krieger-776-1967} showed that these $k$ lie on ``lines of real energy'' in complex $k$-space, which (for our running example) are displayed in Figure \ref{fig:symmetric}(b). The red and blue lines in the figure are the two such lines in our model.
We see that the lines of real energy are continuous and that they only cross a particular energy once such that $E$ can be regarded as a parameter. Let us trace out the trajectory of one of these lines in Figure \ref{fig:symmetric}(b) (the other line is similar). When $E$ is very small (point A), the line is far from the real axis, but approaches the real axis as $E$ increases (toward B). In our example, $\mathrm{Re}(k)$ is constant during this process. Eventually, $E$ hits the edge of a conventional band (B), and the line turns $90^\circ$ to run along the real axis. This is a general result that we discuss in more detail below [point \ref{properties:realE:bandedge}]. $k$ runs along the real axis (B$\to$C) while $E$ continues to increase---forming the conventional band---until $E$ reaches the upper band edge (C). Similar to before, the line of real energy turns $90^\circ$ here and moves off into the complex plane. At some $E$ in the band gap (D) the line turns around and starts heading back toward the real axis. It makes another $90^\circ$ turn at the lower edge (F) of the next band and continues along the real $k$-axis to form the band (F$\to$G). Finally, at the upper edge of the band (G), the line of real energy makes one last $90^\circ$ turn into the complex plane. Further increases in $E$ take the line away from the real axis (toward H). For a slightly different perspective, Figure \ref{fig:symmetric}(c) also plots these lines of real energy in a more traditional style. The right-hand side displays the conventional band structure and the left-hand side shows $|\mathrm{Im}(k)|$ when $k$ is not real.
With this example in mind, let us now briefly discuss some general properties of $E(k)$. A more detailed discussion of these (and other) results is deferred to sections \ref{sec:properties} and \ref{sec:interpretations}. Of greatest importance, Kohn \cite{kohn-809-1959}, Blount \cite{blount-305-1962}, and Krieger \cite{krieger-776-1967} found that $E(k)$ is analytic everywhere except at isolated branch points (see Ref.\ \onlinecite{bk:needham-1997} for a review of branches and branch points). Recalling that $E(k)$ is a multi-valued function, each branch is essentially a band, where a branch point is the boundary between two bands. For reference, there are two branch points in our example that occur at conjugate $k$ values and are marked by black dots in Figure \ref{fig:symmetric}(b,c). It is generally the case that branch points occur in band gaps; that is, they are not on the real axis [see point \ref{properties:general:bp}]. This makes sense: The divide between two bands isn't in either band.
In our example, the branch points were also the ``turn-around points'' for lines of real energy, where the lines reached their maximum distance from the real axis (within the band gap). This is not generally the case \cite{heine-300-1963}, but occurs when there is a mirror plane between the layers [see point \ref{properties:realE:branchpoint} below]. To emphasize this point, Figure \ref{fig:asymmetric} displays the CBS for the same model material with a different choice of ``layer''. Whereas there were only two orbitals in each layer before, there are now four [see Figure \ref{fig:asymmetric}(a)], and the mirror plane between layers is absent. A quick inspection of the ``two-dimensional view'' of CBS in Figure \ref{fig:asymmetric}(c) reveals essentially no differences from Figure \ref{fig:symmetric}(c), reinforcing the idea that the dispersion relation is a material property and independent of the arbitrary choice of layer. Examining the ``three-dimensional view'' [Figure \ref{fig:asymmetric}(b)] shows only one notable difference. The lines of real energy jut around the branch points (and corresponding branch cuts) rather than go through them. Accordingly, the lines of real energy in Figures \ref{fig:symmetric}(b) and \ref{fig:asymmetric}(b) differ only around the branch point and only in $\mathrm{Re}(k)$, which is relatively insignificant to $\mathrm{Im}(k)$ when $\mathrm{Im}(k)\neq0$. As we will see in later sections, even though the dispersion relation (\textit{i.e.},\ the set of eigenvalues from the time-independent Schr\"odinger equation) is largely independent of the choice of layer, the Bloch wavefunctions (the corresponding eigenfunctions) are much more sensitive to this choice.
\begin{figure*}
\resizebox{6.5in}{!}{\includegraphics{fig2-asymmetric}}
\caption{\label{fig:asymmetric}Complex band structure for the same model material as in Figure \ref{fig:symmetric}, but with a different choice of layers that lacks a mirror plane between layers. (a) Schematic of the model material with a layer highlighted in gray. (b) Three-dimensional view of the material's CBS with the new layers. The ``lines of real energy'' are largely indistinguishable from those in Figure \ref{fig:symmetric}(b). The only differences are around the branch points, where the lines now go around the branch points (and corresponding branch cuts) by making jaunts parallel (seemingly) to the real axis. (c) The corresponding two-dimensional view of the material's CBS. Because $\mathrm{Re}(k)$ is not plotted when $\mathrm{Im}(k)\neq0$, there is no apparent difference between this panel and Figure \ref{fig:symmetric}(c). Note that the lattice constant is different with this choice of layer (it is effectively double that from Figure \ref{fig:symmetric}). For ease of comparison, values of $k$ reported here have been shifted and scaled to use the same lattice constant as in Figure \ref{fig:symmetric}.}
\end{figure*}
\section{Unified Complex Band Structure}
\label{sec:unified}
Now that we have some familiarity with the fundamentals of CBS, we turn our focus to computational techniques and discuss methods for calculating CBS. Some general considerations can be found in Refs.\ \onlinecite{jones-443-1966, krieger-776-1967, chang-605-1982, bross-215-1989}, which review the fundamental properties of CBS as well. More specific (and direct) computational procedures have also been developed, as mentioned in section \ref{sec:introduction:background}, but almost always unnecessarily combine formulation with the model or basis set. This, in turn, obscures the inherent generality of CBS and has, perhaps, limited its utility. The goal of this section is to abstract the physical foundations of CBS from the existing literature and to demonstrate that this underlying structure \textit{implies} the existence of CBS. As in section \ref{sec:example}, we focus our attention on one-dimensional CBS.
The key premise of CBS is repetition of a material's ``repeat unit''. The repeat unit may be infinitely tessellated (becoming the ``unit cell''), which results in translational symmetry and conventional band structure. But it may not be. Disorder, perhaps caused by a surface/interface or an impurity, disrupts the repetition. The common factor in all of these cases is that the repeat unit is prevalent throughout the system; that is, \textit{most} layers are identical to the repeat unit.
\subsection{Mathematics of Repetition}
\label{sec:unified:repeat}
As seen in our examples from section \ref{sec:example}, the first step is to identify each ``layer'' within the system. The layer doesn't necessarily have to be an atomic layer, but is simply a way to partition the system and define the repeat units. Mathematically, this is accomplished using orthogonal (Hermitian) projection operators \cite{bk:yanai-2011-ch2}. For example, let $\mathbf{P}_j$ be the projector for layer $j$. Then, $\mathbf{H}_{j,j} \equiv \mathbf{P}_j \mathbf{H} \mathbf{P}_j$ denotes the ``block'' of the Hamiltonian corresponding to layer $j$; likewise,
\begin{equation}
\mathbf{H}_{j,k} \equiv \mathbf{P}_j \mathbf{H} \mathbf{P}_k
\label{eq:def-block}
\end{equation}
is the coupling between layers $j$ and $k$. Because it is built with operators, this projector-based notion of a ``matrix block'' sidesteps any concerns about basis set locality or nonorthogonality and permits a discussion of CBS that is unencumbered by the specific choice of basis set \cite{strandberg-60-1983}. Briefly returning to our examples in section \ref{sec:example}, Figure \ref{fig:symmetric} specified the projectors
\[
\mathbf{P}_n = \mathbf{a}_{n,+}^\dagger \mathbf{a}_{n,+} + \mathbf{a}_{n,-}^\dagger \mathbf{a}_{n,-},
\]
whereas the larger layers in Figure \ref{fig:asymmetric} used
\[
\mathbf{P}_n = \mathbf{a}_{2n+1,+}^\dagger \mathbf{a}_{2n+1,+} + \mathbf{a}_{2n,-}^\dagger \mathbf{a}_{2n,-} + \mathbf{a}_{2n,+}^\dagger \mathbf{a}_{2n,+} + \mathbf{a}_{2n-1,-}^\dagger \mathbf{a}_{2n-1,-}.
\]
We also require (for now) the projectors to be chosen such that each layer only couples to its nearest neighbors. The size of each layer can always be made larger to satisfy this condition, and we will discuss relaxing it in section \ref{sec:extended-coupling}. Mathematically, the Hamiltonian then assumes a block tridiagonal form,
\begin{equation}
\mathbf{H} = \left[
\begin{array}{ccccc}
\mathbf{H}_{1,1} & \mathbf{H}_{1,2} & \mathbf{0} & \mathbf{0} & \cdots \\
\mathbf{H}_{2,1} & \mathbf{H}_{2,2} & \mathbf{H}_{2,3} & \mathbf{0} & \cdots \\
\mathbf{0} & \mathbf{H}_{3,2} & \mathbf{H}_{3,3} & \mathbf{H}_{3,4} & \cdots \\
\mathbf{0} & \mathbf{0} & \mathbf{H}_{4,3} & \mathbf{H}_{4,4} & \cdots \\
\vdots & \vdots & \vdots & \vdots & \ddots \\
\end{array}
\right],
\label{eq:bt-h}
\end{equation}
with $\mathbf{H}_{j,k}=\mathbf{H}_{k,j}^\dagger$ (by Hermiticity). As enumerated in Eq.\ \eqref{eq:bt-h}, the system has a ``surface'' to the left of layer 1 (there is nothing to the left of layer 1). Accordingly, layer 1 is the surface layer, layer 2 the first subsurface layer, and so forth to the ``bulk limit'' at $\infty$ \cite{reuter-034703-2010, reuter-085412-2011, reuter-084707-2013}. In the absence of a surface, the indices would extend to $-\infty$.
If we further neglect other forms of disorder (such as surface reconstructions or impurities), each layer becomes identical to the others. This is the repeat unit. The Hamiltonian is thus also block Toeplitz; that is, blocks are the same along a particular diagonal. We denote $\mathbf{H}_{j,j}=\mathbf{H}_\mathrm{D}$ and $\mathbf{H}_{j,j-1}=\mathbf{H}_\mathrm{S}$ ($\mathbf{H}_{j-1,j}=\mathbf{H}_\mathrm{S}^\dagger$) such that
\begin{equation}
\mathbf{H} = \left[
\begin{array}{ccccc}
\mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} & \mathbf{0} & \cdots \\
\mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} & \cdots \\
\mathbf{0} & \mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger & \cdots \\
\mathbf{0} & \mathbf{0} & \mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \cdots \\
\vdots & \vdots & \vdots & \vdots & \ddots \\
\end{array}
\right].
\label{eq:btbt-h}
\end{equation}
(The ``$\mathrm{D}$'' and ``$\mathrm{S}$'' subscripts stand for ``diagonal'' and ``sub-/super-diagonal'', respectively.) For the time being, we assume that $\mathbf{H}_\mathrm{S}$ is invertible. Section \ref{sec:singular} will discuss relaxing this condition.
For computational purposes (\textit{e.g.},\ the examples in section \ref{sec:example}), we assume that each layer can be sufficiently described by $M$ basis functions, such that the operators $\mathbf{H}_\mathrm{D}$ and $\mathbf{H}_\mathrm{S}$ can be represented by $M\times M$ matrices. Matrix representations for the small layers in Figure \ref{fig:symmetric} are $2\times2$,
\[
\mathbf{H}_\mathrm{D} = \left[ \begin{array}{cc} \alpha & \beta_1^\ast \\ \beta_1 & -\alpha \end{array} \right] \text{ and } \mathbf{H}_\mathrm{S} = \left[ \begin{array}{cc} 0 & \beta_2 \\ \beta_2 & 0 \end{array} \right],
\]
whereas the larger layers in Figure \ref{fig:asymmetric} require $4\times4$ matrices,
\[
\mathbf{H}_\mathrm{D} = \left[ \begin{array}{cccc} \alpha & 0 & \beta_2^\ast & 0 \\ 0 & \alpha & \beta_1^\ast & \beta_2^\ast \\ \beta_2 & \beta_1 & -\alpha & 0 \\ 0 & \beta_2 & 0 & -\alpha \end{array} \right] \text{ and } \mathbf{H}_\mathrm{S} = \left[ \begin{array}{cccc} 0 & 0 & \beta_2 & \beta_1 \\ 0 & 0 & 0 & \beta_2 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right].
\]
Note that $\mathbf{H}_\mathrm{S}$ for the larger layers is not invertible, and provides an example for the discussion in section \ref{sec:singular}. These basis set and matrix details are only necessary for computation and (unless otherwise noted) will not be considered when formulating or discussing the theory of CBS.
\subsection{Block Tridiagonal, Block Toeplitz Matrices}
\label{sec:unified:btbtm}
This block tridiagonal and block Toeplitz structure in the Hamiltonian is foundational to CBS, as it mathematically leads to CBS. We now detail this idea.
Most applications aim to either diagonalize or (essentially) invert the Hamiltonian. The former appears in the context of the time-independent Schr\"odinger equation,
\begin{equation}
\mathbf{H} \ket{\psi} = E \ket{\psi},
\label{eq:schrodinger}
\end{equation}
and the latter in finding the (retarded/advanced) Green function (GF) \cite{bk:economou-2006},
\begin{equation}
\mathbf{G}(E) = \lim_{\eta \to 0^\pm} \left[ (E+ i \eta)\mathbf{I} - \mathbf{H} \right]^{-1}.
\label{eq:gf}
\end{equation}
As foreshadowed, the block tridiagonal and block Toeplitz structure of $\mathbf{H}$ greatly aids in these processes.
To begin, consider the amount of information in $\mathbf{H}$. All that is needed to \textit{completely} describe our system is $\mathbf{H}_\mathrm{D}$, $\mathbf{H}_\mathrm{S}$, and the number of layers (\textit{i.e.},\ the number of block rows/columns). With respect to the size of our system (the number of layers) this is $\mathcal{O}(1)$ information; that is, the amount of information \textit{does not change} with system size. Because $\mathbf{H}$ uniquely determines the eigenvalues (eigenvectors) and the GF, there must only be $\mathcal{O}(1)$ information in these quantities. CBS is this $\mathcal{O}(1)$ information. All reported formulations of CBS access this information to solve specific problems, \textit{e.g.},\ diagonalizing $\mathbf{H}$ for the time-independent Schr\"odinger equation. In this manner, CBS describes, and can be used to understand, all of a material's static and dynamic electronic properties.
Such an analysis of a matrix's information content is not new to the mathematics community, and is one facet of studying (block) quasi-separable (sometimes called rank-structured) matrices \cite{bk:vandebril-2008-v1, bk:vandebril-2008-v2}. As it suffices for our purposes (a more rigorous definition is presented in %
\footnote{A rigorous definition of a block quasi-separable matrix can be found on p. 482 of Ref.\ \onlinecite{bk:vandebril-2008-v1}. Suppose $\mathbf{A}$ is a block matrix with $N$ block rows and $N$ block columns. If $\mathbf{A}$ is block quasi-separable, then its blocks can be written as
\[
\mathbf{A}_{j,k} = \begin{cases}
\mathbf{U}_j \mathbf{T}_{j-1} \mathbf{T}_{j-2} \cdots \mathbf{T}_{k+1} \mathbf{V}_k & \mbox{if } 1\le k < j \le N \\
\mathbf{D}_j & \mbox{if } 1\le j = k \le N \\
\mathbf{P}_j \mathbf{R}_{j+1} \mathbf{R}_{j+2} \cdots \mathbf{R}_{k-1} \mathbf{Q}_k & \mbox{if } 1\le j < k \le N \\
\end{cases},
\]
where the $\mathbf{U}_j$, $\mathbf{T}_j$, $\mathbf{V}_j$, $\mathbf{D}_j$, $\mathbf{P}_j$, $\mathbf{R}_j$, and $\mathbf{Q}_j$ sequences are a set of generators for $\mathbf{A}$.
}%
), a block quasi-separable matrix has certain sub-parts with specific rank. For the present discussion, the sub-parts that contain a sub-diagonal (or super-diagonal) block and the bottom-left (top-right) corner, but do not contain a diagonal block, will have rank equal to that of $\mathbf{H}_\mathrm{S}$. Graphical examples of these sub-parts are
\[
\left[ \begin{array}{cccc}
\mathbf{H}_\mathrm{D} & \tikzmarkin[hor=style grey]{e1} \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} & \mathbf{0} \tikzmarkend{e1} \\
\mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \\
\mathbf{0} & \mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger \\
\mathbf{0} & \mathbf{0} & \mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} \\
\end{array} \right]
\text{ and }
\left[ \begin{array}{cccc}
\mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} & \mathbf{0} \\
\mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \\
\tikzmarkin[hor=style grey]{e2} \mathbf{0} & \mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} & \mathbf{H}_\mathrm{S}^\dagger \\
\mathbf{0} & \mathbf{0} \tikzmarkend{e2} & \mathbf{H}_\mathrm{S} & \mathbf{H}_\mathrm{D} \\
\end{array} \right].
\]
Clearly, block tridiagonal matrices (with or without the block Toeplitz structure) belong to the class of block quasi-separable matrices.
Ramifications of this block quasi-separable structure have been discussed in numerous contexts \cite{bk:vandebril-2008-v1, bk:vandebril-2008-v2}. In many cases, they include analytical characterizations and/or fast numerical algorithms for inverting or diagonalizing a matrix. References \onlinecite{mostoller-552-1979, mostoller-6168-1982, rozsa-447-1989, godfrin-7843-1991, meurant-707-1992, huang-7919-1997, tomfohr-1542-2004, techrpt:jain-2007, koulaei-223-2007, cauley-043713-2011, reuter-014009-2012, boffi-015001-2015} overview specific results for block tridiagonal matrices. In the end, these results stem from finding and exploiting the ``generators'' of the matrix \cite{eidelman-187-2005}; that is, the minimal information needed to generate the entire matrix, its spectral decomposition, or its inverse. The generators are not unique, and some generators may be preferential to others for specific operations (such as inverting or diagonalizing the matrix).
Returning to our block tridiagonal and block Toeplitz Hamiltonian, we see that the set $\{ \mathbf{H}_\mathrm{D}, \mathbf{H}_\mathrm{S}, N\}$ is a generator for $\mathbf{H}$ (where $N$ is the number of layers). We will show in the following sections that the various formulations of CBS produce different generators, as needed for specific applications, but that they all represent the same information (\textit{i.e.},\ CBS). For example, the transfer matrix \cite{lee-355-1981} and the companion matrix \cite{chang-3975-1982} help in diagonalizing $\mathbf{H}$, whereas generators for (in effect) inverting $\mathbf{H}$ to obtain the GF have also been developed \cite{dy-4237-1979, umerski-5266-1997}. In the following sections we derive some of these generators (that is, formulate CBS) and show how they aid in diagonalizing or inverting $\mathbf{H}$. Along the way, we prove that these generators embody the same information (CBS) and then discuss some uses and applications of CBS.
\section{Complex Band Structure from Wavefunctions}
\label{sec:cbs-wf}
Our first formulation of CBS is a wavefunction-based approach that begins with the time-independent Schr\"odinger equation [Eq.\ \eqref{eq:schrodinger}]. In this sense, we start with a particular (real) energy and aim to find the real or complex values of $k$ that produce states with that energy. In the language of complex analysis, we are calculating the multi-valued function $k(E)$, which is essentially the ``inverse'' of the traditional dispersion relation, $E(k)$. In other words, what are the complex (possibly real) wavevectors that produce to states with a given real energy? As we now see, the transfer matrix \cite{lee-355-1981} provides one method for answering this question.
We start by applying a projector (for an arbitrary layer) to the time-independent Schr\"odinger equation. Writing $\ket{\psi_j}\equiv\mathbf{P}_j\ket{\psi}$ and specifically considering layer $n$,
\begin{align}
E \mathbf{P}_n \ket{\psi} & = \mathbf{P}_n \mathbf{H} \ket{\psi}, \nonumber \\
E \ket{\psi_n} & = \mathbf{H}_{n,n-1} \ket{\psi_{n-1}} + \mathbf{H}_{n,n} \ket{\psi_{n}} + \mathbf{H}_{n,n+1} \ket{\psi_{n+1}}, \nonumber \\
& = \mathbf{H}_\mathrm{S} \ket{\psi_{n-1}} + \mathbf{H}_\mathrm{D} \ket{\psi_{n}} + \mathbf{H}_\mathrm{S}^\dagger \ket{\psi_{n+1}},
\label{eq:schrodinger-recurrence}
\end{align}
where we have utilized both the block tridiagonal and block Toeplitz structure of $\mathbf{H}$. If we now define a ``supercell'' wavefunction,
\begin{equation}
\ket{\Psi_{n+1}} \equiv \left[ \begin{array}{c} \ket{\psi_{n+1}} \\ \ket{\psi_n} \end{array} \right],
\label{eq:def-wf-supercell}
\end{equation}
Eq.\ \eqref{eq:schrodinger-recurrence} can be combined with the tautology $\ket{\psi_n}=\ket{\psi_n}$ to write
\begin{align}
\ket{\Psi_{n+1}} & = \left[ \begin{array}{c} -\mathbf{H}_\mathrm{S}^{-\dagger} \mathbf{H}_\mathrm{S} \ket{\psi_{n-1}} + \mathbf{H}_\mathrm{S}^{-\dagger} \left( E \mathbf{I} - \mathbf{H}_\mathrm{D} \right) \ket{\psi_n} \\ \ket{\psi_n} \end{array} \right] \nonumber \\
& = \left[ \begin{array}{cc} \mathbf{H}_\mathrm{S}^{-\dagger} \left( E \mathbf{I} - \mathbf{H}_\mathrm{D} \right) & -\mathbf{H}_\mathrm{S}^{-\dagger} \mathbf{H}_\mathrm{S} \\ \mathbf{I} & \mathbf{0} \end{array} \right] \left[ \begin{array}{c} \ket{\psi_n} \\ \ket{\psi_{n-1}} \end{array} \right] \nonumber \\
& = \mathbf{T} \ket{\Psi_{n}}, \label{eq:supercell-recurrence}
\end{align}
where $^{-\dagger}$ denotes the inverse Hermitian conjugate. In Eq.\ \eqref{eq:supercell-recurrence},
\begin{equation}
\mathbf{T} = \left[ \begin{array}{cc}
\mathbf{H}_\mathrm{S}^{-\dagger} \left( E \mathbf{I} - \mathbf{H}_\mathrm{D} \right) & -\mathbf{H}_\mathrm{S}^{-\dagger} \mathbf{H}_\mathrm{S} \\
\mathbf{I} & \mathbf{0}
\end{array} \right]
\label{eq:transfer-matrix}
\end{equation}
is the transfer matrix, which provides a convenient way to calculate the supercell wavefunction for any layer subject to some base case,
\[
\ket{\Psi_n} = \mathbf{T}^{n-1} \ket{\Psi_1}.
\]
Note that $\mathbf{T}$ depends on $E$ and that transfer matrices of this form have been discussed in more general contexts \cite{molinari-983-1997, molinari-8553-1998, molinari-4081-2003}.
The link between the transfer matrix and CBS results from an invocation of Bloch's theorem [Eq.\ \eqref{eq:bloch-theorem}]. Accordingly,
\[
\ket{\Psi_{n+1}} = \left[ \begin{array}{c} \ket{\psi_{n+1}} \\ \ket{\psi_n} \end{array} \right] = \left[ \begin{array}{c} e^{ika} \ket{\psi_n} \\ e^{ika} \ket{\psi_{n-1}} \end{array} \right] = e^{ika} \ket{\Psi_{n}}.
\]
Substituting this into Eq.\ \eqref{eq:supercell-recurrence} yields
\begin{equation}
e^{ika} \ket{\Psi_{n}} = \mathbf{T} \ket{\Psi_{n}};
\label{eq:transfer-eigval}
\end{equation}
that is, the eigenvalues of $\mathbf{T}$ are essentially the wavevectors that produce states with energy $E$. Consequently, the CBS for a given energy can be obtained from the spectrum of $\mathbf{T}$. As a side note, the CBS of our model system in Figure \ref{fig:symmetric} was obtained from the eigenvalues of $\mathbf{T}$ at each $E$.
\section{Complex Band Structure from Green Functions}
\label{sec:cbs-gf}
In the previous section we used wavefunction arguments to derive the transfer matrix and, ultimately, CBS. We now work through a GF-based approach to CBS based on matrix M\"obius transformations (MMTs) \cite{schwarz-1913-1981, umerski-5266-1997} and surface GFs. In essence, the inverse (\textit{i.e.},\ GF) of a block tridiagonal Hamiltonian can be completely obtained from surface GFs, and MMTs exploit the block Toeplitz structure to provide the surface GFs. This formulation combines and summarizes aspects of Refs.\ \onlinecite{umerski-5266-1997, godfrin-7843-1991, reuter-085412-2011, boffi-015001-2015}, and is a bit more complicated than the derivation of the transfer matrix. To aid in the discussion, we will present various densities of states (DOSs) for our model systems in section \ref{sec:example}. A DOS is related to a retarded GF by \cite{bk:economou-2006}
\begin{equation}
\rho(E) = -\frac{1}{\pi} \mathrm{Im}\left( \mathrm{Tr}\left[ \mathbf{G}(E) \right] \right),
\label{eq:dos}
\end{equation}
where ``$\mathrm{Tr}$'' is the trace.
Section \ref{sec:cbs-gf:mmts} first overviews some pertinent properties of MMTs. Then, section \ref{sec:cbs-gf:sgf} discusses the utility of MMTs in calculating surface GFs. Finally, section \ref{sec:cbs-gf:derive-cbs-gf} puts all of these elements together to arrive at CBS via the GF.
\subsection{Matrix M\"obius Transformations}
\label{sec:cbs-gf:mmts}
The MMT \cite{schwarz-1913-1981, umerski-5266-1997} is a generalization of the M\"obius (bilinear) transformation from complex variables \cite{bk:needham-1997} to matrices. Let $\mathbf{m}_{11}$, $\mathbf{m}_{12}$, $\mathbf{m}_{21}$, $\mathbf{m}_{22}$, and $\mathbf{z}$ be complex $n\times n$ matrices; a MMT is a mapping with the form
\begin{equation}
\mathbf{M} \bullet \mathbf{z} \equiv \left( \mathbf{m}_{11} \mathbf{z} + \mathbf{m}_{12} \right) \left( \mathbf{m}_{21} \mathbf{z} + \mathbf{m}_{22} \right)^{-1}.
\label{eq:matrix-mt}
\end{equation}
Similar to the complex M\"obius transformation \cite{bk:needham-1997}, we can associate with $\mathbf{M}$ the $2n\times 2n$ matrix \cite{umerski-5266-1997}
\[
\mathbf{M} = \left[ \begin{array}{cc} \mathbf{m}_{11} & \mathbf{m}_{12} \\ \mathbf{m}_{21} & \mathbf{m}_{22} \end{array} \right].
\]
A MMT and its matrix representation will be used interchangeably throughout this work. Note that the ``$\bullet$'' in Eq.\ \eqref{eq:matrix-mt} emphasizes that $\mathbf{M}\bullet\mathbf{z}$ is not a canonical matrix-matrix product. Lastly, it is easily verified that the application of MMTs is associative; that is, for any two MMTs $\mathbf{M}_1$ and $\mathbf{M}_2$,
\begin{equation}
\mathbf{M}_1 \bullet \left( \mathbf{M}_2 \bullet \mathbf{z} \right) = \left( \mathbf{M}_1 \mathbf{M}_2 \right) \bullet \mathbf{z},
\label{eq:mmt-associative}
\end{equation}
where $\mathbf{M}_1\mathbf{M}_2$ is the usual matrix-matrix product.
\subsection{Surface Green Functions}
\label{sec:cbs-gf:sgf}
As we will see in section \ref{sec:cbs-gf:derive-cbs-gf}, the GF for a block tridiagonal Hamiltonian can be formulated in terms of surface GFs. This subsection is, therefore, devoted to surface GFs, which have been thoroughly discussed elsewhere \cite{mostoller-552-1979, lopez-sancho-851-1985, umerski-5266-1997, velev-r637-2004, hod-114704-2006, rungger-035407-2008, reuter-085412-2011}. Herein we develop a layer-by-layer approach for surface GFs that parallels the transfer matrix for wavefunctions.
Before proceeding, we need to define the term ``surface GF'' and specify our notation. Using the layer enumeration in Eq.\ \eqref{eq:bt-h}, the surface GF is the $(1,1)$ block of the total system GF; that is,
\begin{equation}
\widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) = \mathbf{P}_1 \mathbf{G}(E) \mathbf{P}_1 \equiv \mathbf{G}_{1,1}(E).
\label{eq:surface-gf}
\end{equation}
Three comments are needed to clarify our notation for surface GFs. First, the tilde signifies a surface GF, as opposed to the full system GF in Eq.\ \eqref{eq:gf}. Second, the surface GF depends on the orientation of the material relative to the surface. For example, the system described in Eq.\ \eqref{eq:bt-h} places the material to the right of the surface, where the surface is effectively between layers 0 and 1. We use the superscript ``$\mathrm{R}$'' to label this case; ``$\mathrm{L}$'' is used when the material is to the left of the surface. For reference, a left surface GF is the bottom-right block of the total system GF. Third, surface GFs for systems containing a finite number of layers will also be important; $\widetilde{\mathbf{G}}_N^\mathrm{L/R}(E)$ denotes the left/right surface GF for a system with $N$ layers. As demonstrated in Eq.\ \eqref{eq:surface-gf}, $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ is the left/right surface GF for a semi-infinite system.
We now show that MMTs can build surface GFs in a layer-by-layer fashion. To overview the procedure, we use L\"owdin partitioning \cite{lowdin-12-1963} (which is essentially an embedding technique \cite{inglesfield-3795-1981, bk:inglesfield-2015}) to construct an effective Hamiltonian for the surface layer that (i) only exists in the surface layer and (ii) is rigorously equivalent to the total system Hamiltonian within the surface layer. This happens at the expense of having an energy-dependent effective Hamiltonian. Because the surface GF is the block of $\mathbf{G}$ for the surface layer, the surface GF is identical to the GF of this effective Hamiltonian. Equation \eqref{eq:sgf:layer-relate} formally states this result, and also relates the surface GF for a system with $N$ layers to the surface GF for a system with $N-1$ layers. MMTs then exploit Eq.\ \eqref{eq:sgf:layer-relate} to obtain the surface GF for an arbitrary number of layers; this realization leads to CBS in the next section. In what follows we detail this idea for right surface GFs, noting that a similar procedure is used for left surface GFs.
For concreteness, Figure \ref{fig:sgf} displays various surface DOSs,
\begin{equation}
\widetilde{\rho}_n^\mathrm{L/R}(E) = \frac{-1}{\pi} \mathrm{Im} \left( \mathrm{Tr}\left[ \widetilde{\mathbf{G}}_n^\mathrm{L/R}(E) \right] \right),
\label{eq:sdos}
\end{equation}
for our ongoing example (using the layers from both Figures \ref{fig:symmetric} and \ref{fig:asymmetric}). When there are only a few layers in the system, the surface DOS is essentially a collection of $\delta$ functions (one for each eigenvalue of $\mathbf{H}$). As the system gets larger, the $\delta$ functions coalesce into bands, ultimately giving rise to the continuous surface DOSs seen for the semi-infinite systems. When there is a mirror plane between the layers (as in Figure \ref{fig:symmetric}), the valence and conduction bands are anti-symmetrically shaped. There is no such similarity between the bands when the mirror plane is absent (\textit{e.g.},\ the layers of Figure \ref{fig:asymmetric}).
\begin{figure}
\resizebox{3.25in}{!}{\includegraphics{fig3-sdos}}
\caption{\label{fig:sgf}Right surface DOSs, given by Eq.\ \eqref{eq:sdos}, for the example systems discussed in section \ref{sec:example}. Left and right columns: Layers chosen using the projectors from Figures \ref{fig:symmetric} and \ref{fig:asymmetric}, respectively. Each row corresponds to systems with a different number of layers. From top to bottom: 1, 2, 10, 20, and the semi-infinite limit. With only 1 layer, the surface DOS shows isolated states (essentially $\delta$ functions). Bands grow as more layers are added, until the surface DOSs converge to the semi-infinite limit. An imaginary energy component of 10~meV was used [see Eq.\ \eqref{eq:gf}].}
\end{figure}
As in the derivation of the transfer matrix, we begin with the time-independent Schr\"odinger equation. To set up L\"owdin partitioning, suppose that $\mathbf{P}_1$ is the projector for our surface layer and $\mathbf{Q}_1\equiv\mathbf{I}-\mathbf{P}_1=\mathbf{P}_2+\mathbf{P}_3+\ldots$\ is the projector for the rest of the system. If we substitute $\mathbf{I}=\mathbf{P}_1+\mathbf{Q}_1$ before and after $\mathbf{H}$ in Eq.\ \eqref{eq:schrodinger}, distribute, and rearrange terms, we get the matrix equation
\begin{equation}
\left[ \begin{array}{cc} \mathbf{P}_1 \mathbf{HP}_1 & \mathbf{P}_1\mathbf{HQ}_1 \\ \mathbf{Q}_1\mathbf{HP}_1 & \mathbf{Q}_1\mathbf{HQ}_1 \end{array} \right] \left[ \begin{array}{c} \mathbf{P}_1 \ket{\psi} \\ \mathbf{Q}_1 \ket{\psi} \end{array} \right] = E \left[ \begin{array}{c} \mathbf{P}_1 \ket{\psi} \\ \mathbf{Q}_1 \ket{\psi} \end{array} \right].
\label{eq:sgf:lowdin}
\end{equation}
The bottom row can be rearranged to produce
\[
\left( E \mathbf{Q}_1 - \mathbf{Q}_1\mathbf{HQ}_1 \right) \mathbf{Q}_1 \ket{\psi} = \mathbf{Q}_1\mathbf{HP}_1 \ket{\psi}.
\]
We observe that the term in parentheses on the left-hand side is the inverse of the GF for the isolated system of all non-surface layers. If we invert this expression \footnote{Note that the inverse of $E\mathbf{Q}_1-\mathbf{Q}_1 \mathbf{HQ}_1$ is not necessarily well-defined because $\mathbf{Q}_1$ is not full rank. The final result, Eq.\ \eqref{eq:sgf:eff-h}, sandwiches this GF between $\mathbf{Q}_1$ projectors, essentially allowing us to use a pseudo-inverse \cite{bk:strang-2006} to mollify the problem.} to obtain the GF for that subsystem (denoted by $\mathbf{G}_{\mathbf{Q}_1}$), we get
\begin{equation}
\mathbf{Q}_1 \ket{\psi} = \mathbf{G}_{\mathbf{Q}_1}(E) \mathbf{Q}_1 \mathbf{HP}_1 \ket{\psi}.
\label{eq:sgf:pq-psi}
\end{equation}
This gives us the wavefunction outside the surface layer ($\mathbf{Q}_1\ket{\psi}$) in terms of the wavefunction in the surface layer ($\mathbf{P}_1\ket{\psi}\equiv\ket{\psi_1}$). Substituting Eq.\ \eqref{eq:sgf:pq-psi} into the top row of Eq.\ \eqref{eq:sgf:lowdin} and rearranging terms yields
\[
\mathbf{P}_1\mathbf{HP}_1 \ket{\psi_1} + \mathbf{P}_1 \mathbf{HQ}_1 \mathbf{G}_{\mathbf{Q}_1}(E) \mathbf{Q}_1 \mathbf{HP}_1 \ket{\psi_1} = E \ket{\psi_1}.
\]
We now have a time-independent Schr\"odinger equation exclusively in the surface layer, where the effective Hamiltonian is
\begin{equation}
\mathbf{P}_1 \mathbf{HP}_1 + \mathbf{P}_1\mathbf{HQ}_1 \mathbf{G}_{\mathbf{Q}_1}(E) \mathbf{Q}_1 \mathbf{HP}_1.
\label{eq:sgf:eff-h}
\end{equation}
Mathematically, this effective Hamiltonian is related to Schur complements \cite{bk:zhang-2005}. The second, energy-dependent term is sometimes called a self-energy, and can be interpreted as an embedding potential or as an open-system boundary condition.
The surface GF for our total system is then obtained by inverting this effective Hamiltonian within the surface layer's subspace (where $\mathbf{P}_1$ is the identity). If our system has $N$ layers, we would obtain the surface GF for a system with $N$ layers,
\[
\widetilde{\mathbf{G}}_N^\mathrm{R}(E) = \left[ E \mathbf{P}_1 - \mathbf{P}_1\mathbf{HP}_1 - \mathbf{P}_1\mathbf{HQ}_1 \mathbf{G}_{\mathbf{Q}_1}(E) \mathbf{Q}_1 \mathbf{HP}_1 \right]^{-1}.
\]
The block tridiagonal structure of $\mathbf{H}$ now allows significant simplifications to this expression. First, $\mathbf{P}_1\mathbf{HQ}_1=\mathbf{H}_{1,2}\mathbf{P}_2$ because the surface layer (\textit{e.g.},\ layer 1) only couples to layer 2. A similar statement holds for $\mathbf{Q}_1\mathbf{HP}_1$. Then,
\[
\widetilde{\mathbf{G}}_N^\mathrm{R}(E) = \left[ E \mathbf{P}_1 - \mathbf{H}_{1,1} - \mathbf{H}_{1,2} \mathbf{P}_2 \mathbf{G}_{\mathbf{Q}_1}(E) \mathbf{P}_2 \mathbf{H}_{2,1} \right]^{-1}.
\]
Second, layer 2 of the total system is a surface layer of the isolated $\mathbf{Q}_1$ subsystem, meaning that $\mathbf{P}_2\mathbf{G}_{\mathbf{Q}_1}(E)\mathbf{P}_2$ is a surface GF for that subsystem, which has one fewer layer. Thus,
\[
\widetilde{\mathbf{G}}_N^\mathrm{R}(E) = \left[ E \mathbf{P}_1 - \mathbf{H}_{1,1} - \mathbf{H}_{1,2} \widetilde{\mathbf{G}}_{N-1}^\mathrm{R}(E) \mathbf{H}_{2,1} \right]^{-1}.
\]
Finally, exploiting the block Toeplitz structure and, for simplicity, replacing $\mathbf{P}_1$ by the identity, we get
\begin{equation}
\widetilde{\mathbf{G}}_N^\mathrm{R}(E) = \left[ E \mathbf{I} - \mathbf{H}_\mathrm{D} - \mathbf{H}_\mathrm{S}^\dagger \widetilde{\mathbf{G}}_{N-1}^\mathrm{R}(E) \mathbf{H}_\mathrm{S} \right]^{-1}.
\label{eq:sgf:layer-relate}
\end{equation}
Equation \eqref{eq:sgf:layer-relate} relates the surface GF for a system with $N$ layers to the surface GF for a system with $N-1$ layers. After some minor rearrangement, we can also show that a MMT describes this relationship between $\widetilde{\mathbf{G}}_N^\mathrm{R}(E)$ and $\widetilde{\mathbf{G}}_{N-1}^\mathrm{R}(E)$. Starting from Eq.\ \eqref{eq:sgf:layer-relate},
\begin{subequations}
\label{eq:sgf-mmt-r}
\begin{align}
\widetilde{\mathbf{G}}_N^\mathrm{R}(E) & = \mathbf{H}_\mathrm{S}^{-1} \left[ (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-1} - \mathbf{H}_\mathrm{S}^\dagger \widetilde{\mathbf{G}}_{N-1}^\mathrm{R}(E) \right]^{-1} \nonumber \\
& = \left[ \begin{array}{cc} \mathbf{0} & \mathbf{H}_\mathrm{S}^{-1} \\ -\mathbf{H}_\mathrm{S}^\dagger & (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-1} \end{array} \right] \bullet \widetilde{\mathbf{G}}_{N-1}^\mathrm{R}(E), \label{eq:sgf-mmt-r-onestep} \\
& = \left[ \begin{array}{cc} \mathbf{0} & \mathbf{H}_\mathrm{S}^{-1} \\ -\mathbf{H}_\mathrm{S}^\dagger & (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-1} \end{array} \right]^{N} \bullet \widetilde{\mathbf{G}}_{0}^\mathrm{R}(E), \label{eq:sgf-mmt-r-recur}
\end{align}
\end{subequations}
where this last step utilizes associativity of MMTs and we define $\widetilde{\mathbf{G}}_0^\mathrm{R}(E)\equiv\mathbf{0}$. For notational convenience moving forward, we give this MMT the symbol
\[
\mathbf{M}_\mathrm{R} = \left[ \begin{array}{cc} \mathbf{0} & \mathbf{H}_\mathrm{S}^{-1} \\ -\mathbf{H}_\mathrm{S}^\dagger & (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-1} \end{array} \right].
\]
Similar to the transfer matrix, $\mathbf{M}_\mathrm{R}$ depends on $E$.
This process can be repeated for left surface GFs. The end result is similar,
\begin{subequations}
\label{eq:sgf-mmt-l}
\begin{align}
\widetilde{\mathbf{G}}_N^\mathrm{L}(E) & = \mathbf{M}_\mathrm{L} \bullet \widetilde{\mathbf{G}}_{N-1}^\mathrm{L}(E), \label{eq:sgf-mmt-l-onestep} \\
& = \mathbf{M}_\mathrm{L}^N \bullet \widetilde{\mathbf{G}}_{0}^\mathrm{L}(E), \label{eq:sgf-mmt-l-recur}
\end{align}
\end{subequations}
where $\widetilde{\mathbf{G}}_0^\mathrm{L}(E)\equiv\mathbf{0}$ and
\[
\mathbf{M}_\mathrm{L} = \left[ \begin{array}{cc} \mathbf{0} & \mathbf{H}_\mathrm{S}^{-\dagger} \\ -\mathbf{H}_\mathrm{S} & (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-\dagger} \end{array} \right].
\]
\subsection{Derivation of CBS}
\label{sec:cbs-gf:derive-cbs-gf}
From Eq.\ \eqref{eq:gf}, calculating the total system GF is tantamount to inverting the Hamiltonian. Mathematically, we thus want to consider the inverse of a block tridiagonal and block Toeplitz matrix. Theories for inverting a block tridiagonal matrix have been discussed numerous times in various contexts \cite{rozsa-447-1989, godfrin-7843-1991, meurant-707-1992, huang-7919-1997, techrpt:jain-2007, koulaei-223-2007, bk:vandebril-2008-v1}, and very recently \cite{reuter-014009-2012, boffi-015001-2015} the block Toeplitz structure with them. In the end, the block quasi-separable structure of the block tridiagonal and block Toeplitz Hamiltonian leads to a structured GF, which we now formulate. As we will soon see, surface GFs play a key role, and, from the previous discussion [Eqs.\ \eqref{eq:sgf-mmt-r} and \eqref{eq:sgf-mmt-l}], MMTs can be used to obtain all of the surface GFs. In this sense, the MMTs $\mathbf{M}_\mathrm{L}$ and $\mathbf{M}_\mathrm{R}$ are generators for $\mathbf{G}(E)$; that is, they embody the material's CBS.
To see this, let us first consider the diagonal blocks of the GF, $\mathbf{G}_{n,n}(E)$. A surface GF is one such block, but there many others unless the system is trivially small. As with the surface GFs above, we use L\"owdin partitioning \cite{lowdin-12-1963} to obtain an expression for $\mathbf{G}_{n,n}(E)$. Suppose $\mathbf{P}_n$ is the projector for our layer of interest (which is not a surface layer). Similar to before, we define $\mathbf{Q}_n=\mathbf{I}-\mathbf{P}_n$ to be the projector for all other layers of the system. A comparable statement to Eq.\ \eqref{eq:sgf:lowdin} can then be made where all $1$ subscripts are replaced by $n$. Following the same logic, we obtain an effective Hamiltonian for layer $n$ and ultimately arrive at
\begin{align*}
\mathbf{G}_{n,n}(E) & = \left[ E \mathbf{P}_n - \mathbf{P}_n\mathbf{HP}_n - \mathbf{P}_n\mathbf{HQ}_n \mathbf{G}_{\mathbf{Q}_n}(E) \mathbf{Q}_n\mathbf{HP}_n \right]^{-1} \\
& = \left[ E \mathbf{I} - \mathbf{H}_\mathrm{D} - \mathbf{P}_n\mathbf{HQ}_n \mathbf{G}_{\mathbf{Q}_n}(E) \mathbf{Q}_n\mathbf{HP}_n \right]^{-1},
\end{align*}
where we have again replaced $\mathbf{P}_n$ with the identity in the energy term.
The key difference here is that there are two disconnected parts of the $\mathbf{Q}_n$ subsystem. Layers $1,2,\ldots,n-1$ are one set, and layers $n+1,n+2,\ldots$\ are the other. Because layer $n$ only couples to layers $n-1$ and $n+1$ in the total system,
\[
\mathbf{P}_n \mathbf{HQ}_n = \mathbf{H}_{n,n-1} \mathbf{P}_{n-1} + \mathbf{H}_{n,n+1} \mathbf{P}_{n+1},
\]
and similarly for $\mathbf{Q}_n \mathbf{HP}_n$. Thus,
\begin{align*}
\mathbf{G}_{n,n}(E) = \left[ E \mathbf{I} \right. & - \mathbf{H}_\mathrm{D} \\*
& \left. - \left( \mathbf{H}_{n,n-1} \mathbf{P}_{n-1} + \mathbf{H}_{n,n+1} \mathbf{P}_{n+1} \right) \mathbf{G}_{\mathbf{Q}_n}(E) \left( \mathbf{P}_{n-1} \mathbf{H}_{n-1,n} + \mathbf{P}_{n+1}\mathbf{H}_{n+1,n} \right) \right]^{-1}.
\end{align*}
After distributing, each of these ``$\mathbf{PGP}$'' terms simplifies. First (and second), $\mathbf{P}_{n\pm1} \mathbf{G}_{\mathbf{Q}_n}(E) \mathbf{P}_{n\mp1}=\mathbf{0}$ because layers $n-1$ and $n+1$ do not belong to the same part of the $\mathbf{Q}_n$ subsystem (\textit{i.e.},\ they are decoupled in the subsystem). Third, layer $n-1$ is a surface layer for the subsystem containing layers $1,2,\ldots,n-1$; therefore (as in the surface GF discussion), $\mathbf{P}_{n-1}\mathbf{G}_{\mathbf{Q}_n}(E)\mathbf{P}_{n-1} = \widetilde{\mathbf{G}}_{n-1}^\mathrm{L}(E)$. Fourth, and similarly, layer $n+1$ is a surface layer for the subsystem containing layers $n+1,n+2,\ldots$\ such that $\mathbf{P}_{n+1}\mathbf{G}_{\mathbf{Q}_n}(E)\mathbf{P}_{n+1}=\widetilde{\mathbf{G}}_{N-n}^\mathrm{R}(E)$, where we assume there are $N$ layers in the total system. Putting these results together,
\begin{equation}
\mathbf{G}_{n,n}(E) = \left[ E \mathbf{I} - \mathbf{H}_\mathrm{D} - \mathbf{H}_\mathrm{S} \widetilde{\mathbf{G}}_{n-1}^\mathrm{L}(E) \mathbf{H}_\mathrm{S}^\dagger - \mathbf{H}_\mathrm{S}^\dagger \widetilde{\mathbf{G}}_{N-n}^\mathrm{R}(E) \mathbf{H}_\mathrm{S} \right]^{-1}.
\label{eq:gf-blocks:diagonal}
\end{equation}
Using our convention that $\widetilde{\mathbf{G}}_0^\mathrm{L/R}(E)=\mathbf{0}$, this equation also handles the cases when $n$ labels a surface layer; compare to Eq.\ \eqref{eq:sgf:layer-relate}.
It is clear from Eq.\ \eqref{eq:gf-blocks:diagonal} that calculating a diagonal block of the GF essentially reduces to calculating surface GFs. Note also that the appropriate surface GF(s) should be replaced by $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ if the material has an infinite or semi-infinite number of layers.
Tying this result back to our running example, Figure \ref{fig:pdos} shows projected DOSs for various layers of our model system, using both choices of layers from Figures \ref{fig:symmetric} and \ref{fig:asymmetric}. Each layer's projected DOS is an instance of Eq.\ \eqref{eq:dos},
\begin{equation}
\rho_n(E) = \frac{-1}{\pi} \mathrm{Im}\left( \mathrm{Tr}\left[ \mathbf{G}_{n,n}(E) \right] \right),
\label{eq:pdos}
\end{equation}
and has been used \cite{kalkstein-85-1971, brasher-4868-1980, inglesfield-57-1987, hod-233401-2007, reuter-034703-2010, reuter-085412-2011, reuter-084707-2013, dang-155307-2014} to study quantum size effects. The projected DOS for layer 1 (\textit{i.e.},\ the surface layer) is, as expected, the surface DOS for the semi-infinite system. As we move away from the surface, oscillations appear in the projected DOSs, which eventually converge to the bulk DOS (essentially layer $\infty$) at a rate determined by CBS [point \ref{interpretations:surfaces:bulkboundary} below]. The van Hove singularities \cite{van-hove-1189-1953} at the band edges of the bulk DOS signify of a one-dimensional material. Although the choice of layer is critically important for surface-related properties, it does not matter for bulk properties.
\begin{figure}
\resizebox{3.25in}{!}{\includegraphics{fig4-pdos}}
\caption{\label{fig:pdos}Projected DOSs, given by Eq.\ \eqref{eq:pdos}, for various layers of the example systems discussed in section \ref{sec:example}. Left and right columns: Layers chosen using the projectors from Figures \ref{fig:symmetric} and \ref{fig:asymmetric}, respectively. Each row corresponds to the projected DOS of a different layer in a semi-infinite system. From top to bottom: Layers 1, 2, 3, 10, 20, and the bulk limit (essentially layer $\infty$). Layer 1 is the surface layer and, as expected, its projected DOS is identical to the surface DOS for a semi-infinite system (see Figure \ref{fig:sgf}). Oscillations appear in the projected DOSs as we proceed from the surface to the bulk; they eventually dampen away, producing the bulk DOS, which does not depend on the choice of layer. An imaginary energy component of 10~meV was used [see Eq.\ \eqref{eq:gf}]. For comparison purposes, the projected DOSs for the large layers (right column) were scaled by a factor of $1/2$, which accounts for the different layer sizes.}
\end{figure}
Let us now consider the off-diagonal blocks of the GF. Expressions for these blocks can be straightforwardly obtained from Dyson's equation \cite{bk:economou-2006}, which, similar to L\"owdin partitioning, has us break up the system. Unlike L\"owdin partitioning, one of our ``partitions'' will only contain the coupling between two layers; the other ``partition'' contains everything else. The idea is best presented through example. Suppose we want to calculate $\mathbf{G}_{n+1,n}(E)$. We regard the coupling between layers $n$ and $n+1$,
\[
\mathbf{V} = \mathbf{P}_n\mathbf{HP}_{n+1} + \mathbf{P}_{n+1}\mathbf{HP}_n,
\]
as a perturbation on the Hamiltonian $\mathbf{H}_0$, which describes layers $1,2,\ldots,n$ decoupled from layers $n+1,n+2,\ldots$. In other words, $\mathbf{H}_0=\mathbf{H}-\mathbf{V}$. Dyson's equation relates the GF for the total system ($\mathbf{G}$) to the GF for the unperturbed system ($\mathbf{G}_0$),
\begin{equation}
\mathbf{G}(E) = \mathbf{G}_0(E) + \mathbf{G}(E)\mathbf{V}\mathbf{G}_0(E).
\label{eq:dyson}
\end{equation}
Applying this to the $(n+1,n)$ block,
\[
\mathbf{G}_{n+1,n}(E) = \mathbf{P}_{n+1}\mathbf{G}_0(E)\mathbf{P}_n + \mathbf{P}_{n+1}\mathbf{G}(E) \left( \mathbf{P}_n\mathbf{HP}_{n+1} + \mathbf{P}_{n+1}\mathbf{HP}_n \right) \mathbf{G}_0(E)\mathbf{P}_n.
\]
Because layers $n$ and $n+1$ are decoupled in $\mathbf{H}_0$, $\mathbf{P}_{n+1}\mathbf{G}_0(E)\mathbf{P}_n=\mathbf{0}$. Layer $n$ is also a surface layer in $\mathbf{H}_0$ such that $\mathbf{P}_n\mathbf{G}_0(E)\mathbf{P}_n=\widetilde{\mathbf{G}}_n^\mathrm{L}(E)$. Thus,
\begin{align*}
\mathbf{G}_{n+1,n}(E) & = \mathbf{P}_{n+1}\mathbf{G}(E) \mathbf{P}_{n+1}\mathbf{HP}_n \mathbf{G}_0(E)\mathbf{P}_n \\
& = \mathbf{G}_{n+1,n+1}(E) \mathbf{H}_\mathrm{S} \widetilde{\mathbf{G}}_n^\mathrm{L}(E).
\end{align*}
This procedure of isolating the coupling between two layers and applying Dyson's equation can be iterated, leading to the general result
\begin{subequations}
\label{eq:gf-blocks}
\begin{equation}
\mathbf{G}_{m,n}(E) = \mathbf{G}_{m,n+1}(E)\mathbf{H}_\mathrm{S}\widetilde{\mathbf{G}}_n^\mathrm{L}(E) \text{ when } m>n.
\label{eq:gf-blocks:subleft}
\end{equation}
Similar logic can be used when $m<n$, and results in
\begin{equation}
\mathbf{G}_{m,n}(E) = \mathbf{G}_{m,n-1}(E) \mathbf{H}_\mathrm{S}^\dagger \widetilde{\mathbf{G}}_{N-n+1}^\mathrm{R}(E) \text{ when } m < n.
\label{eq:gf-blocks:superright}
\end{equation}
Finally, Dyson's equation can equivalently be written as $\mathbf{G}(E) = \mathbf{G}_0(E) + \mathbf{G}_0(E)\mathbf{V}\mathbf{G}(E)$, which produces the equivalent expressions
\begin{align}
\mathbf{G}_{m,n}(E) & = \widetilde{\mathbf{G}}_{N-m+1}^\mathrm{R}(E) \mathbf{H}_\mathrm{S} \mathbf{G}_{m-1,n}(E) \text{ when } m > n, \label{eq:gf-blocks:subright} \\
\mathbf{G}_{m,n}(E) & = \widetilde{\mathbf{G}}_m^\mathrm{L}(E) \mathbf{H}_\mathrm{S}^\dagger \mathbf{G}_{m+1,n}(E) \text{ when } m < n. \label{eq:gf-blocks:superleft}
\end{align}
\end{subequations}
These equations can also be extended to infinite or semi-infinite systems by replacing the appropriate surface GFs with $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$.
Equations\ \eqref{eq:gf-blocks:diagonal} and \eqref{eq:gf-blocks} demonstrate that every block of the GF can be straightforwardly obtained from surface GFs. Then, because the surface GFs are described by the MMTs $\mathbf{M}_\mathrm{L}$ and $\mathbf{M}_\mathrm{R}$, these MMTs transitively generate the entire GF \cite{boffi-015001-2015}.
This last statement suggests that $\mathbf{M}_\mathrm{L}$ and $\mathbf{M}_\mathrm{R}$ access the material's CBS in a way that facilitates matrix inversion; that is, $\mathbf{M}_\mathrm{L}$ and $\mathbf{M}_\mathrm{R}$ provide a GF-based route to CBS. We demonstrate this by showing some common structure between these MMTs and the transfer matrix $\mathbf{T}$ in Eq.\ \eqref{eq:transfer-matrix}. Mathematically, it is straightforward to prove that $\mathbf{M}_\mathrm{L}$ is similar to $\mathbf{T}$,
\[
\mathbf{M}_\mathrm{L} = \left[ \begin{array}{cc} \mathbf{0} & \mathbf{I} \\ \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \end{array} \right] \mathbf{T} \left[ \begin{array}{cc} \mathbf{0} & \mathbf{H}_\mathrm{S}^{-\dagger} \\ \mathbf{I} & \mathbf{0} \end{array} \right];
\]
meaning that $\mathbf{M}_\mathrm{L}$ has the same eigenvalues as $\mathbf{T}$. From section \ref{sec:cbs-wf}, these eigenvalues are essentially the material's CBS. Likewise, $\mathbf{M}_\mathrm{R}$ is similar to $\mathbf{T}^{-1}$,
\[
\mathbf{M}_\mathrm{R} = \left[ \begin{array}{cc} \mathbf{I} & \mathbf{0} \\ \mathbf{0} & \mathbf{H}_\mathrm{S} \end{array} \right] \mathbf{T}^{-1} \left[ \begin{array}{cc} \mathbf{I} & \mathbf{0} \\ \mathbf{0} & \mathbf{H}_\mathrm{S}^{-1} \end{array} \right].
\]
In other words, $\lambda=e^{ika}$ is an eigenvalue of $\mathbf{T}$ (or $\mathbf{M}_\mathrm{L}$) if and only if $1/\lambda=e^{-ika}$ is an eigenvalue of $\mathbf{M}_\mathrm{R}$. For reference,
\[
\mathbf{T}^{-1} = \left[ \begin{array}{cc} \mathbf{0} & \mathbf{I} \\ - \mathbf{H}_\mathrm{S}^{-1} \mathbf{H}_\mathrm{S}^\dagger & \mathbf{H}_\mathrm{S}^{-1}(E \mathbf{I} - \mathbf{H}_\mathrm{D}) \end{array} \right].
\]
This relationship between $\mathbf{T}$, $\mathbf{M}_\mathrm{L}$, and $\mathbf{M}_\mathrm{R}$ is the multi-band generalization of a likewise statement for one-band materials, where $\mathbf{T}$, $\mathbf{M}_\mathrm{L}$, and $\mathbf{M}_\mathrm{R}$ are $2\times 2$ matrices and complex M\"obius transformations were used \cite{sanchez-soto-191-2012}.
\section{Mathematical Properties of Complex Band Structure}
\label{sec:properties}
Equipped with insight from the example in section \ref{sec:example} and these wavefunction- and GF-based routes to CBS, we now discuss some mathematical properties of CBS. Many of the results in this section and the next have been reported in the various derivations of CBS. For brevity, we do not necessarily cite and/or compare every development of the results, but instead focus on the results themselves. Each point will first state the main result. In most cases, more information, and possibly justification, will then be given.
\subsection{The Dispersion Relation}
\label{sec:properties:general}
There are a handful of properties that apply regardless of the material's dimensionality. More detailed discussions of these results can be found in Refs.\ \onlinecite{blount-305-1962, heine-300-1963, krieger-776-1967, prodan-035128-2006}.
\begin{enumerate}[label=(\ref{sec:properties}.\arabic*),align=left,series=properties]
\item \label{properties:general:analytic-dispersion} The dispersion relation $E(\vec{k})$ is analytic \cite{bk:needham-1997} except at $\vec{k}$ where multiple bands intersect (\textit{i.e.},\ branch points). Proofs are presented in Refs.\ \onlinecite{blount-305-1962, des-cloizeaux-a685-1964, krieger-776-1967, prodan-035128-2006}.
\item \label{properties:general:reciprocal} If $\vec{g}$ is a reciprocal lattice vector (and thus real), then $E(\vec{k})=E(\vec{k} + \vec{g})$. This result is easily verified by using Bloch's theorem in the time-independent Schr\"odinger equation.
\item \label{properties:general:conjugation} $E(\vec{k})=E(\vec{k}^\ast)^\ast$, where $^\ast$ denotes complex conjugation. Similar to the previous point, this result can also be verified using Bloch's theorem in the time-independent Schr\"odinger equation. Furthermore, if $E$ is real, then both $\vec{k}$ and $\vec{k}^\ast$ are in the material's CBS. This symmetry between $\vec{k}$ and $\vec{k}^\ast$ can be observed in Figures \ref{fig:symmetric} and \ref{fig:asymmetric} from our example.
\item \label{properties:general:bp} Branch points occur at non-real $\vec{k}$ when the potential (and thus Hamiltonian) are Hermitian \cite{prodan-035128-2006}. However, branch points can be shifted to real $\vec{k}$ if the potential is non-Hermitian, as might happen in (for example) high-energy electron diffraction \cite{buxton-3941-1977}.
\end{enumerate}
\subsection{Heine's ``Lines of Real Energy''}
\label{sec:properties:realE}
Heine \cite{heine-300-1963} derived several results about the ``lines of real $E$'' when $\vec{k}$ is only complex in one dimension that were subsequently classified by Chang \cite{chang-605-1982}. These results are all visible in our example from section \ref{sec:example} (Figures \ref{fig:symmetric} and \ref{fig:asymmetric}), which may be a useful companion.
\begin{enumerate}[resume*=properties]
\item \label{properties:realE:winding} Away from branch points and band edges, lines of real energy do not merge, split, begin, or terminate. Heine \cite{heine-300-1963} demonstrated these results by considering the winding number \cite{bk:needham-1997} of the function $E(k)-E(k_0)$ for some wavevector $k_0$ that is not a branch point or band edge. There is only one zero of this function inside some infinitesimally-small, closed contour around $k_0$, meaning that the function accumulates a phase of $2\pi$ going around the contour. Consequently, there are two points with real energies on the contour: One for the line to enter and one for it to leave. If a line merged, split, began, or terminated, there would have to be an odd number of real energies on the contour.
\item \label{properties:realE:bandedge} At a band edge, two lines of real energy meet and take $90^\circ$ turns into the complex plane. A proof is straightforward. Suppose $k_0$ is a band edge. $E(k)$ has an extremum along the real axis at $k_0$, meaning $E'(k_0)=0$. Then, because $E(k)$ is analytic at $k_0$, we know that
\begin{align*}
E(k) & = E(k_0) + E'(k_0) (k-k_0) + \frac{E''(k_0)}{2} (k-k_0)^2 + \mathcal{O}[(k-k_0)^3] \\
& \approx E(k_0) + \frac{E''(k_0)}{2} (k-k_0)^2
\end{align*}
for $k$ near $k_0$. Because $k_0$ is in a real band, $E(k_0)$ and $E''(k_0)$ are both real, implying that $(k-k_0)^2$ must be real for $E(k)$ to be real. Thus, $k-k_0$ is either real (meaning $k$ is also in the conventional band) or purely imaginary (corresponding to a $90^\circ$ turn). From this reasoning, we also see that a band edge $k_0$ is a saddle point of $E(k)$.
\item \label{properties:realE:branchpoint} Lines of real energy do not generally go through branch points. In other words, if $k_\mathrm{bp}$ is a branch point, $E(k_\mathrm{bp})$ is not usually real. See the example in Figure \ref{fig:asymmetric}. The one major exception \cite{heine-300-1963} to this statement occurs when there is a mirror plane between between the layers, as demonstrated in Figure \ref{fig:symmetric}. $E(k_\mathrm{bp})$ will be real in this case. Either way \cite{heine-300-1963},
\begin{equation}
E(k) \approx E(k_\mathrm{bp}) + C \left( k - k_\mathrm{bp} \right)^{1/2},
\label{eq:dispersion-bp}
\end{equation}
for $k$ near $k_\mathrm{bp}$ and some (generally complex) constant $C$. Note that Eq.\ \eqref{eq:dispersion-bp} is only valid in the likely case that two bands meet at the branch point [see point \ref{interpretations:general:bp}]. If $m$ bands were to intersect at $k_\mathrm{bp}$, the exponent $1/2$ would be replaced by $1/m$ \cite{krieger-776-1967}.
\end{enumerate}
\subsection{One-Dimensional Complex Band Structure}
\label{sec:properties:1d}
Because one-dimensional CBS is related to the eigenvalues of the transfer matrix $\mathbf{T}$ (alternatively and equivalently, the MMTs $\mathbf{M}_\mathrm{L}$ and $\mathbf{M}_\mathrm{R}$), we focus on the spectral decomposition of $\mathbf{T}$.
\begin{enumerate}[resume*=properties]
\item \label{properties:1d:nonhermitian} The transfer matrix is not Hermitian. As desired, this means that $\mathbf{T}$ can have complex eigenvalues.
\item \label{properties:1d:pairedevals} The eigenvalues of $\mathbf{T}$ come in pairs: If $\lambda$ is an eigenvalue, so is $1/\lambda^\ast$. To see this, we verify that $\mathbf{T}$ is similar to $\mathbf{T}^{-\dagger}$ (mathematically, $\mathbf{T}$ is complex symplectic),
\[
\mathbf{T} = \left[ \begin{array}{cc} \mathbf{0} & -\mathbf{H}_\mathrm{S}^{-\dagger} \\ \mathbf{H}_\mathrm{S}^{-1} & \mathbf{0} \end{array} \right] \mathbf{T}^{-\dagger} \left[ \begin{array}{cc} \mathbf{0} & \mathbf{H}_\mathrm{S} \\ -\mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \end{array} \right].
\]
\item \label{properties:1d:evaldistribution} $\mathbf{T}$ has an even (possibly zero) number of eigenvalues with unit magnitude, and an equal number of eigenvalues with magnitudes greater than and less than 1. As follows from Fact 2.14.13 of Ref.\ \onlinecite{bk:bernstein-2009},
\[
\mathrm{det}\left(\mathbf{T}\right)=\mathrm{det}\left(\mathbf{H}_\mathrm{S}^{-\dagger}\mathbf{H}_\mathrm{S}\right) = \frac{\mathrm{det}(\mathbf{H}_\mathrm{S})}{\mathrm{det}(\mathbf{H}_\mathrm{S})^\ast}.
\]
Thus, $|\mathrm{det}(\mathbf{T})|=1$, meaning the product of all eigenvalues is 1 (in magnitude). Appealing to the eigenvalue pairing in point \ref{properties:1d:pairedevals}, it is straightforward to see the conclusion. Suppose that $\lambda$ is an eigenvalue of $\mathbf{T}$ with unit magnitude. Then, $1/\lambda^\ast$ has unit magnitude and is also an eigenvalue. Similar reasoning reveals that there are an equal number of eigenvalues with $|\lambda|<1$ and $|\lambda|>1$.
\item \label{properties:1d:eval1realE} Eigenvalues with unit magnitude [see points \ref{interpretations:1d:propagate} and \ref{interpretations:1d:evanescent} below] are forbidden when $E$ is not real. This situation arises, for example, when numerically evaluating the limit in Eq.\ \eqref{eq:gf} and a small imaginary component is present in the energy. In essence, the imaginary energy breaks symmetry so that the paired eigenvalues are no longer degenerate. Half of the states will have $|\lambda|<1$ and the other half will have $|\lambda|>1$ \cite{velev-r637-2004}. Note that, as expected, $|\lambda|\to1$ as $\mathrm{Im}(E)\to0^\pm$ for these eigenvalues.
A proof of this result is straightforward and proceeds by contradiction. Suppose that $\lambda$ is an eigenvalue of $\mathbf{T}$ ($|\lambda|=1$) when $\mathrm{Im}(E)\neq0$. Then, from Eq.\ \eqref{eq:transfer-matrix} and Fact 2.14.13 in Ref.\ \onlinecite{bk:bernstein-2009},
\begin{eqnarray}
0 & = & \mathrm{det} \left[ \lambda \mathbf{I} - \mathbf{T} \right] \nonumber \\
& = & \mathrm{det} \left[ \begin{array}{cc} \lambda \mathbf{I} - \mathbf{H}_\mathrm{S}^{-\dagger} (E\mathbf{I} - \mathbf{H}_\mathrm{D}) & \mathbf{H}_\mathrm{S}^{-\dagger} \mathbf{H}_\mathrm{S} \\ -\mathbf{I} & \lambda \mathbf{I} \end{array} \right] \nonumber \\
& = & \mathrm{det} \left[ -\lambda \mathbf{H}_\mathrm{S}^{-\dagger} \right] \mathrm{det} \left[ E \mathbf{I} - \lambda \mathbf{H}_\mathrm{S}^\dagger - \mathbf{H}_\mathrm{D} - \frac{1}{\lambda} \mathbf{H}_\mathrm{S} \right]. \nonumber
\end{eqnarray}
The first determinant is nonzero because $|\lambda|=1$ and $\mathbf{H}_\mathrm{S}^{-\dagger}$ is invertible. Considering the second determinant, we can write $\lambda=e^{i\theta}$ ($\theta$ is real) such that
\begin{equation}
0 = \mathrm{det} \left[ E \mathbf{I} - e^{i\theta} \mathbf{H}_\mathrm{S}^\dagger - \mathbf{H}_\mathrm{D} - e^{-i\theta} \mathbf{H}_\mathrm{S} \right].
\label{eq:proof-lambda}
\end{equation}
Recalling that $\mathbf{H}_\mathrm{D}$ is Hermitian, it is easily verified that $e^{i\theta} \mathbf{H}_\mathrm{S}^\dagger + \mathbf{H}_\mathrm{D} + e^{-i\theta} \mathbf{H}_\mathrm{S}$ is also Hermitian. Then, because Eq.\ \eqref{eq:proof-lambda} can only hold when $E$ is an eigenvalue of $e^{i\theta} \mathbf{H}_\mathrm{S}^\dagger + \mathbf{H}_\mathrm{D} + e^{-i\theta} \mathbf{H}_\mathrm{S}$, we reach a contradiction. Hermitian operators must have real eigenvalues and $E$ is not real. Thus, eigenvalues of $\mathbf{T}$ cannot have unit magnitude when $E$ is not real.
\item \label{properties:1d:diagonalizable} $\mathbf{T}$ may not be diagonalizable. In general, the transfer matrix is not normal, meaning $\mathbf{TT}^\dagger\neq\mathbf{T}^\dagger\mathbf{T}$. Consequently, $\mathbf{T}$ is not guaranteed to have a complete set of eigenvectors and any eigenvectors it does possess are not required to be mutually orthogonal \cite{bk:golub-2012}. Rather than diagonalizing $\mathbf{T}$, a Jordan or Schur decomposition \cite{bk:golub-2012} (which both generalize the notion of ``diagonalization'') may be more appropriate for analytical or numerical considerations, respectively. Additional structure of the eigenvectors can be found in Ref.\ \onlinecite{biczo-1992-1985}.
\end{enumerate}
\section{Physical Interpretations and Applications of Complex Band Structure}
\label{sec:interpretations}
The mathematical properties of CBS discussed in the last section set the stage for physical interpretations of CBS, which we now discuss. As before, we first state the main result for each point and then offer some discussion or justification. Some of the justifications are detailed (particularly those in Sec.\ \ref{sec:interpretations:surfaces}), but are not necessary for proceeding to subsequent points. We provide these details as references for the interested reader.
\subsection{The Dispersion Relation}
\label{sec:interpretations:general}
\begin{enumerate}[label=(\ref{sec:interpretations}.\arabic*),align=left,series=interpretations]
\item \label{interpretations:general:branch} Each branch of the multi-valued dispersion relation corresponds to a band of the material. In this context, a ``band'' includes the real band from conventional band structure, as well as some complex wavevectors. Prodan \cite{prodan-035128-2006} further discusses, with examples of the various Riemann surfaces, branches and branch points of $E(k)$ when $k$ is strictly one-dimensional.
\item \label{interpretations:general:bp} Branch points of $E(\vec{k})$ occur at the boundaries between two bands. Consequently, they will only appear in band gaps. It may be possible that branch points will appear inside a band for higher-dimensional (not one-dimensional) materials \cite{mahboob-201307r-2004, schleife-012014-2009}; however, in the absence of general tools for higher-dimensional CBS, heuristics were used to calculate those branch point energies \footnote{Heuristic branch point energies for higher-dimensional materials were obtained by averaging over a heuristic branch point energy for each real wavevector in the material's first Brillouin zone. It is possible that this average is above the lower band edge (which comes from a specific $\vec{k}$) of the conduction band or below the upper band edge of the valence band.}.
\item \label{interpretations:general:character} If $\vec{k}_\mathrm{bp}$ is a branch point [and therefore in a band gap from point \ref{interpretations:general:bp}], states near $\vec{k}_\mathrm{bp}$ with $E<E(\vec{k}_\mathrm{bp})$ will have more ``valence band character'', and similarly for ``conduction band character'' when $E>E(\vec{k}_\mathrm{bp})$ \cite{appelbaum-4973-1974, rehr-448-1974, rehr-1981-1974}. Using chemistry nomenclature, states near the branch point with $E=E(\vec{k}_\mathrm{bp})$ are ``non-bonding''.
\end{enumerate}
\subsection{One-Dimensional Complex Band Structure}
\label{sec:interpretations:1d}
\begin{enumerate}[resume*=interpretations]
\item \label{interpretations:1d:propagate} Eigenvalues of $\mathbf{T}$ with unit magnitude correspond to the bulk-propagating states from conventional band structure. Briefly, if $\lambda=e^{ika}$ and $|\lambda|=1$, then $\mathrm{Im}(k)=0$. Because $k$ is real, $\lambda$ and $1/\lambda^\ast$ are degenerate eigenvalues. If $E$ is outside of a band, there will not be any of these states or eigenvalues.
\item \label{interpretations:1d:evanescent} Eigenvalues of $\mathbf{T}$ with non-unit magnitudes correspond to the evanescent states that are excluded from conventional band structure. Depending on orientation, states corresponding to $|\lambda|<1$ grow or decay from one layer to the next, whereas states with $|\lambda|>1$ exhibit the opposite behavior. From point \ref{properties:1d:evaldistribution}, there are an equal number of states with $|\lambda|>1$ and $|\lambda|<1$.
\item \label{interpretations:1d:defective} The states corresponding to defective eigenvalues of $\mathbf{T}$ (\textit{i.e.},\ the degenerate eigenvalues that lack a complete set of eigenvectors, thereby prohibiting $\mathbf{T}$ from being diagonalized) are not well understood. Some results \cite{blount-305-1962, cottey-1235-1971, cottey-2583-1972, cottey-2591-1972, buxton-3941-1977, biczo-31-1979, biczo-347-1981, biczo-1992-1985, biczo-503-1998, reuter-034703-2010, reuter-084707-2013} suggest that $\mathbf{T}$ is defective at a band edge or a branch point, or when a surface state exists [point \ref{interpretations:surfaces:states}]. More work needs to be performed to understand the significance of defective $\mathbf{T}$.
\item \label{interpretations:1d:dynamics} CBS determines the dynamic properties of electrons in a material \cite{des-cloizeaux-a685-1964}, which are described by the off-diagonal blocks of the GF. To see this, we apply general properties of MMTs \cite{boffi-015001-2015}. Equation \eqref{eq:gf-blocks:subright} takes a lower-triangular block of the GF and moves one layer down, and can be approximated as
\[
\mathbf{G}_{j,k}(E) \approx \widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) \mathbf{H}_\mathrm{S} \mathbf{G}_{j-1,k}(E),
\]
when $j$ is sufficiently large [see point \ref{interpretations:surfaces:converge} below]. Then, using Eq.\ \eqref{eq:gf-fp},
\begin{eqnarray}
\mathbf{G}_{j,k}(E) & = & \left( \mathbf{M}_\mathrm{R} \bullet \widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) \right) \mathbf{H}_\mathrm{S} \mathbf{G}_{j-1,k}(E) \nonumber \\
& = & \mathbf{H}_\mathrm{S}^{-1} \left[ -\mathbf{H}_\mathrm{S}^\dagger \widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) + (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-1} \right]^{-1} \mathbf{H}_\mathrm{S} \mathbf{G}_{j-1,k}(E). \nonumber
\end{eqnarray}
We now invoke a general property of MMTs (Thm.\ 2 of Ref.\ \onlinecite{boffi-015001-2015}): The eigenvalues of $-\mathbf{H}_\mathrm{S}^\dagger \widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) + (E\mathbf{I} - \mathbf{H}_\mathrm{D}) \mathbf{H}_\mathrm{S}^{-1}$ are exactly the eigenvalues of $\mathbf{M}_\mathrm{R}$ with $|\lambda|>1$ [assuming $\mathrm{Im}(E)\neq0$]. In this sense, the material's CBS determines the behavior of the GF as we move off the diagonal; these are the dynamical properties of electrons. If $E$ is outside of a band, the electron transition probabilities fall off rapidly with each additional layer because there are no propagating states. Conversely, if $E$ is in a band, there is at least one eigenvalue satisfying $|\lambda|\approx1$, and the off-diagonal blocks of the GF are slow to decay. As a final comment, a similar analysis can be performed for each part of Eq.\ \eqref{eq:gf-blocks}, and leads to the same result.
\end{enumerate}
\subsection{Surfaces}
\label{sec:interpretations:surfaces}
Surfaces were one of the first applications of CBS. Notably, and as discussed in section \ref{sec:cbs-gf}, CBS is closely linked to the surface GF for a semi-infinite system \cite{garcia-moliner-1789-1969, dy-4237-1979, lee-4997-1981, lee-355-1981, umerski-5266-1997, velev-r637-2004, rungger-035407-2008, reuter-085412-2011, bravi-155445-2014}. Herein we describe some additional results related to surfaces.
\begin{enumerate}[resume*=interpretations]
\item \label{interpretations:surfaces:fp} The surface GF (for a semi-infinite system) can be directly obtained from the eigenvectors of the transfer matrix \cite{lee-4997-1981} or the MMTs $\mathbf{M}_\mathrm{L}$ and $\mathbf{M}_\mathrm{R}$ \cite{umerski-5266-1997, reuter-085412-2011}; see Eq.\ \eqref{eq:surface-gf-mmt}. (Generalized eigenvectors may also be needed if the transfer matrix is not diagonalizable.) This explains why the surface GFs in Figure \ref{fig:sgf} are different for the two choices of layers. Even though the eigenvalues of $\mathbf{M}_\mathrm{R}$, $\mathbf{M}_\mathrm{L}$, or $\mathbf{T}$ (\textit{i.e.},\ the dispersion relation) are mostly independent of the choice of layer (see Figures \ref{fig:symmetric} and \ref{fig:asymmetric}), the eigenvectors are much more sensitive.
To elaborate, consider a semi-infinite material with (right) surface GF $\widetilde{\mathbf{G}}_\infty^\mathrm{R}(E)$. When applied to a surface GF, the MMT $\mathbf{M}_\mathrm{R}$ produces the surface GF for a system with an additional layer. This new system, however, is identical to the original semi-infinite system; thus,
\begin{equation}
\widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) = \mathbf{M}_\mathrm{R} \bullet \widetilde{\mathbf{G}}_\infty^\mathrm{R}(E).
\label{eq:gf-fp}
\end{equation}
Mathematically, the surface GF for a semi-infinite system is a fixed point of $\mathbf{M}_\mathrm{R}$.
We previously showed \cite{boffi-015001-2015} that the fixed points of a MMT are related to the MMT's invariant subspaces \cite{bk:gohberg-2006}. In other words, a fixed point corresponds to a collection of eigenvectors of the MMT (and possibly generalized eigenvectors if the MMT is not diagonalizable). Physically, we are interested in the half of the eigenvectors that correspond to right- (or left-)propagating states; that is, those with eigenvalues greater than (or less than) 1 in magnitude. Note that, if there are eigenvalues with unit magnitude (\textit{i.e.},\ $E$ is in a band) such that half of the eigenvalues are not greater (less) than 1 in magnitude [point \ref{properties:1d:pairedevals}], it is well known that the surface GF is not well-defined. Adding a small imaginary part to the energy [point \ref{properties:1d:eval1realE}] eliminates this problem.
A proof of the upcoming result is lengthy and requires other properties of MMTs that we will not discuss here. For brevity, we simply state the result and direct the interested reader to Ref.\ \onlinecite{boffi-015001-2015} for more details. We also have to impose a basis set and assume that $\mathbf{H}_\mathrm{D}$ and $\mathbf{H}_\mathrm{S}$ are accurately represented by $M\times M$ matrices \footnote{The requirement of a basis in point \ref{interpretations:surfaces:fp} can be lifted by appealing to the projective space interpretation of a MMT. More details can be found in Ref.\ \onlinecite{boffi-015001-2015}.}. Let $\mathbf{U}$ be the $2M\times M$ matrix whose columns are the $M$ (generalized) eigenvectors of $\mathbf{M}_\mathrm{R}$ with eigenvalues greater than 1 in magnitude (recall that $\mathbf{M}_\mathrm{R}$ will be a $2M\times 2M$ matrix). Then,
\begin{equation}
\widetilde{\mathbf{G}}_\infty^\mathrm{R}(E) = \mathbf{U}_1 \mathbf{U}_2^{-1},
\label{eq:surface-gf-mmt}
\end{equation}
where $\mathbf{U}_1$ is the first $M$ rows of $\mathbf{U}$ and $\mathbf{U}_2$ is the bottom $M$ rows of $\mathbf{U}$. Likewise, if $\mathbf{V}$ is the $2M\times M$ matrix of eigenvectors of $\mathbf{M}_\mathrm{L}$ with eigenvalues greater than 1 in magnitude, then
\[
\widetilde{\mathbf{G}}_\infty^\mathrm{L}(E) = \mathbf{V}_1 \mathbf{V}_2^{-1},
\]
where $\mathbf{V}_1$ and $\mathbf{V}_2$ are similarly defined.
\item \label{interpretations:surfaces:converge} CBS determines the rate at which $\widetilde{\mathbf{G}}_N^\mathrm{L/R}(E)$ converges to $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ with increasing $N$ (if it converges), which is part of an examination of quantum size effects. This result is also relevant to numerical methods that approximate $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ by $\widetilde{\mathbf{G}}_N^\mathrm{L/R}(E)$ for $N$ sufficiently large \cite{lopez-sancho-851-1985, velev-r637-2004}. Specifically, $\widetilde{\mathbf{G}}_N^\mathrm{L/R}(E)$ converges to $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ as $\mathcal{O}(|\lambda_M|^{2N})$, where $\lambda_M$ is the largest (in magnitude) eigenvalue of $\mathbf{M}_\mathrm{L/R}$ satisfying $|\lambda_M|\le1$. As implied, $\widetilde{\mathbf{G}}_N^\mathrm{L/R}(E)$ does not converge to the semi-infinite limit when $|\lambda_M|=1$. In this case, $E$ is in a band and $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ is not well-defined.
Starting from Eq.\ \eqref{eq:sgf-mmt-r-recur} and using the Jordan form of $\mathbf{M}_\mathrm{R}=\mathbf{U}_\mathrm{R}\mathbf{J}\mathbf{U}_\mathrm{R}^{-1}$,
\[
\widetilde{\mathbf{G}}_N^\mathrm{R}(E) = \left( \mathbf{U}_\mathrm{R} \mathbf{J}^N \mathbf{U}_\mathrm{R}^{-1} \right) \bullet \widetilde{\mathbf{G}}_0^\mathrm{R}(E).
\]
Similar to the discussion in point \ref{interpretations:surfaces:fp}, we assume that $\mathrm{Im}(E)>0$ such that half of the eigenvalues in $\mathbf{J}$ are greater than 1 (in magnitude) and the other half are less than 1 (in magnitude); see points \ref{properties:1d:evaldistribution} and \ref{properties:1d:eval1realE}. We can then order the eigenvalues in $\mathbf{J}$ so that each element of $\mathbf{U}_\mathrm{R}^{-1}\bullet \widetilde{\mathbf{G}}_0^\mathrm{R}(E)$ is multiplied by an eigenvalue that is smaller than 1 (in magnitude) and simultaneously divided by an eigenvalue that is larger than 1 (in magnitude) when $\mathbf{J}^N$ is applied. Consequently, each element of $(\mathbf{J}^{N}\mathbf{U}_\mathrm{R}^{-1})\bullet \widetilde{\mathbf{G}}_0^\mathrm{R}(E)$ exponentially converges to $0$ with a rate determined by CBS. The slowest element to approach zero converges as $\mathcal{O}(|\lambda_M|^{2N})$.
\item \label{interpretations:surfaces:states} CBS and the (semi-infinite) surface GF detail the existence of surface states and their asymptotic decay rates. Surface states appear at energies where there are no bulk-propagating states (that is, no $\lambda$ have unit magnitude) and $\widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E)$ has a pole \cite{dy-4237-1979, lee-4988-1981}. The latter condition is satisfied when, mathematically (with a momentary abuse of notation),
\[
0 = \mathrm{det}\left[ \left( \widetilde{\mathbf{G}}_\infty^\mathrm{L/R}(E) \right)^{-1} \right].
\]
From Eq.\ \eqref{eq:surface-gf-mmt}, this condition rigorously becomes
\begin{equation}
0 = \mathrm{det}\left( \mathbf{U}_2 \right) \text{ or } 0 = \mathrm{det}\left( \mathbf{V}_2 \right),
\label{eq:surface-state-condition}
\end{equation}
with $\mathbf{U}_2$ and $\mathbf{V}_2$ defined as in point \ref{interpretations:surfaces:fp}.
Last, if there exists a surface state at energy $E$, the state asymptotically decays (in magnitude) at a rate of $-\ln|\lambda_{M}|$ per layer \cite{lee-4997-1981}, where $\lambda_M$ is the largest eigenvalue (in magnitude) satisfying $|\lambda_M|<1$. Physically, $\lambda_M$ corresponds to the most slowly decaying evanescent state from the material's CBS.
\item \label{interpretations:surfaces:bulkboundary} CBS describes the general decay of surface effects \cite{reuter-085412-2011, reuter-084707-2013}, which helps explore quantum size effects. As depicted in Figure \ref{fig:pdos}, the diagonal blocks of the GF (which provide access to the projected DOSs) are one way to investigate this behavior. From Eq.\ \eqref{eq:gf-blocks:diagonal}, diagonal blocks of the GF are easily calculated from surface GFs for systems with the appropriate numbers of layers. More specifically, $\mathbf{G}_{j,j}(E)$ requires the surface GFs for the system with all layers to the left of layer $j$ [$\widetilde{\mathbf{G}}_{j-1}^\mathrm{L}(E)$] and for the system with all layers to the right of layer $j$ [$\widetilde{\mathbf{G}}_{N-j}^\mathrm{R}(E)$]. We thus see that surface effects persist as long as these two surface GFs do not resemble the surface GFs for semi-infinite systems; the ``bulk'' material would see semi-infinite systems on both sides. Because point \ref{interpretations:surfaces:converge} describes how CBS determines the convergence of these surface GFs to their semi-infinite limits, the subsurface properties of a material are also tied to CBS.
\end{enumerate}
\subsection{Interfaces}
\label{sec:interpretations:interfaces}
Many of the CBS results for surfaces readily generalize to interfaces \cite{mostoller-552-1979, brasher-4868-1980, schulman-2346-1983}, where two materials come into contact with each other.
\begin{enumerate}[resume*=interpretations]
\item \label{interpretations:interfaces:levelalignment} CBS describes the band ``lineup'' between the materials on both sides of the interface when neither material is metallic \cite{tersoff-4874-1984, margaritondo-2526-1985, demkov-195306-2005, monch-1113724-2011}. In almost every case, the two materials will have different workfunctions, meaning that electrons will spontaneously flow from one material to the other when the interface forms. The Fermi energies of the ``donor'' and ``acceptor'' materials will locally fall and rise near the interface, respectively, until there is no more charge transfer across the interface. This results in an interface dipole. Tersoff \cite{tersoff-4874-1984} posited that charge is transferred from ``conduction-like'' states of the donor to ``valence-like'' states of the acceptor, implying that equilibration occurs when ``non-bonding'' states from both materials are the next to be used. Using point \ref{interpretations:general:character}, the band lineup is thus determined by the branch point energies of the materials. This idea seems to work in many, but not all, cases \cite{monch-1113724-2011}.
\item \label{interpretations:interfaces:migs} Metal-induced gap states (MIGSs) at metal-semiconductor interfaces arise from semiconductor states with complex $k$ \cite{tersoff-465-1984}. In the band gap, these states penetrate several layers into the metal, but decay according to $\mathrm{Im}(k)$. Using similar reasoning to point \ref{interpretations:interfaces:levelalignment}, the metal's Fermi energy is pinned at or near the branch point energy of the semiconductor. This interpretation has also been used to describe the level alignment between an oligomeric molecule and metal surfaces for molecular electronics applications \cite{tomfohr-245105-2002, tomfohr-59-2002, tomfohr-1542-2004, wang-016401-2004}, although it may not always be appropriate \cite{lee-215204-2007}.
\end{enumerate}
\section{Singular Coupling Operators}
\label{sec:singular}
Sections \ref{sec:cbs-wf} and \ref{sec:cbs-gf} both assumed that the interlayer coupling operator, $\mathbf{H}_\mathrm{S}$, was nonsingular and, subsequently, used $\mathbf{H}_\mathrm{S}^{-1}$ in deriving CBS. Analytically, $\mathbf{H}_\mathrm{S}$ is probably not singular---there is infinitesimally weak coupling---but, in practice, it may be poorly conditioned such that numerically calculating or using $\mathbf{H}_\mathrm{S}^{-1}$ is error-prone \cite{bk:trefethen-1997, bk:golub-2012}. More physically, $\mathrm{rank}(\mathbf{H}_\mathrm{S})$ can be roughly interpreted as the number of ``bonds'' between layers \cite{dwivedi-134304-2016} such that, for example, $\mathbf{H}_\mathrm{S}$ will be \mbox{(near-)singular} when the repeat unit is too big. The choice of layers in Figure \ref{fig:asymmetric} exemplifies such a case; as seen in Figure \ref{fig:asymmetric}(a), the $-\alpha$ orbitals in layer $n$ (white circles) do not directly couple to the $+\alpha$ orbitals in layer $n+1$ (red circles).
Such \mbox{(near-)singularity} of $\mathbf{H}_\mathrm{S}$ has an immediate physical interpretation that will be helpful below. To see this, note that the transfer matrix $\mathbf{T}$ tends to singularity as $\mathbf{H}_\mathrm{S}$ tends to singularity (\textit{vide infra}). Then, because an eigenvalue $\lambda$ of the transfer matrix is related to a (complex) wavevector by $\lambda=e^{ika}$, (near-)singularity of $\mathbf{T}$ (\textit{i.e.},\ $\lambda\simeq0$) implies the existence of Bloch states with $\mathrm{Im}(k)\gg0$. Point \ref{properties:1d:pairedevals} further showed that $1/\lambda^\ast$ will be an eigenvalue of $\mathbf{T}$, meaning that there is also a state with $\mathrm{Im}(k)\ll 0$. These highly evanescent states rapidly decay (grow) from one layer to the next and are not likely to be physically important because of their strong localization \cite{osbourn-2124-1979, ram-mohan-6151-1988, boykin-8107-1996, mavropoulos-1088-2000, khomyakov-195402-2004, sorensen-155301-2008, sorensen-205322-2009, vergniory-544-2010, reuter-085412-2011}. It is possible that they can be approximated or even neglected.
In the end, there are several good reasons why the layers may be chosen such that $\mathbf{H}_\mathrm{S}$ is \mbox{(near-)singular}. For example, such a layer spacing could be the smallest that (i) requires only nearest-neighbor coupling between layers and (ii) makes all layers identical. These conditions correspond to the block tridiagonal and block Toeplitz structures of $\mathbf{H}$, respectively. Fortunately, several strategies have been developed for addressing the \mbox{(near-)singularity} of $\mathbf{H}_\mathrm{S}$ when formulating CBS \cite{chang-3975-1982, biczo-51-1985, mailhiot-8360-1986, brand-607-1987, bowen-2754-1995, boykin-7670-1996, boykin-8107-1996, hjort-5245-2000, tomfohr-245105-2002, fagas-268-2004, khomyakov-195402-2004, velev-r637-2004, li-194113-2006, rungger-035407-2008, reuter-085412-2011, dwivedi-134304-2016}. We discuss four in turn.
\subsection{Choose Smaller Layers}
\label{sec:singular:smaller-layers}
If large layers cause $\mathbf{H}_\mathrm{S}$ to be (near-)singular (see our example in Figure \ref{fig:asymmetric}), an obvious solution is to choose smaller layers. The problem is that smaller layers could lead to deviations from the block tridiagonal and block Toeplitz structure of the Hamiltonian, which were essential to deriving CBS in sections \ref{sec:cbs-wf} and \ref{sec:cbs-gf}. We defer our discussion of going beyond the block tridiagonal structure to section \ref{sec:extended-coupling} and focus here on relaxing the block Toeplitz structure.
In such a case, the Hamiltonian would instead have a periodic block Toeplitz structure \cite{chang-3975-1982, mostoller-6168-1982, schulman-2346-1983, rozsa-35-1992}, where the blocks along each diagonal cycle through some finite number of distinct blocks. Such a system is essentially a superlattice. For example, there are two different small layers that alternate in Figure \ref{fig:asymmetric}; the Hamiltonian is structured as
\[
\mathbf{H} = \left[
\begin{array}{cc|cc|c}
\mathbf{H}_{\mathrm{D,1}} & \mathbf{H}_{\mathrm{S},1}^\dagger & \mathbf{0} & \mathbf{0} & \cdots \\
\mathbf{H}_{\mathrm{S},1} & \mathbf{H}_{\mathrm{D},2} & \mathbf{H}_{\mathrm{S},2}^\dagger & \mathbf{0} & \cdots \\ \hline
\mathbf{0} & \mathbf{H}_{\mathrm{S},2} & \mathbf{H}_{\mathrm{D},1} & \mathbf{H}_{\mathrm{S},1}^\dagger & \cdots \\
\mathbf{0} & \mathbf{0} & \mathbf{H}_{\mathrm{S},1} & \mathbf{H}_{\mathrm{D},2} & \cdots \\ \hline
\vdots & \vdots & \vdots & \vdots & \ddots \\
\end{array}
\right],
\]
where the lines show that the block Toeplitz structure is restored when two small layers are grouped together. We can address this structure by defining two transfer matrices (similarly for MMTs): One moves from a layer of type 1 to a layer of type 2 and \textit{vice versa}. Using the associativity of matrix multiplication (MMT action), we can produce an ``aggregate'' transfer matrix (MMT) for the periodic structure \cite{schulman-2346-1983, ghahramani-1102-1989, reuter-084707-2013}.
\subsection{Approximate the Highly Evanescent States}
\label{sec:singular:approximate}
We can approximate the very evanescent states, which lead to the \mbox{(near-)singularity}, by states that are less evanescent \cite{mailhiot-8360-1986, brand-607-1987, rungger-035407-2008, reuter-085412-2011}. As long as the new states still decay (grow) much more rapidly than the physically-important states, the error can be negligible (or at least controllable). In essence, we compute the singular value decomposition \cite{bk:strang-2006} of $\mathbf{H}_\mathrm{S}$ and artificially increase the small singular values. This results in an approximate coupling operator $\overline{\mathbf{H}}_\mathrm{S}$ that has a condition number determined by the amount of increase. $\overline{\mathbf{H}}_\mathrm{S}$ is thus invertible. The derivations in sections \ref{sec:cbs-wf} and \ref{sec:cbs-gf} are then readily applicable using $\overline{\mathbf{H}}_\mathrm{S}$ instead of $\mathbf{H}_\mathrm{S}$.
\subsection{Use a Generalized Eigenvalue Problem}
\label{sec:singular:geneig}
The invertibility of $\mathbf{H}_\mathrm{S}$ was not needed until the final steps of deriving the transfer matrix [between Eqs.\ \eqref{eq:schrodinger-recurrence} and \eqref{eq:supercell-recurrence}]. Instead, we can rearrange Eq.\ \eqref{eq:schrodinger-recurrence} to obtain
\[
\mathbf{H}_\mathrm{S}^\dagger \ket{\psi_{n+1}} = (E\mathbf{I}-\mathbf{H}_\mathrm{D})\ket{\psi_n} - \mathbf{H}_\mathrm{S}\ket{\psi_{n-1}},
\]
where we stop just before inverting $\mathbf{H}_\mathrm{S}^\dagger$. As in section \ref{sec:cbs-wf}, we combine this equation with the tautology $\ket{\psi_n}=\ket{\psi_n}$ to produce \cite{biczo-51-1985, boykin-7670-1996, boykin-8107-1996, fagas-268-2004, khomyakov-195402-2004, velev-r637-2004, li-194113-2006, rungger-035407-2008}
\[
\left[ \begin{array}{c} \mathbf{H}_\mathrm{S}^\dagger \ket{\psi_{n+1}} \\ \ket{\psi_n} \end{array} \right] = \left[ \begin{array}{cc} E\mathbf{I}-\mathbf{H}_\mathrm{D} & -\mathbf{H}_\mathrm{S} \\ \mathbf{I} & \mathbf{0} \end{array} \right] \left[ \begin{array}{c} \ket{\psi_n} \\ \ket{\psi_{n-1}} \end{array} \right].
\]
Converting to the ``supercell'' wavefunction,
\[
\left[ \begin{array}{cc} \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \\ \mathbf{0} & \mathbf{I} \end{array} \right] \ket{\Psi_{n+1}} = \left[ \begin{array}{cc} E\mathbf{I} - \mathbf{H}_\mathrm{D} & -\mathbf{H}_\mathrm{S} \\ \mathbf{I} & \mathbf{0} \end{array} \right] \ket{\Psi_{n}},
\]
and invoking Bloch's theorem [Eq.\ \eqref{eq:bloch-theorem}], we get
\begin{equation}
e^{ika} \left[ \begin{array}{cc} \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \\ \mathbf{0} & \mathbf{I} \end{array} \right] \ket{\Psi_{n}} = \left[ \begin{array}{cc} E\mathbf{I} - \mathbf{H}_\mathrm{D} & -\mathbf{H}_\mathrm{S} \\ \mathbf{I} & \mathbf{0} \end{array} \right] \ket{\Psi_{n}}.
\label{eq:wf-gen-eigval}
\end{equation}
We thus avoid inverting $\mathbf{H}_\mathrm{S}^{\dagger}$ at the expense of having to solve a generalized eigenvalue problem \cite{bk:golub-2012} for CBS.
Generalized eigenvalue problems are more sophisticated than standard eigenvalue problems in that they admit both finite and infinite eigenvalues. For the present discussion, the generalized eigenvalue problem has a zero or infinite eigenvalue when
\[
\left[ \begin{array}{cc} E\mathbf{I} - \mathbf{H}_\mathrm{D} & -\mathbf{H}_\mathrm{S} \\ \mathbf{I} & \mathbf{0} \end{array} \right] \text{ or } \left[ \begin{array}{cc} \mathbf{H}_\mathrm{S}^\dagger & \mathbf{0} \\ \mathbf{0} & \mathbf{I} \end{array} \right]
\]
is singular, respectively \cite{bk:golub-2012}. Both of these cases arise when $\mathbf{H}_\mathrm{S}$ is singular, giving our justification that $\mathbf{T}$ tends to singularity as $\mathbf{H}_\mathrm{S}$ becomes singular. Note that, when $\mathbf{H}_\mathrm{S}$ is near-singular (such that the transfer matrix is analytically well-defined), $\mathbf{T}$ will have very large (in magnitude) eigenvalues that correspond to the nearly infinite eigenvalues of Eq.\ \eqref{eq:wf-gen-eigval}. In practice, these eigenvalues may lead to numerical instabilities when applying CBS \cite{mailhiot-8360-1986, ko-9945-1988, ting-985-1999}, which partly motivates the next strategy.
\subsection{Exclude the Highly Evanescent States}
\label{sec:singular:exclude}
Finally, we can build on the previous idea and remove the 0 and infinite eigenvalues. In essence, we deflate \cite{bk:trefethen-1997} the generalized eigenvalue problem so that only the states of interest remain \cite{brand-607-1987, khomyakov-195402-2004, rungger-035407-2008, sorensen-155301-2008, sorensen-205322-2009, dwivedi-134304-2016} by projecting the highly evanescent states out of the problem. Unlike in section \ref{sec:singular:approximate}, these states are not approximated; rather, they are completely removed from consideration. After deflation, the resulting (and approximate) $\mathbf{H}_\mathrm{S}$ is invertible such that the tools in sections \ref{sec:cbs-wf} and \ref{sec:cbs-gf} are again applicable. Compared to the generalized eigenvalue problem in section \ref{sec:singular:geneig}, this procedure requires a bit of extra numerical effort to perform the deflation, but ultimately produces smaller matrices for the CBS computations.
\section{Extended Coupling Between Layers}
\label{sec:extended-coupling}
As discussed in the previous section, \mbox{(near-)singularity} of the inter-layer coupling operator $\mathbf{H}_\mathrm{S}$ may prompt us to choose smaller layers, which may, in turn, break the block tridiagonal structure of $\mathbf{H}$. This is but one reason why the block tridiagonal structure may be too restrictive, and in this section we discuss extended couplings. In a broad sense, a block banded matrix (which includes block tridiagonal and block pentadiagonal matrices, \textit{etc}.)\ remains block quasi-separable, meaning all of the ideas about generators---and CBS---are still applicable. As we now see, extended couplings do not fundamentally change anything; they only require more complicated algebra.
As an illustrative example, we first consider a system with nearest-neighbor and next-nearest-neighbor couplings before discussing the general case. Accordingly, the Hamiltonian becomes block pentadiagonal (and is still block Toeplitz),
\[
\mathbf{H} = \left[
\begin{array}{cccccc}
\mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \mathbf{H}_{\mathrm{S},2}^\dagger & \mathbf{0} & \mathbf{0} & \cdots \\
\mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \mathbf{H}_{\mathrm{S},2}^\dagger & \mathbf{0} & \cdots \\
\mathbf{H}_{\mathrm{S},2} & \mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \mathbf{H}_{\mathrm{S},2}^\dagger & \cdots \\
\mathbf{0} & \mathbf{H}_{\mathrm{S},2} & \mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \cdots \\
\mathbf{0} & \mathbf{0} & \mathbf{H}_{\mathrm{S},2} & \mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \cdots \\
\vdots & \vdots & \vdots & \vdots & \vdots & \ddots \\
\end{array}
\right].
\]
We will assume $\mathbf{H}_{\mathrm{S},2}$ is nonsingular, noting that all of the ideas discussed in section \ref{sec:singular} are generalizable if it is not.
Let us first focus on diagonalizing $\mathbf{H}$. Following section \ref{sec:cbs-wf}, we can derive a similar statement to Eq.\ \eqref{eq:schrodinger-recurrence},
\[
E\ket{\psi_n} = \mathbf{H}_{\mathrm{S},2} \ket{\psi_{n-2}} + \mathbf{H}_{\mathrm{S},1} \ket{\psi_{n-1}} + \mathbf{H}_\mathrm{D} \ket{\psi_n} + \mathbf{H}_{\mathrm{S},1}^\dagger \ket{\psi_{n+1}} + \mathbf{H}_{\mathrm{S},2}^\dagger \ket{\psi_{n+2}}.
\]
This equation, along with a tautology for each layer ($\ket{\psi_j}=\ket{\psi_j}$), leads to
\begin{equation}
\left[ \begin{array}{c}
\ket{\psi_{n+2}} \\
\ket{\psi_{n+1}} \\
\ket{\psi_n} \\
\ket{\psi_{n-1}} \\
\end{array} \right] = \left[ \begin{array}{cccc}
-\mathbf{H}_{\mathrm{S},2}^{-\dagger} \mathbf{H}_{\mathrm{S},1}^\dagger & -\mathbf{H}_{\mathrm{S},2}^{-\dagger} (E\mathbf{I} - \mathbf{H}_\mathrm{D}) & -\mathbf{H}_{\mathrm{S},2}^{-\dagger} \mathbf{H}_{\mathrm{S},1} & -\mathbf{H}_{\mathrm{S},2}^{-\dagger} \mathbf{H}_{\mathrm{S},2} \\
\mathbf{I} & \mathbf{0} & \mathbf{0} & \mathbf{0} \\
\mathbf{0} & \mathbf{I} & \mathbf{0} & \mathbf{0} \\
\mathbf{0} & \mathbf{0} & \mathbf{I} & \mathbf{0} \\
\end{array} \right] \left[ \begin{array}{c}
\ket{\psi_{n+1}} \\
\ket{\psi_n} \\
\ket{\psi_{n-1}} \\
\ket{\psi_{n-2}} \\
\end{array} \right].
\label{eq:pentadiagonal-recurrence}
\end{equation}
The matrix on the right-hand side is essentially the companion matrix of Ref.\ \onlinecite{chang-3975-1982}, which is a ``multi-layer'' generalization of the transfer matrix. We denote this companion matrix by $\mathbf{C}$ in what follows. We can also generalize the supercell wavefunction by making it larger,
\[
\ket{\Psi_{n+1}} \equiv \left[ \begin{array}{c}
\ket{\psi_{n+2}} \\
\ket{\psi_{n+1}} \\
\ket{\psi_n} \\
\ket{\psi_{n-1}} \\
\end{array} \right],
\]
such that Eq.\ \eqref{eq:pentadiagonal-recurrence} becomes
\[
\ket{\Psi_{n+1}} = \mathbf{C} \ket{\Psi_{n}}.
\]
Finally, Bloch's theorem again specifies that $\ket{\Psi_{n+1}}=e^{ika}\ket{\Psi_{n}}$, such that
\begin{equation}
e^{ika} \ket{\Psi_{n}} = \mathbf{C} \ket{\Psi_{n}}.
\label{eq:companion}
\end{equation}
Mirroring the transfer matrix for a Hamiltonian with nearest-neighbor couplings, the eigenvalues of $\mathbf{C}$ produce the CBS for a Hamiltonian with extended couplings.
Turning now to the inversion problem (\textit{i.e.},\ calculating the GF), there are fewer direct results. We know of only one study that has explicitly investigated the inverses of block pentadiagonal matrices \cite{koulaei-223-2007}, and it derived recurrence relations for the blocks of the inverse that are similar to Eq.\ \eqref{eq:gf-blocks}. That said, and as one might expect, these recurrence relations are more complicated. It is not immediately obvious how to access CBS from them when the block Toeplitz structure is also present. More work needs to be performed to derive generators for the inverses of block pentadiagonal and block Toeplitz matrices.
Generalizing to higher-order extended couplings, the Hamiltonian becomes block banded and block Toeplitz. For example, if a layer couples to its nearest $C$ neighbors on either side,
\[
\mathbf{H} = \left[
\begin{array}{cccccccc}
\mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \mathbf{H}_{\mathrm{S},2}^\dagger & \cdots & \mathbf{H}_{\mathrm{S},C}^\dagger & \mathbf{0} & \mathbf{0} & \cdots \\
\mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \cdots & \mathbf{H}_{\mathrm{S},C-1}^\dagger & \mathbf{H}_{\mathrm{S},C}^\dagger & \mathbf{0} & \cdots \\
\mathbf{H}_{\mathrm{S},2} & \mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \cdots & \mathbf{H}_{\mathrm{S},C-2}^\dagger & \mathbf{H}_{\mathrm{S},C-1}^\dagger & \mathbf{H}_{\mathrm{S},C}^\dagger & \cdots \\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots & \vdots \\
\mathbf{H}_{\mathrm{S},C} & \mathbf{H}_{\mathrm{S},C-1} & \mathbf{H}_{\mathrm{S},C-2} & \cdots & \mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \mathbf{H}_{\mathrm{S},2}^\dagger & \cdots \\
\mathbf{0} & \mathbf{H}_{\mathrm{S},C} & \mathbf{H}_{\mathrm{S},C-1} & \cdots & \mathbf{H}_{\mathrm{S},1} & \mathbf{H}_\mathrm{D} & \mathbf{H}_{\mathrm{S},1}^\dagger & \cdots \\
\mathbf{0} & \mathbf{0} & \mathbf{H}_{\mathrm{S},C} & \cdots & \mathbf{H}_{\mathrm{S},2} & \mathbf{H}_{\mathrm{S},2} & \mathbf{H}_\mathrm{D} & \cdots \\
\vdots & \vdots & \vdots & & \vdots & \vdots & \vdots & \ddots \\
\end{array}
\right].
\]
As above, we assume $\mathbf{H}_{\mathrm{S},C}$ is nonsingular, again referring the reader to the techniques developed in section \ref{sec:singular} if it is not. From this structure of $\mathbf{H}$, we can similarly derive the companion matrix
\[
\mathbf{C} = \left[ \begin{array}{ccccccc}
-\mathbf{H}_{\mathrm{S},C}^{-\dagger} \mathbf{H}_{\mathrm{S},C-1}^\dagger & -\mathbf{H}_{\mathrm{S},C}^{-\dagger} \mathbf{H}_{\mathrm{S},C-2}^\dagger & \cdots &\mathbf{H}_{\mathrm{S},C}^{-\dagger} (E\mathbf{I} - \mathbf{H}_\mathrm{D}) & \cdots & -\mathbf{H}_{\mathrm{S},C}^{-\dagger} \mathbf{H}_{\mathrm{S},C-1} & -\mathbf{H}_{\mathrm{S},C}^{-\dagger} \mathbf{H}_{\mathrm{S},C} \\
\mathbf{I} & \mathbf{0} & & & & \\
& \mathbf{I} & \ddots & & & \\
& & \ddots & \mathbf{0} & & \\
& & & \mathbf{I} & \ddots & \\
& & & & \ddots & \mathbf{0} \\
& & & & & \mathbf{I} & \mathbf{0} \\
\end{array} \right],
\]
where omitted blocks are $\mathbf{0}$ (after the first block row, only the block diagonal and block sub-diagonal are shown). Using an expanded supercell wavefunction,
\[
\ket{\Psi_{n+1}} \equiv \left[ \begin{array}{c}
\ket{\psi_{n+C}} \\
\ket{\psi_{n+C-1}} \\
\vdots \\
\ket{\psi_{n-C+1}} \\
\end{array} \right],
\]
we can again write $\ket{\Psi_{n+1}} = \mathbf{C} \ket{\Psi_{n}}$, which becomes Eq.\ \eqref{eq:companion} with the application of Bloch's theorem. Analogously, the eigenvalues of our generalized $\mathbf{C}$ still access the material's CBS. Regarding the GF, algorithms for inverting block banded matrices have been developed and discussed in the mathematical literature \cite{rozsa-447-1989, bk:vandebril-2008-v1}, but do not provide such a clear connection to CBS when the block Toeplitz structure is also imposed. We know of one work \cite{bini-431-1988} that specifically considers the inverses of block banded and block Toeplitz matrices. As in the block pentadiagonal case, future investigations into this problem are needed to determine a precise relation between CBS and the GF.
\section{Conclusions}
\label{sec:conclusions}
Complex band structure is an intrinsic material property for describing the physics of materials that have a prevalent repeat unit but lack perfect translational symmetry. Although it was first developed to answer fundamental questions regarding the implications of complex-valued wavevectors in Bloch's theorem, it subsequently found numerous applications throughout condensed matter physics, ranging from surfaces and interfaces to topological materials. The development of CBS, however, has generally been limited by, and sometimes in spite of, this broad applicability. Many studies independently redeveloped the main concepts and results of CBS, often with little comparison or connection to previous efforts. As a result, there are several seemingly disparate formulations of CBS in the literature.
Our primary objective in this work was to develop the interpretation that CBS is the minimal information that describes all of a material's static and dynamic electronic properties. This result is a mathematical implication of the block tridiagonal and block Toeplitz structure of a one-dimensional material's Hamiltonian (with generalizations discussed in sections \ref{sec:singular} and \ref{sec:extended-coupling}). In this way, each formulation of CBS taps into this intrinsic material information to address a specific physical problem.
Using this understanding of CBS and a consistent notation, we proceeded to unify many of CBS's reported properties and applications (sections \ref{sec:properties} and \ref{sec:interpretations}). Instead of viewing them as implications of CBS, we showed that they are intrinsic material properties and inherently described by CBS. In this way, our discussion showcased the applicability and consistency of the various CBS formulations. It also occurred at the level of operators, not matrices, to obviate unnecessary discussions of basis-set effects. We ultimately hope that this work provides an accessible introduction to CBS, outlines its utlity, and encourages its further development.
Analyzing CBS in terms of the Hamiltonian's information content exposes two potential future directions. (i) Higher-dimensional CBS, where $\vec{k}$ has complex components in more than one dimension, needs to be studied in more detail. In the absence of impurities (for example), each unit cell will have the same diagonal block ($\mathbf{H}_\mathrm{D}$) and will couple to its neighbors with a direction-dependent element (generalizing $\mathbf{H}_\mathrm{S}$). Collectively, this is still $\mathcal{O}(1)$ information; the key difference is that the material's topology is no longer linear. CBS is conceptually unchanged, and we need only build on fundamental discussions by Blount \cite{blount-305-1962} and Heine \cite{heine-300-1963} to better access CBS. Along these lines, some non-linear material topologies have recently been investigated, including Cayley trees \cite{jiang-057202-2012} and cylinders \cite{nguyen-275301-2016}. (ii) Adding impurities to the Hamiltonian, where the information becomes $\mathcal{O}(n)$ in the number of defects, would also be an interesting endeavor. CBS alone will no longer fully describe the defected material's electronic structure, but there should be insightful and relatively inexpensive techniques to incorporate the disorder. We recently reported \cite{reuter-014009-2012} some mathematical preliminaries on the topic.
\begin{acknowledgments}
I am grateful to Caroline Taylor, Thorsten Hansen, Jay Bardhan, Mark Ratner, and Jefferson Bates for helpful conversations. This research was supported by startup funds from the Institute for Advanced Computational Science at Stony Brook University. The figures were prepared using the SciDraw package \cite{caprio-107-2005}.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Soft X-ray transients (SXTs) are a subclass of low mass X-ray binaries (LMXBs) which alternate between quiescent periods lasting years and bright outbursts which can reach the Eddington limit and last typically for weeks \citep[see e.g.][for reviews of the observations]{ts96,csl97,yy15}. These bright sources were discovered soon after the first X-ray satellites came into operation. With the advent of more sensitive X-ray telescopes such as \textit{Chandra}, \textit{XMM-Newton} and \textit{Swift}, fainter and fainter sources were discovered \citep[see e.g.][]{hww04,mpb05,pgb05,swd05} during observations of galactic fields and particularly the galactic centre. \citet{wzr06} then introduced two new classes of transient sources, the so-called faint X-ray transients, and very faint X-ray transients (VFXTs) that have peak luminosities in the range 10$^{36-37}$ erg s$^{-1}$, and 10$^{34-36}$ erg s$^{-1}$ respectively. These VFXTs are part of the very faint X-ray binaries group (VFXBs) which are mostly transients with quiescent luminosities below 10$^{33}$ erg s$^{-1}$ \citep{hbd15}. Despite the fact that they were discovered more than ten years ago, their nature still remains elusive, probably in part because of their faintness, and also because most of them are located in highly absorbed regions making the identification of optical counterparts difficult. It is in particular not clear that the groups of faint and very faint transients are homogeneous, and searching for a model accounting for properties of all VFXBs might be inappropriate.
It is widely accepted that the transient nature of the classical soft X-ray transients (SXTs), which are much brighter in outburst than VFXBs and can reach luminosities of order of the Eddington limit, is due to the thermal/viscous instability of the accretion disc in regions where its effective temperature is of order of 6,000 -- 8,000 K \citep{kr98,dhl01}. This occurs when the mass-transfer rate from the secondary lies in some range $[\dot{M}_{\rm crit}^-(r_{\rm in}) - \dot{M}_{\rm crit}^+(r_{\rm out})]$ \citep{lasota-01}, where $r_{\rm in}$ and $r_{\rm out}$ are the inner and outer disc radius respectively, and $\dot{M}_{\rm crit}^-$ and $\dot{M}_{\rm crit}^+$ are the critical mass-transfer rates for the instability to occur at a given point in the disc ($\dot{M}_{\rm crit}^-$ is the maximum value for the disc to stay on the cold, stable branch, and $\dot{M}_{\rm crit}^+$ is the minimum value for the disc to stay on the hot branch). The observed luminosities of persistent and transient X-ray binaries show that this in indeed the case \citep{vp96,cfd12} if the accretion disc is X-ray self-irradiated.
It is tempting to assume that faint and very faint outbursts could be due to the same instability that is responsible for the bright outbursts of SXTs. Such a model has to explain the faintness of the outbursts and take into account the low average mass accretion rate, which for VFXBs lies in the range $3 \times 10^{-13}$ -- $1 \times 10^{-10}$ M$_\odot$ yr$^{-1}$ \citep{dw09}. This low mass-transfer rate is directly related to the nature of the secondary and its evolutionary status, and, to a lesser extent, to the primary mass.
\citet{k00} suggested that faint transients contain low mass secondaries that have evolved past the minimum orbital period. \citet{kw06} proposed later that VFXBs were either formed with a brown dwarf or a planetary companion, or could be the end products of massive primordial binaries which lead to systems containing an intermediate mass black hole, of the order of 1000 M$_\odot$. The latter explanation would of course not hold for VFXBs that exhibit type I X-ray bursts, and must therefore contain a neutron star. As noted by \citet{hbd15}, many of VFXBs have exhibited type I X-ray bursts, for example XMM J174457-2850.3 \citep{dwr14} and therefore do not harbor a black hole. \citet{mp13} proposed that the weak mass-transfer rates in VFXBs can be explained if these systems are detached, being in a similar state as cataclysmic variables in the period gap. This occurs when the secondary star becomes fully convective as a result of mass loss; angular momentum losses are reduced and the slightly bloated secondary shrinks, returning to thermal equilibrium; mass transfer can only occur via the faint secondary wind. The mass-transfer rate is difficult to estimate, but \citet{mp13} quote values of order of $10^{-16} - 10^{-14}$ M$_\odot$ yr$^{-1}$, which are much lower than the values of the order of $10^{-13}$ M$_\odot$yr$^{-1}$
inferred from observations of VFXTs \citep{dw10}.
If the transient nature of the faint X-ray sources and VFXBs were to be attributed to the thermal/viscous disc instability, the observational difference between bright transients and VFXBs would have to be due either to a small accretion disc for the latter, (the critical mass-transfer rates $\dot{M}_{\rm crit}^+$ and $\dot{M}_{\rm crit^-}$ and the outburst peak luminosity scale as $r^{2.65}$), and/or to the fact that only a small fraction of the disc is involved in the instability. The smallest accretion discs are found in ultracompact binaries. These systems form a subclass of low-mass X-ray binaries (LMXBs) with very short orbital periods (typically less than $\sim$ 1hr). Unless the system has followed a very special evolutionary track, the secondary star is hydrogen poor in order to fit inside its Roche lobe, being either a degenerate helium star or the core of a carbon-oxygen or oxygen-neon-magnesium white dwarf \citep[see e.g.][]{N08}. Compared to hydrogen-rich binaries, this change in chemical composition significantly increases the temperature for which the disc becomes unstable, and hence the critical mass-transfer rates, but the global picture remains unchanged \citep{ldk08}.
In this paper, we investigate in detail the nature and observational consequences of the disc instability model (DIM) applied to ultracompact binaries, in particular taking into account the effects of self-irradiation by accretion-produced X-rays. \cite{ldk08} estimated the stability conditions for hydrogen deficient ultracompact binaries, taking into account these effects, but have not produced light curves that can be directly compared with observations. This is needed in particular when one is interested in the faint end of the outburst luminosity distribution, for which the outermost parts of the disc may remain in the cold state during an outburst, and the peak luminosity is far from reaching the maximum luminosity one can derive using simple analytical estimates. \cite{kld12} did calculate the light curves predicted by the DIM in the case of degenerate helium secondaries, but, as they were interested in AM CVn stars, they did not consider disc irradiation. We show here that the observational properties (peak luminosities, duration, recurrence time) of many of the VFXB transients can be accounted for by the DIM provided, however, that the viscosity in quiescence is slightly smaller (typically a factor of between two and four) than in hydrogen--rich binaries. This does not imply that all VFXBs are ultracompact binaries -- some, such as for example \object{AX J1745.6-2901} which has an orbital period of 8.4 hr, are obviously not. Other systems, such as \object{1RXH J173523.7-354013} for which an H$\alpha$ line has been detected \citep[][note that this particular object is a persistent source, with an X-ray luminosity of order of $2 \times 10^{35}$ erg s$^{-1}$]{djt10} may contain a hydrogen rich secondary and are therefore not ultracompact. It does, however, mean that the short and faint outbursts observed in a number of VFXTs can be explained by the DIM when applied to ultracompact binaries.
We leave for future work the much needed application of the DIM to the case where the donor star is a carbon-oxygen degenerate star, and in particular the confirmation that the C/O discs would behave in a manner similar to hydrogen rich discs.
\section{DIM for irradiated helium accretion discs}
\subsection{Vertical structure}
The effect of X-ray irradiation on the local disc vertical structure at a given radius $r$ is merely to change the surface boundary condition:
\begin{equation}
T_{\rm surf}^4 = T_{\rm eff}^4 + T_{\rm irr}^4,
\end{equation}
where $T_{\rm eff}$ is the effective temperature corresponding to the heat flux generated by viscous dissipation is the disc, $T_{\rm surf}$ is the temperature at the disc surface, defined as the place where the optical depth is 2/3, and $T_{\rm irr}$ is the local irradiation temperature at point $r$ given by
\begin{equation}
\sigma T_{\rm irr}^4 = C \frac{L_{\rm X}}{4 \pi r^2},
\end{equation}
with
\begin{equation}
L_{\rm X} = \epsilon \dot{M}_{\rm in} c^2.
\end{equation}
$\dot{M}_{\rm in}$ is the mass accretion rate at the inner disc radius, $\epsilon$ is the accretion efficiency, typically of order of 0.1. The coefficient $C$ measures the fraction of the X-ray luminosity which is intercepted by the disc; it could be in principle calculated in a self consistent way, but this turns out to be a formidable task. The coefficient $C$ also incorporates the fact that only a fraction of the X-rays is absorbed below the photosphere and is hence able to affect the thermal structure of the disc; X-rays absorbed in the upper photosphere would rather be responsible for the formation of a hot corona whose interaction with the underlying cooler layers is complex. Based on observations, \cite{dhl01} showed that with $C = 5 \times 10^{-3}$, the model reproduces reasonably well the light curves of SXTs. It has been shown, however, that models are rather insensitive to the value of C, as long as $5\times 10^{-3} \lesssim C \lesssim 2\times 10^{-2}$ \citep[see][but we note that in this paper $C$ contains accretion efficiency $\epsilon$]{ldk08}, so for convenience we will assume in this paper that $C=0.01$.
The disc vertical structure equations have been solved on a grid of parameters : surface density $\Sigma$, central disc temperature $T_{\rm c}$, radius $r$ and irradiation temperatures $T_{\rm irr}$, as described in \cite{ldk08}. Figure \ref{scurve} (dashed lines) shows the classical helium-disc thermal equilibrium S-curves. They have the same behaviour as in the hydrogen case, except that the effective temperatures of the turning points are higher, because the ionization potential of helium is larger than that of hydrogen.
As for cataclysmic variables and low-mass X-ray binaries, we assume that the Shakura-Sunyaev parameter $\alpha$ is a function of the disc central temperature, so that it is constant and equal to $\alpha_{\rm h}$ on the hot stable branch, and to $\alpha_{\rm c}$ on the cool lower branch; to smoothly connect the two branches, we take:
\begin{equation}
\log(\alpha) = \log(\alpha_{\rm c}) + \left[ \log(\alpha_{\rm h}) - \log(\alpha_{\rm c})\right] \times \left[ 1 + \left(\frac{4 \; 10^4 \; \rm K}{T_{\rm c}} \right)^{10}\right]^{-1}
\label{alpha}
\end{equation}
where $T_{\rm c}$ is the disc mid-plane temperature. The choice of this function is rather arbitrary, and is found to have little influence on our results, provided that the two stable branches including the turning points are unaffected; it only impacts the detailed structure of the transition fronts. This relation is similar to the one used in \cite{hmdlh98}, except that we now have a transition temperature of 40,000K instead of 25,000 K, reflecting the higher temperatures found for the turning points. Note also that the exponent for the temperature dependence is slightly different from the one in \citet{hmdlh98} (10 instead of 8). This change was dictated by the necessity to have a sharp enough transition between the hot and cold branches, to avoid affecting the stable parts of these branches, and an index of 8 would have resulted in a cold-branch maximum $\Sigma$ value on the the S curve obtained using Eq. (\ref{alpha}) slightly smaller than the value obtained with a constant $\alpha$ equal to $\alpha_{\rm c}$. Figure \ref{scurve} shows that with our choice for the transition temperature, the resulting S curve does indeed correspond to the case $\alpha = \alpha_{\rm c}$ on the cool branch, and $\alpha = \alpha_{\rm h}$ on the hot branch. It can also be seen that such a choice does not artificially generates unstable branches when $T_{\rm irr}$ is large enough to stabilize the disc. Finally note that whereas the effective temperature of both turning points is weakly dependent on radius, the disc mid-plane temperature is practically constant \citep[see the analytical fits in][]{ldk08}, explaining why the temperature scaling in Eq. (\ref{alpha}) can be taken to be independent of $r$. In the DIM, an \textit{ad hoc} jump in $\alpha$'s value is assumed in order to reproduce disc outbursts properties, their amplitude in particular. It has been found recently by \citet{coleman16} that intermittent convection appearing near the ionization temperature increases the viscosity parameter by the amount required by the DIM, thus providing a physical mechanism for the ansatz used.
\begin{figure}
\centering
\includegraphics[angle=-90,width=\columnwidth]{scurve}
\caption{Two examples of local S curves -- effective temperature vs. surface density -- in a He disc when irradiation is included. The upper blue set of curves shows the case $T_{\rm irr} = 25,000$ K and $r=5 \times 10^{8}$ cm, the lower red set of curves corresponds to $T_{\rm irr} = 5,000$ K and $r= 10^{10}$ cm. In both cases, the solid curve corresponds to $\alpha$ given by Eq. (\ref{alpha}), while the dotted curves are for $\alpha = 0.05$ and $\alpha = 0.2$ (rightmost and leftmost curves respectively). The primary mass is 1.4 M$_\odot$. }
\label{scurve}%
\end{figure}
\subsection{Time evolution}
The maximum outburst luminosity is obtained assuming that during the outburst, the disc is close to a steady state with a mass accretion rate equal to the critical value $\dot{M}_{\rm crit}^+$ for the disc to stay on the hot stable branch \citep{ldk08}:
\begin{equation}
L_{\rm max} \simeq 3.5 \times 10^{37} \left( \frac{P_{\rm orb}}{1 \; \rm h} \right)^{1.67} \; \rm erg \; s^{-1},
\label{lmax}
\end{equation}
in the case of pure helium discs, assuming a 1.4 M$_\odot$ primary; $L_{\rm max}$ is almost independent of $\alpha_{\rm h}$. The upper limit is reached when the heating front is able to propagate up to the outer edge of the disc; for small values of the mass-transfer rate, this need not be the case, as the outer part of the disc could remain on the cold, stable branch even during outbursts. The outer parts of unirradiated discs may remain cold when the mass-transfer rate is less than $\dot{M}_{\rm crit}^-$. However, irradiation enables the heating front to propagate further away during outbursts; using Eq. (\ref{lmax}), the peak irradiation temperature at the outburst maximum can reach:
\begin{equation}
T_{\rm irr, max} = 37,000 \; \left( \frac{P_{\rm orb}}{1 \; \rm h} \right)^{1/3} \left( \frac{C}{10^{-2}} \right)^{1/4} \; K,
\end{equation}
where we have assumed that the disc radius is 0.6 times the orbital separation \citep{p77}, and the primary mass is that given above. The maximum irradiation temperature at the outer edge of the disc is comparable, albeit slightly smaller than the critical temperature of 40,000 K at which the transition between $\alpha_{\rm c}$ and $\alpha_{\rm h}$ occurs. This significantly softens the condition $\dot{M}_{\rm tr} > \dot{M}_{\rm crit}^-$ as a condition for the heating front to reach the outer disc edge.
One therefore expects that systems with orbital periods close to one hour and relatively large mass-transfer rates will undergo bright outbursts approaching the Eddington limit, and thus not very different from those found in standard SXTs; systems with shorter orbital periods and/or low mass-transfer rates should exhibit only faint outbursts.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{lc_xte1b}
\caption{X-ray luminosity of an ultracompact transient system with a He disc, for different values of the viscosity parameters $\alpha_{\rm c}$ and $\alpha_{\rm h}$. Top panel: $\alpha_{\rm c} = 0.02$ and $\alpha_{\rm h}=0.2$; the blue curve shows the case where the inner disc is truncated at a variable radius with time, according to Eq. (\ref{eq:rtrunc}), and the red curve is for a fixed inner radius ($2 \times 10^8$ cm). The mass-transfer rate is $3 \times 10^{15}$ g s$^{-1}$. Intermediate panel: changing $\alpha_{\rm c}$. Same as above for a truncated disc, with $\alpha_{\rm c} = 0.04$ (red curve) and $\alpha_{\rm c}=0.01$ (blue curve); $\alpha_{\rm h}=0.2$ is unchanged. In the case of the blue curve, the mass-transfer rate has been reduced by a factor 3. Bottom panel: changing $\alpha_{\rm h}$. Same as top panel, with $\alpha_{\rm h}=0.1$. $\alpha_{\rm c} = 0.02$ is unchanged}
\label{lc-He}%
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{lc_xte_zoom}
\caption{Detailed outburst profile for system shown in the upper panel of Fig. \ref{lc-He}, with truncation of the inner disc at a variable radius. The X-ray luminosity is shown in the lower panel, and the upper panel shows the inner disc radius (solid black curve) and the position of the transition front (thin solid blue line).}
\label{lc-He-outburst}%
\end{figure}
In order to follow the time evolution of these systems, we use the numerical scheme of \cite{hmdlh98}, modified to include heating by irradiation as described in \citet{dhl01}. We have assumed that either the inner disc radius is set to a fixed, small value (ideally, this should be the radius of the neutron star or that of the innermost stable orbit for a black hole or a very compact neutron star; this is numerically not possible, and we instead typically take $r_{\rm in} = 2 \times 10^8$ cm), or assume, as in the case of standard bright X-ray transients, that it is truncated at a value which depends on the accretion rate at the inner edge. In this case, we set, unless otherwise noted:
\begin{equation}
r_{\rm in} = 2.0 \times 10^8 \left( \frac{\dot{M}_{\rm acc}}{10^{16} \; \rm g s^{-1}} \right)^{-2/7} \; \rm cm,
\label{eq:rtrunc}
\end{equation}
which is formally identical to the case of truncation by a magnetic dipolar field, with a magnetic dipole moment of $2 \times 10^{29}$ Gcm$^3$; one should note however that the magnetic moment of a neutron star is not large enough to disrupt the accretion flow at those large distances; truncation would have to be caused by another mechanism such as evaporation of the inner disc. Whatever the reason, the detailed functional dependence of $r_{\rm in}$ versus $\dot{M}_{\rm acc}$ is of little importance \citep{mhln00}: what matters is the value of $r_{\rm in}$ during quiescence.
The outer disc radius is determined by the tidal torques, which can be written as
\begin{equation}
T_{\rm tid} = c_{\rm tid} \Omega_{\rm orb} \nu \Sigma f(r/a),
\end{equation}
where $c_{\rm tid}$ is a constant, $\Omega_{\rm orb}$ the orbital frequency, $\nu$ the viscosity, $\Sigma$ the surface column density, $a$ the orbital separation and $f$ a function describing the radial dependence of the torque. Several prescriptions can be used for $f$; based on \citet{pp77}, \citet{hmdlh98} used $f(r/a)=(r/a)^5$. This prescription has the disadvantage of allowing the outer disc radius to increase to large values, possibly beyond the Roche lobe.
One could also assume that the tidal torques are negligibly small as long as the disc stays within the tidal truncation radius $r_{\rm tid}$, so that $T_{\rm tid} (r) =0$ for $r < r_{tid}$, and become arbitrarily large for $r>r_{\rm tid}$, thereby ensuring that the radius does not exceed $r_{\rm tid}$. This approach was chosen by \citet{vh08}. For practical numerical reasons, we have chosen instead:
\begin{equation}
f(r/a)=e^{K(r-r_{\rm tid})/a}
\end{equation}
where $K$ is a constant. This formulation is identical to \citet{vh08} when $K$ is large. Here, we take $K=50$ which is numerically convenient and ensures that the outer disc radius does not exceed $r_{\rm tid}$ by more than a couple of percent. We did check that our numerical results do not depend significantly on the assumed value for $K$, provided that $K$ is not small. In this way, the accretion disc does not extend significantly beyond the tidal truncation radius, but is allowed to be smaller. This prescription is therefore very different from assuming that the outer disc radius is fixed, which is unphysical and leads to light curves which are quite different from those with variable outer disc radii \citep{hmdlh98}.
\subsubsection{Bright outbursts}
Figure \ref{lc-He} (top panel) shows the time evolution of a system with orbital period 0.7 hr, primary mass 1.4 M$_\odot$, tidal truncation radius 1.8 $\times$ 10$^{10}$ cm, with and without truncation of the inner disc. The viscosity parameter $\alpha$ is taken to be 0.02 in the cold state, and 0.2 in the hot state. The mass-transfer rate is 3 $\times$ 10$^{15}$ g s$^{-1}$, and we have used $C=0.01$. As can be seen, truncation does not significantly affect the outburst properties of the system, in stringent contrast with standard SXTs with longer orbital periods. The recurrence time is slightly reduced when $r_{\rm in}$ is kept fixed to a small value, by about 25\%, the peak luminosity is also reduced (by about 20\%), and the outburst duration is unchanged. The most noticeable difference is the quiescence luminosity, which is approximately constant at $L_{\rm X} \sim 10^{33}$ erg s$^{-1}$ in the case of disc truncation and may be as low as $10^{31}$ erg s$^{-1}$ otherwise, with significant variations between outbursts. This difference is simply due to the fact that in quiescence, the mass accretion rate onto the compact object must be smaller than $\dot{M}_{\rm crit}^-(r_{\rm in})$, which rapidly increases with the truncation radius.
These orbital parameters as well as the mass-transfer rate are appropriate for the case of for example \object{XTE J1751-305}, which is a transient accreting millisecond pulsar whose 42 minutes orbital period has been determined from the Doppler variation of the spin period of the neutron star \citep{mss02}. We find that the outburst properties are not very different from the observed ones \citep[peak luminosity of order of $2 \times 10^{37}$ erg s$^{-1}$ for a distance of 2 kpc, outburst duration $\sim$ 10--15 days;][]{mss02}. This source underwent four outbursts over a timespan of 12 years \citep{hjw09}; the recurrence time produced by our model is of order of 1 yr, too short by a factor of three. Longer recurrence times can be obtained with smaller values of $\alpha_{\rm c}$; we found that with $\alpha_{\rm c} = 0.01$, one obtains a recurrence time of order of 3 years; the outburst peak luminosity is also increased by a factor of three, its duration being almost unaffected. Reducing the mass-transfer rate by a factor od three results in the light curve shown in the intermediate panel of Fig.\ref{lc-He} (blue curve). One should note that the ratio $\alpha_{\rm h}/\alpha_{\rm c}$ is at least of the order of ten, and larger by a factor of between roughly two and four. than the one used in models of dwarf novae or longer period transient LMXBs. One need to have a relatively low value of $\alpha_{\rm c}$ in order to account for the observed recurrence times, of the order of one year or more in ultracompact binaries, whereas $\alpha_{\rm h}$ cannot be too small in order to obtain outbursts whose duration does not exceed by far the observed values.
The intermediate and lower panel of Fig. \ref{lc-He} illustrate the effects of changing the viscosity. As mentioned above, the main impact of changing in $\alpha_{\rm c}$ is on the recurrence time: increasing (decreasing) $\alpha_{\rm c}$ by a factor of two results in a decrease (increase) of the recurrence time by a factor of three, larger than the relative change in $\alpha_{\rm c}$ because the viscosity is proportional to $\alpha c_{\rm s} H$, $c_{\rm s}$ and $H$ being the mid-plane sound speed and the vertical scale height respectively, which both decrease when $\alpha$ (and hence the central temperature) is reduced. Figure \ref{lc-He} also shows that modifying $\alpha_{\rm h}$ changes the duration of the outburst \citep[see e.g.,][]{KL12}, with relatively little influence on the recurrence time. The duration of the large outbursts shown in the bottom panel of Fig. \ref{lc-He} is of order of 60 days, much longer than the observed duration of XTE J1751-305 outbursts. It should finally be noted that when the ratio $\alpha_{\rm h}/\alpha_{\rm c}$ becomes too large, a sequence of alternating long and short outbursts appear. In the case $\alpha_{\rm c} = 0.04$, $\alpha_{\rm h} = 0.2$ corresponding to the case of classical dwarf novae, a sequence of bright and faint outbursts is observed, with a short recurrence time (100 days, 200 days between bright outbursts), as expected.
Figure \ref{lc-He-outburst} shows the outburst profile in the case of disc truncation, as well as the position of the transition front. It should be stressed out that the heating and cooling fronts are not as steep as in the standard (i.e. low mass primary as in CVs, and weak or no disc irradiation) model; such broad transition fronts are also found in the standard SXT disc instability model \citep[see e.g. Fig. 8 in][]{dhl01}. When the first cooling front forms, it is quite narrow, but cannot propagate down to the inner disc edge, and is reflected into a heating front that propagates outwards. Several such reflections occur during which the width of the cooling front gradually increases. This happens because irradiation has a strong impact on the disc temperature $T_{\rm c}$, since $T_{\rm c}$ has to be larger than $T_{\rm irr}$ and can be very close to this value. But $T_{\rm irr}$ is a non local quantity, given by the accretion rate at the inner disc edge, and the radial temperature profile is largely determined by $T_{\rm irr}$. This is generic to strongly irradiated disc. On the other hand, for values of $C$ smaller than the ones considered in this paper, typically $C \leq 5 \times 10^{-4}$, this would not happen and the cooling front would be narrow. Here we define the transition radius as the point where $\alpha = (\alpha_{\rm c} + \alpha_{\rm h}) /2$.
One can note oscillations in the position of the transition front, as is also the case in long period SXTs \citep{dhl01}. These oscillations are here barely noticeable in the light curve so practically unobservable as such, but see below. The heating front propagates throughout the entire disc, which remains fully in the hot state for a few days before a cooling front can start propagating from the outer edge. This cooling front is unable to propagate down to the innermost radius, but is instead reflected into a heating front; the succession of heating and cooling fronts makes the decay much longer than if a cooling front were propagating freely. Interestingly, the decay is very roughly exponential, even during the phases where the disc is not kept in a hot state by irradiation. This shows that outbursts with exponential decays need not be produced according to \cite{kr98} model, as is sometimes incorrectly implied in the literature: \citet{kr98} showed that exponential decays are obtained during phases where the whole disc remains illuminated, on the hot branch and the cooling front propagation is quenched, while linear decays occur when the cooling front propagates freely. We have shown, however, that in many cases, when the cooling front propagation is controlled by irradiation, the decay from outburst is still close to exponential. Finally, it is worth noting that a significant fraction (about 30\%) of the disc mass is accreted during an outburst.
\subsubsection{Very faint outbursts}
Faint outbursts are expected when either the accretion disc is very small or when the mass-transfer rate is so low that the entire disc cannot be brought into a hot state. Figure \ref{lc_vfxb} shows such a situation; the orbital parameters are those of Fig. \ref{lc-He}, but the mass-transfer rate is reduced to 10$^{14}$ gs$^{-1}$, with a variable or fixed small inner radius. A zoom of the outbursts is shown in Fig. \ref{lc_vfxb_zoom}.
The recurrence time between outbursts is 1.0 yr in the case of disc truncation, almost unchanged from the higher mass-transfer rate case (320 days), but the peak luminosity is reduced by more than one order of magnitude, and the duration is also shorter -- about 10 days. When the disc inner radius is set to $2 \times 10^8$ cm, the recurrence time is 180 days, slightly shorter than the value found for higher $\dot{M}$ (240 days). In both cases, a small fraction of the disc is accreted during an outburst (2.8\% in case of disc truncation, 2\% otherwise, to be compared to $\sim$30\% for the high $\dot{M}$ case), which is numerically challenging as one must follow more than 100 outbursts for the system to relax from the initial conditions. Interestingly, the outburst profile does not deviate much from an exponential; this is again due to the fact that the cooling front cannot propagate freely down to the inner disc radius, but instead a series of reflections between heating/cooling fronts occurs. The impact (in terms of rebrightenings) of these reflections on the light curve shape is, however, negligible from the observational point view. The heating front never propagates beyond about 10$^{10}$ cm in both cases, and the outer disc remains in a cold, stable state. As for Fig. \ref{lc-He} and for the same reason, the quiescent level is much larger when the inner disc is truncated than when the disc inner radius is kept fixed at a small value. Note also that for the truncated disc, the mass accretion rate onto the neutron star is not much smaller than the mass-transfer rate from the secondary; the disc is therefore not very far from being stable on the cold branch, and 40\% of the mass transferred between outbursts is accreted onto the neutron star during quiescence. This helps in having a recurrence time significantly longer than in the case where the inner disc radius is set to $2 \times 10^8$ cm. This would support the idea that in these systems, as in bright transients \citep{dhl01,bz16}, accretion discs are truncated during quiescence, which could in principle be tested with high quality broad-band X-ray observations of these systems. While we do favour this hypothesis, one should keep in mind that other alternatives exist, because the viscosity parameter is not as constrained as in H-rich discs (see below).
\begin{figure}
\centering
\includegraphics[angle=-90,width=\columnwidth]{lc_vfxb}
\caption{Light curves for low mass-transfer rates. For the black and red curves, $\dot{M}_{\rm tr} = 10^{14}$ gs$^{-1}$, the orbital parameters and the viscosity being identical to those of Fig. \ref{lc-He}. The black line shows the case of a truncated disc with an inner disc radius given by Eq. (\ref{eq:rtrunc}). The red curve corresponds to a fixed inner radius of 2 $\times$ 10$^8$ cm. The green and blue curves correspond to $\dot{M}_{\rm tr} = 10^{13}$ gs$^{-1}$, the inner radius being equal to 0.15 times the value given by Eq. \ref{eq:rtrunc}. The viscosity parameters are $\alpha_{\rm h}=0.2$. $\alpha_{\rm c} = 0.01$ (green curve) and $\alpha_{\rm h}=0.1$. $\alpha_{\rm c} = 0.01$ (blue curve).}
\label{lc_vfxb}%
\end{figure}
\begin{figure}
\centering
\includegraphics[angle=-90,width=\columnwidth]{lc_vfxb_zoom}
\caption{Zoom of the outbursts shown in Fig. \ref{lc_vfxb}. For a better legibility, the origin of the time axis is not the same as in Fig. \ref{lc_vfxb}.}
\label{lc_vfxb_zoom}%
\end{figure}
Longer recurrence time can be obtained for lower values of $\alpha_{\rm c}$. Figure \ref{lc_vfxb} shows the case where $\alpha_{\rm c}$ is reduced to 0.01, for $\alpha_{\rm h}$ equal to 0.1 or 0.2. In both cases, the mass-transfer rate is $10^{13}$ g s$^{-1}$. The recurrence time is 2.8 yr for $\alpha_{\rm h}=0.2$ and 1.8 yr when $\alpha_{\rm h}=0.1$. The front propagates to a maximum distance of $4.8 \times 10^9$ cm in the first case, and to $4.1 \times 10^9$ cm in the second, and the outbursts are therefore extremely faint, with peak luminosities of $3.2 \times 10^{35}$ and $1.4 \times 10^{35}$ erg s$^{-1}$ respectively, and have short durations.
It is a general characteristic of the model that the faintest outbursts are also the shortest ones, simply because these correspond to the smallest extension of the hot, inner part of the disc. Longer outbursts could in principle be obtained by reducing the viscosity in the hot state $\alpha_{\rm h}$. Whereas there are good reasons to expect that $\alpha_{\rm c}$ should be different for discs in ultracompact binaries and in longer period systems, there is no obvious reason for such a change in $\alpha_{\rm h}$. The DIM would therefore have difficulties in explaining long and very faint outbursts. In particular those from \object{AX J1745.6-2901} lasting more than 80 weeks, or from \object{CXOGC J174538.0-290022} lasting more than 30 weeks \citep{dw10}. These long outbursts could still be explained by the DIM, but an additional ingredient, such as an increase of the mass-transfer rate due to the illumination of the secondary star is needed. This effect has been invoked to account for the long duration of superoutbursts in SU UMa stars \citep{hlw00,smak08}, even though its magnitude is unclear \citep[see e.g.][]{vh07,c15}.
\begin{figure}
\centering
\includegraphics[angle=-90,width=\columnwidth]{lc_vvfb}
\caption{Outbursts profiles for low mass-transfer rates ( Here, $\dot{M}_{\rm tr} = 5 \times 10^{14}$ gs$^{-1}$), and for $\alpha_{\rm c} = 0.04$ and $\alpha_{\rm h}=0.2$. The disc is truncated with an inner disc radius given by Eq. (\ref{eq:rtrunc}).}
\label{lc_vvfb}%
\end{figure}
We also considered the case where $\alpha_{\rm c}$ = 0.04, keeping $\alpha_{\rm h}$ unchanged to 0.2, which are standard values for cataclysmic variables and most probably for (helium) AM CVn stars \citep{KL12}. We have assumed here a tidal truncation radius of $1.0 \times 10^{10}$ cm, a mass-transfer rate of $5 \times 10^{14}$ g s$^{-1}$, and $C$ is kept equal to 0.01. The resulting profile is shown in Fig. \ref{lc_vvfb}; one can see sequence of very faint (10$^{35}$ erg s$^{-1}$) and faint (10$^{36}$ erg s$^{-1}$) outbursts. As expected, the recurrence time is reduced as compared to the case with $\alpha_{\rm c} = 0.02$, and is of the order of 100 days between the main outbursts. The duration of faint outbursts is of the order of 10 days, that of the faintest ones is 4--5 days only; the entire disc is brought into the hot state during large outbursts, whereas the heating front can reach at most 40\% of the disc radius during the faintest ones.
Interestingly, the peak luminosity of the faintest outbursts is close to that observed during the intermediate state of very faint X-ray transients, such as for example \object{XMM J174457-2850.3} \citep{dwr14}. During these states, the X-ray luminosity varies between typically $10^{33} - 10^{34}$ erg s$^{-1}$; they have been tentatively explained as resulting from the interaction of the accretion flow and the magnetic field of the neutron star. The extremely faint outbursts expected from the DIM would be very difficult to identify, and they would contribute to the variability observed during the intermediate state.
\section{The case of H discs}
\begin{figure}
\centering
\includegraphics[angle=-90,width=\columnwidth]{lc_hdisk}
\caption{X-ray light curve in the case of a irradiated hydrogen disc. The orbital parameters and the viscosity are the same as in Fig. \ref{lc-He} and \ref{lc_vfxb}; the mass-transfer rate is $\dot{M}_{\rm tr} = 10^{14}$ gs$^{-1}$, again as in Fig. \ref{lc_vfxb} }
\label{lc_h}%
\end{figure}
Hydrogen rich companions are not expected in ultracompact binaries, as systems with H rich secondaries should have a minimum orbital period around 80 minutes. It is, however, interesting to consider the case of H--rich secondaries, as the comparison with the He--rich case helps understanding the specificities of ultracompact binaries. They are also expected to behave in a similar way to systems containing a carbon-oxygen white-dwarf donor. Such systems should not be infrequent \citep[see e.g.][]{NJ10}, but modelling them is hampered by the unavailability of opacities for CO mixtures at all temperatures. \cite{mph01} have shown that accretion discs made of carbon and oxygen are subject to the same instability as hydrogen and helium discs, but have constructed thermal equilibrium curves only for pure carbon or pure oxygen above a few thousand degrees. Examining their Fig. 2 shows that the thermal instability for carbon and oxygen discs is triggered at about the same effective temperature as for hydrogen rich discs, which is not unexpected since the ionisation potential of those elements are similar, and much lower than that of helium.
Figure \ref{lc_h} shows the case of a hydrogen disc with the same parameters as for the black curve in Fig. \ref{lc_vfxb}, i.e. $M_1 = 1.4$ M$_\odot$, $M_2=0.035$ M$_\odot$, $P_{\rm orb}=0.7$ hr, and with a variable inner radius given by Eq. (\ref{eq:rtrunc}). We use $\alpha_{\rm c}= 0.02$ and $\alpha_{\rm h}= 0.2$, $C=0.01$, and $\dot{M}_{\rm tr} = 10^{14}$ gs$^{-1}$, also unchanged. Although it is impossible for a hydrogen rich secondary star to fit inside its Roche lobe for these orbital parameters, we chose not to change them to compare the effect of changing the disc composition only, and also because this particular model would be relevant for C/O degenerate secondaries if indeed carbon/oxygen discs do behave in a similar way as hydrogen rich discs. As can be seen the outbursts are significantly weaker than in the hydrogen poor case; they are so weak that the disc is always truncated, even at the outburst peak: the inner disc radius is always larger than 2.5 10$^{8}$ cm. The heating front reaches the outer disc edge only for the largest outbursts; for the weakest ones, the heating front does not propagate to distances larger than $1.1 \times 10^{10}$ cm.
\section{Steady sources}
Several ultracompact X-ray binaries are steady \citep{hvn12}, and these are sources bright enough for their disc to sit entirely on the hot stable branch \citep{ldk08}. However, some VFXBs are also found to be steady with luminosities of order of a few times $10^{34} - 10^{35}$ erg s$^{-1}$\citep[see e.g.][]{icm05,dsm07,djt10,adw13}. These sources could in principle be on the hot stable branch, provided that their orbital period is short enough. Assuming a primary mass of 1.4 M$_\odot$, and using Eqs. (12) and (14) from \citet{ldk08}, one finds that this happens for orbital periods shorter than 3.3 $L_{35}^{0.60}$ minutes in the case of helium discs and 13.2 $L_{35}^{0.63}$ min in the case of a mixed composition ($X=0.1$, $Y=0.9$), where $L_{35}$ is the X-ray luminosity in 10$^{35}$ erg s$^{-1}$ units. These orbital periods are extremely short and the mass-transfer rates expected from angular momentum losses due to gravitational radiation are orders of magnitude larger than the observed values.
The other option is that these sources stay on the cold stable branch. This is possible if the truncation radius is large enough, so that $\dot{M}_{\rm tr} < \dot{M}_{\rm crit}^-(r_{\rm in})$. For a pure helium disc, this happens when $r_{\rm in} > 3.1 \times 10^9 L_{35}^{0.38}$ cm. The situation would then be similar to that of \object{WZ Sge} for which \citet{hlh97} proposed that the source would stay on the cold stable branch during quiescence. WZ Sge outbursts would then be triggered if and when, for some reason, the mass-transfer rate increases above $\dot{M}_{\rm crit}^-(r_{\rm out})$. Mass transfer variations are observed in cataclysmic variables and LMXBs; for such a scenario to work one simply needs the average mass-transfer rate not to be much lower than $\dot{M}_{\rm crit}^-(r_{\rm out})$. \cite{lny96} and \cite{mnl99} suggested the existence of (almost) permanently quiescent X-ray binaries with large truncation radii. The first such system might have been recently found by \citet{tetarenkoetal16}. If this scenario were also applicable to quiescent VFXBs, it would raise the possibility that these sources undergo very rare outbursts.
There are a few faint quasi-persistent sources which stay active for decades but which sometimes undergo deep quiescence with levels $< 10^{33}$ erg s$^{-1}$ \citep{dsm07,ash15}. In principle, these sources should not be on the hot branch during the quasi-steady phase, for reasons explained above. If they are on the cool branch during these quasi-steady states, the transition to very faint quiescence cannot be due to the disc instability. The only remaining possibility is a drop in mass transfer from the secondary. These drops, possibly due to large stellar spots, have been invoked in H-rich CVs to account for some peculiar systems \citep[see e.g.][]{hl14}, but it remains to be shown that this mechanism is applicable to very low-mass helium secondaries.
\citet{hbd15} suggested that the persistent VXFBs can explained by the propeller effect enabling only a small fraction of the accretion flow to reach the neutron star surface. The effects of the magnetic field is complex, and the ability of the propeller mechanism to produce the observed X-ray luminosity remains to be investigated (see below).
\section{Conclusions}
We have shown that the disc instability model applied to ultracompact binaries produces outbursts whose characteristics are very similar to those of the faint and very faint outbursts observed in VFXBs, provided we use the same value of the viscosity parameter $\alpha_{\rm h}$ as in cataclysmic variables and in longer period X-ray binaries, and a smaller (factor a few) value of $\alpha_{\rm c}$ in quiescence. This is not surprising because the non-ideal-MHD effects that are likely affecting the magnetically driven turbulence in quiescent, poorly ionised discs \citep[e.g.][]{Bai14,Lesuretal14}, might depend on their chemical composition \citep[see, however,][]{kld12}. In any case, contrary to $\alpha_{\rm h}$, whose value is well determined from observations which consistently produced a value of $0.1 - 0.2$ \citep[see e.g.][]{KL12}, the value of $\alpha_{\rm c}$ does not have the same status because its observational determination is less direct than that of $\alpha_{\rm h}$. Various estimates give values from 0.0001 to 0.04 \citep[see e.g.][]{lasota-01}. The outburst faintness results from the small size of the accretion disc and from the low mass-transfer rate.
In the case of a 1.4 M$_\odot$ neutron star accreting from a helium degenerate secondary, the secular mean of the mass-transfer rate is \citep{hnv12}:
\begin{equation}
\dot{M}_{\rm tr}= 6.66 \times 10^{13} P_{\rm hr}^{-5.32} \; \rm g s^{-1} .
\end{equation}
For the 0.7 hr orbital period considered here, this gives $\dot{M}_{\rm tr} = 4.4 \times 10^{14}$ gs$^{-1}$. Of course, the actual mass-transfer rate from the secondary need not be equal to the secular mean, but it is worth noting that the values used in this paper are within factors $\sim$ a few equal to the secular mean.
We also predict that many of these systems with low mass-transfer rates could exhibit two classes of outbursts: those where the heat front is able to bring the whole disc in a hot state and are therefore relatively bright and last longer than the much fainter ones for which the heating front is unable to reach the outer disc edge. It is possible that these extremely faint outbursts are not detected as such because of their faintness -- in order to increase the sensitivity of short {\it Swift} exposures, data had to be added within time frames of 2-4 weeks --, and that they could be partly responsible for the luminosity variations observed during the intermediate state of very faint X-ray transients such as \object{XMM J174457-2850.3} \citep{dwr14}. One should, however, notice that the duty cycle we produce is small, and that it would be difficult to account for all of the observed luminosity variations observed during the intermediate state.
Very faint outbursts can also be produced in long period binaries, provided that the mass-transfer rate is low enough. For the $\dot{M}_{\rm tr} = 10^{13}$ gs$^{-1}$ cases presented in Section 2.2.2, the heating front never reached distances larger than a third of the disc size, and the outer parts of the disc were therefore in a steady state on the cool branch; the outburst properties are therefore independent of the disc size, and the same light curves would have been obtained for the same mass-transfer rate for longer orbital periods.
Very faint outbursts have been observed in several bright transients, such as for example \object{Aql X-1} \citep{ccd14}, \object{XTE J1701-462} \citep{f10}, \object{KS 1741-293} \citep{dw13} or SAX J1750.8-2900 \citep{wd13}. The possibility that these transients are of the same nature as those mentioned in this paper, with heat-front not able to propagate throughout the disc is interesting, but needs to be investigated in more details. This is clearly out of the scope of this paper.
One should also notice that the truncated helium disc that we have considered here is stable when the mass-transfer rate is lower than $\dot{M}_{\rm crit}^-(r_{\rm in})$; for the parameters we have used here, this corresponds to X-ray luminosities of $10^{33} - 10^{34}$ erg s$^{-1}$. Such stable low luminosity systems have been observed in the case of AM CVn systems \citep[see e.g.][and references therein]{Bildstenetal06} and are predicted by the DIM \citep{smak83,kld12}.
For a given orbital period, one therefore obtains the following sequence for decreasing the mass-transfer rates: for high $\dot{M}_{\rm tr}$, the system is stable on the hot branch; at lower $\dot{M}_{\rm tr} < \dot{M}_{\rm crit}^+(r_{\rm out})$, (relatively) bright outbursts are observed, during the entire disc is being brought in the hot state; for lower $\dot{M}_{\rm tr}$, one obtains a sequence of bright and faint outbursts during which the heating front does not reach the outer disc edge, and for still lower $\dot{M}_{\rm tr}$, only faint outbursts are obtained. Finally, for $\dot{M}_{\rm tr} < \dot{M}_{\rm crit}^-(r_{\rm in})$, the system is stable on the cold branch. In all cases, the outbursts are expected to be short, and depend mainly on $\alpha_{\rm c}$, and, for the outbursts able to reach the outer disc edge, on the orbital period. In all cases, faint outbursts are expected to be short if no additional ingredient, such as an increase of the mass-transfer rate due to the illumination of the secondary, is added.
We have not taken into account the interaction of the neutron star magnetic field with the accreted matter. This magnetic field could be the reason for the disc truncation; it could also prevent accretion onto the neutron star via the propeller effect, thereby accounting for the observed transition between the millisecond pulsar state and the LMXB state \citep[see e.g.][]{pt15}. However, the recent observation of X-ray pulsations in the accreting millisecond X-ray pulsar \object{PSR J1023+0038} while the system was in quiescence with an X-ray luminosity of $3 \times 10^{33}$ erg s$^{-1}$ \citep{abp15} or of \object{XSS J12270-4859} which has similar properties \citep{pmb15} shows that even at those low mass--accretion rates, the propeller mechanism does not always prevent accretion. It is also worth noting that the light cylinder radius, equal to $4.8 \times 10^6 P_{\rm ms}$ cm, where $P_{\rm ms}$ is the spin period of the millisecond pulsar, is smaller than the truncation radius given by Eq. (\ref{eq:rtrunc}) for millisecond rotation periods. The effect of the neutron star magnetic field is thus more complex than usually assumed, and it is unlikely that simple recipes can be used to decide if the magnetic field is able to disrupt the accretion disc or not.
Finally, one should keep in mind that the neither the VFXB nor the very faint X-ray transient categories are homogeneous. Apparently not all VFXBs are ultracompact and probably not all very faint outbursts can be explained by the DIM in ultracompact binaries.
\begin{acknowledgements}
We thank the anonymous referee for very useful comments which helped to improve this paper. This work was supported by a National Science Centre, Poland grant 2015/19/B/ST9/01099. JPL was supported by a grant from the French Space Agency CNES.
\end{acknowledgements}
|
\section{INTRODUCTION}
\label{sec:intro}
The development of extreme adaptive optics and high-contrast imaging techniques for ground-based telescopes (e.g. GPI \cite{GPI2006}, SPHERE \cite{SPHERE2008}, and SCExAO \cite{Martinache2009}) has enabled the detection and characterization of several young, giant exoplanets \cite{Marois2008,Lafreniere2008,Lagrange2009,Macintosh2015}. However, planets within the detection limits of current instruments are relatively rare \cite{Bowler2016}. The next generation of ground-\cite{Macintosh2006,Kasper2010} and space-based \cite{Noecker2016,Feinberg2014,Dalcanton2015} telescopes will be capable of detecting fainter, older, less massive planets at smaller angular separation from their host stars, thereby providing access to planet populations with significantly higher occurrence rates. Additionally, thorough spectral characterization will be possible for many of these targets thanks to rapid technological developments for precise control and calibration of unwanted stellar radiation, including dedicated coronagraphs for diffraction suppression, wavefront control, as well as new observing and post-processing strategies.
Detection and characterization of faint planets requires an optical system that isolates the light from the planet from noise associated with starlight. A coronagraph accomplishes this by manipulating the amplitude and phase of the incoming light such that the diffracted starlight is suppressed or removed optically prior to detection. Several elegant coronagraph designs exist that provide various levels of suppression and planet throughput \cite{Kuchner2002,Kasdin2003,Codona2004,Mawet2005,Foo2005,Guyon2005,Soummer2005,Trauger2007,Guyon2010}. The performance and complexity of each depends on the shape of the telescope pupil. Large, segmented apertures present a unique challenge; the coronagraph masks must be designed to account for the diffraction owing to discontinuities in the aperture including the secondary mirror, spider support structures, and gaps between mirror segments.
\cite{Mawet2011_improved,Mawet2013_ringapod,Carlotti2014,Ruane2015_SPIE,Ruane2015_LPM,Pueyo2013,Mazoyer2015,Guyon2014,Balasubramanian2015,Trauger2016,Zimmerman2016} (Also see Zimmerman et al. and Guyon et. al., these proceedings).
The vortex coronagraph (VC) \cite{Mawet2005,Foo2005} has been demonstrated to provide high sensitivity to planets at small angular separations \cite{Serabyn2010}. However, complicated aperture shapes limit the performance of the conventional VCs \cite{Mawet2010b} and thereby drive the complexity of the optical design \cite{Mawet2011_improved,Mawet2013_ringapod,Carlotti2014,Ruane2015_SPIE,Ruane2015_LPM} and/or requirements for wavefront control \cite{Pueyo2013,Mazoyer2015}. This work overcomes this technical challenge by introducing a gray-scale apodizer to the VC that acts to suppress polychromatic, diffracted starlight at angular separations $<10\lambda/D$ potentially down to the $10^{-10}$ level on a segmented aperture telescope similar to those proposed for a future LUVOIR flagship mission \cite{Dalcanton2015,SCDA}.
\newpage
\section{CORONAGRAPH OPTIMIZATION} \label{sec:opt}
\begin{figure}[t]
\centering
\includegraphics[width=0.85\linewidth]{OpticalDiagram.pdf}
\caption{Schematic of a coronagraph instrument, with deformable mirrors DM1 and DM2, a pupil-plane apodizer~$A$, focal plane mask~$\Omega$, and a Lyot stop~$\Theta$ in the downstream pupil. The black arrows represent powered optics.}
\label{fig:diagram}
\end{figure}
\begin{figure}[b]
\centering
\hspace{1.25cm}
\includegraphics[width=0.7\linewidth]{RAVC_fig.pdf}\hspace{0.2cm}
\includegraphics[trim={0 -10mm 0 0},clip,scale=0.7]{colorbarsforfig3.pdf}\hspace{0.2cm}
\includegraphics[trim={0 -10mm 0 0},clip,scale=0.7]{colorbarsforfig4.pdf}
\caption{(a) Charge 2 and (b) charge 4 ring-apodized vortex coronagraphs designed for an annular pupil with central obscuration ratio $R_0=0.14$. The dotted line in the Lyot stop (LS) indicates the boundary of the full geometric pupil.}
\label{fig:baseline}
\end{figure}
\subsection{The optical system and baseline design}
A high-contrast imaging coronagraph instrument is made up of a wavefront control sub-system with one or two deformable mirrors (DMs) and a series of coronagraphic masks arranged between powered optics. Fig. \ref{fig:diagram} shows an example system with two DMs, a pupil plane apodizer $A$, a focal plane mask $\Omega$, and a Lyot stop $\Theta$ in the downstream pupil. In general, a coronagraph may also include masks or surfaces displaced from the pupil and focal planes and/or in additional optical relays.
The designs presented here are based on the ring-apodized vortex coronagraph (RAVC) \cite{Mawet2013_ringapod}, which provides theoretically ideal and achromatic on-axis starlight cancellation with an annular aperture. The analytically-inspired design is shown in Fig. \ref{fig:baseline}. The apodizer is a semi-transparent ring with amplitude transmittance $t$ that extends from $R_\mathrm{gray}$ to $R$, where $R_\mathrm{gray}$ is the inner radius of the gray ring and $R$ is the outer radius of the pupil. The focal plane mask has complex transmittance $\exp(il\phi)$, where $l$ is an even, nonzero integer known as the charge and $\phi$ is the azimuthal angle. The Lyot stop is a annulus with inner and outer radii $R_\mathrm{gray}$ and $R$, assuming there is no magnification between the pupils.
For $l=2$, the transmittance of the gray ring is given by $t = 1- (R_0/R_\mathrm{gray})^2$, where $R_0$ is the ratio of the inner and outer radii of the full annular aperture. For $l=4$, there is an additional narrow black ring from $R_\mathrm{black}$ to $R_\mathrm{gray}$, $R_\mathrm{gray} = \sqrt{(R_\mathrm{black}^4-R_0^4)/(R_\mathrm{black}^2-R_0^2)}$, and the transmittance of the gray ring is given by $t = (R_\mathrm{black}^2-R_0^2)/R_\mathrm{gray}^2$. The remaining free parameter is varied to maximize the throughput.
Several coronagraph throughput definitions may be found in the literature. In this work, we use two measures: the total energy throughput (i.e. the fraction of energy from a point source that transmits through the Lyot stop) and the fraction of energy within 0.7$\lambda/D$ of the source position, which is roughly the half-width half-maximum of an ideal point spread function (PSF). The latter definition is the relevant quantity for detecting point sources in noisy data using aperture photometry, as is the case in common post-processing approaches for exoplanet detection. We note, however, that this may be a conservative estimate for throughput provided advanced matched-filtering and local deconvolution techniques can make use of the planet light outside of the PSF core, which will be the topic of future studies.
\begin{figure}
\centering
\includegraphics[height=0.322\linewidth]{TipTiltSensitivity.pdf}
\includegraphics[trim={0.6cm 0 0 0},clip,height=0.322\linewidth]{LeakFiniteSource_analyt.pdf}
\includegraphics[trim={0.6cm 0 0 0},clip,height=0.322\linewidth]{Thpt_RAVC4_compare.pdf}
\caption{Sensitivity of the vortex coronagraph to (a) tip-tilt, (b) stellar angular size, and (c) central obscuration size. (a)-(b) The total energy throughput for an unobstructed circular pupil. (c) Throughput, averaged over source positions 3-5 $\lambda/D$, of a charge 4 ring-apodized vortex coronagraph (RAVC) for a simple annular aperture with central obscuration ratio $R_0$. Both the total energy and the energy within 0.7 $\lambda/D$ of the source position are shown.}
\label{fig:baselinethpt}
\end{figure}
The value of the charge $l$ controls the off-axis throughput as well as sensitivity to tip-tilt, jitter, and stellar angular size (see Fig. \ref{fig:baselinethpt}a-b) \cite{Mawet2010b}. Although $l=2$ provides the best throughput for planets at small angular separations from their host stars, the performance will be heavily degraded on next-generation ground- and space-based telescopes owing to low-order aberrations and the partial resolution of the star. To hedge against this, we choose to use a minimum charge of $l=4$ in our designs. $l>4$ may be used in the future to reduce the sensitivity to such aberrations as needed at the cost of off-axis throughput.
The maximum throughput of an RAVC has a strong dependence on the size of the central obscuration. Fig.~\ref{fig:baselinethpt}c shows the throughput, averaged over angular separations 3-5~$\lambda/D$, as a function of $R_0$ in the $l=4$ case. Future segmented aperture telescopes, including the Thirty Meter Telescope (TMT) and potential space missions, will have $R_0$ values $\sim$14\%. The throughput loss with increasing $R_0$ is due to both the apodizer transmittance $t$ and the shape of the Lyot stop.
We emphasize that the throughput values shown in Fig.~\ref{fig:baselinethpt}c are strictly for the RAVC and may be improved with alternate apodizer functions\cite{Carlotti2014} (Fogarty et al. 2016, in prep.) and/or modifications to the complex pupil field using low-loss techniques \cite{Pueyo2013,Mazoyer2015}. However, for the new coronagraphs presented below, the RAVC architecture serves as the initial condition of our apodizer optimization algorithm and therefore the throughput generally follows the trend shown.
\subsection{Optimization method}
\label{subsec:method}
For arbitrary apertures, we use an iterative numerical optimization method to determine the optimal gray-scale apodizer to achieve a dark hole in the on-axis starlight at the final image plane. The problem is written in linear algebraic form as
\begin{equation}
\underset{w}\min\left(||QCw||^2 + b||w - P A||^2\right),
\end{equation}
where $Q$ is a matrix that represents the dark hole region in the image plane, $C$ is the coronagraph operator that propagates the field from the apodizer plane to the final image plane, $P$ is the original telescope pupil function, $A$ is the current apodizer, and $w$ is the so-called auxiliary field in the apodizer plane which strikes a balance (as regulated by the weight $b$) between the field needed to generate a zero-valued dark hole and the physical field in the apodizer plane. Assuming the pupil and focal planes are related by Fourier transform propagation operators $F$, the coronagraph operator may be written $C=F \Theta F^{-1} \Omega F$, where $\Omega$ and $\Theta$ represent the focal plane mask and Lyot stop transmittance, respectively. The solution to the minimization problem is
\begin{equation}
w=(b I + C^\dagger Q C)^{-1} b P A.
\end{equation}
To reduce computation time, the dimensionality of the inverted (square) matrix is reduced from the number of samples in the pupil plane to the number of samples in the dark hole region, by use of the Woodbury matrix identity:
\begin{equation}
w = \left[ I - C^\dagger Q(b Q + Q C C^\dagger Q)^{-1}Q C\right] P A.
\end{equation}
Since $w$ may be a complex function with infinite support, the physical apodizer is taken to be $A = |w|$ and $A$ is thresholded such that samples where $A>1$ are set to one and non-zero values outside of the original telescope pupil $P$ support are set to zero. A new auxiliary field is calculated based on the updated pupil field, and the process is repeated. The matrix $Q C C^\dagger Q$ is calculated once for each choice of focal plane mask, Lyot stop, and dark hole region as follows:
\begin{equation}
C C^\dagger = (F \Theta F^{-1} \Omega F) (F^{-1} \Omega^\dagger F \Theta^\dagger F^{-1}),
\end{equation}
\begin{equation}
C C^\dagger = F \Theta F^{-1} |\Omega|^2 F \Theta^\dagger F^{-1}.
\end{equation}
Since the focal plane mask has phase-only transmittance (i.e. $|\Omega|^2=I$), this matrix simplifies to
\begin{equation}
C C^\dagger = F |\Theta|^2 F^{-1}
\end{equation}
and only depends on the squared modulus of Lyot stop function.
We note that this algorithm may lead to very low throughput for poor choices of focal plane masks, Lyot stops, and apodizer initial conditions. The RAVC is a suitable initial design that leads to high throughput, even at at relatively small angular separations (potentially $<3~\lambda/D$). Another benefit of the RAVC, is that the numerical apodizer solution applies to all wavelengths.
In general, the starlight suppression in the dark hole region $Q$ is achieved at the cost of off-axis throughput. Assuming a good initial condition is chosen, the apodizer achieves a dark hole in the on-axis starlight without major losses. Finding the optimal combination of throughput and starlight suppression provided by the coronagraph masks requires performance metrics that relate these quantities to the sensitivity of the instrument to faint point sources in the presence of noise.
\subsection{Optimization metrics: minimizing estimated integration time}
The proposed optimization metrics are based on estimates of the signal-to-noise ratio (SNR) achieved in a given integration time for a typical planet, given by
\begin{equation}
\mathrm{SNR} = \frac{\epsilon \Delta t \Delta\lambda \eta_c\Phi_p}{\sqrt{\sigma^2_\mathrm{phot}+\sigma^2_\mathrm{det}+\sigma^2_\mathrm{spk}}},
\end{equation}
where the numerator represents the number of photo-electrons generated by planet light, $\epsilon = T q A_\mathrm{tel}$, $T$ is the telescope transmission, $q$ is the quantum efficiency, $A_\mathrm{tel}$ is collecting area of the telescope, $\Delta t$ is the effective integration time, $\Delta\lambda$ is the spectral bandwidth, $\eta_c$ is the coronagraph throughput, and $\Phi_p$ is the photon flux from the planet at the telescope aperture (photons per unit area per unit time per unit wavelength).
The throughput of the coronagraph $\eta_c$ and the designed starlight suppression factor $s$ are defined as the fraction of planet and stellar energy that is incident on a single resolution element (a circle with assumed radius of 0.7 $\lambda/D$) for an ideal system without optical aberrations or atmospheric turbulence. Similarly, we define $\eta_0$ as the throughput without the coronagraphic masks in the system.
The photon noise term may be expressed as $\sigma^2_\mathrm{phot} = \epsilon \Delta t \Delta\lambda \hat{\Phi}$, where $\hat{\Phi} = \eta_c\Phi_p + s\Phi_\mathrm{star} + \Phi_b + \hat{\Phi}_\mathrm{spk}$ is the total flux within the resolution element, and $\Phi_\mathrm{star}$, $\Phi_b$, and $\hat{\Phi}_\mathrm{spk}$ are the photon fluxes owing to the stellar diffraction, background, and speckles, respectively. $\Phi_p$, $\Phi_\mathrm{star}$, and $\Phi_b$ are defined at the telescope aperture, whereas $\hat{\Phi}_\mathrm{spk}$ is defined in the image plane with respect to $\eta_0\Phi_\mathrm{star}$ and is treated separately from the diffracted starlight $s\Phi_\mathrm{star}$.
The detector noise is given by $\sigma^2_\mathrm{det} = i_d \Delta t + \sigma^2_\mathrm{read}$, where $i_d$ is the dark current and $\sigma^2_\mathrm{read}$ is the read noise averaged over many frames, which we approximate as $\sigma^2_\mathrm{read} \approx N_r^2 \epsilon \Delta t \Delta\lambda \hat{\Phi}_\mathrm{max}/W = \dot{R} \Delta t$, where $N_r$ is the read-out noise for each frame, $\hat{\Phi}_\mathrm{max}$ is the maximum photon flux in the image plane, and $W$ is the full well depth of the detector.
The speckle flux is split into contributions owing to dynamic and quasi-static aberrations: $\hat{\Phi}_\mathrm{spk}=\hat{\Phi}_\mathrm{spk,dyn}+\hat{\Phi}_\mathrm{spk,qs}$. The former corresponds to residual wavefront distortions owing to the atmosphere with an average lifetime $\tau_\mathrm{dyn} = D/v$, where $D$ is outer diameter of the telescope aperture and $v$ is the wind speed. The latter is the slowly-varying aberration term owing to thermal and mechanical distortions in the system, which generate quasi-static speckles with lifetimes, $\tau_\mathrm{qs}$, on the order of hours. The effective speckle noise, excluding photon noise contributions, is approximated by
\begin{equation}
\sigma^2_\mathrm{spk} = \Delta t (\epsilon \Delta\lambda)^2 \tau_\mathrm{dyn} \mathrm{var}(\hat{\Phi}_\mathrm{spk}),
\end{equation}
where
\begin{equation}
\tau_\mathrm{dyn}\mathrm{var}(\hat{\Phi}_\mathrm{spk})=\tau_\mathrm{dyn}\hat{\Phi}_\mathrm{spk,dyn}^2 + \tau_\mathrm{qs} \hat{\Phi}_\mathrm{spk,qs}^2 + 2 s \Phi_\mathrm{star} \left(\tau_\mathrm{dyn}\hat{\Phi}_\mathrm{spk,dyn} + \tau_\mathrm{qs}\hat{\Phi}_\mathrm{spk,qs}\right) + 2 \tau_\mathrm{dyn} \hat{\Phi}_\mathrm{spk,dyn} \hat{\Phi}_\mathrm{spk,qs},
\end{equation}
and $\mathrm{var}(\hat{\Phi}_\mathrm{spk})$ denotes the speckle flux variance, derived in Soummer et al. (2007) \cite{Soummer2007}. For a space telescope, the contribution of quickly-varying speckles is negligible, and the speckle noise term reduces to
\begin{equation}
\sigma^2_\mathrm{spk} = \Delta t (\epsilon \Delta\lambda)^2 \left( \tau_\mathrm{qs} \hat{\Phi}_\mathrm{spk,qs}^2 + 2 \tau_\mathrm{qs} s \Phi_\mathrm{star} \hat{\Phi}_\mathrm{spk,qs}\right).
\end{equation}
Recently developed methods to mitigate the speckle noise level via on-sky speckle nulling \cite{Martinache2014}, angular differential imaging \cite{Marois2006}, and sophisticated post-processing algorithms \cite{Soummer2012} significantly reduce the $\sigma^2_\mathrm{spk}$ term. We approximate these gains by using an effective quasi-static speckle flux level $g_\mathrm{sn} g_\mathrm{pp}\hat{\Phi}_\mathrm{spk,qs}$ in our calculations, where $g_\mathrm{sn}$ is the gain achieved through speckle nulling and $g_\mathrm{pp}$ is the gain achieved through speckle estimation and removal in post-processing.
\begin{figure}[t]
\centering
\includegraphics[trim={0 0 1.6cm 0},clip,height=0.345\linewidth]{exptime_Keck_allv2.pdf}\hfill
\includegraphics[trim={0.7cm 0 1.6cm 0},clip,height=0.345\linewidth]{exptime_TMT_allv2.pdf}\hfill
\includegraphics[trim={0.7cm 0 1.25cm 0},clip,height=0.345\linewidth]{exptime_LUVOIR_allv2.pdf}
\caption{Estimates of integration time $\log_{10}(\Delta t)$ needed for a 5$\sigma$ detection with designed starlight suppression factor $s$ and throughput $\eta_c$ assuming planet:star contrast (a) $\Phi_p/\Phi_\mathrm{star}=10^{-5}$ with Keck in L band, (b) $\Phi_p/\Phi_\mathrm{star}=10^{-6}$ with TMT in K band, and (c) $\Phi_p/\Phi_\mathrm{star}=10^{-10}$ with a future LUVOIR space telescope in V band. $\eta_c$ and $s$ are solely properties of the coronagraph design and refer to the performance under ideal conditions. The effect of $s$ on speckle noise is described by Eqn. \ref{eqn:dt_spk}. The assumptions used for these calculations may be found in the appendix.}
\label{fig:exp_time}
\end{figure}
\newpage
Since $\sigma^2_\mathrm{phot}$, $\sigma^2_\mathrm{det}$, and $\sigma^2_\mathrm{spk}$ increase linearly with $\Delta t$ under these assumptions, we can solve for the integration time needed to reach an SNR threshold $\Gamma$:
\begin{equation}
\Delta t = \Gamma^2\left(\Delta t_\mathrm{phot} + \Delta t_\mathrm{det} + \Delta t_\mathrm{spk}\right),
\end{equation}
where $\Delta t_\mathrm{phot}$, $\Delta t_\mathrm{det}$, and $\Delta t_\mathrm{spk}$ are the scale times to overcome each noise term:
\begin{equation}
\Delta t_\mathrm{phot} = \frac{1}{(\eta_c \Phi_p)^2}\frac{\hat{\Phi}}{\epsilon \Delta\lambda},
\end{equation}
\begin{equation}
\Delta t_\mathrm{det} = \frac{1}{(\eta_c \Phi_p)^2}\frac{\dot{R} + i_d }{(\epsilon \Delta\lambda)^2},
\end{equation}
\begin{equation}
\Delta t_\mathrm{spk}=\frac{1}{(\eta_c \Phi_p)^2}\left\{ \begin{array}{ll}
\tau_\mathrm{dyn} \mathrm{var}(\hat{\Phi}_\mathrm{spk}) & \text{for ground-based telescopes,} \\
\tau_\mathrm{qs} \hat{\Phi}_\mathrm{spk,qs}^2 + 2 \tau_\mathrm{qs} s \Phi_\mathrm{star} \hat{\Phi}_\mathrm{spk,qs} & \text{for space telescopes.} \\
\end{array} \right.
\label{eqn:dt_spk}
\end{equation}
In the calculations shown in Fig. \ref{fig:exp_time}, it can be seen designing the coronagraph masks to have a smaller value of $s$ leads to shorter integration times until the dominant speckle level is reached. The integration time scales as
\begin{equation}
\Delta t \propto \left\{ \begin{array}{ll}
s/\eta_c^2 & s>\hat{\Phi}_\mathrm{spk,0}/ (2 \eta_0 \Phi_\mathrm{star}) \\
1/\eta_c^2 & s<\hat{\Phi}_\mathrm{spk,0}/ (2 \eta_0 \Phi_\mathrm{star}) \\
\end{array} \right.,
\end{equation}
where $\hat{\Phi}_\mathrm{spk,0}$ is the dominant speckle flux term: $\hat{\Phi}_\mathrm{spk,0} \sim \hat{\Phi}_\mathrm{spk,dyn}$ for ground-based and $\hat{\Phi}_\mathrm{spk,0} \sim \hat{\Phi}_\mathrm{spk,qs}$ for space-based applications. This transition occurs at $s \sim 10^{-6}-10^{-5}$ for ground-based telescopes and $s \sim 10^{-11}-10^{-10}$ in space. The integration time depends strongly on $\eta_c$ in both regimes and therefore the throughput is the single most important coronagraph design parameter for minimizing the integration time.
\section{Apodized vortex coronagraphs for segmented apertures} \label{sec:solutions}
Using the method outlined in section \ref{subsec:method}, we optimized apodizing pupil masks for a charge 4 vortex coronagraph on the Thirty Meter Telescope (TMT) and potential future space telescopes \cite{SCDA}.
\begin{figure}[t]
\centering
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{TMT/EP.pdf}\hfill
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{TMT/APOD.pdf}\hfill
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{TMT/LS.pdf}\hspace{1mm}
\includegraphics[trim={0 -8mm -5mm 0 },clip,scale=0.65]{colorbarsforfig3.pdf}\\
\includegraphics[height=0.28\linewidth]{TMT/radproPSF.pdf}\hfill
\includegraphics[height=0.275\linewidth]{TMT/Thpt.pdf}\hfill
\includegraphics[height=0.28\linewidth]{TMT/iPSF.pdf}
\includegraphics[trim={0 -8mm 0 0},clip,scale=0.65]{colorbarsforfig5.pdf}\\
\caption{Apodizer for Thirty Meter Telescope (TMT) aperture. (a)~Pupil amplitude $P$ prior to the apodizer. (b)~Gray-scale apodizer mask $A$ in pupil. (c)~Lyot stop $\Omega$. (d)~Azimuthal average of the on-axis PSF. (e)~Throughput of the coronagraph masks, normalized the to telescope throughput. (f)~On-axis PSF (log irradiance) for an ideal wavefront, normalized to the peak of the telescope PSF, without optical aberrations.}
\label{fig:TMT}
\end{figure}
\subsection{Apodizer for the Thirty Meter Telescope (TMT)}
The goal of ground-based coronagraphic instruments is to minimize integration time by reducing the residual starlight factor $s$ below the dynamic speckle noise level, while maintaining as much throughput as possible. For TMT, we assume $\hat{\Phi}_\mathrm{spk,dyn}=10^{-5}\eta_0\Phi_\mathrm{star}$ and therefore pinned speckles are sufficiently mitigated if $s \ll 5\times10^{-6}$. We design our masks to achieve an ideal value of $s = 10^{-9}$, which leaves a considerable margin for additional slowly-varying speckles caused by manufacturing errors and unforeseen aberrations in the optical system. We also find that the resulting throughput is relatively insensitive to the design value of $s$ when an RAVC is used as the initial condition ($\sim\!6\%$ loss with respect to the initial RAVC throughput).
An apodizer solution for TMT and its corresponding performance is shown in Fig. \ref{fig:TMT}. The coronagraph is made up of three masks: a gray-scale apodizer $A$ in the pupil (see Fig. \ref{fig:TMT}a-b), a charge 4 vortex phase mask in the focal plane $\Omega$, and an annular Lyot stop $\Theta$ (see Fig. \ref{fig:TMT}c). The azimuthal average of the on-axis PSF, shown in Fig. \ref{fig:TMT}d, is at the $\sim\!10^{-9}$ level throughout the dark hole when normalized to the telescope PSF, and therefore $s \approx 10^{-9}$ within an annulus from 3 to 10~$\lambda/D$. The throughput (see Fig. \ref{fig:TMT}e) increases from small to large angular separations across the dark hole region. The energy within the planet PSF core is greater than 15\% for angular separations $>3~\lambda/D$. Although the throughput is calculated for point sources displaced along a single direction, the throughput should be approximately the same in all directions, and the starlight is uniformly suppressed throughout the dark hole (see Fig. \ref{fig:TMT}f).
\begin{figure}[p]
\centering
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{clippedhex4_w_spiders/EP.pdf}\hfill
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{clippedhex4_w_spiders/APOD.pdf}\hfill
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{clippedhex4_w_spiders/LS.pdf}\hspace{1mm}
\includegraphics[trim={0 -8mm -5mm 0 },clip,scale=0.65]{colorbarsforfig3.pdf}\\
\includegraphics[height=0.28\linewidth]{clippedhex4_w_spiders/radproPSF.pdf}\hfill
\includegraphics[height=0.275\linewidth]{clippedhex4_w_spiders/Thpt.pdf}\hfill
\includegraphics[height=0.28\linewidth]{clippedhex4_w_spiders/iPSF.pdf}
\includegraphics[trim={0 -8mm 0 0},clip,scale=0.65]{colorbarsforfig5.pdf}\\
\caption{Same as Fig. \ref{fig:TMT}, but for a segmented aperture with four-rings of hexagonal mirrors and thick spiders. The dotted lines in (e) show the throughput achieved with the spiders removed.}
\label{fig:clippedhex4_w_spiders}
\vspace*{\floatsep}
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{piewedge8_w_spiders/EP.pdf}\hfill
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{piewedge8_w_spiders/APOD.pdf}\hfill
\includegraphics[trim={0 0 0 0},clip,height=0.28\linewidth]{piewedge8_w_spiders/LS.pdf}\hspace{1mm}
\includegraphics[trim={0 -8mm -5mm 0 },clip,scale=0.65]{colorbarsforfig3.pdf}\\
\includegraphics[height=0.28\linewidth]{piewedge8_w_spiders/radproPSF.pdf}\hfill
\includegraphics[height=0.275\linewidth]{piewedge8_w_spiders/Thpt.pdf}\hfill
\includegraphics[height=0.28\linewidth]{piewedge8_w_spiders/iPSF.pdf}
\includegraphics[trim={0 -8mm 0 0},clip,scale=0.65]{colorbarsforfig5.pdf}\\
\caption{Same as Fig. \ref{fig:clippedhex4_w_spiders}, but for a segmented aperture with eight pie-wedge mirrors and thick spiders. The dotted lines in (e) show the throughput achieved with the spiders removed.}
\label{fig:piewedge8_w_spiders}
\end{figure}
\subsection{Apodizers for future segmented-aperture space telescopes}
A major driver for investing in coronagraphic instruments on space telescopes is that the speckle level will be far lower without the contributions from the quickly varying atmosphere ($\hat{\Phi}_\mathrm{spk,dyn}=0$ and $\hat{\Phi}_\mathrm{spk,qs}=10^{-9}\eta_0\Phi_\mathrm{star}$). Therefore, the goal for space-based applications is to reduce $s$ as low as $10^{-10}$ at the highest possible throughput. Here, we consider two potential aperture types for future segmented space telescopes: a segmented primary with four-rings of hexagonal mirrors (see Fig. \ref{fig:clippedhex4_w_spiders}) and one with eight pie-wedge mirrors (see Fig. \ref{fig:piewedge8_w_spiders}). The spiders in each case are chosen to be co-aligned with discontinuities between segments.
In the case of the hexagonally-segmented pupil (Fig.~\ref{fig:clippedhex4_w_spiders}), the optimized apodizer reduces the starlight to the $10^{-10}$ level while providing 12\% PSF core throughput at 3~$\lambda/D$. In the case of the pie-wedge aperture (Fig.~\ref{fig:piewedge8_w_spiders}), the PSF core throughput is 20\% at 3~$\lambda/D$. The gain in throughput is mostly attributed to the size of central obscuration, which is slightly smaller in the case of the pie-wedge aperture ($R_0=0.17$ versus $R_0=0.14$).
We find that the discontinuities in the pupil have a much smaller effect on the throughput achieved. At $s\approx10^{-10}$, the throughput difference with and without spiders is only a couple of percent, as indicated by the dotted lines in Figs. \ref{fig:clippedhex4_w_spiders}e and \ref{fig:piewedge8_w_spiders}e.
All of the apodized vortex solutions shown are theoretically broadband; that is, the suppression and throughput shown are expected at all wavelengths provided the masks have constant complex transmittance over the passband. The practical broadband performance will only be limited by manufacturing constraints.
\newpage
\section{CONCLUSIONS AND FUTURE OUTLOOK} \label{sec:concl}
We have presented apodized pupil vortex coronagraphs designed for ground- and space-based telescopes with segmented apertures. in each case, the coronagraph masks are optimized such that the estimated integration time is minimized, in the presence of noise.
The design goals in terms of starlight suppression and throughput for ground- and space-based applications depend mostly strongly on the expected speckle noise characteristics. In case of TMT, the coronagraph masks suppress diffracted starlight to the $10^{-9}$ level, assuming no aberrations in the system, which is well below the expected speckle noise level. For space telescopes, we show solutions that push the diffraction down to the $10^{-10}$ level while maintaining sufficient throughput to significantly reduce integration time estimates.
The throughput achieved is relatively insensitive to aperture discontinuities. In the case of the space telescopes, the throughput with and without thick spiders only differs by a couple of percent. However, the central obscuration size has a much larger effect. The best performance is expected for telescopes with relatively small secondary mirrors.
Although the theoretical solutions are independent of wavelength, manufacturing errors will ultimately limit the performance of these coronagraph designs. Pathways to approach the ideal performance are available thanks to successful demonstrations of broadband vortex phase masks based on liquid crystal polymers \cite{Mawet2009} and sub-wavelength gratings \cite{Delacroix2013} as well as gray-scale ring apodizers for vortex coronagraphs \cite{Mawet2014}. Detailed simulations are underway to study the chromaticity of gray-scale apodizers fabricated using various methods. Outcomes of these studies will inform a second generation of the presented coronagraph designs that incorporate known material properties.
The methods employed here may be readily generalized to include optimization of the deformable mirror shapes to achieve broadband starlight suppression, potentially at high throughput. A comprehensive exploration of apodizer solutions and designs that also make use two deformable mirrors will be the topic of an upcoming paper.
\newpage
|
\section{\@startsection{section}{1}%
\z@{.7\linespacing\@plus\linespacing}{.5\linespacing}%
{\bfseries
\centering
}}
\def\@secnumfont{\bfseries}
\makeatother
\usepackage{graphicx}
\newcommand{\textcolor}{\textcolor}
\newcommand{\mathbb C}{\mathbb C}
\newcommand{\mathbb D}{\mathbb D}
\newcommand{\mathbb R}{\mathbb R}
\newcommand{\mathbb N}{\mathbb N}
\newcommand{\mathbb Q}{\mathbb Q}
\newcommand{\mathbb Z}{\mathbb Z}
\newcommand{\mathbf{B}}{\mathbf{B}}
\newcommand{\mathbf{E}}{\mathbf{E}}
\newcommand{\mathbf{G}}{\mathbf{G}}
\newcommand{\mathbf{P}}{\mathbf{P}}
\newcommand{\mathbf{\Gamma}}{\mathbf{\Gamma}}
\newcommand{{\bf 1}}{{\bf 1}}
\newcommand{{\bf 2}}{{\bf 2}}
\newcommand{\mathcal B}{\mathcal B}
\newcommand{\mathcal C}{\mathcal C}
\newcommand{\hat {\mathcal C}}{\hat {\mathcal C}}
\newcommand{\mathcal E}{\mathcal E}
\newcommand{\mathcal F}{\mathcal F}
\newcommand{\mathcal H}{\mathcal H}
\newcommand{\mathcal I}{\mathcal I}
\newcommand{\mathcal J}{\mathcal J}
\newcommand{\mathcal L}{\mathcal L}
\newcommand{\mathcal N}{\mathcal N}
\newcommand{\mathcal Q}{\mathcal Q}
\newcommand{\mathcal S}{\mathcal S}
\newcommand{\mathcal T}{\mathcal T}
\newcommand{\mathcal W}{\mathcal W}
\newcommand{\mathcal Z}{\mathcal Z}
\newcommand{\mathcal R}{\mathcal R}
\newcommand{\alpha}{\alpha}
\newcommand{\gamma}{\gamma}
\newcommand{\Gamma}{\Gamma}
\newcommand{\delta}{\delta}
\newcommand{\varepsilon}{\varepsilon}
\newcommand{\iota}{\iota}
\newcommand{\kappa}{\kappa}
\newcommand{\lambda}{\lambda}
\newcommand{\Lambda}{\Lambda}
\newcommand{\omega}{\omega}
\newcommand{\Omega}{\Omega}
\newcommand{\Sigma}{\Sigma}
\newcommand{\sigma}{\sigma}
\newcommand{\theta}{\theta}
\newcommand{\varphi}{\varphi}
\newcommand{\zeta}{\zeta}
\newcommand{\left(}{\left(}
\newcommand{\right)}{\right)}
\newcommand{\left[}{\left[}
\newcommand{\right]}{\right]}
\newcommand{\left\{}{\left\{}
\newcommand{\right\}}{\right\}}
\newcommand{\left|}{\left|}
\newcommand{\right|}{\right|}
\newcommand{\left\langle}{\left\langle}
\newcommand{\right\rangle}{\right\rangle}
\newcommand{\left\langle}{\left\langle}
\setlength{\textheight}{19.5 cm}
\setlength{\textwidth}{14 cm}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{remark}
\newtheorem{remark}{Remark}
\numberwithin{equation}{section}
\setcounter{page}{1}
\begin{document}
\title[FSDE with discontinuous diffusion]{Fractional stochastic differential equation with discontinuous diffusion}
\author[J.Garz\'on]{Johanna Garz\'on}
\address{Departamento de Matem\'aticas, Universidad Nacional de Colombia, Carrera 45 No 26-85 Bogot\'a, Colombia. }
\email{<EMAIL>}
\author[J.A. Le\'on]{Jorge A. Le\'on}
\address{Departamento de Control Autom\'atico, Cinvestav-IPN,
Apartado Postal 14-740, 07000 M\'exico D.F., Mexico}
\email{<EMAIL>}
\author[S.Torres]{Soledad Torres}
\address{Facultad de Ingenier\'ia, CIMFAV Universidad de Valpara\'iso, Casilla 123-V, 4059 Valparaiso, Chile. }
\email{<EMAIL>}
\begin{abstract}
In this paper we study a stochastic differential equation driven by a fractional Brownian motion with a discontinuous coefficient. We also give an approximation to the solution of the equation. This is a first step to define a fractional version of the skew Brownian motion.
\end{abstract}
\maketitle
\medskip\noindent
{\bf Mathematics Subject Classifications (2000)}: Primary 60H15; secondary 65C30.
\medskip\noindent
{\bf Keywords:} Fractional Brownian motion, Fractional calculus,
Pathwise differential equations, Young integral
\allowdisplaybreaks
\section{Introduction}
In this article we show existence and uniqueness for a stochastic differential equation driven by a fractional Brownian motion with Hurst parameter $H >1/2$ and a discontinuous diffusion coefficient.
Recall that a fractional Brownian motion $B^H$, is a centered Gaussian process with covariance structure given by:
$$
R_H(t,s) = \mathbb{E}(B_t^H B_s^H) = \frac12 \left\{ |t|^{2H} + |s|^{2H}| - |t-s|^{2H} \right\} .
$$
The particular case $H= 1/2$ results the Brownian motion.
The equation we will consider in this article is given by:
\begin{equation}\label{DISC}
x_t = x_0 + \int_0^t \sigma (x_s) d B^H_s , \quad t \geq 0;
\end{equation}
where $\sigma$ is a discontinuous function given by
$$\sigma (x) = \frac{1}{\alpha} 1\{x \geq 0\} + \frac{1}{1-\alpha} 1 \{x < 0\}, $$
for some $\alpha \in (0,1)$. Note that,
without loss of generality, we can assume that $\alpha < 1/2$.
The following picture shows the function $\sigma$.
\newpage
\begin{figure}[htbp]
\centering
{\includegraphics[width=53mm] {ALPHA1.eps}}
\caption{$\sigma(x)$ for $\alpha=0.4$}\label{}
\end{figure}
Nakao \cite{Nakao} in 1972, solved the problem of pathwise uniqueness of solutions of SDE driven by Brownian motion, ($H=1/2$), and with a diffusion coefficient uniformly positive and of bounded variation on any compact interval. The author proved the pathwise uniqueness holds for such equation. After that, many works have been developed in this direction, in order to show existence and/or uniqueness for SDE with general coefficients in the diffusion. We can refer here \cite{Ouknine} and \cite{Lejay} and the references therein among others.
In the case $H > 1/2$, the only cases of stochastic differential equations driven a fractional Brownian motion with discontinuous coefficients which have been studied are those corresponding to discontinuous drift coefficient (for $H>1/2$). Regarding that, in \cite{MN}, the authors studied a H\"older continuous drift except on a finite numbers of points. Other class of discontinuity in SDE driven a fractional Brownian motion is related to add to a Poisson process for the SDE. In \cite{BM}, the authors proved the strong solution of this kind of SDE driven by fBm and Poisson point process extending the results given in \cite{MN}.
The main interest in working in this type of equations is related with the problem to define a fractional version of the Skew Brownian motion. In the Brownian motion framework the Skew Brownian motion appeared as a natural generalization of the Brownian motion. The Skew Brownian motion is a process that behaves like a Brownian motion except that the sign of each excursion is chosen using an independent Bernoulli random variable of parameter $p \in (0 , 1)$. For $p = 1/2$ the process corresponds to a Brownian motion. This process is a markov process, semi-martingale which is a strong solution to some Stochastic Differential Equation (SDE) with local time, (see \cite{Lejay} for a survey).
\begin{equation}\label{SB}
X_t = x + B_t + (2p - 1) L^0_t(X).
\end{equation}
In the case of Brownian motion and from It$\hat{o}$-Tanaka formula, equation (\ref{DISC}) and (\ref{SB}) are equivalent. In the context of fractional Brownian motion, since Tanaka formula only exists for fractional Brownian motion and some functionals on it, there is no relation between both equations. Is in this sense and until our knowledge this work corresponds to the first step in this direction.
We organized our paper as follows. In Section 2, we give some preliminares related to fractional calculus. Section 3 is devoted to analyze the problem according to the initial condition and the Scheme approximation. The main results is presented in Section 4. A simulation study is reported in last Section.
\section{Preliminaries}
This section is devoted to introduce an extension of Young's integral defined by
Z\"ahle in \cite{zahle}. Before giving the definition of this integral, we
establish some notations and definitions that we use in this paper.
Consider $0\le a<b\le T$, $\alpha\in(0,1)$, $p>1$ and $f\in L^{p}([0,T])$. For
$t\in[a,b]$ we set
\begin{equation}\label{def:der+}
D_{a+}^{\alpha} f({t})
=L^p-\lim_{\varepsilon\downarrow0}\frac{1}{\Gamma(1-\alpha)}
\left(
\frac{f_{t}}{(t-a)^{\alpha}}
+ \alpha \int_{a}^{t-\varepsilon} \frac{f_{t}-f_{r}}{(t-r)^{1+\alpha}} \, dr
\right),
\end{equation}
in case that this limit is well-defined,
where we use the convention $f_r=0$ on $[a,b]^c$. In this case
$D_{a+}^{\alpha} f$ is called the left-fractional derivative of $f$ of order $\alpha$.
Similarly, for $f\in L^{p}([0,T])$ and $t\in[a,b]$, the right-fractional
derivative of $f$ of order $\alpha$ is introduced as
\begin{equation}\label{def:der-}
D_{b-}^{\alpha} f(t)
=L^p-\lim_{\varepsilon\downarrow0}\frac{1}{\Gamma(1-\alpha)}
\left(
\frac{f_{t}}{(b-t)^{\alpha}}
+ \alpha \int_{t+\varepsilon}^{b} \frac{f_{t}-f_{r}}{(r-t)^{1+\alpha}} \, dr
\right).
\end{equation}
It is not difficult to see that, as a consequence of the proof of
\cite[Theorem 13.2]{Sam}, the fact that $f$, $\frac{f(\cdot)}
{(\cdot-a)^{\alpha}}$ and $\int_a^{\cdot}\frac{f(\cdot)-f_r}{(\cdot-r)^{1+\alpha}}dr$
(resp. $\frac{f(\cdot)}
{(b-\cdot)^{\alpha}}$ and $\int_{\cdot}^b\frac{f(\cdot)-f_r}{(r-\cdot)^{1+\alpha}}dr$)
belong to $L^p([a,b])$ implies that $D_{a+}^{\alpha} f$ (resp.
$D_{b-}^{\alpha} f$) is well-defined and
\begin{equation}\label{eq:def-Da-plus}
D_{a+}^{\alpha} f({t})
=\frac{1}{\Gamma(1-\alpha)}
\left(
\frac{f_{t}}{(t-a)^{\alpha}}
+ \alpha \int_{a}^{t} \frac{f_{t}-f_{r}}{(t-r)^{1+\alpha}} \, dr
\right),
\end{equation}
(resp.
$$D_{b-}^{\alpha} f({t})
=\frac{1}{\Gamma(1-\alpha)}
\left(
\frac{f_{t}}{(b-t)^{\alpha}}
+ \alpha \int_{t}^{b} \frac{f_{t}-f_{r}}{(r-t)^{1+\alpha}} \, dr
\right)).$$
The space of all the $\alpha$-H\"older continuous functions on $[a,b]$ is denoted by
$C^{\alpha}([a,b])$. Then if $f\in C^{\alpha}([a,b]) $, the norm of $f$ is defined as follows
$$\left\|f\right\|_{\alpha, [a,b]}:= \left\|f\right\|_{\infty} + \sup_{a\leq s < t \leq b} \frac{|f(t)-f(s)|}{(t-s)^{\lambda}},$$
where $ \left\|f\right\|_{\infty}= \sup_{a\leq t \leq b} |f(t)|$.
Note that if $f$ belongs to
$C^{\alpha+\varepsilon}([a,b])$, with
$\varepsilon>0$, then (\ref{eq:def-Da-plus}) is true.
Let $g, f\in L^p([0,T])$ be two functions
and
$g_r^{b-}=g_r-g_{b-}$. In this case we say that
$f$ is integrable with respect to $g$ if and only if
$D_{a+}^{\alpha} f$ and $D_{b-}^{1-\alpha} g^{b-}$ exist, and
$(D_{a+}^{\alpha} f)D_{b-}^{1-\alpha} g^{b-}\in L^{1}([a,b])$. In this case we
define the integral $\int_a^b f\, dg$ in the following way
\begin{equation}\label{eq:def-intg-frac}
\int_{a}^{b} f_{r}\, dg_{r}
:=
\int_{a}^{b} (D_{a+}^{\alpha} f)(r) D_{b-}^{1-\alpha}g^{b-}(r) \, dr.
\end{equation}
It is well-known that if $f\in C^{\mu}([a,b])$ and
$g\in C^{\beta}([a,b])$, with $\mu+\beta>1$, then it
can be checked that $\int_{a}^{b} f_{r}\, dg_{r}$ is well-defined, and that it coincides with the Young's integral defined as a limit of Riemann sums.
\section{The Stochastic Differential Equation }
In this section we study the solution of the stochastic differential equation driven by a fBm with a discontinuous diffusion coefficient (\ref{DISC}) as an approximation of a SDE.
The integral in (\ref{DISC}) is defined pathwise as a Young integral.
\subsection{Initial condition }
We divide our problem according to the initial condition $x_0$ in the following three cases:
\begin{enumerate}
\item[i)] $ x_0 > 0$,
\item[ii)] $x_0 < 0$,
\item[iii)] $x_0 = 0$.
\end{enumerate}
Note that if equation (\ref{DISC}) has a continuous solution $x$, then
it satisfies the following:
{\bf Case i)} The continuity of $x$ implies that
there exits $t_0 > 0$ such that $x_t> 0$ on $(0, t_0)$. Hence,
\begin{eqnarray*}
x_t = x_0 + \int_0^t \frac{1}{\alpha} d B_s^H = x_0 + \frac {1}{\alpha} B_t^H \ , \quad t \leq t_0.
\end{eqnarray*}
Let $\tilde{t}_0$ be the first instant such that $x_{\tilde{t}_0} = 0$. Then,
\begin{eqnarray}
\label{eq2}x_t = x_0 + \frac{1}{\alpha} B^H_{\tilde{t}_0} + \int_{\tilde{t}_0}^t \sigma (x_s) d B^H_s= \int_{\tilde{t_0}}^t \sigma (x_s) d B_s^H \ , \quad t \geq \tilde{t}_0.
\end{eqnarray}
{\bf Case ii)} Similarly, we have that, there is $t_1 > 0$ such that $x_t < 0$ on $(0 , t_1)$.
Consequently,
\begin{eqnarray*}
x_t = x_0 + \int_0^t \sigma (x_s) d B^H_s = x_0 + \frac {1}{1 - \alpha} B_t^H \ , \quad t \leq t_1.
\end{eqnarray*}
Again, let $\tilde{t}_1$ be the first time such that $x_{\tilde{t}_1} = 0$. Then,
\begin{eqnarray} \label{eq3}
x_t = \int_{\tilde{t}_1}^t \sigma (x_s) d {B}_s^H \quad , \quad t \geq \tilde{t}_1.
\end{eqnarray}
So, we only need to consider the existence of a unique continuous solution to
equation (\ref{eq3}), for any ${\tilde{t}_1}\ge 0$.
\subsection{ Notation and auxiliary results } In this section, we introduce the sequence of continuous functions $\{\sigma_n : \mathbb{R} \rightarrow \mathbb{R}_+ : n \in \mathbb{N} \}$ converging to $\sigma$ given by:
\begin{eqnarray*}
\sigma_n(x):= \begin{cases} \sigma(x) & \text{if} \ x\notin (-1/n, 0) \\
\frac{1}{\alpha} + n \frac{1 - 2\alpha}{\alpha(1 - \alpha)} x & \text{otherwise}.
\end{cases}
\end{eqnarray*}
Figure 2 shows the approximation function.
\begin{figure}[h]
{\includegraphics[width=150mm] {SIGMAN.eps}}
\caption{$\sigma^n(x)$ for $\alpha =0.4$.}
\label{sigman}
\end{figure}
First, note that, $\sigma_n$ is a Lipschitz continuous function with constant $n \frac{1 - 2\alpha}{\alpha (1 - \alpha)}$, it means,
\begin{eqnarray}\label{eq4}
| \sigma_n (x) - \sigma_n (y) | \leq n \frac{1 - 2\alpha}{\alpha (1 - \alpha)} |x - y| \quad , \quad x,y \in \mathbb{R}.
\end{eqnarray}
Now we define, for $a \in \mathbb{R}$ fixed,
\begin{eqnarray}
\label{deflambdan}
\Lambda_n (x) = \int_a^x \frac{ds}{\sigma_n (s)},\quad x\in\mathbb{R}.
\end{eqnarray}
This integral is well defined due to $\sigma_n>0$. Moreover, $\Lambda_n \in C^1(\mathbb{R})$ and $\Lambda'_n (x) = \frac{1}{\sigma_n (x)} > 0$, for all $x\in\mathbb{R}$.
Therefore $\Lambda_n$ is a strictly increasing function and uniformly bounded in compacts. As a consequence, $\Lambda_n$ has an inverse $\Lambda_n^{-1}$.
Set
\begin{equation}\label{def:L-1}
\Lambda (x) = \int_a^x \frac{ds}{\sigma (s)} , \quad x \in \mathbb{R}.
\end{equation}
\begin{remark}
Since $\Lambda$ is a strictly increasing function, then $\Lambda^{-1}$ exists.
\end{remark}
We will need the following auxiliary result later on.
\begin{lemma}\label{propaprox}
\label{lemma1} There exists a positive constant $C_\alpha$
such that, for $x\in\mathbb{R}$ and $n\in\mathbb{N}$,
$$\left|\Lambda_n(x)-\Lambda(x)\right|\le \frac{C_{\alpha}}{n}\quad
\hbox{\rm and}\quad
\left|\Lambda_n^{-1}(x)-\Lambda^{-1}(x)\right|\le \frac{C_{\alpha}}{n}.$$
\end{lemma}
\begin{proof}
The definitions of $\sigma$ and $\sigma_n$, together with (\ref{deflambdan}) and
(\ref{def:L-1}), imply that
$$\Lambda(x)=\alpha(x-a)1_{\{x\ge 0\}}+((1-\alpha)x-a\alpha)1_{\{x< 0\}}$$
and
\begin{eqnarray*}
\Lambda_n(x)&=&\alpha(x-a)1_{\{x\ge 0\}}\\
&&
+\left(-a\alpha-\frac{\alpha(1-\alpha)}{n(1-2\alpha)}\left(\log(\frac{1}{\alpha})
-\log(\frac{1}{\alpha}+n\frac{1-2\alpha}{\alpha(1-\alpha)}(\frac{-1}{n}\vee x)
\right)\right.\\
&&+\left.(1-\alpha)((\frac{-1}{n}\wedge x)+\frac{1}{n}\right)1_{\{x< 0\}}.
\end{eqnarray*}
Consequently,
\begin{equation}\label{calL-1}
\Lambda^{-1}(x)=\left(\frac{x}{\alpha}+a\right)1_{\{x\ge -a\alpha\}}+
\left(\frac{x+a\alpha}{1-\alpha}\right)1_{\{x\le -a\alpha\}}\end{equation}
and
\begin{eqnarray*}
\Lambda^{-1}_n(x)&=&\left(\frac{x}{\alpha}+a\right)1_{\{x\ge -a\alpha\}}
\\
&&+\frac{1}{\alpha}\left(\exp\left(\frac{n(1-2\alpha)(x+a\alpha)}
{\alpha(1-\alpha)}\right)-1\right)\frac{\alpha(1-\alpha)}{
n(1-2\alpha)}1_{\{\alpha_{n}<x<-a\alpha\}}\\
&&+\left(\frac{x-\alpha_n}{1-\alpha}-\frac{1}{n}\right)
1_{\{x\le \alpha_n\}},
\end{eqnarray*}
with
$$\alpha_n=-a\alpha-\frac{\alpha(1-\alpha)}{n(1-2\alpha)}\left(\log(\frac{1}{\alpha})
-\log(\frac{1}{1-\alpha})\right).$$
Now, it is easy to finish the proof.
\end{proof}
\begin{remark}
\label{lipschitz} $\Lambda^{-1}$ is a Lipschitz function on $\mathbb{R}$.
Indeed,
\begin{eqnarray}
\label{eq7}
| \Lambda_n^{-1} (x) - \Lambda_n^{-1} (y) | &\leq & \sup_{z \in K, n\in \mathbb{N}} | (\Lambda_n^{-1} (z))'| |x - y|\notag\\
&=& \left( \sup_{z \in K, n\in \mathbb{N}} | \sigma_n (\Lambda_n^{-1} (z)) | \right)
|x - y|
\notag\\
&\leq& \left(\frac{1}{\alpha} + \frac{1}{1 - \alpha} \right) |x-y|.
\end{eqnarray}
Thus the claim is a consequence of Lemma \ref{lemma1}.
\end{remark}
We want to see that $x_t = \Lambda^{-1} (B_t^H - B_a^H + \bar{z})$,
where $\bar{z} = \Lambda (0)$,
is a solution of the equation
\begin{eqnarray}
\label{eq8}
x_t = \int_a^t \sigma (x_s) d B_s^H \quad, \quad t \geq a.
\end{eqnarray}
In order to study the uniqueness of the solution to $(\ref{eq8})$, we consider the following auxiliary result.
\begin{lemma}
\label{lemma4}
Let $\tilde{\alpha} > 1 - H$ and $\gamma < H$ such that $\tilde{\alpha} > 1 - \gamma$. Then,
\begin{eqnarray*}
| D_{t-}^{1- \widetilde{\alpha}} \left( B^H \right)^{t-}_s|
&\leq&
C_{\tilde{\alpha}} ||B^H||_{\gamma,[0,T]} (t-s)^{\tilde{\alpha} + \gamma - 1}\\
&\leq&
C_{\tilde{\alpha}} ||B^H||_{\gamma,[0,T]} T^{\tilde{\alpha} + \gamma - 1},
\quad 0 \leq s \leq t \leq T,
\end{eqnarray*}
\end{lemma}
where $(B^H)_s^{t-} = B_s^H - B^H_t$.
\begin{proof}
By definition of $D_{t-}^{1-\alpha}$ (see equalities (\ref{def:der+})
and (\ref{eq:def-Da-plus})), we have
\begin{eqnarray*}
\bigg| D_{t-}^{1- \widetilde{\alpha}} \left( B^H \right)_s^{t-} \bigg| &\leq& \frac{1}{\Gamma (\tilde{\alpha})} \bigg( \frac{|B_s^H - B_t^H|}{(t-s)^{1-\tilde{\alpha}}} + (1-\tilde{\alpha}) \int_s^t \frac {|B_s^H - B_r^H|}{(r - s)^{2-\tilde{\alpha}}} dr \bigg)\\
&\leq& C_{\tilde{\alpha}} \bigg( ||B^H||_{\gamma,[0,T]} |t - s|^{\tilde{\alpha} + \gamma - 1} + ||B^H||_{\gamma,[0,T]} \int_s^t (r - s)^{\tilde{\alpha} + \gamma - 2} dr \bigg)\\
&\leq& C_{\tilde{\alpha}} ||B^H||_{\gamma,[0,T]} |t - s|^{\tilde{\alpha} + \gamma - 1}. \nonumber \\
\end{eqnarray*}
Consequently, the proof is complete.
\end{proof}
\subsection{Existence and uniqueness for equation (\ref{DISC}) }
In this section we state the main results of this article. The first one is related to the existence of a solution of the SDE (\ref{DISC}), and the second one with the uniqueness
of the solution.
\begin{theorem}{\bf Existence:}
\label{theorem1}Let $\gamma \in (\frac{1}{2}, H )$. Then there exists a pathwise solution $x \in C^{\gamma} ([0,T])$ to the equation
\begin{eqnarray}\label{eq18}
x_t = x_0 + \int_0^t \sigma (x_s) d B_s^H , \quad t \in [0,T].
\end{eqnarray}
\end{theorem}
\begin{proof} By (\ref{eq2}) and (\ref{eq3}), we only need to show that the equation
$$x_t=\int_a^t \sigma(x_s)dB^H_s,\quad t\ge a,$$
has a solution for every $a\ge 0$. Towards this end, we
observe that, for $\alpha<1/2$, (\ref{calL-1}) implies that $\Lambda^{-1}$ is a convex
function. Then, Remark 3.5 in \cite{AMV} (see also Theorem 2.1 in \cite{MSV})
yields
$$\Lambda^{-1}(B_t^H-B_a^H+\bar{z})-\Lambda^{-1}(\bar{z})=
\int_a^t(\Lambda^{-1})^{'}_+(B_s^H-B_a^H+\bar{z})dB^H_s,\quad t\ge a.$$
Since $(\Lambda^{-1})^{'}_+(x)=\sigma(x-\bar{z})$, setting
$x_t=\Lambda^{-1}(B_t^H-B_a^H+\bar{z})$, then we have
$$x_t=\int_a^t\sigma(B_s^H-B_a^H)dB^H_s.$$
Finally, we also have that $\sigma(x)=\sigma(\Lambda^{-1}(x
+\bar{z}))$. Thus the proof is complete.
\end{proof}
Now we want to see that ($\ref{eq18}$) has a unique solution $x \in C^\gamma ([0,T])$, for $\gamma \in (\frac{1}{2} , H)$. Towards this end, in the following result we will use the approximation $\sigma_n$, because we will take advantage of \cite{zahle} (Theorem 4.2.1),
which requires the Lipschitz property of $\sigma_n$.
Note that we have figure out a solution of ($\ref{eq18}$) such that
\begin{eqnarray}\label{eq20}
1_{\{a \leq r < s < T\}} \frac{\sigma (x_s) - \sigma (x_r)}{(s-r)^{1+\tilde{\alpha}}} \in L^1([0,T]^2),
\end{eqnarray}
with probability 1.
\begin{lemma}\label{le:lnYi}
Let $\gamma \in (\frac{1}{2} , H )$ and $x$ a $\gamma$-H\"older continuous solution to ($\ref{eq18}$) such that ($\ref{eq20}$) holds. Then,
\[\Lambda_n (x_t) = \Lambda_n (x_0) + \int_0^t \frac{\sigma (x_s)}{\sigma_n (x_s)} d B_s^H, \quad t \in [0,T].\]
\end{lemma}
\begin{proof}
By \cite{zahle} (Theorem 4.2.1) and (\ref{eq4}), we obtain
\[\int_0^t \frac{d x_s}{\sigma_n (x_s)} = \lim\limits_{|\pi| \to 0} \sum_{j=0}^{m-1} \frac{x_{s_{i+1}} - x_{s_i}}{\sigma_n (x_{s_i})},\]
where $\pi$ is a partition of [0,t] of the form $\pi = \{0 = s_0 < s_1<\cdots< s_m = t\}.$
Therefore, with the convention $\sigma_n^{\pi} (s) = \sum_{j=1}^{m-1} \frac{1_{[s_i,s_{i+1}]}(s)}{\sigma_n (x_{s_i})}$, we get
\begin{eqnarray}\label{eq21}
\int_0^t \frac{d x_s}{\sigma_n (x_s)} &=& \lim\limits_{|\pi| \to 0} \sum_{j=0}^{m-1} \int_{s_i}^{s_{i+1}} \frac{\sigma(x_r)}{\sigma_n (x_{s_i})} d B_r^H \notag\\
&=& \lim\limits_{|\pi| \to 0} \int^t_0 \sigma(x_r) \sigma_n^{\pi} (r) dB_r^H.
\end{eqnarray}
On the other hand, the definition of Young integral (\ref{eq:def-intg-frac}) allows to establish
\begin{eqnarray*}\lefteqn{
\left| \int^t_0 \left ( \frac{\sigma(x_s)}{\sigma_n (x_{s})} - \sigma_n^{\pi}(s) \sigma (x_{s}) \right ) dB^H_s \right |} \\
&\le& C_{\tilde{\alpha}} \left | \int^t_0 \sigma (x_{s})( D_{0+}^{\tilde{\alpha}}( \frac{1}{\sigma_n(x_s)} - \sigma_n^{\pi} (s) ) )( D_{t-}^{1 - \tilde{\alpha}} B^H)_{s}^{t-} ds \right| \\
&&+ \tilde{\alpha} \left | \int^t_0 \int^s_0 \left( \frac{1}{\sigma_n(x_r)} - \sigma_n^{\pi} (r) \right ) \frac{\sigma(x_s) - \sigma (x_r)}{(s-r)^{1+
\tilde{\alpha}}} dr ( D_{t-}^{1 - \tilde{\alpha}} B^H)_{s}^{t-} ds \right|\\
&\le& C_{\tilde{\alpha},T} ||B^H||_{\gamma, [0,T] } \bigg ( \int^T_0 \left | D_{0+}^{ \tilde{\alpha}} \left ( \frac{1}{\sigma_n(x_s)} - \sigma_n^{\pi}(s) \right) \right| ds \big. \\
&&+ \int^T_0 \int^s_0 \left | \frac{1}{\sigma_n(x_r)} - \sigma_n^{\pi}(r) \right | \frac{ | \sigma(x_s) - \sigma(x_r) |}{(s-r)^{1+\tilde{\alpha}}} dr ds \bigg ),
\end{eqnarray*}
where last inequality follows from Lemma $\ref{lemma4}$. Then, the result is a consequence of Z\"ahle \cite{zahle} (Theorems 4.1.1 and 4.3.1).
\end{proof}
Note that the solution $x$ to equation ($\ref{eq18}$) is such that, there exists a random variable $G$ such that
\begin{eqnarray*}
| x_s - x_t | \le G |s-t|^\gamma.
\end{eqnarray*}
In the following result we set
\begin{eqnarray*}
f_{\gamma }(s) = \mathbb{E}( G^{\frac{\tilde{\alpha} + \varepsilon}{\gamma}} |x_s|^{-\frac{( \tilde{\alpha} + \varepsilon)}{\gamma}}).
\end{eqnarray*}
\begin{lemma}
Let $\gamma \in (\frac{1}{2}, H)$, $1-H<1-\gamma<\tilde{\alpha}<\gamma$, $x$ a $\gamma$-H\"older continuous solution of (\ref{eq18}) such that (\ref{eq20})
and $\int_0^T1_{\{x_s=0\}}ds=0$
hold, $f_{\gamma} \in L^1 \left( [0,T]\right)$ for some $\varepsilon>0$ small enough,
$G^{1-\eta} \in L^q$ and $\mathbb{P} \left( a < x_s \le b \right) \le g(s) (b - a)$, where $g \in L^{\frac{1}{p}} \left( [0,T] \right)$ with $\frac{1}{p} + \frac{1}{q} = 1$ and
$p\in(1,\frac{1}{1-\eta})$ for some $\eta<1-\frac{{\tilde{\alpha}}}{\gamma}$.
Then,
\begin{eqnarray*}
\Lambda (x_t) - \Lambda (x_0) = B_t^H, \ t\in [0,T].
\end{eqnarray*}
\end{lemma}
\begin{remark}
Note that the solution to (\ref{eq18}) given in Theorem \ref{theorem1} satisfies
the assumptions of this result.
\end{remark}
\begin{proof}
From Lemmas $\ref{lemma4}$, (\ref{lemma1}) and \ref{le:lnYi}, we only need to prove that:
\begin{eqnarray}\label{eq22}
\int^T_0 s^{-\tilde{\alpha}} \left| \frac{\sigma(x_s)}{\sigma_n(x_s)} - 1 \right | ds \longrightarrow 0, \quad \text{as} \ n \rightarrow \infty,
\end{eqnarray}
and
\begin{eqnarray}\label{eq23}
I_{n_k} =\int^T_0 \int^s_0 \left| \frac{\sigma(x_s)}{\sigma_{n_k}(x_s)} - \frac{\sigma(x_r)}{\sigma_{n_k}(x_r)} \right| (s-r)^{-(1+\tilde{\alpha})} dr ds \rightarrow 0, \quad \text{as} \ n_k \longrightarrow \infty,
\end{eqnarray}
for some subsequence$\{n_k:k\in\mathbb{N}\}$, w.p. 1. For ($\ref{eq22}$), we have:
\begin{eqnarray*}
\int^T_0 s^{-\tilde{\alpha}} \left| \frac{ \sigma(x_s) - \sigma_n(x_s)}{\sigma_n(x_s)} \right| ds \leq C_\alpha \int^T_0 s^{-\tilde{\alpha}} | \sigma(x_s) - \sigma_n(x_s) | ds.
\end{eqnarray*}
Thus the dominated convergence theorem implies that ($\ref{eq22}$) holds.
For (\ref{eq23}), we have that
$$\left|\frac{\sigma(x_s)}{\sigma_n(x_s)} - \frac{\sigma(x_r)}{\sigma_n(x_r)} \right |
\rightarrow 0,\quad \hbox{\rm as }\ n\rightarrow\infty,\ \hbox{\rm w.p. 1},$$
and
\begin{eqnarray}\label{eq24}
\left| \frac{\sigma(x_s)}{\sigma_n(x_s)} - \frac{\sigma(x_r)}{\sigma_n(x_r)} \right | &=& \left| \frac{ \sigma(x_s) \sigma_n(x_r) - \sigma(x_r) \sigma_n(x_s)}{\sigma_n(x_s) \sigma_n(x_r)} \right| \nonumber \\
& \le & C_\alpha | \sigma(x_s)\sigma_n(x_r) - \sigma(x_r) \sigma_n(x_s) | \nonumber \\
& \le & C_\alpha | \sigma(x_s) \big | | \sigma_n(x_r) - \sigma_n(x_s) | + C_\alpha | \sigma_n(x_s) \big | | \sigma(x_s) - \sigma(x_r) | \nonumber \\
& \le & C_\alpha \left( | \sigma_n(x_r) - \sigma_n(x_s) \big | + | \sigma(x_s) - \sigma(x_r) | \right) \nonumber \\
& = & I_1 (s,r) + I_2(s,r).
\end{eqnarray}
Let us start analizing the term $I_2(s,r)$. In this case we have the following:
\begin{eqnarray}
\label{I1}\lefteqn{
I_{\{0\le r\le s\le T\}} I_2(s,r) (s-r)^{-(1+\tilde{\alpha})} } \nonumber \\
& \le& C_\alpha I_{\{0\le r\le s\le T\}}
|r-s|^{-1 - \tilde{\alpha}} \left( 1_{\{x_s < 0 < x_r\} } +
1_{\{x_r < 0 < x_s\} } \right).
\end{eqnarray}
Since over $ \{ x_s < 0 < x_r \}$, $|x_s| < |x_r - x_s|$ and $|x_r - x_s|
\le G |r-s|^{\gamma}$.
Also we have
$ |r - s| ^{-1} \le G^{\frac{1}{\gamma}} |x_s|^{-\frac{1}{\gamma}}$.
This inequality leads to:
\begin{equation}
\label{eq:inq1}
I_{\{0\le r\le s\le T\}} I_2(s,r) (s-r)^{-(1+\tilde{\alpha})} I_{\{x_s<0<x_r\}}
\le C_{\alpha}|r-s|^{-1+\varepsilon}G^{\frac{\tilde{\alpha}+\varepsilon}{\gamma}}
|x_s|^{-\frac{\tilde{\alpha}+\varepsilon}{\gamma}}.
\end{equation}
Similarly we can obtain,
\begin{equation}
\label{eq:inq2}
I_{\{0\le r\le s\le T\}} I_2(s,r) (s-r)^{-(1+\tilde{\alpha})} I_{\{x_r<0<x_s\}}
\le C_{\alpha}|r-s|^{-1+\varepsilon}G^{\frac{\tilde{\alpha}+\varepsilon}{\gamma}}
|x_r|^{-\frac{\tilde{\alpha}+\varepsilon}{\gamma}}.
\end{equation}
Now, we show that
\begin{equation}\label{eq:mG}
I_1(s,r)\le C_{\alpha} n^{1-\eta}G^{1-\eta}(s-r)^{\gamma(1-\eta)}
\end{equation}
holds. By
inequality (\ref{eq4}), we can establish
\begin{eqnarray*}\lefteqn{
\big| \sigma_n ( x_s) - \sigma_n(x_r) \big|} \\
&=& \big| \sigma_n ( x_s) - \sigma_n(x_r) \big|^{1-\eta}\big| \sigma_n ( x_s) - \sigma_n(x_r) \big|^\eta \\
&=& n^{1-\eta} \left |\frac{1-2\alpha}{(1-\alpha)\alpha} \right|^{1-\eta} \left|x_s -
x_r
\right |^{1-\eta}
\big| \sigma_n ( x_s) - \sigma_n(x_r) \big|^\eta \\
&\leq& C_{\alpha, \eta} n^{1-\eta} \left |x_s-x_r\right|^{1-\eta},
\end{eqnarray*}
which allows us to see that (\ref{eq:mG}) is true.
On the other hand,
\begin{eqnarray*}\lefteqn{
\mathbb{E}\left(I_{\{x_r<-\frac{1}{n}<x_s<0\}}G^{1-\eta}\right)}\\
&\le&\left(\mathbb{E}G^{q(1-\eta)}\right)^{1/q}\left(\mathbb{P}\left(x_s\in(-\frac{1}{n},0)
\right)\right)^{1/p}
\le Cn^{-\frac{1}{p}}g_s^{\frac{1}{p}}.
\end{eqnarray*}
Hence, (\ref{eq:mG}) implies
$$\mathbb{E}\left(\int_0^T\int_0^sI_1(s,r)(s-r)^{-(\tilde{\alpha}+1)}
I_{\{x_r<-\frac{1}{n}<x_s<0\}}drds\right)\rightarrow 0,$$
as $n\rightarrow\infty$. Moreover,
this analysis also implies
$$\mathbb{E}\left(\int_0^T\int_0^sI_1(s,r)(s-r)^{-(\tilde{\alpha}+1)}
\left(I_{\{x_s<-\frac{1}{n}<x_r<0\}}+
I_{\{-\frac{1}{n}<x_r,x_s<0\}}\right)drds\right)\rightarrow 0,$$
as $n\rightarrow\infty$.
Finally, we can see that inequalities (\ref{eq:inq1}) and (\ref{eq:inq2}) are also true
when we write $I_1(s,r)$ instead of $I_2(s,r)$. Therefore, the result follows
from (\ref{eq24})-(\ref{eq:inq2}) and the dominated convergence theorem.
\end{proof}
\section{Numerical Results}
By (\ref{eq4}) and the mean value theorem we have that $\sigma_n (\Lambda^{-1}_n(x))$ to be a Lipschitz function, then Z\"ahle (\cite{zahle}, theorem 4.3.1) implies:
\begin{eqnarray} \label{eq5}\lefteqn{
\Lambda_n^{-1}(B_t^H - B_a^H + z_n) - \Lambda_n^{-1} (z_n)}\notag \\
&=& \Lambda_n^{-1} (B_t^H - B_a^H + z_n) \notag\\
&=& \int_a^t \sigma_n \left(\Lambda_n^{-1} (B_s^H - B_a^H + z_n) \right) d B_s^H,
\quad t\ge a,
\end{eqnarray}
with $z_n= \Lambda_n (0)$.
Therefore, $x_t^{(n)} = \Lambda_n^{-1} (B_t^H - B_a^H + z_n) $ is a solution to the equation
\begin{eqnarray}\label{eq6}
x_t^{(n)} = \int_a^t \sigma_n (x_s^{(n)} ) d B_s^H , \quad t \geq a.
\end{eqnarray}
Consequently, by Lemma \ref{propaprox}, we can aproximate the solution of equation (\ref{DISC}) by the solution of equation (\ref{eq6}).
In this section we show some simulations for the unique solution of equation (\ref{DISC}) for different values of $H$ and $\alpha$. We use the approximation given in (\ref{eq6}). We also present the limit case $H=0.5$, and $a=0$.
The following figures correspond to the particular case the Skew Brownian motion (SBm) as the solution of a SDE with discontinuous coefficient in the diffusion. Also, the case (c), $\alpha =0.5$ corresponds to the Brownian motion process.
\begin{figure}[htbp]
\centering
\subfigure[SBm for $\alpha = 0.99$. ]{\includegraphics[width=45mm] {SBM099.eps}}
\subfigure[SBm for $\alpha = 0.01$.]{\includegraphics[width=45mm]{SBM001.eps}}
\subfigure[SBm for $\alpha = 0.5$.]{\includegraphics[width=45mm]{SBM05.eps}}
\label{SBm}
\end{figure}
The following pictures shows samples for the unique solution of (\ref{DISC}), for different values of $H$ and $\alpha$.
\begin{figure}[htbp]
\centering
\subfigure[$x_t$ for $H=0.75$ and $\alpha = 0.1$]{\includegraphics[width=45mm] {H075a01.eps}}
\subfigure[$x_t$ for $H=0.75$ and $\alpha = 0.5$]{\includegraphics[width=45mm]{H075a05.eps}}
\subfigure[$x_t$ for $H=0.75$ and $\alpha = 0.99$]{\includegraphics[width=45mm]{H075a099.eps}}
\label{}
\end{figure}
\begin{figure}[htbp]
\centering
\subfigure[$x_t$ for $H=0.95$ and $\alpha = 0.1$]{\includegraphics[width=45mm] {H095a001.eps}}
\subfigure[$x_t$ for $H=0.95$ and $\alpha = 0.5$]{\includegraphics[width=45mm]{H095a05.eps}}
\subfigure[$x_t$ for $H=0.95$ and $\alpha = 0.99$]{\includegraphics[width=45mm]{H095a099.eps}}
\label{}
\end{figure}
\newpage
{\bf Acknowledgments}
J. Garz\'on was partially supported by the Project HERMES 16930.
Jorge A. Le\'on was partially supported by
CONACyT grant 220303. S. Torres was partially supported by the Project PIA ACT1112 C and Fondecyt Grant 1130586. This
work was also developed as part of Inria Chile-CIRICproject Communication and information research \& innovation center, 10CEII-9157.
\bibliographystyle{amsplain}
|
\section{Introduction\label{sec:Introduction}}
\
About three years ago, Cachazo, He and Yuan (CHY) proposed a compact
formula for computing tree-level amplitudes in both Yang-Mills and
gravity theories \cite{Cachazo:2013hca}. There was an increasing
interest then to find a string origin of those results given their
known connection to string amplitudes at the low-energy limit.
Soon after that work, Mason and Skinner introduced the so-called ambitwistor
string \cite{Mason:2013sva}, which could be viewed as an $\alpha'\to0$
limit of the usual string and provided a clear derivation of the CHY
formulae for $D=10$ Yang-Mills and NS-NS supergravity.
Taking advantage of the pure spinor formalism's manifest supersymmetry,
Berkovits proposed its ambitwistor version in \cite{Berkovits:2013xba},
which was explicitly shown in \cite{Gomez:2013wza} to provide the
supersymmetric version of the CHY amplitudes.
When extended to curved backgrounds, one would expect that consistency
of the ambitwistor string should put the target
space fields on-shell. In \cite{Adamo:2014wea}, Adamo \emph{et al} demonstrated
that the nonlinear equations of motion of the NS-NS background arise
as anomalies of the worldsheet supersymmetry algebra. In the pure
spinor case, Chandia and Vallilo investigated the type II background
\cite{Chandia-Vallilo} and realized that Berkovits' original proposal
for the infinite tension string was incomplete and had to be modified
in order to obtain the usual background constraints coming from
the pure spinor formalism. By performing a semi-classical analysis,
they were able to reproduce the known results of \cite{Berkovits:2001ue}
with the introduction of the extra condition of BRST-closedness of
$\mathcal{H}$, a generalized particle-like Hamiltonian.
The ideas in \cite{Chandia-Vallilo} were further explored by one
of the authors in \cite{Jusinskas:2016qjd} and it was shown that
the new model, although still chiral, could be interpreted in terms
of two sectors resembling the usual left and right-movers of the superstring.
This construction was also extended to the heterotic case, providing
a sensible description of the massless heterotic spectrum in this
$\alpha'\to0$ limit. This was achieved by incorporating the observed
sectorization in the heterotic BRST charge, which was then redefined
to be
\begin{equation}
Q=\oint\{\lambda^{\alpha}d_{\alpha}+\bar{c}T_{+}-\bar{b}\bar{c}\partial\bar{c}\},\label{eq:BRSTintro}
\end{equation}
where $\lambda^{\alpha}$ is the pure spinor ghost, $d_{\alpha}$
is the improved worldsheet realization of the superderivative introduced
in \cite{Chandia-Vallilo}, ($\bar{b},\bar{c}$) are the
reparametrization ghosts and $T_{+}$ accounts for one of the sectorized
energy-momentum-like tensors, which are defined in terms of $\mathcal{H}$
and the full energy-momentum tensor $T$ as
\begin{equation}
T_{\pm}\equiv\frac{1}{2}(T\pm\mathcal{H}).
\end{equation}
As we show in the present work, the problem of finding the constraints
on the heterotic background is somewhat more natural than in type
II, in that $\mathcal{H}$ enters the BRST charge $Q$ itself, \emph{cf}.
\eqref{eq:BRSTintro}, and the background constraints all come from
the sole requirement that $Q$ be nilpotent. In a general heterotic
background, the action for the sectorized model and the generalized
particle-like Hamiltonian will be cast as
\begin{eqnarray}
S & = & \frac{1}{2\pi}\int d^{2}z\{\mathcal{P}_{a}\bar{\Pi}^{a}+d_{\alpha}\bar{\Pi}^{\alpha}-\Pi^{A}\bar{\Pi}^{B}B_{BA}+\bar{\Pi}^{A}A_{A}^{I}J_{I}+w_{\alpha}\bar{\nabla}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C},\label{eq:actionintro}\\
\mathcal{H} & = & -\frac{1}{2}\mathcal{P}_{a}\mathcal{P}^{a}-\frac{1}{2}\Pi^{a}\Pi_{a}+d_{\alpha}\Pi^{\alpha}+w_{\alpha}\nabla\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}\nonumber \\
& & -\Pi^{A}A_{A}^{I}J_{I}-d_{\alpha}W^{\alpha I}J_{I}-\lambda^{\alpha}w_{\beta}U_{\alpha}^{\hphantom{\alpha}\beta I}J_{I}.
\end{eqnarray}
The vielbein appears through $\Pi^{A}=\partial Z^{M}E_{M}^{\hphantom{M}A}$,
mapping the curved superspace coordinates $Z^{M}$, to the generalized
superspace invariants with flat (super) indices $A$. The Lorentz
connection $\Omega_{AB}^{\hphantom{AB}C}$, enters the covariant derivative
$\nabla$. The super Kalb-Ramond field is denoted by $B_{AB}$, while
$A_{A}^{I}$, $W^{\alpha I}$ and $U_{\alpha}^{\hphantom{\alpha}\beta I}$
represent the super Yang-Mills background. All the worldsheet fields
above will be detailedly introduced in section \ref{sec:heterotic}.
By performing a classical analysis and computing the generalized Poisson
brackets associated to $S$, we will show that classical nilpotency
of the BRST charge \eqref{eq:BRSTintro} imposes some constraints
on the torsion $T_{AB}^{\hphantom{AB}C}$, the 3-form field strength
$H_{ABC}$, the curvature tensor $R_{ABC}^{\hphantom{ABC}D}$, and
the super Yang-Mills field strength $F_{AB}^{I}$, given by\begin{subequations}\label{eq:hetconstraints}
\begin{equation}
\lambda^{\alpha}\lambda^{\beta}T_{\alpha\beta}^{\hphantom{\alpha\beta}A}=\lambda^{\alpha}\lambda^{\beta}H_{A\alpha\beta}=\lambda^{\alpha}\lambda^{\beta}\lambda^{\gamma}R_{\alpha\beta\gamma}^{\hphantom{\alpha\beta\gamma}\delta}=\lambda^{\alpha}\lambda^{\beta}F_{\alpha\beta}^{I}=0,
\end{equation}
in addition to the so-called holomorphicity constraints\footnote{This name can be misleading here, as the infinite tension limit is
described by the chiral action \eqref{eq:actionintro} and holomorphicity
of the BRST current is trivial.}
\begin{equation}
T_{\alpha a}^{\hphantom{\alpha a}\beta}=T_{\alpha(ab)}=T_{\alpha\beta b}-H_{\alpha\beta b}=H_{ab\alpha}=\lambda^{\alpha}\lambda^{\beta}R_{\alpha a\beta}^{\hphantom{\alpha a\beta}\gamma}=0,
\end{equation}
and
\begin{eqnarray}
F_{\alpha a}^{I} & = & T_{\alpha\beta a}W^{\beta I},\\
\nabla_{\alpha}W^{\beta I}-T_{\alpha\gamma}^{\hphantom{\alpha\gamma}\beta}W^{\gamma I} & = & U_{\alpha}^{\hphantom{\alpha}\beta I},\\
F_{\alpha\beta}^{I} & = & \frac{1}{2}W^{\gamma I}H_{\alpha\beta\gamma},\\
\lambda^{\alpha}\lambda^{\beta}\nabla_{\alpha}U_{\beta}^{\hphantom{\beta}\gamma I} & = & -\lambda^{\alpha}\lambda^{\beta}R_{\delta\alpha\beta}^{\hphantom{\delta\alpha\beta}\gamma}W^{\delta I}.
\end{eqnarray}
\end{subequations}All together, the constraints in \eqref{eq:hetconstraints}
imply the supergravity and super Yang-Mills equations of motion of
the heterotic background, as explained in \cite{Berkovits:2001ue}.
This work is organized as follows. Section \ref{sec:heterotic} presents
the sectorized model introduced in \cite{Jusinskas:2016qjd} for the
heterotic infinite tension string. Starting with a brief review of
Berkovits' original proposal, we will show how the BRST charge was
modified to make the sector description manifest and determine the
classical conditions for its nilpotency. In section \ref{sec:background},
we will discuss the coupling to the heterotic background. For pedagogical
reasons, we will analyze first the pure supergravity coupling and
extend the results including super Yang-Mills next, explaining in
detail how the known background constraints are obtained in the classical
analysis. Section \ref{sec:discussion} discusses the particularities
of the sectorized approach and presents some future directions to
follow. The reader is advised to go through the appendix \ref{sec:notation}
first, as the superspace conventions used here are compactly listed
there. Appendix \ref{sec:currents} contains perhaps the simplest
worldsheet model for the gauge sector with $SO(32)$ group and provides
some of the ingredients used in the main body of the text.
\section{The free heterotic string with infinite tension\label{sec:heterotic}}
\
The heterotic pure spinor string is described in the $\alpha'\to0$
limit by the chiral action
\begin{equation}
S=\frac{1}{2\pi}\int d^{2}z\{P_{a}\bar{\partial}X^{a}+p_{\alpha}\bar{\partial}\theta^{\alpha}+w_{\alpha}\bar{\partial}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C}.\label{eq:heteroticaction}
\end{equation}
$X^{a}$ and $\theta^{\alpha}$ are the $\mathcal{N}=1$ superspace
coordinates with conjugate momenta $P_{a}$ and $p_{\alpha}$, with
$a=0,\ldots,9$ and $\alpha=1,\ldots,16$ denoting the flat vector
and spinor indices respectively. The ghost sector is represented by
the usual reparametrization ghosts, $\bar{b}$ and $\bar{c}$,
the pure spinor $\lambda^{\alpha}$, satisfying $(\lambda\gamma^{a}\lambda)=0$,
and its conjugate $w_{\alpha}$. The gamma matrices satisfy $\{\gamma^{a},\gamma^{b}\}=2\eta^{ab}$,
where $\eta^{ab}$ is the $SO(1,9)$ metric. The gauge sector is encoded
in $S_{C}$. Note that $S$ has no conformal anomaly and its energy-momentum
tensor is given by
\begin{equation}
T=-P_{a}\partial X^{a}-p_{\alpha}\partial\theta^{\alpha}-w_{\alpha}\partial\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C},
\end{equation}
where $T_{C}$ is the gauge sector energy-momentum tensor with central
charge $c=16$.
In \cite{Berkovits:2013xba}, the action \eqref{eq:heteroticaction}
was provided with the BRST charge
\begin{equation}
Q=\oint\{\lambda^{\alpha}[p_{\alpha}-\frac{1}{2}(\gamma^{a}\theta)_{\alpha}P_{a}]+\bar{c}T-\bar{b}\bar{c}\partial\bar{c}\}.\label{eq:BRSTnathan}
\end{equation}
However, it does not correctly describe the expected massless heterotic
spectrum, in particular it fails to reproduce the gauge transformations
of the supergravity states, which are directly related to the invariance
of the theory under general coordinate transformations.
Following the ideas of \cite{Chandia-Vallilo}, an alternative BRST
charge was proposed in \cite{Jusinskas:2016qjd} by one of the authors.
We will review this construction now.
\subsection{Review: sectorization and BRST cohomology}
\
Perhaps the first observation hinting at the inadequacy of the BRST
charge \eqref{eq:BRSTnathan} is the existence of an extra nilpotent
symmetry of the action \eqref{eq:heteroticaction}, also linear in
$\lambda^{\alpha}$, generated by
\begin{equation}
\mathcal{K}=\oint\,(\lambda\gamma_{a}\theta)[\partial X^{a}+\frac{1}{2}(\theta\gamma^{a}\partial\theta)].
\end{equation}
To consistently absorb $\mathcal{K}$ in the BRST charge, the supersymmetry
charges have to be redefined to
\begin{equation}
q_{\alpha}\equiv\oint\{p_{\alpha}+\frac{1}{2}(P_{a}-\partial X_{a})(\gamma^{a}\theta)_{\alpha}-\frac{1}{12}(\theta\gamma_{a}\partial\theta)(\gamma^{a}\theta)_{\alpha}\},
\end{equation}
which in turn brings forth the new invariants:\begin{subequations}\label{eq:flatsusyinvariants}
\begin{eqnarray}
\Pi^{a} & = & \partial X^{a}+\frac{1}{2}(\theta\gamma^{a}\partial\theta),\\
\mathcal{P}_{a} & \equiv & P_{a}-\frac{1}{2}(\theta\gamma_{a}\partial\theta),\\
d_{\alpha} & \equiv & p_{\alpha}-\frac{1}{2}P_{a}(\gamma^{a}\theta)_{\alpha}+\frac{1}{2}\Pi^{a}(\gamma_{a}\theta)_{\alpha}.
\end{eqnarray}
\end{subequations}Note that the operators $P_{a}^{\pm}$ of \cite{Jusinskas:2016qjd}
would be written here as $P_{a}^{\pm}=\mathcal{P}_{a}\pm\Pi_{a}$.
The action and its energy-momentum tensor can be expressed in terms
of the above invariants as
\begin{eqnarray}
S & = & \frac{1}{2\pi}\int d^{2}z\{\mathcal{P}_{a}\bar{\Pi}^{a}+d_{\alpha}\bar{\partial}\theta^{\alpha}+w_{\alpha}\bar{\partial}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C}\nonumber \\
& & -\frac{1}{4\pi}\int d^{2}z\{\Pi^{a}(\theta\gamma_{a}\bar{\partial}\theta)-\bar{\Pi}^{a}(\theta\gamma_{a}\partial\theta)\},\label{eq:hetaction}\\
T & = & -\mathcal{P}_{a}\Pi^{a}-d_{\alpha}\partial\theta^{\alpha}-w_{\alpha}\partial\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}.\label{eq:heteroticT}
\end{eqnarray}
Although not manifestly, $S$ is invariant under supersymmetry. Consider
a transformation with constant parameter $\xi^{\alpha}$, then
\begin{eqnarray}
\delta S & = & \frac{1}{4\pi}\int d^{2}z\{\bar{\Pi}^{a}(\xi\gamma_{a}\partial\theta)-\Pi^{a}(\xi\gamma_{a}\bar{\partial}\theta)\}.\nonumber \\
& = & \frac{1}{4\pi}\int d^{2}z\{(\xi\gamma_{a}\theta)[\bar{\partial}\Pi^{a}-\partial\bar{\Pi}^{a}]\}\nonumber \\
& = & \frac{1}{2\pi}\int d^{2}z\{(\xi\gamma_{a}\theta)(\bar{\partial}\theta\gamma^{a}\partial\theta)\},
\end{eqnarray}
Using the property $(\gamma_{\alpha\beta}^{a}\gamma_{\gamma\lambda}^{b}+\gamma_{\alpha\gamma}^{a}\gamma_{\beta\lambda}^{b}+\gamma_{\alpha\lambda}^{a}\gamma_{\gamma\beta}^{b})\eta_{ab}=0$,
the integrand in the last line can be rewritten as
\begin{equation}
(\xi\gamma_{a}\theta)(\bar{\partial}\theta\gamma^{a}\partial\theta)=\frac{1}{3}\bar{\partial}[(\xi\gamma_{a}\theta)(\theta\gamma^{a}\partial\theta)]-\frac{1}{3}\partial[(\xi\gamma_{a}\theta)(\theta\gamma^{a}\bar{\partial}\theta)],
\end{equation}
which proves the invariance of the action $S$ up to boundary terms.
We will also define the operator
\begin{equation}
\mathcal{H}\equiv-\frac{1}{2}\mathcal{P}_{a}\mathcal{P}^{a}-\frac{1}{2}\Pi_{a}\Pi^{a}+d_{\alpha}\partial\theta^{\alpha}+w_{\alpha}\partial\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C},\label{eq:heteroticH}
\end{equation}
which is the heterotic analogous of the generalized particle-like
Hamiltonian for the type II case of \cite{Chandia-Vallilo}. Using
these operators, it was shown in \cite{Jusinskas:2016qjd} that the
chiral action $S$ can be interpreted in terms of two sectors $(+)$
and $(-)$ with characteristic energy-momentum-like tensors
\begin{equation}
T_{\pm}\equiv\frac{1}{2}(T\pm\mathcal{H}),\label{eq:T+-}
\end{equation}
such that\begin{subequations}\label{eq:T+T-het}
\begin{eqnarray}
T_{+} & = & -\frac{1}{4}\eta^{ab}(\mathcal{P}_{a}+\Pi_{a})(\mathcal{P}_{b}+\Pi_{b})-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C},\\
T_{-} & = & \frac{1}{4}\eta^{ab}(\mathcal{P}_{a}-\Pi_{a})(\mathcal{P}_{b}-\Pi_{b})-d_{\alpha}\partial\theta^{\alpha}-w_{\alpha}\partial\lambda^{\alpha}.
\end{eqnarray}
\end{subequations}
The new BRST charge makes the sectorization of the theory explicit
and is given by
\begin{equation}
Q=Q_{\lambda}+Q_{+},\label{eq:heteroticBRST}
\end{equation}
with\begin{subequations}\label{eq:partshetBRST}
\begin{eqnarray}
Q_{\lambda} & \equiv & \oint\,\lambda^{\alpha}d_{\alpha},\\
Q_{+} & \equiv & \oint\{\bar{c}T_{+}-\bar{b}\bar{c}\partial\bar{c}\}.
\end{eqnarray}
\end{subequations}$Q_{\lambda}$ is very similar to the usual (left-moving)
pure spinor BRST charge while $Q_{+}$ is composed by the familiar
BRST charge coming from the reparametrization symmetry plus an analogous
contribution with the operator $\mathcal{H}$, \emph{cf}. equation
\eqref{eq:T+-}.
The massless spectrum of the heterotic string consists of non-abelian
super Yang-Mills and $\mathcal{N}=1$ supergravity, respectively described
by the vertex operators\begin{subequations}\label{eq:heteroticvertices}
\begin{eqnarray}
U_{SYM} & = & \lambda^{\alpha}\bar{c}A_{\alpha}^{I}J_{I},\\
U_{SG} & = & \lambda^{\alpha}\bar{c}A_{\alpha}^{a}(\mathcal{P}_{a}+\Pi_{a}),
\end{eqnarray}
\end{subequations}where $J_{I}$ corresponds to (holomorphic) generators
of the $SO(32)$ or $E(8)\times E(8)$ current algebra, with $I$
denoting the adjoint representation of the gauge group. BRST-closedness
of $U_{SYM}$ and $U_{SG}$ with respect to \eqref{eq:heteroticBRST}
provides the known superfield equations of motion at the linearized
level,\begin{subequations}
\begin{eqnarray}
\gamma_{abcde}^{\alpha\beta}D_{\alpha}A_{\beta}^{I} & = & 0,\\
\gamma_{abcde}^{\alpha\beta}D_{\beta}A_{\alpha}^{f} & = & 0,\\
\partial^{b}\partial_{b}A_{\alpha}^{a}-\partial^{a}\partial_{b}A_{\alpha}^{b} & = & 0.
\end{eqnarray}
\end{subequations}The gauge transformations of the superfields, given
by\begin{subequations}
\begin{eqnarray}
\delta_{\Sigma}A_{\alpha}^{I} & = & D_{\alpha}\Sigma^{I},\\
\delta_{\Sigma}A_{\alpha}^{a} & = & D_{\alpha}\Sigma^{a}+\partial^{a}\Sigma_{\alpha},
\end{eqnarray}
\end{subequations}can be written in terms of BRST-exact expressions,
as expected. More details can be found in \cite{Jusinskas:2016qjd}.
Next, we will discuss the classical equations associated to the nilpotency
of the BRST-charge \eqref{eq:heteroticBRST} to establish the basis
for the curved background analysis of section \ref{sec:background}.
\subsection{Classical analysis\label{sub:Classical-analysis}}
\
In order to determine the classical conditions to be imposed on the
background, it might be useful to understand their meaning in the
flat case. Recall that the heterotic action can be cast as
\begin{eqnarray}
S & = & \frac{1}{2\pi}\int d^{2}z\{\mathcal{P}_{a}\bar{\Pi}^{a}+d_{\alpha}\bar{\partial}\theta^{\alpha}+w_{\alpha}\bar{\partial}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C}\nonumber \\
& & -\frac{1}{4\pi}\int d^{2}z\{\Pi^{a}(\theta\gamma_{a}\bar{\partial}\theta)-\bar{\Pi}^{a}(\theta\gamma_{a}\partial\theta)\},
\end{eqnarray}
with $\mathcal{P}_{a}$ and $d_{\alpha}$ being supersymmetric invariants
defined in terms of the conjugate momenta of $X^{a}$ and $\theta^{\alpha}$
respectively, \emph{cf}. equation \eqref{eq:flatsusyinvariants}.
It is convenient, however, to treat them as independent variables.
The above action is just one step behind the curved space one that
we will define in the next section.
The BRST symmetry is described by the charge displayed in \eqref{eq:heteroticBRST}.
To compute the classical BRST transformations of the worldsheet variables,
we will rewrite $Q$ in terms of the fields $\{X^{a},\theta^{\alpha},\lambda^{\alpha},\bar{c}\}$,
collectively denoted by $\phi$, and their canonical conjugates, which
are given in terms of $\{\mathcal{P}_{a},d_{\alpha},w_{\alpha},\bar{b}\}$.
The latter will be denoted by $\hat{P}_{\phi}$ and are usually defined
with respect to $\tau$, the worldsheet time. We will use the Minkowski
parametrization with $z=\sigma-\tau$ and $\bar{z}=\sigma+\tau$,
where $\sigma\in[0,2\pi)$ denotes the spatial coordinate. The derivatives
can then be cast as
\begin{equation}
\begin{array}{cc}
\partial=\frac{1}{2}(\partial_{\sigma}-\partial_{\tau}), & \bar{\partial}=\frac{1}{2}(\partial_{\sigma}+\partial_{\tau}).\end{array}
\end{equation}
With this convention, the canonical momenta will be defined to be
\begin{equation}
\hat{P}[\phi]\equiv2\pi\Bigg(\frac{\delta S}{\delta(\bar{\partial}\phi)}-\frac{\delta S}{\delta(\partial\phi)}\Bigg),\label{eq:defcanonical}
\end{equation}
leading to the following identifications:\begin{subequations}
\begin{eqnarray}
\hat{P}[X^{a}] & = & \mathcal{P}_{a}+\frac{1}{2}(\theta\gamma_{a}\partial_{\sigma}\theta),\nonumber \\
& \equiv & \hat{P}_{a}\label{eq:conjugateX}\\
\hat{P}[\theta^{\alpha}] & = & -d_{\alpha}-\frac{1}{2}(\mathcal{P}_{a}-\partial_{\sigma}X_{a})(\gamma^{a}\theta)_{\alpha},\nonumber \\
& \equiv & \hat{P}_{\alpha}\label{eq:conjugatetheta}\\
\hat{P}[\lambda^{\alpha}] & = & w_{\alpha},\\
\hat{P}[\bar{c}] & = & -\bar{b}.
\end{eqnarray}
\end{subequations}
The fundamental Poisson brackets are simply given by
\begin{equation}
\left\{ \hat{P}[\phi'(\sigma')],\phi(\sigma)\right\} _{P.B.}=-\delta_{\phi,\phi'}\delta(\sigma-\sigma').
\end{equation}
Therefore, the BRST transformations of the worldsheet fields are easily
computed when the BRST charge is written in terms of $\hat{P}$ and
$\phi$. For example, $Q_{\lambda}$ in \eqref{eq:partshetBRST} is
expressed as
\begin{equation}
Q_{\lambda}=\oint d\sigma\lambda^{\alpha}[-\hat{P}_{\alpha}-\frac{1}{2}\hat{P}_{a}(\gamma^{a}\theta)_{\alpha}+\frac{1}{2}\Pi_{\sigma}^{a}(\gamma_{a}\theta)_{\alpha}].
\end{equation}
Concerning the nilpotency of the BRST charge $Q$, it can be stated
as
\begin{equation}
Q_{\lambda}^{2}+\{Q_{\lambda},Q_{+}\}+Q_{+}^{2}=0.
\end{equation}
Because $Q_{\lambda}$ is independent of the reparametrization ghosts,
each term in the equation above should vanish separately.
Therefore, following the classical construction just presented, it is easy
to demonstrate that $Q$ is nilpotent if and only if\begin{subequations}\label{eq:nilpotency-conditions}
\begin{eqnarray}
Q_{\lambda}^{2} & = & 0,\label{eq:nilpotencylambda2}\\
\{\lambda^{\alpha}d_{\alpha}(\sigma'),T_{+}(\sigma)\}_{P.B.} & = & 0,\label{eq:nilpotencyTlambda}\\
\{T_{+}(\sigma'),T_{+}(\sigma)\}_{P.B.} & = & 2T_{+}\,\partial_{\sigma}\delta(\sigma'-\sigma)+\partial_{\sigma}T_{+}\,\delta(\sigma'-\sigma).\label{eq:nilpotencyTT}
\end{eqnarray}
\end{subequations}
In flat space, it is straightforward to see that all these relations
are satisfied. In the next section they will be our guidelines for
nontrivial backgrounds. The difference then will be how the background
manifests itself in the definition of the conjugate momenta, in particular
\eqref{eq:conjugateX} and \eqref{eq:conjugatetheta}, which contain
the fundamental ingredients of the BRST charge, $\mathcal{P}_{a}$
and $d_{\alpha}$.
\section{Classical consistency of the heterotic background\label{sec:background}}
\
In this section we will show how the nilpotency conditions discussed
above ultimately impose constraints on the heterotic background, providing
the expected supergravity and super Yang-Mills equations of motion
in superspace detailedly presented in \cite{Berkovits:2001ue} for
the pure spinor superstring.
After understanding how the infinite tension string couples to the
heterotic background, we will be able to build the operator set necessary
for our analysis. The supergravity sector is presented alone beforehand
for two reasons. First, to the best of our knowledge, there is no
good description for $\mathcal{N}=1$ (heterotic) supergravity in
any ambitwistor string so far. So this will be a good test for the
modifications discussed in \cite{Jusinskas:2016qjd} for the sectorized
string. Second, the generalization from flat space is straightforward
and it will help establish the curved superspace language that is
extensively used. Next, we will turn on the super Yang-Mills background
and extend the results.
\subsection{Supergravity background and constraints}
\
The curved superspace generalization of \eqref{eq:hetaction} is given
by
\begin{equation}
S=\frac{1}{2\pi}\int d^{2}z\{\mathcal{P}_{a}\bar{\Pi}^{a}+d_{\alpha}\bar{\Pi}^{\alpha}-\Pi^{A}\bar{\Pi}^{B}B_{BA}+w_{\alpha}\bar{\nabla}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C}.\label{eq:hetaction_curved}
\end{equation}
The vielbein $E_{M}^{\hphantom{M}A}$, and the Lorentz connection
$\Omega_{AB}^{\hphantom{AB}C}$, enter the action through the generalized
superspace invariants and the covariant derivative\footnote{Due to the pure spinor constraint the action has a gauge symmetry
with parameter $\varphi_{a}$ of the form $\delta_{\varphi}w_{\alpha}=\varphi_{a}(\gamma^{a}\lambda)_{\alpha}$.
Therefore, to work only with gauge invariant quantities we must impose
$\lambda^{\alpha}(\gamma^{a}\lambda)_{\beta}\Omega_{A \alpha}^{\hphantom{A \alpha}\beta}=0$.}, given by\begin{subequations}
\begin{eqnarray}
\bar{\Pi}^{A} & = & \bar{\partial}Z^{M}E_{M}^{\hphantom{M}A},\\
\bar{\nabla}\lambda^{\alpha} & = & \bar{\partial}\lambda^{\alpha}+\lambda^{\beta}\bar{\Pi}^{A}\Omega_{A\beta}^{\hphantom{A\beta}\alpha},
\end{eqnarray}
\end{subequations}and analogous expressions for $\Pi^{A}$ and $\nabla\lambda^{\alpha}$,
where $Z^{M}$ denotes the curved $\mathcal{N}=1$ superspace coordinates
$X^{m}$ and $\theta^{\mu}$. The curved vector and spinor indices
are being respectively denoted by $m=0,\ldots,9$ and $\mu=1,\ldots,16$.
Notice that in this language the supermetric $G_{MN}$ is written
in terms of the flat metric as $G_{MN}=E_{M}^{\hphantom{M}a}E_{N}^{\hphantom{N}b}\eta_{ab}$.
The coupling with the Kalb-Ramond superfield can be easily written
with explicit curved space indices,
\begin{eqnarray}
S_{B} & = & -\frac{1}{2\pi}\int d^{2}z\{\Pi^{A}\bar{\Pi}^{B}B_{BA}\}\nonumber \\
& = & -\frac{1}{2\pi}\int d^{2}z\{\partial Z^{M}\bar{\partial}Z^{N}B_{NM}\},
\end{eqnarray}
with
\begin{equation}
B_{AB}=(-1)^{A(B+N)}E_{B}^{\hphantom{B}N}E_{A}^{\hphantom{A}M}B_{MN}.
\end{equation}
This form is more suitable to show the gauge invariance of the action
with respect to the transformations $\delta B_{MN}=\partial_{M}\Sigma_{N}-(-1)^{MN}\partial_{N}\Sigma_{M}$.
More details on the conventions used here can be found in Appendix
\ref{sec:notation}. Concerning the dilaton superfield, it plays no
role in the classical description and this can be seen from the fact
that its coupling to the action naively vanishes in the $\alpha'\to0$ limit.
Following the analysis of subsection \ref{sub:Classical-analysis},
$\mathcal{P}_{a}$ and $d_{\alpha}$ can be viewed as independent
objects invariant under supesymmetry, and the flat space limit of
$S$ is recovered when we express them in terms of regular variables,
\emph{cf.} \eqref{eq:flatsusyinvariants}, together with the non-vanishing
components of $E$ and $B$ in that limit:
\begin{equation}
\begin{array}{cc}
E_{m}^{\hphantom{m}a}=\delta_{m}^{a}, & E_{\mu}^{\hphantom{\mu}a}=-\frac{1}{2}(\gamma^{a}\theta)_{\mu}\\
E_{\mu}^{\hphantom{\mu}\alpha}=\delta_{\mu}^{\alpha}, & B_{m\mu}=-B_{\mu m}=\frac{1}{2}E_{m}^{\hphantom{m}a}(\gamma_{a}\theta)_{\mu}.
\end{array}
\end{equation}
The energy-momentum tensor of the curved space action is given by
\begin{equation}
T=-\mathcal{P}_{a}\Pi^{a}-d_{\alpha}\Pi^{\alpha}-w_{\alpha}\nabla\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C},\label{eq:SUGRAT}
\end{equation}
and the curved version of $\mathcal{H}$ in \eqref{eq:heteroticH}
is simply
\begin{equation}
\mathcal{H}=-\frac{1}{2}\mathcal{P}_{a}\mathcal{P}^{a}-\frac{1}{2}\Pi^{a}\Pi_{a}+d_{\alpha}\Pi^{\alpha}+w_{\alpha}\nabla\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}.\label{eq:SUGRAH}
\end{equation}
The BRST-charge in the curved background has the same structure of
\eqref{eq:heteroticBRST} and the presence of the background can be
seen through the canonical conjugates of the superspace coordinates
$Z^{M}$, denoted by $\hat{P}_{M}$. Using the definition \eqref{eq:defcanonical},
one obtains
\begin{equation}
\hat{P}_{M}=E_{M}^{\hphantom{M}a}\mathcal{P}_{a}-E_{M}^{\hphantom{M}\alpha}d_{\alpha}+\partial_{\sigma}Z^{N}B_{NM}+w_{\alpha}\lambda^{\beta}\Omega_{M\beta}^{\hphantom{M\beta}\alpha},\label{eq:curvedcanonical}
\end{equation}
which enables us to rewrite $\mathcal{P}_{a}$ and $d_{\alpha}$ as\begin{subequations}\label{eq:curvedPdconjugates}
\begin{eqnarray}
d_{\alpha} & = & -E_{\alpha}^{\hphantom{\alpha}M}\hat{P}_{M}+(\Pi^{A}+\bar{\Pi}^{A})B_{A\alpha}+w_{\gamma}\lambda^{\beta}\Omega_{\alpha\beta}^{\hphantom{\alpha\beta}\gamma},\label{eq:dcurved}\\
\mathcal{P}_{a} & = & E_{a}^{\hphantom{\alpha}M}\hat{P}_{M}-(\Pi^{A}+\bar{\Pi}^{A})B_{Aa}-w_{\gamma}\lambda^{\beta}\Omega_{a\beta}^{\hphantom{a\beta}\gamma}.\label{eq:Pcurved}
\end{eqnarray}
\end{subequations}To go from \eqref{eq:curvedcanonical} to \eqref{eq:curvedPdconjugates},
we have used the inverse vielbein $E_{A}^{\hphantom{A}M}$, such that
$E_{A}^{\hphantom{A}M}E_{M}^{\hphantom{M}B}=\delta_{A}^{B}$ and $E_{M}^{\hphantom{N}A}E_{A}^{\hphantom{A}N}=\delta_{M}^{N}$.
For the BRST-charge to be nilpotent, and thus well-defined as such,
the background superfields need to satisfy a number of constraints.
To find these constraints, we begin by computing the transformations
of the worldsheet fields under the action of $Q_{\lambda}$. Using
the graded Poisson brackets
\begin{eqnarray}
\{\hat{P}_{M}(\sigma'),Z^{N}(\sigma)\}_{P.B.} & = & -\delta_{M}^{N}\delta(\sigma-\sigma^{\prime}),\label{eq:canonicalZP}\\
\{w_{\alpha}(\sigma'),\lambda^{\beta}(\sigma)\}_{P.B.} & = & -\delta_{\alpha}^{\beta}\delta(\sigma-\sigma^{\prime}),\label{eq:poissonlambda}
\end{eqnarray}
we obtain the following transformations\footnote{For a reasonably detailed exposition of similar calculations, see
\cite{Chandia:2006ix}.} \begin{subequations}\label{eq:curvedbrsttransf}
\begin{eqnarray}
\delta\lambda^{\alpha} & = & -\lambda^{\beta}\Lambda_{\beta}^{\hphantom{\beta}\alpha},\\
\delta w_{\alpha} & = & \Lambda_{\alpha}^{\hphantom{\alpha}\beta}w_{\beta}+\epsilon d_{\alpha},\\
\delta\Pi^{a} & = & -\Pi^{b}\Lambda_{b}^{\hphantom{b}a}-\epsilon\lambda^{\alpha}\Pi^{A}T_{A\alpha}^{\hphantom{A\alpha}a},\\
\delta\Pi^{\alpha} & = & -\Pi^{\beta}\Lambda_{\beta}^{\hphantom{\beta}\alpha}+\epsilon\nabla\lambda^{\alpha}-\epsilon\lambda^{\beta}\Pi^{A}T_{A\beta}^{\hphantom{A\beta}\alpha},\\
\delta\mathcal{P}_{a} & = & \Lambda_{a}^{\hphantom{a}b}\mathcal{P}_{b}-\epsilon\lambda^{\beta}T_{\beta a}^{\hphantom{\beta a}b}\mathcal{P}_{b}+\epsilon\lambda^{\gamma}T_{\gamma a}^{\hphantom{\gamma a}\beta}d_{\beta}+\epsilon\lambda^{\beta}\Pi^{A}H_{A\beta a}-\epsilon\lambda^{\gamma}\lambda^{\beta}w_{\delta}R_{\beta a\gamma}^{\hphantom{\beta a\gamma}\delta},\\
\delta d_{\alpha} & = & \Lambda_{\alpha}^{\hphantom{\alpha}\beta}d_{\beta}+\epsilon\lambda^{\beta}T_{\beta\alpha}^{\hphantom{\beta\alpha}b}\mathcal{P}_{b}-\epsilon\lambda^{\gamma}T_{\gamma\alpha}^{\hphantom{\gamma\alpha}\beta}d_{\beta}-\epsilon\lambda^{\beta}\Pi^{A}H_{A\beta\alpha}+\epsilon\lambda^{\gamma}\lambda^{\beta}w_{\delta}R_{\beta\alpha\gamma}^{\hphantom{\beta\alpha\gamma}\delta},\\
\delta\Omega_{\alpha}^{\hphantom{\alpha}\beta} & = & \nabla\Lambda_{\alpha}^{\hphantom{\alpha}\beta}-\epsilon\lambda^{\gamma}\Pi^{A}R_{A\gamma\alpha}^{\hphantom{A\gamma\alpha}\beta}.
\end{eqnarray}
\end{subequations}Here $\epsilon$ is a constant anticommuting parameter
and we have defined
\begin{equation}
\begin{array}{cc}
\Lambda_{A}^{\hphantom{A}B}\equiv\epsilon\lambda^{\alpha}\Omega_{\alpha A}^{\hphantom{\alpha A}B}, & \Omega_{\alpha}^{\hphantom{\alpha}\beta}\equiv\Pi^{A}\Omega_{A\alpha}^{\hphantom{A\alpha}\beta}.\end{array}
\end{equation}
Now we can compute the transformation of $\lambda^{\alpha}d_{\alpha}$,
whose vanishing is equivalent to the first condition displayed in
\eqref{eq:nilpotency-conditions}:
\begin{equation}
\delta(\lambda^{\alpha}d_{\alpha})=\epsilon\lambda^{\alpha}\lambda^{\beta}[T_{\alpha\beta}^{\hphantom{\alpha\beta}a}\mathcal{P}_{a}-T_{\alpha\beta}^{\hphantom{\alpha\beta}\gamma}d_{\gamma}-\Pi^{A}H_{A\alpha\beta}+w_{\delta}\lambda^{\gamma}R_{\alpha\beta\gamma}^{\hphantom{\alpha\beta\gamma}\delta}].\label{eq:SUGRAdeltaQlambda}
\end{equation}
Hence the first set of constraints required for the nilpotency of
$Q$ is:
\begin{equation}
\lambda^{\alpha}\lambda^{\beta}T_{\alpha\beta}^{\hphantom{\alpha\beta}A}=\lambda^{\alpha}\lambda^{\beta}H_{A\alpha\beta}=\lambda^{\alpha}\lambda^{\beta}\lambda^{\gamma}R_{\alpha\beta\gamma}^{\hphantom{\alpha\beta\gamma}\delta}=0.\label{eq:nilpotencySUGRAconstraints}
\end{equation}
Nilpotency of the BRST charge also requires $\delta T_{+}$ to vanish.
This is just another way of stating the condition \eqref{eq:nilpotencyTlambda}.
The operator $T_{+}$ is obtained from the definition \eqref{eq:T+-}
and the curved versions of $T$ and $\mathcal{H}$, respectively \eqref{eq:SUGRAT}
and \eqref{eq:SUGRAH}. It can be cast as
\begin{equation}
T_{+}=-\frac{1}{4}\eta^{ab}(\mathcal{P}_{a}+\Pi_{a})(\mathcal{P}_{b}+\Pi_{b})-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}.\label{eq:T+curved}
\end{equation}
Now, to compute $\delta T_{+}$ we just have to use the transformations
of $\mathcal{P}_{a}$ and $\Pi^{a}$ in \eqref{eq:curvedbrsttransf}
and the result is
\begin{eqnarray}
\delta T_{+} & = & -\frac{1}{2}\epsilon(\mathcal{P}^{a}+\Pi^{a})\lambda^{\alpha}d_{\beta}T_{\alpha a}^{\hphantom{\alpha a}\beta}+\frac{1}{2}\epsilon(\mathcal{P}^{a}\mathcal{P}^{b}-\Pi^{a}\Pi^{b})\lambda^{\alpha}T_{\alpha ab}-\frac{1}{2}\epsilon\lambda^{\alpha}\mathcal{P}^{a}\Pi^{b}H_{\alpha ab}\nonumber \\
& & +\frac{1}{2}\epsilon(\mathcal{P}^{a}+\Pi^{a})\lambda^{\alpha}\Pi^{\beta}(T_{\alpha\beta a}-H_{\alpha\beta a})+\frac{1}{2}\epsilon(\mathcal{P}^{a}+\Pi^{a})\lambda^{\alpha}\lambda^{\beta}w_{\gamma}R_{\beta a\alpha}^{\hphantom{\beta a\alpha}\gamma}.
\end{eqnarray}
For this expression to vanish, we need to impose another set constraints:
\begin{equation}
T_{\alpha a}^{\hphantom{\alpha a}\beta}=T_{\alpha(ab)}=T_{\alpha\beta b}-H_{\alpha\beta b}=H_{ab\alpha}=\lambda^{\alpha}\lambda^{\beta}R_{\alpha a\beta}^{\hphantom{\alpha a\beta}\gamma}=0.\label{eq:holoSUGRAconstraints}
\end{equation}
In the usual pure spinor superstring, this set comes from the holomorphicity
of the BRST charge \cite{Berkovits:2001ue}.
Finally, rewriting $T_{+}$ in terms of the canonical conjugates and
using the Poisson brackets of \eqref{eq:canonicalZP} together with
\begin{eqnarray}
\{\bar{b}(\sigma'),\bar{c}(\sigma)\}_{P.B.} & = & \delta(\sigma'-\sigma),\label{eq:poissonbc}\\
\{T_{C}(\sigma'),T_{C}(\sigma)\}_{P.B.} & = & 2T_{C}\,\partial_{\sigma}\delta(\sigma'-\sigma)+\partial_{\sigma}T_{C}\,\delta(\sigma'-\sigma),\label{eq:poissonTc}
\end{eqnarray}
we can show that
\begin{equation}
\{T_{+}(\sigma'),T_{+}(\sigma)\}_{P.B.}=2T_{+}\,\partial_{\sigma}\delta(\sigma'-\sigma)+\partial_{\sigma}T_{+}\,\delta(\sigma'-\sigma).
\end{equation}
Therefore, the three conditions of \eqref{eq:nilpotency-conditions}
for classical nilpotency of the BRST charge in a supergravity background
are all satisfied provided the constraints displayed in \eqref{eq:nilpotencySUGRAconstraints}
and \eqref{eq:holoSUGRAconstraints}.
\subsection{Turning on the super Yang-Mills background}
\
In order to find the remaining heterotic background constraints, we
need to consider the case in which the super Yang-Mills fields are
present. We will introduce the minimal coupling between the gauge
potential $A_{M}^{I}$ and the currents $J_{I}$, such that the action
has the form
\begin{eqnarray}
S & = & \frac{1}{2\pi}\int d^{2}z\{\mathcal{P}_{a}\bar{\Pi}^{a}+d_{\alpha}\bar{\Pi}^{\alpha}-\Pi^{A}\bar{\Pi}^{B}B_{BA}+\bar{A}^{I}J_{I}+w_{\alpha}\bar{\nabla}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C},\label{eq:hetaction_SYM}
\end{eqnarray}
where $\bar{A}^{I}\equiv\bar{\partial}Z^{M}A_{M}^{I}$.
Note this is equivalent to replacing $\mathcal{P}_{a}\to\mathcal{P}_{a}+A_{a}^{I}J_{I}$
and $d_{\alpha}\to d_{\alpha}-A_{\alpha}^{I}J_{I}$ in the action
\ref{eq:hetaction_curved}, with $A_{A}^{I}=E_{A}^{\hphantom{A}M}A_{M}^{I}$.
The gauge invariance of the action \eqref{eq:hetaction_SYM} with
respect to the super Yang-Mills background is straightforward to demonstrate.
Consider the gauge transformations with superparameter $\Sigma^{I}$,
\begin{equation}
\delta_{\Sigma}A_{M}^{I}=\partial_{M}\Sigma^{I}+[A_{M},\Sigma]^{I},\label{eq:SYMgauge}
\end{equation}
where
\begin{equation}
[A_{M},\Sigma]^{I}=f_{JK}^{\hphantom{JK}I}A_{M}^{J}\Sigma^{K},
\end{equation}
and $f_{JK}^{\hphantom{JK}I}$ denotes the structure constants of
the gauge group. Substituting the transformation \eqref{eq:SYMgauge}
in the variation of the action and integrating by parts, one obtains
\begin{equation}
\delta_{\Sigma}S=-\frac{1}{2\pi}\int d^{2}z(\bar{\partial}J_{I}+f_{IJ}^{\hphantom{IJ}K}\bar{A}^{J}J_{K})\Sigma^{I}.
\end{equation}
The expression inside the parentheses is just the equation of motion
for the current $J_{I}$ in the presence of the super Yang-Mills source
and thus vanishes at the classical level. More details on this construction
can be found in appendix \ref{sec:currents} where an explicit realization
of the gauge sector is given for the group $SO(32)$ in terms of worldsheet
fermions. For convenience, we can introduce the super field strength
of $A_{M}^{I}$, given by
\begin{eqnarray}
F_{MN}^{I} & = & \partial_{M}A_{N}^{I}-(-1)^{MN}\partial_{N}A_{M}^{I}+f_{JK}^{\hphantom{JK}I}A_{M}^{J}A_{N}^{K},\label{eq:superfieldstrength}
\end{eqnarray}
which transforms covariantly under \eqref{eq:SYMgauge},
\begin{equation}
\delta_{\Sigma}F_{MN}^{I}=[F_{MN},\Sigma]^{I}.
\end{equation}
The coupling to the super Yang-Mills background changes the energy-momentum
tensor to
\begin{equation}
T=-\mathcal{P}_{a}\Pi^{a}-d_{\alpha}\Pi^{\alpha}-w_{\alpha}\nabla\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}-\Pi^{A}A_{A}^{I}J_{I},\label{eq:SYMT}
\end{equation}
and we also expect the operator $\mathcal{H}$ to be modified accordingly.
It was suggested in \cite{Chandia-Vallilo} that fluctuations of the
background would be manifested through $\mathcal{H}$. Therefore,
inspired by the superstring integrated vertex, we propose
\begin{eqnarray}
\mathcal{H} & = & -\frac{1}{2}\mathcal{P}_{a}\mathcal{P}^{a}-\frac{1}{2}\Pi^{a}\Pi_{a}+d_{\alpha}\Pi^{\alpha}+w_{\alpha}\nabla\lambda^{\alpha}-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}\nonumber \\
& & -\Pi^{A}A_{A}^{I}J_{I}-d_{\alpha}W^{\alpha I}J_{I}-\lambda^{\alpha}w_{\beta}U_{\alpha}^{\hphantom{\alpha}\beta I}J_{I},\label{eq:SYMH}
\end{eqnarray}
where $W^{\alpha I}$ and $U_{\alpha}^{\hphantom{\alpha}\beta I}$
are background superfields\footnote{As before, we must impose $\lambda^{\alpha}(\gamma^{a}\lambda)_{\beta}U_{\alpha}^{\hphantom{\alpha}\beta I}=0$ in order to respect the gauge invariance implied by the pure spinor constraint.} which will be related to \eqref{eq:superfieldstrength}. Again, using
\eqref{eq:T+-}, $T_{+}$ can be cast as
\begin{eqnarray}
T_{+} & = & -\frac{1}{4}\eta^{ab}(\mathcal{P}_{a}+\Pi_{a})(\mathcal{P}_{b}+\Pi_{b})-\bar{b}\partial\bar{c}-\partial(\bar{b}\bar{c})+T_{C}\nonumber \\
& & -\Pi^{A}A_{A}^{I}J_{I}-\frac{1}{2}d_{\alpha}W^{\alpha I}J_{I}-\frac{1}{2}\lambda^{\alpha}w_{\beta}U_{\alpha}^{\hphantom{\alpha}\beta I}J_{I},\label{eq:SYMT+}
\end{eqnarray}
and we now have all the ingredients to analyze the BRST symmetry in
this background.
The modification of the action entails a change in the classical BRST
transformations of the worldsheet fields. This is clearly seen from
the canonical conjugates of the superspace coordinates, which now
have a linear dependence on the gauge field:
\begin{equation}
\hat{P}_{M}=E_{M}^{\hphantom{M}a}\mathcal{P}_{a}-E_{M}^{\hphantom{M}\alpha}d_{\alpha}+\partial_{\sigma}Z^{N}B_{NM}+w_{\alpha}\lambda^{\beta}\Omega_{M\beta}^{\hphantom{M\beta}\alpha}+A_{M}^{I}J_{I}.\label{eq:curvedSYMcanonical}
\end{equation}
To compute these transformations we will use the fundamental brackets
of \eqref{eq:canonicalZP}, \eqref{eq:poissonlambda}, \eqref{eq:poissonbc}
and \eqref{eq:poissonTc}, together with\begin{subequations}
\begin{eqnarray}
\{T_{C}(\sigma'),J_{I}(\sigma)\}_{P.B.} & = & J_{I}\partial_{\sigma}\delta(\sigma-\sigma')+\partial_{\sigma}J_{I}\delta(\sigma'-\sigma),\\
\{J_{I}(\sigma'),J_{J}(\sigma)\}_{P.B.} & = & f_{IJ}^{\hphantom{IJ}K}J_{K}\delta(\sigma'-\sigma),
\end{eqnarray}
\end{subequations}which are derived in Appendix \ref{sec:currents}.
Considering first $Q_{\lambda}$, it is clear that most of the transformations
displayed in \eqref{eq:curvedbrsttransf} remain unchanged, except
for $\delta\mathcal{P}_{a}$ and $\delta d_{\alpha}$, which are now
given by\begin{subequations}
\begin{eqnarray}
\delta\mathcal{P}_{a} & = & \Lambda_{a}^{\hphantom{a}b}\mathcal{P}_{b}-\epsilon\lambda^{\beta}T_{\beta a}^{\hphantom{\beta a}b}\mathcal{P}_{b}+\epsilon\lambda^{\gamma}T_{\gamma a}^{\hphantom{\gamma a}\beta}d_{\beta}\nonumber \\
& & +\epsilon\lambda^{\alpha}\Pi^{A}H_{A\alpha a}-\epsilon\lambda^{\gamma}\lambda^{\beta}w_{\delta}R_{\beta a\gamma}^{\hphantom{\beta a\gamma}\delta}-\epsilon\lambda^{\beta}F_{\beta a}^{I}J_{I},\\
\delta d_{\alpha} & = & \Lambda_{\alpha}^{\hphantom{\alpha}\beta}d_{\beta}+\epsilon\lambda^{\beta}T_{\beta\alpha}^{\hphantom{\beta\alpha}b}\mathcal{P}_{b}-\epsilon\lambda^{\gamma}T_{\gamma\alpha}^{\hphantom{\gamma\alpha}\beta}d_{\beta}\nonumber \\
& & -\epsilon\lambda^{\beta}\Pi^{A}H_{A\beta\alpha}+\epsilon\lambda^{\gamma}\lambda^{\beta}w_{\delta}R_{\beta\alpha\gamma}^{\hphantom{\beta\alpha\gamma}\delta}+\epsilon\lambda^{\beta}F_{\beta\alpha}^{I}J_{I},
\end{eqnarray}
\end{subequations}where
\begin{equation}
F_{AB}^{I}=(-1)^{A(B+N)}E_{B}^{\hphantom{B}N}E_{A}^{\hphantom{A}M}F_{MN}^{I}.
\end{equation}
We can now easily compute the transformation of $\lambda^{\alpha}d_{\alpha}$
and the result is
\begin{equation}
\delta(\lambda^{\alpha}d_{\alpha})=\epsilon\lambda^{\alpha}\lambda^{\beta}[T_{\alpha\beta}^{\hphantom{\alpha\beta}a}\mathcal{P}_{a}-T_{\alpha\beta}^{\hphantom{\alpha\beta}\gamma}d_{\gamma}-\Pi^{A}H_{A\alpha\beta}+w_{\delta}\lambda^{\gamma}R_{\alpha\beta\gamma}^{\hphantom{\alpha\beta\gamma}\delta}+F_{\alpha\beta}^{I}J_{I}].
\end{equation}
Thus, together with the constraints displayed in \eqref{eq:nilpotencySUGRAconstraints}
we also need to impose
\begin{equation}
\lambda^{\alpha}\lambda^{\beta}F_{\alpha\beta}^{I}=0,\label{eq:nilpotencySYM}
\end{equation}
in order to satisfy the first nilpotency condition, \emph{cf}. equation
\eqref{eq:nilpotencylambda2}.
Next, to compute $\delta T_{+}$ and evaluate the condition \eqref{eq:nilpotencyTlambda}
it is worth noting that $T_{C}$ and $J_{I}$ now have nonvanishing
transformations with respect to $Q_{\lambda}$, given by\begin{subequations}
\begin{eqnarray}
\delta T_{C} & = & J_{I}\partial\Sigma^{I},\\
\delta J_{I} & = & -f_{IJ}^{\hphantom{IJ}K}\Sigma^{J}J_{K},
\end{eqnarray}
\end{subequations}where we have defined the gauge-like parameter
$\Sigma^{I}\equiv\epsilon\lambda^{\alpha}A_{\alpha}^{I}$. The introduction
of $\Sigma^{I}$ is convenient when we look at the variations of the
background superfields,\begin{subequations}
\begin{eqnarray}
\delta A^{I} & = & \nabla\Sigma^{I}-\epsilon\lambda^{\alpha}\Pi^{A}F_{A\alpha}^{I},\\
\delta W^{\alpha I} & = & -W^{\beta I}\Lambda_{\beta}^{\hphantom{\beta}\alpha}-[\Sigma,W^{\alpha}]^{I}+\epsilon\lambda^{\beta}\nabla_{\beta}W^{\alpha I},\\
\delta U_{\alpha}^{\hphantom{\alpha}\beta I} & = & \Lambda_{\alpha}^{\hphantom{\alpha}\gamma}U_{\gamma}^{\hphantom{\gamma}\beta I}-U_{\alpha}^{\hphantom{\alpha}\gamma I}\Lambda_{\gamma}^{\hphantom{\gamma}\beta}-[\Sigma,U_{\alpha}^{\hphantom{\alpha}\beta}]^{I}+\epsilon\lambda^{\gamma}\nabla_{\gamma}U_{\alpha}^{\hphantom{\alpha}\beta I},
\end{eqnarray}
\end{subequations}which can be interpreted in terms of a gauge-like
transformation with parameter $\Sigma$, a Lorentz-like transformation
with parameter $\Lambda$, and a superspace translation, with $\nabla_{\alpha}$
denoting the covariant derivative with respect to the local symmetries,
\emph{e.g.}
\begin{equation}
\nabla_{\beta}W^{\alpha I}=D_{\beta}W^{\alpha I}+\Omega_{\beta\gamma}^{\hphantom{\beta\gamma}\alpha}W^{\gamma I}+[A_{\beta},W^{\alpha}]^{I}.
\end{equation}
Gathering all these results and using the supergravity constraints
of \eqref{eq:holoSUGRAconstraints}, we obtain
\begin{eqnarray}
\delta T_{+} & = & \frac{1}{2}\epsilon\lambda^{\alpha}(\mathcal{P}^{a}-\Pi^{a})[F_{\alpha a}^{I}-T_{\alpha\beta a}W^{\beta I}]J_{I}-\frac{1}{2}\epsilon\lambda^{\alpha}\lambda^{\beta}w_{\gamma}[\nabla_{\alpha}U_{\beta}^{\hphantom{\beta}\gamma I}+R_{\delta\alpha\beta}^{\hphantom{\delta\alpha\beta}\gamma}W^{\delta I}]J_{I}\nonumber \\
& & +\epsilon\lambda^{\alpha}\Pi^{\beta}[F_{\alpha\beta}^{I}+\frac{1}{2}H_{\alpha\beta\gamma}W^{\gamma I}]J_{I}+\frac{1}{2}\epsilon\lambda^{\alpha}d_{\beta}[\nabla_{\alpha}W^{\beta I}-T_{\alpha\gamma}^{\hphantom{\alpha\gamma}\beta}W^{\gamma I}-U_{\alpha}^{\hphantom{\alpha}\beta I}]J_{I}\nonumber \\
& & -\frac{1}{2}\epsilon\lambda^{\alpha}F_{\alpha\beta}^{I}W^{\beta J}J_{I}J_{J},\label{newvariationH}
\end{eqnarray}
Hence, we have to further impose the following constraints:\begin{subequations}\label{eq:holoSYM}
\begin{eqnarray}
F_{\alpha a}^{I} & = & T_{\alpha\beta a}W^{\beta I},\\
\nabla_{\alpha}W^{\beta I}-T_{\alpha\gamma}^{\hphantom{\alpha\gamma}\beta}W^{\gamma I} & = & U_{\alpha}^{\hphantom{\alpha}\beta I},\\
F_{\alpha\beta}^{I} & = & \frac{1}{2}W^{\gamma I}H_{\alpha\beta\gamma},\label{eq:F=00003DWH}\\
\lambda^{\alpha}\lambda^{\beta}\nabla_{\alpha}U_{\beta}^{\hphantom{\beta}\gamma I} & = & -\lambda^{\alpha}\lambda^{\beta}R_{\delta\alpha\beta}^{\hphantom{\delta\alpha\beta}\gamma}W^{\delta I}.
\end{eqnarray}
\end{subequations}Note that the last line of \eqref{newvariationH}
vanishes automatically after the identification in \eqref{eq:F=00003DWH}.
As a final consistency check, it is not difficult to show that $T_{+}$
satisfies
\begin{equation}
\{T_{+}(\sigma'),T_{+}(\sigma)\}_{P.B.}=2T_{+}\,\partial_{\sigma}\delta(\sigma'-\sigma)+\partial_{\sigma}T_{+}\,\delta(\sigma'-\sigma),
\end{equation}
which demonstrates the last necessary condition for the nilpotency
of the BRST charge at the classical level.
\section{Discussion\label{sec:discussion}}
\
It is possible to show that the constraints displayed in \eqref{eq:nilpotencySUGRAconstraints},
\eqref{eq:holoSUGRAconstraints}, \eqref{eq:nilpotencySYM} and \eqref{eq:holoSYM}
imply the ten-dimensional supergravity and super Yang-Mills equations
of motion. Instead of presenting these results, which for the pure
spinor superstring were originally obtained and detailedly studied
in \cite{Berkovits:2001ue}, we will discuss the particularities of
the infinite tension string model.
As presented in subsection \ref{sub:Classical-analysis}, there are
in principle three independent conditions to check in order to ensure
classical nilpotency of the BRST charge,\begin{subequations}\label{eq:nilpotencyconditions}
\begin{eqnarray}
Q_{\lambda}^{2} & = & 0,\\
\{\lambda^{\alpha}d_{\alpha}(\sigma'),T_{+}(\sigma)\}_{P.B.} & = & 0,\\
\{T_{+}(\sigma'),T_{+}(\sigma)\}_{P.B.} & = & 2T_{+}\,\partial_{\sigma}\delta(\sigma'-\sigma)+\partial_{\sigma}T_{+}\,\delta(\sigma'-\sigma).
\end{eqnarray}
\end{subequations}The first one is identical to the condition on
the left-moving BRST charge of the usual pure spinor superstring and
not surprisingly provides the so-called nilpotency constraints, given
by
\begin{equation}
\lambda^{\alpha}\lambda^{\beta}T_{\alpha\beta}^{\hphantom{\alpha\beta}A}=\lambda^{\alpha}\lambda^{\beta}H_{A\alpha\beta}=\lambda^{\alpha}\lambda^{\beta}\lambda^{\gamma}R_{\alpha\beta\gamma}^{\hphantom{\alpha\beta\gamma}\delta}=\lambda^{\alpha}\lambda^{\beta}F_{\alpha\beta}^{I}=0,
\end{equation}
exactly as in \cite{Berkovits:2001ue}. The second condition can be
stated as
\begin{equation}
[Q_{\lambda},T_{+}]=0,\label{eq:QlambdaT+discussion}
\end{equation}
and leads to the constraints of \eqref{eq:holoSUGRAconstraints} and
\eqref{eq:holoSYM}, which were obtained in \cite{Berkovits:2001ue}
by requiring holomorphicity of the BRST current. In the present model,
holomorphicity plays no role when it comes to imposing constraints.
This is so because the heterotic background coupled action, given
by
\begin{equation}
S=\frac{1}{2\pi}\int d^{2}z\{\mathcal{P}_{a}\bar{\Pi}^{a}+d_{\alpha}\bar{\Pi}^{\alpha}-\Pi^{A}\bar{\Pi}^{B}B_{BA}+\bar{A}^{I}J_{I}+w_{\alpha}\bar{\nabla}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C},
\end{equation}
is still chiral. Therefore the remaining constraints should manifest
themselves through the condition \eqref{eq:QlambdaT+discussion}.
To interpret it, it might be useful to recall that conformal symmetry
is preserved at the classical level, such that $[Q_{\lambda},T]=0$.
We are then left with
\begin{eqnarray}
[Q_{\lambda},\mathcal{H}] & = & 0,
\end{eqnarray}
\emph{cf}. the definition \ref{eq:T+-}. This is precisely the \emph{ad
hoc} condition used in \cite{Chandia-Vallilo} for the type II construction.
Here, however, it is naturally embedded in the BRST operator. From
the sectorized point of view, the condition above is equivalent to
the conservation of the BRST charge separately in each sector.
Concerning the third condition in \eqref{eq:nilpotencyconditions},
it was verified to hold \emph{independently} of the background constraints.
This is partially connected to the classical conformal symmetry but
we do not have a clear understanding so far. It implies, for example,
that
\begin{equation}
\{\mathcal{H}(\sigma'),\mathcal{H}(\sigma)\}_{P.B.}=2T\,\partial_{\sigma}\delta(\sigma'-\sigma)+\partial_{\sigma}T\,\delta(\sigma'-\sigma).
\end{equation}
We expect this relation to hold in the type II case as well.
An interesting observation is that the background can be absorbed
by a field redefinition in the action, such that
\begin{equation}
S=\frac{1}{2\pi}\int d^{2}z\{\bar{\partial}Z^{M}\boldsymbol{P}_{M}+w_{\alpha}\bar{\partial}\lambda^{\alpha}+\bar{b}\bar{\partial}\bar{c}\}+S_{C},
\end{equation}
where
\begin{eqnarray}
\boldsymbol{P}_{M} & \equiv & E_{M}^{\hphantom{M}a}(\mathcal{P}_{a}+\Pi^{A}B_{Aa}+A_{a}^{I}J_{I}+w_{\gamma}\lambda^{\beta}\Omega_{a\beta}^{\hphantom{a\beta}\gamma})\nonumber \\
& & -E_{M}^{\hphantom{M}\alpha}(d_{\alpha}-\Pi^{A}B_{A\alpha}-A_{\alpha}^{I}J_{I}-w_{\gamma}\lambda^{\beta}\Omega_{\alpha\beta}^{\hphantom{\alpha\beta}\gamma}).
\end{eqnarray}
In this case, one can work with $\boldsymbol{P}_{M}$ as a fundamental
field and rewrite the BRST charge by expressing $d_{\alpha}$ and
$\mathcal{P}_{a}$ as functions of $\boldsymbol{P}_{M}$ and the other
worldsheet fields. Therefore we have instead a free action and the
heterotic background appears as a deformation of the BRST charge,
supporting the observation made by Chandia and Vallilo that the vertex
operators in this model could be seen as flutuations of $\mathcal{H}$.
Needless to emphasize, quantum consistency of the theory would be
much easier to verify in this approach, similarly to what was done
in \cite{Adamo:2014wea} for the NS-NS background. As in there, we
expect the dilaton superfield to start playing a fundamental role
in the quantum formulation of the theory.
It should be noted that in the original ambitwistor strings, either
in RNS or with pure spinors, the heterotic supergravity sector has
some unsolved issues. For example, Mason and Skinner \cite{Mason:2013sva}
computed the n-particle tree level amplitude and could not interpret
them in terms of standard space-time gravity, with the 3-point amplitude
suggesting a $(\textrm{Weyl})^{3}$-type vertex. On the other hand,
the supergravity vertex of \cite{Jusinskas:2016qjd} in the sectorized
string seems to provide the correct OPE structure in the 3-point amplitude,
resembling the usual heterotic pure spinor string up to numerical
factors. This subject deserves a deeper investigation and might shed
some light on the model. Naturally, if we go to 4-point amplitudes
or higher we need also the integrated vertices. We still do not have
a simple proposal for such operators. However, there are interesting
hints pointing out that the holomorphic sectorization can be extended
the bosonic string and to the Ramond-Neveu-Schwarz and Green-Schwarz
formalisms \cite{Chandia-Jusinskas}. We hope that understanding this
construction will provide a better basis to approach the problem of
the integrated vertex operator in the infinite tension limit using
pure spinors. Once this step is taken, we will finally be able to
compute the tree level amplitudes to compare them with the Cachazo-He-Yuan
formulae \cite{Cachazo:2013hca} and investigate the modular invariance
of the theory at 1-loop, for example, as done in \cite{Adamo:2013tsa}.
\
\textbf{Acknowledgements:} TA acknowledges financial support from
Conselho Nacional de Desenvolvimento Cient\'ifico e Tecnol\'ogico (CNPq).
The research of TA was also supported in part by the Knut and Alice
Wallenberg Foundation under grant 2015.0083. RLJ would like to thank
the Grant Agency of the Czech Republic for financial support under
the grant P201/12/G028.
|
\section{Introduction}
The statistical anomalies at about $750\ensuremath{\,\text{GeV}}$ in ATLAS\cite{atlas,Aaboud:2016tru} and CMS\cite{CMS:2015dxe,Khachatryan:2016hje} searches for a diphoton resonance (denoted in this text as \digamma) at \roots{13} with about $3\ensuremath{/\text{fb}}$ caused considerable
activity \see{Ellis:2015oso,Franceschini:2015kwy,Strumia:2016wys}. The experiments reported local significances, which incorporate a look-elsewhere effect (LEE, see e.g.,\xspace \refcite{Lyons:2013yja,2008arXiv0811.1663L}) in the production cross section of the \digamma, of $3.9\sigma$ and $3.4\sigma$,
respectively, and global significances, which incorporate a LEE in the production cross section, mass and width of the \digamma, of $2.1\sigma$ and $1.6\sigma$, respectively. There was concern, however, that an overall LEE, accounting for the numerous hypothesis tests of the SM at the LHC, cannot be incorporated, and that the plausibility of the \digamma was difficult to gauge.
Whilst ultimately the \digamma was disfavoured by searches with about $15\ensuremath{/\text{fb}}$\cite{CMS:2016crm,ATLAS:2016eeo}, we directly calculate the relative plausibility of the SM versus the SM plus \digamma in light of ATLAS data available during the excitement, matching, wherever possible,
parameter ranges and parameterisations in the frequentist analyses. The relative plausibility sidesteps technicalities about the LEE and the frequentist formalism required to interpret significances. We calculate the
Bayes-factor \see{Gregory} in light of ATLAS data,
\begin{equation}
\frac{\cond{\text{ATLAS data}}{\text{SM + \digamma}}}{\cond{\text{ATLAS data}}{\text{SM}}}
=
\frac{
\frac{\cond{\text{SM + \digamma}}{\text{ATLAS data}}}{\cond{\text{SM}}{\text{ATLAS data}}}
}{
\frac{\prob{\text{SM + \digamma}}}{\prob{\text{SM}}}
}.
\end{equation}
Our main result is that we find that, at its peak, the Bayes-factor was about $7.7$ in favour of the \digamma.
In other words, in light of the ATLAS \s{13} $3.2\ensuremath{/\text{fb}}$ and \s{8} $20.3\ensuremath{/\text{fb}}$ diphoton searches, the relative plausibility of the \digamma versus the SM alone
increased by about eight. This was ``substantial'' on the Jeffreys' scale\cite{Jeffreys:1939xee}, lying between ``not worth more than a bare mention'' and ``strong evidence.'' For completeness, we calculated that this preference was reversed by the ATLAS \s{13} $15.4\ensuremath{/\text{fb}}$ search\cite{ATLAS:2016eeo}, resulting in a Bayes-factor of about $0.7$. Nevertheless, the interest in \digamma models in the interim was, to some degree, supported by Bayesian and frequentist analyses. Unfortunately, CMS performed searches in numerous event categories, resulting in a proliferation of background nuisance parameters and making replication difficult without cutting corners or considerable
computing power.
\section{Calculation}
The background shape was characterised by a monotonically decreasing function with two free parameters,
\begin{equation}\label{eq:bg_dist}
p_b(m_{\gamma\gamma}) \propto \left[1 - \left(\frac{m_{\gamma\gamma}}{\sqrt{s}}\right)^{\tfrac13}\right]^b \left(\frac{m_{\gamma\gamma}}{\sqrt{s}}\right)^{a}.
\end{equation}
The \roots{8} and \roots{13} backgrounds were described by separate choices of $a$, $b$ and normalisation, $n_b$.
ATLAS modelled the experimental resolution of the signal shape with a double-sided crystal ball (DSCB) function\cite{atlas,Aaboud:2016tru}. In their combined analysis\cite{atlas}, ATLAS accounted for a substantial width by promoting DSCB parameters to functions of the mass and width of the \digamma. Because the details of this treatment were not published, we picked a simpler ansatz for the signal shape and experimental resolution. The \digamma signal was described by a Breit-Wigner or a Gaussian with a width equal to the ATLAS diphoton resolution if the width, $\ensuremath{\Gamma_\digamma}$, was narrower than the ATLAS diphoton resolution, $\sigma$,
\begin{equation}\label{eq:signal_dist}
p_s(m_{\gamma\gamma}) \propto
\begin{dcases}
\frac{1}{(m_{\gamma\gamma}^2 - \ensuremath{m_\digamma}^2)^2 + \ensuremath{\Gamma_\digamma}^2 \ensuremath{m_\digamma}^2} & \ensuremath{\Gamma_\digamma} > \sigma \\
e^{-\frac{(m_{\gamma\gamma} - \ensuremath{m_\digamma})^2}{2\sigma}} & \ensuremath{\Gamma_\digamma} \le \sigma
\end{dcases},
\end{equation}
and normalisation factors, $n_s$, at \s{8} and \s{13}. The ATLAS diphoton resolution was modelled by a linear function of \digamma mass,
\begin{equation}
\sigma \approx 6 \cdot 10^{-3} \ensuremath{m_\digamma} + 0.8 \ensuremath{\,\text{GeV}},
\end{equation}
motivated by information in \refcite{atlas} that it changes from $2\ensuremath{\,\text{GeV}}$ to $13\ensuremath{\,\text{GeV}}$ between masses of $200\ensuremath{\,\text{GeV}}$ and $2\ensuremath{\,\text{TeV}}$. We described the normalisation factors by an expected number of signal events in the \s{13} $3.2\ensuremath{/\text{fb}}$ search and scaling factors, reflecting the decreased cross section at \roots{8} and different integrated luminosities. Thus the SM ansatz and \digamma ansatz were described by six and four parameters, respectively.
We scraped the ATLAS bin counts, $\{n\}$, and bin edges from \refcite{Aaboud:2016tru,ATLAS:2016eeo}. To convert the distributions into expected numbers of events per bin, $\{\lambda\}$, we used analytic integration. Our likelihood function was simply a product of Poisson distributions,
\begin{equation}\label{eq:like}
\mathcal{L} = \prod_i \frac{\lambda_i^{n_i} e^{-\lambda_i}}{n_i!},
\end{equation}
for the expected and observed number of events in each bin in the \s{8} and \s{13} searches. In total, our calculations required about 6 million calls of our likelihood function.
\newcommand{\range}[2]{\ensuremath{\left[#1, #2\right]}}
\begin{table}
\begin{ruledtabular}
\begin{tabular}{@{\extracolsep{2pt}}llll@{}}
\multicolumn{2}{c}{SM} & \multicolumn{2}{c}{\digamma}\\
\cline{1-2} \cline{3-4}
Parameter & Prior & Parameter & Prior\\
\cline{1-2} \cline{3-4}
$a_{13}$, $a_8$ & Flat, \range{-25}{25} & $\ensuremath{m_\digamma} / {1\ensuremath{\,\text{TeV}}} $ & Log, \range{0.2}{2}\\
$b_{13}$, $b_8$ & Flat, \range{-25}{25} & $\alpha$ & Log, \range{5\cdot10^{-6}}{0.1}\\
$n_{b13}$ & Log, \range{5\cdot10^3}{10^4} & $\sigma_{13/8}$ & Flat, \range{2.5}{5}\\
$n_{b8}$ & Log, \range{2\cdot10^4}{3\cdot10^4} & $n_{s13}$ & Log, \range{5}{200}\\
\end{tabular}
\end{ruledtabular}
\caption{\label{tab:priors}Priors for the SM ansatz and \digamma resonance. Subscript numbers refer to \s{8} and \s{13} $3.2\ensuremath{/\text{fb}}$, e.g.,\xspace $n_{b13}$ refers to the number of expected background events at \s{13} with $3.2\ensuremath{/\text{fb}}$.
}
\end{table}
The priors for the parameters were the final ingredients. We picked logarithmic priors for $\ensuremath{m_\digamma}$ and $\alpha\equiv\ensuremath{\Gamma_\digamma} / \ensuremath{m_\digamma}$ that matched the ranges in the ATLAS search: a mass between $200\ensuremath{\,\text{GeV}}$ and $2000\ensuremath{\,\text{GeV}}$, and $\alpha$ between $5\cdot10^{-6}$ and $0.1$. The latter range spans the narrow-width approximation (NWA) to substantial widths. We picked a logarithmic prior for the expected number of \digamma events in the \s{13} $3.2\ensuremath{/\text{fb}}$ search, $n_{s13}$, between $5$ and $200$ events, which reflected the \digamma models anticipated in the experimental search. The number of expected events in the \s{8} search was modelled by a scaling factor between the \s{13} and \s{8} production cross sections, $\sigma_{13/8}$. We picked a linear prior, i.e.,\xspace $p(\sigma_{13/8}) = \text{const.}$, between $2.5$ (corresponding to a light quark initial state) and $5$ (corresponding to gluon fusion). We, of course, included a factor reflecting the decreased integrated luminosity. Since the models were composite (that is, we considered the SM and SM plus \digamma), we anticipated limited sensitivity to the priors for the SM background ansatz. We list all priors in \reftable{tab:priors} and discuss them further in \refsec{sec:priors}.
We supplied the likelihood and priors to \texttt{MultiNest}\footnote{We picked an evidence tolerance of $0.01$ and $1000$ live points per dimension.}\cite{Feroz:2007kg,Feroz:2008xx,2013arXiv1306.2144F}, which performs numerical integration via the nested sampling algorithm\cite{skilling2006,Skilling:2004}, returning a Bayesian evidence, e.g.,\xspace
\begin{equation}\label{eq:example}
\cond{\text{ATLAS data}}{\text{SM}} = \int \mathcal{L}(\bm x) \cdot \cond{\bm x}{\text{SM}} \, d\bm x,
\end{equation}
where the factors in the integrand are the likelihood and prior, respectively, as discussed above, and $\bm x$ denotes the SM ansatz parameter set (see \refapp{app:evidences}) for a complete expression). Finally, we calculated Bayes-factors, which are ratios of Bayesian evidences. For a clear picture of the changing relative plausibility of the \digamma versus the SM, we calculated Bayes-factors for \s{8}, \s{13} $3.2\ensuremath{/\text{fb}}$ and \s{13} $15.4\ensuremath{/\text{fb}}$ separately and combined. We validated our likelihood function and scanning by checking that we approximately reproduced the best-fit \digamma properties and significances reported by ATLAS; see \reffig{fig:best_fits} and \reffig{fig:pdf}. We found local significances of $3.9\sigma$ and $2.1\sigma$ at \s{13} $3.2\ensuremath{/\text{fb}}$ and \s{8}, respectively, and a combined local significance of $4.1\sigma$.\footnote{We assumed that the log-likelihood ratio was $\frac12\chi_1^2$-distributed \see{Cowan:2010js}.} We found best-fits for $(\ensuremath{m_\digamma}, \alpha)$ at about $(745\ensuremath{\,\text{GeV}}, 0.06)$, $(717\ensuremath{\,\text{GeV}}, 0.08)$ and $(739\ensuremath{\,\text{GeV}}, 0.09)$ at \s{13} $3.2\ensuremath{/\text{fb}}$, \s{8} and combined, respectively. We, furthermore, found $1.4\sigma$ tension between the preferred ratio of cross sections in \s{13} $3.2\ensuremath{/\text{fb}}$ and \s{8} data with mass and width fixed to their \s{13} $3.2\ensuremath{/\text{fb}}$ best-fits.
We found no indication that our signal model in \refeq{eq:signal_dist} was an inadequate or poor approximation to the unknown DSCB function.
\begin{figure}
\centering
\subfloat[Combined best-fit and spectrum at \roots{8}.]{
\includegraphics[width=0.8\linewidth]{digamma_8}
\label{fig:digamma_8}
}
\subfloat[Combined best-fit and spectrum at \roots{13} $3.2\ensuremath{/\text{fb}}$.]{
\includegraphics[width=0.8\linewidth]{digamma_13}
\label{fig:digamma_13}
}%
\caption{\label{fig:best_fits} Diphoton spectrum at \protect\subref{fig:digamma_8} \roots{8} and \protect\subref{fig:digamma_13} \roots{13} $3.2\ensuremath{/\text{fb}}$, along with best-fit background and signals.}
\end{figure}
We calculated that the \roots{8} data slightly disfavours the \digamma by a Bayes-factor of about $0.7$. At \roots{13} with $3.2\ensuremath{/\text{fb}}$, the story changes. The \digamma is favoured by about five with the region around $\ensuremath{m_\digamma} \approx 750\ensuremath{\,\text{GeV}}$ dominating, as expected. We calculated that the \s{8} and \s{13} $3.2\ensuremath{/\text{fb}}$ ATLAS data combined favoured the \digamma by about $7.7$. This is greater than a naive multiplication of Bayes-factors; the Bayes-factors cannot be combined by multiplication, as the evidences are dependent. All told, however, a combination of \s{8} and \s{13} with $15.4\ensuremath{/\text{fb}}$ disfavours the \digamma by about $0.7$. Whilst this is evidence against the \digamma, it is ``not worth more than a bare mention'' on the Jeffreys' scale. The evidences and Bayes-factors are summarised in \reftable{tab:evidences}.
\begin{table}
\begin{ruledtabular}
\begin{tabular}{lccc}
& \multicolumn{2}{c}{Evidence} \\
\cline{2-3}
Data & SM & SM + \digamma & Bayes-factor\\
\hline
ATLAS $8\ensuremath{\,\text{TeV}}$ $20.3\ensuremath{/\text{fb}}$ & $2.4\cdot10^{-64}$ & $1.7\cdot10^{-64}$ & $0.71$\\
ATLAS $13\ensuremath{\,\text{TeV}}$ $3.2\ensuremath{/\text{fb}}$ & $6.8\cdot10^{-64}$ & $3.1\cdot10^{-63}$ & $4.6$\\
ATLAS $13\ensuremath{\,\text{TeV}}$ $15.4\ensuremath{/\text{fb}}$ & $2.8\cdot10^{-87}$ & $7.2\cdot10^{-88}$ & $0.26$\\
\hline
$8\ensuremath{\,\text{TeV}}$ + $13\ensuremath{\,\text{TeV}}$ $3.2\ensuremath{/\text{fb}}$ & $1.7\cdot10^{-127}$ & $1.3\cdot10^{-126}$ & $7.7$\\
$8\ensuremath{\,\text{TeV}}$ + $13\ensuremath{\,\text{TeV}}$ $15.4\ensuremath{/\text{fb}}$ & $6.8\cdot10^{-151}$ & $5.0\cdot10^{-151}$ & $0.73$\\
\end{tabular}
\end{ruledtabular}
\caption{\label{tab:evidences}Evidences for the SM ansatz and \digamma resonance. Bayes-factors of greater than one indicate that the \digamma is
favoured.}
\end{table}
\begin{figure*}
\centering
\subfloat[ATLAS \roots{8} $20.3\ensuremath{/\text{fb}}$ data.]{
\setcounter{subfigure}{1}
\includegraphics[width=0.325\linewidth]{mass_width_pdf_8}
\label{fig:8}
}
\phantom{\includegraphics[width=0.325\linewidth]{mass_width_pdf_13_ichep}}
\subfloat[ATLAS \roots{13} $3.2\ensuremath{/\text{fb}}$ data.]{
\setcounter{subfigure}{2}
\includegraphics[width=0.325\linewidth]{mass_width_pdf_13}
\label{fig:13}
}
\subfloat[ATLAS \roots{13} $15.4\ensuremath{/\text{fb}}$ data.]{
\setcounter{subfigure}{4}
\includegraphics[width=0.325\linewidth]{mass_width_pdf_13_ichep}
\label{fig:13_ichep}
}
\subfloat[ATLAS \roots{8} $20.3\ensuremath{/\text{fb}}$ and \s{13} $3.2\ensuremath{/\text{fb}}$ data.]{
\setcounter{subfigure}{3}
\includegraphics[width=0.325\linewidth]{mass_width_pdf}
\label{fig:combined_pdf}
}
\subfloat[ATLAS \roots{8} $20.3\ensuremath{/\text{fb}}$ and \s{13} $15.4\ensuremath{/\text{fb}}$ data.]{
\setcounter{subfigure}{5}
\includegraphics[width=0.325\linewidth]{mass_width_pdf_ichep}
\label{fig:combined_pdf_ichep}
}
\caption{\label{fig:pdf} The posterior pdf (see \refcite{Fowlie:2016hew}) for
the mass and width of \digamma. Left-panels: \protect\subref{fig:8} ATLAS \roots{8} data,
\protect\subref{fig:13} \s{13} $3.2\ensuremath{/\text{fb}}$ data and \protect\subref{fig:combined_pdf} combined. Right panels:
\protect\subref{fig:13_ichep} \s{13} $15.4\ensuremath{/\text{fb}}$ data and \protect\subref{fig:combined_pdf_ichep} \s{8} and \s{13} $15.4\ensuremath{/\text{fb}}$ combined.}
\end{figure*}
\subsection{Prior sensitivity of the Bayes-factor}\label{sec:priors}
There is, of course, that thorny issue that a Bayes-factor is a functional of our priors for the models' parameters. This is most dangerous in cases in which an evidence is sensitive to the interval of a prior for a parameter. Consider a model with a single parameter $\lambda$ with a linear prior in the interval $\Delta \lambda$, and suppose that the likelihood is reasonable only in the interval $\sigma_\lambda$. The evidence would be approximately proportional to $\sigma_\lambda / \Delta \lambda$. For robust inference, it appears that we require prior information indicating a plausible interval. In fact, this is not strictly necessary. Such factors could cancel in a Bayes-factor, even for improper priors, i.e.,\xspace $\Delta \lambda\to\infty$. Furthermore, pragmatically, one could calculate an evidence for a sub-model with $\lambda$ in a region of interest (e.g.,\xspace a region searched by an experiment), $\delta\lambda$, learning whether such a sub-model was confirmed by data. This, in effect, shifts any $1/\Delta\lambda$ factors from an evidence into a prior for a sub-model, i.e.,\xspace
\begin{equation}
\prob{\text{sub-model}} = \prob{\text{model}} \cdot \cond{\lambda\in\delta\lambda}{\text{model}} \propto \delta \lambda / \Delta\lambda.
\end{equation}
The resulting Bayes-factor indicates the relative change in plausibility of the sub-model.
There may, furthermore, be cases in which our prior information cannot uniquely determine a distribution for a parameter upon an interval, i.e.,\xspace several choices of distribution may appear consistent with our prior information. This is problematic; the different choices may lead to different Bayes-factors. However, remarkably, in many cases particular ignorance about a parameter uniquely determines a distribution \see{jaynes2003probability}. If we are ignorant, e.g.,\xspace of the scale of $\lambda$, our prior should be invariant under $\lambda \to A\lambda$, leading to a logarithmic prior, $p(\log\lambda) = \text{const}$.\footnote{Whilst this distribution is improper, in practice we may be in possession of prior information that $\lambda$ cannot be arbitrarily great or small, though we are ignorant of its scale within that interval, leading to a proper distribution.}
Let us consider these issues in detail for all our priors. To quantify sensitivity, we recalculated Bayes-factors for the \s{8} + \s{13} $3.2\ensuremath{/\text{fb}}$ data:
\begin{itemize}
\item Fortunately, because the SM background ansatz was common to each model, any such factors originating from SM background parameters would vanish in a ratio of evidences, i.e.,\xspace a Bayes-factor. Similarly, we anticipated limited sensitivity to the shape of the priors for the SM background ansatz parameters.
\item Whilst this was not the case for the \digamma mass and width, narrower intervals than those in \reftable{tab:priors} would imply that we were in possession of prior information that precluded resonances in regions that were searched by ATLAS. Wider intervals would merely damage the plausibility of the \digamma model by diluting its evidence, though extreme $\alpha\equiv \ensuremath{m_\digamma} / \ensuremath{\Gamma_\digamma}$ could be implausible from the perspective of QFT. Our intervals matched the resonance masses and widths searched for by ATLAS, i.e.,\xspace we considered the change in plausibility of a \digamma resonance searched for by ATLAS. We picked logarithmic priors for the mass and width of the \digamma; anything else would imply prior information favouring particular scales in the intervals in \reftable{tab:priors}.
Nevertheless, with linear priors (which imply that prior information favoured the highest scales permissible) for the \digamma mass and $\alpha \equiv \ensuremath{m_\digamma} / \ensuremath{\Gamma_\digamma}$, we find a Bayes-factor of $39.7$ in favour of the \digamma.\footnote{This Bayes-factors was found by nested sampling. All further Bayes-factors were found by re-weighting an existing chain.} This is about $5$ times greater than previously; dominantly because a linear prior favoured the best-fitting $\alpha \approx 0.09$ by about $10$ relative to a logarithmic prior. From a theoretical perspective, however, a substantial width was, if anything, implausible relative to a narrow width; if we had any reliable prior information about the scale of the \digamma width, it was that should be narrow.
\item We picked a linear prior for the ratio of the \s{13} and \s{8} cross-sections. The interval spanned $2.5$, corresponding to a light-quark initial state, to $5$, corresponding to gluon fusion. This prior was motivated by knowledge about plausible production mechanisms; we were not ignorant of its scale and a light-quark initial state was a priori as plausible as gluon fusion. Nevertheless, we found that a logarithmic prior, which implies that prior information favoured a light-quark initial state, reduced the Bayes-factor by a factor of about $0.9$.
\item We picked a logarithmic prior between $5$ and $200$ for the number of \digamma signal events in the \s{13} $3.2\ensuremath{/\text{fb}}$ search. Whilst we were ignorant of scale within an interval, an extreme number of signal events was implausible from the perspective of QFT (an extreme cross section) and experiments (no other evidence for a resonance with an extreme cross section), resulting in an upper limit. Reducing the maximum number of events to $50$ increased the Bayes-factor by a factor of about $1.6$. Reducing the minimum number of events would, asymptotically, result in an SM background-like model and thus a Bayes-factor of $1$.
\end{itemize}
We stress that our interpretation is that our Bayes-factor is the change in relative plausibility of a \digamma resonance searched for by ATLAS, i.e.,\xspace with a mass in the interval $200\ensuremath{\,\text{GeV}}$ to $2\ensuremath{\,\text{TeV}}$ and $\alpha\equiv \ensuremath{m_\digamma} / \ensuremath{\Gamma_\digamma}$ in the interval $5\cdot 10^{-6}$ to $0.1$ (see \reftable{tab:priors}), and that the priors chosen reflected knowledge about cross sections and widths in QFT, and, in some cases, ignorance of scale or location. The Bayes-factor increased by a factor of about $5$ in the worst-case --- linear priors of the \digamma mass and width; however, it was a somewhat dishonest choice, since it implied that prior information favoured an appreciable \digamma mass and width. In other words, the prior information was sufficient to insure a weak dependence on choices of prior.
\subsection{Posterior distributions of \digamma properties}
The posterior pdfs for the \digamma properties were by-products of our calculations. Considering the combined \s{8} and \s{13} $3.2\ensuremath{/\text{fb}}$ data, for the mass, the posterior mean, median and mode were about $\ensuremath{m_\digamma} \approx 737\ensuremath{\,\text{GeV}}$, with a symmetric two-tailed $68\%$ credible region spanning $724\ensuremath{\,\text{GeV}}$ to $747\ensuremath{\,\text{GeV}}$. For the width, the posterior mean, median and mode in $\log \alpha$ differed, but spanned about $\alpha \approx 0.05$ to $0.08$, with a one-tailed $1\sigma$ ($2\sigma$) lower bound at $0.05$ ($0.004$). Finally, for the ratio of cross sections, the posterior pdf favoured $\sigma_{13/8}\approx 5$, corresponding to gluon fusion, but smaller ratios were permitted with a one-tailed $1\sigma$ ($2\sigma$) lower bound at $3.8$ ($2.8$). In other words, the posterior pdf favoured a mass of about $740\ensuremath{\,\text{GeV}}$, a large width and production by gluon fusion, as expected.
We show posterior pdf on the $(\ensuremath{m_\digamma}, \ensuremath{\Gamma_\digamma})$ plane in \reffig{fig:pdf} for \s{8}, \s{13} $3.2\ensuremath{/\text{fb}}$ and \s{13} $15.4\ensuremath{/\text{fb}}$ data separately and combined. The credible regions at $\ensuremath{m_\digamma}\lesssim500\ensuremath{\,\text{GeV}}$ (not shown) were vulnerable to digitisation errors in the data scraped from the low-mass region in \refcite{Aaboud:2016tru,ATLAS:2016eeo}.
Fortunately, that region ultimately contributed little to the evidences or our conclusions. The three prongs in the pdf at \s{8} in \reffig{fig:8} originated from deficits and excesses that surrounded the excess near $750\ensuremath{\,\text{GeV}}$ in \reffig{fig:digamma_8}. The pdf at \s{13} $3.2\ensuremath{/\text{fb}}$ in \reffig{fig:13} exhibited a single prong around $750\ensuremath{\,\text{GeV}}$ that narrowed once data was combined in \protect\reffig{fig:combined_pdf}.
\section{Conclusions}
Statistical anomalies near $750\ensuremath{\,\text{GeV}}$ in searches for diphoton resonances at the LHC resulted in a frenzy of model building. In the
official analyses, the data was investigated with frequentist techniques. To sidestep issues regarding the interpretation of significances, with Bayesian statistics, we calculated the change in plausibility of the \digamma resonance versus the SM in light of the ATLAS \s{8} $20.3\ensuremath{/\text{fb}}$, \s{13} $3.2\ensuremath{/\text{fb}}$ and \s{13} $15.4\ensuremath{/\text{fb}}$ diphoton searches. There was limited freedom in the choice of priors for the \digamma: we matched, where possible, the ranges of width and mass searched by ATLAS. Since the models were composite, we expected limited sensitivity to the priors for the SM background ansatz and found weak dependence on our choice of priors for the \digamma signal. Ideally, we would have combined ATLAS and CMS data, though the numerous categories
in the latter make a combination challenging. Our \digamma was a toy-model described by a simple Breit-Wigner with mass, width and cross section potentially within reach of the ATLAS search; we would find different Bayes-factors for a theory that precisely predicted the \digamma properties (though know of no such theory).
We calculated that the relative plausibility of the \digamma increased by about $7.7$ in light of the ATLAS data available at the height of the excitement. This should be contrasted with conclusions from frequentist analysis, e.g.,\xspace the probability of obtaining a test statistic as extreme as that observed in the \s{13} $3.2\ensuremath{/\text{fb}}$ search were the SM correct was about $0.02$ ($2.1\sigma$). This Bayes-factor was unimpressive, especially considering that e.g.,\xspace SM precision measurements could quash that preference and that a width of $\ensuremath{\Gamma_\digamma} \approx 0.06 \ensuremath{m_\digamma}$ was somewhat unexpected, a fact not reflected by our priors. On the other hand, a combination with CMS data could have increased the Bayes-factor, though there may have been tension in the preferred width. Considering all data, including \s{13} $15.4\ensuremath{/\text{fb}}$, the \digamma was disfavoured by about $0.7$. As well as aiding our understanding of the $750\ensuremath{\,\text{GeV}}$ anomaly, we hope our calculations serve as a proof of principle for Bayesian model comparison and parameter inference in future experimental searches.
|
\section{Introduction}
It has been argued that grasping, manipulation, and speech are among the most fundamental human
abilities unparalleled by animals \cite{BicchiKumarRamsete}. The importance of manipulation
and grasping is also evidenced by {\em ``the large fraction of the human motor cortex devoted
to manipulation and the number and sensitivity of mechanoreceptors in our palms
and fingertips"} \cite{OkamuraSurvey}.
Starting from these observations it is unsurprising that grasping and manipulation
generated significant interest in the robotics community, and robotic grasping emerged as
an autonomous area of research since the development of the first
robotic hands. Notwithstanding, robotic grasping continues to be one of the most
challenging areas in robotics. Its inherent difficulty is well captured by
Moravec's paradox, {\em ``It is comparatively easy
to make computers exhibit adult level performance on intelligence tests
or playing checkers, and difficult to give them the skills of a one-year old when
it comes to perception and dexterity"} \cite{Moravec1998}.\\
In robotics literature the terms {\em grasping} and {\em manipulation} are often
used interchangeably, but
they indicate different challenges. Grasping deals with the problem of restraining an object with a robotic hand or gripper.
One of the foremost problems in grasping is deciding where the robot end-effector should make contact with the
object to be grasped and which forces should be applied in order to ensure certain properties, like the ability to resist external
disturbances.
Manipulation is instead concerned with the problem of determining how a body can be moved by a robot.
The object to be moved may be firmly held by the robot (hence the connection to grasping), but this does
not necessarily need to be the case. For example, a robot may manipulate an object by pushing.
Manipulation problems can then be addressed using algorithms and models coming from the motion
planning domain \cite{LatombeMotionPlanning,ChosetBook,LaValleBook}. There is no
universally accepted definition for the expression {\em dexterous manipulation}. However,
many authors use the term to indicate manipulation problems where an object is
moved using robotic fingers, either with one or two hands \cite{OkamuraSurvey,DollarICAR2011}.\\
Given the continued advancements in the field, the complexity of the problem, and its practical importance, literature in robotic
grasping is vast and features some very detailed textbooks \cite{MasonBook,MurrayLiSastryManipulation} and
surveys \cite{PrattichizzoHandbook,BicchiKumar,OkamuraSurvey,ShimogaSurvey,BicchiTransactions2000}.
In this primer we focus on multi-fingered hands and we are particularly attentive to performance metrics and
test methods. Moreover, we are interested in methods and algorithms that can be applied
in the real world, as opposed to abstract formulations that are difficult
to use in practical applications. One should notice upfront that the divide
between theory and application is still significant in certain areas of grasping.
\section{Grasp modeling}
\label{sec:formalism}
As in virtually every area of robotics, it is necessary to develop and introduce an appropriate
mathematical formalism to study the grasping problem. It shall be noted that the level of
sophistication achieved in this area is remarkable and a complete discussion is beyond the scope
of this primer. The interested reader is for example referred to \cite{MurrayLiSastryManipulation} for an
in-depth discussion. We therefore limit the number of concepts we introduce to the minimum and we
assume the reader is familiar with basic concepts from classic mechanics.
Our goal is not to offer a complete treatise, but rather to introduce the minimum amount of
notation needed to formalize concepts and algorithms presented later on.
\subsection{Kinematics}
\label{sec:kinematics}
Throughout this primer we assume the object to be grasped is a rigid body $\mathcal{B}$.
Assuming an inertial reference frame $O_{XYZ}$ is given, we need to specify the pose (i.e., position
and orientation) of $\mathcal{B}$ with respect to $O_{XYZ}$. The classic approach is to
rigidly attach a frame to $\mathcal{B}$ (commonly called
{\em body frame} or {\em moving frame}) and to then describe the pose of this body-fixed frame.
In principle, this body-fixed frame could be attached anywhere to $\mathcal{B}$, but in the grasping literature, it is customary to assume this frame is placed at the center of mass of $\mathcal{B}$.
Let $B_{xyz}$ be the body frame attached to $\mathcal{B}$ (see Fig. \ref{fig:bodyframe}).
\begin{figure}[htb]
\centering
\includegraphics[width=5cm]{BodyFrame.pdf}
\caption{World frame and body frame}
\label{fig:bodyframe}
\end{figure}
Three parameters are used to describe the position of the body frame and
three parameters are needed for its orientation. Therefore, the pose of $\mathcal{B}$ is
fully described by a 6-dimensional vector. Different representations can be used for
the orientation (see for example \cite{CraigBook,LaValleBook} for in depth discussions about different
parametrizations for the space of rotations). Generically, the 6-dimensional pose vector\footnote{Throughout this document we assume vectors are column vectors.} for $\mathcal{B}$ is defined as $\mathbf{u}$ (so-called {\em operational space} \cite{OrioloBook}),
\[
\mathbf{u} =
\left[
\begin{array}{c}
\mathbf{p}\\
\mathbf{\Phi}
\end{array}
\right]
\]
\noindent where $\mathbf{p}=[p_x~p_y~p_z]^T$ represents the translational position and $\mathbf{\Phi}$ represents the orientation. The orientation can be expressed through Euler angles\footnote{One could in fact also use a non minimal representation, e.g., quaternions to describe
the orientation. In this case $\mathbf{q}$ would then have 7 dimensions. This option will not be explored.}.
The vector $\mathbf{u}$ is a more compact representation
of the homogeneous transformation matrix
\[
_B^OT =
\left[
\begin{array}{cccc}
& & & p_x\\
& _B^OR & & p_y\\
& & & p_z\\
0 & 0 & 0 & 1
\end{array}
\right]
\]
where $_B^OR$ is the rotation matrix describing $\mathcal{B}$'s orientation with respect to $O_{XYZ}$
and $(p_x,p_y,p_x)$ define the position of the origin of $B_{xyz}$ with respect to $O_{XYZ}$.
It is important to note that one can define suitable mappings between
$\bf u$ and $_B^OT$. Therefore depending on the context, either of these forms will be used.
\\
To study grasping problems a compact representation for the pose of the robotic hand $\mathcal{H}$ is needed, too. The
most convenient model uses a vector of $m$ joint displacements to fully specify the hand placement in space.
Note that $m$ is the overall number of degrees of freedom, which includes both the degrees of freedom
of the arm and of the hand.
It is customary to indicate this $m$-dimensional vector as $\mathbf{q}$ and to
indicate its components as $q_i$, i.e., $\mathbf{q} = [q_1,q_2,\dots,q_m]^T$.
The vector $\mathbf{q}$ is said to belong to the {\em joint space}.
In certain situations it will be necessary to consider just the degrees of freedom of the hand.
Therefore it is convenient to assume $m=a+h$, where $a$ is the number of degrees of freedom
of the arm and $h$ is the number of degrees of freedom of the hand. The vector $\mathbf{q}$
can then be decomposed as $\mathbf{q} =[\mathbf{q}_a^T ~\mathbf{q}_h^T]^T$, where $\mathbf{q}_a$ includes
just the degrees of freedom of the arm and $\mathbf{q}_h$ includes the degrees of freedom of the hand (see
Fig. \ref{fig:RobotWithHand}).
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\linewidth]{./RobotWithHand}
\caption{Vector $\mathbf{q}$ includes both the degrees of freedom defining the arm pose $\mathbf{q}_a$ and the hand configuration
$\mathbf{q}_h$. }
\label{fig:RobotWithHand}
\end{figure}
Given $\mathbf{q}$, forward kinematics can be used to determine the position of any point on the robot arm or hand
(for example using the Denavit-Hartenberg method \cite{CraigBook}).
In grasping applications, the pose of some points are particularly relevant. These are the palm of the
robotic hand, and the tips of the fingers. Once fixed frames are assigned to the these points, their
pose can also be determined through forward kinematics. The hand orientation, in particular, is often
assumed to be the orientation of the frame attached to the palm. This information is important when parametrizing grasps.\\
Grasping relies on contacts between the hand $\mathcal{H}$ and the body $\mathcal{B}$. Although in general
a contact between $\mathcal{H}$ and $\mathcal{B}$ may occur in arbitrarily shaped regions, it is
customary to only model situations where the contact occurs at a single point. More complex scenarios,
like for example contact along a segment, can be dealt with by using an appropriate set of contact points.
The $i$-th contact point is indicated with the letter $\mathbf{c}_i$ and its three dimensional coordinates
can be expressed with respect to $O_{XYZ}$ or $B_{xyz}$.
In the following, different models to characterize friction at a contact point will be
presented. To this end, it is convenient to assign frames to contact points, too. We assume that
for every contact point $\mathbf{c}_i$, the tangent plane to $\mathcal{B}$ is defined. The frame assigned
to $\mathbf{c}_i$ has one axis pointing inside $\mathcal{B}$ along the normal to the tangent plane,
and the other two axis lying on the tangent plane. The normal axis will be indicated as $\mathbf{n}_i$ (see Figure \ref{fig:contact}).
Let $n_c$ be the number of contact points.
The role of contact points is critical as a conduit for transmitting forces and moments on $\mathcal{B}$ by the robotic hand.
\begin{figure}[htb]
\centering
\includegraphics[width=6cm]{Contact.pdf}
\caption{The $i$-th finger of hand $\mathcal{H}$ makes contact with $\mathcal{B}$ at point $\mathbf{c}_i$. At the contact point
a frame is defined with the axis $\mathbf{n}_i$ normal to the plane tangent to $\mathcal{B}$ at point $\mathbf{c}_i$. The other two axes
are on the tangent plane.}
\label{fig:contact}
\end{figure}
\subsection{Motion and forces}
In grasping applications it is common to rely on
{\em screw theory} to describe the motion of rigid objects. Three concepts will be introduced, namely, {\em screw}, {\em twist}, and
{\em wrench}.
Screws, twists, and wrenches can be introduced through differential geometry
and a wealth of results can therefore be used to characterize them.
The book by Murray et al. \cite{MurrayLiSastryManipulation} offers a very thorough
analysis of these concepts based on exponential representations. However, in the following
we stick to a more classic approach. It is important to note that while these concepts
have been widely used, definitions and representations are not always consistent.\\
A result known since 1830 and normally referred to as Chasles's theorem states that any spatial displacement of
a rigid body can be decomposed as a rotation about an axis followed by a translation along the same axis.
A displacement could then be described with three parameters, namely an axis $\mathbf{a}$, an angle $\theta$, and a displacement
$d$. However, it is common not to provide $d$, but to rather specify a pitch $h$, defined as the ratio
between $d$ and $\theta$. Once the angle $\theta$ is given, the corresponding displacement can be computed as $\theta h$.
Collectively, $\mathbf{a}$, $h$, and $\theta$ define a {\em screw} (also called {\em screw motion}).
At this point it should also be clear why the name {\em screw} is used to describe this geometric object\footnote{In \cite{MasonBook}
screws, twists, and related objects are introduced using slightly different definitions. While eventually one can reconcile these
differing representations, we here stick to the definitions given in \cite{MurrayLiSastryManipulation} and \cite{PrattichizzoHandbook}.}.
Note that a screw with infinite pitch corresponds to a pure translation, whereas a screw with 0 pitch
is a pure rotation. \\
\noindent A {\em twist} is the representation of the velocity of a body $\mathcal{B}$ and is a six dimensional vector
\[
\mathbf{t} =
\left[
\begin{array}{c}
\mathbf{v}\\
{\boldsymbol \omega}
\end{array}
\right]
\]
where $\bf v$ is the translational velocity and $\boldsymbol \omega$ is the rotational velocity.
Both $\bf v$ and $\boldsymbol \omega$ are vectors with three components.
\\
Finally, a {\em wrench} is used to model a force $\bf f$ applied to $\bf p$ on the surface of $\mathcal{B}$. The point $\bf f$ may be a contact point in case the force is applied by one
of the fingers, but also a different point in case the force is due to an external load.
The wrench is also
a six dimensional vector including both the force $\bf f$ and the moment $\bf m$ generated by $\bf f$.
\[
\mathbf{w} =
\left[
\begin{array}{c}
\mathbf{f}\\
\bf m
\end{array}
\right].
\]
Note that both twists and wrenches can be described in different reference frames and not only with respect to $O_{XYZ}$.
In particular, they can be expressed in $B_{xyz}$ or a frame associated with a contact point.
Moreover, since $\bf w$ includes a moment component $\bf m$, when defining a wrench it is
necessary to specify the point about which $\bf m$ is defined.
The concept of a
wrench is fundamental to formalize the notion of a force-closed grasp, and to study
how forces are exchanged between the robotic hand and the object through the contact points. If a multi-fingered robotic hand $\mathcal{H}$
makes contact with $\mathcal{B}$ at multiple points, there will be multiple sources of interaction forces, each with its own associated wrench. The mapping between forces and wrenches
is linear, and if all wrenches are expressed with respect to the same frame, the wrenches can be superimposed by addition. Wrench models will be discussed once appropriate friction models are described in the next section.
\subsection{Friction}
Contact points are usually characterized with three different types of friction \cite{BicchiKumar,BurdickHandbook}.
It should be noted up front that this taxonomy is a useful approximation, but does not model all possible contacts
and is defined only for rigid body objects.
\begin{itemize}
\item {\em Frictionless} contact point. In this case the finger can only exert a force along ${\bf n}_i$,
the normal to the surface at the contact point. In particular, the force is directed towards the
inside of the object, i.e., the finger can push but not pull. Therefore, frictionless contact
points introduce a constraint on the force of the type $f_n \geq 0$, where $f_n$ is the
force along the normal.
\item {\em Frictional} contact point. In this case, the finger can exert a force along
the normal to the surface but also along the tangent plane.
Different friction models could be used, but the most commonly used is based on Coulomb's law.
Coulomb's law states that no relative motion between the finger and the object happens as long as
the tangential component of the force $f_t$ satisfies the following inequality
\begin{equation}
f_t \leq \mu f_n
\label{eq:coulomb}
\end{equation}
where $f_n$ is the normal component and $\mu$ is the static friction coefficient characterizing the
coupling between the finger and the object. According to this model, Eq. \eqref{eq:coulomb} defines
the largest tangential force that can be applied. Since this force can be arbitrarily oriented
in the tangent plane, a common way to graphically depict this situation is given by the so-called
{\em friction cone}. The friction cone has its apex at the contact point and is oriented along the
normal to the tangent plane. Its aperture is $\tan^{-1} \mu$ (see Figure \ref{fig:FrictionCone}).
Often, the tangential component $f_t$ is expressed in terms of its
two projections $f_1$ and $f_2$ along the axis defined
on the tangent plane. In this case we can then write $f_t = \sqrt{f_1^2+f_2^2}$ and Eq. \ref{eq:coulomb}
can then be rewritten as $ \sqrt{f_1^2+f_2^2} \leq \mu f_n$. Moreover, the inequality $f_n \geq 0$
must still hold. These conditions define the friction cone, $FC$, at contact point ${\bf p}_i$ (see also Fig. \ref{fig:FrictionCone}):
\[
\textrm{FC}({\bf p}_i) = \left\{ [f_1~f_2~f_n]^T \in \mathbb{R}^3 ~|~\sqrt{f_1^2+f_2^2} \leq \mu f_n \wedge f_n \geq 0\right\}
\]
\begin{figure}[htb]
\centering
\includegraphics[width=5cm]{FrictionCone.pdf}
\caption{Friction cone}
\label{fig:FrictionCone}
\end{figure}
Sliding between the finger and the object should not occur as long as the total force, $F$, exerted by finger is
inside the cone. Note that in many practical situations the cone is approximated with
a pyramid \cite{FerraryCanny,ZhuTransactions2003}.
\item {\em Soft} contact. In this case the finger can not only exert normal and tangential
forces, but also a torsional moment about the normal to the contact point. Note that in this case
we usually talk about a contact area rather than a contact point. For a soft contact the same
inequalities specified for a frictional contact point must be satisfied. Moreover, indicating
with $\tau$ the intensity of the torsional moment, then $\tau \leq \gamma f_n$.
\end{itemize}
Although there is agreement about the provided taxonomy, nomenclature is not standard.
For example, frictionless contact points are also called {\em point contact without friction},
whereas frictional contact points are also referred to as {\em hard finger} \cite{PrattichizzoHandbook}.\\
The type of contact determines then the number of independent wrenches that can be exerted on the
body $\mathcal{B}$ through the contact. For a frictionless contact point, just one wrench
can be exerted, and this is the result of the force applied
along the normal to the tangent on the contact point.
For a frictional contact point, three independent wrenches can be applied. One of these
is again due to the force along the normal to the tangent plane, whereas the other two are
due to the forces exerted on
the tangent plane. Finally, for the soft contact, four independent wrenches can be applied. These
are due to three forces (same as the frictional point), and one torsional moment about the
normal to the contact
point. When the contact is more complex, the number of independent wrenches can still
be determined, though this is more complicated \cite{SalisburyPhDThesis}.\\
In robotic grasping it is customary to assume that the same friction model applies to each of the contact points (e.g., Figure \ref{fig:frictioncones}), but this is not necessarily the case.
\begin{figure}
\centering
\includegraphics[width=6cm]{MultiFinger.png}
\caption{Multifingered grasp with frictional contact points. The same friction model is assumed at each contact point
and therefore friction cones are associated with each finger-object contact.}
\label{fig:frictioncones}
\end{figure}
\subsection{Multi-fingered grasps}
\label{sec:multifingered}
Given a friction model, a multi-fingered grasp can be formally specified using
a representation to determine the type of restraint
obtained by the grasp. When there are $n_c$ contact points
between the fingers and the body, just specifying their coordinates ${\bf c}_1,\dots,{\bf c}_{n_c}$
is not sufficient, because it is also necessary to consider the forces exchanged
through the contacts and how the fingers can be moved to exert these forces.
This information is encoded in two matrices, namely the hand Jacobian ${\bf J}$ and the
grasp map ${\bf G}$. The hand Jacobian ${\bf J}$ is the classic Jacobian
matrix studied in robot kinematics. Let ${\bf q}_h$ be the $h$ degrees of freedom of the hand.
Then the position of the fingertips can be determined thorough
forward kinematics. Forward kinematics is then a function $\textrm{FK}: \mathbb{R}^h \rightarrow \mathbb{R}^{6n_c}$
mapping the $h$ degrees of freedom ${\bf q}_h$ into the frames attached to the fingertips\footnote{In
this case we assumed that the frame attached to each fingertip is represented using the minimal
representation $\bf u$ in operational space.}.
Note that we assumed
there are $n_c$ fingers, to use the same symbol utilized for the number of contact points.
The hand Jacobian J can be mapped from joint velocities q into frame velocities using
the relationship ${\bf J}{\dot{ \bf q}}$.
While $\bf J$ is used to express relationships involving velocities, the grasp matrix $\bf G$
defines how forces exerted at the contact points are mapped into wrenches. $\bf G$
is specified using a friction model that includes the forces and moments that are applied to the object.
The relationship between friction models and wrenches is then discussed in the following.
\begin{enumerate}
\item Consider a frictionless contact point ${\bf c}_1$ and let ${\bf f}_1$ be the force
applied. Because the point is frictionless, ${\bf f}_1$ can only be applied along the normal
to the tangent plane at ${\bf c}_1$ and the wrench is
\[
{\bf w}_1 = \left[
\begin{array}{c}
{\bf f}_1\\
({\bf c}_1-{\bf o}) \times {\bf f}_1)
\end{array}
\right]
\]
where $\bf o$ is the point about which moments are computed. This is typically the origin of $B_{xyz}$.
In section \ref{sec:kinematics} we assumed that a frame is associated to every
contact point with an axis $\bf n$ normal to the tangent point defined at the contact point.
Then, recalling this frame we can rewrite ${\bf w}_1$ as follows, where we evidence the dependence
on the force intensity $f_1$.
\[
{\bf w}_1 = \left[
\begin{array}{c}
0\\
0\\
1\\
-p_y\\
p_x\\
0
\end{array}
\right]
f_1
\]
where $p_x$ and $p_y$ are the coordinates of ${\bf c}_1$ with respect to frame $B_{xyz}$.
In this second form, the wrench matrix is written as the product of a {\em wrench basis} with
a scalar.
\item Next, consider a frictional contact point ${\bf c}_2$. Three forces can be exerted, one
along the normal and the other two on the tangent plane.
Let these forces be ${\bf f}_1$,${\bf f}_2$ and ${\bf f}_3$, where the first two are on
the tangent plane and the third is the normal force. These three forces are not independent
from each other but, rather, are related by the friction coefficient, as
implied by the Coulomb inequality (Eq. \ref{eq:coulomb}).
Stated differently, their sum must lie inside the friction cone FC.
Applying a reasoning similar
to the one we used for the frictionless contact point, we can then write an expression for the
wrench:
\[
{\bf w}_2 = \left[
\begin{array}{ccc}
{\bf f}_1 & {\bf f}_2& {\bf f}_3\\
({\bf c}_1-{\bf o}) \times {\bf f}_1 & ({\bf c}_2-{\bf o}) \times {\bf f}_2 & ({\bf c}_3-{\bf o}) \times {\bf f}_3
\end{array}
\right].
\]
This expression can also be rewritten as the product between a wrench basis and a vector
with the three intensities $f_1,f_2$ and $f_3$:
\[
{\bf w}_2 = \left[
\begin{array}{ccc}
1 & 0& 0\\
0 & 1& 0\\
0 & 0& 1\\
0 & p_z & -p_y\\
-p_z & 0 & p_x\\
p_y & -p_x & 0
\end{array}
\right]
\left[
\begin{array}{c}
f_1\\
f_2\\
f_3
\end{array}
\right]
\]
\item Finally, consider a soft contact point ${\bf c}_3$. In this case the wrench not only includes
three forces as for the frictional point, but also one torsional moment. Let $f_1,f_2$ and $f_3$ be the
intensities of the forces and $m$ the intensity of the moment. The wrench and
the wrench basis are:
\[
{\bf w}_3 =
\left[
\begin{array}{cccc}
{\bf f}_1 & {\bf f}_2& {\bf f}_3 & 0\\
({\bf c}_1-{\bf o}) \times {\bf f}_1 & ({\bf c}_2-{\bf o}) \times {\bf f}_2 & ({\bf c}_3-{\bf o}) \times {\bf f}_3 & {\bf m}
\end{array}
\right]
\]
\[
=
\left[
\begin{array}{cccc}
1 & 0& 0 & 0\\
0 & 1& 0 & 0\\
0 & 0& 1 & 0\\
0 & p_z & -p_y & 0\\
-p_z & 0 & p_x & 0 \\
p_y & -p_x & 0 & 1
\end{array}
\right]
\left[
\begin{array}{c}
f_1\\
f_2\\
f_3\\
\tau
\end{array}
\right]
\]
\end{enumerate}
Given a set of $n_c$ contact points, each defines its own wrench matrix and basis.
The grasp matrix ${\bf G}$
is the matrix obtained by composing the various wrench bases side by side. It is therefore
a matrix with six rows and the number of columns is $n=n_l + 3n_f + 4n_s$, where $n_l$ is the
number of frictionless contact points, $n_f$ is the number of frictional contact points, and
$n_s$ is the number of soft contact points. The grasp matrix $\bf G$ can be seen
as a linear map $G:\mathbb{R}^n \rightarrow \mathbb{R}^6$ mapping the intensity of forces and moments
into the overall wrench acting on $\mathcal{B}$. To exploit this matrix
notation, it is then convenient to use an $n$-dimensional vector ${\bf l}$ including all
the force and moment components. In this case, the resulting wrench can then be written as
$\bf Gl$, where we will implicitly assume that $\bf l$ satisfies the constraints imposed
by the generalized friction code FC.\\
Equipped with this notation, a multifingered grasp $\mathcal{G}$
is then fully specified by ${\bf J}$, $\bf G$,
and the friction coefficients needed to define the friction cone FC. Note that $\bf G$ implicitly provides the contact points ${\bf c}_i$, so these do not need to be specified separately.
\subsection{Form and force closure}
The concepts of form and force closure
emerged as the two main characterizations for a grasp.
Both ideas find their origin in mechanism design and predate robotics and grasping.
In fact, they are more than 100 years old (see \cite{Reuleaux} for a translation
of the original book). However,
grasping problems have stimulated research aiming to produce accurate and efficient tests
to determine if a grasp achieves any of these two properties.
Before dwelling upon the details, it is useful to anticipate
that every form closure grasp is also a force closure grasp, but not viceversa.
\subsubsection{Form closure}
Informally speaking, form closure is obtained when the robotic hand surrounds the object to be grasped
in such a way that the object cannot move without colliding with the hand.
A different way to express the same concept is saying that contacts prevent all possible motions of $\mathcal{B}$,
including infinitesimal motions \cite{TrinkleICRA1992}.
This idea can be formalized
in different ways. For example, one could say that form closure occurs when the object is located at a singular
configuration in its configuration space \cite{MasonBook}. Alternatively, form closure can be analytically defined
as follows \cite{PrattichizzoHandbook}. Let $\mathbf{u}$ be the configuration of the object and $\mathbf{q}$
be configuration of the hand. Assuming there are $n_c$ contact points between the hand and the object, a
gap function $\psi_i(\mathbf{u},\mathbf{q})$, $1\leq i \leq n_c$ can be defined for each of the contact points.
The gap function gives the distance between the hand and the object at the contact point. By definition,
$\psi_i(\mathbf{u},\mathbf{q})=0$ when the contact occurs, becomes positive when the contact breaks
(the finger moves away from the object), and becomes negative when there is penetration (the finger enters
the object). Based on this definition, a grasp achieves form closure if and only if the following implication
holds:
\begin{equation}
\psi( \mathbf{u}+d\mathbf{u},\mathbf{q}) \geq 0 \Rightarrow d\mathbf{u}=0
\label{eq:form_closure}
\end{equation}
where inequalities between vectors have to be interpreted component-wise.
It is easy to see that this equation is equivalent to say that the object cannot move ($d\mathbf{u}=0$), i.e., that it is located
at an isolated point in its configuration space. Note that this definition does not rely on friction, i.e., form closure is
achievable irrespectively of friction. This class of grasps is therefore also indicated as {\em frictionless grasps}.\\
\noindent {\bf Form closure tests.} A practical problem of great importance is knowing whether a certain
set of contacts achieves form closure or not. A related problem is ranking various grasps leading to
force closure from the best to the worst, according to some performance metric, like for example resistance to
external disturbances. Numerous criteria have been developed for the case where planar objects are being grasped
(see e.g., \cite{VanDerStappenIJRR2000}). In this paper we however concentrate exclusively on the three dimensional
scenario.
A result formulated by Somov more than 100 years old states that at least seven contact points
are necessary to obtain form closure
with an arbitrarily shaped object with six degrees of freedom, i.e., for the
general problem of objects moving in three dimensions.
Because this is a necessary condition, this simple criterion provides an easy test to quickly determine
that all grasps relying on six or less points cannot obtain form closure on a generic object.
But the condition is not sufficient,
and grasps with seven contact points can still be insufficient to achieve form closure. More sophisticated
tests are therefore necessary.
However, if one restricts the class of objects being grasped, then different results hold.
For example, seven contacts points are necessary and sufficient for
the special case of a three dimensional polyhedron \cite{Papadimitriou1990}.\\
Form closure tests of different order can be developed and
are introduced as follows.
Consider the first order Taylor
expansion of Eq. \ref{eq:form_closure} around $(\mathbf{u},\mathbf{q})$. This can be written
as (higher order terms have been omitted for brevity)
\[
\psi( \mathbf{u}+d\mathbf{u},\mathbf{q}) =
\psi( \mathbf{u},\mathbf{q}) + \left. \frac{\partial \psi({\bf u},{\bf q})}{\partial {\bf u}} \right|_{({\bf u},{\bf q})}d{\bf u} =
\left. \frac{\partial \psi({\bf u},{\bf q})}{\partial {\bf u}} \right|_{({\bf u},{\bf q})}d{\bf u}
\]
where we used the fact that by definition $\psi( \mathbf{u},\mathbf{q})=0$ at a contact configuration. This expansion can then be
related to Eq. \ref{eq:form_closure} as follows. If there exist $d {\bf u}$ such that
$\frac{\partial \psi({\bf u},{\bf q})}{\partial {\bf u}}d{\bf u}$ has at least one positive component, then
the grasp does not achieve form closure. In fact, in this case $d {\bf u}$ identifies a direction along which
the object can be moved causing one of the components of the gap functions to become positive, hence breaking one of the contacts.
Alternatively, one can consider the negated equivalent\footnote{That is, rather than using
$p \Rightarrow q$ we consider the logically equivalent implication $\neg q \Rightarrow \neg p$} of the implication given in Eq. \ref{eq:form_closure}. This means
that for every choice of $d{\bf u}$ different from 0, then $\frac{\partial \psi({\bf u},{\bf q})}{\partial {\bf u}}d{\bf u}$
has at least one negative component. In this case the grasp then achieves form closure.\\
When neither of these two conditions hold, the first order Taylor expansion is insufficient to classify the grasp.
In this case a Taylor expansion of order two should be considered. If using the second order Taylor expansion,
neither of the two conditions hold, then a third order Taylor expansion should be used, and so on.
For this reason,
in grasping literature it is common to read about form closure tests of order $n$, with reference to the considered
order for the Taylor expansion. Rimon and Burdick provide numerous results discussing the importance of
higher order analysis in restraint analysis \cite{RimonBurdickTRO98,RimonBurdickTRO98Bis,RimonICRA1996}.\\
The first order form closure test leads to a linear programming program whose solution not only determines
whether a set of contact points achieve form closure, but also provides a quantitative measure that can be
used to rank different grasps.
This idea was first introduced by Trinkle in \cite{TrinkleICRA1992,TrinkleTransactions1992}
and the reader is referred to those papers for details.
\subsubsection{Force closure}
Intuitively, a grasp achieves force closure if it can be maintained in spite of external forces acting on the
restrained object. This idea can be rigorously formalized, but it should be immediately noticed that
there exist slightly different definitions for this concept.\\
A concise definition of force closure can be given using the grasp matrix $\bf G$ \cite{MurrayLiSastryManipulation}. A grasp achieves force closure if it can balance
any external wrench ${\bf w}_e$ applied to the object. Algebraically, this means that
for each external wrench ${\bf w}_e$, there exist a vector ${\bf l} \in \textrm{FC}$ such that
\begin{equation}\label{eq:force_closure_condition}
{\bf Gl} = -{\bf w}_e
\end{equation}
From this algebraic definition it is possible to immediately derive algebraic conditions
to determine if a given grasp achieves force closure. For example, a necessary condition
is that ${\textrm{rank}}({\bf G})=6$. However this condition is just necessary but not
sufficient, because ${\bf l}$ is restricted to be in the friction cone FC, and this
is a subset of $\mathbb{R}^n$. A necessary and sufficient condition is that ${\bf G}( \textrm{FC})=\mathbb{R}^6$,
where ${\bf G}( \textrm{FC})$ is the set obtained by applying the grasp map to every point in the friction cone
FC. Another useful characterization used in practice relates the force closure to convex hulls, which is particularly useful when introducing grasp quality metrics.
As formerly stated, the friction cone is often approximated with a regular pyramid (see Fig. \ref{fig:cone_approx}).
\begin{figure}
\centering
\includegraphics[width=5cm]{fcone.pdf}
\caption{Friction cones are commonly approximated with regular pyramids}
\label{fig:cone_approx}
\end{figure}
Using this approximation, each contact
force lying within the friction cone can be written as
a non-negative combination of forces along the boundary of the friction cone, i.e.,
$\mathbf{f}_i = \sum_{j=1}^{k} \alpha_{i,j}\mathbf{f}_{i,j}$ with $\alpha_{i,j} \geq 0$.
Based on this discretization the wrench generated by the $i$th contact
force can then be written as
\[
\mathbf{w}_i=\left[
\begin{array}{c}
\mathbf{f}_i\\
(\mathbf{c}_i - \mathbf{o}) \times \mathbf{f}_i
\end{array}
\right] =
\left[
\begin{array}{c}
\sum_{j=1}^{k} \alpha_{i,j}\mathbf{f}_{i,j}\\
(\mathbf{c}_i - \mathbf{o}) \times \sum_{j=1}^{k} \alpha_{i,j}\mathbf{f}_{i,j}
\end{array}
\right]=
\]
\[
= \sum_{j=1}^{k} \alpha_{i,j}
\left[
\begin{array}{c}
\mathbf{f}_{i,j}\\
(\mathbf{c}_i - \mathbf{o}) \times \mathbf{f}_{i,j}
\end{array}
\right] = \sum_{j=1}^{k}\alpha_{i,j} \mathbf{w}_{i,j}
\]
where each of the $\mathbf{w}_{i,j}$ is called {\em elementary wrench}. The $kn_c$ elementary wrenches
$\mathbf{w}_{i,j}$ can be arranged in a $6\times kn_c$ matrix $\mathcal{G}$ that approximates the {\em grasp matrix}.
As per Eq. \eqref{eq:force_closure_condition}, the grasp achieves force closure if each disturbance wrench $\mathbf{w}_e$
can be countered by forces inside the friction cones.
That is to say, the grasp matrix $\mathcal{G}$
positively spans $\mathbb{R}^6$. In this case, it is known that an equivalent condition for force closure
is that the convex hull of all the $kn_c$ elementary wrenches $\mathbf{w}_{i,j}$ includes the origin
(see e.g., \cite{MurrayLiSastryManipulation}).
\subsubsection{Grasp quality}
From the definitions of form and force closure, it is evident that these
conditions could be satisfied by multiple different grasps.
The availability of multiple solutions to the same problem immediately
raises the issue of solution selection. In particular, it
is natural to look for a {\em grasp quality metric} to rank different
solutions and then pick the best one. This concept has been heavily
investigated for grasps achieving force closure, and in the following
discussion we limit our discussion to this case. The reader is referred to
\cite{GraspAuro2014} for a comprehensive review of grasp metrics. \\
The most commonly used criterion is the so-called Ferrari-Canny
metric \cite{FerraryCanny}. For sake of correctness it should be noticed that similar ideas were already explored in \cite{Kirkpatrick}.
The underlying idea is that given two grasps and an external load,
one should prefer the grasp that can resist the load by exerting the
smallest effort. In general, since the external load is not known, the criterion should be applied in a worst-case manner, i.e.,
by picking the grasp that can be resist {\em any} external load with
the smallest effort. The concept of effort needs to be formalized as well.
In \cite{FerraryCanny} two different definitions are put forward.
\noindent Any wrench exerted on $\mathcal{B}$ through the $n_c$ contacts can be scaled to
belong to the set
\begin{equation}\label{eq:wlinf}
W_{L_{\infty}}=\textrm{CH}\left(\bigoplus_{i=1}^{n_c} \{\mathbf{w}_{i,1}\dots \mathbf{w}_{i,k} \}\right)
\end{equation}
where $\oplus$ represents the Minkovski sum and CH stands for {\em Convex Hull}. From an operative point of view
it is worth observing that the inner Minkovski sum in Eq. \ref{eq:wlinf} gives
a set of finite elements, and therefore $W_{L_{\infty}}$ is the convex hull of
a finite set of elements in $\mathbb{R}^6$.
The grasp metric $Q_{\infty}$ is then defined
as the distance of the closest facet of $W_{L_{\infty}}$ from the origin, i.e.,
\begin{equation}\label{eq:qinfydefinition}
Q_{\infty} = \inf_{\mathbf{x}\in \partial W_{L_{\infty}}} ||\mathbf{x}||_2
\end{equation}
where $\partial W_{L_{\infty}}$ indicates the boundary of $W_{L_{\infty}}$, i.e., the union of its facets.
Physically,
$Q_{\infty}$ aims to minimize the maximum force exerted by any of the $m$ fingers
to resist an arbitrary external wrench.
Alternatively, one could aim at minimizing the sum of forces exerted by all fingers.
To this end, consider
\begin{equation}\label{eq:wl1}
W_{L_1}=\textrm{CH}\left(\bigcup_{i=1}^{n_c} \{\mathbf{w}_{i,1}\dots \mathbf{w}_{i,k} \}\right).
\end{equation}
Similarly to $Q_{\infty}$, the grasp metric $Q_{1}$ is then defined
as the distance of the closest facet of $W_{L_1}$ from the origin
\begin{equation}\label{eq:q1definition}
Q_{1} = \inf_{\mathbf{x}\in \partial W_{L_{1}}} ||\mathbf{x}||_2.
\end{equation}
Both these ideas can be easily formalized considering the convex hull
of the friction cones at the contact points.
Moreover, their computation is also related to the computation of the convex
hull.
In particular, if the
origin of $\mathbb{R}^6$ is not inside the convex hull, the grasp does
not achieve force closure. However, when the origin is inside, the
distance between the origin and the closest face of convex hull
is chosen as the grasp quality, and grasps with larger values are
preferred. Indeed, grasps with larger values of this metric can resist
external loads using wrenches with smaller magnitude.
Various relevant aspects should be mentioned. First, practically speaking, the effort definition used is defined as the magnitude of the largest wrench generated by any finger.
Next, to increase the computational efficiency,
the friction cone is discretized using a representation based
on a regular pyramid. A tradeoff between accuracy and speed evidently
emerges, with more edges on the pyramid leading to a more accurate result,
but also requiring more computational time. As discussed later on, in many
instances the value of the grasp quality metric is used to drive the
grasp planning process, and the quality is then iteratively computed
multiple times. To this end, significant efforts have been devoted to
accelerate the computation of the quality metric. The QuickHull
algorithm \cite{QuickHullArticle} had emerged as the de-facto standard in this
area thanks to its efficiency and its widespread availability. To compute the
metric, one starts computing the convex hull in six dimensions and then
computes the distance between the origin and the hull. However, it
is easy to realize that this approach is wasteful inasmuch as it computes
the full convex hull. In reality, just the face closest to the origin is
needed to compute the metric. Capitalizing on this observation,
we recently introduced the Partial Quick Hull algorithm \cite{Liu2}
that greatly accelerates the computation of these metrics.
Despite its popularity, the Ferrari-Canny metric suffers from various drawbacks.
For example, it is not scale invariant since it depends on the choice of the point about
which torques are computed and it does not take into account the geometry of the
object. Moreover, it is in general too conservative, i.e., it scores a grasp based on
its ability to resist an arbitrary disturbance wrench, whereas in practice the set
of realistic disturbance wrenches is much smaller than the set of all disturbance
wrenches. Starting from these observations, Strandberg and Wahlberg proposed
an alternative metric based on the ability of a grasp to resist wrenches occurring
in practice \cite{strandberg2006}. Their method builds upon the concept of
object wrench space defined by Pollard \cite{Pollard1994}. The object wrench space
is the set of wrenches that can be exerted on an object through the action of a
disturbance force. Evidently this set is different from the set of all wrenches, since
for example is does not contain wrenches with just a torque component but no force
component. Therefore, in \cite{strandberg2006}, it is proposed to score a grasp on
the ability to resist these wrenches. This revised metric overcomes most of the limitations
encountered by the Ferrari-Canny metric, but in practice it has been rarely used
because its computation is demanding. In fact, Borst in \cite{borst2006}
proposed a method to approximate its value, but thse alternate
algorithms did not receive much attention either. Recently, Liu and Carpin \cite{LiuCASE2015}
proposed an exact algorithm to compute the metric proposed in \cite{strandberg2006}.
The algorithm is based on the principle of partial quick hull computation
and builds upon some of the ideas developed in \cite{Liu2}.
\subsection{Grasp parametrizations}
\label{sec:grasp_parametrizations}
A growing number of grasping methods rely on the assumption that a set of grasps is
computed \textit{a priori} and can be queried at run time for retrieval (see Section \ref{sec:algorithms}).
To implement this functionality it is therefore necessary to chose an appropriate representation
of a grasp to simplify storage and retrieval. Note that this cannot just be ${\bf q}$ or ${\bf q}_a$
because it is necessary to somehow also consider the spatial relationship between the hand
and the object to be grasped without necessarily storing the object itself.
An ideal parametrization should be as general as possible and not be tied to a specific robot hand.
Also, one should notice that although these representations are known as {\em grasp} parametrizations,
they should not be confused with the formerly introduced concept of multi-fingered grasp (see
Section \ref{sec:multifingered}).
In comparison, they describe the pose of the hand as it approaches an object before performing the actual grasp.\\
We present two different parametrizations proposed in grasping literature.
Berenson and colleagues \cite{BerensonHumanoids2007} propose to parametrize each
grasp with four parameters and assume a body frame has been attached to palm of the robot hand.
Although the specifics of this frame are not explicitly spelled out, it is convenient to assume that
one of the axes of this frame, say ${\bf n}_p$, is orthogonal to the palm\footnote{It is then
implicitly assumed that it is possible to identify the direction orthogonal to the palm. This may be
easy for certain robotic hands, e.g., the Barret Hand, but less obvious for other devices, so this choice
is not unambiguous. }. The four parameters are the following:
\begin{itemize}
\item the direction of approach of the hand, i.e., ${\bf n}_p$;
\item a point $\bf p$ on the surface of $\mathcal{B}$ towards which the hand is moving;
\item the roll $\theta$ of the hand about ${\bf n}_p$;
\item the preshape of the hand, i.e., ${\bf q}_h$.
\end{itemize}
Note that ${\bf n}_p$ and $\theta$ can be computed from ${\bf q}_a$ through forward kinematics, but also observe
that ${\bf q}_a$ includes additional information that is not necessary for this parametrization.\\
Another parametrization is proposed in \cite{AllenICRA2004} by Pelossof and colleagues. Although specifically
designed for the Barrett hand, it can be generalized and extended for other hands, too.
The Barrett arm and hand has $m=10$ degrees of freedom, with $a=6$ and $h=4$. Note that for the Barrett Hand
three of the four values in ${\bf q}_h$ define the opening of the fingers, whereas the last parameter
gives the spread.
In order to reduce the number of parameters, the authors make some assumptions. The first is
that the palm should be parallel to the surface of the object, i.e., ${\bf n}_p$ should
be orthogonal to the surface. Moreover, because the hand will move along ${\bf n}_p$, the distance
between the object and the hand does not need to be explicitly considered.
It is also assumed that the hand starts with its fingers fully
open, i.e., three of the four values in ${\bf q}_h$ are fixed upfront. Based on these assumptions,
four parameters can be used to parametrize a grasp:
\begin{itemize}
\item two values, say $\eta$ and $\zeta$, to specify the position of the palm of the hand. Note that
because it is assumed that ${\bf n}_p$ is orthogonal to the surface (i.e., it is constrained
to be on a plane) and the distance to the object is not relevant, two parameters suffice.
\item the roll $\theta$ of the hand about ${\bf n}_p$;
\item the spread of the fingers, i.e., the only free parameter in ${\bf q}_h$.
\end{itemize}
A few comments about these two alternatives are needed. The first remark is about the size of the
parametrization itself. While they both nominally rely on four parameters, it is clear that the dimensionality
of the first one is much higher, because both ${\bf n}_p$ and $\bf p$ each require three parameters to be specified.
In addition, the first parametrization offers the possibility to fully specify the preshape with ${\bf q}_h$ and
this comes at the cost of an increase in its dimension. Both representations include the roll angle
about ${\bf n}_p$. In \cite{BerensonHumanoids2007} the authors maintain that their parametrization is
superior to the one proposed in \cite{AllenICRA2004}. In particular they argue that it is advantageous
to relax the constraint that ${\bf n}_p$ is orthogonal to the surface. This claim is not controversial
because indeed it allows to explore a wider range of possibilities.
Furthermore, the parametrization proposed in \cite{BerensonHumanoids2007} offers more
flexibility in specifying the preshape. It is clear however that these benefits come at the cost of increasing
the dimension of the parametrization and potentially of the solution search space, and therefore depending
on the grasp planning algorithm used, the cost may overcome the benefits.
\section{Grasp Planning Algorithms}
\label{sec:algorithms}
Grasp planning algorithms have been investigated since the very introduction of robotic grippers.
Through the years more and more complex problems and situations have been considered.
In terms
of robot actuators, initial studies mostly dealt with parallel jaw grippers, but nowadays a significant amount of
research is devoted to multi-fingered hands with three or more fingers and numerous degrees of freedom.
The complexity of objects considered in grasp problems has also evolved and the focus is now on
grasping objects used in everyday activities by humans, as opposed to early investigations
concerned with planar objects with elementary shapes. At the same time, algorithms considering different degrees of
uncertainty (e.g., uncertainty in the location or shape of the object to be grasped, or imprecise robot motions, etc.)
have been developed. Coherently with the goals of this primer, we will not consider the
problem of grasping a 2D object\footnote{Technically speaking, 2D objects do not exist. This expression
is however commonly used to indicate flat objects in which one of the dimensions is much smaller than the others.}.
In terms of hands, our focus is on multi-fingered hands, but we
will nevertheless consider certain algorithms that have been demonstrated on parallel-jaw grippers
if their applicability to multi-fingered hands is evident. \newline
Research in grasping and existing methods could be categorized in many different ways. We developed
a taxonomy presented in Table \ref{tab:taxonomy} and we structure this
section accordingly\footnote{References
given in Table \ref{tab:taxonomy} are meant to be just a sampling of the vast literature in the field.
They provide starting points to delve into this area.}.
This is just one way of characterizing work in this area, and given the variety of algorithms
we are aware that some will not clearly fit into it. Nevertheless,
for the class of problems relevant for this effort, we found that this table offers a useful
way to look at how the grasping problem can be approached.\\
\begin{table}[htb]
\centering
\begin{tabular}{l|l|l|l|l|l|l}
\hline
\multicolumn{7}{|c|}{{\bf Grasping algorithms}} \\ \hline
\multicolumn{3}{|p{6cm}|}{{\bf Model Based}\newline The geometry or shape of the object to be grasped is given a-priori. Planning happens without integrating sensor information.} & \multicolumn{4}{|p{8cm}|}{{\bf Model-less}\newline Shape is not known a-priori and the robot uses sensors to acquire information about the object to be grasped.} \\ \hline
\multicolumn{2}{|p{4cm}|}{{\bf Geometric only}} & \multicolumn{1}{|p{2cm}|}{{\bf Data Driven}} & \multicolumn{2}{|p{4cm}|}{{\bf Vision based}} & \multicolumn{2}{|p{4cm}|}{{\bf Touch based}}\\ \hline
\multicolumn{1}{|p{2cm}|}{{\bf Consider Object Only} \newline \cite{AllenICRA2007}\cite{LiSastryJournalRA}\cite{AllenICRA2003} \cite{AllenICRA2004}\cite{PonceTransactions1995}\cite{ZhuTransactions2003}} & \multicolumn{1}{|p{2cm}|}{{\bf Consider Object and Environment}\newline \cite{BerensonHumanoids2007}} & \cite{ColumbiaGraspDatabase,AllenDataDrivenGrasping}& \multicolumn{1}{|p{2cm}|}{{\bf Construct model} \newline \cite{collet:2009}\cite{DuneIROS2008}\cite{GallardoICAR} \cite{HubnerICRA2008}\cite{ChinellatoAutRob2008} } &\multicolumn{1}{|p{2cm}|}{{\bf Operate in feature space}\newline \cite{KragicIJRR2012}\cite{SaxenaNgIJRR}}& \multicolumn{1}{|p{2cm}|}{{\bf Construct model} \newline \cite{BierbaumTactile2008}\cite{BierbaumTactile}\cite{GrupenTRO2010}}& \multicolumn{1}{|p{2cm}|}{{\bf Operate in feature space } \cite{AllenLearningGraspStability}\cite{AllenICRA2011}} \\ \hline
\end{tabular}
\caption{A possible taxonomy for grasping algorithms with references to methods falling in each category.
Note that not all referenced papers are fully covered in this primer.}
\label{tab:taxonomy}
\end{table}
A final word on the use of machine learning algorithms is warranted. In general, the use of machine learning
in robotics (including robotic grasping) is increasing at a fast pace.
We believe three reasons
justify this expansion. First, dramatic improvements in computational power and major algorithmic
breakthroughs have produced a wide variety of techniques that can now efficiently solve
hard computational problems. Second, the necessity of developing robots operating in
poorly structured environments or subject to various sources of noise calls for the use
of algorithms and representations inherently capable of dealing with probabilistic uncertainty.
In this area, too, machine learning offers a wide variety of tools. Finally, the ever increasing
availability of inexpensive sensors producing vast amounts of data enables the development
of machine learning algorithms that extract models from the data.
The interplay between machine learning and
robotic grasping appears to be unavoidable in the future and is evidenced by the selected papers
we discuss in the following. The increasing interplay between machine learning and robotics
is well captured by the emerging expression {\em cloud robotics}, whereby the cloud is
used to enable learning algorithms operating over large, remotely stored data sets \cite{GoldbergTASE2015}.
\subsection{Model-based Algorithms}
In this subsection we discuss grasp planning methods sharing the common assumption that a geometric
model for the object to be grasped is provided in advance. Many different approaches
exploiting this knowledge have been explored. At one end of the spectrum, we find methods relying
on geometric models only. Typically, these are models for the hand, the object to be grasped, and the
environment. We name this class {\em geometric} algorithms. At the other end of the spectrum, we find algorithms that exploit also additional knowledge,
like past experience acquired while grasping, pre-computed grasps, or data collected while executing
the grasp plan itself. We dub this second class {\em data driven} methods, and they will
be discussed in the second part.\\
The overarching computational question is to determine an appropriate set of contact points
on the surface of the object being grasped in order to achieve a certain objective, like e.g.,
form closure or force closure. Many parameters need to be given to fully specify an instance
of this problem, like the shape of the object (e.g., polyhedral or with curvature), friction
model, number of fingers of the hand, and so on. Therefore, many different algorithms have
been proposed in the robotics literature. Before dwelling upon the details, it is worthwhile
recalling that we are here interested in planning (or synthesis) algorithms, as opposed
to analysis algorithms. The former aim to produce a set of contact points satisfying certain
constraints, whereas the latter are intended to determine if a given set of points satisfy the
constraints or not.
\subsubsection{Geometric algorithms for grasp point calculation}
Some common features are shared by algorithms that plan grasping points for an arbitrary
number of fingers on three dimensional objects. A common pattern among these
approaches is to formulate the problem as a search of the grasp configuration case
with the intent of optimizing a certain objective function. According to this paradigm,
then, grasp point calculation is the result of an optimization problem. Despite the
apparent simplicity of this paradigm, a certain level of mathematical sophistication
is required to formulate meaningful objective functions.\\
Li and Sastry \cite{LiSastryJournalRA} propose a method to compute a set of
grasp points that aims at
maximizing some grasp quality indexes discussed or introduced in the paper.
Their method works for 3D objects
but has the limitation of being applicable only for objects with a polyhedral boundary.
The approach relies on the screw theory extensively discussed in \cite{MurrayLiSastryManipulation}.
Three different quality indexes are considered. The first two are similar to other
measures discussed in literature, whereas the third one is novel and developed by
the authors. The first one, indicated as $\delta$, is defined as the smallest
singular value of the grasp matrix $\bf G$. The second one, indicated as $\nu$, is
the so-called volume in wrench space and is defined by the following integral
\[
\nu = \int_{G(B^n_1\cap \textrm{FC})} d\nu,
\]
where $B^n_1$ is the unit ball in $\mathbb{R}^n$, $ \textrm{FC}$ is the friction cone, and
$G(B^n_1\cap \textrm{FC})$ is the image of the set through the map defined by the grasping
matrix $\bf G$. Both $\delta$ and $\nu$ are known to have desirable properties in terms of
grasp performance measurement, but Li and Sastry outline drawbacks that render them inappropriate
to drive the grasp-point calculation process. In particular, $\delta$ is not invariant
under a change of the torque origin, and $\nu$ does not capture force closure (which the
authors call stability). In other words, $\nu$ in itself does not discriminate between
grasps achieving closure, but can be used to rank grasps only after they have been
shown to obtain closure. To overcome these limitations, the authors introduce
a third quality measure that is {\em task oriented}. This term indicates that the measure
aims at identifying the ability of a grasp to generate wrenches relevant to the task
being performed. To be precise, tasks are modeled by ellipsoids in the wrench space.
The intuition is that this ellipsoid should capture the type of wrenches necessary
to successfully complete a given task, and to make their case the authors
compare the wrenches that should be exerted to complete a {\em peg-in-hole} task
with the wrenches to perform some writing with a pencil. Evidently, these two tasks
require different wrenches and then should be associated with different ellipsoids.
A task ellipsoid is defined starting from a vector ${\bf a} \in \mathbb{R}^6$ and
a $6\times 6$ positive definite matrix $\bf Q$:
\[
A_{\beta} = \{ {\bf w }\in \mathbb{R}^6, {\bf w}^T{\bf Qw} + {\bf w}^T{\bf a} \leq \beta^2 \}.
\]
For a given $A_{\beta}$ the task oriented quality measure $\mu$ is then defined as
\[
\mu(G) = \sup \{\beta \geq 0, \textrm{such that } A_{\beta} \subset G(B^n_1\cap \textrm{FC})\}.
\]
when ${\bf a}=0$ and a slightly different expression holds when ${\bf a}\neq 0$ (see \cite{LiSastryJournalRA} for details).
Task ellipsoids have pros and cons. The advantage is that their definition is related
to the task and aims at enforcing wrenches that are useful for the task.
The disadvantage is that, as acknowledged by the authors,
their definition extensively relies on experience and in-depth knowledge of the task
being performed.
With the objective of maximizing $\mu(\cdot)$, the grasp point calculation problem
is then formulated as a constrained optimization problem. For a given multifingered
grasp $\mathcal{G}$, let $\bar{\mu}$ be the value of the quality measure computed
on the map induced by $\bf G$. Then the problem is
\[
\max_{{\bf c}_1,\dots,{\bf c}_{n_c}} \bar{\mu}
\]
subject to the constraints that the various ${\bf c}_i$ are placed on the boundaries
of the object $\mathcal{B}$ being grasped. The main problem with this formulation is that
this optimization problem needs to be solved numerically and its solution is computationally hard.
These problems are partially mitigated by the method we propose next.\\
Zhu and Wang \cite{ZhuTransactions2003} propose the first algorithm capable of synthesizing a force
closure grasp on a 3D object with an arbitrary number of contact points and
applicable to 3D objects with piecewise smooth surface.
The idea is very simple yet elegant. A numerical test to determine if a set of contacts
achieve force closure is presented. Under mild assumptions the function defining the test
is differentiable with respect to the grasp configuration. Because such derivatives
can be computed exactly, the space of grasps is then searched using a gradient descent
and a grasp achieving force closure is eventually determined. This method then sidesteps
the heavy numerical optimization problem evidenced in the previous discussion.
The novel force closure test proposed by the authors is based on the so-called $Q$ norm $||\cdot||_Q$,
a special norm different from the commonly used $L_2$ norm. Starting from this special norm (see
\cite{ZhuTransactions2003} for details), two distances between a point $\bf p$ and a convex
polyhedron $A$ are defined:
\[
d^+_Q({\bf p},A) = \min_{{\bf a}\in A}||{\bf a}-{\bf p}||_Q
\]
\[
d^-_Q({\bf p},A) = \min_{{\bf a}\in \partial A}||{\bf a}-{\bf p}||_Q.
\]
\noindent For appropriate choices of $Q$ the two distances $d^+_Q$ and $d^-_Q$ can
be computed in closed form and are differentiable almost everywhere.
Let $\mathcal{G}$ be a multifingered grasp with $n_c$ contact points ${\bf c}_1,\dots,{\bf c}_{n_c}$.
Note that each ${\bf c}_i$ is uniquely defined once we know $\bf q$, therefore
we could also write $\mathcal{G}(\bf q)$ or ${\bf c}_i({\bf q})$.
In addition, because of this dependency when a derivative with respect to the contact point
${\bf c}_i$ is considered, one can also consider the derivative with respect to $\bf q$.
Let ${\bf w}_i$ be the wrench associated with every contact point.
Zhu and Wang prove a novel result showing that a necessary condition to achieve force closure is
$d^+_Q(0,\textrm{co}\{{\bf w}_1,\dots,{\bf w}_n\})$, where $\textrm{co}$ indicates the convex hull.
This condition is necessary but not sufficient. However a necessary and sufficient condition can
be determined using $d^-_Q$ (details are omitted because this derivation is rather involved).
This test is in essence a test on the sign of $d^-_Q$, i.e., the grasp achieves force closure if
and only if $d^-_Q<0$.
The grasping planning algorithm is then implemented by the following iterative procedure
where $dQ(q)$ is a case-by-case function defined as $dQ^+$ or $dQ^-$.
\begin{algorithm}
Let $\mathcal{G}_0$ be the grasp associated with ${\bf q}_0$\;
\Repeat{$d_Q({\bf q}_k < 0)$}{
${\bf q}_{k+1} = {\bf q}_k - \lambda \frac{d Q({\bf q})}{d{\bf q}}$\;
}
\caption{Grasp planning based on $Q$ distance}
\end{algorithm}
We conclude this discussion noting that in addition to the objective of obtaining force closure,
it is also possible to include constraints to the grasp, like for example
imposing contacts in certain areas of the object being grasped. Moreover it is also possible
to fix the location of some contacts (say the contact positions of two fingers) and let the
algorithm plan the contacts for the remaining $n_c-2$ fingers. Finally, the terminating condition
could be replaced to let the algorithm find a local minimum rather than stopping as soon as $Q$
becomes negative. In that case the algorithm not only determines a force-closure grasp, but
it also optimizes the grasp in terms of resistance offered to an external load.
The authors conclude the paper showing how the algorithm can be used to plan contact points for various complex objects.\\
When models for the object and the hand are available, one way to plan a grasp is to rely on physical simulation to determine
if a candidate grasp is good or not.
This approach falls in the {\em geometric only} category because it is purely based on geometric properties and does
not rely on otherwise collected data.
This line of research has been pioneered in Allen's research group \cite{AllenICRA2004} and relies on the subtle
distinction that grasp planning is performed using a simplified object model, but grasp simulation uses a
highly accurate geometric and physical model.
Within this framework, a key intuition is that the great number of possibilities emerging
in grasp planning can be reduced by relying on the concept of {\em pre-shape}, i.e., a fixed number
of general purpose hand configurations. This approach is then coupled with their
GraspIt! simulation environment \cite{Graspit} and aims to be largely hand-independent, as opposed to methods
specifically tailored to a certain class of hands or specific robot hardware. Pre-shapes simplify the complexity on
the hand side because it restricts the search space. However, the approach performs further simplification also
on the object side, because the geometric model is not assumed to be accurate, but rather an approximation based
on a set of primitive shapes, i.e., spheres, boxes, cylinders, and cones. Each primitive shape is associated with
a set of grasping strategies. A grasp strategy defines the pre-shape and is parametrized in a way similar to
those presented in Section \ref{sec:grasp_parametrizations}, i.e., in addition to the spread of the finger, the
position and the orientation of the hand are specified, too. At this point each grasp is tested for quality
using the Ferrari-Canny metric \cite{FerraryCanny}. The evaluation does not take place in the physical world
but is rather performed inside GraspIt! assuming a Barrett Hand mounted on a Puma 560 arm.
The simulation system moves the hand along the approach direction
and eventually closes the fingers, assuming no impediment to the motion emerged. At that point the metric
is computed and the quality of the grasp is returned.
The same principles are further developed in \cite{AllenICRA2007}. The main difference is that rather than relying
on the elementary shapes described above, the authors propose to use a tree of superquadrics (this method
shares many similarities with the system presented in \cite{HubnerICRA2008} that we discuss in Section \ref{sec:modelless_vision}).
The simple idea behind this method is that a set of superquadrics
better approximates the the shape of a complex object than a single superquadric. There is a rich literature in
computer graphics describing how to ``best" approximate a single shape with a superquadric. There also exist
methods to compute hierarchical superquadric decompositions where a superquadric is recursively
split until a certain error threshold is met. The authors maintain that this method
performs poorly when used for grasping applications and propose instead to set a priori a number $k$ of
elementary superquadrics, i.e., superquadrics that cannot be further split. The choice of this parameter
is critical and according to the authors it depends on the dexterity of the hand being used (but not on
the class of objects being grasped). For the Barrett Hand this number has been determined to be $k=6$.
As for the previous method, a set of tentative canonical grasp is considered for each shape, with the advantage
that in this case all shapes are superquadrics of different size because of the decomposition. After
the set of tentative grasps is generated, the approach resembles \cite{AllenICRA2004}, i.e.,
it uses GraspIt! for simulation and the Ferrari-Canny metric for evaluation purposes.\\
\subsubsection{Data driven algorithms}
Goldfeder and co-workers proposed a data driven method \cite{ColumbiaGraspDatabase,AllenDataDrivenGrasping} that relies on the
Columbia Grasp Database. The database, described in the same paper, includes a large number of geometric models
of objects taken from a publically available shape benchmark. Each object is appropriately scaled to fit into a
human or robotic hand, and is further associated with grasp relevant information. In particular, for every object
the database stores three types of information. 1) a {\em pre-grasp}, i.e., the pose of the hand right before the hand contacts the object;
2) the grasp itself, implementing a form closure grasp; 3) quality metrics for the grasp, as defined in \cite{FerraryCanny}.
Note that grasps were computed for two versions of the Barrett hand (with different friction), and a human hand.
The paper further explains how these grasps were automatically pre-computed using the GraspIt! software and other
preexisting grasping methods, but these details are immaterial to the data-driven grasping method they propose.
The algorithm
takes as input $\alpha$, a geometric model of the object to grasp, and returns a set $R$ of tentative grasps for the object. To
compute $R$, the algorithm scans the database for the $k$ most similar objects, according to a shape similarity metric.
Next, for each of the $k$ similar objects, it aligns $\alpha$ with the object and tries to apply the associated
precomputed grasps stored in the database. The final grasp is evaluated using GraspIt! and if the computed quality
exceeds a given quality threshold $\tau$, it is included in $R$, otherwise it is discarded. \\
The method critically depends on three parameters. The first is the shape similarity metric used to
identify similar objects. In the paper the authors use the $L^2$ distance between Zernike descriptors \cite{Zernike}
combined with a $k$ nearest neighbor approach.
The second parameter is $k$, i.e., the number of nearest neighbors to consider. Evidently, in picking $k$ there is
a tradeoff between speed and accuracy. In the papers the authors fix $k=5$. Finally, the threshold $\tau$ is chosen
to decide if a computed grasp shall be included in the result set $R$ or not.
The authors report that the method is capable of computing grasps in about 20 seconds (computational times were given in 2009).
The performance of the algorithm is compared with a ground truth baseline giving the best match in the database. The analysis
provided by the authors show that the selected shape-recall method offers comparable performance. The authors also compare
the data-driven method with another algorithm they formerly proposed \cite{CiocarlieEigenGrasps} and show that the data-driven approach is
faster and more accurate.\\
We conclude by observing that this method shares many features with \cite{BalaguerIJHR2011}, although in \cite{BalaguerIJHR2011} we assumed
the geometric model was acquired on the fly, and then, according to our taxonomy, we implemented a model-less algorithm.
\subsection{Model-less Algorithms}
A significant body of recent research on grasping considers the situation where a geometric
model of the object to be grasped is not available upfront. This problem domain is considered to be
much more challenging, but it also has a much broader applicability
than model-based algorithms and presents the opportunity for broader application sets.
To remedy the lack
of an \textit{a priori} model, these grasping algorithms rely on sensor data acquired at run time\footnote{We
purposefully ignore sensorless algorithms like \cite{GoldbergAlgorithmica1993,GoldbergAlgorithmica2000} as their scope and
applicability falls outside the scope of this primer.}.
Typically, these include vision and tactile sensing. Information acquired at run time
could then be used to build an approximate or incomplete object model on the fly,
or to compute a grasping strategy directly from raw data without attempting to produce a model.
One could retort that an algorithm that builds an approximate model on-the-fly is not
model-less anymore. Here we consider an algorithm model-less if it does not assume the availability
of a model \textit{a priori}. Then, according to this definition, an algorithm building a model on-the-fly
is still model-less.
It goes without saying that the use of machine learning algorithms is prevalent in this area.
\subsubsection{Model-less algorithms based on vision}
\label{sec:modelless_vision}
In this section we present two algorithms capable of synthesizing a grasp for an unknown three dimensional
object based on visual input. While numerous methods were developed for the special case of planar objects,
the three dimensional case is significantly harder and has been considered in its full generality only
recently.\\
\noindent {\bf Methods that construct a model}\\
When a camera system is available, one way to solve the grasping problem is to construct
a model of the object and then plan a grasp based on the recovered model. Given the copious amount
of work in shape reconstruction available in the computer vision literature, it is evident that
many different methods could be engineered around this idea. Therefore, in this primer we will only
analyze a few selected ones. The reader is also referred to \cite{KragicRAS2010} for a more detailed
discussion of this topic.\\
In \cite{DuneIROS2008}, the author proposes a shape estimation method aimed for implementing a {\em ``One click grasping tool"}.
The goal is to extract a geometric model of an object on-the-fly, so that a robotic gripper can then approach and restrain it.
The envisioned application is an assistive device for disabled people, like a robotic arm mounted on a wheel chair.
The underlying assumption is that the object to be grasped is convex (or roughly convex). The shape of the
object is then approximated using a quadric and the parameters of the quadric are identified from multiple views
after the contour of the object has been extracted from the image plane. In order to take images from
different vantage points, it is assumed a camera is mounted on the robotic arm and the authors describe
a method to move the camera around the object (explore) with the objective of minimizing the estimation
error. Once the estimation process for the parameters of the quadric has been completed, the object
is grasped by aligning the gripper with the minimal dimension while at the same time being orthogonal
to the largest axis.\\
Besides methods trying to estimate shapes starting from images or sequence of images, there is also
work attempting to achieve the same goal starting from point clouds or stereoimages.
The method presented in \cite{HubnerICRA2008} is similar to the reference \cite{DuneIROS2008} discussed above,
in the sense that the objective is to find a primitive shape offering a good approximation of the object to
be grasped. The difference is that in \cite{HubnerICRA2008} the input is a point cloud and that rather than
using a quadric, the authors opt for a hierarchical decomposition based on minimum volume bounding boxes (MVBB).
The rationale behind this choice is that boxes are
simple shapes to reason about, but at the same time offer poor approximations of complex shapes resembling
everyday objects that a robot may want to grasp. Therefore a hierarchical decomposition is used in order to
retain the advantages and mitigate the problems. It is worth noting that these hierarchical decompositions
have been extensively used in collision detection algorithms, but with the notable difference that in
that case a geometric representation of the object based on primitive shapes (e.g., triangular meshes) is
often given, whereas the authors aim at processing point clouds. The authors then propose a
{\em fit and split} algorithm to decide if a bounding box should be split and, if that is the case, where.
The method is tested using GraspIt! \cite{Graspit} and the Columbia Grasp Data Base \cite{ColumbiaGraspDatabase}
using a five finger robotic hand. The objective is to evaluate how the proposed method performs in terms
of identifying force-closure grasps. The grasping strategy follows the method given by Pelossof et al. \cite{AllenICRA2004}
and formerly described.\\
\noindent {\bf Methods that operate in feature space}\\
The first algorithm capable of grasping a previously unseen object was presented by Saxena et al. \cite{SaxenaNgIJRR}.
The algorithm was implemented and tested on a Barrett arm equipped with the Barrett hand for tasks like unloading
a dish-washer.
The algorithm processes a single image of an object to be grasped and identifies
in the image the pixels corresponding to good grasping points. This information is then
passed to a robot controller that will attempt to grasp the object by placing the fingers
in the locations corresponding to the identified grasping points.
The training stage of the algorithm relies on a large set of images of objects to be
grasped where good grasping locations have already been identified and marked (e.g., the
handle of a mug). Interestingly, the authors use a database of synthetic images that is
publicly available. During the training stage,
a high-dimension feature vector
for every pixel in the image is created,
and then a classifier discriminating good grasping points from bad grasping points
is trained. For every pixel, a feature vector is created using various image filters computed
in a $5\times 5$ neighborhood of the pixel. Each feature $x_i$ is a vector\footnote{Seventeen
filters are applied on each neighboring pixel in a 5$\times$5 path and on the pixel
on the scaled version of the image in three different sizes, hence 17*24 + 17*3 = 459.} in $\mathbb{R}^{459}$.
Since every pixel in the training images is already labeled as good/bad, each feature
vector $x_i$ can then also be associated with the corresponding binary label $z_i$.
The classifier is based on logistic regression and works as follows. First,
a parameter vector $\theta^* \in \mathbb{R}^{459}$ is learned through maximum likelihood:
\[
\theta^* = \arg \max_{\theta} \prod_i \Pr(z_i|x_i;\theta).
\]
Classification at run time then works as follows. First, the feature vector $x$ is computed
for every pixel in the image of the object to be grasped. Then, the probability that
such pixel corresponds to a good grasping point (i.e., $z_i=1$) is computed via logistic regression:
\begin{equation}
\Pr(z_i = 1|x;\theta^*) = \frac{1}{1+e^{-x^T\theta^*}}.
\label{eq:log_reg}
\end{equation}
This step assigns a probability to every pixel and pixels can then be sorted accordingly
to these probabilities. The identified pixels can then be used to guide the robot hand
and complete the grasp maneuver. In the original paper, this is done by using a
stereo camera providing depth information, but various options are possible.
The notable part of this algorithm is that at run time, the
robot is not bound to grasp objects it saw during the training stage, but can rather
handle arbitrary objects. Despite its merits, this algorithm has two drawbacks.
The minor one is that computing $\theta^*$ is very time consuming. Although
this step is off-line, it takes time to retrain the algorithm when new images
have to be included in the training set. The major problem is that at run time,
a high-dimension feature vector has to be computed for every pixel in the image
in order to then compute Eq. \ref{eq:log_reg} for every pixel. For large images ,
this process become very time consuming and the frame rate is significantly impacted.
It should be noted that, as evidenced in the accompanying videos available on the authors'
websites, the robot largely relies on pinch grasps or grasps where the whole hand
envelops the object. This approach works well in practice, but does not strictly
follow the force/form closure taxonomy usually used in robotic grasping.
Balaguer and Carpin later on showed that the algorithm maintains a good performance
even when using a much smaller feature vector \cite{BenICRA2010}, i.e. 54 features rather
than 459. This reduction is possible by using a smaller window around the pixel, and just
a subset of the original image filters. The advantage of this method is that every image
can then be processed in a much shorter time (a speedup of a factor of 8 was reported).\\
Koostra and colleagues \cite{KragicIJRR2012} also propose a system to grasp unknown objects
based on vision.
The system, called Early Cognitive
Vision (ECV), extracts three-dimensional features that are then used to compute grasps with
two and three fingers. The system works as follows. Two types of features are extracted, namely from
edges and surfaces in the image.
Edges are extracted from the left and right images and grouped into groups like contours.
Groups from the left and right image are then matched to each other.
Surfaces in the images
generate three different types of features that the authors call {\em texlets}, {\em surflings},
and {\em surfaces}. The idea behind these features is to generate a hierarchy of features incorporating
information at higher and higher levels of abstraction.
They represent local properties of a surface with texture information (texlets), which are at the bottom of the hierarachy.
Each texlet includes the center of the patch ${\bf p}$, the normal ${\bf n}$, and the color $c$.
To build a hierarchical representation, a clustering algorithm is then applied to
texlets to obtain the next level of surfling features. As a collection of local planar patches, a surfling
represents a rectangular planar area. Finally, surflings are also grouped together in the highest
level of the hierarchy, i.e., surfaces. As the name suggests, {\em surfaces} aim to represent surfaces
of objects, and color is therefore not included to account for regions with different colors.\\
Once edge and surface features are extracted, a grasp is synthesized for the special cases of
two and three fingers. Rather than relying on a database of predetermined associations between objects
and grasps, the authors consider a set of basic grasps that are associated with specific geometric attributes.
The correct grasp for an object is then determined by analyzing the previously described features.
Three different grasps are considered, namely an encompassing grasp,
a top pinch grasp and side pinch grasp. The positioning of the fingers to implement this grasp is
then determined by processing the extracted contour and surface features.\\
The method has been demonstrated both in simulation and on a real world system using a Schunk parallel
gripper and a Schunk Dexterous Hand (three fingers). A BumbleBee 2 stereo camera was used for image acquisition.
The experimental setup considers various different objects. By definition, and in contrast to the previous
method that required a training stage, all these objects are novel. Emphasis is given on
grasping performance in the presence of clutter, but no information is provided regarding the
computational complexity or the time needed to compute these grasps.
\subsubsection{Model-less algorithms based on tactile sensing}
When a model for the object is not available, a grasp can be synthesized using touch sensors.
In this case, a robotic hand equipped with tactile sensors can acquire information about the object being
grasped before actually attempting a grasp. Similar to vision-based approaches, some tactile-based algorithms reconstruct a model of the object, whereas others do not. Note that these methods to not combine tactile data
with vision data, but rather rely only on touch.
Given the still relatively scarce availability of
affordable, reliable tactile sensors, this line of research is rather novel and there are
much fewer contributions than systems relying on visual input.\\
\noindent {\bf Methods that construct a model}\\
Bierbaum and collaborators have used touch sensors to infer a model of the object to be grasped \cite{BierbaumTactile2008,BierbaumTactile}.
In particular, in \cite{BierbaumTactile2008} they maintain to be the first to have developed a method for tactile
exploration using a multifingered robotic hand, in contrast to previous methods strictly relying on a single finger for exploration.\\
The objective of the presented system is to reconstruct the 3D shape of an unknown object by touching its surface using the fingers.
It is assumed that just the fingertips are equipped with tactile sensors, and that prior information
about size, position, and orientation of the object is also provided.
The key problem to be addressed is how to move the fingers in order to explore the unknown shape of
the object, a problem related to mobile robot exploration. Exploiting this similarity, the authors embrace a method
based on potential fields associated with a set of points generating attractive and repulsive forces.
Initially, using the prior information about the object, a uniform three-dimensional grid is placed around the object.
An attraction point is then created inside every grid cell, and exploration starts based on the associated attractive field.
Whenever a finger crosses a cell without making any contact, the attractive point located inside the cell is eliminated.
When a finger detects a collision inside a cell, a repulsive point is generated. The exploration of the
surface of the object occurs then by following the negated gradient of the potential field function. As with every method based
on potential fields, the main issue to consider is the possibility of getting stuck in a local minima. To cope with this
problem a reconfiguration policy is added. Reconfiguration occurs when a deadlock situation is detected, i.e., when the hand
has not been moving for a while. When this condition is triggered, the role of attractive and repulsive points is temporarily
swapped, so that the hand moves away from the object and is biased towards an unexplored part of the space. While in \cite{BierbaumTactile2008}
the authors just describe the exploration stage, i.e., how a 3D model for the object can be built, in \cite{BierbaumTactile}
a grasping strategy is presented, too. The main challenge to face is that the 3D shape reconstructed during the exploration
stage is inherently discrete, because of the aforementioned grid. This obstacle is overcome using the
method presented in \cite{PertinTroccaz} where geometric features are extracted and grasping is executed trying to place the fingers
on faces offering a sufficiently high grasping score.
The method is tested on a simulation system replicating the 6-DoF Mitsubishi RM 501 robotic arm equipped with a FRH 4 Hand, a hand with
9-DoFs and four fingers.\\
An alternative approach was proposed by Platt et al. \cite{GrupenTRO2010} whereby the system does not need a
perfectly accurate description of the object geometry. Instead, through an iterative refinement process, an initial
grasp is updated until force closure is achieved. The refinement process is guided by a force feedback system.
The system could then be used jointly with a traditional planner operating on an incomplete or noisy model. The set
of contact points provided by the planner could be inaccurate and/or not guarantee force closure, but would
provide a good starting point for the iterative process implemented by the controller. Specifically, the controller
implements a null-space approach where residuals for force and moments are combined to achieve force closure.
This method was tested on a whole arm manipulator (WAM) arm equipped with a Barrett hand retrofitted with load cells located at the fingertips
for force feedback.\\
\noindent {\bf Methods that operate in feature space}\\
Dang et al. use the term {\em blind grasping} to indicate methods where the
robot relies on tactile sensors only, but does not use vision \cite{AllenICRA2011,AllenLearningGraspStability}.
Their method is feature based and does not attempt to use tactile data to reconstruct an approximate model of the object.
The method assumes the availability of a grid of tactile sensors placed on the fingers.
It is assumed the sensors are subdivided into a certain number of pads (e.g., 4), but
this assumption is not critical for the method.
When performing a grasp, each tactile element in the grid will be either inactive (i.e., no contact)
or active. Note that active tactile elements return the actual force, and not just a binary
indication of whether they are activated.
Since the position of the tactile sensors is fixed on the hand, assuming that
the joint values are known, it is then possible to determine
the location of the contact points through forward kinematics. Starting from these premises, the grasping method
has some ideas in common with the work of Saxena \cite{SaxenaNgIJRR} we described in
Section \ref{sec:modelless_vision}. In fact, the method also relies on a training based on a set of labeled grasps
(good/bad), it
computes a feature vector at run time, and uses a classifier to decide if a grasp is good or bad
based on the information gathered during training.
The details are described in the following. During the training stage a large number of grasps is
considered, and for each of them the set of contact points is recorded. To reduce the dimensionality
of this dataset, just the position of the contact points is recorded (information about the direction
of the force is discarded), and points are then clustered using $K$ means. This reduced
dataset is called {\em contact dictionary} in order to evidence a similarity with methods based
on the Bag of Words principle \cite{BoW}. The dictionary is then
\[
D = [{\bf \hat c}_1,{\bf \hat c}_2,\dots,{\bf \hat c}_p]
\]
where each of the ${\bf \hat c}_i$ is a contact point in $\mathbb{R}^3$. Note that these are not
actual contact points, but rather the results of the clustering process. The authors report
that $p=64$ was used in their experiments and provided a good tradeoff between speed and accuracy.
When the robot hand makes contact with an object, a vector of $q$ contact points
is generated, i.e., a grasp $\mathcal{G}$ is represented by a vector
\[
\mathcal{G} = [{\bf c}_1,{\bf c}_2,\dots,{\bf c}_q].
\]
In general, $p \neq q$.
Starting from $D$ and $\mathcal{G}$ a $p$-dimensional feature vector is computed relating the grasp $\mathcal{G}$ to
the dictionary $D$. The feature vector is defined as
\[
F(\mathcal{G},D) =\sum_{i=1}^q H({\bf c}_i,D) \frac{f_{{\bf c}_i}}{S_{{\bf c}_i}}
\]
where $f_{{\bf c}_i}$ is the force sensed at contact point ${\bf c}_i$ and $S_{{\bf c}_i}$ is
the total force sensed. The term $H({\bf c}_i,D)$ is a $p$-dimensional vector measuring how ${\bf c}_i$
relates to the dictionary. To compute its $j$-th component one can use the following function
\[
h_j = \exp \left( - \frac{|| {\bf c}_i-{\bf \hat c}_j||^2}{\sigma^2}\right).
\]
The training stage is completed using the contact information of more than 24000 grasps from the Columbia Grasp
Data Base \cite{ColumbiaGraspDatabase}. For each grasp the feature vector is computed and
and associated with the corresponding binary label (good/bad grasp). Then, a support vector machine
(SVM) is trained based on this data to discriminate between the two classes. Emphasis is given
to reduce the false positive rate, as the authors speculate that from a
practical point of view this kind of error is more severe than a false negative (i.e., not executing
a good grasp because the classifier erroneously considers it to be a bad grasp.) At run time the robot approaches an object
lying on a plane
and grasps it with a tentative grasp $\hat{\mathcal{G}}$. The corresponding feature vector $F(\hat{ \mathcal{G}},D)$
is computed and then classified using the SVM. If the SVM classifies $\hat{\mathcal{G}}$ as a good grasp,
the robot tries lifting the object, otherwise it tries a different grasp.\\
The method is validated using a Barrett Hand with a good performance. However, there are some drawbacks,
too. For example, in order to initially make the tentative grasp $\hat{\mathcal{G}}$ the robot needs
prior information about the location of the object. If the position is not available
or if is noisy, the method will be either inapplicable or doomed to fail.
\section{Dual Arm Grasping}
Various technological developments and practical reasons continue to drive research in
dual arm grasping and manipulation. From a practical point of view, certain
grasping tasks require interaction with larger sized parts that may have physical
properties (weights, inertias .. ) and manipulation requirements that would require two or more robotic arms utilizing simple grippers or even possibly robotic hands. In these scenarios, the use of two arms is then preferred.
Moreover, there is an increasing expectation that robots will work side-by-side
with humans in a variety of everyday tasks ({\em corobots}). In this case robots
have to be able to use the same tools and operate in the same environments designed
for humans. As these designs often assume end-users are equipped with two hands,
a dual arm design will then ease the integration of robots into these environments.
Finally,
continued research in humanoid robotics is naturally linked to dual arm
grasping and manipulation.
Recent surveys on the topic appeared in \cite{KragicDualArm} and \cite{DualArmHandbook}.\\
From a high-level point of view, dual arm grasping can be treated with the same formalism
used for single arm robots. In fact, one can think of the dual arm robot as just one system
with a higher number of degrees of freedom. According to a taxonomy proposed in \cite{KragicDualArm},
this paradigm is called {\em coordinated} operation, and it refers to the situation where
the two arms perform different parts of the same task\footnote{This is opposed to what
they call {\em non-coordinated} operation, where the two arms perform separate tasks and can
therefore be treated separately.}. From a grasping perspective, then, one can define extended
versions for the grasping matrix $\bf G$ and the Jacobian matrix $\bf J$ and then perform
the same analyses presented in Section \ref{sec:formalism} for assessing form and force
closure.
\section{Grasp Performance Benchmarking}
Standardized performance testing, an emerging tool within the robotics community, is proving itself worthy within other robotics community disciplines and offers the benefits of an ``honest broker" to measure how well a system performs in a particular ability \cite{Bonsignorio2014}. Ultimately, physical measurements assess grasping performance using measurement techniques external to the system under evaluation. The results of such evaluations and benchmarks help to match capabilities to end-user needs, as well as to help developers improve their product designs. Section 5.1 discusses existing grasp measures and more abstract grasp properties for robotic hands, tactile sensing, and grasp algorithms within the robotic grasping community. Section 5.2 takes an application approach to overview the grasping problem in order to fully understand the scope of measurements. Finally in section 5.4, the results of our literature review and task analysis are used to formulate some concepts towards a set of performance evaluation and benchmarking techniques for automated grasping systems.
\subsection{Quantitative Grasp Measures}
The physical results of grasping are reported using both qualitative and quantitative data. Qualitative data is a categorical measurement expressed by means of a natural language description where quantitative data is a numerical measurement. Qualitative measures expressing the ability to grasp an object are commonplace and typically use pass/fail indicators along with a description of how well a grasp was performed on a given test (e.g., grasp A is not as stable as grasp B, the object was ejected from the grasp). Another aspect of performance testing is functional vs. non-functional testing. Functional tests evaluate a robotic hand and overall robotic system's ability to perform the grasp required to accomplish a specific task (e.g., holding and operating tools, grasping and turning valves, and operating a door knob after unlocking it with a key) \cite{DARPARC1,DARPAAC,NISTASTM}, while non-functional tests would be designed to measure more general properties of a robot hand outside the scope of an integrated robotic system. Both qualitative and quantitative measures can be used to express the results of functional and non-functional tests. Qualitative measures are easily found in robotic grasping research literature; however, examples of applying quantitative measures to evaluate grasp performance are sparse.
\subsubsection{Volumetric}
The authors of \cite{Kragten2010} propose a benchmark to measure the kinematic ability of a robotic hand to grasp objects. In particular, cylindrical objects were used of increasing diameter (40, 50, 55, 60, 63, 75, 90, 110, 115, and 120 mm) under both pinch grasps and enveloping grasps. In the case of a pinch grasp, the outermost point of the object circumference was placed a distance of L from the palm, and in the case of an enveloping grasp, the object was placed against the palm. Also noted during grasps were the cases where a pinch grasp resulted in the hand pulling the artifact towards the palm resulting in a transition to a final enveloping grasp equilibrium. A performance metric was defined as follows:
\[
Q_{grasp} =
\frac{\frac{\pi}{2}\Delta D_{obj}}{2L+2L_O}
\]
\noindent where $\Delta D_{obj}$ is the difference between the diameter of the largest and smallest graspable object, $L=L_1+ L_2$ is the length of one finger, and $2L_0$ is the palm width of the hand (Figure \ref{fig:volume_test}).
\begin{figure}
\centering
\includegraphics[width=12cm]{volume_test.png}
\caption{Reproduced from \cite{Kragten2010}, initial positions of a freely moving cylindrical object with respect to the palm of a hand to determine the ability to successfully grasp this object. Palm position is represented by the solid object and the pinch position by the dashed object. $D_{obj}$ is the diameter of the object, $2L_0$ is the width of the palm, $L_1$ is the length of the proximal phalanx, $L_2$ is the length of the distal phalanx, $L$ is the length of the finger. A torque $T_a$ applies to the base of the fingers.}
\label{fig:volume_test}
\end{figure}
\subsubsection{Internal Force}
Odhner et al. implemented a test apparatus to test the power grasp capabilities of the iRobot-Harvard-Yale (iHY) Hand, a compliant under-articulated hand that used tendons to actuate finger motion \cite{Odhner2014}. The apparatus was constructed of a split cylinder and a load cell attached at the cylinder center. The apparatus was oriented such that it was symmetric with the fingers and the load cell measured the force exerted between the opposing fingers in the direction of the split. The same test artifact was used to measure both power grasping and finger-tip grasping (Figure \ref{fig:force_test}). \newline
\begin{figure}
\centering
\includegraphics[width=12cm]{force_test_2.png}
\caption{Reproduced from \cite{Odhner2014} split ring test apparatus to measure power grasp (a) and finger-tip force (b).}
\label{fig:force_test}
\end{figure}
Romano et al. presents quantitative testing when evaluating the performance of a novel robotic grasp controller \cite{Romano2011}. The system's ability to control delicate manipulation tasks was evaluated with crushing measures. Crushing was defined as a deformation of 10 mm beyond the initial surface contact. There was no indication as to how these measures were made.
\subsubsection{Resistance to Force and Slip}
A benchmark in \cite{Kragten2010} tests the ability to hold objects. Again using cylindrical objects, the object is placed against the palm of the hand and grasped. The object is slowly moved along a straight line (5 mm/s) in a disturbance direction $\mu$, with the object allowed to move in the perpendicular direction $\nu$ (Figure \ref{fig:slip_test}). The force is measured throughout the pulling direction over several pull directions $\Phi$ and the maximum pull force is recorded for each. A performance metric was designed as follows:
\[
Q_{hold} =
\frac{FL}{T_a}
\]
\noindent where $F$ is the minimum force needed to pull an object out of the hand, $T_a$ is the constant actuation torque applied at the base of the fingers, and L is the total length of the finger. In \cite{Meijneke2011} the authors conduct similar experiments with grasps as in \cite{Kragten2010}, but also independently measure the contact force using a load cell internal to the cylinder and coupled to the hand through a ball bearing protruding through a hole in the cylinder. In these tests, the authors are relating the forces exerted on the cylinder by the hand to the forces required to pull the cylinders from the hand. \newline
\begin{figure}
\centering
\includegraphics[width=12cm]{slip_test.png}
\caption{Reproduced from \cite{Kragten2010}, a schematic for the setup where a cylindrical object with a diameter $D_{obj}$ is pulled out of the hand at a constant slow speed $\omega$ in the direction of $\mu$ while the fingers are at a constant torque $T_a$. The object is free to move in the direction $\nu$, which is perpendicular to $\mu$. The resultant of the contact forces on the object in the pull direction $\mu$ is measured.}
\label{fig:slip_test}
\end{figure}
Romano et al. test a robotic hand's ability to control delicate manipulation tasks using slippage measures \cite{Romano2011}. Slippage was defined was defined in two forms: translation (greater than 5 mm), and rotation (greater than 10 degrees). There was no indication as to how these measures were made. Slip measures were used to evaluate the grasp controller's ability to adjust the minimum grip force necessary to lift an object 10 mm from a table surface and to evaluate the controller's slip response. To evaluate slip response, a cup was stably grasped at a fixed load of 5 N. The cup was loaded incrementally with batches of 15 marbles (about 0.6 N per batch) and the gripper was shaken for two seconds while the cup was observed for slip. The authors also used pressure-sensitive film to capture the forces imposed on an object during placement onto a flat surface.
\subsubsection{Touch Sensitivity}
Dollar et al. presents an experimental setup to test grasp improvements achieved when integrating piezofilm contact sensors with a reactive control algorithm onto the Shape Deposition Manufactured (SDM) Hand \cite{Dollar2010}. The experimental setup consists of a shape artifact constrained to a six-axis force-torque sensor. A nominal grasp pose relative to the position of the artifact to be grasped is determined. Error offsets are then applied to the nominal pose and the forces associated with and without sensor feedback are measured. In addition to the force measurements, a qualitative assessment is applied to measure the success of the grasp. A successful grasp was defined as one where the object was able to be successfully lifted out of the force sensor mount without slipping out of the hand. Grasp success and contact force data were evaluated at 10 mm error increments from the nominal position. Results indicated that the addition of feedback from the contact sensors on the hand decreases the forces applied to the object during the grasp and increases the range of acceptable positioning offsets that still result in a successful grasp. \newline
Based on Dollar's work, SynTouch LLC reports an experiment for comparing the sensitivity of grasping using tactile sensing technologies\footnote{ This work was performed as part of a NIST Small Business Innovative Research (SBIR) contract no. DB1341-13-SE-0300 titled Advanced Tactile Sensing Technology for Robotic Hands.}. Using a spherical object fixed to a force plate, the experiment measures the unbalanced forces acting on the object upon making grasp point contacts. The tests were conducted over a range of closing velocities and varied the position of the object to test how grasps can adapt to positional errors. Results showed higher forces with increasing closing velocities and decreasing sensor compliance and were attributed to the speed of the hand's force control loop using the integrated sensor system. The research also presented a mechanism for using the collected data to determine the range of velocities and position errors a robotic hand system can tolerate for a given peak force.
\subsubsection{Compliance}
The developers of the iHY also developed a test for measuring the compliance of planar and spherical pinch grasps \cite{Odhner2014}. This was accomplished by mounting a 6-axis force-torque sensor to a mill headstock with the iHY hand fixtured in the mill's vice (Figure \ref{fig:stiffness_test}). Disturbance displacements were applied using the three linear axes of the milling machine, and the resultant forces were recorded relative to displacement. Stiffness values were determined by averaging out hysteresis due to tendon friction and the viscoelasticity of polymer pads and flexures by averaging values in both directions of each motion over several cycles.
A linear least squares estimation was used to fit the parameters of a symmetric stiffness matrix K to the data for both the opposed and spherical fingertip grasps.
\[
K=
\left( \begin{array}{ccc}
K_{xx} & K_{xy} & K_{xz} \\
K_{yx} & K_{yy} & K_{yz} \\
K_{zx} & K_{zy} & K_{zz}
\end{array} \right)
\]
\begin{figure}
\centering
\includegraphics[width=12cm]{stiffness_test.png}
\caption{Reproduced from \cite{Odhner2014}, experimental setup for measuring stiffness properties of compliant hand.}
\label{fig:stiffness_test}
\end{figure}
\subsubsection{In-Hand Manipulation}
Odhner et al. reports on experiments used to evaluate the in-hand manipulation capabilities using two fingers of an under-actuated robotic hand \cite{Odhoner2013}. Using several small objects having different width and radius of curvature, manipulation tests were conducted by tracking the position of the object relative to an initial fingertip grasped position. Objects were tracked in six degree-of-freedom space using a commercially available tracking system and the degree of slip was detected by measuring the error between the nominal start and finish object positions after returning the robotic hand fingers to their original fingertip grasp position.
\subsubsection{Grasp Properties and Quantitative Measures}
Grasp synthesis identifies the physical and mechanical properties of grasps and the creation of suitable parameters to quantify them. In a review of grasp synthesis algorithms, Shimoga identified the main properties of grasping as disturbance resistance, dexterity, equilibrium, stability, and dynamic behavior \cite{ShimogaSurvey}. A grasp with good disturbance resistance can withstand disturbances in any direction. This can be accomplished by form-closure (complete kinematical restraint) where a set of grasp points results in finger positioning that constrains an object, or force closure where grasp point forces applied by fingers constrain motion of the object (specific measures for these are proposed by many researchers and are discussed in Section 5.3). A grasp is considered dexterous if the kinematic properties of the robotic hand allow the object to be moved using a controlled and stable method, a concept also referred to as in-hand manipulation. A grasp is in equilibrium if resultant forces and torques applied to an object by finger and external forces equate to zero. A grasp is considered to be stable if any positional errors caused by external forces in finger or object position can be eliminated once the disturbance is removed. Finally, the dynamic behavior of a grasp is defined as the time response of the grasp for changes in its motion or force trajectories.
Similar to Shimoga, Cutkosky presents the properties of force closure, form closure, stability, and manipulability as analytical measures used to describe a grasp \cite{Cutkosky1989}. Cutkosky also presents internal forces, slip resistance, compliance, and connectivity. Internal forces apply to the magnitude and variance of forces that a hand applies to an object while maintaining grasp equilibrium that is described above. Slip resistance is the magnitude of the forces and moments on the object at the onset of slip. The resistance to slipping depends on the configuration of the grasp, on the types of contacts, and on the friction between the object and the fingertips. Compliance (inverse of stiffness) of the grasped object with respect to the hand is a function of grasp configuration, joint actuation, and structural compliances in the links, joints, and fingertips. Finally, connectivity is the number of degrees of freedom between the grasped object with respect to the hand.
Another important property is indicated in \cite{Dollar2010} that we will call grasp sensitivity, the ability of a grasp to conform to deviations in nominal object position without disturbing actual object location prior to achieving form closure with the object. Grasp sensitivity is a property of a force or contact sensing and associated control algorithms that occurs when achieving form closure. Table \ref{table:grasp_properties} consolidates these measures and maps them to the quantitative experimental methods as described in section 5.1. As indicated, there are multiple performance tests that can be used to assess a given measure; and some measures can be supported using several of the experimental methods found in the literature.
\begin{table}
\centering
\caption{Grasp properties and applicable performance tests}
\vspace{.5cm}
\begin{tabular}{ | l | p{7.5cm} | p{3.5cm} |}
\hline \textbf{Grasp Property} & \textbf{Description} & \textbf{Applicable Tests} (from Section 5.1) \\
\hline Form Closure & Ability to spatially constrain an object from moving when the finger joints are locked when assuming contact between the fingers and the object. & 5.1.1 \\
\hline Force Closure & There exists a conical combination of contact forces at the points of contact such that any external wrench applied to the object can be resisted. & 5.1.2, 5.1.3 \\
\hline Dexterity/Manipulability & The ability of the fingers to impart motions to the object using the kinematic properties of the robotic hand allowing the object to be moved using a controlled and stable method. Also called in-hand manipulation. & 5.1.6 \\
\hline Equilibrium & Resultant forces and torques applied to an object by finger and external forces equate to zero. & 5.1.2, 5.1.3 \\
\hline Stability & Ability of the grasp to return to its initial configuration after being disturbed by an external force or moment. & 5.1.2, 5.1.3 \\
\hline Dynamic Behavior & The time response of the grasp for changes in its motion or force trajectories. & 5.1.2, 5.1.3, 5.1.4 \\
\hline Internal Forces & Magnitude and variance of internal grasp forces that a hand applies to an object without disturbing the grasp equilibrium. & 5.1.2 \\
\hline Slip Resistance & Magnitude of the forces and moments on the object at the onset of slip. & 5.1.2, 5.1.3 \\
\hline Compliance & The effective compliance of the grasped object with respect to the hand. & 5.1.5 \\
\hline Connectivity & Number of degrees of freedom between the grasped object and the hand. & 5.1.6 \\
\hline Sensitivity & Ability to conform to deviations in nominal object position without disturbing actual object location prior to achieving final grasp. & 5.1.4 \\
\hline
\end{tabular}
\label{table:grasp_properties}
\end{table}
\subsection{Analysis of a Grasping Task}
As always, breaking down a problem into its parts can provide novel insights towards its solution. In particular, consider the underlying tasks associated with a robotic pick and place operation for a fully integrated multi-fingered robotic hand (see Figure \ref{fig:grasp_task}). Each task in this particular operation\footnote{ This discussion presents the results of an informal grasp metrics working group organized by NIST as well as grasp metrics and test methods research being performed at NIST. More information at http://www.nist.gov/el/isd/grasp.cfm and http://rhgm.org/. See also \cite{RAS2015}.} possesses a number of associated problems that can serve as a basis for extracting performance measures. More specifically, quantifying the performance of a system in handling these problems can help guide and justify the various strategies taken. Furthermore, identifying the significance of particular performance measures towards different grasping tasks would provide valuable knowledge on necessary functionalities and their performance towards task completion. For example, picking up a part and tossing it into a bin requires minimal position accuracy of the grasped object once the grasp component is completed where picking up a part and performing an assembly operation requires much more accurate positioning throughout the task. Thus, quantifying and suggesting a minimal level of performance in the system's ability to control and measure object position in the latter scenario would be critical in predicting operation success.
A plausible outline for the pick and place operation begins with a "best" set of grasp points as determined from a grasp planner. The hand is positioned by a robotic arm to cage an object by establishing an approach trajectory and offsets that are based on the grasp planning stage. During the cage task segment, it is possible for components of the robotic hand to run into obstructions in the vicinity of the object or run into the candidate object to be grasped due to inadequate \textit{clearances}.
\begin{figure}
\centering
\includegraphics[width=12cm]{grasp_task.jpg}
\caption{Pick and place task segmentation and transitions between grasped and un-grasped states (from \cite{RAS2015}.)}
\label{fig:grasp_task}
\end{figure}
During the \textit{constrain} task segment, the object is spatially confined by the grasp at the grasp points. Sensor-less contacts depend on the positioning accuracy of the hand delivery system and/or the \textit{synchronization} of fingertip position in time. Unsynchronized contact requires minimal force contacts to minimize disturbance to the part, if maintaining part position is important, and requires hand sensing capabilities such as tactile or current sensing. Position problems can occur during this task segment that result in a missed \textit{contact point} where the part is not fully constrained and contact movement due to synchronization issues or inadvertent contact by the hand due to \textit{clearance} issues. Clearance issues may also result if the object is too small to be grasped as in the case of a three fingered radial grasp on a cylinder where closing the fingers results in a collision between fingers before making contact with the object.
The load task segment applies the calculated forces required to keep a firm grasp on the object. These forces are most often calculated to obtain an efficient grasp based on the forces required to stabilize the object in the presence of gravitational and inertial forces. Problems during this segment are due to the uncertainty in the kinetics of the system (i.e., object mass), disturbance forces, and torques applied to the object. Uncertainty in the system can lead to the occurrence of \textit{slippage}, damage to the part (\textit{crushing}), or \textit{ejection} when trying to achieve efficient grasp forces.
The pick task segment lifts the object for manipulation. Problems during this segment are due to the errors between gravitational forces and the forces applied during the load task and again result in \textit{slippage}, \textit{ejection}, or \textit{crushing}. Other considerations are variations in object position relative to the robot hand coordinate frame upon picking up the part due to the \textit{compliance} properties of the robot hand.
During the manipulation phase, the part is picked (or lifted) from its grasped position, moved along a trajectory, and placed in a final position. The trajectory could be induced by the robot carrying the hand as well as the robotic hand itself, often referred to as in-hand manipulation. Problems are due to change in part position relative to the hand caused by external forces associated with contacts between the object and the environment throughout the manipulation process, as well as uncertainty in the objects kinetic properties. In addition, fluctuations in the mass of the object as well as exogenous disturbances can occur due to an intermediate assembly operation on the object. All of these force changes can result in \textit{slippage}, \textit{ejection}, or \textit{crushing}.
During the place task segment the object errors are dependent on object positional placement accuracy. In the most lenient case, the object is placed into a bin at a random orientation. In another case the object is positioned on a flat surface where accuracy errors could result in unexpected surface/object contact forces leading to \textit{ejection}, \textit{slip}, or \textit{crushing}. The most complex case is that of assembly, where assembly algorithms are dependent on the positional accuracy of initial contact between the object and the subassembly and the object is subjected to a multitude of external forces throughout the assembly process also leading to \textit{ejection}, \textit{slip}, or \textit{crushing}.
Post manipulation phase, the remaining task segments are a reversal of the task segments that led up to the manipulation phase and the errors associated with these are similar to their counterparts. Here the object is unloaded to the point of zero force contact and released so that the robot hand components clear the object, allowing the hand to be moved to the next operation.
\subsection{Towards Standardized Benchmarks for Robotic Hands}
Robotic hands are an integrated, mechatronic system of sensors, motors, and control algorithms typically ranging from three-fingered to five-fingered anthropomorphic designs having both fully and under articulated linkages with compliant or rigid joints. Designs incorporate a variety of sensing technologies, including simple current sensing at the drive motors, load cells, barometers, hydrophones, pressure transducers, electrodes, cameras, and tactile arrays. Depending on the sensory layouts and particular mechanical implementation, tactile sensing capabilities can include the ability to resolve point of contact, directionality and magnitude of contact forces, as well as other sensing modalities such as vibration and temperature. Control algorithms use these signals to incorporate position, velocity, and force control schemes. This wide scope of performance characteristics requires a modular set of performance metrics and associated test methods that can be chosen based on a defined set of grasp types the hand can perform, as well as a scheme for classifying a hand that includes sensing and control capabilities. Also needed are a common set of test objects (artifacts) to be used along with the test methods. A framework for benchmarking the performance of robotic hands is shown in Figure \ref{fig:framework}.
\begin{figure}[htb]
\centering
\includegraphics[width=10cm]{framework.jpg}
\caption{Framework for standardized benchmarking of robotic hands}
\label{fig:framework}
\end{figure}
Grasp taxonomies for the human hand have been developed towards the understanding of grasps humans commonly use in everyday tasks. Cutkosky performed a study of the grasps used by machinists in a small batch manufacturing operation and developed a taxonomy of grasps to provide insights for the design of versatile robotic hands for manufacturing \cite{Cutkosky1989}. Feix et al. derived a taxonomy of grasps based on a literature review of 14 human grasp studies (including Cutkosky's) from both the robotics and medical communities \cite{Feix2009}. The knowledge of these grasp taxonomies has been applied to the design of robotic and prosthetic hands and provides a basis for describing the grasp types a hand can perform.
Performance tests should encompass general grasping tasks as the one defined in 5.2 above and should contain unit, integrated, and functional test methods. When evaluating the capabilities of a robotic hand, unit and integrated performance tests should be agnostic to the other system components such as the robot arm and perception system. While it is possible to access data directly from a robotic hand and derive the defined metrics, these measurements would be based on the inherent properties of the system under test. Therefore, independent measurement systems must be developed to support testing to allow for comparative metrics between systems without effects such as force accuracies and data latencies.
Unit performance characteristics include kinematic properties such as volumetric capabilities and grasp configurations with associated maximum force capabilities. At the very basic level, primitive geometries such as spheres, cylinders, and cubes can be used to characterize the volumetric capabilities of a hand and maximum pinch and grasp forces can be determined at the bounds of these primitive volumetric capabilities. Individual finger tests can be performed to determine the positional accuracy and repeatability of the finger as well as velocity and acceleration characteristics. Sensors can be tested at their stock sensing modalities for properties such as resolution and sensitivity. For example, in the case of tactile sensors, desired characteristics might include normal and shear sensing capabilities as well as the ability to resolve the direction for forces and spatial resolution.
Integrated system characteristics include tests to evaluate the ability of a hand to withstand external forces while maintaining a good level of grasp efficiency and the ability to make initial contact with an object with minimal disturbance. In addition, tests are needed to characterize the integration of a sensor system. For example, tests are needed to characterize feedback latecy from a tactile sensor, as well as to quantify the grasp efficiency of a hand holding onto an object that is subjected to external disturbances.
Functional tests which include a robot arm and perception system can be standardized if they are defined generically to meet the requirements of an application space or to evaluate the capabilities of different robotic hand technologies for a known application. These functional tests would be best suited for a robotic hand that has evolved through a development process that included both unit and integrated testing to the point that warrants testing for use within an intended application space. Finally, functional tests should include testing within the actual application space to determine if a system is capable for the intended application.
In summary, standardized performance benchmarks for robotic hands offer the benefits of an ``honest broker'' to measure how well a system performs in a particular ability where the results of such evaluations and benchmarks help to match capabilities to end-user needs, as well as to help developers improve their product designs. To date, benchmarks to assess the results of grasp research are primarily qualitative measures; however, there is evidence of quantitative assessments of grasping research results scattered across the community. Standardized benchmarks will require a framework for matching the grasp types that a system under test can perform, as well as matching its sensing and control capabilities to the right set of unit and integrated performance tests in order to perform a thorough evaluation of a robotic hand system. General functional tests designed for particular application spaces are also needed for use prior to functional tests to support the actual application.
\newpage
\section*{Symbols used}
\begin{table}[htb]
\centering
\begin{tabular}{|c|l|}
\hline
Symbol & Meaning\\
\hline
$O_{XYZ}$ & inertial reference frame\\
$B_{xyz}$ & body frame attached to the robot\\
$\mathcal{B}$ & object being manipulated\\
$\mathcal{H}$ & robotic hand\\
${\bf u}$ & configuration of a frame in operational space\\
$h$ & number of degrees of freedom of the hand\\
$a$ & number of degrees of freedom of the arm\\
$m$ & number of degrees of freedom of the robot ($m=a+h$)\\
${\bf q}$ & configuration of the robot (vector in $\mathbb{R}^m$)\\
${\bf q}_a$ & configuration of the arm (vector in $\mathbb{R}^a$)\\
${\bf q}_h$ & configuration of the hand (vector in $\mathbb{R}^h$)\\
${\bf c}_i$ & contact point between $\mathcal{H}$ and $\mathcal{B}$\\
${\bf w}$ & wrench \\
$\mu, \gamma$ & friction coefficients\\
$f_n$ & normal force exerted at a contact\\
FC & friction cone\\
$n_c$ & number of contact points\\
$\bf J$ & hand Jacobian\\
FK & forward kinematics map\\
${\bf G}$ & grasp matrix\\
$\mathcal{G}$ & multifingered grasp\\
$n$ & number of independent forces and moments\\
{\bf l} & vector of independent forces and moments \\
$\psi$ & gap function\\
\hline
\end{tabular}
\caption{Symbols used in this primer}
\end{table}
\newpage
|
\section{Introduction}
The aim of this paper is to solve, through the dynamic programming approach,
a class of stochastic optimal control problems with delay in the control.
Stochastic optimal control problems governed by delay equations with delay in the control are usually harder to study than the ones when the delay appears only in the state (see e.g. \cite[Chapter 4]{BDDM07} in the deterministic case and \cite{GM,GMSJOTA} in the stochastic case).
When one tries to apply the dynamic programming method the main difficulty is the fact that, even in the simplified setting introduced first by Vinter and Kwong \cite{VK} in the deterministic case (see e.g. \cite{GM} for the stochastic case), the associated HJB equation is an infinite dimensional second order semilinear PDE that
does not satisfy the so-called ``structure condition'', which substantially
means that the control can act on the system modifying its dynamics
at most along the same directions along which the noise acts.
The absence of such condition, together with the lack of smoothing properties which is a common feature of problems with delay,
prevents the use of the known techniques, based on BSDE's
(see e.g. \cite{FT2})
or on fixed point theorems in spaces of continuous functions
(see e.g. \cite{CDP1,CDP2,DP3,G1,G2})
or in Gauss-Sobolev spaces
(see e.g. \cite{ChowMenaldi,GGSPA}),
to prove the existence of regular solutions of this HJB equation.\footnote{The viscosity solution technique can still be used (see e.g. \cite{GMSJOTA})
but to prove existence (and possibly uniqueness) of solutions that are merely continuous.}
In the companion paper \cite{FGFM} we proved, using an ad hoc method based on the partial smoothing properties of the Ornstein-Uhlenbeck semigroup, that
the HJB equation associated to the class of problems under study here,
admits a unique {\em mild solution}, i.e. a solution which possesses enough regularity to give sense to the ``candidate optimal feedback map''
which depends on a suitable directional derivative (denoted by $\nabla^B$)
of the solution.
In this paper we start from such a result and exploit it to solve our class of problems. More precisely we prove here:
\begin{itemize}
\item[(A)] an approximation result for the solutions of the HJB equation, i.e. that the mild solutions can be approximated by classical solutions, to which Ito's formula applies (Lemma \ref{lm:approximation});
\item[(B)] a verification theorem for the control problem
(Theorem \ref{teorema controllo});
\item[(C)] the existence, under further assumptions on the Hamiltonian, of optimal controls in feedback form (Theorem \ref{teo su controllo feedback}).
\end{itemize}
These results allows to treat satisfactorily a specific class of state equations and data which arises naturally in many applied problems (see e.g. \cite{FabbriGozzi08,FedTacSicon,GM,GMSJOTA,Kol-Sha,PhamBruder}).
The three points outlined above are the ones followed, in a different context, in \cite{G1}. However the specific features of our problems prevents the use of the same techniques. In particular the approximation result, which is a key tool here,
must be completely reworked, as we explain in Remarks \ref{rm:defpistrictstrong}.
and \ref{rm:Yk}.
We finally note that, similarly to what often happens in the literature (see e.g.
\cite{FedTacSicon,GM,GMSJOTA}) here we treat the case of ``distributed delay''
which gives rise to a bounded control operator in the state equation.
The case of ``pointwise delay'', even if it seems treatable with our approach,
is left for future extensions of our research.
\subsection{Our results in a simple motivating case}
\label{SS:motiv}
To be more clear we now briefly describe our setting and our main result in a special case. Let $(\Omega, \calf, \P)$
be a complete probability space
and consider the following linear controlled Stochastic
Differential Equation (SDE) in $\R$:
\begin{equation}
\left\{
\begin{array}
[c]{l}
dy(s) =a_0 y(s) dt+b_0 u(s) ds +\dis\int_{-d}^0b_1(\xi)u(s+\xi)d\xi+\sigma dW_s
,\text{ \ \ \ }s\in[t,T] \\
y(t) =y_0,\\
u(\xi)=u_0(\xi), \quad \xi \in [-d,0).
\end{array}
\right. \label{eq-contr-ritINTRO}
\end{equation}
Here $W$ is a standard Brownian motion in $\R$, and $(\calf_t)_{t\geq 0}$ is the
augmented filtration generated by $W$.
We assume $a_0,b_0\in \R$, $\sigma >0$. The parameter $d>0$ represents the maximum delay
the control takes to affect the system while $b_1$ is the density function
taking account the aftereffect of the control on the system.
The case treated here is when $b_1\in L^2([-d,0],\R)$ (``distributed delay'')
while a more difficult case which we leave for further research is when $b_1$ is a measure, e.g. a Dirac delta in $-d$ (``pointwise delay'').
The initial data are the initial state $y_0$ and the past history $u_0$ of the control.
The control $u$ belongs to $L^2_{\calf}(\Omega\times [0,T], U)$,
the space of predictable square integrable processes with values in
$U\subseteq \R$, closed.
Such kind of equations is used e.g. to model the effect of advertising on
the sales of a product (see e.g. \cite{GM,GMSJOTA}), the effect of investments
on growth (see e.g. \cite{FabbriGozzi08} in a deterministic case),
or, in a more general setting, to model optimal portfolio problems with execution delay,
(see e.g. \cite{PhamBruder}) or to model the interaction of drugs with tumor cells
(see e.g. \cite{Kol-Sha} p.17 in the deterministic case).
In many applied cases (like the ones quoted above) the goal of the problem is to minimize the following objective functional
\begin{equation}\label{costoconcretoINTRO}
J(t,x_0,u_0;u(\cdot))=\E \int_t^T \left(\bar\ell_0(s,y(s))+\bar\ell_1(u(s))\right)\;ds +\E \bar\phi(y(T)).
\end{equation}
where $\bar\ell_0:[0,T]\times\R\rightarrow \R$, $\bar\ell_1:U\rightarrow \R$ and
$\bar\phi:\R\rightarrow \R$ are continuous functions satisfying suitable assumptions that will be introduced in Section \ref{section-statement}.
It is important to note that here $\bar\ell_0$, $\bar\ell_1$ and $\bar\phi$ do not depend on the past of the state and/or control.
This is a very common feature of such applied problems.
A standard way\footnote{It must be noted that, under suitable restrictions on the data, one can treat optimal control
problems with delay in the control also by a direct approach without transforming them in infinite dimensional problems.
However in the stochastic case such direct approach seems limited to a very special class of cases
(see e.g. \cite{LarssenRisebro03}) which does not include our model and models commonly used in applications
like the ones just quoted.} to approach these delayed control problems, introduced in \cite{VK} for the deterministic case
(see \cite{GM} for the stochastic case) is to reformulate the above linear delay equation as a linear SDE in the Hilbert space
$\calh:=\R \times L^2([-d,0],\R)$, with state variable $Y=(Y_0,Y_1)$ as follows.
\begin{equation}
\left\{
\begin{array}
[c]{l}
dY(s) =AY(s) ds+Bu(s) ds+GdW_s
,\text{ \ \ \ }s\in[t,T] \\
Y(t)=x=(x_0,x_1),
\end{array}
\right. \label{eq-astrINTRO}
\end{equation}
where $A$ generates a $C_0$-semigroup (see next section for precise definitions)
while, at least formally,\footnote{Note that, when $b_1$ is a measure, the above operator
$B$ is not bounded and this makes the problem more difficult.}
\begin{equation}
\label{BGINTRO}
B:\R\rightarrow \calh,\qquad Bu=(b_0 u, b_1(\cdot)u), \; u\in\R,
\qquad
G:\R\rightarrow \calh,\qquad Gx=(\sigma x, 0), \; x\in\R.
\end{equation}
Moreover $x_0=y_0$ while $x_1(\xi)=\int_{-d}^\xi b_1(\varsigma)u_0(\varsigma-\xi)d\varsigma$, ($\xi \in [-d,0]$) i.e.
the infinite dimensional datum $x_1$ depends on the past of the control.
The value function is defined as
$$
V(t,x):=\inf_{u(\cdot)\in L^2_{\calf}(\Omega\times [0,T], U)}
\E \left(\int_t^T \left[\bar\ell(s,Y_0(s))+\bar\ell_1(u(s))\right] ds +\bar\phi(Y_0(T))
\right)
$$
The associated HJB equation (whose candidate solution is the value function) is
\begin{equation}\label{HJBINTRO}
\left\{\begin{array}{l}\dis
-\frac{\partial v(t,x)}{\partial t}=
\frac{1}{2}Tr \;GG^*\nabla^2v(t,x)
+ \< Ax,\nabla v(t,x)\>_\calh
+ \bar H_{min} (\nabla v(t,x)) +\bar\ell_0(t,x),\qquad t\in [0,T],\,
x\in \calh,\\
\\
\dis v(T,x)=\bar\phi(x_0),
\end{array}\right.
\end{equation}
where, defining the current value Hamiltonian $\bar H_{CV}$ as
$$
\bar H_{CV}(p;u):=\<p,Bu\>_\calh+\bar\ell_1(u)=\<B^*p,u\>_\R +\bar\ell_1(u)
$$
we have
\begin{equation}\label{HminINTRO}
\bar H_{min} (p):=
\inf_{u\in U} \bar H_{CV}(p\,;u).
\end{equation}
It is well known, see e.g. \cite{YongZhou99} Section 5.5.1, that, if the value function
$V$ is smooth enough, and if the current value Hamiltonian $\bar H_{CV}$ always admits at least a minimum point, a natural candidate
optimal feedback map is given by $(t,x) \mapsto u^*(t,x)$ where
$u^*(t,x)$ satisfies
\[
\<\nabla V(t,x),Bu^*(t,x)\>_{\R}+\bar\ell_1(u^*(t,x))= \bar H_{min}(\nabla V(t,x)).
\]
i.e. where $u^*(t,x)$ is a minimum point of the function
$u\mapsto \bar H_{CV}(\nabla v(t,x);u)$, $\R\to \R$.
To take account of the presence of $B$ it is convenient to define,
for $z \in \R$,
\begin{equation}\label{HminINTRObis}
H_{min} (z):=
\inf_{u\in U} \{\<z,u\>_\R+\bar\ell_1(u) \}=:\inf_{u\in U} H_{CV}(z\,;u) \}
\end{equation}
so that
\begin{equation}\label{HminINTROter}
\bar H_{min} (p)=
\inf_{u\in U} \bar H_{CV}(p\,;u).
=\inf_{u\in U} \{\<B^*p,u\>_\R+\bar\ell_1(u ) \}
=:H_{min}(B^*p).
\end{equation}
From now on we will use $H_{CV}$ and $H_{min}$ in place of $\bar H_{CV}$ and $\bar H_{min}$ writing $H_{min}(\nabla^B v(t,x))$ in place of $\bar H_{min}(\nabla v(t,x))$ in \myref{HJBINTRO}.
Since $u^*(t,x)$ is a minimum point of the function
$u\mapsto H_{CV}(B^*\nabla v(t,x);u)$, $\R\to \R$,
the minimal regularity required to give sense to such term is the existence
of $B^*\nabla v(t,x)$ which we will call $\nabla^B v(t,x)$ according to the definition and notation used e.g. in \cite{FTGgrad,Mas} and in the companion paper \cite{FGFM}.
In \cite{FGFM} a result of existence and uniqueness of mild solutions for
such equation has been proved. Mild solution (defined through an integral form of
\myref{HJBINTRO} are continuous and such that
$\nabla^B v$ is well defined and continuous, hence the
natural candidate optimal feedback map $u^*(t,x)$ above is well defined.
Notice that if the controlled state equation satisfies the ``structure condition'', meaning that the control affects the system only through
the noise, then by the so called BSDEs approach, see e.g. \cite{FT2}, the fundamental relation and the consequent verification theorem,
can be proved also by applying the Girsanov Theorem. In our case the ``structure condition'' does not hold,
since it would mean that ${\rm Im} B\subseteq {\rm Im} G$. This is an intrinsic feature of control problems with delay in the control
since the fact that $\operatorname{Im} B $ is not contained in $\operatorname{Im} G$ is just due to the presence of the delay in the control.
If the delay in the control disappears, then the structure condition hold, even if delay in the state is present (see e.g. \cite{GM,GMSJOTA,FT2,MasBanach}).
Then solve the problem, as recalled in the beginning of this introduction,
one needs to accomplish the steps (A)-(B)-(C), that we briefly introduce:
\begin{itemize}
\item [(A)] In Lemma \ref{lm:approximation} we show that if we suitably approximate the coefficients of the HJB equation (\ref{HJBINTRO}), we obtain
a sequence of functions $(w_n)_n$ which are strict solutions of the approximating HJB equations, and which
are once differentiable in time and twice differentiable with respect to $x$. Moreover the sequence $(w_n)_n$ converges to $v$, the mild solution
of the HJB equation, in the sense of the $\calk$-convergence, see Definition \ref{k-conv}.
\item [(B)] In Theorem \ref{teorema controllo} we apply the fundamental relation proved in Proposition \ref{prop rel fond}: in order to prove the fundamental relation
it is crucial to apply the Ito formula, and this can be done thanks to the approximation performed in (A) and to a further approximation of the state.
\item [(C)] In Theorem \ref{teo su controllo feedback}, under further regularity assumptions stated in Hypothesis
\ref{ipotesicostoconcretobis}, we solve the closed loop equation and so we show the existence of optimal controls in feedback form and the fact that the value function coincides
with the solution of the HJB equation, see Theorem \ref{teo:v=V}.
\end{itemize}
\noindent
\bigskip
Finally notice that in the present paper we deal with a finite dimensional control delay equation
(\ref{eq-contr-ritINTRO}), that here in the introduction we have presented in dimension one for the sake of simplicity. The same arguments apply
if we consider the case of a controlled stochastic differential equation in an infinite dimensional Hilbert space $\calh_0$ with delay in the control
as follows.
\begin{equation}
\left\{
\begin{array}
[c]{l
dy(t) =A_0 y(t) dt+B_0 u(t) dt +\int_{-d}^0B_1(\xi)u(t+\xi)d\xi+\sigma dW_t
,\text{ \ \ \ }t\in[0,T] \\
y(0) =y_0,\\
u(\xi)=u_0(\xi), \quad \xi \in [-d,0).
\end{array}
\right. \label{eq-contr-rit-infINTRO}
\end{equation}
Here $W$ is a cylindrical Wiener process in another Hilbert space $\Xi$, $A_0$ is the generator of a strongly continuous semigroup in
$\calh_0$, $\sigma \in \call (\Xi,\calh_0)$, and we have to assume some smoothing properties for the Ornstein Uhlenbeck transition
semigroup with drift term given by $A_0$ and diffusion equal to $\sigma$, see Remarks 3.2, 4.7 and 4.12 of \cite{FGFM} for more details.
\subsection{Plan of the paper}
In the first two Sections we present the control problem we treat and we recall some results on the state equation and on the HJB equation proved in \cite{FGFM},
in the last two Sections we solve the control problem. In details, the plan of the paper is the following:
\begin{itemize}
\item in Section \ref{section-prel} we give some notations and we present the problem and the main assumptions;
\item in Section \ref{section-smoothOU} we collect some results proved in \cite{FGFM} and fundamental in solving the control problem:
the partial smoothing property for the Ornstein-Uhlenbeck transition semigroup related to the state equations, and results on
the existence of a mild solution of the HJB equation;
\item in Section \ref{sec-Kstrong} we prove that this mild solution can be approximated by means of a sequence of strong solutions;
\item Section \ref{sec-verifica} is devoted to solve the optimal control problem. In Subsection 5.1
we prove a verification theorem, and finally in Subsection 5.2 we identify the value function of the control problem with the solution
of the HJB equation and we characterize the optimal control by a feedback law.
\end{itemize}
\section{Some preliminary results on the control problem}\label{section-prel}
In this section we collect and synthesize, for the reader's convenience, the basic material on our control problem which has been already exposed in the companion paper \cite{FGFM}.
\subsection{Notation and $C$-derivatives}\label{subsection-notation}
Let $H$, $K$ be real and separable Hilbert spaces, with norms given by $\left| x\right|_{H}$
and by $\left| x\right|_{K}$, respectively, or by $\left| x\right| $, if no
confusion is possible, and by scalar product $\left\langle \cdot,\cdot\right\rangle _H$,
$\left\langle \cdot,\cdot\right\rangle _K$, respectively, or simply
by $\left\langle \cdot,\cdot\right\rangle $. The space $\call(H,K)$ denotes the
space of bounded linear operators from $H$ to $K$ endowed with the usual
operator norm.
In the following, by $(\Omega, \calf, \P)$ we denote a complete probability
space, and by $L^2_\calp(\Omega\times[0,T],H)$
the Hilbert space of all predictable processes
$(Z_t)_{t\in[0,T]}$ with values in $H$, normed by
$\Vert Z\Vert^2 _{L^2_\calp(\Omega\times[0,T],H)}=\E\int_0^T\vert Z_t\vert^2\,dt$.
Next we introduce some spaces of functions. We let $H$ and $Z$ be Hilbert spaces.
By $B_b(H,Z)$ (respectively $C_b(H,Z)$, $UC_b(H,Z)$) we denote the space of all functions
$f:H\rightarrow Z$ which are Borel measurable and bounded (respectively continuous
and bounded, uniformly continuous and bounded).
Given an interval $I\subseteq \R$ we denote by
$C(I\times H,Z)$ (respectively $C_b(I\times H,Z)$)
the space of all functions $f:I \times H\rightarrow Z$
which are continuous (respectively continuous and bounded).
$C^{0,1}(I\times H,Z)$ is the space of functions
$ f\in C(I\times H)$ such that for all $t\in I$
$f(t,\cdot)$ is Fr\'echet differentiable.
By $UC_{b}^{1,2}(I\times H,Z)$
we denote the linear space of the mappings $f:I\times H \to Z$
which are uniformly continuous and bounded
together with their first time derivative $f_t$ and its first and second space
derivatives $\nabla f,\nabla^2f$.
If $Z=\R$ we omit it in all the above spaces.
Next, to introduce some spaces of functions which are differentiable in suitable directions, we recall the definition of $C$-directional derivatives
given in \cite{Mas}, Section 2, and in \cite{FTGgrad}.
\begin{definition}
\label{df4:Gder}
Let $H$, $K$, $Z$ be real Hilbert spaces.
Let $C:K \rightarrow H$ be a bounded linear operator and let $f:H\rightarrow Z$.
\begin{itemize}
\item The $C$-directional
derivative $\nabla^{C}$ at a point $x\in H$ in the direction $k\in K$ is defined
as:
\begin{equation}
\nabla^{C}f(x;k)=\lim_{s\rightarrow 0}
\frac{f(x+s Ck)-f(x)}{s},\text{ }s\in\mathbb{R},
\label{Cderivata}
\end{equation}
provided that the limit exists.
\item We say that a continuous
function $f$ is $C$-G\^ateaux differentiable at a point $x\in H$
if $f$ admits the $C$-directional derivative in every direction $k\in K$ and
there exists a linear operator, called the $C$-G\^ateaux differential, $\nabla^{C}f(x)\in\call(K,Z)$, such that
$\nabla^{C}f(x;k)=\nabla^{C}f(x)k$ for $x \in H$, $k\in K$.
The function $f$ is $C$-G\^ateaux differentiable on $H$ if it is
$C$-G\^ateaux differentiable at every point $x\in H$.
\item We say that $f$ is $C$-Fr\'echet differentiable
at a point $x\in H$ if it is $C$-G\^ateaux differentiable and if the limit
in (\ref{Cderivata}) is uniform for $k$ in the unit ball of $K$. In this case
we call $\nabla^C f(x)$ the $C$-Fr\'echet derivative (or simply the $C$-derivative) of $f$ at $x$.
We say that $f$ is $C$-Fr\'echet differentiable on $H$ if it is $C$-Fr\'echet differentiable at every point $x\in H$.
\end{itemize}
\end{definition}
Note that, in doing the $C$-derivative, one considers only the directions in $H$ selected in the image of $C$.
When $Z=\R$ we have $\nabla^C f(x) \in K^*$. Usually we will identify $K$ with its dual $K^*$ so $\nabla^C f(x)$ will be treated as an element of $K$.
If $f:H\to \R$ is G\^ateaux (Fr\'echet) differentiable on $H$ we have that, given any $C$ as in the definition above, $f$ is $C$-G\^ateaux (Fr\'echet) differentiable on $H$ and
$$
\<\nabla^{C}f(x),k\>_{K} =\<\nabla f(x),Ck\>_{H}
$$
i.e. the $C$-directional derivative is just the usual directional derivative at a point $x\in H$ in direction
$Ck\in H$. Anyway the $C$-derivative,
as defined above, allows us to deal also with functions that are not G\^ateaux differentiable in every direction.
Next we introduce some spaces of functions, see also \cite{FGFM}, Section 2.2.
Now we define suitable spaces of $C$-differentiable functions.
\begin{definition}
\label{df4:Gspaces}
Let $I$ be an interval in $\R$ and let $H$, $K$ and $Z$ be
suitable real Hilbert spaces.
\begin{itemize}
\item
We call $C^{1,C}_{b}(H,Z)$ the space of all functions $f:H\to Z$
which admit continuous and bounded $C$-Fr\'echet derivative. Moreover we call
$C^{0,1,C}_b(I\times H,Z)$ the space of functions $f:I\times H\to Z$
belonging to $C_b(I\times H,Z)$ and such that, for every $t\in I$,
$f(t,\cdot )\in C^{1,C}_b(H,Z)$. When $Z=\R$ we omit it.
\item We call $C^{2,C}_{b}(H,Z)$ the space of all functions $f$ in $C^1_b(H,Z)$
which admit continuous and bounded directional second order derivative $\nabla^C \nabla f$;
by $C^{0,2,C}_b(I\times H,Z)$ we denote the space of functions
$f \in C_b(I\times H,K)$ such that for every $t\in I$,
$f(t,\cdot )\in C^{2,C}_b(H,Z)$. When $Z=\R$ we omit it.
\item Here and in the following point of the Definition we take $Z=\R$. \newline
For any $\alpha\in(0,1)$ and $T>0$ (this time $I$ is equal to $[0,T]$) we denote
by $C^{0,1}_{\alpha}([0,T]\times H)$ the space of functions
$ f\in C_b([0,T]\times H))\cap C^{0,1}((0,T]\times H,Z)$ such that
the map $(t,x)\mapsto t^{\alpha} \nabla f(t,x)$
belongs to $C_b((0,T]\times H,H)$.
The space $C^{0,1}_{\alpha}([0,T]\times H)$
is a Banach space when endowed with the norm
\[
\left\Vert f\right\Vert _{C^{0,1}_{\alpha}([0,T]\times H) }=\sup_{(t,x)\in[0,T]\times H}
\vert f(t,x)\vert+
\sup_{(t,x)\in (0,T]\times H} t^{\alpha }\left\Vert \nabla f(t,x)\right\Vert_{H}.
\]
When clear from the context we will write simply
$\left\Vert f\right\Vert _{C^{0,1}_{\alpha}}$.
\noindent We also denote
by $C^{0,1,C}_{\alpha}([0,T]\times H)$ the space of functions
$ f\in C_b([0,T]\times H)\cap C^{0,1,C}((0,T]\times H)$ such that
the map $(t,x)\mapsto t^{\alpha} \nabla^C f(t,x)$
belongs to $C_b((0,T]\times H,K)$.
The space $C^{0,1,C}_{\alpha}([0,T]\times H)$
is a Banach space when endowed with the norm
\[
\left\Vert f\right\Vert _{C^{0,1,C}_{\alpha}([0,T]\times H) }=\sup_{(t,x)\in[0,T]\times H}
\vert f(t,x)\vert+
\sup_{(t,x)\in (0,T]\times H} t^{\alpha }\left\Vert \nabla^C f(t,x)\right\Vert_{K}.
\]
When clear from the context we will write simply
$\left\Vert f\right\Vert _{C^{0,1,C}_{\alpha}}$.
\item For any $\alpha\in(0,1)$ and $T>0$ we denote by $C^{0,2}_{\alpha}([0,T]\times H)$
the space of functions
$ f\in C_b([0,T]\times H)\cap C^{0,2}((0,T]\times H)$ such that for all $t\in(0,T],\,x\in H$ the map
$ (t,x)\mapsto t^{\alpha} \nabla^C\nabla f(t,x)$ is bounded and continuous as a map
from $(0,T]\times H$ with values in $H$.
The space $C^{0,2}_{\alpha}([0,T]\times H)$ turns out to be a Banach space if it is endowed with the norm
\begin{align*}
\Vert & f \Vert _{C^{0,2}_{\alpha}([0,T]\times H) }\\
&=\sup_{(t,x)\in[0,T]\times H}
\vert f(t,x)\vert+
\sup_{(t,x)\in[0,T]\times H} \left\Vert \nabla f\left( t,x\right) \right\Vert _{H}
+\sup_{(t,x)\in[0,T]\times H} t^{\alpha}
\left\Vert \nabla^2 f\left( t,x\right) \right\Vert _{H\times H}.
\end{align*}
We also denote by $C^{0,2,C}_{\alpha}([0,T]\times H)$
the space of functions
$ f\in C_b([0,T]\times H)\cap C^{0,2,C}((0,T]\times H)$ such that for all $t\in(0,T],\,x\in H$ the map
$ (t,x)\mapsto t^{\alpha} \nabla^C\nabla f(t,x)$ is bounded and continuous as a map
from $(0,T]\times H$ with values in $H\times K$.
The space $C^{0,2,C}_{\alpha}([0,T]\times H)$ turns out to be a Banach space if it is endowed with the norm
\begin{align*}
\Vert & f \Vert _{C^{0,2,C}_{\alpha}([0,T]\times H) }\\
&=\sup_{(t,x)\in[0,T]\times H}
\vert f(t,x)\vert+
\sup_{(t,x)\in[0,T]\times H} \left\Vert \nabla f\left( t,x\right) \right\Vert _{H}
+\sup_{(t,x)\in[0,T]\times H} t^{\alpha}
\left\Vert \nabla^C\nabla f\left( t,x\right) \right\Vert _{H\times K}.
\end{align*}
\end{itemize}
\end{definition}
\subsection{Setting of the problem and main assumptions}\label{section-statement}
In this paper we are concerned with the solution of an optimal control problem related to
an $n$-dimensional controlled equation with delay in the control, that we are going to introduce.\\
In a complete probability space $(\Omega, \calf, \P)$
we consider
\begin{equation}
\left\{
\begin{array}
[c]{l
dy(t) =a_0 y(t) dt+b_0 u(t) dt +\dis\int_{-d}^0b_1(\xi)u(t+\xi)d\xi+\sigma dW_t
,\text{ \ \ \ }t\in[0,T] \\
y(0) =y_0,\\
u(\xi)=u_0(\xi), \quad \xi \in [-d,0),
\end{array}
\right. \label{eq-contr-rit}
\end{equation}
where we assume the following.
\begin{hypothesis}\label{ipotesibasic}
\begin{itemize}
\item[]
\item[(i)] $W$ is a standard Brownian motion in $\R^k$, and $(\calf_t)_{t\geq 0}$ is the
augmented filtration generated by $W$;
\item[(ii)] $a_0\in \call(\R^n;\R^n)$, $\sigma$ is in $\call(\R^k;\R^n)$;
\item[(iii)] the control strategy $u$ belongs to $\calu$ where
$$\calu:=\left\lbrace z\in L^2_{\calp}(\Omega\times [0,T], \R^m):
u(t)\in U \;a.s.\right\rbrace $$
where $U$ is a closed subset of $\R^n$;
\item[(iv)] $d>0$ (the maximum delay the control takes to affect the system);
\item[(v)] $b_0 \in \call(\R^m;\R^n)$;
\item[(vi)]
$b_1\in L^2([-d,0],\call(\R^m;\R^n)).$
\end{itemize}
\end{hypothesis}
\medskip
Notice that assumption (vi) on $b_1$ does not cover the
case of pointwise delay since it is technically complicated to
deal with: indeed it gives rise, as we are going to see in next subsection,
to an unbounded control operator $B$, for this reason we leave the extension
of our approach to this case for further research.
As noticed in \cite{FGFM}, our results can be generalized to the case when the process $y$ is infinite dimensional, more precisely, when $y$
is the solution of the following controlled stochastic differential equation in an infinite dimensional Hilbert space $H$, with delay in the control:
\begin{equation}
\left\{
\begin{array}
[c]{l
dy(t) =A_0 y(t) dt+B_0 u(t) dt +\dis\int_{-d}^0B_1(\xi)u(t+\xi)d\xi+\sigma dW_t
,\text{ \ \ \ }t\in[0,T] \\
y(0) =y_0,\\
u(\xi)=u_0(\xi), \quad \xi \in [-d,0),
\end{array}
\right. \label{eq-contr-rit-inf}
\end{equation}
where $A_0$, $B_0$, $B_1$ are suitable operators, see \cite{FGFM} for more details.
\subsection{Infinite dimensional reformulation}
Following the approach of \cite{VK}, applied in \cite{GM} to the stochastic case, we reformulate equation (\ref{eq-contr-rit}) as an abstract stochastic
differential equation in the Hilbert space $\calh=\R^n\times L^2([-d,0],\R^n)$.
To this end we introduce the operator $A : \cald(A) \subset \calh
\rightarrow \calh$ as follows: for $(y_0,y_1)\in \calh$
\begin{equation}\label{A}
A(y_0 ,y_1 )=( a_0 y_0 +y_1(0), -y_1'), \qquad
\cald(A)=\left\lbrace(y_0,y_1)\in \calh:y_1\in W^{1,2}([-d,0],\R^n), y_1(-d)=0 \right\rbrace.
\end{equation}
We denote by $e^{tA}$ the $C_0$-semigroup generated by $A$: for
$y=(y_0,y_1)\in \calh$,
\begin{equation}
e^{tA} \left(\begin{array}{l}y_0 \\y_1\end{array}\right)=
\left(
\begin{array}
[c]{ll
e^{ta_0 }y_0+\int_{-d}^{0}1_{[-t,0]} e^{(t+s)a_0 } y_1(s)ds \\[3mm]
y_1(\cdot-t)1_{[-d+t,0]}(\cdot).
\end{array}
\right) \label{semigroup}
\end{equation}
We will use, for $N\in \N$ big enough, the resolvent operator
$(N-A)^{-1}$ which can be computed explicitly
giving
\begin{equation}
(N-A)^{-1} \left(\begin{array}{l}y_0 \\y_1\end{array}\right)=
\left(
\begin{array}
[c]{ll
(N-a_0)^{-1}\left[y_0+\int_{-d}^{0} e^{N s } y_1(s)ds\right]
\\[3mm]
\int_{-d}^{\cdot} e^{N(s-\cdot)} y_1(s)ds.
\end{array}
\right) \label{resolvent}
\end{equation}
Similarly, denoting by $e^{tA^*}=(e^{tA})^*$ the $C_0$-semigroup generated by $A^*$,
we have for
$z=\left(z_0,z_1\right)\in \calh $
\begin{equation}
e^{tA^*} \left(\begin{array}{l}z_0 \\z_1\end{array}\right)=
\left(
\begin{array}[c]{ll}
e^{t a_0^* }z_0 \\[3mm]
e^{(\cdot+t) a_0^* }z_0 1_{[-t,0]}(\cdot) +z_1(\cdot+t)1_{[-d,-t)}(\cdot).
\end{array}
\right) \label{semigroupadjoint}
\end{equation}
The infinite dimensional noise operator is defined as
\begin{equation}
\label{G}
G:\R^{k}\rightarrow \calh,\qquad Gy=(\sigma y, 0), \; y\in\R^k.
\end{equation}
The control operator $B$
is bounded and defined as
\begin{equation}
\label{B}
B:\R^{m}\rightarrow \calh,\qquad Bu=(b_0 u, b_1(\cdot)u), \; u\in\R^m
\end{equation}
and its adjoint is
\begin{equation}
\label{B*}
B^*:\calh^* \rightarrow \R^{m},\qquad B^*(x_0,x_1)=
b^*_0 x_0+\int_{-d}^0 b_1(\xi)^*x_1(\xi)d\xi, \; (x_0,x_1)\in\calh.
\end{equation}
Note that, in the case of pointwise delay the last term of the drift in
the state equation \myref{eq-contr-rit} is $u(t-d)$, hence $b_1(\cdot)$ is
a measure: the Dirac delta $\delta_{-d}$.
Hence in this case $B$ is unbounded as it takes values in
$\R^n \times C^*([-d,0],\R^n)$ (here we denote by $C^*([-d,0],\R^n)$ the dual space of $C([-d,0],\R^n)$).
Given any initial datum $(y_0,u_0)\in \calh$ and any admissible control $u\in \calu$ we call $y(t;y_0,u_0,u)$ (or simply $y(t)$ when clear from the context) the unique solution (which comes from standard results on SDE's, see e.g. \cite{IkedaWatanabe} Chapter 4, Sections 2 and 3)
of (\ref{eq-contr-rit}).
Let us now define the process
$Y=(Y_0,Y_1)\in L^2_\calp(\Omega \times [0,T],\calh)$ as
$$
Y_0(t)=y(t), \qquad Y_1(t)(\xi)=\int_{-d}^\xi u(\zeta+t-\xi)b_1(\zeta)d\zeta,
$$
where $y$ is the solution of equation (\ref{eq-contr-rit}) and $u$ is the control process in (\ref{eq-contr-rit}).
By Proposition 2 of \cite{GM},
the process $Y$
is the unique solution of the abstract evolution equation
in $\calh$
\begin{equation}
\left\{
\begin{array}
[c]{l}
dY(t) =AY(t) dt+Bu(t) dt+GdW_t
,\text{ \ \ \ }t\in[ 0,T] \\
Y(0) =y=(y_0,y_1),
\end{array}
\right. \label{eq-astr
\end{equation}
where $y_0=x_0$ and $y_1(\xi)=\dis\int_{-d}^\xi u_0(\zeta-\xi)b_1(\zeta)d\zeta$.
Note that we have $y_1\in L^2([-d,0];\R^n)$
Taking the integral (or mild) form of (\ref{eq-astr}) we have
\begin{equation}
Y(t) =e^{tA}y+\int_0^te^{(t-s)A}B u(s) ds +\int_0^te^{(t-s)A}GdW_s
,\text{ \ \ \ }t\in[ 0,T]. \\
\label{eq-astr-mild
\end{equation}
\subsection{Optimal Control problem}
The objective is to minimize, over all controls in $\calu$,
the following finite horizon cost:
\begin{equation}\label{costoconcreto}
J(t,x,u)=\E \int_t^T \left(\bar\ell_0(s,y(s))+\bar\ell_1(u(s))\right)\;ds +\E \bar\phi(x(T)).
\end{equation}
where $\bar\ell_0:[0,T]\times\R^n\rightarrow \R$ and
$\bar\phi:\R^n\rightarrow \R$ are continuous and bounded while
$\bar\ell_1:U\rightarrow\R$ is measurable and bounded from below.
Referring to the abstract formulation (\ref{eq-astr}) the cost in (\ref{costoconcreto}) can be rewritten also as
\begin{equation}\label{costoconcreto1}
J(t,x;u)=\E \left(\int_t^T \left[\ell_0(s,Y(s))+\ell_1(u(s))\right]\,ds + \phi(Y(T))\right),
\end{equation}
where
$\ell_0:[0,T]\times \calh\rightarrow \R$, $\ell_1:U\to \R$ are defined by setting
\begin{equation}\label{l_0}
\ell_0(t,x):=\bar\ell_0(t,x_0) \quad
\forall x=(x_0,x_1)\in \calh
\end{equation}
\begin{equation}\label{l_1}
\ell_1:=\bar\ell_1
\end{equation}
(here we cut the bar only to keep the notation homogeneous)
while $\phi :\calh\rightarrow \R$ is defined as
\begin{equation}\label{fi0}
\phi(x):=\bar\phi(x_0) \quad
\forall x=(x_0,x_1)\in \calh.
\end{equation}
Clearly, under the assumption above,
$\ell_0$ and $\phi$
are continuous and bounded while $\ell_1$ is measurable and bounded from below.
The value function of the problem is
\begin{equation}\label{valuefunction}
V(t,x):= \inf_{u \in \calu}J(t,x;u).
\end{equation}
As done in Subsection \ref{SS:motiv}, we define the Hamiltonian in a modified way
(see \myref{HminINTRObis}); indeed, for $p\in \calh$, $u \in U$,
we define the current value Hamiltonian $H_{CV}$ as
$$
H_{CV}(p\,;u):=\<p,u\>_{\R^m}+\ell_1(u)
$$
and the (minimum value) Hamiltonian by
\begin{equation}\label{psi1}
H_{min}(p)=\inf_{u\in U}H_{CV}(p\,;u),
\end{equation}
The associated HJB equation with unknown $v$ is then formally written as
\begin{equation}\label{HJBformale1}
\left\{\begin{array}{l}\dis
-\frac{\partial v(t,x)}{\partial t}=\frac{1}{2}Tr \;GG^*\nabla^2v(t,x)
+ \< Ax,\nabla v(t,x)\>_\calh +\ell_0(t,x)+ {H}_{min} (\nabla^Bv(t,x)),\qquad t\in [0,T],\,
x\in D(A),\\
\\
\dis v(T,x)=\phi(x).
\end{array}\right.
\end{equation}
Existence of mild solutions of \myref{HJBformale1} is proved in \cite{FGFM}, and the following assumptions are needed.
\begin{hypothesis}\label{ipotesicostoconcreto}
\begin{itemize}
\item[]
\item[(i)] $\phi\in C_b(\calh)$ and it is given by \myref{fi0} for
a suitable $\phi \in C_b(\R^n)$;
\item[(ii)] $\ell_0\in C_b([0,T] \times\calh)$ and it is given by \myref{l_0} for
a suitable $\bar\ell_0 \in C_b([0,T] \times \R^n)$;
\item[(iii)] $\ell_1:U\rightarrow\R$ is measurable and bounded from below;
\item[(iv)] the Hamiltonian $H_{min}:\R^m \to \R$ is Lipschitz continuous so
there exists $L>0$ such that
\begin{equation}\label{eq:Hlip}
\begin{array}{c}
\vert H_{min }(p_1)-H_{min }(p_2)\vert\leq L \vert p_1-p_2\vert
\quad \forall\,p_1,\,p_2\in\R^m;
\\[1.5mm]
\vert H_{min }(p)\vert\leq L(1 + \vert p\vert )
\quad \forall\,p\in\R^m.
\end{array}
\end{equation}
\end{itemize}
\end{hypothesis}
To get more regular solutions (well defined second derivative $\nabla^B\nabla$,
which will be used to prove existence of optimal feedback controls) we will need
the following further assumption.
\begin{hypothesis}\label{ipotesicostoconcretobis}
\begin{itemize}
\item[]
\item[(i)] $\ell_0$ is continuously differentiable in the variable $x$
with bounded derivative.
\item[(ii)] the Hamiltonian $H_{min}:\R^m \to \R$ is continuously differentiable
and, for a given $L>0$, we have, beyond \myref{eq:Hlip},
\begin{equation}\label{eq:Hlipder}
\begin{array}{c}
\vert \nabla H_{min }(p_1)-\nabla H_{min }(p_2)\vert\leq L \vert p_1-p_2\vert
\quad \forall\,p_1,\,p_2\in\R^m;
\end{array}
\end{equation}
\end{itemize}
\end{hypothesis}
\section{The Ornstein-Uhlenbeck semigroup and the HJB equation}
\label{section-smoothOU}
In the setting of Section \ref{section-statement}
we assume that Hypothesis \ref{ipotesibasic} holds true.
We take $\calh=\R^n \times L^2(-d,0;\R^n)$, $\Xi=\R^k$,
$(\Omega, \calf, \P)$ a complete probability space, $W$ a standard
Wiener process in $\Xi$, $A$ and $G$ as in (\ref{A}) and (\ref{G}).
Then, for $x\in \calh$, we take the Ornstein-Uhlenbeck process $X^x(\cdot)$
given by
\begin{equation}
\left\{
\begin{array}
[c]{l
dX(t) =AX(t) dt+GdW_t
,\text{ \ \ \ }t\ge 0\\
X(0) =x,
\end{array}
\right.\label{ornstein-gen}
\end{equation}
In mild form, the Ornstein-Uhlenbeck process $X^x$ is given by
\begin{equation}
X^x(t) =e^{tA}x +\int_0^te^{(t-s)A}GdW_s
,\text{ \ \ \ }t\ge 0. \\
\label{ornstein-mild-gen}
\end{equation}
$X$ is a Gaussian process, namely for every $t>0$, the law of
$X(t)$ is $\caln (e^{tA}x,Q_t)$, the Gaussian measure with mean $e^{tA}x$ and
covariance operator $Q_t$,
where
\[
Q_t=\int_0^t e^{sA}GG^*e^{sA^*}ds.
\]
The associated Ornstein-Uhlenbeck transition semigroup $R_t$ is defined by setting, for all $f\in B_b(\calh)$,
\begin{equation}
\label{ornstein-sem-gen}
R_t[f](x)=\E f(X^x(t))
=\int_K f(z+e^{tA}x)\caln(0,Q_t)(dz).
\end{equation}
Given any $\bar\phi\in B_b(\R^n)$, we define, as in (\ref{fi0})
a function $\phi \in B_b(\calh)$, by setting
\begin{equation}\label{fi}
\phi(x)=\bar\phi(x_0) \quad
\forall x=\left(
x_0 ,x_1 \right)\in \calh.
\end{equation}
For such functions, the Ornstein-Uhlenbeck semigroup $R_t$ is written as
\begin{equation}
\label{ornstein-sem-spec}
R_t[\phi](x)=\E \phi (X^x(t))=\E \bar\phi ((X^x(t))_0)=\int_\calh\bar\phi((z+e^{tA}x)_0)\caln(0,Q_t)(dz).
\end{equation}
Concerning the covariance operator $Q_t$, by Lemma 4.6 in \cite{FGFM} we have that
$$
\operatorname{Im} Q_t
=\operatorname{Im} Q_t^0\times \left\lbrace 0\right\rbrace\subseteq \R^n
\times \left\lbrace 0\right\rbrace$$
where $Q^0_t$ is the selfadjoint operator in $\R^n$ defined as
\begin{equation}\label{eq:Q^0def}
Q^0_t:=\int_0^t e^{sa_0}\sigma\sigma^*
e^{sa_0^*}\, ds.
\end{equation}
Then for every $(x_0,x_1) \in \calh$ we have
\begin{equation}\label{eq:Q^0Q}
Q_t\left(
x_0, x_1
\right) =\left(
Q^0_t x_0, 0
\right)
\end{equation}
and so
for every
$\phi :\calh\rightarrow\R$ defined in (\ref{fi}) we have
\begin{equation}
\label{eq:ornstein-sem-specbis}
R_t[\phi](x)=\int_{\R^n}\bar\phi \left(z_0+(e^{tA}x)_0 \right)\caln(0,Q_t^0)(dz_0).
\end{equation}
For the Ornstein Uhlenbeck transition semigroup we have the following regularizing property.
\begin{proposition}\label{lemmaderhpdeb}
Assume that Hypothesis \ref{ipotesibasic} holds. Assume moreover that, either
\begin{equation}\label{eq:hpdebreg}
\operatorname{Im}(e^{ta_0}b_0)\subseteq
\operatorname{Im}\sigma, \; \forall t > 0;
\qquad
\operatorname{Im}b_1(s)\in\operatorname{Im}\sigma,
\quad a.e.\, \forall s\in[-d,0]
\end{equation}
or
\begin{equation}\label{eq:hpdebregbis}
\operatorname{Im}\left(e^{ta_0}b_0 +\int_{-d}^0 1_{[-t,0]}e^{(t+r)a_0}b_1(dr)
\right)
\subseteq\operatorname{Im}\sigma,
\quad \forall t>0.
\end{equation}
Then, for any bounded measurable $\phi$ as in (\ref{fi}),
$R_{t}\left[\phi\right]$ is $B$-Fr\'echet differentiable for every
$t>0$, and, for every $h\in \R^m$,
$\<\nabla^{B}(R_{t}\left[\phi\right])(x),k\>_{\R^m}$ is given by
\begin{align}
&\<\nabla^B(R_{t}\left[\phi\right] )(x), k\>_{\R^m}
=\int_{\R^n}
\bar\phi\left(z_0+\left( e^{tA} x\right)_0\right)
\<(Q^0_{t})^{-1/2}
\left( e^{tA} Bk\right)_0, (Q^0_t)^{-1/2}z_0\>_{\R^n}\caln(0,Q^0_t)(dz_0).
\label{eq:formulaDRT1B}
\end{align}
Moreover, for every $k\in \R^m$,
\begin{equation}\label{eq:formulastimaDRT1B}
\vert \<\nabla^{B}(R_{t}\left[\phi\right] )(x),k\>_{\R^m}\vert
\leq \Vert \bar\phi\Vert_\infty
\left\Vert (Q^0_{t})^{-1/2}
\left( e^{tA} B\right)_0
\right\Vert_{\call(R^m,\R^n)} \; \vert k\vert_{\R^m},
\end{equation}so that for all $T>0$ there exists $C_T$ such that
\begin{equation}
\vert\< \nabla^{B}(R_{t}\left[\phi\right]) (x),k\>_{\R^m}\vert
\leq C_T t^{-{1/2}} \Vert \bar\phi \Vert_\infty \;\vert k\vert_{\R^m}.
\label{eq:stimaderBnew}
\end{equation}
\end{proposition}
\subsection{Regular mild solutions of the HJB equation}
\label{sec-HJB}
We now recall results proved in \cite{FGFM} about existence of mild solutions of the HJB equation (\ref{HJBINTRO}).
We also state results about the existence of second order derivatives that will be need in Section
\ref{sec:contr-feedback} to solve our control problem.
First of all we introduce some suitable spaces of differentiable functions in $C_b([0,T]\times\calh)$. We fix $\alpha \in (0,1)$, in the following
we will consider these spaces with $\alpha=1/2$.
\begin{definition}\label{df:Sigma}
Let $T>0$, $\alpha \in (0,1)$.
A function $g\in C_b([0,T]\times \calh)$
belongs to $\Sigma^1_{T,\alpha}$ if there exists a function
$f\in C^{0,1}_{\alpha}([0,T]\times \R^n)$ such that
$$g(t,x)=f\left(t,(e^{tA}x)_0\right),
\qquad \forall (t,x) \in [0,T]\times \calh.
$$
\end{definition}
If $g\in \Sigma^1_{T,\alpha}$, for any $t\in(0,T]$ the function $g(t,\cdot)$ is
both Fr\'echet differentiable and $B$-Fr\'echet differentiable.
Moreover, for $(t,x)\in [0,T]\times \calh$, $h \in \calh$, $k\in \R^m$,
$$
\<\nabla g(t,x),h\>_{\calh}=\<\nabla f\left(t,(e^{tA}x)_0\right),(e^{tA}h)_0\>_{\R^n},
\quad and \quad
\<\nabla^B g(t,x),k\>_{\R^m}=\<\nabla f\left(t,(e^{tA}x)_0\right),(e^{tA}Bk)_0\>_{\R^n}.
$$
This in particular imply that, for all $k\in \R^m$
\begin{equation}
\label{eq:nablaperSigma}
\<\nabla^B g(t,x),k\>_{\R^m}=\<\nabla g(t,x),Bk\>_{\R^n \times L^2([-d,0];\R^n)},
\end{equation}
which also means $B^*\nabla g = \nabla^B g$.
For later notational use we call $\bar f\in C_b((0,T]\times \R^n;\R^m)$
the function defined by
$$
\<\bar f(t,y),k\>_{\R^m}=
t^\alpha\<\nabla f\left(t,y\right),(e^{tA}Bk)_0\>_{\R^n},
\qquad (t,y)\in (0,T]\times \R^n, \quad k \in \R^m,
$$
which is such that
$$
t^\alpha\nabla^B g(t,x)=\bar f\left(t,(e^{tA}x)_0\right).
$$
We also notice that if $g\in \Sigma^1_{T,\alpha}$, then in order to have $g$ $B$-Fr\'echet differentiable it suffices to require $(e^{tA}B)_0$
bounded and continuous.
The set $\Sigma^1_{T,\alpha}$ is a closed subspace of $C^{0,1,B}_{\alpha}([0,T]\times \calh)$.
Next, in analogy to what we have done defining $\Sigma^1_{T,\alpha}$,
we introduce a subspace $\Sigma^2_{T,\alpha}$ of functions
$g \in C^{0,2,B}_{\alpha}([0,T]\times \calh)$
that depends in a special way on the variable $x\in\calh$.
\begin{definition}\label{df:Sigma2}
A function $g\in C_b([0,T]\times \calh)$ belongs to $\Sigma^2_{T,\alpha}$
if there exists a function
$f\in C^{0,2}_{\alpha}([0,T]\times \R^n)$ such that
for all $(t,x) \in [0,T]\times \calh$,
\begin{align*}
&g(t,x)=f\left(t,(e^{tA}x)_0\right).
\end{align*}
\end{definition}
If $g\in \Sigma^2_{T,\alpha}$ then for any $t\in(0,T]$ the function $g(t,\cdot)$
is Fr\'echet differentiable and
$$
\<\nabla g(t,x),h\>_\calh=
\<\nabla f\left(t,(e^{tA}x)_0\right),(e^{tA}h)_0\>_{\R^n},
\quad \hbox{for $(t,x)\in [0,T]\times \calh$, $h \in \calh$.}
$$
Moreover also $\nabla g(t,\cdot)$ is $B$-Fr\'echet differentiable and
$$
\<\nabla^B\left(\nabla g(t,x)h\right)),k\>_{\R^m}=
\<\nabla^2 f\left(t,(e^{tA}x)_0\right)(e^{tA}h)_0,(e^{tA}Bk)_0 \>_{\R^n},
\quad \hbox{for $(t,x)\in [0,T]\times \calh$, $h \in \calh$, $k \in \R^m$.}
$$
We also notice that, since the function $f$ is twice continuously
Fr\'echet differentiable the second order derivatives
$\nabla^B\nabla g$ and $\nabla\nabla^B g$ both exist and coincide:
$$
\<\nabla^B\<\nabla g(t,x),h\>_{\calh},k\>_{\R^m}
=\<\nabla\<\nabla^B g(t,x),k\>_{\R^m},h\>_{\calh}.
$$
Again for later notational use we call
$\bar f_1\in C_b([0,T]\times \R^n;\R^m)$
the function defined by
$$
\<\bar f_1(t,y),h\>_{\R^m}=
\<\nabla f\left(t,y\right),(e^{tA}Bh)_0\>_{\R^n},
\qquad (t,y)\in [0,T]\times \R^n, \quad h \in R^m,
$$
which is such that
$$
\nabla^B g(t,x)=\bar f_1\left(t,(e^{tA}x)_0\right).
$$
Similarly we call
$\bar{\bar f}\in C_b\left((0,T]\times \R^n;\call(\calh,\R^m)\right)$
the function defined by
$$
\<\<\bar{\bar f}(t,y),h\>_\calh,k\>_{\R^m}=
t^\alpha\<\nabla^2 f\left(t,y\right)(e^{tA}h)_0,(e^{tA}Bk)_0 \>_{\R^n}
\qquad (t,y)\in [0,T]\times \R^n, \quad h \in \calh, \; k\in R^m,
$$
which is such that
$$
t^\alpha\nabla^B\nabla g(t,x)=t^\alpha\nabla\nabla^B g(t,x)=
\bar{\bar f}\left(t,(e^{tA}x)_0\right).
$$
When (\ref{eq:hpdebreg}) or
(\ref{eq:hpdebregbis}) hold we know, by Proposition \ref{lemmaderhpdeb},
that the function $g(t,x)=R_t[\phi](x)$
for $\phi$ given by (\ref{fi}) with $\bar\phi$ bounded and continuous,
belongs to $\Sigma^1_{T,1/2}$.
\begin{lemma}\label{lemma reg-convpercontr}
Let
(\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold true.
Let $T>0$ and let $\psi:\R^m\rightarrow\R$ be a continuous
function satisfiying Hypothesis (\ref{ipotesicostoconcreto}), estimates
(\ref{eq:Hlip}).
Then
\begin{itemize}
\item [i)]
for every $g \in \Sigma^1_{T,{1/2}}$, the function
$\hat g:[0,T]\times \calh\rightarrow \R$ belongs to $\Sigma^1_{T,{1/2}}$ where
\begin{equation}\label{iterata-primag}
\hat g(t,x) =\int_{0}^{t}
R_{t-s} [\psi(\nabla^{B}g(s,\cdot))](x) ds.
\end{equation}
Hence, in particular, $\hat g(t,\cdot)$
is $B$-Fr\'echet differentiable for every $t\in (0,T]$ and, for all $x\in \calh$,
\begin{equation}\label{stimaiterata-primag}
\left\vert \nabla ^B(\hat g(t,\cdot))(x) \right\vert_{(\R^m)^*} \leq
C\left(t^{{1/2}}+\Vert g \Vert_{C^{0,1,B}_{{1/2}}}\right).
\end{equation}
If $\sigma$ is onto, then $\hat g(t,\cdot)$
is Fr\'echet differentiable for every $t\in (0,T]$ and, for all $h\in\calh$, $x\in \calh$,
\begin{equation}\label{stimaiterata-primagbis}
\left\vert\nabla(\hat g(t,\cdot))(x) \right\vert_{\calh^*}
\leq
C\left(t^{{1/2}}+\Vert g \Vert_{C^{0,1}_{{1/2}}}\right).
\end{equation}
\item [ii)] Assume moreover that $\psi\in C^1(\R^m)$.
For every $g \in \Sigma^2_{T,{1/2}}$, the function
$\hat g$ defined in (\ref{iterata-primag})
belongs to $\Sigma^2_{T,{1/2}}$. Hence, in particular, the second order derivatives
$\nabla\nabla^B\hat g(t,\cdot)$ and $\nabla^B\nabla\hat g(t,\cdot)$ exist, coincide and
for every $t\in (0,T]$ and, for all $x\in \calh$,
\begin{equation}\label{stimaiterata-secondag}
\left\vert \nabla ^B\nabla(\hat g(t,\cdot))(x) \right\vert_{\calh^*\times(\R^m)^*} \leq
C
\Vert g \Vert_{C^{0,2,B}_{{1/2}}}
\end{equation}
If $\sigma$ is onto, then $\hat g(t,\cdot)$
is twice Fr\'echet differentiable and for every $t\in(0,T]$, for all $h\in\calh$ and $x\in \calh$,
\begin{equation}\label{stimaiterata-secondagbis}
\left\vert\nabla^2(\hat g(t,\cdot))(x) \right\vert_{\calh^*\times \calh^*}
\leq
C
\Vert g \Vert_{C^{0,2}_{{1/2}}}
\end{equation}
\end{itemize}
\end{lemma}
Now we introduce mild solutions to the HJB equation (\ref{HJBformale1}). By applying formally the variation of
constants formula to (\ref{HJBformale1}) we get the integral equation satisfied by the mild solution, that we rewrite here for the reader
convenience:
\begin{equation}
v(t,x) =R_{T-t}[\phi](x)+\int_t^T R_{s-t}\left[
H_{min}(\nabla^B v(s,\cdot))+\ell_0(s,\cdot)\right](x)\; ds,\qquad t\in [0,T],\
x\in H.\label{solmildHJB}
\end{equation}
We use this formula to give the notion of mild
solution for the HJB equation (\ref{HJBformale1}).
\begin{definition}\label{defsolmildHJB}
We say that a
function $v:[0,T]\times \calh\rightarrow\mathbb{R}$ is a mild
solution of the HJB equation (\ref{HJBformale1}) if the following
are satisfied:
\begin{enumerate}
\item $v(T-\cdot, \cdot)\in C^{0,1,B}_{{1/2}}\left([0,T]\times \calh\right)$;
\item equality (\ref{solmildHJB}) holds on $[0,T]\times \calh$.
\end{enumerate}
\end{definition}
\begin{theorem}\label{esistenzaHJB}
Let Hypotheses \ref{ipotesibasic} and \ref{ipotesicostoconcreto} hold and
let (\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold.
Then the HJB equation (\ref{HJBformale1})
admits a mild solution $v$ according to Definition \ref{defsolmildHJB}.
Moreover $v$
is unique among the functions $w$ such that $w(T-\cdot,\cdot)\in\Sigma_{T,1/2}$ and it satisfies, for suitable $C_T>0$, the estimate
\begin{equation}\label{eq:stimavmainteo}
\Vert v(T-\cdot,\cdot)\Vert_{C^{0,1,B}_{{1/2}}}\le C_T\left(\Vert\bar\phi \Vert_\infty
+\Vert\bar\ell_0 \Vert_\infty \right).
\end{equation}
Finally if the initial datum $\phi$ is also continuously $B$-Fr\'echet
(or Fr\'echet) differentiable,
then $v \in C^{0,1,B}_{b}([0,T]\times \calh)$ and, for suitable $C_T>0$,
\begin{equation}\label{eq:stimavmainteobis}
\Vert v\Vert_{C^{0,1,B}_{b}}\le C_T\left(\Vert\phi \Vert_\infty
+\Vert\nabla^B\phi \Vert_\infty+\Vert\ell_0 \Vert_\infty \right)
\end{equation}
(substituting $\nabla^B\phi$ with $\nabla\phi$ if $\phi$ is Fr\'echet differentiable).
If $\sigma$ is onto, then the mild solution of equation (\ref{HJBformale1})
found in the previous theorem
is also Fr\'echet differentiable,
and the following estimate holds true
\begin{equation}\label{stimadiffle-solHJB}
\left\Vert v(T-\cdot,\cdot)\right\Vert _{C^{0,1}_{{1/2}}
}\leq C_T\left(\Vert\bar\phi \Vert_\infty
+\Vert\bar\ell_0 \Vert_\infty \right)
\end{equation}
If the coefficients are regular, we have further regularity on the mild solution, namely:
\begin{itemize}
\item[(i)]
if $\phi$ is continuously differentiable then we have $v \in \Sigma^2_{T,{1/2}}$,
hence the second order derivatives $\nabla^B\nabla v$ and $\nabla\nabla^B v$
exist and are equal.
Moreover there exists a constant $C>0$ such that
\begin{equation}\label{stimanablav}
\vert \nabla v(t,x)\vert \leq C\left( \Vert \nabla\bar\phi\Vert_\infty+
\Vert \nabla\bar\ell_0\Vert_\infty\right),
\end{equation}
\begin{equation}\label{stimanablav^2}
\vert \nabla^B\nabla v(t,x)\vert= \vert \nabla\nabla^B v(t,x)\vert
\leq C\left( (T-t)^{-{1/2}}
\Vert \nabla\bar\phi\Vert_\infty+(T-t)^{{1/2}}
\Vert \nabla\bar\ell_0\Vert_\infty\right).
\end{equation}
Finally, if $\sigma$ is onto, then also $\nabla^2 v$ exists and is continuous
and, for suitable $C>0$,
\begin{equation}\label{stimanablav^2bis}
\vert \nabla^2 v(t,x)\vert
\leq C\left( (T-t)^{-{1/2}}
\Vert \nabla\bar\phi\Vert_\infty+(T-t)^{{1/2}}
\Vert \nabla\bar\ell_0\Vert_\infty\right).
\end{equation}
\item[(ii)]
If $\phi$ is only continuous then the function
$(t,x)\mapsto (T-t)^{1/2}v(t,x)$ belongs to $\Sigma^2_{T,{1/2}}$.
Moreover there exists a constant $C>0$ such that
\begin{equation}\label{stimanablavreg}
\vert \nabla v(t,x)\vert \leq C\left((T-t)^{-1/2} \Vert\bar\phi\Vert_\infty+
\Vert \nabla\bar\ell_0\Vert_\infty\right),
\end{equation}
\begin{equation}\label{stimanablav^2reg}
\vert \nabla^B\nabla v(t,x)\vert= \vert \nabla\nabla^B v(t,x)\vert
\leq C\left( (T-t)^{-1}
\Vert \bar\phi\Vert_\infty+(T-t)^{-1/2}
\Vert \nabla\bar\ell_0\Vert_\infty\right).
\end{equation}
Finally, if $\sigma$ is onto, then also $\nabla^2 v$ exists and is continuous
in $[0,T)\times \calh$
and, for suitable $C>0$,
\begin{equation}\label{stimanablav^2bisreg}
\vert \nabla^2 v(t,x)\vert
\leq C\left( (T-t)^{-1}
\Vert \nabla\bar\phi\Vert_\infty+(T-t)^{-1/2}
\Vert \nabla\bar\ell_0\Vert_\infty\right).
\end{equation}
\end{itemize}
\end{theorem}
\section{$\calk$-convergence and approximations of solutions of the HJB equation}
\label{sec-Kstrong}
We first introduce the notion of $\calk$-convergence, following \cite{Ce95} and \cite{G1}.
\begin{definition}\label{k-conv}
A sequence $(f_n)_{n\geq 0}\in C_b(\calh)$
is said to be $\calk$-convergent to a function $f\in C_b(H)$ (and we shall write
$f_n\overset{\calk}{\rightarrow} f$ or $f=\calk-\lim_{n\rightarrow\infty}f_n$) if for any compact set
$\calk\subset\calh$
$$
\sup_{n\in\N}\Vert f_n\Vert_\infty<+\infty \quad {\rm and } \quad
\lim_{n\rightarrow\infty}\sup_{x\in\calk}\vert f(x)-f_n(x)\vert =0.
$$
\noindent
Similarly, given $I\subseteq \R$, a sequence
$(f_n)_{n\geq 0}\in C_b(I \times \calh)$ is said to be $\calk$-convergent to a function $f\in C_b(I\times \calh)$ (and we shall write again
$f_n\overset{\calk}{\rightarrow} f$ or $f=\calk-\lim_{n\rightarrow\infty}f_n$) if for any compact set
$\calk\in\calh$ and for any $(t,x)\in I\times\calk $ we have
$$
\sup_{n\in\N}\Vert f_n\Vert_\infty<+\infty \quad {\rm and } \quad
\lim_{n\rightarrow\infty}\sup_{(t,x)\in I\times\calk}\vert f(t,x)-f_n(t,x)\vert=0.
$$
\end{definition}
Now we recall the definition (given in \cite{DP3}, beginning of Chapter 7)
of \textit{strict solution}
of a family of Kolmogorov equations.
Consider the following forward Kolmogorov equation with unknown $w$:
\begin{equation}\label{eq:Kforward}
\left\{\begin{array}{l}\dis
\frac{\partial w(t,x)}{\partial t}=\frac{1}{2} Tr \;GG^*\; \nabla^2w(t,x)
+ \< Ax,\nabla w(t,x)\>
+\calf(t,x)
\qquad t\in [0,T],\,
x\in \calh,\\
\\
\dis w(0,x)=\phi(x).
\end{array}\right.
\end{equation}
where
the functions
$\calf:[0,T]\times \calh \to \R$ and $\phi:\calh\to\R$ are bounded and continuous.
\begin{definition}\label{df:strictandpistrong}
By \textit{strict solution} of the Kolmogorov equation (\ref{eq:Kforward}) we mean
a function $w$ such that
\begin{equation}\label{propr-strict-sol}
\left\lbrace
\begin{array}{l}
w \in C_b([0,T]\times \calh)\quad {\rm and}\quad w(0,x)=\phi(x)
\\
w(t,\cdot) \in UC^2_b(\calh), \; \forall t\in [0,T];
\\
w(\cdot, x)\in C^{1}([0,T]),\;\forall x\in D(A)\; \hbox{and $w$ satisfies (\ref{eq:Kforward}).}
\end{array}
\right.
\end{equation}
\end{definition}
Now we prove a key approximation lemma.
\begin{lemma}\label{lm:approximation}
Let Hypothesis \ref{ipotesibasic} and \ref{ipotesicostoconcreto} hold.
Let also (\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold.
Let $v$ be the mild solution of the HJB equation (\ref{HJBformale1})
and set $w(t,x)=v(T-t,x)$ for $(t,x)\in [0,T]\times \calh$.
Then there exist three sequences of functions $(\bar\phi_n)$, $(\bar \ell_{0,n})$ and $(\calf_n)$ such that, for all $n\in \N$,
\begin{equation}\label{eq:defapproxphif}
\bar\phi_n \in C_c^\infty(\R^n), \qquad
\bar \ell_{0,n} \in C_c^\infty([0,T]\times \R^n), \qquad
\calf_n \in C_c^\infty([0,T]\times \calh) \cap \Sigma_{T,{1/2}}
\end{equation}
and
\begin{equation}\label{eq:convapproxphif}
\bar\phi_n \rightarrow \bar\phi ,
\qquad \bar \ell_{0,n} \rightarrow \bar \ell ,
\qquad\calf_n \to H_{min} (\nabla^B w)
\end{equation}
in the sense of $\calk$-convergence.
Moreover, defining $\phi_n(x)=\bar\phi_n(x_0)$, $\ell_{0,n}(s,x)=\bar\ell_{0,n}(s,x_0)$
and
\begin{equation}\label{v_n-mild}
w_n(t,x):= R_t \phi_n + \int_0^t R_{t-s}[\calf_n (s,\cdot)+\ell_{0,n}(s,\cdot)] (x)ds
\end{equation}
the following hold:
\begin{itemize}
\item $w_n \in UC^{1,2}_b ([0,T]\times \calh) \cap \Sigma_{T,{1/2}}$,
\item $w_n$ is a strict solution of
\myref{eq:Kforward} with $\phi_n$ in place of $\phi$ and
$\calf_n + \ell_{0,n}$ in place of $\calf $,
\item we have, in the sense of $\calk$-convergence
(the first in $[0,T]\times \calh$,
the second in $(0,T]\times \calh$),
\begin{equation}\label{eq:convapproxphifsol}
w_n \rightarrow w ,
\qquad t^{{1/2}}\nabla^B w_n \to t^{{1/2}}\nabla^B w.
\end{equation}
\end{itemize}
\end{lemma}
\dim
We divide the proof in three steps.
{\bf Step 1: choosing the three approximating sequences.}
We choose $\bar\phi_n$ and $\bar\ell_{0,n}$ to be
the standard approximations by convolution of $\bar \phi$ and $\bar \ell_0$.
To define $\calf_n$ we observe first that, since $w \in \Sigma^1_{T,{1/2}}$, then the function
$$
(t,x) \mapsto \calf(t,x):= H_{min} (\nabla^B w(t,x))
$$
has the property that there exist $f:[0,T)\times \R^n \to \R$, continuous and bounded, such that
$$
\calf (t,x)=t^{-{1/2}} f(t,(e^{tA}x)_0).
$$
We then let $f_n$ be the approximation by convolution of $f$ and define
$$
\calf_n (t,x)=t^{-{1/2}} f_n(t,(e^{tA}x)_0).
$$
{\bf Step 2: proof that $w_n\in UC^{1,2}_b([0,T]\times\calh) \cap \Sigma^1_{T,{1/2}}$ and
that it is a strict solution.}
The fact that $w_n\in \Sigma_{T,{1/2}}$ follows immediately from \myref{v_n-mild},
Proposition \ref{lemmaderhpdeb} and Lemma \ref{lemma reg-convpercontr}-(i).
Differentiability with respect to the variable $x$ follows by applying the dominated convergence theorem,
while explicitely differentiating $ R_{t}\left[\phi\right] $, namely or any $h\in\calh$
\begin{align*}
\<\nabla R_{t}\left[\phi\right] (x),h\>_H
&=\lim_{\alpha\rightarrow 0}\frac{1}{\alpha}
\left[\int_\calh\phi\left(z+e^{tA}(x+\alpha h)\right)
\caln\left(0,Q_{t}\right) (dz)
-\int_\calh
\phi\left(z+e^{tA}x\right) \caln\left( 0,Q_{t}\right) (dz) \right]
\\[2mm]
& =\int_\calh\lim_{\alpha\rightarrow 0}\frac{1}{\alpha}
\left[\phi\left(z+e^{tA}(x+\alpha h)\right)-\phi\left(z+e^{tA}x\right)\right]
\caln\left(0,Q_{t}\right) (dz)\\[2mm]
& =\int_\calh\<\nabla \phi\left(z+e^{tA}x\right), e^{tA}h\>_H
\caln\left(0,Q_{t}\right) (dz)=R_{t}\left[\<\nabla \phi , e^{tA}h\>_H\right ] (x).
\end{align*}
In a similar way, differentiating twice
we get that, for all $h,k \in \calh$
\begin{align*}
& \<\nabla^2 R_{t}\left[\phi_n\right] (x)h,k\>_\calh=
R_{t}\left[\<\nabla^2 \phi_n e^{tA}h,e^{tA}k\>_\calh\right ] (x).
\end{align*}
Similarly we have, for the convolution term containing $\calf_n$,
\begin{align*}
& \<\nabla \int_0^t R_{t-s}\left[\calf_n(s,\cdot)\right] (x),h\>_\calh
\\[2mm]
& =\lim_{\alpha\rightarrow 0}\frac{1}{\alpha}
\int_0^t\left[\int_{\calh}\calf_n\left(s,z+e^{(t-s)A}(x+\alpha h)\right)
\caln\left(0,Q_{t-s}\right) (dz)
-\int_{\calh}
\calf_n\left(s,z+e^{(t-s)A}x\right) \caln\left( 0,Q_{t-s}\right) (dz) \,ds\right]
\\[2mm]
& =\int_0^t\int_{\calh}\lim_{\alpha\rightarrow 0}\frac{1}{\alpha}
\left[\calf_n\left(s,z+e^{(t-s)A}(x+\alpha h)\right)-\calf_n\left(s,z+e^{(t-s)A}x\right)\right]
\caln\left(0,Q_{t-s}\right) (dz)\,ds\\[2mm]
& =\int_0^t\int_{\calh}\<\nabla \calf_n\left(s,z+e^{(t-s)A}x\right), e^{(t-s)A}h\>_\calh
\caln\left(0,Q_{t-s}\right) (dz)\,ds\\[2mm]
&=\int_0^tR_{t-s}\left[\<\nabla \calf_n(s,\cdot) ,e^{(t-s)A}h\>_\calh\right ] (x)\,ds,
\end{align*}
and also, arguing in the same way,
\begin{align*}
& \<\nabla^2\int_0^t R_{t}\left[\calf_n(s,\cdot)\right] (x)\,ds\, h,k\>\calh=
\int_0^tR_{t-s}\left[\<\nabla^2 \calf_n(s,\cdot) e^{(t-s)A}h,e^{(t-s)A}k\>_\calh\right ] (x)\,ds.
\end{align*}
The convolution term involving $\ell_{0,n}$ is treated exactly in the same way.
The proof that $w_n$ is differentiable with respect to time and that
$w_{nt}\in UC_b([0,T]\times\calh) $ is completely analogous
to what is done in \cite[Theorems 9.23 and 9.25]{DP1})
for homogeneous Kolmogorov equations and we omit it\footnote{The proof
in the nonhomogeneous case can be found in the forthcoming book \cite[Section 4.4]{FabbriGozziSwiech}.}.
By Theorem 5.3 in \cite{CeGo}, see also Theorem 7.5.1 in \cite{DP3} for Kolmogorov equations,
we finally conclude that $w_n$
is a strict solution to equation \ref{eq:Kforward}.
{\bf Step 3: proof of the convergences.}
First we prove that the sequences $(w_n)$ and $(t^{1/2}\nabla^B w_n)$
are bounded uniformly with respect to $n$. Indeed,
by (\ref{v_n-mild}) and by the properties of convolutions,
\begin{align*}
&\vert w_n(t,x)\vert
\leq \Vert \bar\phi_n\Vert_\infty +
\int_0^t\sup_{x\in\calh}\left[\vert\calf_n (s,x)\vert+\vert
\ell_{0,n}(s,x)\vert\right]\,ds
\\
&\leq
\Vert \bar\phi\Vert_\infty + \int_0^t\sup_{y\in\R^n}\left[s^{-{1/2}} \vert
f_n (s,y)\vert+\vert
\bar \ell_{0,n}(s,y)\vert\right]\,ds \leq
\Vert \bar\phi\Vert_\infty + \int_0^t\sup_{y\in\R^n}\left[ s^{-{1/2}}\vert f (s,y)\vert+\vert
\bar \ell_{0}(s,y)\vert\right]\,ds
\end{align*}
Moreover, using Proposition \ref{lemmaderhpdeb} and the results in \cite{FGFM}, Section 5, formula 5.9, with $\psi=identity$, namely
\begin{align}
&\<\nabla ^B \left(\int_{0}^{t}
R_{t-s}\left[\nabla^{B}(g(s,\cdot)\right] ds
\right) (x),k \>_{\R^m}=
\label{eq:derBconvnew}
\\[3mm]
\nonumber
&=
\int_{0}^{t} \int_\calh
s^{-{1/2}} \bar f\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
\<(Q^0_{t-s})^{-{1/2}}
\left( e^{tA} Bk\right)_0, (Q^0_{t-s})^{-{1/2}}z_0\>_{\R^n}
\caln\left(0,Q_{t-s}\right)(dz)ds,
\end{align}
we get
\begin{equation*}
\vert t^{1/2} \nabla^B w_n(t,x)\vert
\leq C\Vert \bar\phi\Vert_\infty
+Ct^{1/2}\int_0^ts^{-{1/2}}(t-s)^{-{1/2}}\sup_{y\in\R^n}\left(\vert
f\left(s, y\right)\vert+ \vert\bar \ell_{0}\left(s, y\right)
\vert\right)
ds,
\end{equation*}
for a suitable $C>0$.
Now, with similar computations, we prove the convergences. Indeed,
\begin{align*}
\vert v_n(t,x)-v(t,x)\vert&\leq \int_\calh\vert \bar\phi_n\left( (e^{tA}z)_0+(e^{tA}x)_0\right)-
\bar\phi\left( (e^{tA}z)_0+(e^{tA}x)_0\right)\vert\caln(0,Q_{t})(dz) \\
& +
\int_0^t\int_\calh\left[\vert s^{-{1/2}} \calf_n \left(s,(e^{sA}z)_0+(e^{tA}x)_0\right)
-
s^{-{1/2}} \calf \left(s,(e^{sA}z)_0+(e^{tA}x)_0\right)\vert\right.\\
&\left.+\vert
\ell_{0,n}\left(s,(e^{sA}z)_0+(e^{tA}x)_0\right)-
\ell_0\left(s,(e^{sA}z)_0+(e^{tA}x)_0\right)
\vert\right]\caln(0,Q_{t-s})(dz)\,ds.
\end{align*}
Since, for every compact set $\calk \subset \calh$ the set
$\{(e^{tA}x)_0, \; t \in [0,T],\, x \in \calk\}$ is compact in $\R^n$,
then by the Dominated Convergence Theorem we get that for any compact set $\calk \subset \calh$
\begin{equation}
\label{eq:convvnnew}
\sup_{(t,x)\in[0,T]\times\calk} \vert v_n(t,x)-v(t,x)\vert\rightarrow 0.
\end{equation}
Moreover, using again Propositions \ref{lemmaderhpdeb}
and \ref{lemma reg-convpercontr}-(i),
for a suitable $C>0$, we get
\begin{align*}
&\vert \nabla^B v_n(t,x)- \nabla^B v(t,x)\vert\\
&\leq C
\int_\calh\left( \bar\phi_n\left( (e^{tA}z)_0+(e^{tA}x)_0\right)-
\bar\phi\left( (e^{tA}z)_0+(e^{tA}x)_0\right)\right)\<(Q^0_{t})^{-{1/2}}
\left( e^{tA} Bh\right)_0, (Q^0_{t})^{-{1/2}}z_0\>_{\R^n}\caln(0,Q_{t})(dz)\\
& +C\int_0^t
\int_\calh \left(\left\vert
s^{-{1/2}} f_n\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
-s^{-{1/2}} f\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
\right\vert\right.\\
&\left. +\left\vert \bar \ell_{0,n}\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
-\bar \ell_{0}\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)\right\vert\right)\\
&\quad \left \vert
\<(Q^0_{t-s})^{-{1/2}}
\left( e^{(t-s)A} Bh\right)_0, (Q^0_{t-s})^{-{1/2}}z_0\>_{\R^n}
\right\vert\caln(0,Q_{t-s})(dz) ds.
\end{align*}
By Proposition \ref{lemmaderhpdeb} we know that for suitable $C>0$,
\[
\int_\calh\vert\<(Q^0_{t})^{-{1/2}}
\left( e^{tA} Bh\right)_0, (Q^0_{t})^{-{1/2}}z_0\>_{\R^n}\vert^2\caln(0,Q_{t})(dz)
\leq C\Vert (Q^0_{t})^{-{1/2}}
\left( e^{tA} Bh\right)_0\Vert^2\leq \frac{C}{t}.
\]
Hence, applying Cauchy-Schwartz inequality we get
\begin{align*}
&\vert \nabla^B v_n(t,x)- \nabla^B v(t,x)\vert\\
&\leq C t^{-{1/2}}
\vert\int_\calh\vert \bar\phi_n\left( (e^{tA}z)_0+(e^{tA}x)_0\right)-
\bar\phi_n\left( (e^{tA}z)_0+(e^{tA}x)_0\right)\vert^2\caln(0,Q_{t})(dz)\vert^{1/2}\\
&+\int_0^t (t- s)^{-{1/2}}
\left( \int_\calh \left(\left\vert
s^{-{1/2}} f_n\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
-s^{-{1/2}} f\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
\right\vert\right.\right.\\
&\left.\left.+ \left\vert \bar \ell_{0,n}\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)
-\bar \ell_{0}\left(s, (e^{sA}z)_0+(e^{tA}x)_0\right)\right\vert\right)^2\caln(0,Q_{t-s})(dz)
\right)^{1/2}.
\end{align*}
Applying the Dominated Convergence Theorem as for the proof of
\myref{eq:convvnnew} we get the final claim
\[
\sup_{x\in(0,T]\times\calk} \vert t^{1/2}\nabla^B v_n(t,x)-t^{1/2}\nabla^B v(t,x)\vert\rightarrow 0, \qquad \hbox{for any compact set $\calk\subset\calh$.}
\]
\qed
\medskip
Notice that, using the terminology of \cite{CeGo,G1}, the above result
implies that a mild solution (\ref{eq:Kforward}) is also a {\em $\calk$-strong solution}.
In general, in an infinite dimensional Hilbert space $H$, existence of $\calk$-strong solutions is not a routine application of the theory of evolution equations, as the operator $\call$ formally introduced in \ref{eq:ell}
is not the infinitesimal generator of a strongly continuous semigroup in the Banach space $C_b(H)$. To overcome this difficulty in the already mentioned paper
\cite{CeGo} the theory of weakly continuous (or $\calk$-continuous)
semigroups has been used.
\begin{remark}\label{rm:defpistrictstrong}
The approximation results proved just above is needed to prove the fundamental identity,
(see next Proposition \ref{prop rel fond}) which is the key point to get the verification theorem and the existence of optimal feedback controls.
The idea is to apply Ito's formula to the approximating sequence $w_n$ composed with
the state process $Y$ and then to pass to the limit for $n\to +\infty$
(see e.g. \cite{G1} or \cite[Section 4.4]{FabbriGozziSwiech}).
However in the literature the approximating sequence is taken more regular,
i.e. the $w_n$ are required to be classical solutions
(see e.g. \cite[Section 6.2, p.103]{DP3}) of \myref{eq:Kforward}.
This in particular means that $\nabla w_n\in D(A^*)$ and this fact is crucial
since it allows to apply Ito's formula without requiring that the
state process $Y$ belongs to $D(A)$, which would be a too strong requirement.
In our case the used approximating procedure does not give rise in general to
functions $w_n$ with $\nabla w_n\in D(A^*)$.
Indeed for our purposes we need that the approximants of the data $\phi,\,\ell_0, \,\calf$ remain all in the space $\Sigma_{T,{1/2}}$; without this, since we only have
``partial'' smoothing,
it is not clear at all how to prove the convergence of the derivative $\nabla^B w_n$
(which is needed when we pass to the limit to prove the fundamental identity in next subsection).
Hence, in particular, since we need that, for all $n\in\N$, $\calf_n\in\Sigma^1_{T,{1/2}}$, and since
$\calf$ is written in terms of $f$, we approximate $f$
by $f_n$, and this procedure gives the approximants $\calf_n$ of $\calf$.
In this way $\calf_n\in \Sigma^1_{T,{1/2}}$ but $\nabla \calf_n \notin D(A^*)$.
Summing up, we are only able to find approximating strict solutions
and not classical solutions. Since the state process $Y$ does not belong to $D(A)$
this fact will force us to introduce suitable regularizations $Y_k$ of it
(see the proof of Proposition \ref{prop rel fond}).
\end{remark}
\begin{remark}\label{rm:defpistrictstrongbis}
Calling $\ell_n:=\ell_{0,n}+\calf_n -H_{min}(\nabla^Bw_n)$
is not difficult to see that the sequence
$w_n$ is a strict solution of the approximating HJB equation
\begin{equation*}
\left\{\begin{array}{l}\dis
\frac{\partial w(t,x)}{\partial t}=
\frac{1}{2}Tr \;GG^*\nabla^2w(t,x)
+ \< Ax,\nabla w(t,x)\>_\calh
+H_{min} (\nabla^B w(t,x)) +\ell_n(t,x),\qquad t\in [0,T],\,
x\in \calh,\\
\\
\dis w(0,x)=\phi_n(x),
\end{array}\right.
\end{equation*}
This means, with the terminology used e.g. in \cite{G1}, that $w$ is a $\calk$-strong solution
of \myref{HJBINTRO}. We do not go deeper into this since here we use the approximation
only as a tool to solve our stochastic optimal control problem.
\end{remark}
\begin{remark}\label{rm:piconv}
It can be noticed, see also \cite{FGFM}, that our results on the HJB equation
could be extended without difficulties to the case
when the boundedness assumption on $\bar \phi$ and $\bar \ell_0$, and so on $\phi$ and $\ell_0$,
is replaced by a polynomial growth assumption:
namely that, for some $N\in \N$, the functions
\begin{equation}\label{eq:polgrowthphil0}
x\mapsto \dfrac{\phi(x)}{1+\vert x\vert ^N},
\qquad
(t,x)\mapsto \dfrac{\ell_0(t,x)}{1+\vert x\vert ^N},
\end{equation}
are bounded.
Also \ref{lm:approximation} can be easily generalized to the case when the data $\phi$
and $\ell_0$ are not bounded but satisfy a polynomial growth condition in the variable $x$
as from \myref{eq:polgrowthphil0}.
Moreover it is still possible to extend the results of Lemma \ref{lm:approximation} to the case when $\phi$
and $\ell_0$ are only measurable. In this case
the approximations would take place in the sense of the $\pi$-convergence, which is weaker than the $\calk$-convergence and towards the $\calk$-convergence has also the disadvantage of being not metrizable. For more on the notion of $\pi$-convergence the reader can see
\cite[Section 6.3]{DP3} (see also \cite{EthierKurtz} and \cite{PriolaStudia}).
\end{remark}
\section{Verification Theorem and Optimal Feedbacks}\label{sec-verifica
The aim of this section is to provide a verification theorem and the existence of optimal feedback controls for our problem.
This in particular will imply that the mild solution $v$ of the HJB equation (\ref{HJBINTRO}) built in Theorem \ref{esistenzaHJB} is equal to the
value function $V$ of our optimal control problem.
The main tool needed to get the wanted results is an identity (often called ``{\em fundamental identity}'', see equation (\ref{relfond})) satisfied by the solutions of the HJB equation. When the solution is smooth enough
(e.g. it belongs to $UC_{b}^{1,2}\left( \left[0,T\right] \times H\right)$)
such identity is easily proved using the Ito's Formula. Since in our case the value
function does not possess this regularity, we proceed by approximation, following the lines of Section \ref{sec-Kstrong}.
Due to the features of our problem (lack of smoothing and of the structure condition)
the methods of proof used in the literature do not apply here.
We will discuss the main issues along the way.
\subsection{The Fundamental Identity and the Verification Theorem}
Now we finally go back to the control problem of Section \ref{section-statement}.
We rewrite here for the reader convenience the state equation (\ref{eq-contr-rit}),
\begin{equation*}
\left\{
\begin{array}
[c]{l}
dy(s) =a_0 y(s) ds+b_0 u(s) ds +\int_{-d}^0b_1(\xi)u(s+\xi)d\xi+\sigma dW_s
,\text{ \ \ \ }s\in[t,T] \\
y(t) =y_0,\\
u(\xi)=u_0(\xi), \quad \xi \in [-d,0),
\end{array}
\right.
\end{equation*}
and its abstract reformulation (\ref{eq-astr}),
\begin{equation*}
\left\{
\begin{array}
[c]{l}
dY(s) =AY(s) ds+Bu(s) ds+GdW_s
,\text{ \ \ \ }s\in[t,T] \\
Y(t)=x=(x_0,x_1).
\end{array}
\right.
\end{equation*}
Similarly the cost functional in (\ref{costoconcreto}) is
\begin{equation*}
J(t,x,u)=\E \int_t^T \left(\bar\ell_0(s,x(s))+\ell_1(u(s))\right)\;ds +\E \bar\phi(x(T))
\end{equation*}
and is rewritten as (see (\ref{costoconcreto1}))
\begin{equation*}
J(t,x;u)=\E \int_t^T \left(\ell_0(s,Y(s))+\ell_1(u(s))\right)\;ds +\E \phi(Y(T)).
\end{equation*}
We notice that throughout this subsection and the following one, in order to avoid further technical difficulties, we keep the probability space $(\Omega,\calf,\P)$
fixed. Nothing would change if we work in the weak formulation, where the probability space can change (see e.g. \cite{YongZhou99}[Chapter 2] and
\cite{FabbriGozziSwiech}[Chapter 2] for more on strong and weak formulations
in finite and infinite dimension, respectively).
We first prove the fundamental identity.
\begin{proposition}\label{prop rel fond}
Let Hypotheses \ref{ipotesibasic} and \ref{ipotesicostoconcreto} hold.
Let also (\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold.
Let $v$ be the mild solution of the HJB equation (\ref{HJBINTRO})
according to Definition \ref{defsolmildHJB}.
Then for every $t\in[ 0,T] $ and $x\in
H$, and for every admissible control $u$, we
have the fundamental identity
\begin{equation}\label{relfond}
v(t,x)
=J(t,x;u)+\E\int_t^T \left[H_{min}(\nabla^B v(s,Y(s)))
- H_{CV}(\nabla^B v(s,Y(s));u(s))\right]\,ds.
\end{equation}
\end{proposition}
\dim
Take any admissible state-control couple $(Y(\cdot),u(\cdot))$,
and let $v_n(t,x):=w_n(T-t,x)$ where $(w_n)_n$ is the approximating sequence
of strict solutions defined
in Lemma \ref{lm:approximation}. We want to apply the Ito formula to
$v_n(t, Y(t))$. Unfortunately, as mentioned in Remark \ref{rm:defpistrictstrong},
this is not possible since the process
$Y(t)$ does not live in $D(A)$. So we approximate it as follows.
Set, for $k \in \N$, sufficiently large,
\begin{equation}\label{eq:Ykdef}
Y_k(s;t,x)=k(k-A)^{-1}Y(s;t,x).
\end{equation}
The process $Y_k$ is in $D(A)$, it converges to $Y$ ($\P$-a.s. and $s\in [t,T]$ a.e.)
and it is a strong solution\footnote{Here we mean strong in the probabilistic sense
and also in the sense of \cite{DP1}, Section 5.6.}
of the Cauchy problem
\begin{equation*}
\left\{
\begin{array}
[c]{l}
dY_k(s) =AY_k(s) ds+B_ku(s) ds+G_kdW_s
,\text{ \ \ \ }s\in [t,T] \\
Y_k(t)=x_k,
\end{array}
\right.
\end{equation*}
where $B_k=k(k-A)^{-1}B$, $G_k=k(k-A)^{-1}G$ and $x_k=k(k-A)^{-1}x$.
Now observe that, thanks to \myref{resolvent} and \myref{B},
the operator $B_k$ is continuous, hence we can apply Dynkin's formula (see e.g.
\cite[Section 1.7]{FabbriGozziSwiech} or \cite[Section 4.5]{DP1})
to $v_{n}(s, Y_k(s))$ in the interval $[t,T]$, getting
\begin{align}\label{Dynkinv^nk}
&\E v_n(Y_k(T)) - v_n(t,x_k)\\ \nonumber &=
\E\int_t^T
\left[v_{nt}(s,Y_k(s))+ \frac{1}{2}Tr \;GG^*\nabla^2v(s,Y_k(s))
+ \< AY_k(s)+B_ku(s),\nabla v_n(s,Y_k(s))\>_\calh\right]ds.
\end{align}
Using the Kolmogorov equation (\ref{eq:Kforward}), whose strict solution is $w_n$,
we then write
\begin{equation}\label{quasirelfondv^nk}
\E\phi_n(Y_k(T)) - v_n(t,x_k)=\E\int_t^T \left[\calf_n (s,Y_k(s))+ \ell_{0,n}(s,Y_k(s))+\<B_k u(s),\nabla v_n(s,Y_k(s))\>_{\R^m} \right]ds
\end{equation}
We first let $k\rightarrow\infty$ in \myref{quasirelfondv^nk}.
Since $\ell_{0,n}$ and $\nabla v_n$ are
bounded functions and since $\calf_n(s,x)$ has a singularity of type $s^{-{1/2}}$ with
respect to time and is bounded with respect to $x$, we can apply
the Dominated Convergence Theorem to all terms but the last getting
\begin{align}\label{quasirelfondv^n}
&\E\phi_n(Y(T)) - v_n(t,x)\\ \nonumber&=\E\int_t^T [\calf_n (s,Y(s))+\ell_{0,n}(s,Y(s))]ds
+\lim_{k \to + \infty}\E\int_t^T\<B_k u(s),\nabla v_n(s,Y_k(s))\>_\calh]ds.
\end{align}
Concerning the last term we observe first that, by \myref{resolvent} and
\myref{B}, as well as by property of the operators $k(k-A)^{-1}$, we have that for every $u\in\R^m$ $$B_ku \to Bu.$$
This in particular implies, by the Banach-Steinhaus Theorem, that
$\{B_k u\}_k$ is uniformly bounded in $\calh$.
Now we use the fact that $\nabla v_n(s,x) \in \calh$
(see \myref{eq:nablaperSigma})
to rewrite the integrand of the last term of \myref{quasirelfondv^n} as
$$
\<B_k u(s),\nabla v_n(s,Y_k(s))-\nabla v_n(s,Y(s))\>_{\calh}
+\<B_k u(s),\nabla v_n(s,Y(s))\>_{\calh}
$$
Thanks to what said above the first term goes to $0$ as $k \to + \infty$ and is dominated while the second term is also dominated and converges to
$\<B u(s),\nabla v_n(s,Y(s))\>_{\calh}$
which, thanks to \myref{eq:nablaperSigma}, is equal to
$\< u(s),\nabla^B v_n(s,Y(s))\>_{\R^m}$ (both convergences are clearly
$\P$-a.s. and $s\in [t,T]$ a.e.). Hence
$$
\lim_{k \to + \infty}\E\int_t^T\<B_k u(s),\nabla v_n(s,Y_k(s))\>_\calh ds
=
\E\int_t^T\<u(s),\nabla^B v_n(s,Y(s))\>_{\R^m}ds
$$
Now we let $n\rightarrow\infty$. By Lemma \ref{lm:approximation}, we know that
$$v_n(t,x)\rightarrow v(t,x) \quad\text{ and }\quad(T-t)^{1/2} \nabla^Bv_n(t,x)\rightarrow (T-t)^{1/2}\nabla^B v(t,x)$$
pointwise. Moreover $v_n(t,x)$, $(T-t)^{1/2}\nabla^Bv^n(t,x)$
are uniformly bounded, so that, by dominated convergence, we get
\[
\E\int_t^T \<u(s),\nabla^Bv_n(s,Y(s))\>_{\R^m}]\,ds\rightarrow
\E\int_t^T \<u(s),\nabla^Bv(s,Y(s))\>_{\R^m}]\,ds.
\]
The convergence
\[
\E\phi_n(x(T)) -\E\int_t^T [\calf_n(s,Y(s))+ \ell_{0,n}(s,Y(s))]ds
\rightarrow
\E\phi(x(T)) -\E\int_t^T [H_{min}(\nabla^B v(s,Y(s)))+ \ell_0(s,Y(s))]ds
\]
follows directly by the construction of the approximating sequences $(\phi_n)_n$,
$(\calf_n)_n$ and $(\ell_{0,n})_n$.
Then, adding and subtracting $\E\dis\int_t^T \ell_1(u(s))ds$
and letting $n\rightarrow\infty$ in (\ref{quasirelfondv^n})
we obtain
\begin{align*}
v(t,x)&
=\E\phi(Y(T))+\E\int_t^T [\ell_0(s,Y(s))+ \ell_1(u(s))]ds\\
&+\E\int_t^T \left[H_{min}(\nabla^B v(s,Y(s)))-
H_{CV}(\nabla^B v(s,Y(s));u(s))\right]
\,ds
\end{align*}
which immediately gives the claim.
\qed
\begin{remark}
\label{rm:Yk}
One may wonder why we approximate the process $Y$ with $Y_k$ as in \myref{eq:Ykdef}
instead of using the Yosida approximants $A_k$ of $A$ as, e.g., in
\cite[p.144]{DP3}. The reason is that we need that $Y_k$ belongs to $D(A)$,
which is not guaranteed if we use Yosida approximants.
A similar procedure is used, in a different context, in
the book \cite{DP1}, in the proof of Theorem 7.7, p. 203.
\end{remark}
We can now pass to prove our Verification Theorem i.e.
a sufficient condition of optimality given in term of the
mild solution $v$ of the HJB equation (\ref{HJBINTRO}).
\begin{theorem}
\label{teorema controllo}
Let Hypotheses \ref{ipotesibasic} and \ref{ipotesicostoconcreto} hold.
Let also (\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold.
Let $v$ be the mild solution of the HJB equation (\ref{HJBINTRO})
whose existence and uniqueness is proved in (\ref{esistenzaHJB}).
Then the following holds.
\begin{itemize}
\item For all $(t,x)\in [0,T]\times \calh$ we have
$v(t,x) \le V(t,x)$, where $V$ is the value function
defined in (\ref{valuefunction}).
\item
Let $t\in [0,T]$ and $x\in \calh$ be fixed.
If, for an admissible control $u^*$, we
have, calling $Y^*$ the corresponding state,
$$
u^*(s)\in \arg\min_{u\in U}H_{CV}(\nabla^B v(s,Y^*(s);u)
$$
for a.e. $s\in [t,T]$, $\P$-a.s., then the pair $(u^*,Y^*)$ is
optimal for the control problem starting from $x$ at time $t$
and $v(t,x)=V(t,x)=J(t,x;u^*)$.
\end{itemize}
\end{theorem}
\dim The first statement follows directly by \eqref{relfond} due to the
negativity of the integrand.
Concerning the second statement, we immediately see that, when $u=u^*$
(\ref{relfond}) becomes
$v(t,x)=J(t,x;u^*)$.
Since we know that for any admissible control $u$
\[
J(t,x;u)\geq V(t,x) \ge v(t,x),
\]
the claim immediately follows.
\qed
\subsection{Optimal feedback controls and $v=V$}
\label{sec:contr-feedback}
We now prove the existence of optimal feedback controls. Under the Hypotheses
of Theorem \ref{teorema controllo} we define, for $(s,y)\in [0,T)\times \calh$,
the {\em feedback map}
\begin{equation}\label{defdiPsi}
\Psi(s,y):=\arg \min_{u\in U} H_{CV}(\nabla^Bv(s,y);u),
\end{equation}
where, as usual, $v$ is the solution of the HJB equation \myref{HJBINTRO}.
Given any $(t,x)\in [0,T)\times \calh$,
the so-called Closed Loop Equation (which here is, in general, an inclusion)
is written, formally, as
\begin{equation}\label{cleinclusion}
\left\{
\begin{array}
[c]{l}
dY(s) \in AY(s) ds+B\Psi\left(s,Y(s)\right) ds+GdW_s
,\text{ \ \ \ }s\in [t,T) \\
Y(t)=x.
\end{array}
\right.
\end{equation}
First of all we have the following straightforward corollary whose proof is immediate from Theorem \ref{teorema controllo}.
\begin{corollary}
\label{cr:optimalfeedback}
Let the assumptions of Theorem \ref{teorema controllo} hold true.
Let $v$ be the mild solution of \myref{HJBformale1}.
Fix $(t,x)\in [0,T)\times H$ and assume that, on $[t,T)\times H$, the map $\Psi$
defined in (\ref{defdiPsi}) admits a measurable selection
$\psi:[t,T)\times H \to \Lambda$ such that the Closed Loop Equation
\begin{equation}
\label{eq:CLEselection}
\left \{
\begin{array}{l}
d Y(s) = AY(s)d s+B\psi\left(s,Y(s)\right) ds+GdW_s
,\text{ \ \ \ }s\in [t,T) \\
Y(t)=x.
\end{array}
\right.
\end{equation}
has a mild solution $Y_\psi(\cdot;t,x)$ (in the sense of \cite[p.187]{DP1}).
Define, for $s \in [t,T)$, $u_\psi (s)=\psi(s,Y_\psi(s;t,x))$.
Then the couple
$(u_\psi(\cdot),Y_\psi(\cdot;t,x))$ is optimal at
$(t,x)$ and $v(t,x)=V(t,x)$.
If, finally, $\Psi(t,x)$ is always a singleton and the mild solution
of \myref{eq:CLEselection} is unique,
then the optimal control is unique.
\end{corollary}
We now give sufficient conditions to verify the assumptions of Corollary
\ref{cr:optimalfeedback}.
First of all define
\begin{equation}\label{defdigammagrandebis}
\Gamma(p):=\left\{ u\in U: \<p,u\>+\ell_1(u)= H_{min }(p)\right\}.
\end{equation}
Then, clearly, we have $\Psi(t,x)=\Gamma(\nabla^B v(t,x))$.
Observe that, under mild additional conditions on $U$ and $\ell_1$
(for example taking $U$ compact or $\ell_1$ of superlinear growth),
the set $\Gamma$ is nonempty for all $p \in \R^m$.
If this is the case then, by \cite{AubFr}, Theorems 8.2.10 and
8.2.11, $\Gamma$ admits a measurable selection, i.e. there exists
a measurable function $\gamma:\R^m \rightarrow U$ with
$\gamma(z)\in \Gamma(z)$ for every $z\in \R^m$.
Since $H_{min }$ is Lipschitz continuous, then $\Gamma$, and so
$\gamma$, must be uniformly bounded. In some cases
studied in the literature this is enough to find an optimal feedback
but not in our case (read on this the subsequent Remark \ref{rm:discfeed}-(ii)).
Hence to prove existence of a mild solution of the closed loop
equation \myref{eq:CLEselection}, as requested in Corollary
\ref{cr:optimalfeedback}, we need more regularity of the feedback
term $\psi(s,y)=\gamma(\nabla^B v(s,y))$.
Beyond the smooth assumptions on the coefficients required in the second part of Theorem \ref{esistenzaHJB}, which give the regularity of
$\nabla^B v(t,x)$, we need the following assumption about the map $\Gamma$.
\begin{hypothesis}\label{hp:lipsel}
The set-valued map $\Gamma$ defined in (\ref{defdigammagrandebis}) is
always non empty; moreover it admits a Lipschitz continuous selection $\gamma$.
\end{hypothesis}
\begin{remark}
\label{rm:discfeed}
\begin{itemize}
\item[(i)]
Notice that the above Hypothesis is verified if we assume that $\ell_1:\R^m\rightarrow \R$ is differentiable with an invertible derivative and that $(\ell_1')^{-1}$ is Lipschitz continuous.
Indeed in this case, see (\ref{psi1}),
the infimum of $H_{CV}$ is achieved at
\[
u=(\ell_1')^{-1}(z), \text{ so that }\Gamma_0(z)=(\ell_1')^{-1}(z).
\]
\item[(ii)]
The problem of the lack of regularity of the feedback law
is sometimes faced (see e.g. in \cite{FT2}) by formulating the optimal control problem in the weak sense (see e.g. \cite{FlSo} or \cite{YongZhou99}, Section 4.2)
and then using Girsanov Theorem to prove existence, in the weak sense, of a mild solution of (\ref{eq:CLEselection}) when the map $\psi$ is only measurable and bounded.
This is not possible here due to the already mentioned absence of the structure condition in the controlled state equation (i.e. the control affects the system not only through the noise).
\end{itemize}
\end{remark}
Taking the selection $\gamma$ from Hypothesis \ref{hp:lipsel}
we consider the closed loop equation
\begin{equation}\label{cle}
\left\{
\begin{array}
[c]{l}
dY(s) =AY(s) ds+B\gamma\left(\nabla^B v(s,Y(s))\right) ds+GdW_s
,\text{ \ \ \ }s\in[t,T] \\
Y(t)=x=(x_0,x_1),
\end{array}
\right.
\end{equation}
and we have the following result.
\begin{theorem}\label{teo su controllo feedback}
Assume that Hypotheses \ref{ipotesibasic},
\ref{ipotesicostoconcreto}, \ref{ipotesicostoconcretobis} and \ref{hp:lipsel} hold true.
Fix any $(t,x)\in [0,T)\times H$.
Let also (\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold.
Then the closed loop equation (\ref{cle}) admits a unique mild solution
$Y_\gamma(\cdot;t,x)$ (in the sense of \cite[p.187]{DP1})
and setting
$$
u_\gamma(s)=\gamma\left(\nabla^{B} v(s,Y_\gamma(s;t,x) )\right), \quad s \in [t,T]
$$
we obtain an optimal control at $(t,x)$. Moreover $v(t,x)=V(t,x)$.
\end{theorem}
\noindent {\bf Proof.} Thanks to Corollary \ref{cr:optimalfeedback}
it is enough to prove the existence and uniqueness of the mild solution of (\ref{cle}). We apply a fixed point theorem to the following integral form of
(\ref{cle}):
\begin{equation}\label{clemild}
Y(s)= e^{(s-t)A}x+
\int_t^s e^{(s-r)A}G\,dW_r
+\int_{t}^s e^{(s-r)A}B\gamma(\nabla^{B}v(r ,\overline{X}_r)))dr.
\end{equation}
By Hypothesis \ref{ipotesicostoconcretobis} and the second part of Theorem \ref{esistenzaHJB}
we get that the mild solution $v$ of the HJB equation (\ref{HJBformale1})
is differentiable, with bounded derivative. Moreover, since, again by the second part of
Theorem \ref{esistenzaHJB},
$v$ admits the second order derivative $\nabla^B\nabla v$, and
$\nabla^B\nabla v=\nabla\nabla^B v$,
we deduce that $t^{1/2}\nabla^B v(t,\cdot)$ is
Lipschitz continuous, uniformly with respect to $t$.
Using this Lipschitz property we can solve (\ref{clemild}) by a fixed point argument. Since the argument to do this is straightforward
we only show how to estimate the difficult term in \myref{clemild}. We have
\begin{align*}
\int_{t}^s& \vert e^{(s-r)A}B\left(\gamma(\nabla^{B}
v(r ,\overline{X}(r)))-\gamma(\nabla^{B}
v(r ,\overline{Y}(r)))\right)\vert_\calh dr
\leq \int_{t}^s C\vert\gamma(\nabla^{B}
v(r ,\overline{X}(r)))-\gamma(\nabla^{B}
v(r ,\overline{Y}(r)))\vert_{\R^m}] dr
\\
&
\leq C\int_{t}^s \vert\nabla^{B}
v(r ,\overline{X}(r)))-\nabla^{B}
v(r ,\overline{Y}(r))\vert_{\R^m}] dr
\leq C\int_{t}^s r^{-{1/2}}\vert \overline{X}(r)-\overline{Y}(r)\vert_\calh dr
\end{align*}
where $C$ is a constant that can change its values from line to line.
\qed
We devote our final result to show that the identification $v=V$ can be done,
using an approximation procedure, also
in cases when we do not know if optimal feedback controls exist.
\begin{theorem}\label{teo:v=V}
Let Hypotheses \ref{ipotesibasic}, \ref{ipotesicostoconcreto} hold.
Let also (\ref{eq:hpdebreg}) or (\ref{eq:hpdebregbis}) hold.
Moreover let Hypotheses \ref{ipotesicostoconcretobis}-(ii)
and \ref{hp:lipsel} hold and let
$\phi$ and $\ell_0$ be uniformly continuous.
Then $v=V$.
\end{theorem}
\dim
We approximate $\ell_0$ and $\phi$ by approximating $\bar\ell_0$ and $\bar\phi$
with standard approximants $\bar\ell_{0,n}$ and $\bar\phi_n$ built
by convolutions.
We set
\begin{equation}\label{costo_n}
J_n(t,x;u)=\E \int_t^T \left(\ell_{0,n}(s,Y(s))+\ell_1(u(s))\right)\;ds +\E \phi_n(Y(T))
\end{equation}
and call $w_n$ the mild solution of the HJB equation
\begin{equation}\label{HJBformale-n}
\left\{\begin{array}{l}\dis
\frac{\partial w(t,x)}{\partial t}=\call [w(t,\cdot)](x) +\ell_{0,n}(t,x)+
H_{min} (\nabla^B w(t,x)),\qquad t\in [0,T],\,
x\in \calh,\\
\\
\dis w(0,x)=\phi_n(x),
\end{array}\right.
\end{equation}
where $\call$ is given by
\begin{equation}\label{eq:ell}
\call[f](x)=\frac{1}{2} Tr \;GG^*\; \nabla^2f(x)
+ \< x,A^*\nabla f(x)\>.
\end{equation}
In particular $w_n$ satisfies the integral equation
\begin{equation}
w_n(t,x) =R_{t}[\phi_n](x)+\int_0^t R_{t-s}[
H_{min}(
\nabla^B w_n(s,\cdot)+\ell_{0,n}(s,\cdot)
](x)\; ds,\qquad t\in [0,T].\
x\in \calh,\label{solmildHJB-forwardn}
\end{equation}
By Theorem \ref{teo su controllo feedback}
calling $v_n(t,x)=w_n(T-t,x)$ we have
\begin{equation}\label{identif-v_n}
v_n(t,x) =V_n(t,x):=\inf_{u\in \calu} J_n(t,x;u).
\end{equation}
and there exists an optimal feedback control $u_n(s)=\psi_n(s,Y(s))$.
Moreover, by Lemma \ref{lm:approximation} we know that
$$
v_n(t,x) \overset{\calk}{\rightarrow} v(t,x).
$$
Now it is enough to prove that $V_n(t,x) \to V(t,x)$ pointwise.
Given $\eps >0$, we have, for $n$ large enough,
\begin{align*}
V_n(t,x)&=\inf_{u\in\calu}
\left[\E \int_t^T \left(\ell_{0,n}(s,Y(s))+\ell_1(u(s))\right)\;ds +\E \phi_n(Y(T))\right]\\
&=\inf_{u\in\calu}\left[\E \int_t^T \left(\ell_{0}(s,Y(s))+\ell_1(u(s))\right)\;ds +\E \phi(Y(T))\right.\\
&\left.+\E \int_t^T \left(\ell_{0,n}(s,Y(s))-\ell_{0}(s,Y(s))\right)\;ds
+\E\left[\phi_n(Y(T)) -\phi(Y(T))\right]\right]\\
&\leq\inf_{u\in\calu}\left[\E \int_t^T \left(\ell_{0}(s,Y(s))+\ell_1(u(s))\right)\;ds +\E \phi(Y(T))\right]
+\varepsilon,
\end{align*}
where the last passage follows by the Dominated Convergence Theorem, and since $\phi$
and $\ell_{0,n}$ are uniformly continuous.
We have shown that
\[
V_n(t,x)\leq V(t,x)+\varepsilon.
\]
Exchanging the role of $V_n$ and $V$ we also find that the reverse inequality holds true. Hence $V_n \to V$ pointwise an the claim follows.
\qed
\begin{remark}\label{rm:unif-cont-feedback}
In Theorem \ref{teo:v=V} we have assumed further uniform continuity on the data.
When $U$ is compact the result still remain true if the data are only continuous.
\end{remark}
|
\section{Introduction}
Superconducting Josephson junctions provide high-precision frequency-to-voltage conversion known as the Shapiro effect~\cite{PhysRevLett.11.80,barone}.
In fact, the modern standard of the volt is based on arrays of Josephson junctions connected to a microwave field with frequency controlled by an atomic frequency standard~\cite{grandstrand:2004,0957-0233-23-12-124002,hamilton2000josephson}.
Moreover, as an essential building block of superconducting quantum interference devices (SQUIDs) Josephson junctions provide magnetic-flux-to-voltage conversion enabling high-precision magnetic-flux measurements~\cite{squidbook}.
Recently, advanced techniques in optical trapping and control initiated studies of Josephson effects in atomic Bose-Einstein condensates (BECs).
Narrow barriers created by blue-detuned laser beams operate as Josephson junctions with highly adjustable properties.
In this way one can produce simple two-well traps~\cite{levy2007ac,PhysRevLett.95.010402}, atomic SQUIDs with one or two junctions~\cite{PhysRevLett.111.205301,PhysRevLett.113.045305,eckel2014hysteresis,1367-2630-17-12-125012}, or even arrays of Josephson junctions in optical lattices~\cite{RevModPhys.78.179,PhysRevLett.100.040404}.
Extensive theoretical and numerical studies investigate various dynamics of BECs with such barriers and determine the requirements for barriers to operate in the Josephson regime~\cite{RevModPhys.73.307,PhysRevA.59.620}. Moreover, simple mathematical models based on the two-mode approximation were develloped and supplemented with corrections for non-linear interactions and asymmetric trap configurations in order to understand atomtronic Josephson physics~\cite{PhysRevA.90.043610}.
Various experimental and theoretical results demonstrate how the quantum nature of the Josephson effect and the existence of critical tunneling currents lead to the creation of a chemical potential difference in two-well systems \cite{levy2007ac,PhysRevLett.95.010402,PhysRevLett.111.205301,PhysRevLett.113.045305}. On the microscopic level vortex-mediated phase slips were observed and used to study such phenomena as persistent currents and quantum hysteresis \cite{PhysRevLett.106.130401,eckel2014hysteresis,abad2015phase,PhysRevA.91.033607,
PhysRevA.91.023607,yakimenko2014generation}.
The existence of critical currents and phase slips in the over-critical region are cornerstones of Josephson physics in BECs.
Nevertheless, some aspects of Josephson physics in atomic BEC still remain little explored. One of them is the Shapiro effect, i.e. the use of atomtronic Josephson devices as precise frequency-to-chemical potential converters.
There are several
theoretical investigations of the Shapiro effect in a trapped BEC with a single Josephson barrier~\cite{KohleSols03,PhysRevLett.95.200401,1367-2630-13-6-065026}. While such an effect has been observed in optical lattices~\cite{PhysRevLett.100.040404},
experimental observations of Shapiro effects in single- or double-junction setups are lacking.
This may be due to the fact that coupling of Josephson oscillations to other excitations with similar frequencies masks the observation of Shapiro resonances. The existence of such coupling effects is well known from superconducting
Josephson systems~\cite{PhysRev.138.A744,Maksimov1999391,PhysRevLett.79.737} as well as from superfluid $^3$He~\cite{PhysRevLett.81.1247}.
In the present work we study resonant coupling between Josephson oscillations and phonon modes in atomic BECs. To this end we propose to investigate an experiment on the basis of a toroidal trap used to realize Josephson junctions for atomtronic SQUIDs~\cite{PhysRevLett.111.205301}.
The trap is divided into two parts by two optical Josephson barriers, and these two parts are initially populated with condensates of different number density.
If this imbalance is small one expects imbalance oscillations around zero mean (often called plasma oscillations in analogy to superconductors).
For larger imbalances one expects Josephson oscillations of the number density imbalance around a non-zero mean, which is also known as macroscopic quantum self-trapping (MQST).
However, if phonon modes can be resonantly exited by the Josephson alternating current, then this coupling will provide a dissipation channel.
This should happen if the phonon frequencies of the trap match the Josephson current frequency.
We therefore expect to observe some characteristic resonant response from the system in this region, which can be observed in the evolution of the population imbalance or the chemical potential difference between two wells.
In this paper we propose an experimentally feasible protocol to observe the dynamical picture outlined above.
The simulations of such an experiment are made using the three-dimensional time-dependent Gross-Pitaevskii equation (GPE). The results are supported by a simplified model based on the Josephson equations.
\section{Theoretical setup}
We consider the toroidal harmonic trap used in Ref.~\cite{PhysRevLett.111.205301} for the experimental realization of atomtronic Josephson junctions. It is descibed by the potential
\begin{equation}
V_{\rm trap}(\mathbf{r}) = \frac12 M \omega_z^2 z^2 + \frac12 M \omega_0^2 (r_\perp - r_0)^2,
\end{equation}
with $r_\perp=\sqrt{x^2+y^2}$, ring radius $r_0=4~\mu$m, and frequencies $\omega_z/2\pi=300$~Hz and $\omega_0/2\pi=570$~Hz.
In comparison to other toroidal traps \cite{PhysRevLett.106.130401,eckel2014hysteresis,PhysRevA.91.013602,PhysRevLett.113.045305,1367-2630-17-12-125012}, this trap is considerably elongated in the $z$ direction.
The advantage of this trap for our purposes is that the nucleation of vortices is suppressed (see also Ref.~\cite{1367-2630-13-4-043008}).
The formation of annular vortices may significantly distort the dynamics we intend to observe. Such a trap configuration requires three-dimensional simulations of the condensate dynamics.
The trap is subdivided by static barriers into two parts created by scanning blue-detuned laser beams across the trap annulus. They can be represented by repulsive potentials $V_1$ and $V_2$ , which are homogeneous in the radial direction and have a Gaussian shape in the tangential direction \cite{PhysRevA.91.033607},
\begin{equation}
V_{i}(\textbf{r}_\perp)=U_{i}\Theta(\textbf{r}_\perp\cdot \textbf{n}_{i})e^{-[\textbf{r}_\perp \times \textbf{n}_{i}]^2 / d^2}.
\end{equation}
The unit vectors $\textbf{n}_{i}$ ($i=1,2$) point radially into the directions of the barriers, $\textbf{r}_\perp=\left(x,y\right)$ is a vector in the $x y$ plane, $\Theta$ is the Heaviside Theta function, and $d=1.2~\mu$m is the $1/e$ half-width of the barrier (corresponding to the $2~\mu$m full width at half maximum in the experiment).
The experimental barrier height is reported to be $44~$nK which corresponds to $U_1/h=U_2/h=917.3~$Hz. As reported in \cite{PhysRevLett.111.205301} and supported by the results discussed below, such barriers operate in the Josephson regime.
The initial population imbalance between the two parts of the trap is produced by an additional tilt potential (not present in the experimental setup of Ref.~\cite{PhysRevLett.111.205301}).
The total external potential $V$ in which the BEC moves therefore consists of the trapping potential, the barrier potential $V_b=V_1+V_2$, and the time-dependent tilt potential,
\begin{equation}
V(\mathbf{r},t) = V_{\rm trap}(\mathbf{r}) + V_b(\mathbf{r}) + V_{\rm tilt}(\mathbf{r},t).
\end{equation}
The initial tilt is linearly switched off within $\tau=0.01~$s after starting the simulation of the BEC dynamics,
\begin{equation}
V_{\rm tilt}(\mathbf{r},t) = \left\{\begin{aligned}
U_0\,x(1-t/\tau), &\qquad t\le\tau \\
0, &\qquad t>\tau.
\end{aligned}\right.
\end{equation}
Then the evolution of the trapped condensate is observed for an additional $0.5~$s.
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{config.eps}
\caption{
Shown on the left is the density distribution of the simulated BEC cloud in $xy$-plane in the tilted toroidal trap with two barriers.
Shown on the right is the creation of the chemical potential difference in the proposed experimental protocol. The figures show the trapping potential along the $x$-axis (blue lines), tilt potential (straight red lines) and Thomas-Fermi populations in each part (blue shaded regions) at two different times.
}\label{fig:config}
\end{figure}
The proposed trap configuration is shown in Fig.~\ref{fig:config}, which illustrates how the tilt potential produces the initial population imbalance between the two parts of the ring.
When the tilt potential is switched off, this initial population imbalance leads to a chemical potential difference $\Delta\mu$.
The BEC is described by a macroscopic wave function $\Psi(\mathbf{r},t)$, which obeys the time-dependent GPE
\begin{equation}\label{GP}
i \hbar \frac{\partial \Psi}{\partial t} = -\frac{\hbar^2}{2 M} \nabla^2 \Psi + (V + g |\Psi|^2) \Psi.
\end{equation}
The wave function is normalized to the total number of atoms
$
N_T=\int|\Psi|^2d\textbf{r},
$
which is chosen to be $N_T=5000$ unless explicitly stated otherwise. The nonlinear coupling $ g = 4 \pi \hbar^2 a_s/{M}$ is given in terms of the mass $M$ of the $^{87}$Rb atom and its $s$-wave scattering length $a_s$.
The populations in each part of the trap can be obtained by integration over each nonoverlapping part $W_i$ separately,
$
N_{i}(t) = \int_{W_{i}} |\Psi(\mathbf{r},t)|^2\,d\mathbf{r},
$
with $i=L$ (left) or $i=R$ (right) and $N_T= N_L(t)+N_R(t)$. The total particle number $N_T$ is an integral of motion of the GP equation (\ref{GP}) and therefore time independent.
The relative population imbalance is then given by
$$
Z(t)=\frac{N_L(t)-N_R(t)}{N_T}.
$$
Analogously, we obtain the local chemical potentials $\mu_L(t)$ and $\mu_R(t)$ of each part of the trapped condensate as well as the chemical potential difference $\Delta\mu(t) = \mu_L(t) - \mu_R(t)$ from
\begin{equation}\label{GPmu}
\mu_i = \frac{1}{N_i} \int_{W_i} \left[ -\frac{\hbar^2}{2 M} | \nabla \Psi |^2 + V |\Psi|^2 + g |\Psi|^4 \right] d\textbf{r}.
\end{equation}
In terms of these quantities one may describe the dynamics of the BEC by a two-mode approximation to the GP equation~\cite{PhysRevA.59.620,PhysRevA.90.043610} often termed Josephson equations,
\begin{eqnarray}\label{JEFs}
\dot Z(t) &=& -\omega_J\sqrt{1-Z(t)^2} \sin[\phi(t)],\nonumber\\
\dot \phi(t) &=&\Delta \mu(t)/\hbar,
\end{eqnarray}
where $\phi$ is the phase difference between the two parts of the condensate.
This set of equations is closed by the relation $\Delta\mu(t) = \hbar\omega_C Z(t)$.
The evolution of the condensate is uniquely determined by the initial population imbalance $Z(0)=Z_0$, the capacitive energy $E_C = 2\hbar\omega_C/N_T$, and the Josephson critical current $\omega_J$, which is related to the Josephson coupling energy $E_J = N_T\hbar\omega_J/2$.
Values for these quantities can be estimated from stationary solutions~\cite{PhysRevA.90.043610}, which we calculate using imaginary-time evolution of the Eq.~(\ref{GP}) with a static tilt potential.
We find $\omega_J/2\pi = 0.68$~Hz and $\omega_C/2\pi=851$~Hz as well as the initial population imbalance $Z_0$ for each value of $U_0$.
The ratio $\omega_C/\omega_J$ is well inside the region $1 \ll \omega_C/\omega_J \ll (N_T/2)^2$, supporting that the barriers operate in the Josephson regime~\cite{RevModPhys.73.307}.
Since the ratio between the barrier hight and the total chemical potential is about $U_{1,2}/\mu \approx 0.8$, the barriers can be considered to be Josephson weak links.
Our GPE simulations clearly support the expected linear relation between $Z$ and $\Delta\mu$ which allows us to use the relative population imbalance as a measure of the chemical potential difference.
In the presentation of our results below we will use either of these two quantities interchangeably, whichever is more illustrative.
From the simulations we also confirm the linear relation between $\dot Z$ and $\sin(\phi)$, another indication that the barriers operate in the Josephson regime.
\subsection*{Excitations in a trapped condensate}
In order to identify the elementary excitations in the trapped BEC system under consideration we employ the Bogolyubov-de-Gennes (BdG) formalism (see e.g.~\cite{1367-2630-17-12-125012,PhysRevA.86.011602,PhysRevA.86.011604}).
We first consider a condensate without tilt and barrier potentials and write the cylindrically symmetric BEC wave function in the form
\begin{equation}
\Psi(\mathbf{r},t) = e^{-i\mu t} \left[ \Psi_0(r_\perp,z) + \delta \Psi(\mathbf{r}_\perp,z,t) \right],
\end{equation}
where $\Psi_0$ is the stationary solution of (\ref{GP}) with chemical potential $\mu$. The perturbation
\begin{equation}
\delta \Psi(\mathbf{r}_\perp,z,t) = u_m (r_\perp,z) e^{-i(\omega t - m \theta)} + v_m^* (r_\perp,z) e^{i(\omega t - m \theta)}
\end{equation}
is characterized by a well-defined azimuthal quantum number $m$ due to the cylindrical symmetry.
We insert this ansatz into Eq.~(\ref{GP}) and linearize it with respect to $u_m$ and $v_m$. From the resulting BdG equations one finds for each $m$ a set of eigenvalues $\omega$.
The smallest eigenvalue for each $m$ determines the lowest branch $\omega_m$ of the excitation spectrum, which is shown in Fig.~\ref{fig:pjres_bdg}.
For small quantum numbers $m$ the dispersion law is linear, which suggests that the excitations are soundlike modes (phonons).
This part of the excitation spectrum may be expressed in terms of the speed of sound $c$ and the ring radius $r_0$, $\omega_m = mc/r_0$. Alternatively, the speed of sound can be related to the average density $\tilde n$ of the condensate, $c=\sqrt{g\tilde n/M}$ \cite{PhysRevA.57.518}.
Estimating the average density as one-half of the peak density of the stationary state, we obtain values for $c$
in very good agreement with those obtained from the BdG spectrum as illustrated by the dashed red line in Fig.~\ref{fig:pjres_bdg}.
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{pjres_bdg.eps}
\caption{Lowest branch of the BdG spectrum for elementary excitations of the condensate in a symmetric ring trap.
The dashed line represents the sound frequencies estimated from the average density.
The inset shows the particle density $|\Psi|^2$ in the $xy$ plane with 25\% of the population in the $m=6$ mode (marked with a red circle).
}\label{fig:pjres_bdg}
\end{figure}
In the presence of barriers phonon excitations become standing waves localized in either of two parts of the condensate with nodes at the barrier's positions.
Both the average density and the angular extent of the two condensate parts $L$ or $R$ can be different, therefore, in principle one needs to consider two different phonon frequencies
$
\Omega_{i} = {2\pi\, c_{\rm i}}/({\alpha_{i}\, r_0}),
$
with $i=\{L,R\}$ denoting the left and right parts, respectively, $\alpha_{i}$ is the angular size of each part, and $c_i$ is the speed of sound in each part.
\section{Phonon-Josephson resonances}\label{sec:mainres}
The frequency of Josephson oscillations is related to the chemical potential difference~[see Eq.~(\ref{JEFs}) and Ref.~\cite{levy2007ac}].
For a system operating in the Josephson regime we therefore expect to observe resonant coupling between Josephson oscillations and phonon modes under the conditions
\begin{equation}\label{eq:res_cond}
\Delta\mu / \hbar = m \Omega_{L,R}, \qquad m=1,2,\ldots .
\end{equation}
To verify this expectation we perform a number of simulations based on the GPE~(\ref{GP}).
We begin with a symmetric barrier configuration, i.e., both barriers are located along the $y$ axis as shown in Fig.~\ref{fig:config}.
The angular extent of both parts is $\alpha_L = \alpha_R = \pi$.
We simulate the dynamics of the GPE (\ref{GP}) using a split-step Fourier transform method.
Each simulation starts with a different initial tilt $U_0$ providing us with different values of initial chemical potential difference $\Delta\mu(t=\tau)$. We then measure the final chemical potential difference as the average value over the last 0.2 s of the evolution in order to eliminate the effects of high-frequency Josephson or plasma oscillations.
Such a series of simulations allows us to study the final $\Delta\mu$ as a function of the initial $\Delta\mu$ and identify those regions where they are different.
The results of such a series of simulations are presented in Fig.~\ref{fig:fiske}.
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{pjres_mu_0.eps}
\caption{
Demonstration of phonon-Josephson resonances.
The solid blue line shows the final chemical potential difference as a function of initial chemical potential difference.
Dashed lines represent estimated resonance positions for phonon modes.
The straight green line is a guide to the eye marking equal initial and final values of $\Delta\mu$.
}\label{fig:fiske}
\end{figure}
There are three different regimes observed in the system's evolution depending on the initial imbalance. These regimes are marked as 1, 2, and 3 in Fig.~\ref{fig:fiske}. Corresponding curves in Fig.~\ref{fig:phonon_zt} show the actual time evolution of the population imbalance.
For small values of the initial population imbalance the system is in the plasma oscillation regime, and the time average of the population imbalance and the chemical potential difference are zero.
With larger initial population imbalance the system switches into the self-trapped regime (MQST), and according to the Josephson dynamics the average population imbalance should remain close to its initial value.
This is indeed the case everywhere except for two distinct regions, where the final population imbalance is reduced due to the resonant dissipation into phonon excitations.
As expected the resonance occurs in regions where the Josephson frequency is close to the frequency of a phonon mode.
The frequencies $\Omega_L$ and $\Omega_R$ of the phonon modes in Fig.~\ref{fig:fiske} are estimated using the density distributions of each initial stationary state.
These frequencies are equal for zero initial tilt and differ for growing population imbalance. We notice that the resonance peak is observed at slightly larger $\Delta\mu$ than expected from Eq.~(\ref{eq:res_cond}).
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{pjres_zt.eps}
\caption{
(Color online) Time evolution of the relative population imbalance for the system in the regimes of plasma oscillations (1), MQST (2), and phonon-Josephson resonance (3).
}\label{fig:phonon_zt}
\end{figure}
In the resonance region the excitation of phonon modes can be clearly seen in the density distribution of the condensate (see Fig.~\ref{fig:phonon_az_dens}).
From the time dependence of the relative population imbalance $Z(t)$ one can see that the generation of phonon modes leads to a decrease of the population imbalance.
This can be understood as a transfer of energy stored in the chemical potential difference to the phonon mode through resonant coupling.
Such phonon-assisted dissipation quickly drives the system out of resonance (within the first $\sim 0.04$~s), lowering the Josephson oscillation frequency.
This breaks the coupling and the population imbalance oscillates around its new average value.
When the system is out of resonance, phonon modes are not excited and the system again shows MQST dynamics.
The latter statement is supported by the trace of the phase difference shown in Fig.~\ref{fig:phonon_az_dens}.
Between the phase slips the phase difference grows linearly, which is typical for the MQST regime.
Its slope and consequently the frequency of phase slips changes after the population imbalance is reduced by
the resonant coupling to phonons, but otherwise the Josephson dynamics is not affected.
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{pjres_combined.eps}
\caption{
Generation of a phonon mode by Josephson oscillations.
The top panel shows the azimuthal distribution of the condensate density $n=|\Psi|^2$ as it changes with time (dark blue regions correspond to the barriers positioned at $\pi/2$ and $-\pi/2$).
The middle and bottom panels show the evolution of population imbalance $Z$ and phase difference $\phi$ within the same time period.
}\label{fig:phonon_az_dens}
\end{figure}
We now study the Josephson dynamics in the proposed trap with a significantly reduced amplitude of the Josephson oscillations. This can be achieved in various ways, e.g. by a modification of the barrier parameters or a reduction of the total number of particles.
By lowering this amplitude we eventually observe a situation when the system is not driven out of resonance. Instead the energy, which was resonantly transferred to the phonon mode, is transferred back into the chemical potential difference, i.e. Josephson oscillations.
As an example we consider a BEC in the same trap as before with only $N_T=2000$ atoms.
Then one finds $\omega_J/2\pi=0.05$~Hz and $\omega_C/2\pi=627$~Hz and the system is still in the Josephson regime.
The amplitude of the Josephson oscillations is reduced by about one order of magnitude.
In this case one does {\it not} observe any noticeable drop of final chemical potential difference in the resonance regions.
The system is not driven out of the resonance conditions and the energy transferred from the Josephson oscillations to the phonon mode can be resonantly transferred back into Josephson oscillations, reducing the energy of the phonon mode and increasing the chemical potential difference.
The system behaves like coupled oscillators showing characteristic beats as shown in Fig.~\ref{fig:pjres_beats}.
\begin{figure}[htbp]
\includegraphics[width=0.95\linewidth]{pjres_beats}
\caption{Beats of the population imbalance for the system with $N_T=2000$ particles in near-resonance conditions. Slow oscillations result from the interference of the phonon mode and Josephson oscillations (fast oscillations of $Z$) with similar frequencies.
}\label{fig:pjres_beats}
\end{figure}
\subsection*{Equivalent description using Josephson equations}
In this subsection we describe the dynamics observed in our Gross-Pitaevskii simulations by equivalent Josephson equations~(\ref{JEFs}).
The system described by the Josephson equations can be considered as an electric circuit consisting of a capacitor with capacitance $1/\hbar\omega_C$ initially charged to $Z_0$ and connected to an ideal Josephson junction. The second Josephson equation represents a second Kirchoff law for such a circuit.
Let us now consider a more complicated circuit (Fig.~\ref{fig:circuit}) that contains in addition two parallel $RLC$ circuits connected in series to the capacitor and Josephson junction. For this case the second Kirchoff law must additionally include voltage drops on these two $RLC$ circuits. This setup results in the set of equations
\begin{figure}[htb]
\includegraphics[width=\linewidth]{eqcirc}
\caption{Equivalent electric circuit for the toroidal condensate with two weak links. It consists of the Josephson junction, the initially charged capacitor with charge $Z$, which represents the total population imbalance, and two $RLC$ circuits, which represent phonon modes in two parts of the condensate.
}\label{fig:circuit}
\end{figure}
\begin{eqnarray}\label{JEFs2}
\dot Z(t) &=& -\omega_J\sqrt{1-Z(t)^2} \sin[\phi(t)]\nonumber\\
\dot \phi(t) &=&\omega_C Z + \frac{u_L}{\hbar} + \frac{u_R}{\hbar}
\end{eqnarray}
with additional equations for $u_L$ and $u_R$:
\begin{equation}\label{eq:kirchhof}
\frac{1}{\mathcal{R}_i} u_i + C_i \dot u_i + \frac{1}{\mathcal{L}_i} \int\limits_0^t u_i(\tau) d\tau = \dot Z(t), \quad i=\{L,R\},
\end{equation}
where $\dot Z(t)$ represents the total current through the $RLC$ circuit and three terms in the left-hand side are partial currents through each branch of the circuit.
The inductance and capacitance in each circuit are chosen to match the calculated phonon frequencies $1/\sqrt{\mathcal{L}_{i}C_{i}}=\Omega_{i}$.
We also consider equal inductances $\mathcal{L}_L=\mathcal{L}_R=\mathcal{L}$ and resistances $\mathcal{R}_L=\mathcal{R}_R=\mathcal{R}$ for simplicity, which means that we have only two free parameters in the model, $\mathcal{L}$ and $\mathcal{R}$. They are obtained by fitting the GPE results.
Such fitting gives $\mathcal{R}/\hbar=8000$ and $\mathcal{L}/\hbar=61$~ms, which results in a damping factor of the $RLC$ circuit $\zeta = \sqrt{\mathcal{L}/C}/2\mathcal{R} \approx 4 \cdot 10^{-3}$.
From the results in Fig.~\ref{fig:pjres_z_eqcirc} one can see that such an equivalent resonating circuit correctly captures the qualitative resonant picture, and even quantitative agreement can be achieved except in the region of high population imbalance, where other excitations become relevant for the condensate dynamics.
It is also worth noticing that a small subharmonic resonance at $\Delta\mu / \hbar = \Omega/2$, which is barely visible in the results of GPE dynamics, is much more pronounced in the equivalent circuit.
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{pjres_mu_eqcirc}
\caption{(Color online) Final chemical potential difference as a function of initial chemical potential difference obtained from the GPE calculation (black line) and Josephson equations with (green line) and without (red dashed line) additional $RLC$ circuits representing phonon modes.
}\label{fig:pjres_z_eqcirc}
\end{figure}
\section{Non-symmetric traps}
In this section we describe simulations of phonon-Josephson resonances in non-symmetric traps either
by choosing the barriers that divide the trap into parts of unequal size, or by placing
barriers with unequal properties into the trap.
\subsection{Asymmetric positions of the barriers}
In the symmetric trap discussed above the resonance frequencies in the left and right parts are very similar.
Therefore, we do not observe separate ``left'' or ``right'' resonances with phonon modes.
In order to separate left and right modes we change the angular positions of the barriers in such a way
that the angular extent of the left part is $\alpha_L = 3/4 \pi$ and that of the right part is $\alpha_R = 5/4 \pi$, or reversed. The right part always has lower density than the left one.
The results of these calculations are presented in Fig.~\ref{fig:pjres_shifted}.
In both cases the resonance regions are clearly observed, however, the positions of the minima of the chemical potential difference do not as accurately match the phonon frequencies as for symmetric barriers.
A possible reason for this discrepancy could be the considerable azimuthal inhomogeneity of the condensate parts in the case of the rotated barriers, while our estimates of the azimuthal sound wave frequencies are based on the assumption of azimuthal homogeneity of the condensate.
This inhomogeneity can distort the sound waves in the condensate and also can produce additional excitations of different nature.
It can also be seen from Fig.~\ref{fig:pjres_shifted} that right resonances are considerably more pronounced then the left ones.
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{pjres_mu_plus.eps}
\includegraphics[width=\linewidth]{pjres_mu_min.eps}
\caption{\footnotesize
Phonon-Josephson resonances in an asymmetric system.
Shown on top corresponds to the barriers rotated by $\pi/8$ to the right ($\alpha_L = 5/4 \pi$, $\alpha_R = 3/4 \pi$).
The bottom corresponds to the barriers rotated by $\pi/8$ to the left ($\alpha_L = 3/4 \pi$, $\alpha_R = 5/4 \pi$).
Notation and labeling are as in Fig.~\ref{fig:fiske}.
}\label{fig:pjres_shifted}
\end{figure}
\subsection{Asymmetric height of the barriers}
We now analyze a setup with one of the barriers operating in the Josephson regime and the other one close to the Fock regime [$\omega_C/\omega_J \sim (N/2)^2$], i.e., tunneling is completely suppressed.
To this end we make a series of simulations with increased height of one of the barriers $U_1=4U_2$.
In this case we effectively have a system with only one Josephson junction as a source of resonant oscillations.
If the interaction between Josephson oscillations and sound modes only happens at one of the junctions, then the period of the lowest resonant phonon mode will not be the time that sound needs to travel from one Josephson barrier the other (as shown in Fig.~\ref{fig:phonon_az_dens}), but the time that it takes to travel from the only Josephson barrier back and forth. The phonon is reflected from the non-Josephson barrier (see Fig.~\ref{fig:pjres_onejj_dens}).
Therefore, we expect to observe an additional series of resonances at half-integer phonon frequencies.
\begin{figure}
\includegraphics[width=\linewidth]{pjres_onejj_dens}
\caption{Azimuthal distribution of the condensate density $n=|\Psi|^2$ as a function of time. The initial conditions correspond to the first observed resonance at $\Delta\mu/h \approx \Omega_{R,L}/2$.
}\label{fig:pjres_onejj_dens}
\end{figure}
The result of this series of simulations is presented in Fig.~\ref{fig:pjres_onejj}.
We observe additional resonaces at half-integer phonon frequencies as strong as the resonances at integer phonon frequencies.
\begin{figure}
\includegraphics[width=\linewidth]{pjres_onejj}
\caption{Phonon-Josephson resonances if one barrier is outside the Josephson regime. Notation and labeling are as in Fig.~\ref{fig:fiske}.
}\label{fig:pjres_onejj}
\end{figure}
The high barrier that we introduced changes the topology of the system.
Such disconnected rings can be considered to be topologically equivalent to singly connected traps.
We therefore expect that a similar resonance picture is observed for a cigar-shaped condensate with just one Josephson junction.
\section{Conclusions}
In the present work we demonstrated that a coupling between phonon modes and Josephson oscillations in BEC can be observed in various traps with one or two Josephson barriers.
The resonant coupling manifests itself as a reduction of the average population imbalance or chemical potential difference during the condensate evolution.
Alternatively, one may think of this effect as a phonon-induced additional tunneling current within the self-trapped Josephson regime.
It provides a dissipation channel even when thermal dissipation is suppressed.
The observed population imbalance shows a ladder like structure with higher-order resonances pronounced equally to the first one.
This indicates that the observation of Shapiro resonances in such systems may be strongly affected for driving frequencies higher then the lowest characteristic phonon frequency.
Our study suggests that Josephson oscillations may couple not only to sound modes but also to other low-energy collective modes (see also Ref.~\cite{Giovanazziphd}).
Such resonant couplings should be taken into account while engineering and interpreting BEC experiments with Josephson barriers.
The observed resonant coupling effect may be used for spectroscopy of phonons as well as for other low-energy collective excitations in Bose-Einstein condensates.
While the realization proposed here only distinguishes modes with well-separated frequencies, the resolution may be improved by fine-tuning the barrier parameters.
\begin{acknowledgments}
The authors are thankful to S. Eckel and A. I. Yakimenko for useful comments and discussions.
\end{acknowledgments}
|
\section{Introduction} \label{sec:intro}
The interstellar medium (ISM) has a highly filamentary and hierarchical structure. On the upper end of this filamentary hierarchy are large-scale gaseous filaments with length up to the order of 100 pc. What is their distribution in our Galaxy and what role do they play in the context of Galactic star formation? Answers to these questions are important for a critical comparison with theoretical studies and numerical simulations of galaxy formation and filamentary cloud formation. The observational key to answer these questions is a homogeneous sample of large filaments across the Galaxy identified in a uniform way.
Studies in the past years have revealed a number of large filaments with a wide range of aspect ratios and morphologies, from linear filaments to a collection of cloud complexes.
\cite{Goodman2014} find that the 80 pc long infrared dark cloud (IRDC) ``\object{Nessie}'' \citep{Jackson2010} in the southern sky can be traced up to 430 pc in the position-position-velocity (PPV) space in $^{12}$CO (1--0), guided by connecting the IR-dark patches presumably caused by high column density regions extincting the otherwise smooth IR background emission from the Galactic plane. They argue that the Nessie runs in the center of the Scutum-Centaurus spiral arm in the PPV space, so termed as a ``bone'' of the Milky Way.
In a follow up study, \cite{Zucker2015} searched the region covered by the
MIPSGAL \citep[\emph{Spitzer}/MIPS Galactic Plane Survey, $|l|<62^\circ, \, |b|<1^\circ$;][]{MIPSGAL} focusing on the PPV loci of arms expected by various spiral arm models, finding 10 bone candidates with length 13--52 pc and aspect ratio 25--150.
\cite{Ragan2014-GFL} and \cite{Abreu2016} extend this ``mid-IR extinction'' method to a blind search, i.e., not restricting to arm loci but the full extend of the observed PPV space. They find 7 and 9 filaments with length 38--234 pc in part of the first
and fourth Galactic quadrants covered by the GRS (Galactic Ring Survey; \citealt{Jackson2006-GRS}), and the ThrUMMS (Three-mm Ultimate Mopra Milky Way Survey; \citealt{Barnes2015-ThrUMMS}), respectively. The aspect ratios of those filaments are not well defined due to the complex morphology, but inferring from the figures in the papers, the typical aspect ratio is much less than 10.
In contrast to the indirect\footnote{Indirect because ``IR-dark'' does not necessarily correspond to a dense cloud; it can be caused by a real ``hole in the sky'' \citep{Jackson2008,Wilcock2012}.}
``mid-IR extinction'' method, \cite{me15} identify large filaments directly based on emission at far-IR wavelengths near the spectral energy distribution (SED) peak of cold filaments. They develop a Fourier Transform filter to separate high-contrast filaments from the low-contrast background/foreground emission. Fitting the SED built up from the multi-wavelength \emph{Herschel}\ data from the Hi-GAL survey \citep{Hi-GAL}, they derive temperature and column density maps, and have used those maps to
select the ``largest, coldest and densest'' filaments. They present 9 filaments with length 37--99 pc and aspect ratio 19--80, identified primarily from the GRS field
These systematic searches have started to uncover the spatial distribution of large filaments in our Galaxy, revealing filaments within and outside major spiral arms. However, with different searching methods and selection criteria, in addition to inherent bias from manual inspection, it is difficult to cross compare the results from these studies. The small sample size also limits the robustness of statistical attempts \citep[e.g., see discussion in][]{me15}.
All the above mentioned searches start from a ``by-eye'' inspection of dust features (either mid-IR extinction or far-IR emission), identify candidate filaments, and then verify the coherence in radial velocity using gas tracers --- spectral line data.
We automate the identification process by applying a customized minimum spanning tree algorithm to the
PPV space. We present the first homogeneous sample of 54 large-scale velocity-coherent filaments in the range of $7.^{\circ}5 \le l \le 194^{\circ}$
(see exact coverage in \autoref{sec:data}).
We derive mass, length, linearity, aspect ratio, velocity gradient and dispersion, temperature, column/volume density, fragmentation, Galactic location and orientation angle. For the first time, we are able to investigate the Galactic distribution of their physical properties, and to estimate the fraction of the ISM confined in large filaments and star formation therein.
We describe the data set in \autoref{sec:data} and present our identification method in \autoref{sec:method}.
The identified sample of filaments and their physical properties and statistics are presented in \autoref{sec:para}, followed by a discussion of the nature and implication of the filaments in \autoref{sec:discuss}. Main conclusions are summarized in \autoref{sec:sum}.
Following the spirit of \cite{me15}, we focus on the densest filaments traced by millimeter dust continuum emission, and not the more diffused CO filaments.
\section{Data: a complete spectroscopic catalog} \label{sec:data}
The Bolocam Galactic Plane Survey (BGPS) is a blind mapping of the northern Galactic plane at 1.1 mm using the Caltech Submillimeter Observatory 10m telescope with an effective resolution of $33''$, revealing over 8400 continuum sources \citep{Aguirre2011_BGPS1,Rosolowsky2010_BGPS2}. Spectroscopic follow-ups carried out by \cite{Schlingman2011} and \cite{Shirley2013} have observed all the 6194 BGPS sources in the longitude range of $7.^{\circ}5 \le l \le 194^{\circ}$ in dense gas tracers $\rm{HCO^+}$ (3--2) and $\rm{N_2H^+}$ (3--2), using the 10m Heinrich Hertz Submillimeter telescope, with a FWHM beamwidth of $15''$. The detection rate is about 50\%, and about 99\% of the detections show a unique velocity component.
From these observations \cite{Shirley2013} compiled a complete spectroscopic catalog of 3126 sources with a single velocity component resolved in $\rm{HCO^+}$ (3--2) and/or $\rm{N_2H^+}$ (3--2). In a typical temperature range of 10--20 K of the BGPS sources \citep{Dunham2011-BGPS-NH3}, these two lines have a critical density of $\sim 10^6$ \,$\rm{cm^{-3}}$\ and
an effective excitation density of the order $10^4 - 10^5$ \,$\rm{cm^{-3}}$\ \citep{Shirley2015}, thus they trace very dense gas.
In the following, we refer to these sources as ``dense BGPS sources''.
At a typical distance of few kpc, a detection of the lines towards a BGPS 1.1 mm continuum peak marks the presence of pc scale dense gas, and a chain of such clumps connected in PPV means a rather prominent structure.
Therefore, this catalog is an excellent data set for searching for velocity-coherent filaments.
Note that the coverage of the BGPS spectroscopic catalog is contiguous in the range of $7^{\circ}.5 \le l \le 90^{\circ}.5$, $|b| \le 0^{\circ}.5$, with latitude coverage flaring up to $|b| \le 1^{\circ}.5$ in several longitude cuts. In the outer Galaxy, four selected regions were observed
($l$ range in [98.85, 100], [110, 112], [132.5, 138.5], and [187.5, 193.5] degrees).
The BGPS spectroscopic survey \citep{Shirley2013} used the version 1 BGPS continuum source catalog \citep{Aguirre2011_BGPS1}. In the release of the version 2 catalog, \cite{Ginsburg2013_BGPSv2} resolved an offset in flux scale:
$S_{\rm v2} = 1.5 S_{\rm v1}$. In this study, we use the flux from v2 (\autoref{sec:para}).
\begin{figure}
\includegraphics[width=.5\textwidth,angle=0]{demo1.pdf}
\includegraphics[width=.5\textwidth,angle=0]{demo2.pdf}
\includegraphics[width=.5\textwidth,angle=0]{demo3.pdf}
\caption{
Demonstration of the customized MST applied to a subset of the Galactic longitude.
\textit{Upper}: a MST.
\textit{Middle}: MSTs with a maximum edge length and minimum number of clumps (criteria 1--2 in \autoref{sec:method}).
\textit{Lower}: MSTs with edge limited to $\Delta v <2$\,km~s$^{-1}$\ (criteria 1--3).
Circles represent BGPS sources color coded by radial velocity as shown in the color bar.
}
\label{fig:demo}
\end{figure}
\section{Method: connecting dots using customized MST} \label{sec:method}
The minimum spanning tree (MST) was first introduced by
Otakar Bor{\textrm{u\kern -1.3ex\raisebox{0.6ex}{$^\circ$}}}vka
to optimize the cost of electrical grids by minimizing its total length. It is now widely used in general optimization of a variety of networks (see review by \citealt{MST2001}).
We adopt the MST algorithm to isolate coherent filaments out of a PPV catalog bearing in mind that ``coherence'' means ``close proximity'' in position \textit{and} velocity.
Our method is demonstrated in \autoref{fig:demo}.
A MST connects all the nodes (BGPS clumps) in a graph with the cost of minimum sum of edge lengths, where ``edge'' refers to the separation between two nodes (\autoref{fig:demo}a).
We customize the MST such that a graph is connected not in one MST, but in a collection of MSTs with the following criteria:
\begin{enumerate}
\item[(1)]
The accepted MST must contain at least 5 BGPS clumps: $N_{\rm cl} \ge 5$.
\item[(2)]
Only edges shorter than a maximum length can be connected ($\Delta L <0.1^{\circ}$, \autoref{fig:demo}b).
\item[(3)]
For any two clumps to be connected, the difference in line-of-sight velocity ($\Delta v$) must be less than 2 \,km~s$^{-1}$\, (\autoref{fig:demo}c).
\end{enumerate}
The maximum edge length and velocity difference (criteria 2--3) are chosen based on characteristics of previously known filaments \citep{me15}.
As in \cite{me15}, we have used the ``\object{Snake}'' nebula, one of the first identified IRDCs \citep{carey98}, as a primary guide to test criteria (2--3). In addition to its prominent IR extinction feature, the Snake is known to have velocity-coherent structures at multiple spatial scales \citep{carey2000,me14,my-Springer-sum}. Together, these properties make the Snake a good test case.
During the tests, we relax criteria (2--3) by increasing $\Delta L$ and $\Delta v$ until the MST starts to (incorrectly) connect unrelated BGPS sources into the Snake. We also require the algorithm to be able to identify as many as possible previously known filaments from other studies (see \autoref{sec:knownFL}), but not to connect unrelated sources.
Criterion (1) is arbitrary, but we note that there is no difference if we set the minimum number of clumps $N_{\rm cl}$ as 5 or 6. It is clear in \autoref{tab:fl} which filaments would be picked up should we increase the required $N_{\rm cl}$.
So far, these MSTs are \textit{coherent} structures in the $(l,b,v)$ space, but not necessarily \textit{filamentary} structures. To further select filaments we introduce a linearity check.
For a given filament, we fit a straight line to the $(l,b)$ data points.
The fitted line represents the filament's major axis.
Linearity is defined as the ratio between
the spread (standard deviation) of data points measured along the filament's major axis to the spread along minor axis: $f_L = {\sigma_{\rm major}}/{\sigma_{\rm minor}}$.
After visual inspection we accept structures with
\begin{enumerate}
\item[(4)] Linearity $f_L >1.5$
\end{enumerate}
as filaments.
Note that the linearity is defined to quantify
how far the structure is away from a linear structure. For a straight line, $f_L \rightarrow \infty$.
For bent or wiggling filaments, which is often the case, the linearity is much smaller than the aspect ratio. For instance, the famous IRDC ``Snake'' (F7 in \autoref{tab:fl}) has an aspect ratio of 43, while its linearity is only 4 because of its sine wiggling in the $(l,b)$ space \citep{me14,me15,my-Springer-g11}.
Finally, after determining the distance (\autoref{sec:dist}) we accept long filaments with
\begin{enumerate}
\item[(5)] Projected length $\ge$10 pc
\end{enumerate}
as large-scale filaments for the interest of this study.
Applying the methodology to the BGPS spectroscopic catalog, we have identified 91 velocity-coherent structures (satisfying criteria 1-3), 70 of which are linear filaments (satisfying criteria 1-4). Among these, 54 are large-scale filaments (satisfying criteria 1-5), including 48 in the first Galactic quadrant where the BGPS coverage is contiguous, and 5/1 in the second/third quadrants respectively, where the BGPS coverage is targeted to known star formation regions.
Of the 54 filaments, only 6 are previously known
(F7, F13, F25, F33, F36, and F41; \autoref{sec:knownFL}).
Clearly, our filaments identification method depends on free parameters (like many other methods) \footnote{Other filament-finding methods include: getfilaments \citep{soft:getfilaments-Menshchikov2013}, DisPerSE \citep{soft:DisPerSE-Sousbie2011,soft-FilTER-Panopoulou2014,LiGX2016-FL}, FilFinder \citep{soft:FilFinder-Koch2015}, Hessian matrix \citep{Schisano2014,Salji2015-FL-Orion}, and Bisous model \citep{soft-Tempel2016-Bisous}.
Our customized MST and the Bisous model work on discrete points, and others work on continuous images.}, hence the identified filaments and their properties depend on the chosen parameters of criteria (1--5). Here we have chosen reasonable criteria to select the \textit{representative} large-scale velocity-coherent filaments based on previously known filaments.
Each of the 54 filaments are plotted in \autoref{fig:rgb} in a two-color view, where the mid-IR 24 or 22\,$\mu\mathrm{m}$\ emission \citep{MIPSGAL,WISE} is shown in cyan and the (sub)millimeter 0.87 or 1.1 mm emission \citep{Csengeri2016-ATLASGAL,Ginsburg2013_BGPSv2} is shown in red. The MST edges are drawn to outline the filaments.
Most (40 out of the 54) filaments are IRDCs, while 14 are IR-bright filaments.
The filaments show a wide range of filamentary morphologies (see next section).
\section{Physical parameters} \label{sec:para}
\autoref{tab:fl} lists physical parameters of the 54 identified filaments.
Column (1) assigns identification numbers running from F1 to F54.
Col. (2--4) flux weighted longitude, latitude (in degree), and local standard of rest (LSR) velocity (\,km~s$^{-1}$). For instance, for a filament containing $n$ clumps, its flux weighted longitude is
$l_{\rm wt} = \frac{\sum_{i=1}^{n} F_i\times l_i}{\sum_{i=1}^{n} F_i}$, where $F_i$ and $l_i$ are the BGPS flux and the longitude of the $i$th clump, respectively.
Col. (5--6) distance (kpc) and its type (see \autoref{sec:dist}).
Col. (7) number of clumps in the filament.
Col. (8--9) length of the filament in degree and pc by summing all the edges in the filament.
Col. (10) velocity gradient (\,km~s$^{-1}$\,pc$^{-1}$): the mean of all the edges.
We caution that both velocity and length are subject to projection effect.
Col. (11) dispersion of the central velocity of all the clumps in the filament (\,km~s$^{-1}$).
Col. (12--13) minimum and maximum temperature of the clumps (see \autoref{sec:T}).
Col. (14-15) filament mass (in unit $10^3$\,$M_\odot$) and linear mass density (\,$M_\odot$/pc).
Mass is computed from the integrated BGPS v2 1.1\,mm dust emission flux measured in a polygon encompassing the filament guided by the MST, adopting the \cite{Ossenkopf1994} dust opacity law and $\beta = 1.5$, and accounting for the different temperatures in the clumps (\autoref{sec:T}).
Col. (16--17) molecular hydrogen column density ($10^{22}$ \,$\rm{cm^{-2}}$) and volume density ($10^3$ \,$\rm{cm^{-3}}$) of the filament. These are estimated by simplifying the filament as a cylinder with a length of the filament length and a diameter of the mean major axes of the clumps.
Col. (18) aspect ratio $f_A$ estimated by dividing filament length with the averaged major axes of the clumps.
Col. (19) linearity $f_L$ (see definition in \autoref{sec:method}).
Col. (20) $R_{\rm gc}$, Galactocentric radius (kpc).
Col. (21) $z$, vertical distance (pc) to the physical Galactic mid-plane
after correction for the Sun's displacement of 25 pc above the mid-plane and the true position of the Galactic Center \citep{Goodman2014,me15}.
Col. (22) $\theta$, orientation angle (degree) between the filament's long axis and the physical Galactic mid-plane. Positive/negative angle means Galactic latitude increases/decreases with increasing longitude (\autoref{fig:demo}c).
Col. (23) Morphology class as defined in \cite{me15}:
L: linear straight or L-shape;
C: bent C-shape;
S: quasi-sinusoidal shape;
X: crossing of multiple filaments;
H: head-tail or hub-filament system.
Some filaments are characterized by more than one class.
Different morphologies may have resulted from different filament formation processes. For example, expansion of bubbles can produce C shaped filaments, collision of bubbles can make S shaped filaments, gravitational contraction of a clump embedded in a sheet can make H type filaments, while turbulence and converging flows can make filaments of any shape.
Col. (24--27) Galactic coordinate boundary of the filament (degree).
The last rows of \autoref{tab:fl} list statistics of these physical parameters.
These include minimum, maximum, median, mean, standard deviation, skewness ($K$), and kurtosis ($K$).
Skewness is a measure of symmetry. A symmetric distribution has $S = 0$, while negative/positive value of $S$ means asymmetric tails with lower/higher values around the mean, respectively.
Kurtosis is a measure of how the distribution is compared to a normal distribution (e.g., Gaussian): a Gaussian distribution has $K=0$, while a negative or positive value $K$ means the distribution is a flatter or more centrally peaked distribution than Gaussian, respectively.
These statistics provide a description of each parameter, which is discussed in \autoref{sec:stat}.
The major sources of uncertainty of the parameters originate from the uncertainties in distance, dust opacity, and projection. Because projection is unknown, we do not correct for it in \autoref{tab:fl}, but we discuss its effect on different parameters here.
The dust opacity tabulated in \cite{Ossenkopf1994} is subject to a factor of 2 uncertainty.
The typical distance uncertainty is 10-30\%, depending on different distance types (see \autoref{sec:dist}).
Distance uncertainty propagates to other parameters. In the following estimation we use a conservative 30\% distance uncertainty.
\textit{Length}:
includes distance uncertainty of 30\% and projection. Due to projection, the length is a lower limit. For a random inclination angle ($\phi$, defined as the angle between line-of-sight and filament's long axis) of 1 radian, $L_{\rm obs} = L \times {\rm sin} (\phi) = 0.84 L$.
\textit{Velocity gradient}:
includes distance uncertainty of 30\% and projection.
Projection affects both velocity and length, in a form of tan($\phi$), thus projection can lead to an over- or underestimation of the velocity gradient.
For $\phi = 30^{\circ}, 57.3^{\circ}, 75^{\circ}$,
the observed velocity gradient is the true value times cos($\phi$)/sin($\phi$) = 1.73, 0.64, 0.27, respectively.
\textit{Mass}:
includes uncertainties in dust opacity, distance (scaling as $d^2$), and flux (10\%, \citealt{Ginsburg2013_BGPSv2}).
Together, these uncertainties propagate to $<$50\% uncertainty in mass.
Note we have accounted the temperature variance across the filament \citep[similar to][]{me15}, an improvement over other studies in which a uniform temperature is \textit{assumed} for all filaments \citep[e.g.][]{Ragan2014-GFL,Abreu2016}.
\textit{Mass per unit length}:
Compared to mass, $M/L$ has a smaller dependence on distance uncertainty (scaling as $d^1$), thus the error is less than that of mass.
But $M/L$ is affected by projection in the form of $1/L$.
\textit{Column density}:
defined as mass/(length$\times$width), the distance error is eliminated, thus the uncertainty in column density is less than that of mass.
Projection affects length but not width, so projection is in the form of $1/L$.
Additionally, simplifying filaments into cylinders can cause uncertainties for some filaments which deviate from a linear structure (with a relatively small linearity $f_L$).
\textit{Volume density}:
defined as mass/(length$\times$width$\times$depth), where depth is assumed to be the same as width.
Compared to mass, volume density has a smaller dependence on distance uncertainty (scaling as $1/d$), thus the uncertainty is less than that of mass.
But volume density is affected by projection in the form of $1/L$.
Similar to column density, simplifying filaments into cylinders can cause additional uncertainties for some filaments which deviate from a linear structure.
\subsection{Distance estimation} \label{sec:dist}
Distance is determined by three methods, listed in decreasing order of robustness:
(1) type P: trigonometric parallax measurements of associated masers from the BeSSeL project \citep[][Sanna, private communication]{BeSSeL,Brunthaler2009,Moscadelli2009,XuY2011,Immer2013,Wu2014-BeSSeL}.
(2) type ML: maximum likelihood distance from Bayesian evaluation of kinematic distance using external data to place priors \citep{Ellsworth2015,Ellsworth2013-dist}.
(3) type KN, KF:
near or far kinematic distance computed using the procedure of \cite{Reid2009} with updated Galactic parameters from \cite{Reid2014}.
Wherever available, we use parallax distance, then Bayesian distance, then kinematic distance.
Of the 54 filaments, 8 are assigned for type P, 39 for ML, 4 for KN, and 3 for KF distances.
When more than one clump in a given filament has ML distance, the median of the clump distances are used, computed after excluding extreme values. This occurs, for example, when a filament is generally IR-dark (therefore more likely to be located at near distance), but a minority of its clumps are IR-bright and so the far distance is assigned by the ML evaluation.
Kinematic distance is used in 8 cases. The distance ambiguity is resolved with the IR emission/extinction (see \autoref{fig:rgb}) \textit{and} the fact that all clumps in a given filament should have a consistent distance computed by the \cite{Reid2009} code.
\subsection{Temperature estimation} \label{sec:T}
We evaluate the temperature of every clump from three resources, listed in order of decreasing priority to search for a match:
(1) gas kinetic temperature determined from $\rm{NH_3}$ for a subset of the BGPS sample \citep{Dunham2011-BGPS-NH3},
(2) same but for a subset of the ATLASGAL (Atacama Pathfinder EXperiment Telescope Large Area Survey of the Galaxy) sample \citep{Wienen2012-NH3}, and
(3) dust color temperature determined by comparing 350\,$\mu\mathrm{m}$\ and 1.1\,mm fluxes for a subset of BGPS sample \citep{Merello2015-BGPS-350um}. We exclude color temperatures with large uncertainties ($>$100\%).
When querying for $\rm{NH_3}$ temperature the clump must match in position and velocity, and for color temperature only position is available for matching.
If a clump does not have a match, we assume a temperature of 15 K based on the average value of BGPS clumps \citep{Dunham2011-BGPS-NH3}.
Among the 54 filaments, 46 have at least one temperature match.
The minimum and maximum temperature values are listed in \autoref{tab:fl}.
\begin{figure*}
\centering
\includegraphics[width=.245\textwidth]{hist_M.pdf}
\includegraphics[width=.245\textwidth]{hist_len.pdf}
\includegraphics[width=.245\textwidth]{hist_dkpc.pdf}
\includegraphics[width=.245\textwidth]{hist_theta.pdf}
\\
\includegraphics[width=.245\textwidth]{hist_l.pdf}
\includegraphics[width=.245\textwidth]{hist_b.pdf}
\includegraphics[width=.245\textwidth]{hist_z.pdf}
\includegraphics[width=.245\textwidth]{hist_Rgc.pdf}
\caption{
Panels (a--d) show histograms of filament mass, length, distance, and orientation angle. The 54 filaments and 13 bones are labeled.
Panels (e--h) plot normalized distribution of $l, b, z$ and $R_{\rm gc}$, comparing filaments with BGPS sources.
``BGPS sources'' are the BGPS clumps covered in the searching field.
``BGPS dense gas'' means the 3126 BGPS clumps with dense gas detection in \cite{Schlingman2011} and \cite{Shirley2013}.
The 1710 ``BGPS w/ distance'' source and 3508 ``BGPS w/ velocity'' sources are from \cite{Ellsworth2015}.
}
\label{fig:hist}
\end{figure*}
\subsection{Statistics of the parameters} \label{sec:stat}
\autoref{fig:hist} presents histograms of selected parameters. Panels (a-d) show mass, length, distance, and orientation angle for the filaments. Panels (e-h) present normalized histograms of $l, \, b, \, z, \, R_{\rm gc}$ where the distribution of filaments is compared to an appropriate BGPS sub-sample when data is available.
Among the 54 identified filaments, F5 stands out as an extreme: 276 pc and $6.4\times10^5$ \,$M_\odot$. Excluding this extreme, the other 53 filaments extend 10--93 pc over the projected sky, with a mass in the order of $10^3 - 10^5$ \,$M_\odot$. For all the 54 filaments, the angular length lies in the range of $0.17^{\circ}-1.50^{\circ}$ with a median of $0.35^{\circ}$.
The column density $N_{\rm H_2}$ lies in the range of $(0.2-16.9) \times 10^{22}$ \,$\rm{cm^{-2}}$\ with a median value of $0.9 \times 10^{22}$ \,$\rm{cm^{-2}}$, and volume density $n_{\rm H_2}$ is in the range $(0.4-22) \times 10^3$ \,$\rm{cm^{-3}}$\ with a median value of $1.8 \times 10^3$ \,$\rm{cm^{-3}}$.
The filaments are 1--2 orders of magnitude denser than the filaments identified by the ``mid-IR extinction'' \citep{Ragan2014-GFL,Zucker2015,Abreu2016}, while as dense as the filaments identified from \emph{Herschel}\ far-IR emission by \cite{me15}.
The mass-length relation is best fitted as ${\rm lg}L = 0.41\times {\rm lg}M - 0.19$. This suggests that the filaments are not $N$ small unrelated filaments that appear to be accidentally connected. Suppose the sub-units have a typical length $l$ and a typical mass $m$. The total length of the filament would then be $L=N\times l$ with mass $M=N\times m$, resulting $L \sim M$ and not $L \sim M^{0.41}$ as observed here.
The mass-length relation implies $M \sim L^{2.4}$, i.e., a fractal dimension of 2.4. This is comparable with the observed GRS molecular cloud catalog (although of whole clouds) in the Milky Way \citep{Roman2010-GRS} and with numerical simulations of supersonic gas turbulence \citep{Federrath2009}. The mass-length scaling of filaments, from large to small-scales, will be discussed in detail in a forthcoming paper (Burkert et al. in prep).
The orientation of the filaments is clearly not random. The distribution of the orientation angles $\theta$ is close to Gaussian ($K = -0.3$), showing a weak but significant concentration towards $0 ^{\circ}$.
The vertical distribution $z$ of the filaments is surprisingly not symmetric with respect to the Galactic plane ($S=-0.7$). It skews towards the negative values, with a mean at a positive value of $z=11.5$ pc. This is likely an observational bias (see \autoref{sec:bias}). Nevertheless, $z$ concentrates towards small values, with 50\% (27/54) filaments located within $|z| \le 20$ pc and 70\% (38/54) within $|z| \le 30$ pc. Towards higher vertical positions the number of filaments decreases much faster than a Gaussian function ($K=2.8$).
Three parameters have a Gaussian-like distribution: the Galactocentric radius $R_{\rm gc}$ ($K=0.0$), the clump velocity dispersion $\sigma (v)$ ($K=-0.01$), and the flux weighted LSR velocity $v_{\rm wt}$ ($K=-0.3$). Notably, the mean velocity gradient along the filament is small ($0.43\pm0.31$ \,km~s$^{-1}$\,pc$^{-1}$), but in broad agreement with simulations \citep{Duarte2016}.
Across the Galaxy, in general, the distribution of filaments follows the number density of BGPS sources, i.e., it is more likely to find a filament where the BGPS sources are crowed. Specifically, the probability density\footnote{The probability density is defined as
$P = N_{\rm bin}/(N_{\rm total} \times \Delta_{\rm bin})$, so that the integral of the histogram will sum to 1.}
of filaments ($P_{\rm Filament}$) largely agrees with the probability density of BGPS sources: $P_{\rm Filament} \sim P_{\rm BGPS}$, as binned in $l, \, b, \, z, \, R_{\rm gc}$ [\autoref{fig:hist}(a-d)].
This is expected because we used the BGPS spectroscopic catalog as input, which is a homogeneous subsample of the BGPS catalog. However, there are interesting exceptions.
First, a significantly lower filament probability ($P_{\rm Filament} ${\raisebox{-.9ex}{$\stackrel{\textstyle<}{\sim}$}}$ 0.5\times P_{\rm BGPS} $) is seen towards $l \sim 30^{\circ}$, the Scutum tangent. The same discrepancy is seen towards $b \sim 0.2^{\circ}$, $z \sim 20$ pc, and $R_{\rm gc} \sim 5$ kpc.
Second, on the other hand, a much higher $P_{\rm Filament}$ is found in the inner Galaxy of $b \sim 10^{\circ}$, towards Galactocentric radius of $R_{\rm gc} \sim 4$ and 8 kpc, and most significantly, towards zero Galactic vertical scale height, $z$.
Third, the averaged $b$ and $z$ for filaments are closer to zero than that of BGPS sources. In another word, the filaments show a more symmetric distribution (than BGPS sources) with respect to the physical Galactic mid-plane.
Are filaments closer to the mid-plane more likely to align with the mid-plane, i.e., $\theta$ approaches $0^{\circ}$? The data does not indicate so (\autoref{fig:theta}):
the Pearson's product-moment correlation coefficient
between $|z|$ and $|\theta|$ is $C_{\rm Pearson} = +$0.26 or $-$0.35 for the filaments and bones respectively, far from a linear correlation ($+$1 or $-$1).
The statistical trends observed in these filaments provide excellent targets for quantitative tests with future theoretical calculations and numerical simulations.
\begin{figure}
\centering
\includegraphics[width=.49\textwidth,angle=0]{theta_z.pdf}
\caption{
Filament orientation angle $|\theta|$ plotted versus the vertical height to the physical Galactic mid-plane $|z|$. The 54 filaments and 13 bones are labeled. For reference, the height of the Sun is $z_\odot = +25$ pc.
}
\label{fig:theta}
\end{figure}
\section{Discussion} \label{sec:discuss}
\subsection{Comparison to previously known filaments} \label{sec:knownFL}
Our MST algorithm finds some filaments previously identified by other methods.
Our search field (\autoref{sec:data}) partially overlaps with previous searches by \cite{Ragan2014-GFL}, \cite{me15}, and \cite{Zucker2015}.
Of the 9 cold and dense prominent filaments presented by \cite{me15}, 7 are in our search field, 3 of which are identified by MST (F7, F25, and F41 correspond to G11, G24, and G49, respectively).
Others are not identified because of a lack of dense BGPS sources (G29, G47, and G64) or too large disruption in velocity space due to active star formation (G28).
Among the 10 ``bone'' candidates presented by \cite{Zucker2015}, 6 are in our search field, 2 of which have dense BGPS sources: BC011.13--0.12 and BC024.95--0.17.
The former corresponds to F7 (the Snake), and the latter is not identified by MST because of too large velocity disruption.
In addition, our MST filament F28 is visible in their Figure 13 and seems to fulfill all their criteria, but was not identified by \cite{Zucker2015}.
Among the 7 giant molecular filaments presented by \cite{Ragan2014-GFL},
6 are partially covered in our field (``partially'' because most of those filaments extend beyond our coverage in $|b|$).
F36 is a small dense part of GMF38.1-32.4a, but note that \cite{Ragan2014-GFL} used a kinematic distance of 3.3--3.7 kpc, a factor of 2 larger than the ML distance of 1.7 kpc.
F19 and F38 fall in the positional coverage of GMF20.0-17.9 and GMF41.0-41.3, respectively,
but outside the velocity ranges.
In addition, F33 is the dense part of the ``massive molecular filament'' G32.02+0.06 presented by \cite{Battersby2012-FL} in a case study.
F13 is part of the IRDC G14.225-0.506 \citep{Busquet2013}.
F31 runs across a well studied IRDC \object{G28.34+0.06}, also known as the ``Dragon'' nebula \citep{me11,me12,my-Springer-sum}. The IRDC is the IR-dark and submm-bright arc, bent towards the bottom of the panel in \autoref{fig:rgb}. The MST filament F31 is a new filament that runs across the IRDC at P1, where a proto-cluster is forming \citep{me11,me12,qz15}. Interestingly, the clump scale magnetic fields \citep{me12} are aligned with F31. At scales of the order of 10 pc, magnetic fields may be shaped by gravity, while on smaller scales (within 1 pc, or clump-scale), the magnetic fields control the formation of a secondary filament, as interpreted in \cite{me12}. The secondary filament is a small part of F31. Dust polarization observations of these filaments are needed to further investigate the role of magnetic fields on the formation and evolution of these filaments.
In summary, our MST method successfully finds previously known filaments where the criteria are satisfied. An important difference between the MST identified filaments and others is that the former contains dense clumps over the \textit{whole} extent, while this is not the case for previously ``by-eye'' identified large filaments (``gaps'' in velocity space are allowed).
The MST method also finds filaments embedded in a crowded PPV space, which are difficult to isolate with human eyes (e.g., F31).
It is noteworthy that, because of the filamenatry and hierarchical nature of the ISM, one can find an arbitrary number of filaments in the same data set using different criteria. Therefore, when presenting a filament sample, it is \textit{equally important} to explicitly list the criteria used to define filaments. For the same reason, when comparing different samples of filaments one has to notice the difference in criteria, otherwise the comparison is misleading.
\subsection{Completeness and bias} \label{sec:bias}
The 54 filaments form the first comprehensive sample of large-scale velocity-coherent gas structures in the northern Galactic plane covered by the BGPS spectroscopic survey. The homogeneous sample allows us to investigate statistical trends (\autoref{sec:stat}) for the first time.
With length in the range of 10--276 pc and average column density above $10^{21}$ \,$\rm{cm^{-2}}$, these filaments are among the densest and largest structures observed in the Galaxy, and provide excellent tracers for Galactic structure and kinematics \citep[e.g.][]{Englmaier1999,Dame2001,Dobbs2012,Reid2014,Vallee2016,Smith2016}.
The completeness of our filament sample largely depends on the data we use. The BGPS continuum catalog is 98\% complete at the 0.4 Jy level \citep{Rosolowsky2010_BGPS2}. The spectroscopic catalog \citep{Shirley2013} contains 50\% of the sources in the survey coverage with dense gas lines detected. As our identification used the spectroscopic catalog, the results are biased to dense clumps. This is evident in the high averaged column density and high linear mass density.
However, we emphasize that this is indeed our goal --- we are interested in the most prominent dense filaments.
On the other hand, given the location of our Sun in the Galaxy, even homogeneous surveys like BGPS or ATLASGAL are biased to structures (a) closer to the Sun and (b) on the same side with respect to the mid-plane as the Sun (\autoref{fig:hist}; \citealt{ATLASGAL}, but see discussion in \citealt{Rosolowsky2010_BGPS2}).
As mentioned in \autoref{sec:stat}, the large filaments show less bias than BGPS sources in $z$, but the distribution of $z$ is indeed not perfectly symmetric. Should we mirror the distribution of $z>0$ to $z<0$, the total number of filaments would increase to 74. That is, 27\% of the filaments may be missed due to this effect.
Although with a much improved completeness compared to previous methods,
the MST approach cannot find all large filaments in our Galaxy. On of the main strength of this method is the repeat-ability compared to manual approaches.
\begin{figure}
\centering
\includegraphics[width=.45\textwidth,angle=0]{arm.pdf}
\includegraphics[width=.45\textwidth,angle=0]{arm2.pdf}\\
\includegraphics[width=.43\textwidth,angle=0]{faceon.pdf}
\caption{
The Galactic distribution of the filaments.
\textit{Upper:}
The longitude-velocity plot showing the spiral arm segments derived from maser parallaxes (\citealt{Reid2014}; Reid et al. 2015, private communication). For simplicity, only related arms (Scutum, Sagittarius, Norma, Local, Perseus, and Outer) are plotted. The color shaded segments are of $\pm 5$ \,km~s$^{-1}$\ width with respect to arm centers. The 54 filaments and 13 bones are labeled. The grey shaded horizontal strips along the x axis depict the searched longitude ranges.
\textit{Middle:} a zoom-in of the upper panel for clarity.
\textit{Lower:}
A ``face-on'' view from the northern Galactic pole. The arm widths (170--630 pc) are from \cite{Reid2014} except for Norma, whose width is not available, and we plot 200 pc width for reference. The solar symbol $\odot$ is plotted at (0, 8.34) kpc.
}
\label{fig:arm}
\end{figure}
\subsection{Galactic distribution and number of large filaments and ``bones'' in the Galaxy}
\label{sec:Gal-dist-bones}
Most of the filaments are associated with major spiral arms (\autoref{fig:arm}), consistent with the observations by \cite{me15}.
Many of them concentrate along the longitude-velocity tracks of the Scutum, Sagittarius and Norma arms, and a few associated with the Local arm, the Perseus arm, and one associated with the Outer arm\footnote{In the $l-v$ view (\autoref{fig:arm}), most filaments follow the spiral arms, while the association is less evident in the face-on view. This originates from the difference in the the distance determination methods for the arm segments (parallax measurements) and the filaments (mainly kinematic distance). The same is seen in e.g. \cite{Abreu2016}.}.
Only a small fraction (11/54, or 20\%)
of the filaments are not within $\pm 5$ \,km~s$^{-1}$\ of any arm structure, and are analogs of ``spurs'' observed in other galaxies.
How many large filaments exist in our Galaxy? Using our method, we have identified 48 filaments in the contiguous coverage of $7.5^{\circ} \le l \le 90.5^{\circ}$. It is reasonable to estimate a similar number of filaments in the fourth quadrant. In the outer Galaxy, the BGPS survey is targeted to several star formation regions, therefore the 6 identified filaments provide an extremely lower limit.
Taking all these into account, and correcting for the bias as discussed in \autoref{sec:bias}, we estimate about 200 velocity-coherent filaments longer than 10 pc and with a global column density above $10^{21}$ \,$\rm{cm^{-2}}$, as the filaments presented in this study.
How many filaments lie in the center of spiral arms and thus sketch out the ``bones'' of the Milky Way? \cite{Goodman2014} argued that the long and skinny IRDC ``Nessie'' lies in the center of the Scutum-Centaurus spiral arm in the $(l,v)$ space, and within $z = \pm 20$ pc from the physical mid-plane.
Following \cite{Goodman2014} and \cite{Zucker2015}, our criteria for a ``bone'', on top of our large-filaments criteria (1--5) (\autoref{sec:method}), are:
\begin{enumerate}
\item[(6)]
Lies in the very center of the physical Galactic mid-plane, with
$|z| \le 20$ pc.
\item[(7)]
Runs almost parallel to arms in the projected sky, with
$|\theta| \le 30^{\circ}$.
\item[(8)]
The flux weighted LSR velocity $v_{\rm wt}$ is within $\pm 5$ \,km~s$^{-1}$\ from spiral arms.
\end{enumerate}
However, the exact structure and position of the spiral arms in our Galaxy are not well established. Diverse models have been derived from a variety of data ranging from atomic, molecular, ionized gas to stars and pulsars \citep[e.g.][]{Reid2014-AR,Hou2014-arm,Vallee2015,Vallee2016}.
{Here we}
have adopted the spiral segments derived from maser parallaxes (\citealt{Reid2014}; Reid et al. 2015, private communication), which have well constrained distances.
In \autoref{fig:arm} we superpose the filaments on the spiral segments.
Among the 54 filaments, 27 fulfill criteria (1-6), 21 fulfill criteria (1-7), and 13 of them also fulfill criterion (8). These 13 filaments (F2, F3, F7, F10, F13, F14, F15, F18, F28, F29, F37, F38, and F48) are ``bones'' according to our definition.
Our criteria for a bone are more strict than \cite{Zucker2015} in terms of velocity coherence and mean column density.
When compared to other filaments, bones do not stand out in mass, length (\autoref{fig:hist}(a--b)), column/volumn density, or temperature.
All the 13 bones are located in the first quadrant (which is not surprising given our search field, see \autoref{sec:data}), making 27\% of the 48 filaments in the same region of blind BGPS survey.
Obviously, owning to disagreement among the many spiral arm models, adopting a different model will lead to different ``bones''. But the fraction of bones in filaments should not change dramatically in a reasonable spiral model.
\subsection{Fraction of ISM confined in large filaments} \label{sec:fractionISM}
Given the importance of filamentary geometry in enhancing massive clustered star formation,
it is of great interest to quantify the fraction of the ISM contained in large filaments and to evaluate the star formation activities therein.
To address this question we consider only the range of $7.5^{\circ} \le l \le 90.5^{\circ}$ where the BGPS and its spectroscopic follow-up are contiguous.
In this field there are 5841 BGPS v1 sources, 2893 having $\rm{HCO^+}$ or $\rm{N_2H^+}$ (3--2) detection which we term as ``dense BGPS sources''.
We identified 48 filaments in this field, which are comprised of 521 BGPS sources. That means 17.7\% (512/2893) of dense BGPS sources, or 8.8\% (512/5841) of all BGPS sources, are confined in large filaments.
If we count BGPS clumps in the bones only, 6.8\% of dense BGPS sources, or 3.4\% of all BGPS sources are confined in bones.
Compact 1.1 mm continuum emission of the BGPS sources outline the dense, inner part of much larger and less dense envelopes of molecular clouds \citep{Dunham2011-BGPS-NH3}. Assuming a dense gas mass fraction of 10-20\% \citep[cf.][]{Ragan2014-GFL,Ginsburg2015-W51}, we infer an order of 1\% of the molecular ISM is confined in large filaments, and about 1/3 of this amount confined in bones -- marking spiral arm centers.
\begin{figure*}
\centering
\includegraphics[width=.46\textwidth,angle=0]{frag.pdf}
\includegraphics[width=.52\textwidth,angle=0]{sf.pdf}
\caption{
\textit{Left:}
Mean clump mass versus mean edge length for the 54 filaments, as compared to IRDCs and theoretical predictions.
The magenta line is \textit{not} a fit to the data points, instead, it depicts
{cylindrical}
fragmentation appropriate for the filaments (see text). In comparison, the resolved fragmentation of IRDCs \citep[data from][]{me11,me12,me14,qz11} is consistent with turbulent Jeans fragmentation. The green line, orange line, and associated shaded regions correspond to a range of density and temperature appropriate for IRDCs \citep{me14}.
\textit{Right:}
Clump mass versus radius for BGPS clumps, clumps in large filaments, and clumps in other velocity-coherent structures. The color lines depict various empirical criteria of star formation: blue -- the \cite{Krumholz2008Nature} criterion of massive star formation;
yellow -- average of \cite{Heiderman2010} and \cite{Lada2010} for ``efficient'' star formation;
red -- the \cite{Kauffmann2010} criterion of massive star formation with a correction of the
{adopted}
dust opacity law as in \cite{Dunham2011-BGPS-NH3}.
}
\label{fig:frag}
\end{figure*}
\subsection{An apparent length limit of 100 pc and the longest filaments beyond this limit}
\label{sec:100pc}
\cite{me15} pointed out an apparent upper limit of 100 pc projected length for the longest filaments in their study designed to find cold and dense filaments based on \emph{Herschel}\ far-IR emission.
In this study, we use a different approach without limiting the temperature. Except the extremely long filament F5, all other filaments are indeed shorter than 100 pc. Interestingly, this limit is also seen in \cite{Zucker2015} despite of the different search method. The 100 pc limit seems to be present in filaments with a global column density above the order of $10^{21} - 10^{22}$ \,$\rm{cm^{-2}}$. Relaxing the column density to a lower cut of $<10^{20}$ \,$\rm{cm^{-2}}$, longer filaments start to be picked up in $^{12}$CO/$^{13}$CO (1--0): the the 430 pc ``optimistic Nessie'' \citep{Goodman2014}; the 500 pc ``wisp'' \citep{LiGX2013-FL500pc}, and few filaments by \citealt{Ragan2014-GFL} and \citealt{Abreu2016}. However, those CO filaments have much smaller aspect ratios (typically $\ll$10, see figures in their papers), and the low-$J$ CO gas outlines the relatively diffuse envelopes of denser structures traced by MIR extinction. For example, one filament reported by \cite{Ragan2014-GFL} contains F36 as a small part (\autoref{sec:knownFL}).
Whether 100 pc is a true limit for dense filaments warrants further study. This provides a quantitative test case for numerical simulations \citep[e.g.][]{Falceta2015}.
So far, the 276 pc long filament F5 (\autoref{fig:rgb}) is the only exception of dense ($>10^{21}$ \,$\rm{cm^{-2}}$) filaments longer than 100 pc. Compared to the above mentioned extremely long CO filaments, F5 is at least 10 times denser in average column density, and it may also have less dense extensions similar to the 80 pc ``classic Nessie''.
More systematic searches and comparison to numerical simulations can resolve the true length limit of the longest filaments.
We {emphasize}
that average column density is a crucial parameter in defining the boundary of filaments
and thus the length and aspect ratio. It is also worthwhile to note that our filaments, as defined by a collection of dense BGPS clumps, form the center of larger and less dense structures.
The origin of large velocity coherent filaments is still a mystery. Numerical simulations
of the multiphase ISM in galactic disks demonstrate that the cold, dense gas component tends
to organize itself naturally into a filamentary network
\citep[e.g.][]{Tasker2009,Smith2014}.
Spiral arms can sweep up and compress gas, generating bones \citep{Goodman2014}. Gravitationally unstable disk regions condense into gaseous rings and arcs \citep{Behrendt2015}. In differentially rotating disks, structures like molecular cloud complexes could be sheared into elongated filaments.
Our calculations (Burkert et al. in prep) show that tidal effects of the Milky Way are too weak to affect the maximum length of filaments. This is consistent with our observations, where we find filament lengths do not correlate with Galactocentric radii ($C_{\rm Pearson} = -0.07$).
The maximum filament length of order 100 pc might be related to the timescale of $\tau_{\rm SF} = 10^7$ yrs \citep[e.g.][]{Burkert2013} on which stars form in dense molecular gas and destroy their environment. Typical turbulent velocities on large scales in galactic
disks are of order $\sigma = 10$ \,km~s$^{-1}$\ \citep{Dib2006}. If $\sigma$ is the maximum velocity
with which coherent filaments can grow and if $\tau_{\rm SF}$ denotes the timescale on which they are destroyed again their length is limited by $l = \sigma \times \tau_{\rm SF}$ = 100 pc, in agreement with the observations.
\subsection{Fragmentation of large-scale filaments and subsequent star formation} \label{sec:frag}
By definition, the filaments presented in this study are in the form of a chain of dense clumps physically connected by less dense gas in between. For linear filaments, this geometry resembles a fragmented ``cylinder'' with regularly spaced clumps under the ``sausage instability'' of self-gravity (e.g. F9 in \autoref{fig:rgb}). According to the framework of \cite{Chandra1953}, an isothermal gas cylinder becomes super-critical when its linear mass density exceeds the critical value $(M/L)_{\rm crit}$, and will fragment into a chain of equally spaced fragments with a spacing of $\lambda_\mathrm{cl}$, with each fragment having a mass of $M_\mathrm{cl} = (M/L)_{\rm crit} \times \lambda_\mathrm{cl}$.
In short, the fragmentation is governed by central density and pressure (thermal plus non-thermal).
This framework has been followed by many authors \citep[e.g.][]{Ostriker1964_FL,Nagasawa1987_FL,Bastien1991_FL,Inutsuka1992_FL,Fischera2012_FL}. See \cite{me11,me14} for useful deduction of the formulas. \autoref{fig:frag} (left panel) plots the mean clump mass of each filament with the mean separation between clumps (mean length of the edges in the filament). The observed fragmentation is consistent with the theoretical prediction of cylindrical fragmentation assuming a central density of $1\times 10^4$ \,$\rm{cm^{-3}}$\ and velocity dispersion of 0.4--2.2\,km~s$^{-1}$\ (magenta line).
The spread of the data points around the prediction line may be due to a range of central densities and imperfect cylinder geometry. Recent numerical simulations have shown that geometric bending, which is often seen in observed filaments, can change the regularity of the spacing \citep{Gritschneder2016}, indicating that more theoretical work is required in order to understand the stability and dynamics of filaments.
Dense clumps with a typical mass of $10^3$ \,$M_\odot$\ and typical size of 1 pc (\autoref{fig:frag}, left panel) are, in general, capable of forming a cluster of stars.
Statistically, dense clumps residing within large filaments are slightly denser than clumps elsewhere (see below). in \autoref{fig:frag} (right panel), we plot clump mass versus deconvolved radius (not all BGPS clumps have a valid radius, see \autoref{sec:para}) for three categories of BGPS clumps: I -- the 1710 clumps with well determined distance from \cite{Ellsworth2015}. II -- the 294 clumps in velocity-coherent structures but not large filaments. III -- the 469 clumps in large filaments. Categories I, II, and III have 41.0\%, 39.8\%, and 46.2\% clumps satisfying the \cite{Kauffmann2010} threshold of forming massive stars. Thus, categories I and II are indistinguishable, while in comparison, category III is slightly more favorable for massive star formation.
If we count in mass instead of number of clumps, the fractions are 79.2\% for BGPS sources, 86.3\% for velocity-coherent structures but not large filaments, and 91.0\% for large filaments.
Surprisingly, bones do not show a higher fraction compared to large filaments, either counted in number or mass.
This indicates that local environment such as a velocity-coherent filament
plays a role in enhancing massive star formation. Filaments, in particular, provide a preferred form of geometry to channel mass flows that can inhomogeneously feed star-forming clumps
\citep[e.g.][]{Arzoumanian2011,Peretto2014-SDC13,qz15,Heigl2016,Federrath2016}.
On the other hand, the Galactic environment does not seem to affect local star formation across the few hundreds pc spread of vertical position $z$, consistent with previous studies \citep{Eden2012,Eden2013}.
\section{Summary} \label{sec:sum}
We present an automated method designed to identify large-scale velocity-coherent filaments throughout the Galaxy. The method utilizes a customized MST algorithm, which connects neighboring voxels in the PPV space. We have applied the algorithm to the BGPS spectroscopic catalog in the range $7.^{\circ}5 \le l \le 194^{\circ}$, $|b|<0.5^\circ$.
We have identified a comprehensive catalog of 54 large-scale filaments and derived physical parameters including mass ($\sim 10^3 - 10^5$\,$M_\odot$), length (10--276 pc), linear mass density (54--8625 \,$M_\odot$\,pc$^{-1}$), aspect ratio (18--176), linearity, velocity gradient, temperature, fragmentation, Galactic location and orientation angle.
We investigate the Galactic distribution of these parameters and compare the filaments with an updated Galactic spiral arm model.
\begin{enumerate}
\item
In general, the Galactic distribution of large filaments follows the dense gas traced by BGPS sources. However, there are interesting exceptions to be further explored by a quantitative comparison to theoretical work (\autoref{sec:stat}).
\item
Most of the filaments are associated with major spiral arms including the Scutum, Sagittarius and Norma arms, and a few associated with the Local, Perseus, and Outer arms. About 20\% of the filaments are inter-arm structures, or ``spurs'' (\autoref{sec:Gal-dist-bones}).
The filaments tend to align with Galactic plane, but the tendency does not correlate with vertical height (\autoref{fig:theta}).
\item
The filaments are widely distributed across the Galactic disk, with 50\% located within $\pm$20 pc from the Galactic mid-plane and 27\% runs in the center of major spiral arms, forming the ``bones'' of the Milky Way (\autoref{sec:Gal-dist-bones}).
\item
An order of 1\% of the molecular ISM is confined in large filaments, and about 1/3 of this amount is confined in bones -- marking spiral arm centers (\autoref{sec:fractionISM}).
\item
Massive star formation is more favorable in large filaments compared to elsewhere.
However, Galactic environment is not observed to affect local star formation (\autoref{sec:frag}).
\item
An apparent length limit of 100 pc is observed for filaments with a global column density $N_{\rm H_2}$ higher than $10^{21}$\,$\rm{cm^{-2}}$\ (or optical extinction $A_V \approx 1$ mag). Longer filaments are rarer, with a much lower aspect ratio, and have at least one order of magnitude lower global column density (\autoref{sec:100pc}).
\end{enumerate}
Our method can be applied to 3-dimensional PPV catalogs from observations or PPP catalogs from simulations. This study focuses on the northern Galactic plane covered by the BGPS. In the near future, the 2-dimensional ATLASGAL (inner Galactic plane; \citealt{cat:ATLASGAL-Csengeri2014}) and Hi-GAL (full Galactic plane; \citealt{survey:HiGAL-DR1-Molinari2016}) catalogs will be complemented with $V\rm{_{LSR}}$\ information from currently ongoing spectral line surveys
(SEDIGISM: Structure, Excitation, and Dynamics of the Inner Galactic Interstellar Medium, Schuller et al. submitted;
MWISP: Milky Way Imaging Scroll Painting, \citealt{Jiang2013-MWISP,Sun2015-arm};
Mopra Southern Galactic Plane CO Survey, \citealt{Burton2013-MopraCO};
and
ThrUMMS: Three-mm Ultimate Mopra Milky Way Survey, \citealt{Barnes2015-ThrUMMS}).
By then, we can complete the census of large-scale, velocity filaments in the full Galactic plane.
As increasing number of filaments with various morphologies are published in the literature, we urge our colleagues to explicitly list the criteria used to define their filaments before a commonly agreed, physics driven, definition of filaments can be reached in the community.
The final definition will likely be scale dependent, given the different physics behind large-scale and smaller-scale filaments in the ISM.
\acknowledgments{
\textit{Acknowledgments}
We are grateful to Mark Reid and Tom Dame for providing the spiral arm segments, and Alberto Sanna for the parallax distance to maser G11.11-0.11, before publication.
We thank an anonymous referee for a constructive review report.
K.W. is supported by grant WA3628-1/1 of the German Research Foundation (DFG) through the priority program 1573 (``Physics of the Interstellar Medium'').
Color bars in the figures utilize the \textsc{cubehelix} color scheme introduced by \cite{Green2011-cubehelix}.
This research made use of \textsc{Astropy}, a community-developed core Python package for Astronomy \citep{Astropy}.
This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration.
}
\facility{CSO, SMT, \textit{Spitzer}, \textit{WISE}}
\software{\textsc{Python, Astropy}}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FL_a.png}
\caption{A two-color view of the filaments. For F1 to F42, cyan shows the \emph{Spitzer}\ 24\,$\mu\mathrm{m}$\ emission in logarithmic scale \citep[from the MIPSGAL survey,][]{MIPSGAL}, and red shows the APEX+\textit{Planck} 0.87\,mm emission in linear scale \citep[from the ATLASGAL survey,][]{Csengeri2016-ATLASGAL}. For F43 to F54, cyan is \textit{WISE} 22\,$\mu\mathrm{m}$\ emission \citep{WISE} and red is the BGPS 1.1\,mm emission \citep{Ginsburg2013_BGPSv2}.
The filled circles represent dense BGPS sources with color coded velocity as shown in the color bar. The sources are connected by MST edges (see text).
A scale bar of 5 pc is shown for reference.
}
\label{fig:rgb}
\end{figure*}
\setcounter{figure}{5}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FL_b.png}
\caption{Continued.
}
\end{figure*}
\setcounter{figure}{5}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FL_c.png}
\caption{Continued.
}
\end{figure*}
\setcounter{figure}{5}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FL_d.png}
\caption{Continued.
}
\end{figure*}
\setcounter{figure}{5}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FL_e.png}
\caption{Continued.
}
\end{figure*}
\setcounter{figure}{5}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FL_f.png}
\caption{Continued.
}
\end{figure*}
\setlength\tabcolsep{4.5pt}
\clearpage
\begin{turnpage}
\capstartfalse
\input{fl.tex}
\capstarttrue
\end{turnpage}
\clearpage
\setlength\tabcolsep{4.3pt}
\setcounter{table}{0}
\clearpage
\begin{turnpage}
\capstartfalse
\input{fl2.tex}
\capstarttrue
\end{turnpage}
\clearpage
\bibliographystyle{yahapj}
|
\section{Introduction}
Positronium ($e^+e^-$, Ps) and muonium ($\mu^+e^-$, M) being purely leptonic atomic systems are ideal to study bound-state QED free of finite size effects with hadronic contributions strongly suppressed compared to hydrogen.\cite{savely} They are also used for searches of new physics,\cite{Willmann,oPsInv} tests of fundamental symmetries,\cite{asaiPRL2010} and found their application in material science (see, e.g., Refs.\ \refcite{reviewMuAppl,reviewPsAppl}). In this contribution we will review the status and the prospects of the determination of the 1S-2S transition frequencies for Ps and M which are sensitive probes for the Standard-Model Extension (SME).\cite{Kostelecky} The 1S-2S transition was first observed for Ps in 1982\cite{ChuMills1982} and a few years later (1988) for M.\cite{ChuMills1988} In these two measurements positrons and positive muons were implanted in silica powders in which, by capturing an electron, ground state Ps or M were produced. A fraction of these atoms diffused to the silica surface and were emitted into vacuum where they were excited with pulsed lasers to the 2S state.
Subsequent measurements followed in 1984 for positronium\cite{ChuMills1984} with a more efficient target for Ps production made of Al(111) and in 1994 for muonium\cite{Maas1994} with a higher muon flux (almost two orders of magnitude larger). These upgrades allowed reaching uncertainties of 12 ppb and 21 ppb, respectively.
For Ps, the use of a CW laser allowed a further improvement in the determination of the transition frequency at a level of 2.4 ppb in 1993.\cite{Fee1993} For M the best current measurement (4 ppb) was performed in the year 2000 with a refined laser chirp control and the use of long pulses to reduce the time-of-flight broadening\cite{Meyer2000} by increasing the interaction time.
The results of those measurements are in a good agreement with the QED calculations.\cite{savely}
Both experiments are statistically limited and would profit from new sources with a larger flux of colder Ps and M atoms. This can be achieved by more intense/brighter primary beams and/or improved conversion targets.\cite{Jungmann2006} In particular slower atoms would be a great advantage to reduce systematic effects and increase the signal rate because of the extended interaction time with the laser (the excitation probability is proportional to $t^2$).
\section{Status of positronium 1S-2S measurement}
For positronium, bound-state QED calculations reached a level of 0.5 ppb.\cite{QEDPs} A new measurement is ongoing at ETH Zurich\cite{hype15} and at UC Riverside \cite{mills2016} to improve the experimental accuracy to this level. New targets for stable Ps production were developed\cite{APL} and preliminary results were obtained detecting the annihilation of Ps in the 2S states.\cite{hype15} Those prompted the need for improved S/N ratio with the use of a buffer gas trap.\cite{Surko} Efficient extraction of the pulsed beam\cite{Cooke2015b} to a field-free region was achieved reducing the systematic from electromagnetic fields such as DC and motional Stark and Zeeman shifts. Excitation in Rydberg states with subsequent detection via field ionization
will allow correcting for the second-order Doppler shift expected to be the main systematic effect of the experiment caused by the very light mass of Ps resulting in a high velocity ($10^5$ m/s) even at room temperature. The used positron bunches for Ps production have a narrow time window of $1\,\text{ns}$ and the Ps is emitted from porous silica with a well defined velocity determined by the ground state in the pores.\cite{oPsTOF,CassidyPRA2010} A time-of-flight measurement of the 2S-excited Ps atoms will be performed by their detection at a known distance from
a plate in which field Ps atoms will be ionized. With this it is aimed to obtain the mean emission velocity to $\leq 4\%$ by comparison to simulated time spectra resulting in an uncertainty at a level of $100\,\text{kHz}$. An accuracy of 0.5 ppb seems thus in reach with the available Ps targets. Different schemes to produce colder Ps have been proposed including laser cooling (see, e.g., Ref.\ \refcite{crivelli2014}) and Stark deceleration of Rydberg Ps atoms\cite{Hogan2016} pointing to further possible improvements.
\section{Status of muonium 1S-2S measurement}
Recent advances in the production of M into vacuum\cite{Antognini2012} and spatial confinement of M\cite{MuConf} enable CW spectroscopy with current UV technology and with the existing low-energy muon beam line at PSI.\cite{LEM} This will result in a narrower line (about 1 MHz) compared to pulsed lasers (20 MHz dominated by laser chirps). The main systematic effect due to the residual first-order Doppler shift will be at a negligible level due to the use of an enhancement cavity as in hydrogen spectroscopy.\cite{MPQ} A much higher degree of collinearity of the counter-propagating beams will be granted with this, allowing for a measurement of the 1S-2S transition frequency at the 0.2 ppb level (a factor of 20 improvement over Meyer in 2000\cite{Meyer2000}). This results in a test of bound-state QED with an actual theoretical uncertainty of 0.4 ppb.\cite{pachucki, karshenboim} It will also provide the best verification of charge equality in the first two generations of particles and it will improve the determination of the muon mass at the 40 ppb level (a factor 3 better than extracted from the hyperfine splitting measurement\cite{LiuHFS}). This experiment will be statistically limited, therefore, ongoing efforts to develop a high-flux and high-brightness slow-muon beam line\cite{MuCool,Andreas, Strasser:2014dsa} promise even higher accuracy.
\section{Conclusions}
New improved measurements of the 1S-2S transition in positronium and muonium at a level of 0.5 ppb are ongoing and the results are expected in the next few years providing stringent tests of bound-state QED and the SME. For positronium, searches for the annual variations predicted by the SME can be performed more easily since the experiment is not accelerator based as required for muonium production. The ongoing technological developments in this field are aiming in the not too distant future to reach a precision down to a few ppt.
\section*{Acknowledgments
The authors gratefully acknowledge the organizers of the CPT'16 conference for their kind invitation. This work has been supported by the Swiss National Science Foundation under the grant number 200020\_166286 and by the ETH Zurich Research grant ETH-35 14-1.
|
\section{Introduction}\label{section:introduction}
Consider a random sequence $\mathbf{X}=(X_1,\ldots,X_n)$ taking values in an at most countable set ${S}$.
For concreteness, we may regard ${S}$ as a set of species, so that $\Xbf$ records the species for a random sample of animals from a certain population.
From $\Xbf$, we define an equivalence relation $\sim_{\Xbf}$ on $[n]:=\{1,\ldots,n\}$ by
\begin{equation}\label{eq:equiv-relation}
i\sim_{\Xbf}j\quad\text{if and only if}\quad X_i=X_j.
\end{equation}
We write $\Pi(\Xbf)=\{B_1,B_2,\ldots\}$ to denote the set partition whose blocks $B_1,B_2,\ldots$ are the equivalence classes induced by $\sim_{\Xbf}$.
We sometimes write $\sim_{\Pi}$ in place of $\sim_{\Xbf}$ when convenient.
Suppose now that $\Xbf=(X_1,\ldots,X_n)$ is {\em exchangeable}, meaning that
\[(X_{\sigma(1)},\ldots,X_{\sigma(n)})=_{\mathcal{D}}(X_1,\ldots,X_n)\]
for all permutations $\sigma:[n]\to[n]$, where $=_{\mathcal{D}}$ denotes {\em equality in law}.
Exchangeability of $\Xbf$ induces exchangeability on the equivalence relation $\Pi:=\Pi(\Xbf)$ as in \eqref{eq:equiv-relation} in the sense that $\Pi^{\sigma}=_{\mathcal{D}}\Pi$ for all permutations $\sigma:[n]\to[n]$, with $\Pi^{\sigma}$ defined by
\[i\sim_{\Pi^{\sigma}} j\quad\text{if and only if}\quad\sigma^{-1}(i)\sim_{\Pi}\sigma^{-1}(j).\]
Below we present a theory of general {\em relationally exchangeable structures} which contains the above example as a special case and generalizes Kingman's correspondence for exchangeable random partitions \cite{Kingman1978a}.
To appreciate the generality in which we work, let $S$ again be a countable set but now consider a sequence $\Xbf=(X_1,\ldots,X_n)$ taking values in $S\times S$, so that each $X_i$ is an ordered pair $(c_i,r_i)$.
For a concrete application, we can think of each $X_i=(c_i,r_i)$ as identifying the caller and receiver of a phone call sampled uniformly at random from a telephone call log.
While the data comes in the form of a sequence $\Xbf$, the structure relating the calls can be represented by a network-like object as in Figure \ref{fig:calls}.
Since the sequence $\Xbf$ is exchangeable, the model assigns equal probability to structures that are isomorphic up to relabeling edges as in Figures \ref{fig:calls}(a) and \ref{fig:calls}(b).
\begin{figure}[!t
\includegraphics[scale=0.5]{calls-exch.png}
\caption{(a) Representation of sequence of ordered pairs in $S=\{a,b,c,d,e\}$ with $X_1=(a,b), X_2=(c,a), X_3=(d,e), X_4=(a,c)$. (b) Representation of sequence of ordered pairs in $S=\{a,b,c,d,e\}$ with $X'_1=(c,a), X'_2=(d,e), X'_3=(a,c), X'_4=(a,b)$. The object in (b) can be obtained from that in (a) by reordering $(X_1,X_2,X_3,X_4)$ as $(X_2,X_3,X_4,X_1)$. In both (a) and (b), the edges are labeled according to where they appear in sequence.}
\label{fig:calls}
\end{figure}
Just as we disregarded the species names in passing from $\Xbf$ to its induced equivalence relation $\sim_{\Xbf}$ in \eqref{eq:equiv-relation}, we may also disregard the vertex labels in Figure \ref{fig:calls} to obtain an {\em edge-labeled graph}, as shown in Figure \ref{fig:calls-exch}.
The edge-labeled graph is defined formally as the equivalence class $\Xbf_{\cong}$ of all sequences $\Xbf'$ that yield the same structure as $\Xbf$ after removing vertex labels:
\begin{equation}\label{eq:edge-labeled}
\Xbf_{\cong}:=\{\Xbf':[n]\to S\times{S}: \rho \Xbf'=\Xbf\text{ for some bijection }\rho:S\to S\},
\end{equation}
where here we overload notation by allowing the bijection $\rho:S\to S$ to act on $S\times S$ by $(c,r)\mapsto(\rho(c),\rho(r))$, so that $\rho\Xbf':=\rho\circ\Xbf':[n]\to S\times{S}$ is well-defined by composition of functions.
Exchangeability of $\Xbf$ immediately implies {\em edge exchangeability} of the edge-labeled graph associated to $\Xbf_{\cong}$ \cite{CD2016e2}.
\begin{figure}[!t
\includegraphics[scale=0.5]{email-exch-unlab.png}
\caption{(a) Edge-labeled graph induced by disregarding vertex labels in Figure \ref{fig:calls}(a). (b) Edge-labeled graph induced by disregarding vertex labels in Figure \ref{fig:calls}(b). The graphs in (a) and (b) have equal probability under an edge exchangeable model.}
\label{fig:calls-exch}
\end{figure}
For another example, let $S$ represent a set of scientists and let
$\Xbf=(X_1,\ldots,X_n)$ record names of coauthors on a collection of
$n$ scientific journal articles sampled uniformly from a database
(e.g., arXiv).
Each $X_i$ is a finite subset $\{s_{i,1},\ldots,s_{i,k_i}\}\subset S$.
The assumption of uniform sampling again makes $\Xbf$ an exchangeable
sequence, which in turn induces exchangeability on the associated
edge-labeled hypergraph constructed by disregarding vertex labels in
the induced hypergraph structure in a manner analogous to the
construction of edge-labeled graphs from the equivalence class
\eqref{eq:edge-labeled}.
For another example, let $S$ be a set of Internet Protocol (IP) addresses and let each entry of $\Xbf=(X_1,\ldots,X_n)$ correspond to the path taken by a message sent between two IP addresses over the Internet. Each $X_i$ corresponds to a path $(s_i,a_{i,1},\ldots,a_{i,k_i},t_i)$ from source $s_i$ to target $t_i$ by passing through the intermediate nodes $a_{i,1},\ldots,a_{i,k_i}$, altogether indicating that the path traversed $s$ to $a_{i,1}$ to $a_{i,2}$ and so on until passing from $s_{i,k_1}$ to $t_i$.
If $\Xbf$ was obtained, for example, by uniform random sampling of source-target pairs $(s_i,t_i)$ and then by applying an algorithm, such as traceroute, to obtain a path between $s_i$ and $t_i$, then the sequence of paths $\Xbf$ is exchangeable and, therefore, so is the induced path-labeled structure shown in Figure \ref{fig:path}.
Another example in the realm of networks is to take each $X_i$ to be an $r$-step ego network, obtained, e.g., by snowball sampling a neighborhood of size $r$ from a randomly chosen vertex.
The relationally labeled structure is constructed by piecing together the
neighborhoods obtained from the $r$-neighborhoods of $n$ randomly
chosen ``egos'' in this population. We omit the details of this
example and instead move on to our general treatment.
Myriad other data structures arise according to a similar recipe: let $S$ be a set of elements and $\mathcal{R}$ be a set of relations on the elements in $S$; sample an exchangeable sequence $\Xbf$ taking values in $\mathcal{R}$; construct a structure from $\Xbf$ (as in Figure \ref{fig:calls}) and obtain the corresponding relationally labeled structure by removing the vertex labels (as in Figure \ref{fig:calls-exch}).
The resulting structure is {exchangeable} with respect to relabeling of its relations, a property which we call {\em relational exchangeability}.
We now define this notion formally and prove a generic structure theorem for infinite relationally exchangeable structures. The formal statement is given in Theorem \ref{thm:general}.
\begin{figure}[!t
\includegraphics[scale=0.5]{path-exch.png}
\caption{(a) and (b) both correspond to the structures induced by taking the equivalence class over paths between vertices. In this illustration, all paths have the same source vertex and each path is labeled by a different color, so that the structure in (b) can be obtained by recoloring the paths from (a) (i.e., green to orange, red to green, and orange to red).}
\label{fig:path}
\end{figure}
\section{Relational Exchangeability}\label{section:preliminaries}
Let $(S,\mathcal{S})$ be any Borel space. A countably infinite sequence of $S$-valued random variables $\Xbf=(X_1,X_2,\ldots)$ is {\em exchangeable} if $\Xbf^{\sigma}=_{\mathcal{D}}\Xbf$ for every permutation $\sigma:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits$, where $=_{\mathcal{D}}$ denotes {\em equality in distribution} and $\Xbf^{\sigma}:=(X_{\sigma(1)},X_{\sigma(2)},\ldots)$ is the reordering of $\Xbf$ according to $\sigma$.
By de Finetti's theorem \cite{deFinetti1937}, the distribution of every such countably infinite sequence can be expressed as a mixture of independent, identically distributed (i.i.d.)~sequences.
In particular, with $\mathcal{P}(S)$ denoting the space of probability measures on $(S,\mathcal{S})$, there is a unique probability measure $\phi$ on $\mathcal{P}(S)$ such that the distribution of $\Xbf$ can be expressed as
\begin{equation}\label{eq:definetti}
\mathbb{P}(\Xbf\in\cdot)=\int_{\mathcal{P}(S)}\nu^{\infty}(\cdot)\phi(d\nu),
\end{equation}
where $\nu^{\infty}$ denotes the infinite product measure induced by $\nu$ on $S^{\infty}$. We call the measure $\phi$ in \eqref{eq:definetti} the {\em de Finetti measure} of $\Xbf$.
de Finetti's theorem figures into our treatment of relationally exchangeable structures, which we now define. Note that all of the examples in Section \ref{section:introduction}, and many more that could arise in practice, involve relational structures as they are defined in the coming section. Example \ref{ex:example} provides a concrete illustration.
\subsection{Relational structures}
Let $\ar:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits\cup\{0\}$ be a \emph{signature}
such that $\ar(j)=0$ implies $\ar(k)=0$ for all $k\geq j$.
An {\em $\ar$-structure with domain $A$} is a collection $\mathcal{A}=(R_1^{\mathcal{A}}, R_2^{\mathcal{A}},\ldots)$, where $A$ is a set and each $R_j^{\mathcal{A}}\subseteq A^{\ar(j)}$ is a relation of arity $\ar(j)$ on $A$.
We adopt the convention $A^0:=\emptyset$ so that $R_j^{\mathcal{A}}=\emptyset$ whenever $\alpha(j)=0$.
We write $\mathop{\mathrm{dom}}\nolimits\mathcal{A}:=A$ for the {\em domain} of $\mathcal{A}$
and $\mathop{\textsc{Fin}}\nolimits_{\ar}$ for the set of all $\alpha$-structures
$\mathcal{A}$ with finite
$\mathop{\mathrm{dom}}\nolimits\mathcal{A}\subset\mathop{\mathbb{Z}_{}}\nolimits=\{\ldots,-1,0,1,\ldots\}$
such that $\sum_{j=1}^{\infty}|R_j^{\mathcal{A}}|<\infty$, where $|S|$
denotes the cardinality of a set $S$. Writing $\mathop{\mathrm{dom}}\nolimits\mathcal{A}\mathop{\subset_f}\mathop{\mathbb{Z}_{}}\nolimits$ to denote that $\mathop{\mathrm{dom}}\nolimits\mathcal{A}$ is a finite subset of $\mathop{\mathbb{Z}_{}}\nolimits$, we have
\[\mathop{\textsc{Fin}}\nolimits_{\ar}:=\{\mathcal{A}=(R_1^{\mathcal{A}},R_2^{\mathcal{A}},\ldots):
\mathop{\mathrm{dom}}\nolimits\mathcal{A}\mathop{\subset_f}\mathop{\mathbb{Z}_{}}\nolimits\text{ and }\sum_{j=1}^{\infty}|R_j^{\mathcal{A}}|<\infty\}.\]
Condition $\sum_{j\geq1}|R_j^{\mathcal{A}}|<\infty$ implies that to every $\mathcal{A}\in\mathop{\textsc{Fin}}\nolimits_{\ar}$ there is an $r_{\mathcal{A}}:=\max\{k\geq1: R_k^{\mathcal{A}}\neq\emptyset\}<\infty$, making it convenient to sometimes express $\mathcal{A}\in\mathop{\textsc{Fin}}\nolimits_{\ar}$ as $(R_1^{\mathcal{A}},\ldots,R_{r_{\mathcal{A}}}^{\mathcal{A}})$
by omitting the infinite sequence of empty relations after $r_{\mathcal{A}}$.
For any $A^*\supseteq A$ and any injection $\rho:A^*\to A'$ there is an induced action on $\mathcal{A}=(R_1^{\mathcal{A}},R_2^{\mathcal{A}},\ldots)$ by $\mathcal{A}\mapsto\rho(\mathcal{A})=(R_1^{\rho(\mathcal{A})}, R_2^{\rho(\mathcal{A})},\ldots)$ with $\mathop{\mathrm{dom}}\nolimits\rho(\mathcal{A})=\rho(A):=\{\rho(a): a\in A\}$ and
\begin{equation}\label{eq:relabel}(\rho(a_1),\ldots,\rho(a_{\ar(j)}))\in R_j^{\rho(\mathcal{A})}\quad\text{if and only if}\quad(a_1,\ldots,a_{\ar(j)})\in R_j^{\mathcal{A}}\end{equation}
for each $j=1,2,\ldots$.
For $n\geq1$ and any at most countable subset of finite $\ar$-structures $\mathcal{R}\subseteq\mathop{\textsc{Fin}}\nolimits_{\ar}$ for which each $\mathcal{A}\in\mathcal{R}$ has $\mathop{\mathrm{dom}}\nolimits\mathcal{A}\subseteq\mathop{\mathbb{Z}_{}}\nolimits_+:=\{1,2,\ldots\}$, we consider the set $\mathop{\mathcal{R}_{[n]}^{\cong}}\nolimits$ of equivalence classes of $\mathcal{R}$-valued sequences $\xbf:[n]\to\mathcal{R}$ obtained as follows.
Any $\xbf:[n]\to\mathcal{R}$ determines a {\em relationally labeled structure} by removing the labels of the elements contained in each of the $\xbf(i)$ while maintaining the integrity of the overall structure induced by $\xbf$.
Formally, we define the {\em relationally labeled structure induced by $\xbf$} as the equivalence class
\begin{equation}\label{eq:cong}\xbf_{\cong}:=\{\xbf':[n]\to\mathcal{R}: \rho \xbf'= \xbf\text{ for some bijection }\rho:\mathop{\mathbb{Z}_{}}\nolimits\to\mathop{\mathbb{Z}_{}}\nolimits\},\end{equation}
where $\rho \xbf:[n]\to\mathcal{R}$ is defined by $(\rho \xbf)(i)=\rho(\xbf(i))$ as in \eqref{eq:relabel} above.
We write $\mathop{\mathcal{R}_{[n]}^{\cong}}\nolimits$ as the set of all $\xbf_{\cong}$ constructed from
some $\xbf:[n]\to\mathcal{R}$ in this way.
\begin{rmk}[Notation]
Below we reserve lowercase letters $x,y,\ldots$ for generic (non-random) objects and uppercase letters $X,Y,\ldots$ for random objects. We use bold letters $\xbf,\ybf,\ldots$ for generic (non-random) functions $\xbf:[n]\to\mathcal{R}$ and $\Xbf,\Ybf,\ldots$ for random functions $\Xbf:[n]\to\mathcal{R}$.
\end{rmk}
\begin{example}\label{ex:example}
In this example, we demonstrate how the structures in Section \ref{section:introduction} fit into the framework of $\alpha$-structures. For this we write $\mathcal{A}=(R_1^{\mathcal{A}},R_2^{\mathcal{A}},\ldots)$ as an $\alpha$-structure with $\mathop{\mathrm{dom}}\nolimits\mathcal{A}=A$ and $\alpha$ varying according to the context.
For $\ar(1)=1$ and $\ar(k)=0$ for $k\geq2$, each $\ar$-structure
$\mathcal{A}$ is determined by a subset $R_1^{\mathcal{A}}\subseteq A$.
If we then choose $\mathcal{R}\subseteq\mathop{\textsc{Fin}}\nolimits_{\ar}$ to consist of all
singleton subsets of $\mathop{\mathbb{N}_{}}\nolimits$, so that each $R_1^{\mathcal{A}}$ has the
form $\{i\}$, $i\geq1$, then the equivalence class of
$\xbf:[n]\to\mathcal{R}$ in \eqref{eq:cong} corresponds to the
equivalence relation induced by $\xbf$ as in \eqref{eq:equiv-relation}.
For $\ar(1)=2$, $\ar(k)=0$ for $k\geq2$, each $\ar$-structure
$\mathcal{A}$ with $\mathop{\mathrm{dom}}\nolimits\mathcal{A}=A$ corresponds to a binary relation $R_1^{\mathcal{A}}\subseteq A^2$.
Taking $\mathcal{R}\subseteq \mathop{\textsc{Fin}}\nolimits_{\ar}$ to consist of all structures
with $R_1^{\mathcal{A}}=\{(i,j)\}$ for $i\neq j$ corresponds to sampling
from the phone call database in Section \ref{section:introduction}:
each sampled $X_k=\{(i,j)\}$ corresponds to a caller-receiver pair
$(i,j)$. The equivalence class $\xbf_{\cong}$ gives an edge-labeled
(directed) graph as in Figure \ref{fig:calls-exch}. (To get an
undirected graph we would take each $R_1^{\mathcal{A}}$ to consist of symmetric pairs $\{(i,j),(j,i)\}$.)
Taking $\alpha(j) = j$ for all $j\geq1$ means that any
$\alpha$-structure is determined by a collection
$R_j^{\mathcal{A}}\subseteq A^j$ of subsets of $j$-tuples for every
$j\geq1$.
If we take $\mathcal{R}\subset\mathop{\textsc{Fin}}\nolimits_{\ar}$ to consist only
of those $\alpha$-structures of the form
$R_k^{\mathcal{A}}=\{(a_{\sigma(1)},a_{\sigma(2)},\ldots,a_{\sigma(k)}):
\text{permutations }\sigma:[k]\to[k] \}$ for some $k\geq1$ and
$R_j^{\mathcal{A}}=\emptyset$ for all $j\neq k$, then the elements
$\mathcal{A}\in\mathcal{R}$ can be used to represent the set of
coauthors in a sample of articles.
Also for $\alpha(j) = j$ for all $j \geq 1$, if we take
$\mathcal{R}\subset\mathop{\textsc{Fin}}\nolimits_{\ar}$ to consist only of those
$\alpha$-structures of the form
$R_k^{\mathcal{A}}=\{(a_1,a_2,\ldots,a_{k})\}$ for some $k\geq1$ and
$R_j^{\mathcal{A}}=\emptyset$ for all $j\neq k$, then each
$\mathcal{A}\in\mathcal{R}$ corresponds to a path from $a_1$ to
$a_{k}$, as discussed at the end of Section
\ref{section:introduction}.
\end{example}
Our discussion below specializes to the case of random structures labeled by the countable set $\mathop{\mathbb{N}_{}}\nolimits$,
which are those structures constructed just as in \eqref{eq:cong} but for a countable sequence $\xbf:\mathop{\mathbb{N}_{}}\nolimits\to\mathcal{R}$.
(Note well the difference between the index set $\mathop{\mathbb{N}_{}}\nolimits$ and the domain of the structures in $\mathcal{R}$,
which we take to be $\mathop{\mathbb{Z}_{}}\nolimits_{+}$. The domain labeling is ``quotiented out'' in \eqref{eq:cong}, while the indexing $\mathop{\mathbb{N}_{}}\nolimits$ remains, serving as the ``edge labels'' in the associated interpretation as an edge-labeled graph in Figure \ref{fig:calls-exch}.)
We write $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ to denote the set of such structures and we equip $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ with the Borel $\sigma$-field associated to its product-discrete topology induced by the following metric. For any $\xbf_{\cong}\in\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ and $n\geq1$, we define the restriction $\mathop{\mathbf{R}_n}\nolimits\xbf_{\cong}$ of $\xbf_{\cong}$ to $\mathop{\mathcal{R}_{[n]}^{\cong}}\nolimits$ as the structure obtained by taking any $\xbf'\in\xbf_{\cong}$ (i.e., $\xbf':\mathop{\mathbb{N}_{}}\nolimits\to\mathcal{R}$ for which there exists $\rho$ such that $\rho\xbf'=\xbf$), restricting its domain to $[n]$ by $\xbf'_{|[n]}:[n]\to\mathcal{R}$, $i\mapsto \xbf'(i)$, and putting $\mathop{\mathbf{R}_n}\nolimits\xbf_{\cong}:=(\xbf'_{|[n]})_{\cong}$ as in \eqref{eq:cong}. By the definition of $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$, it is clear that this is well-defined and does not depend on the specific choice of $\xbf'\in\xbf_{\cong}$. We then define the metric on $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ by
\[d(\xbf_{\cong},\xbf'_{\cong}):=1/(1+\sup\{n\geq1: \mathop{\mathbf{R}_n}\nolimits\xbf_{\cong}=\mathop{\mathbf{R}_n}\nolimits\xbf'_{\cong}\}),\quad\xbf_{\cong},\xbf'_{\cong}\in\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits,\]
with the convention that $1/\infty=0$, under which $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ is compact.
For any permutation $\sigma:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits$, we define the {\em relabeling of $\xbf_{\cong}\in\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ by $\sigma$} as the structure $\xbf_{\cong}^{\sigma}$ obtained by first choosing any $\xbf'\in\xbf_{\cong}$ and putting $\xbf_{\cong}^{\sigma}:=(\xbf'\circ\sigma)_{\cong}$, where $\xbf'\circ\sigma:\mathop{\mathbb{N}_{}}\nolimits\to\mathcal{R}$ is defined by usual composition of functions, $(\xbf'\circ\sigma)(i):=\xbf'(\sigma(i))$. It is once again clear that this does not depend on the specific choice of representative $\xbf'\in\xbf_{\cong}$ since the actions of $\rho:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits$ and $\sigma:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits$ commute for all $\xbf'$, i.e., $\rho(\xbf'\circ\sigma)=(\rho\xbf')\circ\sigma$.
\begin{defn}[Relational exchangeability]\label{defn:relational}
A random structure $\Xbf_{\cong}\in\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ is {\em relationally exchangeable} if $\Xbf^{\sigma}_{\cong}=_{\mathcal{D}} \Xbf_{\cong}$ for all permutations $\sigma:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits$.
\end{defn}
\subsection{Representation theorem}
Our main theorem establishes a generic construction for infinite relationally exchangeable structures $\Xbf_{\cong}$ in $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$. The construction proceeds by sampling a sequence $X_1,X_2,\ldots$ conditionally i.i.d.\ in a related space $\mathcal{R}^\star$ and then modifying these observations to obtain a new sequence $\Xbf^{\dagger}:=(X^\dagger_1,X^\dagger_2,\ldots)$. We then construct the equivalence class $\Xbf^{\dagger}_{\cong}$ based on the sequence $X^\dagger_1,X^\dagger_2,\ldots$ as in \eqref{eq:cong}. Since stating the theorem formally requires some new ideas and notation, we give the construction and key ideas of the proof prior to stating the result in Theorem \ref{thm:general}.
As above, let $\ar:\mathop{\mathbb{N}_{}}\nolimits\to\mathop{\mathbb{N}_{}}\nolimits\cup\{0\}$ and $\mathcal{R}\subseteq\mathop{\textsc{Fin}}\nolimits_{\ar}$ be an at most countable set of $\ar$-structures such that $\ar(j)=0$ implies $\ar(k)=0$ for all $k\geq j$ and each $\mathcal{A}\in\mathcal{R}$ has $\mathop{\mathrm{dom}}\nolimits\mathcal{A}\mathop{\subset_f}\mathop{\mathbb{Z}_{}}\nolimits_+$. Since $\mathcal{R}$ is at most countable, we may fix an ordering $\mathcal{R}=\{\mathcal{R}_n\}_{n\geq1}$ of its elements.
Given any $\mathcal{A}=(R_1^{\mathcal{A}},R_2^{\mathcal{A}},\ldots)\in\mathop{\textsc{Fin}}\nolimits_{\ar}$ and any $[0,1]$-valued sequence $(T_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$, we define $T\mathcal{A}:\equiv T\circ\mathcal{A}$ as the $\alpha$-structure $T\mathcal{A}$ with $\mathop{\mathrm{dom}}\nolimits(T\mathcal{A}):=\{T_i: i\in\mathop{\mathrm{dom}}\nolimits\mathcal{A}\}$ and relations $R_j^{T\mathcal{A}}$ given by
\begin{equation}\label{eq:xi-structure}
(T_{a_1},\ldots,T_{a_{\alpha(j)}})\in R_j^{T\mathcal{A}}\quad\text{if and only if}\quad(a_1,\ldots,a_{\alpha(j)})\in R_j^{\mathcal{A}}\end{equation}
for each $j=1,2,\ldots$.
We then define
\[\mathcal{R}_{[0,1]}:=\bigcup_{n\geq1}\{T\mathcal{R}_n: T=(T_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}\in[0,1]^{\mathop{\mathbb{Z}_{}}\nolimits}\}\]
as the set of $\ar$-structures obtained by associating $[0,1]$-valued labels to the elements of $\mathcal{R}$.
More generally, if $(\Xi_{i})_{i\in\mathop{\mathrm{dom}}\nolimits\mathcal{A}}$ is a collection of subsets $\Xi_i\subseteq[0,1]$, then we define $\Xi\mathcal{A}$ by
\[\Xi\mathcal{A}:=\{T\mathcal{A}: T_i\in\Xi_i\text{ for each }i\in\mathop{\mathrm{dom}}\nolimits\mathcal{A}\}.\]
We equip $\mathcal{R}_{[0,1]}$ with the $\sigma$-field on $\mathcal{R}_{[0,1]}$ generated by all sets of the form $\Xi\mathcal{A}$ with $\mathcal{A}\in\mathcal{R}$ and $\Xi=(\Xi_i)_{i\in\mathop{\mathrm{dom}}\nolimits\mathcal{A}}$ a collection of Borel subsets of $[0,1]$.
From any $\xbf_{\cong}\in\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ and any sequence $\xi=(\xi_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ of i.i.d.\ Uniform$[0,1]$ random variables, we write $\xi\xbf_{\cong}$ to denote a random $\mathop{\mathcal{R}_{[0,1]}}\nolimits$-valued sequence $(Z_i)_{i\geq1}$ obtained by first taking any representative $\xbf'\in\xbf_{\cong}$ and then putting $Z_i=\xi\xbf'(i)$ for each $i\geq1$, with $\xi\xbf'(i)$ defined as in \eqref{eq:xi-structure}.
Since $\xbf_{\cong}$ is fixed in this example, each $\xi \xbf'(i)$ corresponds to an assignment of Uniform$[0,1]$ random labels to the elements in the domain of $\xbf'(i)$. Since $\xi_i\neq\xi_j$ with probability 1 for all $i\neq j$, it is immediate that
the distribution of $\xi\xbf_{\cong}$ does not depend on the manner in which the representative $\xbf'$ is chosen.
Now for any infinite relationally exchangeable structure $\Xbf_{\cong}\in\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ and a sequence $\xi$ of Uniform$[0,1]$ random variables independent of $\Xbf_{\cong}$, $\xi \Xbf_{\cong}=(Z_i)_{i\geq1}$ defines a random sequence obtained by putting $Z_i=\xi\xbf'(i)$ for a representative $\xbf'\in\xbf_{\cong}$ on the event $\Xbf_{\cong}=\xbf_{\cong}$.
By exchangeability of $\Xbf_{\cong}$, the sequence $\mathop{\mathbf{Z}_{}}\nolimits=(Z_i)_{i\geq1}$ obtained in this way is also exchangeable and de Finetti's theorem implies that $\mathop{\mathbf{Z}_{}}\nolimits$ is distributed as an i.i.d.\ sequence from a random measure $\nu$ on $\mathop{\mathcal{R}_{[0,1]}}\nolimits$, as described in \eqref{eq:definetti}.
Given $\nu$, we define the {\em propensity of $u\in[0,1]$ in $\mathop{\mathbf{Z}_{}}\nolimits$} by
\begin{equation}\label{eq:nu-star}\nu^\star(u):=\nu(\{\mathcal{A}\in\mathcal{R}_{[0,1]}: u\in\mathop{\mathrm{dom}}\nolimits\mathcal{A}\}),\end{equation}
which equals the conditional probability of the event $\{u\in \mathop{\mathrm{dom}}\nolimits Z_j\}$ given $\nu$, for each $j\geq1$.
It is clear that each $u\in[0,1]$ appears in either 0, 1, or infinitely many of the relations $(Z_i)_{i\geq1}$ with probability 1. First, if $\nu^\star(u)>0$, then the strong law of large numbers implies that $u$ occurs in infinitely many of the relations $Z_i$ with proportion $\nu^\star(u)$. If $\nu^\star(u)=0$ and $u\in Z_i$ for some $i$, then the probability that $u$ occurs in $Z_j$ is $\nu^\star(u)=0$ independently for each $j\neq i$ and therefore $u$ appears only once in $\mathop{\mathbf{Z}_{}}\nolimits$ with probability 1.
Clearly the set $\mathcal{U}=\{u:\nu^\star(u)>0\}$ is at most countable since $|\mathop{\mathrm{dom}}\nolimits(Z_i)|<\infty$ for each $i=1,2,\ldots$. We can, therefore, order the elements of $\mathcal{U}$ as $u_1,u_2,\ldots$ such that $\nu^\star(u_j)\geq\nu^\star(u_{j+1})$ for $j\geq1$, breaking ties $\nu^\star(v)=\nu^\star(w)$ as follows. For each $\mathcal{A}\in\mathcal{R}$, we define
\[\nu^\star(u;\mathcal{A}):=\nu(\{T\mathcal{A}: T=(T_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}\in[0,1]^{\mathop{\mathbb{Z}_{}}\nolimits}\text{ such that } u\in\mathop{\mathrm{dom}}\nolimits(T\mathcal{A})\}),\quad u\in[0,1],\]
to be the measure assigned to the subset of $\mathcal{R}_{[0,1]}$ whose structure is consistent with $\mathcal{A}$ and which contains $u$ in its domain.\label{tiebreak}
Assuming $\nu^\star(v)=\nu^\star(w)$, we assign the smaller label to $v$ if there is some $k\geq1$ such that $\nu^\star(v;\mathcal{R}_j)=\nu^\star(w;\mathcal{R}_j)$ for all $j<k$ and $\nu^\star(v;\mathcal{R}_k)>\nu^\star(w;\mathcal{R}_k)$.
If $\nu^\star(v;\mathcal{R}_j)=\nu^\star(w;\mathcal{R}_j)$ for all $j\geq1$, then we label $v$ and $w$ in increasing order, i.e., if we are to assign labels $j$ and $j+1$ to $v$ and $w$ and if $v<w$, then we shall put $u_j=v$ and $u_{j+1}=w$; otherwise, we put $u_j=w$ and $u_{j+1}=v$.
The ordering $\mathcal{U}=(u_j)_{j\geq1}$ is thus uniquely determined by $\nu$ and the fixed ordering of $\mathcal{R}$ chosen at the outset.
Now given $\mathcal{U}$, for any $\mathcal{A}\in\mathcal{R}_{[0,1]}$ we define $\mathcal{A}^\star$ (suppressing the dependence on $\nu$) by replacing each occurrence of $u_j\in\mathcal{U}$ by $j$ and replacing each occurrence of $v'\notin\mathcal{U}$ in $\mathcal{A}$ by a unique non-positive integer $z(v')=0,-1,-2,\ldots$ so that for $v',v''\notin\mathcal{U}$ and both in $\mathop{\mathrm{dom}}\nolimits\mathcal{A}$ with $v'\leq v''$ we have $z(v')\leq z(v'')$ and the $z(v')$ are chosen to be the largest possible non-positive integers that satisfy this condition. (For example, if $v_1<v_2<v_3$ are the only elements in $\mathop{\mathrm{dom}}\nolimits\mathcal{A}$ that are not in $\mathcal{U}$, we assign $z(v_3)=0$, $z(v_2)=-1$, and $z(v_1)=-2$.)
Every $\mathcal{A}\in\mathcal{R}_{[0,1]}$ thus corresponds to a unique such $\mathcal{A}^\star$ and we define $\mathcal{R}^\star$ as the set of all structures $\mathcal{A}^\star$ obtained in this way. Note that since $\mathcal{R}$ is at most countable, so is $\mathcal{R}^\star$. Also, although we have constructed $\mathcal{R}^\star$ using $\mathcal{U}$ (and therefore $\nu$), the set $\mathcal{R}^\star$ depends only on $\mathcal{R}$, justifying the term {\em $\mathcal{R}$-simplex} in our definition of $\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ in \eqref{eq:FR} below.
The elements of $\mathcal{R}^\star$ are thus $\ar$-structures $\mathcal{A}^\star$ with $\mathop{\mathrm{dom}}\nolimits\mathcal{A}^\star\subseteq\mathop{\mathbb{Z}_{}}\nolimits$.
We define the $\mathcal{R}$-simplex $\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ by
\begin{equation}\label{eq:FR}\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits:=\left\{(f_{\mathcal{B}})_{\mathcal{B}\in\mathcal{R}^\star}:\ f_{\mathcal{B}}\geq0\text{ and } \sum_{\mathcal{B}\in\mathcal{R}^\star}f_{\mathcal{B}}=1\right\},\end{equation}
on which we equip the metric
\[d_{\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits}(f,f'):=\sum_{\mathcal{B}\in\mathcal{R}^\star}|f_{\mathcal{B}}-f'_{\mathcal{B}}|,\quad f,f'\in\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits,\]
and the associated Borel $\sigma$-field.
Any $f\in\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ determines a unique probability measure $\varepsilon_f$ on $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$ by first drawing $\Xbf=(X_1,X_2,\ldots)$ i.i.d.\ from
\begin{equation}\label{eq:f-induced}\mathbb{P}(X_i=\mathcal{B}\mid f)=f_{\mathcal{B}},\quad\mathcal{B}\in\mathcal{R}^\star,\end{equation}
and then constructing $\Xbf^\dagger=(X^\dagger_1,X^\dagger_2,\ldots)$ from $\Xbf$ as follows.
We initialize by putting $m_0=0$.
For each $n\geq1$, given $m_{n-1}$, we replace the non-positive elements in $\mathop{\mathrm{dom}}\nolimits X_n$ according to the rule:
\begin{itemize}
\item[(i)] If $\mathop{\mathrm{dom}}\nolimits X_n$ has no non-positive elements, then put $m_n=m_{n-1}$ and $X^\dagger_n:=X_n$.
\item[(ii)] If $\mathop{\mathrm{dom}}\nolimits X_n$ has non-positive elements $0,\ldots,-k$ for $k\geq0$, then define $X^\dagger_n$ by replacing each occurrence of $-i$, for $0\leq i\leq k$, in $X_n$ by $m_{n-1}-i$, then putting $m_{n}=m_{n-1}-k-1$ and keeping positive elements unchanged. See Example \ref{example:partition} below for an illustration.
\end{itemize}
Note the distinction between the use of non-positive elements in construction of $X_i$ and $X_i^\dagger$ respectively.
The non-positive elements of each $X_i \in \mathcal{R}^\star$ serve to denote the non-recurring particles that appear \emph{only within} this particular relation.
The non-positive elements of $X_i^\dagger$, on the other hand, serve to denote the non-recurring particles across \emph{all} relations in the
sequence~$\Xbf^\dagger$.
We define $\varepsilon_f$ to be the distribution of $\Xbf^\dagger_{\cong}$ constructed by applying \eqref{eq:cong} to the sequence $\Xbf^\dagger$.
From $\nu$, we define $f^{\nu}=(f_{\mathcal{B}}^{\nu})_{\mathcal{B}\in\mathcal{R}^\star}\in\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ by
\begin{equation}\label{eq:canonical}f^{\nu}_{\mathcal{B}}:=\nu(\{\mathcal{A}\in\mathcal{R}_{[0,1]}: \mathcal{A}^\star=\mathcal{B}\}),\quad\mathcal{B}\in\mathcal{R}^\star.\end{equation}
This choice of $f^{\nu}=(f^{\nu}_{\mathcal{B}})_{\mathcal{B}\in\mathcal{R}^\star}$ is uniquely determined by $\nu$ and the fixed ordering of $\mathcal{R}$.
Conversely, given $f=(f_{\mathcal{B}})_{\mathcal{B}\in\mathcal{R}^\star}\in\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ we construct a measure $\nu_f$ on $\mathcal{R}_{[0,1]}$ to be the distribution of $\xi Y$ from \eqref{eq:xi-structure} for $Y$ drawn from distribution \eqref{eq:f-induced} and $\xi=(\xi_{i})_{i\in\mathbb{Z}}$ i.i.d.\ Uniform$[0,1]$ independent of $Y$.
Proposition~\ref{prop:nu} proves that the above procedure
does not alter the random, relationally labeled structure.
\begin{prop}\label{prop:nu}
Let $\Xbf_{\cong}$ be relationally exchangeable and let $\theta=(\theta_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ be i.i.d.\ Uniform$[0,1]$ independent of $\Xbf_{\cong}$. Then $(((\theta\Xbf_{\cong})^{\star})^{\dagger})_{\cong}=\Xbf_{\cong}$ a.s., where $(\theta\Xbf_{\cong})^\star$ denotes the application of $^{\star}:\mathcal{R}_{[0,1]}\to\mathcal{R}^\star$ to each component of the sequence $\theta\Xbf_{\cong}$.
\end{prop}
\begin{proof}
Let $\theta=(\theta_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ be i.i.d.\ Uniform$[0,1]$ independently of $\Xbf_{\cong}$.
Each event $\Xbf_{\cong}=\xbf_{\cong}$ gives rise to a probability measure $\nu$ on $\mathcal{R}_{[0,1]}$ through the de Finetti measure of $\theta\xbf_{\cong}$; see \eqref{eq:definetti}. Let $\mathcal{U}=(u_i)_{i\geq1}$ be the ordered subset of $[0,1]$ corresponding to the atoms of $\nu^\star$ as in \eqref{eq:nu-star}. Since $\theta$ are i.i.d.\ Uniform$[0,1]$, we have
\[\mathbb{P}\{\theta_i\neq\theta_j\text{ for all }i\neq j\}=1,\]
which implies that distinct $i,j\in\mathop{\mathbb{Z}_{}}\nolimits$ are labeled distinctly in $\theta\xbf_{\cong}$ with probability 1.
In particular, the non-positive labels in $(\theta\xbf_{\cong})^\star$ account for all those elements that appear in only one entry of the sequence $\theta\xbf_{\cong}$.
It follows that $((\theta\xbf_{\cong})^\star)^{\dagger}\in\xbf_{\cong}$ with probability 1 and, thus, $(((\theta\xbf_{\cong})^\star)^{\dagger})_{\cong}=\xbf_{\cong}$ with probability 1 for all possible outcomes $\Xbf_{\cong}=\xbf_{\cong}$.
\end{proof}
\begin{example}\label{example:partition}
To illustrate the above procedure, let $\mathcal{R}$ consist of the singleton sets $(\{i\})$, $i\geq1$, just as in the first part of Example \ref{ex:example}, written as $\{1\},\{2\},\ldots$ for simplicity.
Let $\mathcal{U}=(u_i)_{i\geq1}$ be a countable subset of $[0,1]$.
From $\mathcal{R}=\{\{1\},\{2\},\ldots\}$, we obtain $\mathcal{R}^\star=\{\{0\},\{1\},\ldots\}$ since for any sequence $T=(T_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ and any $\{i\}$, $i\geq1$, the transformed relation $T\{i\}\equiv\{T_i\}$ has either $T_i\in\mathcal{U}$ or $T_i\not\in\mathcal{U}$. If $T_i=u_k\in\mathcal{U}$, then $\{T_i\}^\star=\{k\}$ for $k\geq1$; and if $T_i=v\not\in\mathcal{U}$, then $\{T_i\}^\star = (\{0\})$.
Given $(f_{\{0\}},f_{\{1\}},\ldots)\in\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$, we generate a sequence $X_1,X_2,\ldots$ of singleton sets i.i.d.\ as in \eqref{eq:f-induced}, from which we obtain $X_1^{\dagger},X_2^{\dagger},\ldots$ by reassigning occurrences of $0$ to the greatest negative integer that has not yet appeared in the sequence.
For example, if the sequence begins $\{4\},\{0\},\{2\},\{0\},\{2\},\{2\},\{0\},\ldots$, then we reassign labels to obtain $\{4\},\{0\},\{2\},\{-1\},\{2\},\{2\},\{-2\},\ldots$.
Delabeling according to \eqref{eq:cong} (or equivalently
\eqref{eq:equiv-relation}) yields the equivalence classes $\{1\}$,
$\{2\}$, $\{3,5,6\}$, $\{4\}$, $\{7\}$.
For another example, let $\mathcal{R}$ consist of ordered pairs $\{(i,j)\}$, $1\leq i\neq j<\infty$, so that we are in the phone call example of Section \ref{section:introduction}. Let $\mathcal{U}=(u_i)_{i\geq1}$ be a countable subset of $[0,1]$.
Then $\mathcal{R}=\{\{(i,j)\}: i\neq j\geq1\}$ and
\[\mathcal{R}^{\star}=\{\{(i,0)\}: i\geq1\}\cup\{\{(0,i)\}: i\geq1\}\cup\{\{(0,-1)\},\{(-1,0)\}\}\cup\mathcal{R}\] since any $T=(T_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ and any $\{(i,j)\}\in\mathcal{R}$ is transformed by $T\{(i,j)\}=\{(T_i,T_j)\}$, which may have $T_i=u_{i'}\in\mathcal{U}$ and $T_j=u_{j'}\in\mathcal{U}$ in which case $\{(T_i,T_j)\}^{\star}=\{(i',j')\}$. If $T_i=u_{i'}\in\mathcal{U}$ and $T_j\not\in\mathcal{U}$, then $\{(T_i,T_j)\}^{\star}=\{(i',0)\}$. If $T_i\not\in\mathcal{U}$ and $T_j=u_{j'}$ then $\{(T_i,T_j)\}^{\star}=\{(0,j')\}$. If $T_i,T_j\not\in\mathcal{U}$ then $\{(T_i,T_j)\}^{\star}$ will equal $\{(0,-1)\}$ or $\{(-1,0)\}$ depending on whether $T_j<T_i$ or $T_i<T_j$, respectively.
\end{example}
The above construction gives the following representation theorem.
\begin{thm}\label{thm:general}
Let $\ar$ be a signature, $\mathcal{R}\subseteq\mathop{\textsc{Fin}}\nolimits_{\ar}$, and fix an ordering $\mathcal{R}=(\mathcal{R}_n)_{n\geq1}$.
Let $\Xbf_{\cong}$ be a relationally exchangeable random structure in $\mathop{\mathcal{R}_{\mathbb{N}}^{\cong}}\nolimits$.
Then there exists a probability measure $\phi$ on $\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ such that $\Xbf_{\cong}\sim\varepsilon_{\phi}$, where
\begin{equation}\label{eq:epsilon-phi}\varepsilon_{\phi}(\cdot)=\int_{\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits}\varepsilon_f(\cdot)\phi(df).\end{equation}
\end{thm}
A canonical version of the measure $\phi$ in Theorem \ref{thm:general}
can be constructed as in \eqref{eq:canonical} above, and this is the measure we construct in the following proof.
\begin{proof}
We proceed by constructing $\phi$ as in \eqref{eq:canonical} and showing that $\Ybf_{\cong}^{\dagger}\sim\varepsilon_{\phi}$ satisfies $\Ybf_{\cong}^{\dagger}=_{\mathcal{D}}\Xbf_{\cong}$.
To see this, we first let $\psi$ be the de Finetti measure on $\mathcal{P}(\mathcal{R}_{[0,1]})$ (i.e., the space of probability measures on $\mathcal{R}_{[0,1]}$) associated to $\xi\Xbf_{\cong}$ for $\xi=(\xi_{i})_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ i.i.d.\ Uniform$[0,1]$ independently of $\Xbf_{\cong}$. In particular, the distribution of $\xi\Xbf_{\cong}$ is conditionally i.i.d.\ from $\nu\sim\psi$.
The measure $\psi$ determines a measure $\phi$ on $\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ through \eqref{eq:canonical}.
Given $f\sim\phi$, we construct $\Ybf=(Y_1,Y_2,\ldots)$ as conditionally i.i.d.\ from the distribution in \eqref{eq:f-induced} and $\Ybf^\dagger$ as in (i) and (ii) above. Let $\mathcal{U}=(u_i)_{i\geq1}$ be the atoms of $\theta\Ybf^{\dagger}$ for $\theta=(\theta_i)_{i\in\mathop{\mathbb{Z}_{}}\nolimits}$ i.i.d.\ Uniform$[0,1]$ independent of $\Ybf^{\dagger}$, let $\zeta=(\zeta_{k})_{k\leq0}$ be i.i.d.\ Uniform$[0,1]$ independently of everything else, and define $\xi_f=(\xi_i)_{i\in\mathbb{Z}}$ by
\[\xi_i=\left\{\begin{array}{cc} u_i,& i\geq1,\\ \zeta_i,&
\text{otherwise.}\end{array}\right.\]
First note that the conditional distribution of $\xi_f\Ybf^\dagger$
given $f$ is the same as the conditional law of ${\theta}\Ybf^{\dagger}$
given $\nu_f$, since the $\xi_i$ for $i\geq1$ were constructed from the atoms of $\theta\Ybf^{\dagger}_{\cong}$. Writing $\mathcal{D}(X\mid Y)$ to denote the conditional distribution of $X$ given $Y$, we thus have
\begin{align}
\mathcal{D}(\xi_f\Ybf_{\cong}^{\dagger}\mid f)&=\mathcal{D}(\theta\Ybf^{\dagger}\mid\nu_f)\label{eq:1}\quad\text{and}\\
\mathcal{D}(\xi_f\Ybf_{\cong}^{\dagger}\mid f)&=\mathcal{D}(\theta\Xbf_{\cong}\mid\nu_f),\label{eq:2}
\end{align}
where \eqref{eq:2} follows from the construction of $\phi$ from the de Finetti measure $\psi$ of $\xi\Xbf_{\cong}$.
From \eqref{eq:1} and \eqref{eq:2} it follows that
\begin{align*}
\mathcal{D}((((\xi_f\Ybf_{\cong}^{\dagger})^{\star})^{\dagger})_{\cong}\mid f)&=\mathcal{D}((((\theta\Ybf^{\dagger})^{\star})^{\dagger})_{\cong}\mid\nu_f)\quad\text{and}\\
\mathcal{D}((((\xi_f\Ybf^{\dagger}_{\cong})^\star)^{\dagger})_{\cong}\mid f)&=\mathcal{D}(((({\theta}
\Xbf_{\cong})^\star)^{\dagger})_{\cong}\mid \nu_f),
\end{align*}
and thus
\[\mathcal{D}((((\theta\Ybf^{\dagger})^{\star})^{\dagger})_{\cong}\mid\nu_f)=\mathcal{D}(((({\theta}
\Xbf_{\cong})^\star)^{\dagger})_{\cong}\mid \nu_f).\]
By Proposition \ref{prop:nu}, we have
$(((\theta\Ybf^{\dagger}_{\cong})^{\star})^\dagger)_{\cong}=\Ybf^{\dagger}_{\cong}$
a.s.\ and
$((({\theta}\Xbf_{\cong})^\star)^{\dagger})_{\cong}=\Xbf_{\cong}$,
implying $\Ybf^{\dagger}_{\cong}=_{\mathcal{D}}\Xbf_{\cong}$ as desired.
\end{proof}
\subsection{Special cases}
As discussed in Example \ref{example:partition}, the case in which $\mathcal{R}$ corresponds to the singleton sets $\{i\}$ for $i\geq1$ gives $\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ equal to the ranked simplex
\[\Delta^{\downarrow}:=\left\{(f_0,f_1,\ldots): f_1\geq f_2\geq\cdots\geq0,\ f_0\geq0,\ \sum_{i\geq0}f_i=1\right\}.\]
To see this note that $\mathcal{R}_{[0,1]}$ is the set of singleton elements of $[0,1]$, i.e.,
\[\mathcal{R}_{[0,1]}:=\{\{u\}: u\in[0,1]\}.\]
An exchangeable sequence $\Xbf=(X_1,X_2,\ldots)$ in $\mathcal{R}_{[0,1]}$ gives rise to a random countable subset $\mathcal{U}\subset[0,1]$ of elements that appear infinitely often among the $X_i$ and its complement $[0,1]\setminus\mathcal{U}$ consisting of all elements appearing at most once among the $X_i$. In the construction of $\Xbf^\dagger$ from $\Xbf$ described above, any occurrence $X_i=\{u\}$ for $u\notin\mathcal{U}$ gives rise to $X_i^{\star}=\{0\}$, explaining why the definition of $\mathop{\mathcal{F}_{\mathcal{R}}}\nolimits$ in \eqref{eq:FR} corresponds to the simplex of elements $(f_i)_{i\geq0}$.
Following Example \ref{ex:example}, the case of edge exchangeable random (directed) graphs corresponds to $\mathcal{R}\subset\mathop{\textsc{Fin}}\nolimits_{\ar}$ with $\ar(1)=2$ and $\ar(k)=0$ for $k\geq2$, with each $\mathcal{A}\in\mathcal{R}$ having $R_1^{\mathcal{A}}=\{(i,j)\}$, $i\neq j$. In this case, $\mathcal{R}_{[0,1]}$ corresponds to all pairs $(u,v)$ for $u,v\in[0,1]$. An exchangeable sequence $\Xbf$ in $\mathcal{R}_{[0,1]}$ determines a random subset $\mathcal{U}\subset[0,1]$. Now, in each $X_i$ there can be $0$, $1$, or $2$ elements $v\notin\mathcal{U}$. Occurrences of such elements are replaced by $0$ and $-1$, as explained in Example \ref{example:partition}.
In general, we call the occurrences of any $u\notin\mathcal{U}$ in the sequence $\Xbf$ {\em blips}. These elements are merely a `blip' in the overall sequence $\Xbf$---they occur only for an instant and then never again. Our labeling convention in defining $\mathcal{R}^\star$ is that the non-positive labels correspond to the blips. Theorem \ref{thm:general} describes how blips arise in general relationally exchangeable structures.
\section{Concluding}
Recent work in the statistical analysis of network data underscores the significance of relationally labeled structures in applications, as many data structures which are typically represented graphically, such as social networks and networks detailing email correspondence and professional collaborations, arise from a process by which interactions or relations accumulate within a population of otherwise indistinguishable individuals.
The work in \cite{CD2016e2} focused on the case of edge and hyperedge exchangeable random graphs, but the additional examples in Section \ref{section:introduction} involving repeated path sampling and snowball sampling are also highly relevant in networks applications.
The representation theorem serves two immediate statistical purposes.
First, the representation characterizes a general class of nonparametric statistical models of potential interest in the aforementioned applications.
Second, it establishes that vertices arrive in size-biased random order in relationally exchangeable structures, explaining why the common assumption of exchangeable vertex labeling, as presented in graphon models \cite{LovaszSzegedy2006}, is not tenable in many applications.
We reserve discussion of these practical implications for other work; see \cite{CD2016e2,CD2015AOS}.
|
\section{INTRODUCTION}
\label{sec:intro}
The direct imaging of exoplanets is more sensitive to planets beyond $\sim$5-10 AU. Although direct imaging has seen less planet detections than radial velocity or transit techniques, the past eight years have revealed a handful of directly imaged planets, including multiple planets around HR 8799 \cite{8799_1, 8799_2}, HD 95086 b\cite{95086}, Beta Pic b\cite{beta_pic}, and most recently 51 Eri b\cite{51eri}.
Detailed characterization of these existing planets with extremely large telescopes (ELTs) may be a difficult task, since these systems may lie either at the edge of or outside of the typical $\sim$0.4 arcsecond ELT adaptive optics (AO) Nyquist control region when observing in the near infrared\cite{nfiraos}. This region is set by the deformable mirror (DM) actuator pitch projected onto the telescope pupil, and for a square grid DM is a $(N_\text{act})(\lambda/D)\times(N_\text{act})(\lambda/D)$ region around the on-axis PSF, where $N_\text{act}$ is the number of actuators in width across the telescope pupil, $\lambda$ is the wavelength of light, and $D$ is the telescope diameter\cite{snwfc}. Thus, with classical, single conjugate AO (SCAO), uncorrected atmospheric turbulence and and quasi-static speckles will lower the planet signal to noise ratio (SNR).
A recent technique has recently been proposed to allow wavefront control outside the Nyquist control region, called super-Nyquist wavefront control (SNWFC)\cite{snwfc}. The main hardware component in this technique requires the use of a super-Nyquist element in an AO system, such as a mild pupil plane diffraction grating with a spacing between lines that is smaller than the DM actuator pitch relative to the pupil size. The pupil plane imprint creates a PSF copy in the focal plane that is outside the DM Nyquist region, allowing wavefront control to work in a similar $(N_\text{act})(\lambda/D)\times(N_\text{act})(\lambda/D)$ control region around this super-Nyquist PSF copy.
In this paper, we present the results of a laboratory experiment and simulations for a future experiment to show that it is possible to use SNWFC on an ELT AO system system to directly image already known and new exoplanets. In \S\ref{lab} we describe the laboratory experiment design (\S\ref{lab_design}), simulations of expected lab performance (\S\ref{lab_sims}), and results in the lab (\S\ref{lab_results}). In \S\ref{scc} we describe the setup and results of our simulation using the self coherent camera (SCC)\cite{scc1,scc2}. In \S\ref{conclusion}, we summarize our results and discuss future work.
\section{DETERMINISTIC LABORATORY EXPERIMENT}
\label{lab}
We first design, simulate, and test a simple deterministic laboratory experiment that demonstrates the possible performance improvement using SNWFC. The goal here is to show that performance improvement is possible in an idealized case to motivate more realistic SNWFC simulations (\S \ref{scc}), the future testing of this technique in a laboratory, and the future ELT applications for high contrast imaging of exoplanets.
\subsection{Experiment Design}
\label{lab_design}
In this section we describe the methodology and algorithm setup for our deterministic speckle nulling experiment. Figure \ref{fig: experiment_diagram} shows a schematic layout for our SNWFC experiment.
\begin{figure} [h]
\begin{center}
\begin{tabular}{c}
\includegraphics[height=7cm]{experiment_diagram.pdf}
\end{tabular}
\end{center}
\caption{A schematic diagram of the focal plane for our deterministic SNWFC laboratory experiment. \enquote{$\alpha$} represents the sine wave position angle, counter clockwise from the +$x$ axis. \enquote{cpa} is the number of cycles per aperture for a given DM sine wave, and it represents radial location of the sine wave PSF copy in the focal plane in units of $\lambda/D$ (the DM sine wave angular frequency is $f=2\pi(\text{cpa})/(D_\text{DM})$,where $D_\text{DM}$ is the DM pupil plane diameter). }
\label{fig: experiment_diagram}
\end{figure}
The basic structure involves the placement of three sine waves on the DM:
\begin{itemize}
\item \textbf{spot}: Shown in blue in Figure \ref{fig: experiment_diagram}, the spot represents the effect of a PSF copy from a pupil plane sine wave phase plate at a higher frequency than the DM Nyquist frequency, thus creating a copy of the on-axis PSF outside the AO control region. In the absence of a phase plate for our experiment, we use the highest DM sine wave frequency to represent the spot. We note that with this design there is no actual super-Nyquist element, since the spot is, by definition, sub-Nyquist. However, one can instead imagine a system with a lower order DM where the DM Nyquist frequency creates the speckle, in which case a super-Nyquist phase plate creates the spot. In this case, SNWFC is needed to null the speckle using the spot. But, the super-Nyquist phase plate for a lower order DM and our real DM spot sine wave have the same optical effect, and so for the purposes of this experiment we are still demonstrating SNWFC, but instead using less than the full DM to represent the Nyquist region.
\item \textbf{speckle}: Shown in purple in Figure \ref{fig: experiment_diagram}, the speckle is at a slightly lower frequency than the spot, again sub-Nyquist in this experiment, and is meant to represent speckle noise hiding the signal from a planet at the \textbf{point of interest} (Figure \ref{fig: experiment_diagram}).
\item \textbf{anti-speckle}: Shown in red in Figure \ref{fig: experiment_diagram}, the goal of this experiment is to copy the anti-speckle sine wave from the spot so that electric field in the focal plane at the point of interest is minimized, allowing the off axis planet light at that point to then be seen. Unlike the speckle intensity, the planet light is not removed because it is incoherent with the stellar light.
\end{itemize}
The frequency, position angle, amplitude, and phase of the spot and speckle are user-defined parameters. In order to cancel the speckle at the point of interest, the anti-speckle frequency and position angle are already determined geometrically based on the user-defined spot and speckle parameters. The phase and amplitude of the anti-speckle are then determined iteratively using the following methodology and procedure:
\begin{enumerate}
\item Using the image plane intensity, which is approximately a measure of amplitude squared, the anti-speckle amplitude is then approximately
\begin{equation}
a_{\text{anti-speckle}}=(f) a_{\text{speckle}}\sqrt{\frac{\text{im}_{\text{spot}}}{\text{im}_{\text{speckle}}}}
\label{eq: fudge}
\end{equation}
in an image, im, with only two sine waves, the spot and speckle, where $a$ is the sine wave amplitude. The fudge parameter, $f$ is equal to 1 in this step and explained further below in step 3.
\item The anti-speckle phase is unknown, so we simply loop through -$\pi$ to $\pi $ in phase on the anti-speckle in order to find the phase that minimizes image plane intensity at the point of interest.
\item Using $f=1$ in equation \ref{eq: fudge} is only a rough estimate for $a_\text{anti-speckle}$ because we are using a linear approximation of the wavefront ($\text{wavefront}=a\; e^{i\; \phi}\sim a(1+i\phi)$, so $|\text{wavefront}|^2\sim O(a^2)$) to null the speckle, neglecting the higher order terms. So, to correct for these higher order terms we add a fudge factor as necessary in the anti-speckle amplitude, where $0.5\lesssim f\lesssim$2.
\end{enumerate}
In simulations (\S\ref{lab_sims}) and in the lab (\S\ref{lab_results}), finding a precise optimal anti-speckle phase is done iteratively, first finding a rough estimate, and then using a finer grid spacing in phase around that rough estimate to get a more precise value. The same iterative procedure is applied in finding the amplitude fudge factor. Throughout the remainder of \S\ref{lab_design}, we calculate contrast in a given image normalized to the peak value in that same image, since we do not use a coronagraph in this laboratory experiment.
\subsection{Simulations}
\label{lab_sims}
We ran simulations of two different sine wave configurations: one where the point of interest lies on a bright Airy ring of the on-axis PSF, and one where the point of interest lies on a nearby dark ring. The initial and final parameters for both simulations are shown in Table \ref{tab: parameters}. We use an image size of 2048$\times$2048 pixels, beam ratio of 4, and a circular greyscale pupil (instead of a binary mask, the pupil mask is mean binned 10$\times$10 from an original binary mask of image size of 20480$\times$20480 to prevent numerical pixelation effects).
\begin{table}[ht]
\caption{Initial simulation and final parameters for our three sine wave deterministic speckle nulling procedure for both our dark ring and Airy ring simulations. The spot and speckle parameters are set initially, while the anti-speckle phase and amplitude are determined iteratively via the focal plane wavefront sensing scheme described in \S\ref{lab_design}.}
\label{tab: parameters}
\begin{center}
\begin{tabular}{|l|l|l|l|l|}
\hline
\rule[-1ex]{0pt}{3.5ex} \textbf{sine wave} & \textbf{$\alpha$ (degrees)} & \textbf{cpa} & \textbf{phase (radians)} & \textbf{amplitude (radians)} \\
\hline
\multicolumn{5}{|c|}{Airy ring simulation} \\
\hline
\rule[-1ex]{0pt}{3.5ex} spot & 1.0 & 4.0 & 0 & 0.13 \\
\hline
\rule[-1ex]{0pt}{3.5ex} speckle & 28.7 & 2.95 & 0 & 0.064 \\
\hline
\rule[-1ex]{0pt}{3.5ex} anti-speckle & 135.0 & 2.0 & -2.57 & 0.94 \\
\hline
\multicolumn{5}{|c|}{dark ring simulation} \\
\hline
\rule[-1ex]{0pt}{3.5ex} spot & 1.0 & 4.0 & 0 & 0.13 \\
\hline
\rule[-1ex]{0pt}{3.5ex} speckle & 35.0 & 2.57 & 0 & 0.064 \\
\hline
\rule[-1ex]{0pt}{3.5ex} anti-speckle & 142.0 & 2.4 & 0.84 & 0.85 \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[!h]
\begin{center}
\includegraphics[width=1.0\textwidth]{sim_bench_images}
\end{center}
\caption{Images of our Airy ring (first row) and dark ring (second row) deterministic speckle nulling simulations. The left column shows the initial configuration with only two sine waves (spot+speckle), and the right column shows the configuration with all three sine waves after reaching the minimum anti-speckle phase and amplitude.}
\label{fig: sim_images}
\end{figure}
The results of our simulations after anti-speckle phase and amplitude iterations are shown in Figure \ref{fig: sim_images}. Figure \ref{fig: a} shows the results after only phase iterations, to be compared later with our laboratory results in \S\ref{lab_results}. Our two simulations show that changing the point of interest from a bright Airy ring to a nearby dark ring changes the achievable contrast as well as the absolute anti-speckle phase. The final contrasts reached after anti-speckle amplitude iterations (equation \ref{eq: fudge}) for the Airy ring and dark ring simulations are $1.1\times10^{-5}$ and $2.8\times10^{-7}$, respectively. After amplitude iterations for the Airy ring simulation we found $f=1.786$, whereas for the dark ring simulation we found $f=1.004$. The fudge factor results are expected, since higher order terms in the Taylor expansion of the PSF fall off more quickly in the dark ring compared to the Airy ring\cite{perrin_psf}.
\subsection{Laboratory Results}
\label{lab_results}
We use a simple SCAO setup in the Adaptive Optics Laboratory at the National Research Council, Astronomy and Astrophysics (NRC), shown in Figure \ref{fig: bench_pic}. We use a monochromatic 655 nm fiber-fed laser diode as a light source, an $11\times11$ actuator ALPAO DM (circular pupil, 97 total actuators), a $32\times40$ subaperture Imagine Optic HASO Shack Hartman wavefront sensor (SHWFS), ultimately using a 29 subaperture diameter to remove edge effects.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=1.0\textwidth]{bench}
\end{center}
\caption{The NRC Adaptive Optics Laboratory wavefront sensor bench. The common path is in blue, the SHWFS path is in purple, the non-common path to the science camera (SC) is in red, and a separate pyramid WFS path (not fully shown) is in green.}
\label{fig: bench_pic}
\end{figure}
We apply sine waves to the DM in a closed loop procedure, using the expected SHWFS slopes for the desired sine wave amplitude, frequency, and phase, described below. The $x$ and $y$ SHWFS slopes for a DM sine wave at a subaperture, $i$ (using 1 indexing), are the mean values of the derivative of the wavefront phase along $x$ and $y$, respectively:
\begin{align}
\label{eq: shwfs_slopes}
x_\text{slope}(i)&=\left\langle \frac{\partial \Phi }{\partial x}\right\rangle_i \nonumber \\
&= \frac{\int _{x_i-d/2}^{x_i+d/2}\int _{y_i-d/2}^{y_i+d/2}\frac{\partial \Phi }{\partial x}d x d y}{\int _{x_i-d/2}^{x_i+d/2}\int _{y_i-d/2}^{y_i+d/2}d
x d y}, \\
y_\text{slope}(i)&=\left\langle \frac{\partial \Phi }{\partial y}\right\rangle_i \nonumber \\
&= \frac{\int _{x_i-d/2}^{x_i+d/2}\int _{y_i-d/2}^{y_i+d/2}\frac{\partial \Phi }{\partial y}d x d y}{\int _{x_i-d/2}^{x_i+d/2}\int _{y_i-d/2}^{y_i+d/2}dx d y}, \nonumber
\end{align}
where $\Phi$ is the wavefront phase, $d$ is the size of one subaperture ($d=D_\text{pup}/n_\text{subapertures}$, where $D_\text{pup}$ and $n_\text{subapertures}$ are the diameter and number of subapertures across the SHWFS pupil, respectively), and $(x_i,y_i)$ is the center of subaperture $i$ in units of pupil diameter\footnote{Using this coordinate system, the lower left corner of the lower left subaperture in use is $(x_1,y_1)=(0,0)$ and the upper right corner of the upper right subaperture is $(x_{n_\text{subapertures}},y_{n_\text{subapertures}})=(D_\text{pup},D_\text{pup})$.}. After calibrating a flat wavefront with the reference slopes from closing the loop on zeros, a sine wave on the DM imparts a wavefront phase shift of
\begin{equation}
\Phi= a\; \text{sin}\left(\frac{2\pi}{D_\text{pup}}\text{cpa}(x\; \text{cos}(\alpha)-y\; \text{sin}(\alpha))-\phi\right),
\label{eq: sine}
\end{equation}
where $a,\; \alpha,\; \text{and } \phi$ are the sine wave the amplitude, position angle, and phase, respectively. Combining equations \ref{eq: shwfs_slopes} and \ref{eq: sine} allows us to apply a sine wave on the DM in closed loop at any amplitude, frequency (sub-Nyquist), position angle, and phase.
\begin{figure} [!h]
\begin{center}
\includegraphics[height=8cm]{bench_images.pdf}
\end{center}
\caption{ \label{fig: bench}
Laboratory science camera images from our deterministic SNWFC experiment, using the Airy ring simulation parameters from Table \ref{tab: parameters}. The left image shows the initial speckle contrast for the spot+speckle DM sine wave combination, and the right image shows the spot+speckle+anti-speckle combination at the minimum anti-speckle phase of $\phi=-0.12$. The images are background/dark-subtracted, hence the negative contrast scale.}
\end{figure}
\begin{figure} [!h]
\begin{center}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1.0\textwidth]{phase_sim_iterations}
\caption{}
\label{fig: a}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1.0\textwidth]{bench_phase_iterations}
\caption{}
\label{fig: b}
\end{subfigure}
\end{center}
\caption{Speckle contrast plotted vs. anti-speckle phase (radians) at the point of interest. (a) Noiseless simulations for our SNWFC laboratory experiment, showing both the replicated laboratory setup, where the speckle location lies on an Airy ring, and a second simulation showing the effect of moving the speckle location into a nearby dark ring. (b) Results from our laboratory data. The minimum contrast and phase are both in between our two simulations in (a), suggesting that noise on the bench is moving the speckle location closer to a dark ring from the original Airy ring location.}
\label{fig: experiment_sims}
\end{figure}
Using this methodology and the same parameters as our Airy ring simulation in Table \ref{tab: parameters} (but not the anti-speckle phase and amplitude, which we again determine iteratively), we implemented the deterministic speckle nulling procedure described in \ref{lab_design}, but after two phase iterations we reached the SC detector noise floor at the point of interest, and so we did not do any amplitude iterations. The before and after images comparing the spot+speckle with the minimum spot+speckle+anti-speckle are shown in Figure \ref{fig: bench}, and our laboratory phase iterations are shown in Figure \ref{fig: b}.
We found a final contrast at the point of interest of $5.6\times10^{-4}$, using $a_\text{anti-speckle}=88$ nm and $\phi_\text{anti-speckle}=-0.12$ (from equations \ref{eq: fudge} and \ref{eq: sine}, respectively).
Comparing the final anti-speckle phase and contrast from our lab results in Figure \ref{fig: b} to our simulation results in Figure \ref{fig: a} suggests that in the lab, the speckle lies somewhere in between a bright and dark ring. Despite our laboratory setup identical to the Airy ring simulation, we suspect that noise on the bench distorts the on-axis Airy ring locations so that the speckle is located near a dark ring, suggested by the deeper contrast and closer absolute phase shift for the dark ring simulation in Figure \ref{fig: a}.
\section{SIMULATIONS USING THE SELF-COHERENT CAMERA}
\label{scc}
Although our deterministic speckle nulling procedure in \S\ref{lab} was designed to show that SNWFC is possible experimentally, we would like in the future to further test SNWFC by creating a more realistic dark hole as opposed to nulling a single speckle. In the absence of infrastructure for more stable lab conditions, a higher order DM, and a super-Nyquist phase plate, we instead run simulations to represent the possible performance on a future ELT high contrast imaging instrument.
The SCC, first developed by Baudoz et al. (2006) and evolving most recently to the design and algorithm in Mazoyer et al. (2014), is a wavefront control method to remove speckles in the focal plane\cite{scc1,scc2}. Use of an off-axis hole placed in the Lyot stop causes fringes in the focal plane, but only from the stellar light. Isolating these fringed speckles is done using an amplitude mask on the complex-valued modulus transfer function (MTF), yielding a complex-valued image, $I^-$, which contains phase and amplitude information of only the fringed stellar speckles. By constructing a set reference images in $I^-$ (i.e., an interaction matrix) composed of sines and cosines (each with zero phase) at every $\lambda/D$ interval within a symmetric half-Nyquist region (offset by $0.5\lambda/D$ to preserve symmetry), a least squares fit to an aberrated target image (e.g., with phase and amplitude errors in a sub-Nyquist half-DH region) yields the negative sine and cosine amplitude coefficients to apply to the DM such that the fringed speckles in the focal plane are minimized. See Mazoyer et al. (2014) and references therein for a more detailed description of the SCC wavefront control algorithm.
We choose to use the SCC wavefront control method instead of, e.g., electric field conjugation (EFC)\cite{efc1}, because of the algorithm timescale to converge. It is shown experimentally in in Mazoyer et al. (2014) that the SCC algorithm converges immediately after one iteration\footnote{We note that Mazoyer et al. (2014) do apply a second iteration after the first dark hole is achieved by recalibrating a new interaction matrix, offset with the coefficients already obtained from the initial least squares. We do not perform this second calibration in our simulations, as this is likely done to correct for laboratory limitations that we do not consider in our simulations, such as DM hysteresis, influence functions, fitting error, etc.}, and only relies on the time required to compute the target image correlation vector (i.e., after the initial daytime calibration has been done to compute and invert the interaction matrix). In contrast, EFC, even algorithmically, takes tens of iterations to converge, as in Give'On et al. (2007). With the already calibrated interaction matrix, the SCC timescale to compute the target image correlation vector, sine and cosine amplitude coefficients, and resulting corrected image is $\sim$30 seconds on our 2.7 GHz Mac desktop using one core, but this could ultimately run much faster with a fully parallel code.
The main component of super-Nyquist wavefront control using the SCC, or with any other wavefront control method, requires an optical element that creates a PSF copy outside of the DM Nyquist region. In Thomas et al. (2015), they consider the use of (1) influence functions that are narrow relative to the DM actuator pitch, such as the known manufacturing residual print-through pattern on many DMs, and (2) a mild transmissive diffraction grating (i.e., at a super-Nyquist frequency), and perform SNWFC simulations using option (1). Here, we consider the use of a super-Nyquist sine wave phase plate, initially proposed in Marois et al. (2006) and Sivaramakrishnan et al. (2006) for astrometric and photometric calibration purposes \cite{marois06, sivaramakrishnan}.
We use a 60 nm amplitude sine wave phase plate to create the SCC reference array described above and in Mazoyer et al. (2014), thus copying every sine and cosine from the sub-Nyquist region to the analogous super-Nyquist region. The construction of these reference images and the corresponding least-squares covariance matrix, was done without noise, using 21 nm amplitude sine waves \cite{loci}. This interaction matrix procedure could be implemented on an instrument during daytime calibrations, provided the conditions are stable between daytime calibration and night time observations.
\begin{figure} [!h]
\begin{center}
\includegraphics[height=8cm]{scc}
\end{center}
\caption{ \label{fig: scc}
Before (left) and after (right) SCC SNWFC simulations, using 40 nm rms wavefront error, 1\% amplitude error, 60 nm amplitude super-Nyquist phase plate (using $\text{cpa}=1.1*N_\text{act}$), a 32$\times$32 DM, and an APLC. The contrast improves by a factor of $\sim$5 after dark hole correction.}
\end{figure}
Similar to Marois et al. (2012), our target image contained a 40 nm rms phase screen wavefront error and 1\% amplitude errors, both with a -2 power law in spatial frequency\cite{christian_nfiraos}. To simulate the effect of an AO system, we removed the first 21 Zernike polynomials within the DM Nyquist region, using a least squares on the target image with a set of Zernike reference images. We used a 32 by 32 actuator DM, beam ratio of 4 pixels, image plane dimensions of 1012 by 1012, wavelength of 1.65 $\mu$m, and sine wave phase plate cpa of $1.1\left(N_\text{act}\right)$, placing the super-Nyquist region just beyond the sub-Nyquist region. We used the GPI aperture and apodized Lyot coronagraph (APLC) design \cite{remi}. To determine the separation and diameter of the off-axis hole in the Lyot stop that makes the SCC, we used the equations from Mazoyer et al. (2014) ($\epsilon_0>1.5\;D_\text{pupil}$, and $D_R<(1.22\sqrt{2}/N_\text{actuators})D_\text{pupil}$, respectively)\cite{scc2}.
Our simulation results are shown in Figure \ref{fig: scc}. The DH contrast in the raw image is $7.3\times10^{-7}$, and in the SCC image it goes down to $1.6\times10^{-7}$, a gain of $\sim$5 in contrast. We chose a half DH region left of the on-axis PSF, on the left side of the super-Nyquist region, but more generally the phase plate position angle and choice of which side of the spot to build a half DH can be set by the user, if, e.g., the planet location is already known. Images are converted to units of contrast by normalizing the values in the coronagraphic image to the peak value in the initial non-coronagraphic image. To calculate contrast in the super-Nyquist DH-region we take the robust standard deviation\cite{robust} in the contrast-normalized image within an aperture that removes $1\; \lambda/D$ off each edge of the half-DH (to remove edge effects) and $5\;\lambda/D$ radially around the super-Nyquist spot (to remove brighter diffraction effects from the spot that are not present in the raw image without a spot).
\section{CONCLUSION}
SNWFC allows already known and new exoplanets to be directly imaged at a higher SNR than is otherwise possible with the current AO design for ELTs. Our main conclusions in testing this technique are as follows:
\begin{itemize}
\item We have demonstrated in the lab that a deterministic SNWFC speckle nulling scheme increases performance, consistent with simulations.
\item We demonstrated in simulation that SNWFC performance improvement is possible using the SCC after only one iteration, which is of particular interest to wavefront control on shorter timescales for ground-based telescopes.
\end{itemize}
There is still a lot of unexplored parameter space to get a better understanding of how well SCC-based SNWFC would perform on a telescope. Some additional factors in our simulations that are beyond the scope of this initial exploratory paper but worth further research are performance in polychromatic light, using a segmented pupil, using different coronagraphs and/or apodization, dependence on higher super-Nyquist phase plate frequencies, dependence on using an inner and outer scale in chosen wavefront error power law, dependence on a higher zero pad sampling for the off-axis SCC Lyot hole, and additional improvement based on diffraction suppression techniques\cite{efc_l}. Ultimately, the next step before this technique can be tested on-sky is to demonstrate it in the lab.
\label{conclusion}
\acknowledgments
We gratefully acknowledge research support of the Natural Sciences and Engineering Council (NSERC) of Canada.
|
\section{Introduction}
\label{sec:introduction}
Perfectly matched layers \citep{Berenger:1994ua} are a well-developed method for simulating wave propagation in unbounded media enabling the use of a reduced computational domain without having to worry about spurious boundary reflections. B\'{e}renger showed that by adding specific conductivity parameters to Maxwell's equations, perfect matching and decaying of the propagating waves in the PML could be achieved \citep{Berenger:1994ua}. An alternative method is to assume that the material contained within the PML is a uniaxial anisotropic media \citep{Sacks:1995gs,Roden:1997fo,Gedney:1996ub}, generally referred to as the uniaxial PML approach. A third method, with greater generality and flexibility, is the complex coordinate stretching approach \citep{Chew:1994dn}. In fact, the conductivity parameter introduced by B\'{e}renger \citep{Berenger:1994ua} can be thought of as a damping parameter in a stretch function that transforms the spatial coordinate in the layer to the complex plane.
Subsequent to these electromagnetic wave applications, many PML formulations have been introduced for elastic wave propagation \citep{Collino:2001vt,Hastings:1996um,Drossaert:2007fi,Chew:1996wk,Appelo:2006wz,MezaFajardo:2008dx,Kucukcoban:2011tn}. Amongst these, the split-field formulations that are typically described by systems of first order equations with double (for 2D) or triple (for 3D) the number of stress-velocity physical equations (nine equations in all for the 3D case) . Second-order formulations uses one physical field variable (usually the displacement) along with extra auxiliary variables that are typically needed to obtain the time-domain equations from the frequency-domain equations. Mathematically, it has been proven that certain second-order PMLs are strongly well-posed, while the first-order type is only weakly well-posed \citep{Deinega:2011er,Abarbanel:2002wm,Duru:2012va}.
There are other advantages for choosing second-order formulations. The second-order displacement elastic wave equation emerges directly from Newton's second law \citep {Duru:2012va}, unlike the fist-order stress-velocity elastic wave equation which introduces a new non-physical wave mode with zero velocity \citep{Duru:2012va, Appelo:2006wz}. Moreover, the second-order PML formulations are more readily implemented in common numerical schemes that are based on second-order displacement wave equations \citep{Komatitsch:2007bz, Li:2010tu}. However, deriving second-order PML formulations is less trivial than that for first-order ones, especially in the time-domain where many auxiliary variables are needed. The problem becomes more complex for the 3D modeling which would partially explain the dearth of second order formulations in 3D.
There have been a number of papers that describe the formulation of time-domain wave propagation in 3D fluid media using PMLs (see of example \citep{Hu:2007jb,kaltenbacher2013modified}) but there are relatively few that address the same problem for anisotropic, inhomogeneous elastic media, especially those attempting a second-order formulation. In previous works \citep{Assi:2013tp, Assi:2015tv}, the authors introduced a compact second-order time-domain PML formulations for the elastic wave equation in 2D which has only four auxiliary variable. Recently Lee and Shin \citep{Koria:2015th} introduced an unsplit PML formulation for isotropic media or media with vertical axis of symmetry (VTI). Their 2D derivation was based on second-order elastic wave equations, and the final formulation followed closely the one given in Assi and Cobbold \citep{Assi:2013tp, Assi:2015tv}. It should be noted that the final form of the PML formulation, and not the way it wavs derived, that governs its robustness and other characteristics. Additionally, Lee and Shin extended their formulation to 3D VTI media for which they presented numerical results \citep{Koria:2015th}.
The purpose of this paper is to derive a time-domain second-order formulation to model elastic wave propagation in an unbounded three-dimensional general anisotropic inhomogeneous solid. As will be seen the formulation results in a system of equations that are applicable throughout the computational domain. In the physical domain, the complex stretch function is simply set to unity. To demonstrate the application of our formulation, propagation from a spherical transient source embedded in a highly anisotropic medium (the mineral Olivine) is illustrated. Extension of the formulation to include a viscoelastic medium that can be represented by a Kelvin-Vogit model \citep{Meyers:2008,Banks:2011gj}, is presented in \autoref{app:viscoelasticity}.
\section{Background and materials}
\label{sec:background}
\subsection{Elastic waves in solids}
\label{subsec:theoElastic}
Wave propagation in linear elastic solids can be described using Newton's second law, along with Hook's law and a linear approximation for the strain. These lead to the following second-order formulation of the elastic wave equation:
\begin{equation}
\rho\frac{\partial^{2}u_{i}}{\partial t^{2}}-\sum\limits _{j=1}^{3}\frac{\partial}{\partial x_{j}}\left(\sum\limits _{k,l=1}^{3}C_{ijkl}\frac{\partial u_k}{\partial x_l}\right)=0,
\label{eq:waveElastic}
\end{equation}
where $t\in\mathbb{R}^{+} $ is time, $\mathbf{x}\in\mathbb{R}^{3}$ is the space variable, $u_i(\mathbf{x},t)$ are the components of particle displacement vector. Moreover, $\rho(\mathbf{x})$ is the solid mass density and $C_{ijkl}(\mathbf{x})$ are the components of the fourth order elasticity tensor with the following symmetry properties: $C_{ijkl}=C_{ijlk}=C_{jikl}$, and $C_{ijkl}=C_{klij}$. The source of energy that excites the elastic medium can be added as a load vector, $F_i(\mathbf{x},t)$, to the right-hand side (RHS) of \eqref{eq:waveElastic}.
In general, the elasticity tensor, $C_{ijkl}$, has 81 components, but due to the above symmetries, the maximum number of independent parameters is 21. For the special case of isotropic solids, the elasticity tensor can be described by two independent parameters such as the Lam\'{e} coefficients, $\lambda(\mathbf{x})$ and $\mu(\mathbf{x})$. In terms of these two coefficients, the elasticity tensor can be written as:
\begin{equation}
C_{ijkl}=\lambda\delta_{ij}\delta_{kl}+\mu(\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}),
\label{eq:HookIsotropic}
\end{equation}
where $\delta_{ij}$is the Kronecker delta function.
For the purpose of plane wave and Fourier analyses, the harmonic wave solutions of the following form:
\begin{equation}
\mathbf{u}=\mathbf{A}\exp{\left[i(\mathbf{k\cdot x}-\omega t)\right]},
\label{eq:harmonicWave}
\end{equation}
will be considered for the elastic wave equation as given by \eqref{eq:waveElastic}. In this equation, $\mathrm{\mathbf{A}} \in \mathbb{C}^{3}$
is the constant amplitude polarization vector, $\mathrm{\mathbf{k}}\in\mathbb{R}^{3}$ is the wavevector, $\omega\in\mathbb{C}$ is the angular frequency, and $i^2=-1$.
\subsection{Complex coordinates stretching}
\label{subsec:complex}
\sloppy{To obtain a PML formulation for a given wave equation, the complex coordinate stretching \citep{Chew:1994dn} can expressed as a coordinate transform: ${\mathbf{x}\mapsto\tilde{\mathbf{x}}:\mathbb{R}^3\to\mathbb{C}^3}$.} Since $\tilde{\mathbf{x}}=\mathbf{x}$ in the physical domain and the PML region is assumed to be homogeneous, then $\tilde{\mathbf{x}}$ appears only in the form of spatial partial derivatives in the PDEs. Given a field variable $u$, then using the chain rule: $\frac{\partial u} {\partial x_j}= \sum_{k=1}^{3} \frac{\partial u}{\partial \tilde{x}_k} \;\frac{\partial \tilde{x}_k}{\partial x_j}$, which reduces to $\frac{\partial u} {\partial x_j}= \frac{\partial u}{\partial \tilde{x}_j} \;\frac{\partial \tilde{x}_j}{\partial x_j}$ since $\tilde{x}_j$ depends only on $x_j$. As a result, defining the complex stretch function by $s_{j}\left(x_{j}\right)=\frac{\partial\tilde{x}_{j}}{\partial x_{j}}$ suffices to perform transformation:
\begin{equation}
\frac{\partial}{{\partial\tilde{x}}_{j}}=\frac{1}{s_{j}\left(x_{j}\right)}\frac{\partial}{{\partial x}_{j}}.
\label{eq:derivativeTransform}
\end{equation}
The two-parameter complex stretch function introduced by Fang and Wu \citep{Fang:1996tc} in their generalized PML (GPML) is adopted in this paper. This function is given by
\begin{equation}
s_{j}\left(x_{j}\right)=\alpha_{j}\left(x_{j}\right)\left[1+i\,\frac{\beta_{j}\left(x_{j}\right)}{\omega}\right],
\label{eq:CSF}
\end{equation}
where the $\beta_{j}\geq0$ is the damping parameter responsible for damping the propagating wave inside the PML. In this equation, the scaling parameter, $\alpha_{j}>0$, is responsible for either stretching ($\alpha_{j}>1$) or compressing ($0<\alpha_j<1$) the coordinate. It should be noted that in the physical domain, where $\tilde{x}_{j}\left(x_{j}\right)=x_{j}$, $\beta_{j}=0$ and $\alpha_{j}=1$.
Appropriate choices are now needed for the stretch function parameters $\alpha_{j}(x_{j})$ and $\beta_{j}(x_{j})$. Despite the absence of a rigorous methodology for their choice \citep{Chew:1996wk,Kucukcoban:2011tn}, polynomial functions are often used as indicated below for the damping parameter:
\begin{equation}
\beta_{j}\left(x_{j}\right) =\begin{cases}
0 & \text{if}\left|x_{j}\right|<x_{0}\\
\beta_{0_{j}}\left(\dfrac{\left|x_{j}\right|-x_{0}}{d}\right)^{n} & \text{if }x_{0}\leq\left|x_{j}\right|\leq x_{0}+d,
\end{cases}
\label{eq:beta}
\end{equation}
$\!\!$where \textit{d} is the thickness of the PML, ${2x}_{0}$ is the dimension of the square physical domain centered at the origin, $n$ is the polynomial order, $\beta_{0_{j}}$ is a constant that represent the maximum values of $\beta_{j}$. The value of this parameter needs to be specified. It is helpful to express the value of $\beta_{0_{j}}$ in terms of a desired amplitude reflection coefficient ($R_{j})$ caused by the reflection from the outer boundary of the PML. It can be shown that for normal incidence and assuming $\alpha_{j}=1$,
\begin{equation}h
\beta_{0_{j}}=\frac{c_{\mathrm{max}}\left(n+1\right)\ln\left(1/R_{j}\right)}{2d}.
\label{eq:reflection}
\end{equation}
Quadratic polynomial, corresponds to $n=2$, will be used in this work unless mentioned otherwise.
Without loss of generality, the scaling parameter, $\alpha_{j}$, is set to unity in this work. This parameter can be readily introduced back to any PML formulation that is derived using the complex stretch function in \eqref{eq:CSF}, by replacing each $\partial/\partial x_j$ by $\partial/ \alpha_j\partial x_j$ in the PDEs.
\section{Formulation of PML for elastic wave propagation}
\label{sec:formulation}
With the help of the above background, our time-domain PML formulation can be introduced for the wave propagation in unbounded media. All parameters, namely, $\rho$, $C_{ijkl}$, $\beta_{j}$, and $s_{j}$, are assumed to be space dependent throughout the derivation leading to a variable-coefficient PML formulation. Since the stretch function also depends on the frequency, the derivation starts in the frequency domain.
First, we take Fourier transforms of the elastic wave equation \eqref{eq:waveElastic}, and then transform the spatial coordinates using complex stretching, ${\mathbf{x}\mapsto\tilde{\mathbf{x}}}$, as introduced in \autoref{subsec:complex}. These steps lead to:
\begin{equation}
\left(-i\omega\right)^{2}\hat{u}_i\,\rho=\sum\limits _{j=1}^{3}\frac{\partial}{\partial \tilde{x}_{j}}\left(\sum\limits _{k,l=1}^{3}C_{ijkl}\frac{\partial \hat{u}_k}{\partial \tilde{x}_l}\right),
\label{eq:derevationElastic1}
\end{equation}
where $\hat{\Box}$ denotes the Fourier transform in time. The need to solve this differential equation along contours in the complex plane can be avoided by inverse transforming the complex-stretched coordinates back to the original spatial coordinates using \eqref{eq:derivativeTransform}. This is followed by multiplying the equation by $s_1\, s_2 \, s_3$, leading to:
\begin{equation}
s_1s_2s_3\,\left(-i\omega\right)^{2}\hat{u}_i\,\rho=\sum\limits _{j=1}^{3}\frac{s_1s_2s_3}{s_j}\frac{\partial}{\partial x_{j}}\left(\sum\limits _{k,l=1}^{3}C_{ijkl}\frac{1}{s_l}\frac{\partial \hat{u}_k}{\partial x_l}\right).
\label{eq:derevationElastic2}
\end{equation}
Expanding $s_1s_2s_3$ according to \eqref{eq:CSF} while assuming $\alpha_j=1$, the left-hand side (LHS) of the above equation becomes:
\begin{equation}
\rho\,\left[ \left(-i\omega\right)^{2}+(-i\omega) \left(\beta_1+\beta_2+\beta_3\right)+ \left(\beta_1\,\beta_2+\beta_2\,\beta_3+\beta_3\,\beta_1 \right)+\frac{\beta_1\beta_2\beta_3}{(-i\omega)} \right] \,\hat{u}_i.
\label{eq:derevationElastic31}
\end{equation}
Here, it is helpful to introduce the variable
\begin{equation}
U_i(\mathbf{x},t)=\int_{0}^{t}u_i(\mathbf{x},\tau)\, d\tau,
\label{eq:derevationHistory}
\end{equation}
whose Fourier transform is given by $\hat{U}_j\mathbf{\mathrm{(}x},\omega)=\hat{u}_i(\mathbf{x},\omega)/(-i\omega)+\pi\,\hat{u}_i(\mathbf{x},0)\delta(\omega)$. However, the second term vanishes since the stretch function \eqref{eq:CSF} is not defined for the static case of $\omega=0$ \citep{Kucukcoban:2011tn}. Consequently, substituting $\hat{u}_{i}=-i\omega\,\hat{U}_{i}$ in the last term of \eqref{eq:derevationElastic31} and taking inverse Fourier transform $(-i\omega\Rightarrow\thinspace\partial/\partial t)$ of this, results in
\begin{equation}
\rho\,\left[ \frac{\partial^{2}u_i}{\partial t^2}+\left(\beta_1+\beta_2+\beta_3\right)\frac{\partial u_i}{\partial t}+ \left(\beta_1\,\beta_2+\beta_2\,\beta_3+\beta_3\,\beta_1 \right)\,u_i+\beta_1\beta_2\beta_3\,U_i\right].
\label{eq:derevationElastic32}
\end{equation}
It should be noted that $s_1(x_1)s_2(x_2)s_3(x_3)/s_j(x_j)=\prod_{i\neq j} \,s_i(x_i)$, and hence does not depend on $x_j$, enabling the RHS of \eqref{eq:derevationElastic2} to be written as:
\begin{equation}
\sum\limits _{j=1}^{3}\frac{\partial}{\partial x_{j}}\left(\sum\limits _{k,l=1}^{3}\frac{s_1s_2s_3}{s_j s_l}\,C_{ijkl}\,\frac{\partial \hat{u}_k}{\partial x_l}\right).
\label{eq:derevationElastic4}
\end{equation}
After some manipulations, it can be shown that
\begin{equation}
\frac{s_1s_2s_3}{s_j s_l}=1+\frac{\beta_1+\beta_2+\beta_3-\beta_j-\beta_l+\frac{\beta_1\beta_2\beta_3}{-i\omega\,\beta_l}}{-i\omega+\beta_j}
\label{eq:derevationElastic41}
\end{equation}
At this point, we introduce the auxiliary variables, $w_{ij}(\mathbf{x},t)$ such that their Fourier transform
\begin{equation}
\hat{w}_{ij}(\mathbf{x},\omega)=\sum\limits _{k,l=1}^{3} \frac{\beta_1+\beta_2+\beta_3-\beta_j-\beta_l+\frac{\beta_1\beta_2\beta_3}{-i\omega\,\beta_l}}{-i\omega+\beta_j}\;C_{ijkl}\frac{\partial \hat{u}_k}{\partial x_l}.
\label{eq:derevationElastic42}
\end{equation}
Multiplying the above equations by $-i\omega+\beta_j$ and taking its inverse Fourier transform results in the following time-domain auxiliary equations:
\begin{equation}
\frac{\partial w_{ij}}{\partial t}+\beta_{j}\, w_{ij} =\sum_{k,l=1}^3\left(\left(\beta_1+\beta_2+\beta_3-\beta_j-\beta_l\right)\,C_{ijkl}\,\frac{\partial u_k}{\partial x_l}+ \frac{\beta_1\beta_2\beta_3}{\beta_l}C_{ijkl}\,\frac{\partial U_k}{\partial x_l}\right),
\label{eq:derevationAuxiliary}
\end{equation}
and the RHS of \eqref{eq:derevationElastic2} becomes
\begin{equation}
\sum\limits _{j=1}^3\frac{\partial}{\partial x_j}\left(\sum\limits _{k,l=1}^3 C_{ijkl}\frac{\partial u_k}{\partial x_l}+w_{ij}\right).
\label{eq:derevationElastic43}
\end{equation}
This concludes our derivation, so that the final second-order time-domain PML formulation for elastic wave propagation in three-dimensional anisotropic solid can be written as
\begin{subequations}
\begin{align}
\rho\left(\frac{\partial^{2}u_i}{\partial t^2}+a\,\frac{\partial u_i}{\partial t}+b\, u_i +c\, U_i\right) & =\sum\limits _{j=1}^3\frac{\partial}{\partial x_j}\left(\sum\limits _{k,l=1}^3 C_{ijkl}\frac{\partial u_k}{\partial x_l}+w_{ij}\right)\\[12pt]
\frac{\partial w_{ij}}{\partial t}+\beta_{j}\, w_{ij} & =\sum_{k,l=1}^3\left(\tilde{C}_{ijkl}\,\frac{\partial u_k}{\partial x_l}+\breve{C}_{ijkl}\,\frac{\partial U_k}{\partial x_l}\right)\\[12pt]
\frac{\partial U_i}{\partial t}&=u_i ,
\end{align}
\label{eq:pmlElastic}
\end{subequations}
where $a(\mathbf{x})=\beta_1+\beta_2+\beta_3$, $b(\mathbf{x})=\beta_1\,\beta_2+\beta_2\,\beta_3+\beta_3\,\beta_1$, $c(\mathbf{x})=\beta_1\,\beta_2\,\beta_3$, $\tilde{C}_{ijkl}(\mathbf{x})=(a-\beta_j-\beta_l)\,C_{ijkl}$, and $\breve{C}_{ijkl}(\mathbf{x})=(c \,/\, \beta_l)\, C_{ijkl}$.
\section{Numerical Methods and Results}
\label{sec:numerical}
For our studies, the source of excitation was a 1~mm radius sphere, centered at the origin and embedded in an infinite 3D medium. To model the infinite medium we assumed a cubic physical domain of 2.0 cm$^{3}$ that is centered at the origin and surrounded by a 2.0~mm PML. The boundary of the sphere was assumed to vibrate with a displacement, whose normalized time-dependence is given by the first derivative of a Gaussian, i.e., by
\begin{equation}
u_{0}\left(t\right)=-\sqrt{2e}\thinspace\pi f_{0}\left(t-t_{0}\right)\thinspace e^{-\pi^{2}f_{0}^{2}\left(t-t_{0}\right)},
\label{eq:source}
\end{equation}
where $f_{0}$ is the dominant frequency and $t_{0}$ is a source delay time. All numerical experiments used $f_{0}=1$ MHz and $t_{0}=1~\mu$s.
The simulations were performed using COMSOL with the second-order Lagrange finite elements employing a cubic mesh for the PML region and the default tetrahedral shape in the physical domain (see \autoref{fig:ThreeDGeometry}a). For discussing the mesh dimensions and time discretization, it is helpful to define the minimum and maximum characteristic wave speeds associated with the medium by $c_{\text{min}}$ and $c_{\text{max}}$. The mesh size is governed by the shortest wavelength of significance for the propagating pulse, i.e., by $c_{\text{min}}/f_c$.
Specifically, the mesh size was assumed to be given by
\begin{equation}
h_{0}=\frac{c_{\text{min}}}{Nf_{0}},\label{eq:meshSize}
\end{equation}
which for the second-order accurate finite elements used, corresponds to $2N$ degrees of freedom per wavelength. As illustrated in \autoref{fig:ThreeDGeometry}a, the mesh employed uses a PML whose thickness consists of just four elements. As will be seen this is sufficient to ensure virtually complete absorption of the various incident waves. For the time discretization, a second-order generalized alpha method, as defined by Chung and Hulbert \citep{Chung:1993io}, was used with $\rho_{\infty}$= 0.75. The step duration was $0.9h_{0}/c_{\text{max}}$, which is just less than the time needed for the fastest wave to travel through the smallest mesh dimension.
To test the validity of our formulation and the accuracy with which our finite element simulations describe the propagating pulse, the exact solution for a monochromatic compressional wave caused by a 1-mm radius sphere whose surface vibrated normal to the surface \citep{Beltzer:1988ut} was used. The sphere was assumed to be embedded in an unbounded isotropic solid (glass). By multiplying this monochromatic solution with the Fourier transform of equation \eqref{eq:source} and then taking the inverse Fourier transform, the time-domain analytic solution was obtained. Good agreement with the numerically calculated response provided convincing evidence for the validity of our 3D formulation, though the results are not presented here.
\begin{figure}[H]
\centering
\mbox{\includegraphics[width=1.0\textwidth]{fig1.png}}
\protect\caption{(a) Showing the 3D grid used in numerical calculations. (b) Snapshot at $t=2.5~\mu$s.illustrating the 3D response for the mineral Olivine.}
\label{fig:ThreeDGeometry}
\end{figure}
To illustrate the application of our general formulation, we chose to present the results for a highly anisotropic medium. In particular, we chose to examine the 3D response when the above source, in \eqref{eq:source}, is contained within in a single crystal of olivine (Mn$_{2}$SiO$_{4}$). Olivine is a mineral with an orthorhombic structure and nine independent elasticity components whose measured parameters at 25$^{\circ}$C are given by \citep{sumino1979elastic,bass1984elasticity}: $c_{11}=2.58$, $c_{22}=1.66$, $c_{33}=2.07$, $c_{44}=0.45$, $c_{55}=0.56$, $c_{66}=0.58$, $c_{12}=0.87$, $c_{13}=0.95$, $c_{23}=0.92$ Mbar. Unlike the isotropic case, fast and slow waves propagate in anisotropic solids even if the excitation is normal to the sphere's surface. Nevertheless, in order to observe a clearer presence of these different waves, we decided to to excite the medium by vibrating the sphere's surface at $45^\circ$ to the normal in the polar direction. Namely, the Dirichlet boundary condition at the surface of the sphere is set to:
\begin{equation}
\frac{1}{\sqrt{2}}(\hat{\mathbf{n}}+\hat{\mathbf{t}}_{\phi})\,u_{0}\left(t\right),
\label{eq:Dirichlet}
\end{equation}
where $u_0$ is defined in \eqref{eq:source}, $\hat{\mathbf{n}}$ is the normal unit vector, and $\hat{\mathbf{t}}_{\phi}$ is the tangential unit vector in the $\phi$ direction, and $\phi$ is the polar angle that varies from 0 to $\pi$ away from the $x_3$-axis. At $\phi=90^\circ$ for example, $\hat{\mathbf{t}}_{\phi}$ is in the negative $x_3$-direction, hence, the quasi-longitudinal wave is expected to be dominant on the $x_1$-$x_2$~plane.
The results of the simulations are presented as density and vectors plots that represent the magnitudes and the directions of the normalized displacement field. While \autoref{fig:ThreeDGeometry}b provides a snapshot of the propagating waves in a 3D format, such an image is difficult to interpret. The three sets of snapshots for three different planes, as shown in the nine panels of \autoref{fig:Planes2D}, provides much more detailed information. These snapshots show 2D plots of the field on the thee principal planes at 1~$\mu$s, 2.2~$\mu$s, and 3.5~$\mu$s. The first column shows the displacement field on the $x_1$-$x_2$~plane, wherein, as expected, the fast wave is dominant. Meanwhile, on the other two planes, the thee waves; the quasi-longitudinal (fast) and the two quasi-shear (slow), are clearly present as shown in the 2.2~$\mu$s snapshots. At this time, the fast wave is being effectively absorbed by the PML, while the slow waves are being absorbed in the 3.5~$\mu$s snapshot.
\begin{figure}[H]
\centering
\mbox{ \includegraphics[width=1.0\textwidth]{fig2.png}}
\caption{Sets of snapshots of the waves produced by the 1-mm radius source excited by the waveform given by equation \eqref{eq:source}, at three different times for three different planes.}
\label{fig:Planes2D}
\end{figure}
The effectiveness of the PML to absorb all the incident energy can be obtained by looking at the manner in which the energy in the physical domain evolves over time. Since the total energy in the physical domain is the sum of the kinetic and potential energy, it can be calculated from
\begin{equation}
E (t)=\frac{1}{2}\int\limits _{\Omega}\left[\rho\sum\limits _{j=1}^{3}\left(\dfrac{\partial u_{j}}{\partial t}\right)^{2}+\sum\limits _{i,j,k,l=1}^{3}C_{ijkl}\dfrac{\partial u_{i}}{\partial x_{j}}\dfrac{\partial u_{k}}{\partial x_{l}}\right]\mathrm{d}\Omega.\label{eq:TotalEnergy}
\end{equation}
\autoref{fig:EnergyDecay} shows that the total energy decays to a negligible level in less than 5$~\mu$s.
\begin{figure}[H]
\centering
\mbox{\includegraphics[width=0.7\textwidth]{Fig3.png}}
\protect\caption{The decay of energy as calculated from \eqref{eq:TotalEnergy} in the physical domain for the mineral Olivine showing that the PML acts as a near ideal absorber.}
\label{fig:EnergyDecay}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
Using PML approach we have addressed the problem of wave propagation in an unbounded, linear anisotropic solid in three dimensions. A time-domain second order PDE has been derived using complex coordinate stretching. The advantages of our time-domain formulation is the fact that it covers the more general inhomogeneous anisotropic case using a small number of equations. Specifically, three second-order equations of the displacement field and nine auxiliary equations, along with the time histories of the displacement field. This simplifies the problem and reduces the computational resources needed. Moreover, use can be made of a wider variety of second-order numerical schemes.
\section*{Acknowledgments}
We wish to thank the Natural Sciences and Engineering Research Council (NSERC) for their support [grant number 3247-2012]. We also wish to thank Pooya Bidari from Ryerson University for drawing to our attention the importance of developing a viscoelastic 3D model.
\newpage
\bibliographystyle{IEEEtranN}
|
\section{Introduction}
The concept of microlensing, whereby a background star is gravitationally lensed by a foreground object, was conceived by \citet{Einstein:36} however it was not until \citet{Paczynski:86} that modern microlensing experiments were born.
One of the primary observational microlensing measurements is the optical depth. The optical depth is defined as the fraction of stars whose projection lies within the Einstein radius of a lens, and it is these stars who see a significant ($>3/\sqrt{5}$) brightening due to microlensing. For a star at distance \ensuremath{D_{\rm s}}\xspace the optical depth $\tau$ is given by \citep[{\it e.g.}\xspace][]{Kiraga:94}
\begin{equation}
\tau(\ensuremath{D_{\rm s}}\xspace) = \frac{4\pi G }{c^2} \int_0^\ensuremath{D_{\rm s}}\xspace \ensuremath{\rho _{\rm l}}\xspace(\ensuremath{D_{\rm l}}\xspace) \left(\frac{1}{\ensuremath{D_{\rm l}}\xspace} - \frac{1}{\ensuremath{D_{\rm s}}\xspace} \right) \ensuremath{D_{\rm l}}\xspace^2 \, d\ensuremath{D_{\rm l}}\xspace \label{eq:tau} \end{equation}
where $\ensuremath{\rho _{\rm l}}\xspace(\ensuremath{D_{\rm l}}\xspace)$ is the density of lenses. That the optical depth depends only on the density of lenses, and not on their distribution of masses or velocities, makes it theoretically very elegant.
An original motivation for microlensing studies was to explore the possibility that the Galactic dark halo could be composed of Massive Astrophysical Compact Halo Objects (MACHOs) \citep{Paczynski:86}. The first results from the MACHO survey measuring the optical depth towards the LMC suggested, based on $\sim 10$ events, that 20\% of the Galactic halo could be composed of MACHOs with average mass $\sim 0.4\ensuremath{{M_\odot}}\xspace$ \citep{Alcock:00,Bennett:05}. In contrast the EROS survey found that less than 8\% of the halo could be composed of MACHOs with mass $\sim 0.4\ensuremath{{M_\odot}}\xspace$ \citep{Tisserand:07}. The OGLE-III survey was much larger in both duration and coverage and found stringent limits of the contribution of MACHOs: a fraction less than $7\%$ for sub-solar lenses \citep{Wyrzykowski:11}. Hence, the optical depth observed towards the LMC has turned out to be too small for the halo to contain more than a small fraction of dark objects able to microlense, most likely because dark matter is composed of particles with low mass, so that they have an Einstein radius far too small to produce a significant brightening \citep[{\it i.e.}\xspace $\lesssim 10^{-7}\ensuremath{{M_\odot}}\xspace$, ][]{Paczynski:86}.
This instead makes microlensing of Milky Way bulge stars a unique tool for unravelling the structure of galaxies because only stellar matter is able to produce microlensing events. Typically the mass profiles of Galaxies can be constrained dynamically, but this may be distributed between baryonic and dark matter \citep[{\it e.g.}\xspace][]{Courteau:14}. In contrast we are able to use bulge microlensing to break the stellar to dark matter degeneracy in the inner Milky way ($R \lesssim 5\ensuremath{\,{\rm kpc}}\xspace$) in a manner that is generally not possible in external galaxies without additional assumptions such as on the stellar mass-to-light ratio \citep{Iocco:11}.
Microlensing towards the bulge was first considered by \citet{Paczynski:91} and \citet{Griest:91bulge}. Initial predictions were based on bulge models fitted to COBE data \citep{Han:95,Bissantz:97,Evans:02,Bissantz:02}. However the measured optical depths were significantly higher than these models predicted. OGLE measured $\tau=(3.3\pm1.2)\times 10^{-6}$ \citep{Udalski:94} and MACHO $\tau =(3.9\substack{+1.8 \\ -1.2})\times 10^{-6}$ \citep{Alcock:97}. These early optical depths were higher even than the theoretical bounds: no model could be constructed that reproduced $\tau$ without overshooting the rotation curve \citep{Binney:00}. Later estimates with improved treatment of blending through difference image analysis (DIA) \citep{Alcock:00bulge} and focusing on brighter resolved giants \citep{Popowski:05,Hamadache:06} produced reduced optical depths, so that at least some possible Galactic models were consistent with the data \citep[{\it e.g.}\xspace][]{Wood:05}, and actually brought the optical depth into line with the earlier predictions.
Pioneering studies from OGLE \citep{Udalski:94} and MACHO \citep{Alcock:97} were based on a handful of microlensing events. Since then several thousand microlensing events have been detected by recent and ongoing microlensing surveys such as EROS-2 \citep{Afonso:03,Hamadache:06}, MOA-2 \citep{Sumi:11,Sumi:13,Sumi:16}, MACHO \citep{Popowski:05}, WeCAPP \citep{Lee:15}, OGLE-III \citep{Wyrzykowski:15}, and OGLE-IV \citep{Udalski:15}. These larger samples, the convergence between measurements of recent optical depths for bright resolved stars with fainter unresolved stars \citep{Sumi:16}, and the new made-to-measure models of the bulge \citep{Portail:15}, prompts us to revisit of the constraints provided by bulge microlensing data on Galactic models. We use recent microlensing data to constrain the amount of stellar matter towards the Galactic bulge, and consequently the fractional contributions of stellar and dark matter in the inner Galaxy. The data is primarily taken from the MOA-II survey \citep{Sumi:13} which was recently updated by \citet{Sumi:16} with improved estimates of the effective number of monitored stars. We use the data from \citet{Sumi:16} throughout, referring to it as the revised MOA-II data. We additionally consider and cross check our results with data from the EROS-II survey \citep{Hamadache:06}.
\begin{table*}
\caption{Summary description of the model Galaxy used to predict microlensing quantities.}
\label{tab:model}
\begin{tabular}{llll}
\hline
& Property & Fiducial & Variations \\
\hline
Bulge Model & N-body model from \citet{Portail:15} & M90 \ & M80, M85 \\
\hline
\multirow{4}{*}{Disk Model} & Disk scale length, $R_d$ & 2.6\ensuremath{\,{\rm kpc}}\xspace & Range: $1.9-3.4\ensuremath{\,{\rm kpc}}\xspace$\\
& Solar neighbourhood disk scale height, $H_\odot$ & 0.3\ensuremath{\,{\rm kpc}}\xspace & --- \\
& Inner disk scale height , $H_{4.5}$ & 0.18\ensuremath{\,{\rm kpc}}\xspace & 0.3\ensuremath{\,{\rm kpc}}\xspace \\
& Local stellar surface density, $\Sigma_\odot$ & $38\, \ensuremath{{M_\odot}}\xspace\,\ensuremath{\,{\rm kpc}}\xspace^{-2}$ & --- \\
\hline
\multirow{4}{*}{Stellar Population} & Source stellar population & 10Gyr, Baade's Window MDF from \citet{Zoccali:08} & --- \\
& Isochrones & $\alpha$-enhanced BASTI \citep{Pietrinferni:04} & --- \\
& IMF & \citet{Kroupa:01} & \citet{salpeter:55} \\
& Remnant masses & Prescription from \citet{Maraston:98} & --- \\
\hline
\end{tabular}
\end{table*}
This paper proceeds as follows:
In \autoref{sec:models} we describe our models of the distribution of stellar matter in the Milky Way.
In \autoref{sec:microtheory} we describe how we calculate microlensing properties such as the optical depth and timescale distribution which can then be subsequently compared to the data in sections \ref{sec:data} and \ref{sec:eros}. We compute the resultant rotation curves consistent with data in \autoref{sec:rotcurve}, and briefly investigate the timescale distribution of events in \autoref{sec:tedistdiscuss}. We discuss the consequences of the results in \autoref{sec:discuss}, in particular the limits these place on the halo contribution, and place our Milky Way constraints in the context of external galaxies. We conclude in \autoref{sec:conc}.
\section{Model}
\label{sec:models}
Using the microlensing data we constrain the distribution of stellar matter in a model Milky Way galaxy. Our model consists of N-body bulges together with parametric exponential stellar disks. The properties of these models are summarised in \autoref{tab:model}.
\subsection{Bulge}
\label{sec:bulgemodels}
In the central region we use the N-body models constructed by \citet{Portail:15}. These models were constructed by the made-to-measure method to match the three dimensional density of red clump stars measured by \citet{Wegg:13} together with the radial velocity measurements provided by the BRAVA survey \citep{Kunder:12}. This combination of data highly constrains the total mass in a central box of size $(\pm 2.2 \times \pm 1.4 \times \pm 1.2 \ensuremath{\,{\rm kpc}}\xspace)$ to be $(1.84\pm 0.07 )\times 10^{10} \ensuremath{{M_\odot}}\xspace$. The proportions of stellar and dark matter to this total is however not well dynamically constrained. The models have stellar masses in the range $(1.35 - 1.6)\times 10^{10}\ensuremath{{M_\odot}}\xspace$. We use the same naming convention as \citet{Portail:15}, the models are labeled M80-M90 corresponding to the initial levels of disk contribution of 80-90\% at $R=2.2R_d$.
As our fiducial model we use model M90, and consider the other variants, M80 and M85, when noted. As described in \citet{Portail:15}, these models are expected to cover the range of possible stellar masses for the bulge on the basis of the stellar population and mass-to-light ratio for reasonable IMFs. We use model M90 as our fiducial model because it is consistent with the mass-to-light and mass-to-red clump ratios of a 10Gyr old Kroupa IMF. As we show in \autoref{sec:tedistdiscuss}, a Kroupa IMF also results in a timescale distribution of microlensing events close to that observed, making this model the most self-consistent.
\subsection{Disk}
\label{sec:diskmodels}
The N-body models of \citet{Portail:15} were fitted to the bulge where they are expected to be accurate dynamical models. Outside the bulge however they were not constrained, and therefore do not there accurately represent the Milky Way. In particular they have a disk scale length of $1.1-1.2\ensuremath{\,{\rm kpc}}\xspace$, significantly shorter than measured in the Milky Way {\it e.g.}\xspace typical values for inner MW disk based on NIR and surface mass density are 2.5\ensuremath{\,{\rm kpc}}\xspace: \citet{Binney:97}; 2.4\ensuremath{\,{\rm kpc}}\xspace: \citet{Bissantz:02}; 2.6\ensuremath{\,{\rm kpc}}\xspace: \citet{Juric:08}; 2.15\ensuremath{\,{\rm kpc}}\xspace: \citet{Bovy:13}; see \citet{Ortwin:araa} for a fuller discussion. This is because the self-consistent pure N-body models from which they were derived have large bar to disk scale length ratios, and therefore when the bar is scaled to the Milky Way short disk scale lengths and very low stellar densities in the solar neighbourhood result. We therefore truncate the N-body model at a radius \ensuremath{R_{\rm cut}}\xspace, using the N-body models inside, and a simple analytic disk model outside.
We consider the stellar disk as a double exponential
\begin{equation}
\rho = \frac{\Sigma_\odot}{2 H(R)} \exp \left( \frac{R_0 - R}{R_d} \right) \exp \left( - \frac{|z|}{H(R)} \right)
\end{equation}
where $R$ is the galactocentric radius, $R_0$ is the solar galactocentric radius, $R_d$ is the disk scale length, $z$ the distance from the Galactic plane, and $H$ the disk scale height.
$\Sigma_\odot$ is the stellar surface density in the solar neighbourhood and $H_\odot$ the scale height in the solar neighbourhood. The local disk is relatively well constrained and here we use fiducial local disk parameters of $\Sigma_0=38\, \ensuremath{{M_\odot}}\xspace\,\ensuremath{\,{\rm kpc}}\xspace^{-2}$ \citep{Bovy:13} and $H_\odot=0.3\ensuremath{\,{\rm kpc}}\xspace$ \citep{Juric:08}.
The disk towards the inner Galaxy is considerably less well constrained, both in terms of its increase in surface density (parameterised here though the disk scale length $R_d$) and the disk scale height.
There is evidence that the Milky Way disk scale height may decrease inside the solar circle (\citealt{Kent:91,LopezCorredoira:02}, the 180\ensuremath{\,{\rm pc}}\xspace thin bar in \citealt{Wegg:15}). We therefore allow a flaring of the disk between $4.5\ensuremath{\,{\rm kpc}}\xspace$ and the solar neighbourhood parameterised through
\begin{equation}
H(R)=
\begin{cases}
H_\odot + (R - R_0)\frac{H_\odot - H_{4.5}}{R_0 - 4.5\ensuremath{\,{\rm kpc}}\xspace} & \mbox{if } R > 4.5\ensuremath{\,{\rm kpc}}\xspace \\
H_{4.5} & \mbox{if } R \le 4.5\ensuremath{\,{\rm kpc}}\xspace
\end{cases} ~ .
\end{equation}
We consider two choices: either $H_{4.5}=180\ensuremath{\,{\rm pc}}\xspace$ \citep[the long bar scale height found in][]{Wegg:15}, or that the scale height is constant and $H_{4.5}=300\ensuremath{\,{\rm pc}}\xspace$.
We choose this simple disk model since the microlensing results are most sensitive to the stellar disk inwards towards the bulge, where the disk density and profile is highly uncertain, and therefore more complex models are not justified.
As our fiducial value of $R_d$ we use 2.6\ensuremath{\,{\rm kpc}}\xspace \citep{Juric:08}. When considering variations on this we use a range of 1.9\ensuremath{\,{\rm kpc}}\xspace to 3.4\ensuremath{\,{\rm kpc}}\xspace considering this to cover the range deemed reasonable by recent data \citep{Ortwin:araa}.
We switch between the disk and bulge models at a radius \ensuremath{R_{\rm cut}}\xspace which is chosen so that the azimuthally averaged surface density is continuous. This choice is made so that the rotation curve is smooth. A typical value of \ensuremath{R_{\rm cut}}\xspace is $1.6\ensuremath{\,{\rm kpc}}\xspace$ from the fiducial model.
When velocities of the disk are required we use a Schwarzschild velocity distribution with mean velocity taken from the rotation curve of \citet{Sofue:09}. For the local velocity dispersions we use values from \citet{Dehnen:1998} and extrapolate these inwards using an exponential increase $\exp (-R/R_\sigma)$ with $R_\sigma=2\times4.37\ensuremath{\,{\rm kpc}}\xspace$ \citep{Lewis:89}. Our results are not sensitive to these choices. This is because they are based on the optical depth which is not highly sensitive to the velocities and resultant timescale distribution as shown by the IMF independence in \autoref{sec:minoraxis}.
Although we use an analytic model for the disk, in practice we produce N-body realisations when computing microlensing quantities for consistency with the N-body bulge model.
\subsection{Stellar Population}
\label{sec:pop}
Because the majority of stars microlensed are in the bulge, when the luminosity function is required we use the $\alpha$-enhanced BASTI isochrones from \citet{Pietrinferni:04} for 10Gyr, the MDF measured by \citet{Zoccali:08} towards Baade's window, and a \citet{Kroupa:01} IMF. We show in \autoref{sec:minoraxis} the effect of altering the IMF to the \citet{salpeter:55} IMF and consider the effects of the IMF on the timescale distribution in \autoref{sec:tedistdiscuss}. For the lens mass distribution a remnant mass prescription is required. For this we use the same prescription for white dwarfs, black holes and neutron stars as \citet{Maraston:98}: White dwarfs are formed from initial masses $M_{\rm i}<8.5\ensuremath{{M_\odot}}\xspace$ with remnant mass $M_{\rm f}=0.077M_{\rm i}+0.48\ensuremath{{M_\odot}}\xspace$, neutron stars of mass $M_{\rm f}=1.4\ensuremath{{M_\odot}}\xspace$ result from $8.5\ensuremath{{M_\odot}}\xspace \le M_{\rm i} < 40\ensuremath{{M_\odot}}\xspace$, and black holes of mass $M_{\rm f}=0.5 M_{\rm i}$ result from $M_{\rm i} \ge 40\ensuremath{{M_\odot}}\xspace$. Throughout we assume that white dwarfs, black holes and neutron stars receive no kick at birth and are distributed as their progenitors.
\section{Microlensing Parameters}
\label{sec:microtheory}
\subsection{Optical Depth}
\label{sec:tautheory}
In our N-body models, for a source particle $i$ at distance $\ensuremath{D_i}\xspace$ the optical depth $\tau_i$ can be calculated as a Monte-Carlo integral over the foreground lens particles in a small field around the star of interest ({\it cf.}\xspace \autoref{eq:tau})
\begin{equation}
\tau_i = \frac{4\pi G}{c^2 \omega} \sum_j M_j \left(\frac{1}{D_j} - \frac{1}{D_i} \right) \label{eq:taumc}
\end{equation}
where $M_j$ is the mass of particle $j$ and the sum runs over all particles in a pencil beam of solid angle $\omega$ with distance $D_j<D_i$. We use this notation throughout.
The values of optical depth measured by observational surveys however are an average over observable stars. Typically two values are given: An optical depth for all sources, and one confined to the brighter giants in the region of the colour-magnitude diagram which contains red clump giants (RCGs). These two values can differ both because surveys may have problems with blending of unresolved fainter stars which is less important for the brighter RCGs, and also because their luminosity function and therefore weighting by distance is different. Since RCGs are bright enough to be observed throughout the bulge then it is typically assumed that there is no weighting by distance. For the all-source microlensing a slightly different weighting is typically utilised. Intrinsically fainter stars are more numerous than their brighter counterparts and therefore in a magnitude limited sample of bulge giants the more distant stars are down weighted by $\approx r^{-1.4}$ \citep{Portail:15}. \cite{Kiraga:94} for example parameterise this weighting by as a power law luminosity function $L^\beta$ so that the number of detectable stars in a field brighter than some faint magnitude limit is
\begin{equation}
dn \propto \ensuremath{\rho _{\rm s}}\xspace(\ensuremath{D_{\rm s}}\xspace) \ensuremath{D_{\rm s}}\xspace^2 \ensuremath{D_{\rm s}}\xspace^{2\beta} \, d\ensuremath{D_{\rm s}}\xspace ~.
\label{eq:betaparam}
\end{equation}
Stars visible throughout the bulge therefore have $\beta=0$ while other bulge giants have $\beta\approx -0.7$.
From \autoref{eq:tau} the measured optical depths are therefore the weighted average over observable stars
\begin{dmath}
\langle \tau(\beta) \rangle = \frac{\int \ensuremath{\rho _{\rm s}}\xspace(\ensuremath{D_{\rm s}}\xspace) \ensuremath{D_{\rm s}}\xspace^{2+2\beta} \tau(\ensuremath{D_{\rm s}}\xspace) \, d\ensuremath{D_{\rm s}}\xspace}{\int \ensuremath{\rho _{\rm s}}\xspace(\ensuremath{D_{\rm s}}\xspace) \ensuremath{D_{\rm s}}\xspace^{2+2\beta} \, d\ensuremath{D_{\rm s}}\xspace} ~.\label{eq:tauavg}
\end{dmath}
Practically this can be calculated by taking the weighted average of \autoref{eq:taumc} \citep{Fux:97a}
\begin{dmath}
\langle \tau(\beta) \rangle = \frac{4\pi G}{c^2 \omega} \left( \sum_i M_i D_i^{2\beta} \right)^{-1} \sum_{i,j} M_i M_j D_i^{2\beta} \left(\frac{1}{D_j} - \frac{1}{D_i} \right) \label{eq:tauavgmc} ~.
\end{dmath}
However we now show that the simple power law parametrisation \autoref{eq:betaparam} does not adequately capture the variation of $\tau$ because the luminosity function is not sufficently close to a power law. This was demonstrated by \citet{Stanek:95} and \citet{Wood:07} for source stars with magnitudes similar to bulge RCGs, and is also true near the main sequence turnoff.
An improvement over \autoref{eq:betaparam} is to use a luminosity function constructed from isochrones. In this case the number of stars with distances between $\ensuremath{D_{\rm s}}\xspace$ and $\ensuremath{D_{\rm s}}\xspace+d\ensuremath{D_{\rm s}}\xspace$ and between extinction corrected apparent \I-band\xspace magnitude $I_0$ and $I_0+dI_0$ will be proportional to
\begin{equation}
dn \propto \ensuremath{\rho _{\rm s}}\xspace(\ensuremath{D_{\rm s}}\xspace) \ensuremath{D_{\rm s}}\xspace^2 \Phi_{\rm I}(I_0-5\log[\ensuremath{D_{\rm s}}\xspace/10{\rm pc}]) \, d\ensuremath{D_{\rm s}}\xspace \, d\ensuremath{{I}}\xspace_0 ~,
\end{equation}
where $\Phi_{\rm I}$ is the \I-band\xspace luminosity function ({\it i.e.}\xspace the number of stars per unit mass with absolute \I-band\xspace magnitudes between $M_I$ and $M_I+dM_I$ is $\Phi_{\rm I}(M_I) \, dM_I$). The optical depth as a function of magnitude is therefore
\begin{equation}
\langle \tau(I_0) \rangle = \frac{\int \tau(\ensuremath{D_{\rm s}}\xspace) \ensuremath{\rho _{\rm s}}\xspace(\ensuremath{D_{\rm s}}\xspace) \Phi_{\rm I}(I_0-5\log[\ensuremath{D_{\rm s}}\xspace/10{\rm pc}]) \ensuremath{D_{\rm s}}\xspace^2 \, d\ensuremath{D_{\rm s}}\xspace}{\int \ensuremath{\rho _{\rm s}}\xspace(\ensuremath{D_{\rm s}}\xspace) \Phi_{\rm I}(I_0-5\log[\ensuremath{D_{\rm s}}\xspace/10{\rm pc}]) \ensuremath{D_{\rm s}}\xspace^2 \, d\ensuremath{D_{\rm s}}\xspace} ~.
\end{equation}
This can be written in terms of distance modulus $\mu$ as
\begin{equation}
\langle \tau(I_0) \rangle = \frac{\int \tau(\mu) \Delta(\mu) \Phi_{\rm I}(I_0-\mu) \, d\mu}{\int \Delta(\mu) \Phi_{\rm I}(I_0-\mu) \, d\mu} \label{eq:tauvsI0}
\end{equation}
where $\Delta(\mu)=\ensuremath{\rho _{\rm s}}\xspace \ensuremath{D_{\rm s}}\xspace^3$ written as a function of $\mu$. If the optical depth as a function of extincted source magnitude $I=I_0+A_I(D_s)$ is required, then this expression becomes
\begin{equation}
\langle \tau(I) \rangle = \frac{\int \tau(\mu_I) \Delta_I(\mu_I) \Phi_{\rm I}(I-\mu_I) \, d\mu_I}{\int \Delta_I(\mu_I) \Phi_{\rm I}(I-\mu_I) \, d\mu_I} \label{eq:tauvsI}
\end{equation}
where $\mu_I=\mu+A_I$ and
\begin{equation}
\Delta_I(\mu_I) = \frac{\ensuremath{\rho _{\rm s}}\xspace \ensuremath{D_{\rm s}}\xspace^3}{\left[1+ \frac{dA_I}{d\log D_s} \right]} ~. \label{eq:deltaI}
\end{equation}
We use the expressions in Equations \ref{eq:tauvsI0}--\ref{eq:deltaI} to calculate optical depths that are more accurate and representative of the data. An alternative approach is to produce Monte Carlo realisations of mock fields like the Besan\c{c}on model \citep{Robin:03} and simulate the microlensing properties of these \citep{Kerins:09,Penny:13,Awiphan:16}. In this work however we wish to consider a large range of Galactic models, and compute their microlensing properties, to decide which are consistent with the microlensing data. For this we use the fast but accurate method provided above.
For the brightest stars we expect departures from these expressions because of the finite source effect: if the magnification changes significantly over the angular diameter of the source, then the overall magnification is smaller than the point source approximation used in \autoref{eq:tau}, and therefore the optical depth is smaller. The brightest bulge giants are also the largest and so would most significant for these. A typical RCG in the bulge however has angular radius 6$\,\mu$as, more than an order of magnitude smaller than the typical Einstein radius of bulge microlensing events. Finite source effects are there important only for the rare highest magnification events of giants, and do not significantly effect the overall optical depths for RCGs, and are negligible for MS stars.
\begin{figure}
\includegraphics[width=\linewidth]{bwtau-crop}
\caption{\label{fig:tauvsImag} The microlensing optical depth as function of \I-band\xspace magnitude in the direction of Baade's window ($l=1.0^\circ$, $b=-3.9^\circ$). The optical depth of our fiducial model is shown as the solid black line calculated though \autoref{eq:tauvsI0}. The labeled red lines are the optical depth parameterised through the $\beta$ parameter as in \autoref{eq:tauavg}. The dashed grey line is the magnitude distribution predicted by the model in this field plotted against the right axis.}
\end{figure}
Practically for the N-body models considered we compute \autoref{eq:tauvsI0} from the particles with the following procedure: up to an unimportant normalisation $\Delta(\mu)$ is just the particle masses, $M_i$, binned in equal distance modulus increments. The denominator is computed by convolving this with the luminosity function through an FFT. The numerator is computed similarly by convolution with the luminosity function after binning $M_i \tau_i$ in equal distance modulus increments. This procedure is efficient, and accurate provided that the bin size is chosen sufficiently narrow to capture the smallest scale variations in the luminosity function and distance distribution. \autoref{eq:tauvsI} can be similarly computed with each particle mass weighted by the additional denominator in \autoref{eq:deltaI}. We show in \autoref{fig:tauvsImag} a representative plot of the optical depth as a function of \I-band\xspace magnitude towards Baade's window.
Several features are clear regarding the optical as function of magnitude in \autoref{fig:tauvsImag}: For bright stars the optical depth is low. This is because, brighter than the tip of the red giant branch of the bulge, the observed stars are disk stars with low optical depth. The prominent dip, and subsequent rise in optical depth at $I_0 \approx 14$ is due to the numerous RCGs, which can also be seen in the luminosity function. Because RCGs are approximate standard candles, stars towards the bright end of the bump in the luminosity function are likely to be towards the front of the bulge, and subsequently have a low optical depth. By contrast stars towards the faint side of the bump are more likely to lie at the distant side of the bulge and therefore have a high optical depth. Fainter than this the optical depth drops, continuing to drop as disk main sequence stars which have low optical depth become numerous. The optical depth finally rises again near $I_0 \approx18$ when the main sequence turn off of the bulge is reached, and bulge dwarfs dominate the sample.
\begin{figure}
\includegraphics[width=\linewidth]{ogle3tauvsmag-crop}
\caption{The data with error bars are an approximate number of microlensing events per star detected by the OGLE-III survey \citep{Wyrzykowski:15} as function of extinction corrected $K_s$-band magnitude. The extinction corrected K-band magnitude distribution of microlensing events is divided by the magnitude distribution of all stars. Since the efficiency of microlensing detection changes slowly with magnitude then variations are a proxy for the change in optical depth with magnitude. We plot as the solid black line predicted shape from our fiducial model towards Baade's window. On the right axis we plot in grey the magnitude distribution predicted by the model (dashed line) together with the magnitude distribution observed (solid line) to illustrate the position of the bump due to RCGs in the bulge and the relative position of the jump in rate. All curves were arbitrarily renormalised because the field-by-field microlensing efficiencies and number of monitored sources is not yet known for the OGLE-III survey. \label{fig:tauvsKmagOGLE}}
\end{figure}
The feature at $I_0 \approx 14$ was predicted by \citet{Stanek:95} and elaborated on by \citet{Wood:07}. We demonstrate in \autoref{fig:tauvsKmagOGLE} that it is also present in the current data. We use the OGLE-III data because it is has a higher statistical significance being a larger sample than the MOA-II sample. However the full efficiencies of the large OGLE-III sample of microlensing events has not yet been computed. Instead in this figure we plot a number which should approximately track the changes in optical depth: the number of events, divided by the magnitude distribution, as a function of extinction corrected $K_s$-band magnitude in 0.2\ensuremath{\,{\rm mag}}\xspace bins. The resultant number should be similar to the event rate per star, which is proportional to optical depth for a constant timescale distribution (\autoref{sec:tetheory}).
The $K$-band data is taken from VVV by cross matching to the OGLE-III events. Extinction is corrected using the $H-K_s$ colour \citep[NICE method,][]{Lada:94} and a \citet{Nishiyama:09} extinction law. The extinction corrected magnitude distribution was computed using the 100 nearest stars to each microlensing event so that it is representative. Because the efficiency changes slowly with magnitude the resultant number should change in proportion to the event rate per star, at least over narrow magnitude ranges like the feature at the magnitude of bulge red clump giants. This feature is visible in the data as the increase in event rate at $K_0 \approx 13$.
Because of the changes in optical depth as a function of magnitude, and the inability of a single $\beta$ value to capture these changes, we urge the use of more accurate expressions such as those in Equations \ref{eq:tauvsI0}--\ref{eq:deltaI} when comparing to microlensing data.
Unfortunately a significant contribution to the optical depth arises from long duration microlensing events which are difficult to detect. As a result the error on the total microlensing optical depth can be strongly influenced by these long events. Instead what is often measured are quantities such as $\tau_{200}$, the optical depth excluding events with timescale longer than 200 days. This is also the case for the optical depth of the revised MOA-II data \citet{Sumi:16}. We therefore proceed to estimate the timescale distribution of microlensing events.
\subsection{Event Timescales and Microlensing Rates}
\label{sec:tetheory}
The event timescale is important to consider for two reasons. First in the lowest order description of microlensing the measurable parameters for each event are the source magnitude, and the event amplification, and its duration. While the amplification is geometrical and provides no useful information on the Galaxy, the timescale depends on the relative velocity of the lens and source together with the mass of the lens and therefore has physical importance. It has for example therefore been used to provide constraints on the mass spectrum of lenses \citep[{\it e.g.}\xspace][]{Han:96,CalchiNovati:08}, and low mass objects because these generate characteristically short timescale events \citep{Sumi:11}. Secondly, as mentioned above, microlensing surveys are of finite duration and therefore insensitive to long time scale events. Because these events contribute a significant amount of the total optical depth it is better to consider the optical depth excluding long duration events. The disadvantage is that both the lens mass distribution and a dynamical model is required.
The timescale $t_E$ of a microlensing event is given by
\begin{align}
\ensuremath{t_{\rm E}}\xspace &=R_E/{\rm v} = \frac{1}{\rm v} \sqrt{\frac{4 G M_{\rm l} \ensuremath{D_{\rm l}}\xspace^2}{c^2} \left(\frac{1}{\ensuremath{D_{\rm l}}\xspace} - \frac{1}{\ensuremath{D_{\rm s}}\xspace} \right)} \\
&=40 \, {\rm days} \left(\frac{200\ensuremath{\,{\rm km}\,{\rm s}^{-1}}\xspace}{\rm v}\right) \left(\frac{M}{M_\odot}\right)^{1/2} \left( \frac{\ensuremath{D_{\rm s}}\xspace}{10 \ensuremath{\,{\rm kpc}}\xspace} \right)^{1/2} \frac{ \left[ x(x -1) \right]^{1/2} }{0.5} \nonumber
\end{align}
where $R_E$ is the Einstein radius, $M_{\rm l}$ the mass of the lens, $x\equiv \ensuremath{D_{\rm l}}\xspace/\ensuremath{D_{\rm s}}\xspace$, and $v$ is the transverse velocity of the lens relative to the line of sight toward the source star. It is given by
\begin{equation}
\bld{v} = \bld{v}_{\rm l} - \bld{v}_{\rm o}+(\bld{v}_{\rm o} - \bld{v}_{\rm s})\frac{\ensuremath{D_{\rm l}}\xspace}{\ensuremath{D_{\rm s}}\xspace}
\end{equation}
where $\bld{v}_{\rm l}$, $\bld{v}_{\rm o}$ and $\bld{v}_{\rm s}$ are the transverse velocities of the lens, observer and source respectively.
We can calculate the event time distribution for a mono-mass lens population by considering that the optical depth is \citep{Paczynski:91}
\begin{equation}
\tau=\frac{\pi}{2}\int \ensuremath{t_{\rm E}}\xspace \Gamma(\log \ensuremath{t_{\rm E}}\xspace) \, d\log \ensuremath{t_{\rm E}}\xspace~,
\label{eq:gammadef}
\end{equation}
where $\Gamma(\log \ensuremath{t_{\rm E}}\xspace)$ is the event rate as a function of $\log \ensuremath{t_{\rm E}}\xspace$ {\it i.e.}\xspace the rate of microlensing events in the range $\log \ensuremath{t_{\rm E}}\xspace \rightarrow \log \ensuremath{t_{\rm E}}\xspace+d\log\ensuremath{t_{\rm E}}\xspace$ is $\Gamma(\log \ensuremath{t_{\rm E}}\xspace)$.
Therefore, in the $\beta$ parameterisation of the luminosity function, to calculate the event time distribution from the N-body model we calculate for each lens-source pair $(j,i)$ in a field the event time, $t_{E,ij}$, and give each the weight
\begin{equation}
d\Gamma_{ij}= \frac{2}{\pi} \frac{\tau_i}{t_{E,ij}} \frac{M_i D_i^{ 2\beta}}{\sum_i M_i D_i^{2\beta}} ~.
\label{eq:dgammaij}
\end{equation}
Binning these weights as a function of $t_E$ gives the timescale distribution.
To extend to a mass spectrum then consider a lens mass distribution $\Phi({\log~} M_l)$. The corresponding probability of an star being microlensed by mass $M_l$ at a point in time is $\propto R_E^2 \Phi({\log~} M_l) \propto M_l \Phi({\log~} M_l)$. To compute the event time distribution we can first evaluate the event time distribution for a $1\ensuremath{{M_\odot}}\xspace$ population, defining this $\gamma({\log~} \ensuremath{t_{\rm E}}\xspace)$. The multi-mass event distribution can then be deduced by considering that the event timescale is proportional to $\sqrt{M_l}$ with weighting $1/\sqrt{M_l} \times M_l \Phi({\log~} M_l)$. The additional $1/\sqrt{M_l}$ weighting arises because the rate is related to the optical depth through an additional factor of $1/\ensuremath{t_{\rm E}}\xspace$ (\autoref{eq:gammadef}).
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{rate2d-crop}
\caption{The microlensing event rate per star as function of \I-band\xspace magnitude and event timescale $\log t_E$ in the direction of Baade's window ($l=1.0^\circ$, $b=-3.9^\circ$). The same assumptions on the stellar population as in \autoref{fig:tauvsImag} was used. \label{fig:bw2deventrate}}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth]{tedist-crop}
\caption{The predicted timescale distribution in the direction of Baade's window ($l=1.0^\circ$, $b=-3.9^\circ$). We show as the red line the event rate for a mono-mass $1\ensuremath{{M_\odot}}\xspace$ population. The $\sqrt{M_l}$ weighted lens mass distribution is shown as the dashed blue line. When convolved these functions give the black line: the event rate for the assumed mass spectrum of lenses. The rates in this plot, $\Gamma$ are computed per star, per unit ${\log~} \ensuremath{t_{\rm E}}\xspace$, for stars with $I_0 < 16.7$. Although the timescale distribution is fairly insensitive to magnitude, its normalisation is. The same assumptions on the stellar population as in \autoref{fig:tauvsImag} was used. In the lens mass function we use the assumptions for remnants described in \autoref{sec:pop} and the resultant white dwarfs, neutron stars are visible as the spikes at $0.6\ensuremath{{M_\odot}}\xspace$, $1.4\ensuremath{{M_\odot}}\xspace$. We use this prescription throughout the paper when timescales are required. \label{fig:tedist}}
\end{figure}
The event rate is therefore given by \citep{Han:96}
\begin{align}
\Gamma({\log~} \ensuremath{t_{\rm E}}\xspace) &= \int \gamma({\log~} \ensuremath{t_{\rm E}}\xspace') \Phi({\log~} M_l) \times \label{eq:teconv} \\
&\qquad \qquad \sqrt{M_l} \delta ({\log~} (\ensuremath{t_{\rm E}}\xspace' \sqrt{M_l}) - {\log~} \ensuremath{t_{\rm E}}\xspace) d\,{\log~} \ensuremath{t_{\rm E}}\xspace' d\,{\log~} M_l \nonumber \\
&= \int 10^{\frac{1}{2} {\log~} M_l} \gamma({\log~} \ensuremath{t_{\rm E}}\xspace - \frac{1}{2} {\log~} M_l) \Phi({\log~} M_l) d\,{\log~} M_l ~. \nonumber
\label{eq:teconv}
\end{align}
Defining $x={\log~} t_E$ and $y=\frac{1}{2} {\log~} M_l$ then it is clear the event timescale is a convolution
\begin{equation}
\Gamma(x) = 2 \int dy 10^{y} \Phi(2y) \gamma(x - y) ~. \label{eq:teconvobv}
\end{equation}
We can therefore easily compute the event timescale distribution by computing and binning the rate of each particle pair from \autoref{eq:dgammaij} and convolving with the appropriately scaled and weighted mass spectrum using an FFT.
This provides a recipe for calculation of the event rate and timescale distribution for the $\beta$ parameterisation. To calculate the rate as function of source magnitude, $\Gamma(\log \ensuremath{t_{\rm E}}\xspace,I_0)$, instead a two dimensional convolution is required. In this case we compute for particle pair the weight
\begin{equation}
d\Gamma_{ij}= \frac{2}{\pi} \frac{\tau_i M_i}{t_{E,ij}} ~.
\label{eq:dgammaij2}
\end{equation}
This is then two dimensionally binned in equal increments of $\mu$ and $\log \ensuremath{t_{\rm E}}\xspace$. We then convolve this with a kernel of the luminosity function in the distance modulus direction by reference to \autoref{eq:tauvsI0}, and the scaled mass distribution in the timescale direction by reference to \autoref{eq:teconv}. After normalisation with the magnitude distribution, this provides $\Gamma(\log \ensuremath{t_{\rm E}}\xspace,I_0)$, the event rate per star as a function of timescale and magnitude, the primary microlensing observables.
We show in \autoref{fig:bw2deventrate} the event rate $\Gamma(\log \ensuremath{t_{\rm E}}\xspace,I_0)$ in the direction of Baade's window. While the event rate is a strong function of source magnitude, $I_0$, the timescale distribution is not. Only in the tilt of the red clump region is there a visible change.
We show in \autoref{fig:tedist} the event timescale distribution. This shows how this distribution arises through convolution of the mono-mass timescale distribution of the dynamical model, with the distribution of lens masses. The resultant timescale distribution is a useful probe of the mass distribution lenses in particular which we briefly consider in \autoref{sec:tedistdiscuss}.
\begin{figure*}
\includegraphics[width=\linewidth]{sumifidmaps-crop}
\caption{Comparison of the predicted optical depth, $\tau_{200}$ of our fiducial model (bulge model M90, together with a disk of scale length $R_d=2.6\ensuremath{\,{\rm kpc}}\xspace$ and an inner scale height of $H_{4.5}=0.18\ensuremath{\,{\rm kpc}}\xspace$) compared to the MOA-II data corrected by the revision in \citep{Sumi:16}. The data and the model are averaged in the same manner: an Gaussian average with $\sigma=0.4\deg$ of fields within $1\deg$. We show from left to right comparison of the optical depth, the event rate per star, and the mean timescale $\langle \ensuremath{t_{\rm E}}\xspace \rangle$ in days. The optical depth and event rate maps show good qualitative agreement. The mean timescale map has large observational error due to the large statistical variance of timescale caused by the rare very long timescale events. The timescale map is better constrained by the larger sample of \citet[][{\it e.g.}\xspace Figures 12, 14 and 16]{Wyrzykowski:15} and we discuss its distribution in \autoref{sec:tedistdiscuss}. \label{fig:sumifidmaps} }
\end{figure*}
\section{Constraints from MOA-II Microlensing Data}
\label{sec:data}
In this section we compare our Milky Way models to the sample of microlensing data from the MOA-II survey \citep{Sumi:13}. Throughout we use the revised microlensing optical depth and rate data corrected for the effective number of monitored stars according to \citet{Sumi:16}.
Both the original \citep{Sumi:13} and the revised MOA-II data \citep{Sumi:16} split their sample of microlensing events an all-source sample, defined as sources with $I<20\ensuremath{\,{\rm mag}}\xspace$, and an red clump giant (RCG) sample restricted to $I<17.5\ensuremath{\,{\rm mag}}\xspace$. The selection criteria of MOA-II was optimised to detect short period events, however bright long timescale events are not efficiently detected: the longest period event in the RCG sample is only 55 days. Because these events make a significant contribution to the optical depth in the RCG sample we therefore consider only the all-source sample where the issue is not significant, as do \citet{Sumi:16}. The all-source sample contains events with $t<200\,{\rm days}$ and therefore throughout we compute and compare the data with the same cut. In \autoref{sec:eros} we compare to the EROS-II microlensing optical depth data measured by \citep{Hamadache:06} of bright stars around the red clump from to show the results are consistent.
\subsection{Microlensing Maps}
\label{sec:maps}
Here we compute maps of the optical depth, timescale and event rate of our fiducial model to compare to the maps in \cite{Sumi:16}.
We use a simple exponential model for the extinction:
\begin{equation}
\frac{dA_I}{ds} = a_I \times \exp ( -|z|/z_0 )
\end{equation}
with coefficients $a_I=0.5\ensuremath{\,{\rm mag}}\xspace\ensuremath{\,{\rm kpc}}\xspace^{-1}$ and $z_0=160\ensuremath{\,{\rm pc}}\xspace$ \citep{Nataf:13}. We have tried other values for these coefficients, and found that with the magnitude cuts selected by \citet{Sumi:13} the results are generally insensitive to these choices. For the latitudes probed by current microlensing experiments the source stars lie at $z \gg z_0$ and therefore extinction is largely a foreground screen. The changes due to extinction therefore arise mostly because extinction reduces the unextincted magnitude limit (see {\it e.g.}\xspace \autoref{fig:tauvsImag}). This therefore changes the optical depth by altering the range of magnitudes over which the measured optical depth is an average. Extinction has a much greater effect on the rate per unit area, and we therefore choose not to model this quantity.
We compute the microlensing properties of the models in each of the 1536 subfields of MOA-II using the procedure described in \autoref{sec:tetheory} to predict $\Gamma(\log \ensuremath{t_{\rm E}}\xspace, I)$. We then average the optical depth, weighted by the model magnitude distribution, over the range $14<I<20$ to represent the MOA-II magnitude range.
We exclude events with $\ensuremath{t_{\rm E}}\xspace>200\ensuremath{\,{\rm days}}\xspace$ when computing the optical depth {\it i.e.}\xspace we calculate $\tau_{200}$. Both the event rate and mean event timescale have no exclusion of $\ensuremath{t_{\rm E}}\xspace>200\ensuremath{\,{\rm days}}\xspace$ events. Maps comparing the fiducial model with the data of the resultant optical depth $\tau_{200}$, the event rate $\Gamma$, and the mean event timescale $\langle \ensuremath{t_{\rm E}}\xspace \rangle$ are shown in \autoref{fig:sumifidmaps}.
\subsection{Minor Axis Profile}
\label{sec:minoraxis}
\begin{figure}
\includegraphics[width=0.9\linewidth]{sumicomp-crop}
\caption{Comparison of the predicted microlensing properties for $|l|<5\deg$ with the revised MOA-II data. The upper panel shows the optical depth, and the lower the microlensing rate per star. The fiducial model, summarised in \autoref{tab:model}, is the solid black line. The coloured lines correspond to changing one feature of the fiducial model: blue to a constant $H=0.3\ensuremath{\,{\rm kpc}}\xspace$ scale height disk, green to bulge mode M80 with a lower stellar mass in the bulge, and red to a $R_d=2\ensuremath{\,{\rm kpc}}\xspace$ disk scale length. The orange line corresponds to the effect of changing to a \citet[][]{salpeter:55} IMF. The impact of the uncertain IMF on the rate leads us to constrain the models using the optical depth. \label{fig:minor_sumicomp}}
\end{figure}
The maps in \autoref{fig:sumifidmaps} show qualitatively good agreement. Here we use the minor axis profile of the microlensing data in the range $|l|<5\deg$ to make \emph{quantitative} comparison to Galactic models and hence determine which models are allowed by the data. The revised MOA-II data is taken from Table 2 of \citet{Sumi:16}. We take the maps computed from the models in the previous section and bin the $|l|<5\deg$ microlensing properties in the same manner as the data. The result for the fiducial model compared to the data is shown as the black solid line in \autoref{fig:minor_sumicomp}. We show both the optical depth and rate predictions. The fiducial model matches the optical depth well with a $\chi^2$ of $7.4$ over the 11 data points {\it i.e.}\xspace 0.7 per data point.
We also show the effect of variations to our fiducial parameters in \autoref{fig:minor_sumicomp}. In particular we show the effect of changing the IMF from our fiducial \citet{Kroupa:01} IMF, to a \citet{salpeter:55} IMF. This effects the event rate significantly, but hardly changes the optical depth. This is because the IMF directly alters the event timescale distribution and therefore directly alters the rates. However the IMF only enters the $\tau_{200}$ optical depth through the timescale distribution altering the contribution of events beyond the 200 day cutoff. For this reason we focus on constraints provided by the optical depth alone, and discuss the timescale distribution separately in \autoref{sec:tedistdiscuss}. As shown in that section our fiducial \citet{Kroupa:01} IMF has a slightly shorter timescale distribution than the observed. Using an IMF with less low mass stars matches the timescale distribution better, and would therefore also match better the rate profile data in the lower panel of \autoref{fig:minor_sumicomp}.
\subsection{Constraining the Model}
\label{sec:Rdconstaint}
In \autoref{fig:Rdconstaint} we show the $\chi^2$ between the data and model, computed using the data in \autoref{fig:minor_sumicomp}, for each bulge model as a function of disk scale length, $R_d$. The different lines correspond to variations of the bulge dark matter fraction and disk scale length. We consider only $R_d \ge 1.9\ensuremath{\,{\rm kpc}}\xspace$ since shorter disk scale lengths produce a baryonic rotation clearly larger than observed, with a peak $V_c>240\ensuremath{\,{\rm km}\,{\rm s}^{-1}}\xspace$ as computed in \autoref{sec:rotcurve}.
We consider the bulge dynamical models with different bulge dark matter fractions (M80-M90) and consider both the flared disk with our fiducial inner scale height of $H_{4.5}=0.18\ensuremath{\,{\rm kpc}}\xspace$ and a constant disk scale height of $H_{4.5}=H_\odot=0.3\ensuremath{\,{\rm kpc}}\xspace$. There is some degeneracy between disk scale length and bulge model: The bulge model with lower stellar matter, M80, requires short disk scale lengths of $(2.1 \pm 0.2)\ensuremath{\,{\rm kpc}}\xspace$ optical depths (at $1\sigma$). However model M85 with a larger stellar fraction would require $(2.3 \pm 0.2)\ensuremath{\,{\rm kpc}}\xspace$, and model M90 with the highest stellar mass fraction prefers even longer disks. This degeneracy occurs because the optical depth, \autoref{eq:tau}, measures the density of stellar matter towards the bulge, however for the same optical depth the lenses can be placed either in the bulge region or the foreground disk. We therefore consider the different bulge models and disk models together in the following sections.
The flared disks with inner scale height of $H_{4.5}=0.18\ensuremath{\,{\rm kpc}}\xspace$ are preferred at $>1\sigma$ over the constant scale height $H_{4.5}=0.3\ensuremath{\,{\rm kpc}}\xspace$ disks. This is because the constant scale height disks produce slightly flatter optical depth vs. latitude profiles than the data. For the same bulge model the constant disk scale height models require longer disk scale lengths. This is because most of the optical depth points lie higher than one disk scale height from the Galactic plane. Therefore a larger inner disk scale height places more mass at latitudes where it contributes to the optical depth and reduces the need for additional stellar mass in the disk.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{c2models-crop}
\caption{Allowed disk scale lengths as a function of $R_d$ for the revised MOA-II optical depth data. $\chi^2$ is calculated from the difference between model and data as in \autoref{fig:minor_sumicomp}. Different colours correspond to bulge models with different stellar masses M90 having the highest stellar mass and M80 the lowest. The dashed lines correspond to a constant $0.3\ensuremath{\,{\rm kpc}}\xspace$ scale height disk and the solid lines to a flared disk with inner scale height $H_{4.5}=0.30\ensuremath{\,{\rm kpc}}\xspace$. \label{fig:Rdconstaint}}
\end{figure}
\section{Comparison to EROS-II Microlensing Optical Depths}
\label{sec:eros}
In this section we consider if the models consistent with the revised MOA-II optical depth data are also consistent with the EROS-II optical depths. \citet{Hamadache:06} measured the EROS-II optical depth from a sample of bright resolved events. Only events magnifying stars in a selection box centred on the red clump in color-magnitude were considered. We replicate this in the models by computing the magnitude distribution in each of the 66 EROS-II bulge fields and in these fitting for the magnitude of the simulated red clump. We then compute the optical depth of stars within $1\, {\rm mag}$ of this to simulate the selection box in each field, and combine the optical depths across fields to simulate the optical depth vs latitude profile measured by \citet{Hamadache:06}.
The resultant profiles are shown compared to the data in \autoref{fig:minor_eroscomp}. The same fiducial model fits the EROS-II data with a $\chi^2=3.5$ across 5 data points, or 0.7 per data point. In \autoref{fig:erosRdconstaint} we show the models that are consistent with the data. Comparison with \autoref{fig:Rdconstaint} shows that the two data sets constrain the models similarly, and that models consistent with the MOA-II data are also consistent with the EROS-II data. The constraints provided by the EROS-II data are however weaker. This is a result of the smaller number of events considered: 120 in EROS-II vs 427 in MOA-II.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{eroscomp-crop}
\caption{Comparison of the predicted microlensing optical depth for $|l|<5\deg$ compared to the EROS-II data \citep{Hamadache:06}. The fiducial model is the black line and the other lines correspond to the same models as in \autoref{fig:minor_sumicomp}. \label{fig:minor_eroscomp}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{erosc2-crop}
\caption{The $\chi^2$ of the Galactic models as a function of $R_d$ for \citet{Hamadache:06} optical depth data. $\chi^2$ is calculated from data points as in the examples in \autoref{fig:minor_eroscomp}. The lines correspond to the same models as in \autoref{fig:Rdconstaint}. \label{fig:erosRdconstaint}}
\end{figure}
Initial microlensing studies yielded optical depths that were larger than theoretical predictions \citep{Udalski:94,Alcock:97}, and even theoretical bounds \citep{Binney:00}. Some these high optical depths resulted from an inadequate treatment of blending of faint stars, a problem that was not serious in the samples of brighter RCG stars \citep{Popowski:05}. The situation improved with DIA analysis and better treatment of blended stars. It is reassuring then that the models which match the revised MOA-II all-source data, also match the optical depths of the bright resolved RCG sample from \citet{Hamadache:06}, albeit with larger statistical errors associated with their smaller sample.
\section{Constraining the Galactic Rotation Curve with Microlensing Data}
\label{sec:rotcurve}
Each of the models consistent with the microlensing data, as computed in \autoref{sec:Rdconstaint}, has an associated stellar mass and, thus, stellar rotation curve. We include in addition a gas component of surface mass $13\,M_\odot\ensuremath{\,{\rm pc}}\xspace^{-2}$ with scale height 130\ensuremath{\,{\rm pc}}\xspace and scale length $2R_d$ \citep{Bovy:13}. We plot in \autoref{fig:rotcurve} the resultant rotation curves consistent with the revised MOA-II data and those consistent with the EROS-II data \citep[{\it cf.}\xspace Fig 16 of][]{Ortwin:araa}. We also plot the circular velocity taken from the compilation of gas dynamics of \citet{Sofue:09} scaled to $V_0=238\ensuremath{\,{\rm km}\,{\rm s}^{-1}}\xspace$ \citep{Schonrich:12,Reid:14} and $R_0=8.3\ensuremath{\,{\rm kpc}}\xspace$ \citep{Sotiris:14}.
When performing disk-halo decompositions in external disk galaxies there is a well known stellar mass-to-light vs. dark matter degeneracy \cite[{\it e.g.}\xspace][]{vanAlbada:85,Courteau:14}. The maximal disk hypothesis breaks this by assuming the disk contributes the maximum level possible by the data. Under this assumption typical disk contributions to the circular velocity at $R=2.2R_d$ in external disk galaxies are $f_v=(0.85\pm0.10)$ \citep{Sackett:97}. The point $R=2.2R_d$ is chosen as the comparison point since this is the location of the maximum of the rotation curve of a thin exponential disk.
The specific value used as the boundary between maximal and sub-maximal disks is arbitrary. Maximal disks necessarily contribute less than $f_v=1$ because the halo is significant at large radii, and its density cannot reasonably decrease inwards. However the lower end of $f_v=(0.85\pm0.10)$ given by \citet{Sackett:97} for maximal disk fits results from bulge-disk decompositions where the bulge is significant and not included in {\ensuremath{f_v}\xspace}~. Motivated by this, and following \citet{Courteau:14}, we consider ${\ensuremath{f_v}\xspace}=0.85$ as the boundary for maximal disk models.
Microlensing of Milky Way bulge stars offers a unique way to break the baryonic to dark matter degeneracy, and therefore estimate dark matter fractions and test the maximal disk hypothesis. However making a similar comparison for the Milky Way is complicated by the Galaxy not being a pure disk. Recently it has been realised that the bulge of the Milky Way is consistent with a minimal contribution from a classical bulge \citep[{\it e.g.}\xspace $<8\%$ of the disk mass by][]{Shen:10}. Instead the the bulge is a box/peanut bulge which formed via secular evolution from the disk. We therefore consider instead the fractional baryonic contribution at the peak of the baryonic rotation curve, since this seems most analogous to disk-halo decompositions in external galaxies \citep[as also advocated by][for external galaxies]{Courteau:14} and use the same boundary ${\ensuremath{f_v}\xspace}=0.85$ for maximality. The total circular velocity used for comparison is the weighted average of points within 500\ensuremath{\,{\rm pc}}\xspace of this computed similarly to \autoref{fig:rotcurve}.
For each model admitted by the data we compute this maximality: the fractional baryonic {\ensuremath{f_v}\xspace}. From the extremes of the models consistent with the revised MOA-II data at $1\sigma$ we find a maximality {\ensuremath{f_v=(0.88\pm0.07)}\xspace}. Performing the same process with the models consistent with the EROS-II data gives {\ensuremath{f_v=(0.9\pm0.1)}\xspace}~at $1\sigma$, the wider range of allowed curves resulting from the smaller number of microlensing events. These levels of baryonic contribution therefore place the Milky Way on the boundary between maximal disk fits and sub-maximal disks.
At $2\sigma$ a wider range of models are allowed. No useful upper limit on the maximality can be derived because very short disk scale lengths are allowed which would result in a baryonic contribution that even exceeds the total rotation curve. Instead only a lower limit of $>0.75$ can be given. However the models with high stellar to dark matter fractions in the bulge admit very long disk scale lengths since, for these bulges at $2\sigma$, the MOA-II optical depths can be reproduced almost by the bulge alone. This can result in baryonic rotation curves whose peak lies very close to the Galactic centre at $<2\ensuremath{\,{\rm kpc}}\xspace$, and which then slowly decline outwards. Computing the maximality at this position for these models is suboptimal for two reasons
\begin{inparaenum}
\item the circular velocity from \citet{Sofue:09} here is very uncertain here because the bar influences the gas orbits making them highly non-circular
\item the maximality is useful as a number that encodes the interplay between baryonic and dark matter in galaxies, however making the comparison at such a small fraction of the disk scale-length reduces its utility.
\end{inparaenum}
We therefore also consider a second Galactocentric radius for comparison. The models consistent with the data at $1\sigma$ have their baryonic peaks at $(4-5)\ensuremath{\,{\rm kpc}}\xspace$ from the Galactic centre. If we compute the maximality at $4\ensuremath{\,{\rm kpc}}\xspace$ the maximality constraint would be slightly lower: ${\ensuremath{f_v > 0.72}\xspace}$ at $2\sigma$ and this is the value we consider more robust. From EROS-II no useful constraints can be extracted at $2\sigma$.
\begin{figure}
\includegraphics[width=\linewidth]{hires_rotcurve}
\caption{Bands are the range of baryonic rotation curves of the models consistent with the revised MOA-II data \citep{Sumi:16} at 1 (red) and $2\sigma$ (pink). The baryonic rotation curves of the models consistent with the EROS-II data \citep{Hamadache:06} at $1\sigma$ level span the cross hatched area. The models plotted here have a range of bulge dark matter fraction (M80-M90), disk scale length ($R_d$) and inner disk scale height ($H_{4.5}$) but are all consistent with the microlensing optical depths. Data is taken from the compilation of gas kinematics by \citet{Sofue:09} scaled to $R_0=8.3\ensuremath{\,{\rm kpc}}\xspace$ \citep{Sotiris:14} and $V_0=238\ensuremath{\,{\rm km}\,{\rm s}^{-1}}\xspace$ \citep{Schonrich:12} (see \citealt{Ortwin:araa} for a fuller discussion of these values). \label{fig:rotcurve}}
\end{figure}
\section{Timescale Distribution}
\label{sec:tedistdiscuss}
Although the focus of this work are the constraints that surveys of microlensing in the bulge place on Galactic structure, which is best revealed by the optical depth, we briefly consider the timescale distribution in this section. While the optical depth has not yet been computed from the OGLE-III survey, the timescale distribution has by \citet{Wyrzykowski:15}. In the upper panel of \autoref{fig:tevssurvey} we show a comparison of this with the timescale distribution measured by MOA-II.
\begin{figure}
\includegraphics[width=\linewidth]{tevssurvey}
\caption{In the above panel we show the efficiency corrected timescale distribution of the OGLE-III survey (blue) compared to the MOA-II sample (red). We also show the best fitting log-normal distributions to the surveys as the curves. In the lower panel we compare the efficiency corrected OGLE-III timescale distribution to our fiducial model with different IMFs: A \citet{salpeter:55} IMF (short dashed line), a \citet{Zoccali:00} IMF (long dashes), a \citet{Kroupa:01} IMF (dash-dot) and the log-normal IMF of \citet{Calamida:15} (dotted). \label{fig:tevssurvey}}
\end{figure}
The two samples agree in their timescale distribution remarkably well. Fitting a log-normal distribution to each we find the mean log timescales are $<\log \ensuremath{t_{\rm E}}\xspace>=(1.275\pm0.008)$ and $(1.25\pm0.03)$ for the OGLE-III survey and MOA-II all source sample respectively, in statistical agreement. Likewise the standard deviation of the log timescales are $\sigma(\log \ensuremath{t_{\rm E}}\xspace)=(0.409\pm0.006)$ and $(0.41\pm0.02)$, also in agreement.
Note that a log-normal distribution does not capture the short timescale events present in the MOA-II sample. These are presumed to arise from microlensing by planetary mass objects \citep{Sumi:11}. It is also not expected to capture the asymptotic distribution of short or long timescale events \citep{Mao:96} but, given the current data, is still a useful parameterisation. This is particularly true since it provides the first two moments of the log timescale distribution. The log timescale distribution is a convolution of the two PDFs in \autoref{eq:teconvobv}, resulting from the dynamical model, and the lens mass distribution. Because of this the means and variances add (and indeed all cumulants) to give the mean and variance of the log timescale distribution. The mean log timescale and its variance are therefore straightforward to interpret and consider the effects of other mass distributions or dynamical models.
In the lower panel of \autoref{fig:tevssurvey} we show the timescale distribution predicted from our model for four different IMFs: those from \citet{salpeter:55} (a single power law with slope $\alpha=2.3$), a \citet{Zoccali:00} IMF ($\alpha=1.3$ for $M<1\ensuremath{{M_\odot}}\xspace$), a \citet{Kroupa:01} IMF, and log-normal IMF fitted for the bulge by \citet{Calamida:15} (log-normal with $M_c=0.25\ensuremath{{M_\odot}}\xspace$, $\sigma=0.5$). All IMFs use lower and upper mass limits of $0.01\ensuremath{{M_\odot}}\xspace$ and $100\ensuremath{{M_\odot}}\xspace$ respectively, aside from the \citet{salpeter:55} IMF which uses a lower limit of $0.1\ensuremath{{M_\odot}}\xspace$. For all we use the remnant prescriptions of \citet{Maraston:98} above the turnoff of a $10\ensuremath{\,{\rm Gyr}}\xspace$ population as described in \autoref{sec:pop}.
Given that we have not tailored the model to the timescale distribution it is reassuring that the \citet{Kroupa:01} and particularly the \citet{Calamida:15} IMFs together with the dynamical model reproduce the timescale distribution fairly well. There is a discrepancy at long timescales which the model under predicts. As a result, with the \citet{Calamida:15} IMF the model predicts $<\log \ensuremath{t_{\rm E}}\xspace>=1.21$ slightly less than the observed $<\log \ensuremath{t_{\rm E}}\xspace>=(1.275 \pm 0.008)$. The most uncertain region of the mass function is the brown dwarf region. That the short timescale distribution agrees with the OGLE-III data for the \citet{Calamida:15} log-normal IMF suggests that our model requires a low number of brown dwarfs to be present in the bulge compared to rising power law IMFs like the \citet{Kroupa:01} IMF, similar to local estimates \citep[{\it e.g.}\xspace][]{Andersen:08}.
More detailed comparison is beyond the scope of this work and would require consideration of variations of the kinematical model, particularly outside the central region where the N-body model was constrained by \citet{Portail:15}. We are presently constructing of made-to-measure models of the entire inner Galaxy, including the inner disk, and we will model the timescale distribution and place more robust constraints on the IMF when that is complete. We emphasise that the results outside this section are based on the optical depth, which is insensitive to the timescale distribution. This is demonstrated in \autoref{sec:minoraxis} by the insensitivity of the optical depth to the much shorter timescales produced by the \citet{salpeter:55} IMF.
\section{Discussion}
\label{sec:discuss}
\subsection{A high baryonic fraction in the inner Milky Way}
The quantity best constrained by the microlensing optical depth is the stellar mass per unit solid angle towards the bulge and inner Galaxy. We find that there is a degeneracy in where that mass is placed between the bulge and the foreground disk: models with lower stellar fractions in the bulge require more prominent foreground disks and vice versa.
The scale length of the Milky Way disk is highly uncertain. Earlier estimates using optical wavelengths tended to favour larger scale lengths \citep[{\it e.g.}\xspace 3.5\ensuremath{\,{\rm kpc}}\xspace by][]{Bahcall:80}, while more recent measurements in the NIR and large photometric parallax surveys have favoured lower measurements typically in the range 2-3\ensuremath{\,{\rm kpc}}\xspace ( {\it e.g.}\xspace \citealt{Binney:97,Bissantz:02,Robin:03,Juric:08,Bovy:13}; see \citealt{Ortwin:araa} for a fuller discussion).
Despite this uncertainty, in all the models which reproduce the revised MOA-II optical depths, the baryonic contribution to the rotation curve at its peak is high: {\ensuremath{f_v=(0.88\pm0.07)}\xspace}~at $1\sigma$ and ${\ensuremath{f_v > 0.72}\xspace}$ at $2\sigma$. These are consistent with the constraints from EROS-II of {\ensuremath{f_v=(0.9\pm0.1)}\xspace}. As discussed in \autoref{sec:rotcurve}, we consider maximal disks to have ${\ensuremath{f_v}\xspace} > 0.85$ \citep{Courteau:14}. These levels of baryonic contribution therefore place the Milky Way on the boundary between maximal disk fits and sub-maximal disks .
The DiskMass survey has measured the vertical velocity dispersions of nearly face on disk galaxies. Together with the statistically determined scale height from a sample of side on galaxies, this directly measures the disk mass under the assumption of a locally isothermal disk. They find that their sample is generally sub-maximal with a fractional baryonic contribution at $2.2R_d$ of $0.57\pm0.07$ \citep{Martinsson:13}. However the measured light weighted stellar kinematics will be biased towards younger stars as opposed to old dynamically relaxed populations. Measuring the vertical kinematics of the Milky Way locally suggests that this correction could be a factor of $\sim 2$ \citep{Aniyan:16}. In addition the resultant IR mass-to-light of the disk mass survey is also a factor of $\sim 2$ smaller than that estimated through other methods \citep{McGaugh:15}. The derived maximal disk in this work would therefore not necessarily make the Milky Way unusual in the context of external disk galaxies. In addition in external barred galaxies a further possibility exists to break the baryonic to dark matter degeneracy. The stellar distribution can be determined from the light, while the non-circular gas motions constrain the shape of the effective potential, and therefore the more spherical dark matter, in the bar. Using this method several studies have found maximal or near maximal disks in the inner regions of barred galaxies \citep{Weiner:01,Weiner:04,Perez:04,Sanchez:08}.
\subsection{Consequences for the Milky Way's dark halo}
\begin{figure}
\includegraphics[width=\linewidth]{rotadiabaticgrey}
\caption{We show the range of baryonic rotation curves consistent with the revised MOA-II data at the $1\sigma$ level together with a selection of possible dark matter haloes. The baryonic rotation curves are shown as the cross hatched area, the dark matter profiles as the black lines, and the resultant range of profiles of the total rotation curve in grey. The top left panel shows an NFW profile with concentration $c_{200}=9.0$. The upper right an NFW halo, also with $c_{200}=9.0$, adiabatically contracted through the prescription \citet{Blumenthal:86}. The lower left panel shows the contraction suggested by \citet{Gnedin:04} for a smaller but still cosmologically possible initial $c_{200}=7.0$, and the lower right the contraction fitted by \citet{Abadi:10} for $c_{200}=9.0$. \label{fig:rotadiabatic} }
\end{figure}
The high baryonic contribution to the rotation curve in the inner Galaxy seen in \autoref{fig:rotadiabatic}, and the resultant high levels of disk maximality, are not however inconsistent with the current understanding of the $\Lambda$CDM paradigm. An NFW dark matter halo with a Milky Way halo mass $M_{\rm 200}=1.1 \times 10^{12} M_\odot$ and radius $R_{\rm 200}=270\ensuremath{\,{\rm kpc}}\xspace$ \citep{Ortwin:araa} and a concentration of $c_{\rm 200}=9.0$ motivated by cosmological simulations (\citealt{Correa:15c} with Planck 2015 cosmology \citealt{PlanckCosmo:15}) gives the profile shown in the top left panel of \autoref{fig:rotadiabatic}. Adding this to the range of baryonic contributions allowed by the revised MOA-II data at $1\sigma$ gives rotation curves that remain below the total rotation curve.
\begin{figure*}
\includegraphics[width=0.7\linewidth]{fullmaps}
\caption{The microlensing optical depth of the fiducial model across the entire Milky Way bulge (top left), compared to the same variants considered earlier in the paper ({\it e.g.}\xspace in \autoref{fig:minor_sumicomp}). We show the effect of reducing the disk scale length to $R_d=2\ensuremath{\,{\rm kpc}}\xspace$ from $R_d=2.6\ensuremath{\,{\rm kpc}}\xspace$ in the top right, using a bulge model with a higher dark matter fraction and lower stellar mass in the bulge (M80) in the lower left, and a model with a constant scale height of $H_{4.5}=0.3\ensuremath{\,{\rm kpc}}\xspace$ in the lower right. All maps are the average optical depth over stars with $14<I_0<19$. \label{fig:fullmaps}}
\end{figure*}
Adiabatically contracted versions of this halo using the prescription of \citet{Blumenthal:86} are generally not consistent with the revised MOA-II data at the $1\sigma$ level. In this prescription the adiabatic invariant is $r M(<r)$ which results in rotation curves that overshoot that observed in the inner Galaxy (top right panel \autoref{fig:rotadiabatic}). This is true even for lower concentrations as small as $c_{\rm 200}=7.0$, which is the expected cosmological halo-to-halo scatter at $1\sigma$ level \citep{Maccio:08}. It is also true for the contraction detirmined by \citet{Piffl:15} since it results in more contracted haloes than \citet{Blumenthal:86}. \citet{Binney:15} similarly concluded that halos with this level of contraction, in conjunction with the local dark matter density, could not both be consistent with the Milky Way's rotation curve and earlier microlensing optical depths.
However other milder contraction prescriptions are consistent with the revised MOA-II data at the $1\sigma$ level. The prescription of \citet{Gnedin:04} is consistent only if the initial halo has lower concentrations ($c_{\rm 200}=7.0$ is shown in the lower left panel of \autoref{fig:rotadiabatic}). The prescription of \citet{Abadi:10} results in less contracted halos, and as a result is consistent with an initial $c_{\rm 200}=9.0$ (lower right panel of \autoref{fig:rotadiabatic}). All haloes were contracted using a baryonic exponential scale length of $2.6\ensuremath{\,{\rm kpc}}\xspace$ and contraction of the \citet{Gnedin:04} and \citet{Abadi:10} halos was implemented through the $\Gamma$ prescription of \citet{Dutton:07,Dutton:11} with $\Gamma=0.8$ and $0.4$ respectively. All contraction prescriptions can be consistent with the range of rotation curves allowed by the MOA-II data at the $2\sigma$ level, and the EROS-II data.
It has been shown that strong feedback could alter the central profiles of Milky Way sized haloes \citep{Maccio:11}. The results here suggest that it is not \emph{required} in the centre of Milky Way sized galaxy haloes in a similar way to dwarf galaxies. The microlensing results however do not rule it out, for example the feedback in \citet{Chan:15} for $10^{12} M_\odot$ mass haloes results in central profiles similar to the pre-contracted NFW haloes which would still be consistent. We would expect our results to also be consistent with other recent simulations \citep[{\it e.g.}\xspace][]{Marinacci:14,Schaller:15} which generally produce only mildly contracted haloes.
We conclude that the levels of baryonic contribution towards the inner Galaxy, {\ensuremath{f_v}\xspace}, are high, but not inconsistent with the estimates of contribution of CDM haloes in the inner region of disk galaxies, provided the level of baryonic contraction is not too extreme.
There is however possible tension between the low baryonic fractions in the inner galaxy measured here, with recent estimates of the local dark matter density. \citet{Piffl:14} measured the local dark matter density to be $0.0126 q^{-0.89} \ensuremath{{M_\odot}}\xspace \ensuremath{\,{\rm pc}}\xspace^{-3} \pm 15\%$ in agreement with $0.014 \ensuremath{{M_\odot}}\xspace \ensuremath{\,{\rm pc}}\xspace^{-3}$ measured by \citet{Bienayme:14}. These more recent measurements utilising the exquisite RAVE data are higher than most over that past decade \citep[see][and references therein]{Read:14}. The haloes shown in \autoref{fig:rotadiabatic} however have local dark matter densities in the range $0.005-0.008 \ensuremath{{M_\odot}}\xspace pc^{-3}$.
The NFW profile considered by \citet{Piffl:14} would be consistent with the range of models considered here and would reproduce also local dark matter density. However this model had an uncomfortably high concentration of $c=20$, significantly higher than expected in dark matter only simulations for Milky Way sized halos. Instead it seems more natural that a less concentrated halo experienced a mild degree of contraction due to growth of the baryonic disk, and this increased the dark matter density at $\sim R_0$. Some tension remains however between the high baryonic fraction in the inner Galaxy measured here, the local dark matter density, and even mild adiabatic contraction prescriptions. Contracting a halo with $c_{200}=15.$, towards the upper end of the cosmological values, with the prescription of \citet{Abadi:10} results in rotation curves just consistent with the microlensing results here. This however gives a solar dark matter density of $0.01 \ensuremath{{M_\odot}}\xspace pc^{-3}$, still somewhat lower than the estimates with RAVE. We therefore encourage direct comparison between the high baryonic fraction in the inner Galaxy required by the microlensing here, the local dark matter density, and recent simulations of Milky Way like galaxies in the $\Lambda$CDM paradigm such as \citet{Marinacci:14}.
\subsection{Microlensing as a tool for Galactic structure}
The measured optical depths close to the plane are particularly important in driving the need for high stellar contributions in the inner Galaxy and low dark matter fractions, because the majority of the stellar mass lies here. It is therefore important to confirm the optical depth measurements in this region.
We anticipate that the constraints in this work could be verified and improved on by the larger sample of 3560 events, and the better detection efficiency of longer events, provided by the OGLE-III sample \citep{Wyrzykowski:15}. Of these 2047 lie inside $|b|<3\deg$ and 546 inside $|b|<2\deg$, therefore the optical depths measured by MOA-II in this region can be verified. The optical depth and event rates require detailed field-by-field efficiency calculations that are not yet complete, but we encourage their computation.
Microlensing in the Milky Way is a unique tool to break the dark-matter to baryonic matter degeneracy, however since a large fraction of the mass of the disk lies at these low latitudes measurements are essential in this challenging region. A further important tool to this goal will be microlensing surveys in the NIR. To help guide the design and assess the impact of future surveys we provide in \autoref{fig:fullmaps} maps of the optical depth across the entire Galactic bulge.
Recently \citet{Awiphan:16} also modelled the MOA-II data within the Besan\c on model, finding optical depths smaller than the observed data, particularly close to the Galactic plane. Some of this discrepancy was resolved by the revised MOA-II data, but the optical depths are remain lower than those observed. The dynamical models used in this work have higher bulge mass than those in \citet{Awiphan:16} which resolves the discrepancy. We note that while the bulge dynamical models do not include a nuclear bulge component inside $1\deg$ \citep{Launhardt:02}, since this is significant only inside $|b|<1\deg$ it does not effect the MOA-II measurements which lie outside this.
The optical depth is fundamentally a weighted mean of the density of lenses along the line-of-sight towards each source star (see {\it e.g.}\xspace \autoref{eq:tau}). Because of this the microlensing data of the bulge constrains the stellar mass distribution towards centre of the Galaxy. The weighting of this density measurement is most sensitive to events halfway between the source and lens, and because of our position in the Galaxy is weighted to be sensitive to the stellar density at $R \sim 4\ensuremath{\,{\rm kpc}}\xspace$. Since this is likely to be near the position of the peak in the stellar rotation it makes the microlensing optical depth an attractive method to measure the disk maximality. A disadvantage is that since there is effectively only one line-of-sight (that towards the bulge) assumptions on the form of the density are needed to convert the optical depth to a stellar density constraint. In this work we assume the disk is axisymmetric, however this assumption is likely to be violated to some extent, and certainly inside the radius of the Galactic bar of $(5.0\pm0.2)\ensuremath{\,{\rm kpc}}\xspace$ \citep{Wegg:15}. Our fiducial model has a disk mass between 2.2\ensuremath{\,{\rm kpc}}\xspace and 5\ensuremath{\,{\rm kpc}}\xspace of $2.1\times10^{10}\ensuremath{{M_\odot}}\xspace$ while the mass of the bar outside the bulge derived by \citet{Wegg:15} is $1\times10^{10}\ensuremath{{M_\odot}}\xspace$. As a result we expect a significant fraction of the mass in this range may be part of the bar and therefore non-axisymmetric. Fortunately our line-of-sight to the bulge lies at an angle $(27\pm 2)\deg$ to the bar \citep{Wegg:13}, and as a result we are neither looking along or perpendicular to the bar. In addition the microlensing optical depth could be sensitive to the presence of a ring or trailing structures at the end of the bar, since this would lie almost halfway to the bulge, where the microlensing is most sensitive to the density of lenses.
We note that microlensing models of the bulge, such as that presented here, are important not just for the constraints they provide on the galaxy. Planet detection through microlensing has become as important tool to probe distant and low mass planets, and this is an important component of the upcoming EUCLID and WFIRST missions \citep{Beaulieu:11}. The expected yields and targeting requires microlensing models such as those described in this work.
\section{Conclusions}
\label{sec:conc}
Our fiducial model does an excellent job of predicting the optical depth, rates, and timescale distributions of both the MOA-II and EROS-II microlensing data. This model, constructed by adding an axisymmetric disk to the N-body bulges of \citet{Portail:15}, has a disk scale length of 2.6\ensuremath{\,{\rm kpc}}\xspace and a bulge dark matter fraction of 12\%.
Our dynamical models also match the timescale distribution of microlensing events in the OGLE-III survey very well for a \citet{Kroupa:01} or particularly a \citet{Calamida:15} log-normal IMF. Preliminary investigations suggest that a low number of brown dwarfs is required to in order to not overestimate the number of short duration events in the OGLE-III survey.
Microlensing in the Milky Way is a unique tool for breaking the degeneracy between dark matter and stellar mass-to-light. It is sensitive to the stellar density in the inner Milky Way where the dark matter contribution is most uncertain, and where the interplay between baryonic and dark matter are expected to be strongest.
By varying the bulge stellar to dark matter fraction and the disk scale length and height we find a range of models consistent with the revised MOA-II data. In particular there is a degeneracy between the amount of stellar matter in the bulge and in the foreground disk. However the resultant rotation curves allowed by our models which match the microlensing data require a high baryonic fraction and permit only a limited amount of dark matter in the inner galaxy. The disk maximality, defined as the baryonic contribution to the rotation curve at its peak, is {\ensuremath{f_v=(0.88\pm0.07)}\xspace}\ at $1\sigma$, and $>72\%\xspace$ at $2\sigma$. Maximal disk fits in external galaxies find $(0.85\pm0.10)$ \citep{Sackett:97} and so this places the Milky Way near the boundary between maximal and sub-maximal disks.
These high baryonic fractions in the inner Galaxy, and high levels of disk maximality are consistent with the NFW profiles predicted by pure $\Lambda$CDM simulations of Milky Way like halos. The $1\sigma$ bounds from the revised MOA-II data are inconsistent with strongly adiabatically contracted haloes. However more recent results suggest less strongly adiabatically haloes. Together with lower concentrations in line with the expected halo-to-halo concentration scatter produces haloes consistent with the data without the need for strong feedback altering the inner profiles of Milky Way sized haloes.
Our results are driven by the optical depths in the revised MOA-II sample. The results in this work could be confirmed and improved though the estimation of optical depths from the larger OGLE-III survey, particularly close to the Galactic plane where much of the mass of the inner disk lies.
\input{maximality.bbl}
\bsp
\label{lastpage}
\end{document}
|
\section*{Introduction}
\markright{\thesection.~Introduction}
In the present paper we investigate geometric characteristics of compact metric spaces, which can be described in terms of Gromov--Hausdorff distances to simplexes, i.e., to finite metric spaces such that all their nonzero distances are equal to each other. In~\cite{TuzMST-GH} these distances were used to calculate the length of edges of a minimum spanning tree constructed on a finite metric space. In the present paper we generalize the results from ~\cite{TuzMST-GH} to the case of arbitrary compact metric spaces. It turns out that these Gromov--Hausdorff distances depend on some geometrical characteristics of finite partitions of compact metric spaces; some of the characteristics can be considered as a natural analogue of the lengths of edges of minimum spanning trees. We calculate the Gromov--Hausdorff distances from an arbitrary compact metric space to a simplex of sufficiently small or sufficiently large diameter, see Theorems~\ref{thm:spec}, \ref{thm:small1}, and \ref{thm:small2}. For a finite $n$-point metric space we find the distances to an arbitrary simplex consisting of at least $n-1$ points (Theorems~\ref{thm:dist-n-simplex-same-dim}, \ref{thm:dist-n-simplex-bigger-dim}, and \ref{thm:minus_one}). Nevertheless, the general problem of calculating the distance from an arbitrary metric space to an arbitrary simplex remained unsolved yet. We demonstrate non-triviality of the problem by presenting a few examples in the end of the paper. In particular, we show that the set of all distances from a compact metric space to all simplexes is not a metric invariant, i.e., such collections can coincide for non-isometric finite metric spaces. Moreover, we construct an example
of infinite (continuum) set of pairwise non-isometric finite metric spaces having the same collection of those distances.
In this paper we use the technique of irreducible optimal correspondences~\cite{IvaNikolaevaTuz, IvaIliadisTuz, IvaTuzIrreducible}. We show that to calculate the Gromov--Hausdorff distance from a compact metric space $X$ to an $n$-point simplex, where $n$ is less than or equal to the cardinality of $X$, one can consider only those correspondences which generates partitions of the space $X$ into $n$ nonempty disjoint subsets, see Theorem~\ref{thm:m_less_than_n_corresp}. This result has enabled us to advance essentially in calculations of concrete Gromov--Hausdorff distances.
\section{Preliminaries}
\markright{\thesection.~Preliminaries}
For an arbitrary set $X$ by $\#X$ we denote its \emph{cardinality}.
Let $X$ be an arbitrary metric space. The distance between its points $x$ and $y$ is denoted by $|xy|$. If $A,B\subset X$ are nonempty, then we put $|AB|=\inf\bigl\{|ab|:a\in A,\,b\in B\bigr\}$. If $A=\{a\}$, then we write $|aB|=|Ba|$ instead of $|\{a\}B|=|B\{a\}|$.
For a point $x\in X$ and a real number $r>0$ by $U_r(x)$ we denote the open ball of radius $r$ centered at $x$; for any nonempty $A\subset X$ and $r>0$ we put $U_r(A)=\cup_{a\in A}U_r(a)$.
\subsection{Hausdorff and Gromov--Hausdorff distances}
For nonempty $A,\,B\subset X$ we put
$$
d_H(A,B)=\inf\bigl\{r>0:A\subset U_r(B)\ \&\ B\subset U_r(A)\bigr\}=\max\{\sup_{a\in A}|aB|,\,\sup_{b\in B}|Ab|\}.
$$
This value if called the \emph{Hausdorff distance between $A$ and $B$}. It is well-known~\cite{BurBurIva} that the Hausdorff distance is a metric on the family of all nonempty closed bounded subsets of $X$.
Let $X$ and $Y$ be metric spaces. A triple $(X',Y',Z)$ that consists of a metric space $Z$ and its subsets $X'$ and $Y'$ isometric to $X$ and $Y$, respectively, is called a \emph{realization of the pair $(X,Y)$}. The \emph{Gromov--Hausdorff distance $d_{GH}(X,Y)$ between $X$ and $Y$} is the infimum of real numbers $r$ such that there exists a realization $(X',Y',Z)$ of the pair $(X,Y)$ with $d_H(X',Y')\le r$. It is well-known~\cite{BurBurIva} that the $d_{GH}$ is a metric on the family ${\cal M}$ of isometry classes of compact metric spaces.
For various calculations of the Gromov--Hausdorff distances, the technique of correspondences is useful.
Let $X$ and $Y$ be arbitrary nonempty sets. Recall that a \emph{relation\/} between the sets $X$ and $Y$ is a subset of the Cartesian product $X\times Y$. By ${\cal P}(X,Y)$ we denote the set of all \textbf{nonempty\/} relations between $X$ and $Y$. Let us look at each relation $\sigma\in{\cal P}(X,Y)$ as at a multivalued mapping, whose domain may be less than $X$. Then, similarly with the case of mappings, for any $x\in X$ and any $A\subset X$ their images $\sigma(x)$ and $\sigma(A)$ are defined, and for any $y\in Y$ and any $B\subset Y$ their preimages $\sigma^{-1}(y)$ and $\sigma^{-1}(B)$ are also defined.
A relation $R\in{\cal P}(X,Y)$ is called a \emph{correspondence}, if the restrictions of the canonical projections $\pi_X\:(x,y)\mapsto x$ and $\pi_Y\:(x,y)\mapsto y$ onto $R$ are surjective. By ${\cal R}(X,Y)$ we denote the set of all correspondences between $X$ and $Y$.
Let $X$ and $Y$ be arbitrary metric spaces. The \emph{distortion $\operatorname{dis}\sigma$ of a relation $\sigma\in{\cal P}(X,Y)$} is the value
$$
\operatorname{dis}\sigma=\sup\Bigl\{\bigl||xx'|-|yy'|\bigr|: (x,y),(x',y')\in\sigma\Bigr\}.
$$
\begin{prop}[\cite{BurBurIva}]
For any metric spaces $X$ and $Y$ we have
$$
d_{GH}(X,Y)=\frac12\inf\bigl\{\operatorname{dis} R:R\in{\cal R}(X,Y)\bigr\}.
$$
\end{prop}
For finite metric spaces $X$ and $Y$ the set ${\cal R}(X,Y)$ is finite as well, therefore there always exists an $R\in{\cal R}(X,Y)$ such that $d_{GH}(X,Y)=\frac12\operatorname{dis} R$. Every such correspondence $R$ is called \emph{optimal}. Notice that the optimal correspondences exist also for any compact metric spaces $X$ and $Y$, see~\cite{IvaIliadisTuz}. The set of all optimal correspondences between $X$ and $Y$ is denoted by ${\cal R}_{\operatorname{opt}}(X,Y)$. Thus, for compact metric spaces $X$ and $Y$ we have ${\cal R}_{\operatorname{opt}}(X,Y)\ne\emptyset$.
The inclusion relation generates a partial order on ${\cal R}(X,Y)$: $R_1\le R_2$, iff $R_1\subset R_2$. The relations minimal with respect to this order are called \emph{irreducible}, and the remaining ones are referred as \emph{reducible}. In~\cite{IvaTuzIrreducible} it is proved that for any compact metric spaces $X$ and $Y$ there always exists an irreducible optimal correspondence $R$. By ${\cal R}_{\operatorname{opt}}^0(X,Y)$ we denote the set of all irreducible optimal correspondences between $X$ and $Y$. As it was mentioned above, ${\cal R}_{\operatorname{opt}}^0(X,Y)\ne\emptyset$.
The next well-known facts can be easily proved by means of the correspondences technique. For any metric space $X$ and any positive real $\lambda>0$ let $\lambda X$ stand for the metric space which is obtained from $X$ by multiplication of all the distances by $\lambda$.
\begin{prop}[\cite{BurBurIva}]\label{prop:GH_simple}
Let $X$ and $Y$ be metric spaces. Then
\begin{enumerate}
\item\label{prop:GH_simple:1} If $X$ is a single-point metric space, then $d_{GH}(X,Y)=\frac12\operatorname{diam} Y$\rom;
\item\label{prop:GH_simple:2} If $\operatorname{diam} X<\infty$, then
$$
d_{GH}(X,Y)\ge\frac12\big|\operatorname{diam} X-\operatorname{diam} Y\big|;
$$
\item\label{prop:GH_simple:3} $d_{GH}(X,Y)\le\frac12\max\{\operatorname{diam} X,\operatorname{diam} Y\}$, in particular, for bounded $X$ and $Y$ it holds $d_{GH}(X,Y)<\infty$\rom;
\item\label{prop:GH_simple:4} For any $X,Y\in{\cal M}$ and any $\lambda>0$ we have $d_{GH}(\lambda X,\lambda Y)=\lambda d_{GH}(X,Y)$. Moreover, for $\lambda\ne1$ the unique invariant space is the single-point one. In other words, the multiplication of a metric by $\lambda>0$ is a homothety of ${\cal M}$ centered at the single-point space.
\end{enumerate}
\end{prop}
\subsection{A few elementary relations}
The next relations will be useful for concrete calculations of Gromov--Hausdorff distances.
\begin{prop}\label{prop:max_abs}
For any nonnegative $a$ and $b$ the following inequality holds\rom:
$$
\max\big\{a,|b-a|\big\}\le\max\{a,b\}.
$$
\end{prop}
\begin{proof}
Indeed, if $b\ge a$, then $|b-a|=b-a\le b$. If $b\le a$, then $|b-a|=a-b\le a$.
\end{proof}
\begin{prop}\label{prop:many_abs_dif}
Let $A\subset\mathbb R$ be a nonempty bounded subset, and let $t\in\mathbb R$. Then
$$
\sup_{a\in A}|t-a|=\max\{t-\inf A,\sup A-t\}=\Big|t-\frac{\inf A+\sup A}{2}\Big|+\frac{\sup A-\inf A}{2}.
$$
\end{prop}
\begin{proof}
Consider the segment $[\inf A,\sup A]$. If $t$ is placed to the left side of the segment middle point, i.e., $t\le(\inf A+\sup A)/2$, then the value $\sup_{a\in A}|t-a|$ is achieved at the right end of the segment, i.e., it is equal to $\sup A-t$; also, $t-\inf A\le|t-\inf A|\le\sup A-t$. Therefore, for such $t$ the proposition holds. One can similarly consider the case $t\ge(\inf A+\sup A)/2$.
\end{proof}
\begin{prop}\label{prop:many_abs_dif_and_t}
Let $A\subset\mathbb R$ be a nonempty bounded subset, $\inf A\ge0$, and let $t\in\mathbb R$. Then
$$
\sup_{a\in A}\big\{t,|t-a|\big\}=\max\{t,\sup A-t\}.
$$
\end{prop}
\begin{proof}
Let $A'=\{0\}\cup A$, then $\sup_{a\in A'}|a-t|=\sup_{a\in A}|a-t|$. By Proposition~\ref{prop:many_abs_dif}, we have
$$
\sup_{a\in A'}|a-t|=\max\{t,\sup A-t\}.
$$
\end{proof}
\section{Minimum Spanning Trees}
\markright{\thesection.~Minimum spanning trees}
To calculate the Gromov--Hausdorff distance between finite metric spaces, minimum and maximum spanning trees turn out to be useful. Also, the edges lengths of these trees turn out to be closely related to some geometrical properties of various partitions of the ambient space.
Let $G=(V,E)$ be an arbitrary (simple) graph with the vertex set $V$ and the edge set $E$. If $V$ is a metric space, then the \emph{length $|e|$ of edge $e=vw$ of the graph $G$} is defined as the distance $|vw|$ between the ending vertices $v$ and $w$ of this edge; also, the \emph{length $|G|$ of the graph $G$} is defined as the sum of all its edges lengths.
Let $M$ be a finite metric space. We define the number $\operatorname{mst}(M)$ as the length of the shortest tree of the form $(M,E)$. This value is called the \emph{length of minimum spanning tree on $M$}; a tree $G=(M,E)$ such that $|G|=\operatorname{mst}(M)$ is called a \emph{minimum spanning tree on $M$}. Notice that for any $M$ there exists a minimum spanning tree on it. The set of all minimum spanning trees on $M$ is denoted by $\operatorname{MST}(M)$.
\subsection{$\operatorname{mst}$-spectrum of a finite metric space}
Notice that minimum spanning tree may be defined not uniquely. For $G\in\operatorname{MST}(M)$ by $\sigma(G,M)$ we denote the vector, whose components are the lengths of edges of the tree $G$, ordered descendingly. If it is clear which metric is used, then we write $\sigma(G)$ instead of $\sigma(G,M)$. The next result is well-known.
\begin{prop}\label{prop:mst-spect}
For any $G_1,G_2\in\operatorname{MST}(M)$ we have $\sigma(G_1)=\sigma(G_2)$.
\end{prop}
Proposition~\ref{prop:mst-spect} explains correctness of the following definition.
\begin{dfn}
For any finite metric space $M$, by $\sigma(M)$ we denote the vector $\sigma(G,M)$ for arbitrary $G\in\operatorname{MST}(M)$, and we call this vector by the \emph{$\operatorname{mst}$-spectrum of the space $M$}.
\end{dfn}
\begin{constr}
For any set $M$ by ${\cal D}_k(M)$ we denote the family of all possible partitions of the $M$ into $k$ its nonempty subsets. Suppose now that $M$ is a metric space and $D=\{M_1,\ldots,M_k\}\in{\cal D}_k(M)$. Put
$$
\alpha(D)=\min\bigl\{|M_iM_j|:i\ne j\bigr\}.
$$
\end{constr}
The next result is proved in~\cite{TuzMST-GH}.
\begin{prop}\label{prop:spect-calc}
Let $M$ be a finite metric space and $\sigma(M)=(\sigma_1,\ldots,\sigma_{n-1})$. Then
$$
\sigma_k=\max\bigl\{\alpha(D):D\in{\cal D}_{k+1}(M)\bigr\}.
$$
\end{prop}
\subsection{$\operatorname{mst}$-spectrum of an arbitrary metric space}\label{subsec:mst-spector}
Now we generalize the concept of $\operatorname{mst}$-spectrum by means of Proposition~\ref{prop:spect-calc}.
\begin{dfn}
For any metric space $X$ and $k\in{\mathbb N}$ we put $\sigma_k=\sup\{\alpha(D):D\in{\cal D}_{k+1}(X)\}$ if ${\cal D}_{k+1}(X)\ne\emptyset$, and $\sigma_k=0$ otherwise. We call the set $\sigma(X)=\{\sigma_1,\sigma_2,\ldots\}$ by \emph{$\operatorname{mst}$-spectrum of $X$}.
\end{dfn}
\begin{rk}
If $\#X=n$, then $\sigma_k=0$ for $k\ge n$.
\end{rk}
In~\cite{IT_MST}, for metric spaces which can be connected by a finite length tree (see~\cite{INT_MST} for definitions), a necessary condition of minimum spanning trees existence is obtained. It follows from~\cite[Theorem~1]{IT_MST} that for a metric space $X$ which can be connected by a minimum spanning tree $G$ of a finite length, all the edges of $G$ are exact in the following sense. For each edge $e$ and the corresponding vertex sets $X_1$ and $X_2$ of the trees forming the forest $G\setminus e$, the length of $e$ equals to the distance between $X_1$ and $X_2$.
Moreover, for $G$ described above and any $\delta>0$ there are finitely many edges of $G$, whose lengths are more than or equal to $\delta$. This enables us to order the edges in such a way that their lengths decrease monotonically. Let $\{e_1,e_2,\ldots\}$ be such an order, and put $r_i=|e_i|$.
\begin{lem}\label{lem:spec_inf}
For any positive integer $k$ consider the partition $D=\{X_1,\ldots,X_{k+1}\}$ of $X$ into vertex sets of the trees forming the forest $G\setminus\{e_1,\ldots,e_k\}$. Then $\alpha(D)=|e_k|$.
\end{lem}
\begin{proof}
Each $p_i\in X_i$ and $p_j\in X_j$, $i\ne j$, are connected by a path in $G$, and the path contains at least one of the edges $e_p$, $p\le k$. Since $G$ is minimal, then $|p_ip_j|\ge|e_p|\ge|e_k|$, thus $|X_iX_j|\ge|e_k|$ and, therefore, $\alpha(D)\ge|e_k|$. On the other hand, if we choose $X_i$ and $X_j$ in such a way that $e_k$ connects them, then $\alpha(D)\le|X_iX_j|=|e_k|$, because the edge $e_k$ is exact.
\end{proof}
\begin{lem}\label{lem:spec_inf_2}
Let $D'=\{X'_1,\ldots,X'_{k+1}\}$ be an arbitrary partition of $X$, then $\alpha(D)\le\alpha(D')$.
\end{lem}
\begin{proof}
Denote by $E'$ the set of all edges of $G$ connecting different elements of partition $D'$. The set $E'$ contains at least $k$ edges (otherwise, the graph $G$ is disconnected), hence $\min_{e\in E'}|e|\le|e_k|$. Let $e'\in E$ satisfy $|e'|=\min_{e\in E'}|e|$, and let $X'_i$ and $X'_j$ be those elements of $D'$ which are connected by $e'$. Then $\alpha(D')\le|X'_iX'_j|\le|e'|\le|e_k|=\alpha(D)$.
\end{proof}
Lemmas~\ref{lem:spec_inf} and~\ref{lem:spec_inf_2} imply the following result.
\begin{cor}\label{cor:spec_inf}
If there exists a minimum spanning tree connecting a metric space $X$, and $\{e_1,e_2,\ldots\}$ are the edges of this tree ordered descendingly, then $\sigma_k=|e_k|$.
\end{cor}
\section{Maximum Spanning Trees}
\markright{\thesection.~Maximum spanning trees}
Let $M$ be a finite metric space. \emph{Maximum spanning tree $G$ on $M$} is a longest tree among all trees of the form $(M,E)$. By $\operatorname{xst}(M)$ we denote the length of a maximum spanning tree on $M$ , and by $\operatorname{XST}(M)$ we denote the set of all maximum spanning trees on $M$.
The next construction is useful for description of relations between minimum and maximum spanning trees.
Let $X$ be an arbitrary not pointwise bounded metric space. Choose any $d\ge2\operatorname{diam} X$ and define on $X$ a new distance function: $\rho(x,y)=d-|xy|$ for any $x\ne y$, and $\rho(x,x)=0$ for any $x$.
\begin{lem}
The function $\rho$ is a metric on $X$.
\end{lem}
\begin{proof}
Indeed, it is obvious that $\rho$ is positively definite and symmetric. To verify the triangle inequalities, choose any pairwise distinct points $x,y,z\in X$, then
$$
\rho(x,y)+\rho(y,z)-\rho(x,z)=d-|xy|+d-|yz|-d+|xz|\ge d-2\operatorname{diam} X\ge0.
$$
If two of these points coincide, say, if $y=z\ne x$, then $|xy|=|xz|$ and $\rho(z,y)=0$, hence
$$
\bigl|\rho(x,z)-\rho(z,y)\bigr|=\bigl|d-|xz|-0\bigr|=d-|xz|=d-|xy|=\rho(x,y)=d-|xz|+0=\rho(x,z)+\rho(z,y).
$$
\end{proof}
The set $X$ with the metric $\rho$ defined above is denoted by $d-X$.
Let $M$ be a finite metric space, $\#M=n$, and $N=n(n-1)/2$. Denote by $\rho(M)=(\rho_1,\ldots,\rho_N)$ the vector constructed from nonzero distances in $M$, ordered descendingly.
\begin{rk}
If $M$ is a finite metric space, and $\rho(M)=(\rho_1,\ldots,\rho_N)$, then $\rho(d-M)=(d-\rho_N,\ldots,d-\rho_1)$.
\end{rk}
\subsection{Duality}
The next Proposition describes a duality between minimum and maximum spanning trees.
\begin{prop}\label{prop:min-max-trees}
Let $M$ be a finite metric space, and $n=\#M$. Then a tree $G=(M,E)$ is a minimum spanning tree on $M$, iff it is a maximum spanning tree on $d-M$.
\end{prop}
\begin{proof}
Let $\rho$ stand for the distance function on $d-M$. Then $\rho(G)=\sum_{e\in E}\rho(e)=d(n-1)-|G|$.
\end{proof}
For $G\in\operatorname{XST}(M)$, by $\Sigma(G,M)$ we denote the vector constructed from the lengths of edges of the tree $G$, ordered ascendingly. If it is clear which metric is in consideration, then we write $\Sigma(G)$ instead of $\Sigma(G,M)$.
Proposition~\ref{prop:min-max-trees} implies the next result.
\begin{cor}\label{cor:spectrum-duality}
Let $M$ be a finite metric space, $\#M=n$, $d\ge2\operatorname{diam} M$, and $G\in\operatorname{MST}(M)$. Let us denote also by $d$ the vector of the length $n-1$, all whose components equal $d$. Then $G\in\operatorname{XST}(d-M)$ and $\Sigma(G,d-M)+\sigma(G,M)=d$.
\end{cor}
Corollary~\ref{cor:spectrum-duality} implies an analogue of Proposition~\ref{prop:mst-spect}.
\begin{prop}\label{prop:mst-spect-dual}
For any $G_1,G_2\in\operatorname{XST}(M)$ we have $\Sigma(G_1)=\Sigma(G_2)$.
\end{prop}
Proposition~\ref{prop:mst-spect-dual} motivates the next definition.
\begin{dfn}
For any finite metric space $M$ by $\Sigma(M)$ we denote the value $\Sigma(G,M)$ for an arbitrary $G\in\operatorname{XST}(M)$, and we call this value by \emph{$\operatorname{xst}$-spectrum of the space $M$}.
\end{dfn}
Let $X$ be an arbitrary metric space, and $A,B\subset X$ be its nonempty subsets. Put
$$
|AB|'=\sup\bigl\{|ab|:a\in A,\,b\in B\bigr\}.
$$
If the metric on $X$ is denoted by $\rho$ as well, we put $|AB|'=\rho(A,B)'$.
\begin{constr}
For a set $M$, let ${\cal C}_k(M)$ stand for the family of all coverings of the set $M$ consisting of $k$ nonempty subsets. Now, let $M$ be a metric space, and $C=\{M_1,\ldots,M_k\}\in{\cal C}_k(M)$. Put $\beta(C,M)=\max\bigl\{|M_iM_j|':i\ne j\bigr\}$. It it is clear which metric is in consideration, we write $\beta(C)$ instead of $\beta(C,M)$.
\end{constr}
\begin{lem}\label{lem:dual-dists-sets}
Let $A,B\subset X$ be nonempty subsets of a bounded metric space, and $\rho$ be the metric of the space $d-X$. Then $|AB|'=d-\rho(A,B)$.
\end{lem}
\begin{proof}
Indeed,
\begin{multline*}
|AB|'=\sup\bigl\{|ab|:a\in A,\,b\in B\bigr\}=\sup\bigl\{d-\rho(a,b):a\in A,\,b\in B\bigr\}=\\
=d-\inf\bigl\{\rho(a,b):a\in A,\,b\in B\bigr\}=d-\rho(A,B).
\end{multline*}
\end{proof}
\begin{lem}\label{lem:beta-alpha-dual}
For any $D=\{M_1,\ldots,M_k\}\in{\cal D}_k(M)$ we have $\beta(D,M)=d-\alpha(D,d-M)$.
\end{lem}
\begin{proof}
Indeed, let $\rho$ be the metric of the space $d-M$. Then, by Lemma~\ref{lem:dual-dists-sets}, it holds
\begin{multline*}
\beta(D,M)=\max\bigl\{|M_iM_j|':i\ne j\bigr\}=\max\bigl\{d-\rho(M_i,M_j):i\ne j\bigr\}=\\
=d-\min\bigl\{\rho(M_i,M_j):i\ne j\bigr\}=d-\alpha(D,d-M).
\end{multline*}
\end{proof}
Up to the end of this section, $M$ stands for a finite metric space consisting of $n$ points, and $\Sigma(M)=(\Sigma_1,\ldots,\Sigma_{n-1})$.
The next result is an analogue of Proposition~\ref{prop:spect-calc}.
\begin{prop}\label{prop:spect-calc-dual}
We have $\Sigma_k=\min\bigl\{\beta(D,M):D\in{\cal D}_{k+1}(M)\bigr\}$.
\end{prop}
\begin{proof}
Choose an arbitrary $d\ge2\operatorname{diam} M$, then $\sigma(d-M)=d-\Sigma(M)$. Put $\sigma(d-M)=(\sigma_1,\ldots,\sigma_{n-1})$ and $\Sigma(M)=(\Sigma_1,\ldots,\Sigma_{n-1})$. By Proposition~\ref{prop:spect-calc} and Lemma~\ref{lem:beta-alpha-dual}, we have
\begin{multline*}
\Sigma_k=d-\sigma_k=d-\max\bigl\{\alpha(D,d-M):D\in{\cal D}_{k+1}(d-M)\bigr\}=\\ =\min\bigl\{d-\alpha(D,d-M):D\in{\cal D}_{k+1}(d-M)\bigr\}=\min\bigl\{\beta(D,M):D\in{\cal D}_{k+1}(d-M)\bigr\}.
\end{multline*}
\end{proof}
\begin{rk}
Under the above notations, it holds $\operatorname{diam} M=\Sigma_{n-1}$.
\end{rk}
In the next two Propositions, we use $G=(M,E)\in\operatorname{XST}(M)$, and the edges $e_i\in E$ are supposed to be ordered in such a way that $|e_i|=\Sigma_i$. Moreover, for the convenience reason, we put $|e_0|=\Sigma_0=0$.
\begin{prop}\label{prop:diams-of-xms-partition}
Let $\{M_1,\ldots,M_{k+1}\}\in{\cal D}_{k+1}(M)$ be a partition into the vertex sets of the trees forming the forest $F=G\setminus\{e_{n-k},\ldots,e_{n-1}\}$. Then for each $i$ we have $\operatorname{diam} M_i\le|e_{n-k-1}|=\Sigma_{n-k-1}$.
\end{prop}
\begin{proof}
Choose arbitrary $x,y\in M_i$. If $xy\in E$, then $|xy|\le|e_{n-k-1}|$ by the order we have chosen on $E$. If $xy\not\in E$, then consider the unique path $\gamma$ in $G$ connecting $x$ and $y$. Since $M_i$ is the set of vertices of a tree from the forest $F$, then for each edge $e_j$ of this path it holds $j\le n-k-1$, and, therefore, $|e_j|\le|e_{n-k-1}|$. Since the tree $G$ is maximal, then $|xy|\le|e_j|$ for some $j\le n-k-1$, hence $\operatorname{diam} M_i\le|e_{n-k-1}|$.
\end{proof}
\begin{prop}\label{prop:dual-dists-of-xms-partition}
Let $\{M_1,\ldots,M_{k+1}\}\in{\cal D}_{k+1}(M)$ be a partition into the vertex sets of the trees forming the forest $F=G\setminus\{e_1,\ldots,e_k\}$. Then for every $i\ne j$ we have $|M_iM_j|'\le|e_k|=\Sigma_k$.
\end{prop}
\begin{proof}
Indeed, consider arbitrary $M_i, M_j$, $i\ne j$, and let $P_i\in M_i$, $P_j\in M_j$ be arbitrary points. Consider the unique path $\gamma$ in the tree $G$ connecting $P_i$ and $P_j$. This path contains at least one of the edges thrown out. Let it be $e_p$. If $P_iP_j$ is not an edge in $G$, then $G\cup P_iP_j$ contains a unique cycle $\gamma\cup P_iP_j$, and the maximality of $G$ implies that each edges of the path $\gamma$ is not shorter than $P_iP_j$. In particular, $|e_k|\ge|e_p|\ge |P_iP_j|$. If $P_iP_j$ is an edge of $G$, then it coincides with an edge we threw out, say, with $e_p$, and again we have $|e_k|\ge|e_p|=|P_iP_j|$. Thus, $|M_iM_j|'=\max|P_iP_j|\le|e_k|$.
\end{proof}
\subsection{$\operatorname{xst}$-spectrum of an arbitrary metric space}
Similarly with the section~\ref{subsec:mst-spector}, let us generalize the concept of $\operatorname{xst}$-spectrum by means of Proposition~\ref{prop:spect-calc-dual}.
\begin{dfn}
For any metric space $X$ and any $k\in{\mathbb N}$ we put $\Sigma_k=\inf\{\beta(D):D\in{\cal D}_{k+1}(X)\}$ if ${\cal D}_{k+1}(X)\ne\emptyset$, and $\Sigma_k=\infty$ otherwise. The set $\Sigma(X)=\{\Sigma_1,\Sigma_2,\ldots\}$ we call the {\em $\operatorname{xst}$-spectrum of $X$}.
\end{dfn}
\begin{rk}
If $\#X=n$, then $\Sigma_k=\infty$ for $k\ge n$.
\end{rk}
\section{Calculation of Distances between Compact Metric Space and Finite Simplexes}
\markright{\thesection.~Calculation of distances between compact metric space and finite simplexes}
We call a metric space $X$ a \emph{simplex}, if all its nonzero distances are the same. Notice that a simplex $X$ is compact, iff it is finite. A simplex consisting of $n$ vertices on the distance $\lambda$ from each other is denoted by $\lambda\Delta_n$. For $\lambda=1$ the space $\lambda\Delta_n$ is denoted by $\Delta_n$ for short.
\subsection{Distances from a finite metric space to simplexes with greater numbers of points}
\begin{thm}\label{thm:dist-n-simplex-bigger-dim}
Let $M$ be a finite metric space, $n=\#M$. Then for every $m\in{\mathbb N}$, $m>n$, and $\lambda>0$ we have
$$
2d_{GH}(\lambda\Delta_m,M)=\max\{\lambda,\operatorname{diam} M-\lambda\}.
$$
\end{thm}
\begin{proof}
Choose an arbitrary $R\in{\cal R}(\lambda\Delta_m,M)$. Since $m>n$, then there exists $x\in M$ such that $\#R^{-1}(x)\ge2$, hence $\operatorname{dis} R\ge\lambda$ and, therefore, $2d_{GH}(\lambda\Delta_m,M)\ge\lambda$.
Put $M=\{x_1,\ldots,x_n\}$ and let $R$ be the correspondence $\bigl\{(i,x_i)\bigr\}_{i=1}^{n-1}\cup\{n,\ldots,m\}\times\{x_n\}$. Then
$$
\operatorname{dis} R=\max\biggl[\lambda,\,\max_{i\ne j}\Bigl\{\bigl||x_ix_j|-\lambda\bigr|\Bigr\}\biggr]=\max\{\lambda,\operatorname{diam} M-\lambda\},
$$
where the second equality follows from Proposition~\ref{prop:many_abs_dif_and_t}. This implies that $2d_{GH}(\lambda\Delta_m,M)\le\max\{\lambda,\operatorname{diam} M-\lambda\}$.
If $\operatorname{diam} M\le 2\lambda$, then $\max\{\lambda,\operatorname{diam} M-\lambda\}=\lambda$, hence $2d_{GH}(\lambda\Delta_m,M)=\lambda$, q.e.d.
If $\operatorname{diam} M>2\lambda$, then $\max\{\lambda,\operatorname{diam} M-\lambda\}=\operatorname{diam} M-\lambda$. Choose a pair $x,y\in M$ such that $\operatorname{diam} M=|xy|$. Take an arbitrary $R\in{\cal R}(\lambda\Delta_m,M)$. Then one of the following conditions holds:
\begin{enumerate}
\item there exists $i\in\lambda\Delta_m$ such that $(i,x),(i,y)\in R$, but then $\operatorname{dis} R\ge\operatorname{diam} M>\operatorname{diam} M-\lambda$;
\item there exist $i\ne j$ such that $(i,x),(j,y)\in R$, and then $\operatorname{dis} R\ge\operatorname{diam} M-\lambda$.
\end{enumerate}
Thus, for any $R\in{\cal R}(\lambda\Delta_m,M)$ we have $\operatorname{dis} R\ge\operatorname{diam} M-\lambda$, therefore, in the case under consideration the equality $2d_{GH}(\lambda\Delta_m,M)=\operatorname{diam} M-\lambda$ is valid.
\end{proof}
\subsection{Distances from a compact metric space to simplexes with no greater number of points}
\begin{thm}\label{thm:m_less_than_n_corresp}
Let $X$ be a compact metric space. Then for every $m\in{\mathbb N}$, $m\le\#X$, and $\lambda>0$ there exists an $R\in{\cal R}_{{\operatorname{opt}}}^0(\lambda\Delta_m,X)$ such that the family $\bigl\{R(i)\bigr\}$ is a partition of $X$. In particular, if $m=\#X$, then one can take a bijection as an optimal correspondence $R$.
\end{thm}
\begin{proof}
Let $R\in{\cal R}(\Delta_m,X)$ be an arbitrary irreducible correspondence. Since $R$ is irreducible, the condition $R(j)\cap R(k)\ne\emptyset$ for some $j\ne k$ implies $R(j)=R(k)=\{x\}$ for some $x\in X$. In particular, if $\#R(i)>1$ for some $i$, then $R(i)$ does not intersect any $R(p)$, $p\ne i$. Let us introduce the following notation: for $\sigma\in{\cal P}(\Delta_m,X)$ we put $\operatorname{np}(\sigma)$ to be equal to the number of pairs $\{j,k\}$ such that $j\ne k$ and $\sigma(j)\cap\sigma(k)\ne\emptyset$. Clearly that for $\sigma\in{\cal R}(\Delta_m,X)$ the condition $\operatorname{np}(\sigma)=0$ is equivalent to that the family $\bigl\{\sigma(i)\bigr\}$ forms a partition $X$.
Suppose that the family $\bigl\{R(i)\bigr\}$ is not a partition. We show that in this case one can always reconstruct the correspondence $R$ in such a way that the resulting correspondence ${\tilde R}$ becomes an irreducible one with $\operatorname{dis}{\tilde R}\le\operatorname{dis} R$ and $\operatorname{np}({\tilde R})<\operatorname{np}(R)$. If $\operatorname{np}({\tilde R})>0$, then we put $R={\tilde R}$, and repeat this procedure. After a finite number of steps we will get a correspondence ${\tilde R}$ such that $\operatorname{np}({\tilde R})=0$ and $\operatorname{dis}{\tilde R}\le\operatorname{dis} R$, q.e.d.
So, let for some $j\ne k$ it holds $R(j)=R(k)=\{x\}$. Since $m\le\#X$, then there exists $i$ such that $\#R(i)>1$, hence $i\not\in\{j,k\}$ and $R(i)$ does not intersect any other $R(p)$. Choose an arbitrary point $x_i\in R(i)$, and construct a new correspondence ${\tilde R}$ which coincides with $R$ at all elements of the simplex, except $i$, $j$, $k$, and
$$
{\tilde R}(i)=R(i)\setminus\{x_i\}, \quad {\tilde R}(j)=\{x_i\}, \quad {\tilde R}(k)=\{x\}.
$$
Clearly that ${\tilde R}\in{\cal R}(\Delta_m,X)$ since $R$ is uniquely defined on all elements of the simplex, and each element of $X$ goes to at least one element of the simplex. Besides that, the correspondence ${\tilde R}$ is still irreducible, which can be verified directly.
Further, to estimate $\operatorname{np}({\tilde R})$ let us notice that among all $R(p)$ only $R(i)$ and $R(j)$ are changed, thus, it is sufficient to investigate how the intersections with these two sets changes. Since $R(i)$ does not intersect the remaining $R(p)$, and since ${\tilde R}(i)\sqcup{\tilde R}(j)=R(i)$, then ${\tilde R}(i)$ and ${\tilde R}(j)$ does not intersect the remaining ${\tilde R}(p)$. Besides that, ${\tilde R}(i)\cap{\tilde R}(j)=\emptyset$. Thus, the number of the remaining $R(p)$ intersecting with $R(i)$ is the same as the number of the remaining ${\tilde R}(p)$ intersecting with ${\tilde R}(i)$ (and it is equal $0$). Concerning $R(j)$, the intersection $R(j)\cap R(k)$ is nonempty (and, perhaps, there are some other nonempty intersections with $R(j)$). However, ${\tilde R}(j)$ does not intersect any of the remaining ${\tilde R}(p)$, hence $\operatorname{np}({\tilde R})<\operatorname{np}(R)$.
Let us prove that $\operatorname{dis}{\tilde R}\le\operatorname{dis} R$.
Put
$$
M(R,p,q)=\max\Bigl\{\big| |x_px_q|-1\big|:x_p\in R(p),\,x_q\in R(q),\,p\ne q\Bigr\}.
$$
Recall that
$$
\operatorname{dis} R=\max\big\{\delta_R=1,\,\max_p\operatorname{diam} R(p),\,\max_{p\ne q}M(R,p,q)\big\},
$$
and
$$
\operatorname{dis}{\tilde R}=\max\big\{\delta_{\tilde R},\,\max_p\operatorname{diam}{\tilde R}(p),\,\max_{p\ne q}M({\tilde R},p,q)\big\}.
$$
Clearly that $\delta_{{\tilde R}}\le 1\le\operatorname{dis} R$ and $\operatorname{diam}{\tilde R}(p)\le\operatorname{diam} R(p)\le\operatorname{dis} R$ for all $p$. To complete the proof it remains to show that for any $p$ and $q$ the inequality $M({\tilde R},p,q)\le\operatorname{dis} R$ holds.
If $p$ and $q$ are not contained in $\{i,j,k\}$, then $M({\tilde R},p,q)=M(R,p,q)\le\operatorname{dis} R$.
Now, suppose that one of the indices $p$ and $q$, say $p$, is not contained in $\{i,j,k\}$, but the remaining one is contained. In this case,
\begin{description}
\item[] $M({\tilde R},p,k)=M(R,p,k)\le\operatorname{dis} R$, because ${\tilde R}(p)=R(p)$ and ${\tilde R}(k)=R(k)$;
\item[] $M({\tilde R},p,i)\le M(R,p,i)\le\operatorname{dis} R$, because ${\tilde R}(i)\subset R(i)$;
\item[] $M({\tilde R},p,j)\le M(R,p,i)\le\operatorname{dis} R$, because ${\tilde R}(j)\subset R(i)$.
\end{description}
Finally, consider the case $\{p,q\}\subset\{i,j,k\}$. We have
\begin{description}
\item[] $M({\tilde R},i,k)\le M(R,i,k)\le\operatorname{dis} R$, because ${\tilde R}(i)\subset R(i)$ and ${\tilde R}(k)=R(k)$;
\item[] $M({\tilde R},j,k)\le M(R,i,k)\le\operatorname{dis} R$, because ${\tilde R}(j)\subset R(i)$ and ${\tilde R}(k)=R(k)$;
\end{description}
and
\begin{multline*}
M({\tilde R},i,j)=\max\Bigl\{\big| |x_i'x_i|-1\big|:x_i'\in {\tilde R}(i)=R(i)\setminus\{x_i\}\Bigr\}\le\\
\le\max\Bigl\{1,\big| |x_i'x_i|-1\big|:x_i'\in {\tilde R}(i)\Bigr\}\le\max\Bigl\{1,\max\big\{ |x_i'x_i| :x_i'\in {\tilde R}(i)\big\}\Bigr\}\le\\
\le\max\big\{1,\operatorname{diam} R(i)\big\}\le\operatorname{dis} R,
\end{multline*}
where the second inequality holds according to Proposition~\ref{prop:max_abs}.
Thus, all the values from the expression for $\operatorname{dis} {\tilde R}$ do not exceed $\operatorname{dis} R$, hence $\operatorname{dis}{\tilde R}\le\operatorname{dis} R$, q.e.d.
\end{proof}
Theorem~\ref{thm:m_less_than_n_corresp} helps to get a useful formula for Gromov--Hausdorff distance between a compact metric space $X$ and a finite simplex such that the number of points in the simplex does not exceed the cardinality of $X$.
\begin{constr}
For an \textbf{arbitrary\/} metric space $X$, $m\le\#X$, and $D=\{X_1,\ldots,X_m\}\in{\cal D}_m(X)$ we put $R_D=\sqcup\,\bigl(\{i\}\times X_i\bigr)$. Notice that for any $D'\in{\cal D}_m(X)$ which differs from $D$ by renumbering of the elements of the partition $D$, we have $\operatorname{dis} R_D=\operatorname{dis} R_{D'}$.
\end{constr}
\begin{notation}
For $D=\{X_1,\ldots,X_m\}\in{\cal D}_m(X)$ let us put
$$
\operatorname{diam} D=\max\{\operatorname{diam} X_1,\ldots,\operatorname{diam} X_m\}.
$$
\end{notation}
\begin{prop}\label{prop:disRD}
Let $X$ be an arbitrary metric space, and $m\in{\mathbb N}$, $m\le\#X$. Then for any $\lambda>0$ and $D\in{\cal D}_m(X)$ it holds
$$
\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\lambda-\alpha(D),\,\beta(D)-\lambda\}.
$$
\end{prop}
\begin{proof}
Let $D=\{X_1,\ldots,X_m\}$. By definition of distortion,
$$
\operatorname{dis} R_D=\sup\{\operatorname{diam} D,\,\bigl|\lambda-|xy|\bigr|:x\in X_p,\,y\in X_q,\,1\le p<q\le m\}.
$$
By Proposition~\ref{prop:many_abs_dif}, we have
$$
\sup\{\bigl|\lambda-|xy|\bigr|:x\in X_p,\,y\in X_q,\,1\le p<q\le m\}=\max\bigl\{\lambda-\alpha(D),\,\beta(D)-\lambda\bigr\}.
$$
\end{proof}
\begin{prop}\label{prop:GH-dist-RD}
Let $X$ be a compact metric space. Then for every $m\in{\mathbb N}$, $m\le\#X$, and $\lambda>0$ it holds
$$
2d_{GH}(\lambda\Delta_m,X)=\inf_{D\in{\cal D}_m(X)}\operatorname{dis} R_D.
$$
\end{prop}
\begin{proof}
By Theorem~\ref{thm:m_less_than_n_corresp}, there exists an $R\in{\cal R}_{{\operatorname{opt}}}^0(\lambda\Delta_m,X)$ such that the family $\bigl\{R(i)\bigr\}$ is a partition of $X$. Thus, $d_{GH}(\lambda\Delta_m,X)$ is achieved at some $R_D$, $D\in{\cal D}_m(X)$.
\end{proof}
\begin{cor}\label{cor:GH-dist-alpha-beta}
Let $X$ be a compact metric space, and $m\in{\mathbb N}$, $m\le\#X$. Then for any $\lambda>0$ we have
$$
2d_{GH}(\lambda\Delta_m,X)=\inf\Bigl\{\max\bigl(\operatorname{diam} D,\,\lambda-\alpha(D),\,\beta(D)-\lambda\bigr):D\in{\cal D}_m(X)\Bigr\}.
$$
\end{cor}
\subsection{Distance from finite metric space to simplexes with the same number of points}
For any metric space $X$ we put
$$
\varepsilon(X)=\inf\bigl\{|xy|:x,y\in X,\,x\ne y\bigr\}.
$$
Notice that $\varepsilon(X)\le\operatorname{diam} X$, and the equality holds, iff $X$ is a simplex. Besides that, if $M$ is a finite metric space consisting of $n$ points, and $\sigma(M)=(\sigma_1,\ldots,\sigma_{n-1})$, then $\varepsilon(M)=\sigma_{n-1}$.
\begin{thm}\label{thm:dist-n-simplex-same-dim}
Let $M$ be a finite metric space, $\#M=n$, $\sigma(M)=(\sigma_1,\ldots,\sigma_{n-1})$, $\Sigma(M)=(\Sigma_1,\ldots,\Sigma_{n-1})$, $\lambda>0$. Then
$$
2d_{GH}(\lambda\Delta_n,M)=\max\{\lambda-\sigma_{n-1},\,\Sigma_{n-1}-\lambda\}=\max\bigl\{\lambda-\varepsilon(M),\,\operatorname{diam} M-\lambda\bigr\}.
$$
More exactly, if $\sigma_{n-1}+\Sigma_{n-1}\le2\lambda$, then $2d_{GH}(\lambda\Delta_n,M)=\lambda-\sigma_{n-1}$, otherwise $2d_{GH}(\lambda\Delta_n,M)=\Sigma_{n-1}-\lambda$.
\end{thm}
\begin{proof}
For each $D\in{\cal D}_n(X)$ it holds $\operatorname{diam} D=0$. Besides that, for all such $D$ we have $\alpha(D)=\sigma_{n-1}$ and $\beta(D)=\Sigma_{n-1}$. It remains to apply Corollary~\ref{cor:GH-dist-alpha-beta}.
\end{proof}
\subsection{Distance from a compact metric space to simplexes having at most the same number of points}
In~\cite{TuzMST-GH} the following result is proved.
\begin{prop}[\cite{TuzMST-GH}]\label{mst-spector-GH}
Let $X$ be a finite metric space, $\lambda\ge2\operatorname{diam} X$, and $\sigma(X)=(\sigma_1,\ldots,\sigma_{n-1})$. Then $2d_{GH}(\lambda\Delta_{k+1},X)=\lambda-\sigma_k$ for all $k=1,\ldots,n-1$.
\end{prop}
The next theorem generalizes Proposition~\ref{mst-spector-GH} to the case of compact metric spaces and also weaken the restrictions on the parameter $\lambda$.
\begin{thm}\label{thm:spec}
Let $X$ be a compact metric space, $\sigma(X)=\{\sigma_1,\sigma_2,\ldots\}$ be the $\operatorname{mst}$-spectrum of $X$, and $\lambda\ge\operatorname{diam} X+\sigma_k$. Then $2d_{GH}(\lambda\Delta_{k+1},X)=\lambda-\sigma_k$ for all $k\in{\mathbb N}$, $k+1\le\#X$.
\end{thm}
\begin{proof}
Put $m=k+1$ and choose an arbitrary partition $D\in{\cal D}_m(X)$. Due to Proposition~\ref{prop:disRD}, we have
$$
\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\lambda-\alpha(D),\,\beta(D)-\lambda\}.
$$
Notice that $\operatorname{diam} D\le\operatorname{diam} X$ and $\beta(D)-\lambda\le\operatorname{diam} X-\lambda<\operatorname{diam} X$. Further, since $\lambda\ge\operatorname{diam} X+\sigma_k$, then $\lambda-\alpha(D)\ge\lambda-\sigma_k\ge\operatorname{diam} X$, and hence $\operatorname{dis} R_D=\lambda-\alpha(D)$. So, due to Proposition~\ref{prop:GH-dist-RD}, we have $2d_{GH}(\lambda\Delta_m,X)=\lambda-\sigma_k$.
\end{proof}
\begin{notation}
Let $X$ be an arbitrary metric space. Put $d_m(X)=\inf\bigl\{\operatorname{diam} D:D\in{\cal D}_m(X)\bigr\}$, if ${\cal D}_m(X)\ne\emptyset$, and $d_m(X)=\infty$ otherwise.
\end{notation}
\begin{rk}
If $X$ is a finite metric space and $n=\#X$, then $d_n(X)=0$.
\end{rk}
Recall that a \emph{clique\/} in a simple graph is any its subgraph which is a complete graph. A graph is said to be an \emph{$m$-clique}, if it contains a spanning subgraph which is a disjoint union of $m$ cliques. Notice that each graph having $n$ vertices is $n$-clique.
Let $X$ be an arbitrary metric space, and $\delta\ge0$. By $G_\delta(X)$ we denote the graph with the vertex set $X$, where $v,w\in X$, $v\ne w$, are connected by edge iff $|vw|\le\delta$ (this graph is infinite, generally speaking). Define the number $\delta_m(X)$ to be equal to the infimum of $\delta\ge0$ such that $G_\delta(X)$ is $m$-clique. If there is no such $\delta$, then we put $\delta_m(X)=\infty$.
\begin{prop}
For an arbitrary metric space $X$ the equality $d_m(X)=\delta_m(X)$ holds.
\end{prop}
\begin{proof}
Indeed, ${\cal D}_m(X)=\emptyset$ iff $\#X<m$, and the latter is equivalent to nonexistence of an $m$-clique graph with the vertex set $X$. So, in this case the required equality holds.
Let $\#X\ge m$. The condition $d_m(X)=\infty$ is equivalent to the fact that for any partition $D=\{X_i\}\in{\cal D}(X)$ we have $\operatorname{diam} X_i=\infty$ for some $i$. The latter condition is equivalent to nonexistence of an $m$-clique subgraph in the complete graph with the vertex set $X$, such that all its edges do not exceed some value $\delta$. Thus, in this case the required equality holds as well.
Now, let $d_m(X)<\infty$. Consider the family of partitions $D_i\in{\cal D}_m(X)$ such that $d_i=\operatorname{diam} D_i\to d_m(X)$. Then the graph $G_{d_i}$ contains an $m$-clique subgraph such that each $D_i$ lies in some its clique. The latter implies that $\delta_m(X)\le d_m(X)$.
Conversely, let $\delta_i$ be a decreasing sequence such that $\delta_i\to\delta_m(X)$. By $G_i$ we denote some $m$-clique subgraph of $G_{\delta_i}$. Let $H_i$ be the subgraph in $G_i$ that is equal to the disjoint union of $m$ cliques. Denote by $X_i^p$ the vertex sets of the cliques of the graph $H_i$, then $\operatorname{diam} X_i^p\le\delta_i$. Put $D_i=\{X_i^p\}$, and we get $D_i\in{\cal D}_m(X)$ and $\operatorname{diam} D_i\le\delta_i$, so $d_m(X)\le\delta_m(X)$.
\end{proof}
\begin{thm}\label{thm:small1}
Let $X$ be a compact metric space, $\sigma(X)=\{\sigma_1,\sigma_2,\ldots\}$ be the $\operatorname{mst}$-spectrum of $X$, and $\lambda<\operatorname{diam} X+\sigma_k$. Assume that for some $k\in{\mathbb N}$, $k+1\le\#X$, the equality $d_{k+1}(X)=\operatorname{diam} X$ is valid. Then $2d_{GH}(\lambda\Delta_{k+1},X)=\operatorname{diam} X$.
\end{thm}
\begin{proof}
Put $m=k+1$. Recall that, in accordance to Proposition~\ref{prop:disRD}, for any $D\in{\cal D}_m(X)$ the relation
$$
\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\lambda-\alpha(D),\,\beta(D)-\lambda\}
$$
is valid. Besides, $\operatorname{diam} D\le\operatorname{diam} X$ and $\beta(D)-\lambda\le\operatorname{diam} X-\lambda<\operatorname{diam} X$. Since $\lambda<\operatorname{diam} X+\sigma_k$, then there exists $D\in{\cal D}_m(X)$ such that $\lambda-\operatorname{diam} X<\alpha(D)$, therefore, for such $D$ we have $\lambda-\alpha(D)<\operatorname{diam} X$. Gathering all those inequalities, we conclude that for such $D$ the inequality $\operatorname{dis} R_D\le\operatorname{diam} X$ holds. Thus, due to Proposition~\ref{prop:GH-dist-RD}, we have $2d_{GH}(\lambda\Delta_m,X)\le\operatorname{diam} X$.
On the other hand, since $d_m(X)=\operatorname{diam} X$, then for any $R_D\in{\cal R}^0_{{\operatorname{opt}}}(\lambda\Delta_m,X)$ we have
$$
\operatorname{diam} X\ge2d_{GH}(\lambda\Delta_m,X)=\operatorname{dis} R_D\ge\operatorname{diam} D\ge d_m(X)=\operatorname{diam} X,
$$
so $2d_{GH}(\lambda\Delta_m,X)=\operatorname{diam} X$.
\end{proof}
\begin{thm}\label{thm:small2}
Let $X$ be a compact metric space. Then for any $m\in{\mathbb N}$, $m\le\#X$, and any $0<\lambda\le(\operatorname{diam} X)/2$ the equality $2d_{GH}(\lambda\Delta_m,X)=\max\{d_m(X),\operatorname{diam} X-\lambda\}$ holds.
\end{thm}
\begin{proof}
Choose an arbitrary partition $D=\{X_i\}\in{\cal D}_m(X)$. Due to proposition~\ref{prop:disRD}, we have
$$
\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\lambda-\alpha(D),\,\beta(D)-\lambda\}.
$$
Notice that $\operatorname{diam} D\le\operatorname{diam} X$, $\lambda-\alpha(D)\le\lambda<\operatorname{diam} X$, and $\beta(D)-\lambda\le\operatorname{diam} X-\lambda<\operatorname{diam} X$, therefore, $\operatorname{dis} R_D\le\operatorname{diam} X$, and if $\operatorname{diam} D<\operatorname{diam} X$, then $\operatorname{dis} R_D<\operatorname{diam} X$. In particular, $2d_{GH}(\lambda\Delta_m,X)<\operatorname{diam} X$ in this case.
Let $d_m(X)=\operatorname{diam} X$, then
$$
\operatorname{diam} X\ge\operatorname{dis} R_D\ge\operatorname{diam} D\ge d_m(X)=\operatorname{diam} X,
$$
hence, $\operatorname{dis} R_D=\operatorname{diam} X$ for all $D\in{\cal D}_m$, and, due to Proposition~\ref{prop:GH-dist-RD}, we have
$$
2d_{GH}(\lambda\Delta_m,X)=\operatorname{dis} R_D=\operatorname{diam} X=\max\bigl\{d_m(X),\operatorname{diam} X-\lambda\bigr\}.
$$
Now, let $d_m(X)<\operatorname{diam} X$. Then, in accordance to the definition of infimum, there exists $D'\in{\cal D}_m(X)$ such that $\operatorname{diam} D'<\operatorname{diam} X$. As we have already mentioned above, the latter implies that $\operatorname{dis} R_{D'}<\operatorname{diam} X$, and hence, $2d_{GH}(\lambda\Delta_m,X)<\operatorname{diam} X$. On the other hand, due to Theorem~\ref{thm:m_less_than_n_corresp}, there exists a $D\in{\cal D}_m(X)$ such that $R_D\in{\cal R}^0_{{\operatorname{opt}}}(\lambda\Delta_m,X)$. But then $\operatorname{dis} R_D\le\operatorname{dis} R_{D'}<\operatorname{diam} X$ and $\operatorname{diam} D<\operatorname{diam} X$.
Also, notice that the inequality $\operatorname{diam} D<\operatorname{diam} X$ implies the equality $\beta(D)=\operatorname{diam} X$. Besides, since
$$
\beta(D)+\alpha(D)=\operatorname{diam} X +\alpha(D)\ge\operatorname{diam} X
$$
in this case, then the assumption $\lambda\le(\operatorname{diam} X)/2$ implies that $\beta(D)+\alpha(D)\ge 2\lambda$, and hence, $\beta(D)-\lambda\ge\lambda-\alpha(D)$ and
$$
\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\beta(D)-\lambda\}.
$$
Thus, $D\in{\cal D}_m(X)$, $R_D\in{\cal R}^0_{{\operatorname{opt}}}(\lambda\Delta_m,X)$, and $\operatorname{diam} D<\operatorname{diam} X$. So, if $\operatorname{diam} D\le\operatorname{diam} X-\lambda$, then $\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\operatorname{diam} X-\lambda\}=\operatorname{diam} X-\lambda=\max\bigl\{d_m(X),\,\operatorname{diam} X-\lambda\bigr\}$.
If $\operatorname{diam} D\ge\operatorname{diam} X-\lambda$, then
$$
\operatorname{dis} R_D=\max\{\operatorname{diam} D,\,\operatorname{diam} X-\lambda\}=\operatorname{diam} D\ge\max\{d_m(X),\,\operatorname{diam} X-\lambda\}.
$$
Let us prove the inverse inequality. To do that, consider a sequence $D_i\in{\cal D}_m(X)$ such that $\operatorname{diam} X>\operatorname{diam} D_i\to d_m(X)$. For each $i$ we have $\beta(D_i)=\operatorname{diam} X$ and $\lambda-\alpha(D_i)\le\beta(D_i)-\lambda$, and hence,
$$
\operatorname{dis} R_{D_i}=\max\{\operatorname{diam} D_i,\operatorname{diam} X-\lambda\}\to\max\bigl\{d_m(X),\operatorname{diam} X-\lambda\bigr\}.
$$
But $R_D\in{\cal R}^0_{{\operatorname{opt}}}(\lambda\Delta_m,X)$, therefore, $\operatorname{dis} R_D\le \operatorname{dis} R_{D_i}$ for any $i$. Passing to the limit, we conclude that $\operatorname{dis} R_D\le \max\bigl\{d_m(X),\,\operatorname{diam} X-\lambda\bigr\}$, q.e.d.
\end{proof}
Now we apply our technique to calculate the distances between a finite $n$-point metric space and the simplex $t\Delta_{n-1}$.
\begin{thm}\label{thm:minus_one}
Let $X$ be a finite metric space, $\#X=n\ge2$, $\lambda>0$. Then
$$
2d_{GH}(\lambda\Delta_{n-1},X)=\max\{\sigma_{n-1}, \lambda-\sigma_{n-2},\Sigma_{n-1}-\lambda\}.
$$
Moreover, the correspondence that takes one of the simplex's elements to a pair of the closest points from $X$, and that is one-to-one on the remaining elements, is optimal.
\end{thm}
\begin{proof}
Apply Corollary~\ref{cor:GH-dist-alpha-beta} again and conclude that the doubled distance sought for is equal to the minimum of distortions of the correspondences $R_D$ generated by partitions $D$ of the space $X$ into $n-1$ subsets. Since $\#X=n$, then all such partitions have the following form: $n-2$ elements of a partition are single points, and one element of the partition consists of two points. Let the two-point element be equal to $\{x_i,x_j\}$. Then
$$
\operatorname{dis} R_D=\max\big\{|x_ix_j|,\,\lambda-\alpha(D),\,\beta(D)-\lambda\big\},
$$
$\alpha(D)=\min\big\{|x_px_q|:p\ne q,\,\{p,q\}\ne\{i,j\}\big\}$, and $\beta(D)=\max\big\{|x_px_q|:p\ne q,\,\{p,q\}\ne\{i,j\}\big\}$, i.e., the minimum and the maximum are taken over all the distances from $X$, except the single one. Notice that if we write down all nonzero distances in a finite metric space in the increasing order, then the resulting sequence has the form $\{\sigma_{n-1},\sigma_{n-2},\ldots,\Sigma_{n-2},\Sigma_{n-1}\}$. Therefore,
$$
\operatorname{dis} R_D=\begin{cases}
\max\{\sigma_{n-1},\,\lambda-\sigma_{n-2},\,\Sigma_{n-1}-\lambda\}, &\text{if $|x_ix_j|=\sigma_{n-1}$},\\
\max\{\Sigma_{n-1},\,\lambda-\sigma_{n-1},\,\Sigma_{n-2}-\lambda\}, &\text{if $|x_ix_j|=\Sigma_{n-1}$},\\
\max\big\{|x_ix_j|,\,\lambda-\sigma_{n-1},\,\Sigma_{n-1}-\lambda\big\}, &\text{in the remaining cases}.
\end{cases}
$$
Since $\Sigma_{n-1}$ is the maximal distance in $X$, and $\sigma_{n-1}$ is the minimal one, then
$$
\max\{\Sigma_{n-1},\,\lambda-\sigma_{n-1},\,\Sigma_{n-2}-\lambda\}=\max\{\Sigma_{n-1},\,\lambda-\sigma_{n-1}\}\ge \max\{\sigma_{n-1},\, \lambda-\sigma_{n-2},\,\Sigma_{n-1}-\lambda\}.
$$
Further, $\max\big\{|x_ix_j|,\,\lambda-\sigma_{n-1},\,\Sigma_{n-1}-\lambda\big\}\ge\max\{\sigma_{n-1},\,\lambda-\sigma_{n-2},\,\Sigma_{n-1}-\lambda\big\}$ because $\sigma_{n-1}\le |x_ix_j|$ for all $i$ and $j$. So,
$$
2d_{GH}(\lambda\Delta_{n-1},X)=\min_D\operatorname{dis} R_D=\max\{\sigma_{n-1},\,\lambda-\sigma_{n-2},\,\Sigma_{n-1}-\lambda\}.
$$
Moreover, the right hand side of the latter equality is the distortion of the correspondence which takes one of the simplex's elements to a pair of the closest elements from the space $X$, and that is one-to-one on the remaining parts of the spaces.
\end{proof}
\subsection{Examples}
However, in general case calculation of the Gromov--Hausdorff distance to a simplex remains a difficult problem. Here we discuss some examples of the Gromov--Hausdorff distance calculation between the simplexes $t\Delta_2$ and a four-point metric space $X=\{x_1,x_2,x_3,x_4\}$ with a distance matrix
$$
\left(\begin{array}{cccc}
0 & a & b & d \\
a & 0 & c & e \\
b & c & 0 & f \\
d & e & f & 0
\end{array}\right).
$$
In accordance to Theorem~\ref{thm:m_less_than_n_corresp} and Corollary~\ref{cor:GH-dist-alpha-beta}, to calculate the distance it suffices to consider only irreducible correspondences $R_D\in{\cal R}(t\Delta_2,X)$ generated by partitions $D=\{X_1,X_2\}$ of the set $X$ into nonempty subsets. Evidently, there are seven such partitions, namely, four into a single element subset and a three-element subset, and three into a pair of two-element subsets. If such a partition is fixed, then in accordance to Proposition~\ref{prop:disRD} its distortion has the following form:
$$
\operatorname{dis} R_D=\max\big\{\operatorname{diam} D,\,t-\alpha(D),\,\beta(D)-t\big\}.
$$
If $X_1=\{x_1\}$, $X_2=\{x_2,x_3,x_4\}$, then
$$
\operatorname{dis} R_D=\max\big\{\max\{c,e,f\},\,t-\min\{a,b,d\},\,\max\{a,b,d\}-t\big\},
$$
and if $X_1=\{x_1,x_2\}$, $X_2=\{x_3,x_4\}$, then
$$
\operatorname{dis} R_D=\max\big\{\max\{a,f\},\,t-\min\{b,c,d,e\},\,\max\{b,c,d,e\}-t\big\}.
$$
To obtain an answer in more concrete form, we need to make some assumptions concerning the distances in $X$.
\subsubsection{Distance function is cumbersome, but can be expressed in terms of the spectra}
Assume that the distances are ordered as follows: $a<e<b<c<f<d$. In this case the spectra are $\sigma=\{b,e,a\}$ and $\Sigma=\{c,f,d\}$. In accordance to the above reasoning, the distance is calculated as the following minimum of seven maxima:
\begin{multline}\label{mult:1}
2d_{GH}(t\Delta_2,X)=\min\Big\{
\max\big\{f,t-a,d-t\big\},
\max\big\{d,t-a,c-t\big\},
\max\big\{d,t-b,f-t\big\}, \\
\max\big\{c,t-e,d-t\big\},
\max\big\{f,t-e,d-t\big\},
\max\big\{b,t-a,d-t\big\},
\max\big\{d,t-a,f-t\big\}
\Big\}.
\end{multline}
Consider the third, fourth, and sixth maxima in the right hand part of Formula~(\ref{mult:1}), and notice that the remaining maxima are greater than or equal to one of these three. This can be verified directly. For example, $\max\big\{d,t-a,c-t\big\}\ge\max\big\{b,t-a,d-t\big\}$, because $d\ge b$ and $d\ge d-t$. Thus,
\begin{equation}\label{mult:1.1}
2d_{GH}(t\Delta_2,X)=\min\Big\{
\max\big\{d,t-b,f-t\big\}, \\
\max\big\{c,t-e,d-t\big\},
\max\big\{b,t-a,d-t\big\}
\Big\}.
\end{equation}
\begin{lem}
For $t\in[0,a+c]$ we have $2d_{GH}(t\Delta_2,X)=\max\big\{b,t-a,d-t\big\}$.
\end{lem}
\begin{proof}
It suffices to verify that in the chosen segment the first two maxima in Formula~(\ref{mult:1.1}) are not less than the third one.
To compare the first and the third maxima, notice that: $b<d$ for all $t$; since $t\le a+c$, then $t-a\le c<d$; and, at last, $d-t\le d$.
To compare the second and the third maxima, notice that: $b<c$ for all $t$; since $t\le a+c$, then $t-a\le c$.
\end{proof}
\begin{lem}
For $t\in[a+c,e+d]$ we have $2d_{GH}(t\Delta_2,X)=\max\big\{c,t-e,d-t\big\}$.
\end{lem}
\begin{proof}
It suffices to verify that in the chosen segment the first and the third maxima in Formula~(\ref{mult:1.1}) are not less than the second one.
To compare the first and the second maxima, notice that: $c<d$ for all $t$; since $t\le e+d$, then $t-e<d$; and, at last, $d-t\le d$.
To compare the third and the second maxima, notice that: $t-a>t-e$ for all $t$; since $t\ge a+c$, then $t-a\ge c$.
\end{proof}
\begin{lem}
For $t\in[e+d,\infty]$ we have $2d_{GH}(t\Delta_2,X)=\max\big\{d,t-b,f-t\big\}$.
\end{lem}
\begin{proof}
It suffices to verify that in the chosen segment the second and the third maxima in Formula~(\ref{mult:1.1}) are not less than the first one.
To compare the second and the first maxima, notice that: $t-e>t-b$, and also $d-t>f-t$ for all $t$; since $t\ge e+d$, then $t-e\ge d$.
To compare the third and the first maxima, notice that: $t-a>t-b$, and also $d-t>f-t$ for all $t$; since $t\ge e+d$, then $t-a>t-e\ge d$.
\end{proof}
To visualize the obtained result let us fix the following values of the distances: $a=3$, $b=4$, $c=5$, $d=6.5$, $e=3.5$, $f=6$ (the triangle inequalities can be verified directly). The graph of the function $f(t)=2d_{GH}(t\Delta_2,X)$ is depicted in Figure~\ref{fig:exam_1}.
\ig{exam_1}{0.33}{fig:exam_1}{Graph of the function $f(t)=2d_{GH}(t\Delta_2,X)$ for $a=3$, $e=3.5$, $b=4$, $c=5$, $f=6$, $d=6.5$.}
Notice that in the case under consideration all the six distances belong to one of the two spectra. Therefore the function $f(t)$ can be expressed in terms of the spectra (in this case all the elements of the spectrum $\sigma$, and also $\Sigma_1$ and $\Sigma_3$, are used). But this is not always true.
\subsubsection{Distance function is easier, but can not be expressed in terms of spectra}
Assume now that the distances are ordered as follows: $a<b<c<d<e<f$. In this case the spectra are $\sigma=\{d,b,a\}$ and $\Sigma=\{d,e,f\}$. The value $c$ does not belong to $\sigma$, because the corresponding edge forms a cycle together with the two edges of the least lengths. Now, the Gromov--Hausdorff distance between $X$ and the simplex $t\Delta_2$ has the form
\begin{multline}\label{mult:2}
2d_{GH}(t\Delta_2,X)=\min\Big\{
\max\big\{f,t-a,d-t\big\},
\max\big\{f,t-a,e-t\big\},
\max\big\{e,t-b,f-t\big\}, \\
\max\big\{c,t-d,f-t\big\},
\max\big\{f,t-b,e-t\big\},
\max\big\{e,t-a,f-t\big\},
\max\big\{d,t-a,f-t\big\}
\Big\}.
\end{multline}
Notice that all the maxima in the right hand part of Formula~(\ref{mult:2}) are not less than the forth maximum $\max\big\{c,t-d,f-t\big\}$, and hence
$$
2d_{GH}(t\Delta_2,X)=\max\big\{c,t-d,f-t\big\}.
$$
To visualize the result, we include the graph of the function $g(t)=2d_{GH}(t\Delta_2,X)$ for some specific values of the distances, see Figure~\ref{fig:exam_2}.
\ig{exam_2}{0.33}{fig:exam_2}{Graph of the function $g(t)=2d_{GH}(t\Delta_2,X)$ for $a=2$, $b=3$, $c=4$, $d=5$, $e=6$, $f=7$.}
In spite of the fact that the function $g$ is simpler than $f$ from the first example, it can not be expressed in terms of the spectra, because $c$ does not belong to them. Notice also that in this case the value $c$ is equal to $\delta_2(X)$ (i.e., to the $2$-clique number of $X$).
\subsubsection{Distances to the simplexes do not distinct non-isometric spaces}
It is well-known that the set of pairwise distances does not completely define the corresponding metric space. One of simple examples can be obtain as follows. On the set $\{x_1,x_2,x_3,x_4\}$ we consider the following two distance matrices that differ in the transposition of the distances $|x_1x_4|$ and $|x_3x_4|$:
$$
S_1=\left(\begin{array}{cccc}
0 & a & b & d \\
a & 0 & c & e \\
b & c & 0 & f \\
d & e & f & 0
\end{array}\right), \qquad
S_2=\left(\begin{array}{cccc}
0 & a & b & f \\
a & 0 & c & e \\
b & c & 0 & d \\
f & e & d & 0
\end{array}\right).
$$
Notice that if all the pairwise distances are close to each other, then the triangle inequalities are valid for the both matrices.
Assume that $a<b<c<d<f<e$, and let $X_i$ be the space with the distance matrix $S_i$, $i=1,\;2$. The sets of the pairwise distances of these spaces are the same, but the spaces are not isometric. For example, the unique minimum spanning tree in $X_1$ is the star--tree centered at the point $x_1$, and the unique minimum spanning tree in $X_2$ is the path $x_2x_1x_3x_4$. But the spectra $\sigma$, the maximum spanning trees, and the spectra $\Sigma$ of these spaces are the same.
\begin{ass}\label{ass:examp_non_isom}
Under the above assumptions, the Gromov--Hausdorff distances from the spaces $X_i$ to any simplex are the same.
\end{ass}
\begin{proof}
Due to Theorem~\ref{thm:dist-n-simplex-bigger-dim}, the distances from $X_i$ to the simplexes consisting of five and more vertices are the same, because $\operatorname{diam} X_1=\operatorname{diam} X_2=e$. The distances to the simplexes $t\Delta_4$ are the same in accordance to Theorem~\ref{thm:dist-n-simplex-same-dim}, because the largest and the smallest distances in the spaces $X_i$ are the same. Further, the distances to the simplexes $t\Delta_3$ are the same in accordance to Theorem~\ref{thm:minus_one}, because the spaces $X_i$ have equal spectra $\sigma$ and $\Sigma$. The distances to the single-point simplex are the same, due to Proposition~\ref{prop:GH_simple}, Item~(\ref{prop:GH_simple:1}), because the diameters of the spaces $X_i$ are the same. It remains to calculate the distances to two-point simplexes. It can be done similarly to the above examples.
For the space $X_1$ the distance can be calculated as follows:
\begin{multline}\label{mult:3}
2d_{GH}(t\Delta_2,X_1)=\min\Big\{\max\big\{e,t-a,d-t\big\},
\max\big\{f,t-a,e-t\big\},
\max\big\{e,t-b,f-t\big\}, \\
\max\big\{c,t-d,e-t\big\},
\max\big\{f,t-b,e-t\big\},
\max\big\{e,t-a,f-t\big\},
\max\big\{d,t-a,e-t\big\}\Big\};
\end{multline}
and for the space $X_2$ the distance has the form
\begin{multline}\label{mult:4}
2d_{GH}(t\Delta_2,X_2)=\min\Big\{\max\big\{e,t-a,f-t\big\},
\max\big\{f,t-a,e-t\big\},
\max\big\{e,t-b,d-t\big\}, \\
\max\big\{c,t-d,e-t\big\},
\max\big\{d,t-b,e-t\big\},
\max\big\{e,t-a,f-t\big\},
\max\big\{f,t-a,e-t\big\}\Big\}.
\end{multline}
Notice that, due to the chosen order, all the maxima in the right hand parts of Formulas~(\ref{mult:3}) and~(\ref{mult:4}) are not less than the fourth maximum $\max\big\{c,t-d,e-t\big\}$, so
$$
2d_{GH}(t\Delta_2,X_1)=\max\big\{c,t-d,e-t\big\}=2d_{GH}(t\Delta_2,X_2).
$$
\end{proof}
\begin{cor}\label{cor:non_isom}
In the space of four-point metric spaces there exists an open subset $U$ such that for any $X\in U$ one can find a four-point metric space $X'$ that is not isometric to $X$, and such that $d_{GH}(X,t\Delta_m)=d_{GH}(X',t\Delta_m)$ for all $t>0$ and $m\in{\mathbb N}$.
\end{cor}
\begin{rk}
Notice that the distances to simplexes from the spaces $X_i$ do not depend on $f$, and so, we have found in fact an infinite (continuum) family of pairwise non-isometric finite metric spaces such that the distances to all simplexes are the same. A natural problem is to describe all such families of metric spaces. In particular, to find out if it is possible to construct similar examples for infinite metric spaces.
\end{rk}
|
\section{Introduction}
\label{sec:introduction}
Searches for new physics at the energy frontier often look for new phenomena at the edge of
distributions. In this kinematical region the knowledge of the Standard Model (SM) background is
typically limited and the challenge is to look for a new resonance where only partial knowledge on
the SM background is available. In this paper we focus in particular on new physics probes based on
the high diphoton invariant mass spectrum. We examine, using the analyses of the 750\,GeV diphoton
resonance as a case study, the strategy currently used by the experimental collaborations in
estimating the dominant SM backgrounds. We employ our methodology on the 13\,TeV LHC dataset to
asses the systematics involved in the current diphoton searches beyond the TeV region\par
In their 2015 data sets, both ATLAS and CMS observed an excess in the diphoton spectrum near
$m_{\gamma \gamma}=750\,\ensuremath{\,\text{GeV}}\,.$ The relevant details of the ATLAS and CMS analyses are described
in~\cite{Aaboud:2016tru, Khachatryan:2016hje}. At face value the local significances for a broad
resonance were given by
\begin{equation}
\label{sigs}
\begin{aligned}[c]
p_{\textrm{ATLAS}}&=4 \times 10^{-5}\,,\\
\sigma_{\textrm{ATLAS}}& = 3.9\,,
\end{aligned}
\qquad
\begin{aligned}[c]
p_{\textrm{CMS}} &= 5\times 10^{-3}\,,\\
\sigma_{\textrm{CMS}} &= 2.6\,,
\end{aligned}
\qquad
\begin{aligned}[c]
p_{\textrm{comb}}&= 1 \times 10^{-6}\,,\\
\sigma_{\textrm{comb}}&= 4.7\,,
\end{aligned}
\end{equation}
where $p_{\textrm{ATLAS, CMS, comb}}$ $(\sigma_{\textrm{ATLAS, CMS, comb}})$ correspond to the local
$p$ value (confidence level) of ATLAS, CMS, and their naive combination.\footnote{We do not discuss
here the global significance as it strongly depends on the lower value of $m_{\gamma \gamma}$
defined for the search region. ATLAS (CMS) chose it to be about 200\,GeV (400\,GeV). Furthermore,
as discussed below, the region below 500\,GeV is dominating the fit to the functional form which
is used to estimate the background. Thus, it is not clear whether one should consider this region
as a control region or as the region of interest for the search itself.}
The local results quoted in Eq.~\eqref{sigs} are quite significant and captured the attention of the
high energy community. Interpreting them naively, one would be lead to one of the following
conclusions
\begin{enumerate}[(i)]
\item this excess is a result of a rare statistical fluctuation;
\item \label{NPclaim} this excess implies a discovery of non-Standard Model dynamics.
\end{enumerate}
As both conclusions are quite extraordinary (certainly the second one), they motivate an
investigation into their robustness. In particular, we raise a third option, to be considered in
conjunction with (i), namely, we ask how unlikely is the possibility that
\begin{enumerate}[(i),resume]
\item \label{BKGclaim} the significance of the excess is overestimated due to underestimating
fake-based backgrounds.
\end{enumerate}
With the inclusion of more data in the analyses the excess eventually vanished~\cite{ATLAS:2016eeo,
CMS:2016crm}, ruling out the new physics hypothesis \ref{NPclaim}. However, the possibility of
claim \ref{BKGclaim} remains unclear, affecting all analyses which rely on a precise knowledge of
the photon faking background and use the same techniques to estimate it.\par
While our conclusion is independent of the 750\,GeV resonance we use it as an example case to
scrutinize the hypothesis of the underestimated background and its implications. First, the main
rationale behind our hypothesis is presented in Sec.~\ref{sec:rationale}, followed by a detailed
description of our approach to background estimation (Sec.~\ref{sec:setup}) and the statistical
treatment of the data (Sec.~\ref{sec:stat-treatm}). The comparison with the full 2016 data set is
presented in Sec.~\ref{sec:post-ichep-update}. Our main conclusions are summarized in
Sec.~\ref{sec:conclusions}. For other relevant works, see
Refs.~\cite{Davis:2016hlw,Bondarenko:2016rvd}.
\section{The rationale}
\label{sec:rationale}
Superficially, the experimental situation related to the diphoton excess was fairly
straightforward. The experiments had reported a relatively narrow ``bump'',
$\Gamma/m\lesssim 6\%\ll1$. Such a bump implies a rise in the differential distribution while, due
to the rapidly falling parton luminosity functions, it is expected that any reasonable
background-related distribution should be a monotonically decreasing function of the invariant
mass. Consequently, the presence of a non-Standard Model feature seemed to have been indicated by
the measurements. While this was qualitatively correct the challenge is to quantify the significance
of the excess. To endow the bump with a significance, one needs to control and quantify the
background.\par
The following approaches can be used to constrain the form of the background:\\
\textbf{I.} \textit{Data-driven approach.} Assuming $\Gamma/m\ll1$ and a featureless monotonic
background, a robust way to constrain it is through \emph{interpolation} via a two-sided side band
analysis. However, this requires to have enough measured events at invariant masses both below the
resonance and above it. In the case of the 750\,GeV excess, there were less than 40 events in all of
the analyses measured with invariant masses above 850 GeV. Such a small number of events does not
allow one to use this method reliably. \\
\textbf{II.} \textit{"First-principle"/Monte-Carlo approach.} There is a rather narrow class of
observables for which the theory has reached an advanced enough level such that we can fully trust
our ability to correctly predict the shape of the background distributions. We believe that the
invariant mass distribution of experimentally measured diphoton events does not (yet) belong to this
selected class of observables. Namely, the continuous diphoton distribution consist of an admixture
of two dominant components: (i) The first is made of two real isolated hard photons. This diphoton
distribution is currently known to next-to-next-to-leading order (NNLO)
accuracy~\cite{Catani:2011qz,Campbell:2016yrh} in perturbative QCD and imposing cuts similar to the
ATLAS spin-0 analysis suggests an overall uncertainty of about 5\% for the invariant mass
distribution~\cite{Campbell:2016yrh}. (ii) An additional important background component is due to
fakes coming mostly from processes involving a hard photon and a jet that passes the various photon
quality and isolation cuts~\cite{Aad:2016xcr}. In addition, depending on these cuts, also the dijet
background could play an important role. The prompt photon-jet cross section is currently known at
next-to-leading order (NLO) in QCD, and several codes are available to produce the relevant
distributions, including JetPhox~\cite{Catani:2002ny} and PeTeR~\cite{Becher:2013vva}. In addition,
QCD threshold resummation at next-to-next-to-next-to-leading logarithmic (N${}^3$LL)
order~\cite{Becher:2012xr,Becher:2015yea} as well as electroweak Sudakov effects are being
included~\cite{Schwartz:2016olw}, resulting in theory uncertainties of about
10-20\%~\cite{Schwartz:2016olw,Aad:2016xcr,Khachatryan:2015ira}. However, a comparison with the
8\,TeV ATLAS measurement~\cite{Aad:2016xcr} shows that at low photon $p_T\sim 50\,$GeV the data
exhibits some level of deviations from the theoretical predictions (a larger uncertainty is found
for the invariant mass distribution, see~\cite{Aad:2013gaa}). In addition, it is important to note,
that the fake rate strongly depends on the quark/gluon ``flavor" of the tagged jet (for some
discussion on jet flavor definitions
see~\cite{Banfi:2006hf,Buckley:2015gua,Larkoski:2014pca,Bauer:2013bza}): intuitively one can
understand the difference through the quark and gluon fragmentation functions to pions. At large
$x$, as required to be able to pass photon isolation criteria, gluon fragmentation to few pions is
much more suppressed (see e.g. Chapter 20 in Ref.~\cite{Agashe:2014kda}). Accordingly, a dedicated
ATLAS study~\cite{ATLAS:2011kuc} found that there is a probability of about $1:2\times10^3$ for a
quark jet to fake a photon, and only $1:2\times10^4$ for a gluon jet to fake a photon, for jets with
$E_T>40\,$GeV. Applying this to the photon-jet background, we also note that subleading jets might
become an important source of fakes if the leading jet is predominantly gluon-initiated.\par
In order to theoretically predict the purity of the diphoton mass distribution, an appropriate
admixture of the diphoton and the photon-jet(s) components needs to be
constructed~\cite{Neufeld:2010fj}. Furthermore, for the latter component, one is required to
convolve the photon-jet distribution with the relevant fragmentation functions or at least tag the
flavor of the jet(s). It is also important to note that the purity is a highly phase space
dependent quantity. Not only does it depend on the ratio of the differential jet-photon and
photon-photon production but also on the jet-to-photon fake rate. The fake rate may exhibit strong
dependence on the differential quantities such as $p_T$ and (pseudo)rapidity $\eta$. For instance,
as discussed below, in the CMS analyses purity is estimated to be better than 90\% in the
(central-central) EBEB event category but only better than 80\% in the (forward-central) EBEE
one. Both experiments consider the purity in an inclusive way. However, in the relevant kinematical
region the data is not sufficient to constrain possibly large deviations from the inclusive purity
estimation (see Fig.~\ref{fig:purity}).\\
\textbf{III.} \textit{Functional-fit approach.} Given the present practical limitations of the
methods \textbf{I} and \textbf{II}, one is lead to a more phenomenological approach in which the
background estimate is obtained by fitting an universal function to control regions in the data and
then \emph{extrapolating} into the signal regions using the fitted functional form. This allows one
to predict the background at relatively high invariant masses in a straightforward
manner. Consequently, both experiments are essentially following this approach in most of their
analyses\footnote{An exception is the ATLAS spin-2 analysis which employs a Monte Carlo approach
(\textbf{II}) with a data-driven estimate of the photon-jet and jet-jet background, see Section
\ref{sec:stat-treatm}.}, although the functional forms used by ATLAS in the spin-0 analysis and by
CMS are slightly different. Thus, the significance of the excess is mostly determined by comparing
measured events to a background estimate predicted by a fitting function.\\
While method \textbf{III} is very transparent and makes the search for bumps easy to analyze, it is
also rather susceptible to systematic effects, in particular a lack of understanding of the physics
modifying the tails of the distributions, as we argue below. The fitting functions used by ATLAS
and CMS are well suited for describing rapidly falling distributions and are fitted to the available
data. With the amount of data in the 2015 data sets, the differentially measured number of events is
abundant in the low invariant mass region and is spare in the high mass region. The extraction of
the functions' parameters is thus dominantly controlled by the low $m_{\gamma \gamma}$ region and
hardly affected by modifications of the invariant mass distribution at diphoton masses of above
roughly 500\,GeV. However, the significance of the excess with respect to the fitting function is
very much affected by such deformations. As it is hard to directly test or predict the correct form
of the diphoton mass distribution this raises the
following questions: \\
\textit{ ${\mathcal{I}}$. Is the experimental signal over background estimation robust against the
presence of deviations from the fitting function predictions at large invariant masses? \\
${\cal II}$. If this is not the case, can one produce smoking-gun predictions to show that indeed
the significance of the excess is being overestimated?}
Let us first focus on point ${\cal I}$. To examine the sensitivity of the significance of the excess
to the variation of the tails of the distributions. We consider a family of background shapes that
are formed by an admixture of the diphoton and photon-jet distributions. We keep the overall
inclusive purity of the samples at 90\% and 80\%, respectively, in accordance with the measured data
at low invariant masses. More specifically, we use two classes of deformations. The first is
derived from a modification of the photon-jet spectrum due to NLO and showering effects combined
with an increased fake rate for larger transverse momenta and pseudo-rapidities of the jets.\par
We then consider a simpler ansatz where we allow the distribution of the $p p \to \gamma j$
component to be reweighted at invariant masses above 500\,GeV such that the purity of events with
large invariant masses is reduced leading to a controlled deviation from the functional fit. In the
following section we provide a detailed description of our approach. We also provide some tests of
our procedure to check that our method complies with public data (below and above the resonance
region) and is passing the relevant statistical tests. We then report how the significance is
affected by the amount of rescaling of the distributions of fakes. Finally we can use our ansatz to
address item ${\cal II}$ and provide smoking guns to test our hypothesis on overestimating the
excess significance. With the full statistics of the 2016 data sets at hand it would be fairly easy
to eliminate our hypothesis.
\section{Reducible and irreducible backgrounds}
\label{sec:setup}
The main background to the diphoton signal is the irreducible $ p p \to \gamma\gamma$ background. We
consider in the following: ATLAS spin-0 and spin-2, and CMS $13\,\ensuremath{\,\text{TeV}}$ EBEB and EBEE categories with
magnets on. We generate the diphoton invariant mass spectrum at NNLO with MCFM version 8.0
\cite{Campbell:1999ah,Campbell:2011bn,Campbell:2015qma,Boughezal:2016wmq,Campbell:2016yrh} applying
the cuts as described in the respective analyses, see Table \ref{tab:cuts}. The main contribution to
the reducible background is the $p p \to \gamma j$ production where the hard jet is a quark jet that
is wrongly reconstructed as a photon. We generate this background at leading order (LO) with
MadGraph5 version 5.2~\cite{Alwall:2014hca}. We note that at LO the $p p \to \gamma j$ sample is
dominated by quark jets, which as already mentioned, lead to a much larger fake rate than gluon
jets.
\begin{table}[tb]
\centering
\begin{tabular}[t]{|l|c|c|c|c|}
\hline
Analysis & ATLAS spin-0 & ATLAS spin-2 & CMS EBEB & CMS EBEE \\\hline
$m_{\gamma\gamma}$ & $> 150\,\ensuremath{\,\text{GeV}}$ & $>200\,\ensuremath{\,\text{GeV}}$ & $>230\,\ensuremath{\,\text{GeV}}$ & $>330\,\ensuremath{\,\text{GeV}}$ \\
$p_{T,1}$ & $>0.4\,m_{\gamma\gamma}$ & $>55\,\ensuremath{\,\text{GeV}}$ & $>75\,\ensuremath{\,\text{GeV}}$ & $>75\,\ensuremath{\,\text{GeV}}$ \\
$p_{T,2}$ & $>0.3\,m_{\gamma\gamma}$ & $>55\,\ensuremath{\,\text{GeV}}$ & $>75\,\ensuremath{\,\text{GeV}}$ & $>75\,\ensuremath{\,\text{GeV}}$ \\
$\left|\eta_1\right|$ & $<2.37$ & $<2.37$ & $<1.44$ & $<1.44$ \\
$\left|\eta_2\right|$ & $<2.37$ & $<2.37$ & $<1.44$ & $1.57<\eta_2<2.5$ \\
$\left|\eta\right|$ excluded & $1.37<\eta_{1,2}<1.52$ & $1.37<\eta_{1,2}<1.52$ & n.a. & n.a. \\\hline
$\sigma_{\gamma\gamma}$ [pb] (NNLO) & 2.7 & 1.9 & 0.52 & 0.23\\
$\sigma_{\gamma j}$ [pb] (NLO) & 1400 & 1000 & 250 & 130 \\
$\left.{\sigma_{\gamma j}}{}/{\sigma_{\gamma\gamma}}{}\right|_{m>500\,\ensuremath{\,\text{GeV}}_{}}$ & 510 &
670
& 470 & 640
\\\hline
\end{tabular}
\caption{Cuts of the analyses where the subscript refers to the hardest and second hardest photon
candidate, the cross section of the $p p \to \gamma\gamma$ sample passing these cuts (calculated
at NNLO with MCFM) and of the $p p \to \gamma j$ sample at hadron level (before applying any
photon mistag rate), calculated at NLO with MadGraph5\_aMC@NLO, showered with
Pythia~\cite{Sjostrand:2006za} and the jets clustered with anti-$k_T$, $R=0.4$ algorithm using
FastJet~\cite{Cacciari:2011ma}. In the last line the ratios of the two distributions in the
invariant mass region above 500\,GeV are given.}
\label{tab:cuts}
\end{table}
The reconstructed diphoton distribution is a mixture of $p p \to \gamma\gamma$ and
$p p \to \gamma j$ invariant mass distributions. Let us define a short-hand notation for the
normalized invariant mass distribution
\begin{equation}
\label{eq:2}
w_{\gamma X}\equiv \frac{1}{\sigma_{\gamma X}} \times \frac{\textnormal{d}\sigma_{\gamma X}}{{\textnormal{d} m_{\gamma X}}}\,,
\end{equation}
with $X=\gamma,j$. The mixed distribution $w_\textnormal{mix}$ is a function of the normalisation
$N_\textnormal{mix}$ and a parameter $\mathcal{R}$, that controls the shape modification of $w_{\gamma j}$
and will be defined Eq.~\eqref{eq:1}. We write $w_\textnormal{mix}$ as
\begin{equation}
\label{eq:5}
w_\textnormal{mix}(N_{\textrm{mix}}) =N_{\textrm{mix}} \left[\mathcal{P}w_{\gamma\gamma}+(1-\mathcal{P})w_{\gamma j} \right]\,,
\end{equation}
where $\mathcal{P}$ is the inclusive purity of the sample. We set $\mathcal{P}=90\%$ (80\%) for the
ATLAS and CMS EBEB (EBEE) analyses, which is within the reported error bands. We will assume, that
$w_{\gamma\gamma}$ is obtained by normalizing the MCFM diphoton invariant mass distribution. As for
$w_{\gamma j}$, following the rationale described in Section \ref{sec:rationale}, we modify in two
different ways as we now describe in detail.
\subsection{QCD and jet-fake dependence of the diphoton shape}
First, we calculate a photon-jet mass dependent K-factor using MadGraph5\_aMC@NLO, showered and
hadronized with Pythia~\cite{Sjostrand:2006za} and jets clustered with an anti-$k_T$, $R=0.4$
algorithm~\cite{Cacciari:2008gp} using FastJet~\cite{Cacciari:2011ma}. We note that in the NLO
distribution we only consider the hardest jet of the event and we do not record its flavor. This
step may be potentially improved by the use of an IR-safe jet flavor definition,
see~\cite{Banfi:2006hf,Buckley:2015gua,Larkoski:2014pca,Bauer:2013bza}. Next, in order to model the
dependence of the fake rate on the pseudo rapidity and the transverse momentum we use the following
simplified ansatz for the jet rejection $r(p_T,\eta)$:
\begin{equation}
r(p_T,\eta) = {\textrm{max}} \left\{ \frac{r_0}{1+ p_T/p_T^0 + \eta/\eta^0}, r_{\textrm{min}}\right\}\,,
\end{equation}
where the functional form is motivated by the kinematical dependence of the jet-rejection rates as
estimated by ATLAS~\cite{ATLAS:2011kuc} and the parameter values $p_T^0 = 30$\,GeV, $\eta^0 = 4$ are
chosen to reproduce the rejection rate ratios between the lowest and highest lying $\eta$ and $p_T$
bins within uncertainties. Finally, $r_0/r_{\textrm{min}}$ is fairly uncertain as estimates of
rejection rates at very high $p_T$ and $\eta$ are not publicly available, but reproducing
experimental purity estimates in the forward region~\cite{Khachatryan:2016hje} leads to values in
the wide range $r_0/r_{\textrm{min}} \in [3,12]$\,. The resulting reweighting factors compared to
the LO partonic $m_{\gamma j}$ distribution obtained from MadGraph5, $w_{\gamma j}^{\textnormal{MG}}$, at
both steps applied successively ($w_{\gamma j}^{\textrm{NLO}}/w_{\gamma j}^{\textrm{MG}}$ and
$w_{\gamma j}^{\textrm{NLO}\,\times\,\textrm{fakes}}/w_{\gamma j}^{\textrm{MG}}$) are shown in
Fig.~\ref{fig:reweightingfactor}. We observe that with our choice of fake rate parameters, the
largest reweighting factors close to 3 are obtained above $m_{\gamma j} > 800$\,GeV for the ATLAS
spin-2 cuts. However, all experimental categories are affected by a reweighting factor which is a
combination of NLO, hadronization and faking effects, and which increases with the photon-jet
invariant mass until it saturates at some point. This suggests a simple functional form for the
effective photon-jet spectrum deformation which we discuss next.
\begin{figure}[tb]
\centering
\begin{subfigure}[t]{0.48\linewidth}
\begin{minipage}[t]{1.0\linewidth}
\vspace{5pt}\includegraphics[width=\textwidth]{figures/NLOtimesFakesReweighting_ATLAS}
\end{minipage}
\end{subfigure}
\begin{subfigure}[t]{0.50\linewidth}
\begin{minipage}[t]{1.0\linewidth}
\vspace{0pt}\includegraphics[width=\textwidth]{figures/NLOtimesFakesReweighting_CMS}
\end{minipage}
\end{subfigure}
\caption{Reweighting factor for $w_{\gamma j}$ with respect to the LO Monte Carlo distribution as
a function of the invariant mass for ATLAS (left) and CMS (right). The dashed lines show the
reweighting factor to modify the LO parton distribution $w_{\gamma j}^\textnormal{MG}$ to the NLO shape, including the effects of
hadronization. The solid lines include in addition the reweighting due to a phase space
dependent fake rate.}
\label{fig:reweightingfactor}
\end{figure}
\subsection{Effective shape deformation}
In our effective ansatz for the deformation of the photon-jet spectrum, we focus on the invariant
mass region above $m_{\gamma\gamma} > 500$\,GeV below which the experiments have sufficient
statistics to control the mass distributions and calibrate their analyses and choose a simple,
linear form
\begin{align}
\label{eq:1}
w_{\gamma j}(\mathcal{R})&=\frac{w_{\gamma j}^{\textrm{MG}}}{N(\mathcal{R})} \left[ 1+\mathcal{R} \times
\begin{cases}
0 & m_{\gamma j} < 500\,\ensuremath{\,\text{GeV}}\\
\frac{\left.w_{\gamma j}^\textnormal{MG}\right|_{m_{\gamma j}=500\,\ensuremath{\,\text{GeV}}}}{w_{\gamma j}^\textnormal{MG}}-1 & 500\,\ensuremath{\,\text{GeV}} \leq
m_{\gamma j} \leq 760\,\ensuremath{\,\text{GeV}}\\
\frac{\left.w_{\gamma j}^\textnormal{MG}\right|_{m_{\gamma j}=500\,\ensuremath{\,\text{GeV}}}}{\left.w_{\gamma
j}^\textnormal{MG}\right|_{m_{\gamma j}=760\,\ensuremath{\,\text{GeV}}}} -1 & 760\,\ensuremath{\,\text{GeV}} < m_{\gamma j}
\end{cases}
\right]\,,
\end{align}
to roughly account for an overall kinematic dependence of the fake rate, hadronization and higher
order effects $N(\mathcal{R})$ is an $\mathcal{R}$-dependent normalisation factor. We define
$w_{\gamma j}(\mathcal R)$ such that $w_{\gamma j}(\mathcal{R}=0)$ corresponds to the partonic LO
distribution. Choosing $\mathcal{R}=1$, the shape of $w_{\gamma j}$ is unmodified for
$m_{\gamma j}<500\,\ensuremath{\,\text{GeV}}$, then flat up to $m_{\gamma j}=760\,\ensuremath{\,\text{GeV}}$, just above the observed peak of
the apparent excess, and finally is rescaled by the ratio of the differential cross sections at
$m_{\gamma j}=500\,\ensuremath{\,\text{GeV}}$ and $m_{\gamma j}=760\,\ensuremath{\,\text{GeV}}$ for $m_{\gamma j}>760\,\ensuremath{\,\text{GeV}}$. The
$\mathcal{R}$-dependence of $N(\mathcal{R})$ is chosen such that the integral over $w_{\gamma j}$ is
always 1, independent of the value of $\mathcal{R}$. For the ATLAS spin-2 and the two CMS analyses,
the intervals in the above equation are shifted by $10\,\ensuremath{\,\text{GeV}}$ to larger values due to the different
binning in these searches.\par
In the left panel of Fig.~\ref{fig:ReweightATLAS0}, we show the normalized differential
$p p \to \gamma j$ cross section for the ATLAS spin-0 analysis as a function of the invariant mass
for several choices of $\mathcal{R}$. In the right panel, the $\mathcal{R}$-dependent reweighting
factor of Eq.~\eqref{eq:1} is shown. Since the spin-0 analysis applies the strongest cuts on the
transverse momenta of the photon candidates ($0.4\,m_{\gamma\gamma}$ and $0.3\,m_{\gamma\gamma}$,
respectively) its distribution is the steepest. Thus the reweighting factor of this analysis is the
largest being almost 7 above 770\,GeV. The maximal reweighting factors for the other analyses are
just above 6. We verified that increasing the flat region by 20\,GeV has only a small impact on the
reported results.
\begin{figure}[tb]
\centering
\begin{subfigure}[t]{0.55\textwidth}
\begin{minipage}[t]{\textwidth}
\vspace{0pt}\includegraphics[height=141pt]{figures/0GaJ_Envelopes}
\end{minipage}
\caption{}
\label{fig:NormalizedDists}
\end{subfigure}
\begin{subfigure}[t]{0.42\textwidth}
\begin{minipage}[t]{1\textwidth}
\vspace{1pt}\includegraphics[height=136pt]{figures/0ReweightingFactor}
\end{minipage}
\caption{}
\label{fig:reweightingfunction}
\end{subfigure}
\caption{Left: Normalized invariant mass distributions after the ATLAS spin-0 cuts for the
$p p \to \gamma j$ background with several choices for the interpolating parameter $\mathcal{R}$. Right:
Corresponding reweighting factor.}
\label{fig:ReweightATLAS0}
\end{figure}
In Fig.~\ref{fig:MixedDists} and Fig.~\ref{fig:1MixedDists}, we show the resulting invariant mass
distributions $w_{\textnormal{mix}}$ for the ATLAS spin-0 and CMS EBEB analysis, respectively, on top of the
normalized distribution as measured in the 2015 data set.
\begin{figure}[tb]
\begin{subfigure}[t]{0.58\linewidth}
\centering
\vspace{0pt}\includegraphics[height=130pt]{figures/0MixedDistributions}
\caption{ATLAS spin-0}
\label{fig:MixedDists}
\end{subfigure}
\begin{subfigure}[t]{0.4\linewidth}
\centering
\vspace{0pt}\includegraphics[height=130pt]{figures/0chisquared}
\caption{ATLAS spin-0}
\label{fig:chisquared}
\end{subfigure}
\begin{subfigure}[t]{0.6\linewidth}
\centering
\vspace{0pt}\includegraphics[height=132pt]{figures/1MixedDistributions}
\caption{CMS EBEB}
\label{fig:1MixedDists}
\end{subfigure}
\begin{subfigure}[t]{0.39\linewidth}
\centering
\vspace{0.8pt}\includegraphics[height=130pt]{figures/CMSchisquared}
\caption{CMS EBEB and EBEE combined}
\label{fig:CMSchisquared}
\end{subfigure}
\caption{Top left: Combined distributions $w_\textnormal{mix}$ for the ATLAS spin-0 cuts, all with an
overall purity of 90\%. The distribution obtained from the 2015 data sets is shown in
blue. Bottom left: Corresponding distribution for the CMS EBEB analysis. Top right:
$\Delta\chi^2$ of fit to the ATLAS spin-0 distribution as a function of $\mathcal{R}$. The 1-
and 2-$\sigma$ regions are indicated by the thin lines. Bottom right: corresponding plot for the
combined CMS EBEB and EBEE analyses.}
\end{figure}\par
In addition to the average purity of the full sample, ATLAS and CMS try to estimate the purity as a
function of the diphoton invariant mass. This local purity is given by
\begin{equation}
\label{eq:3}
\mathcal{P}^i=\frac{\mathcal{P} w^i_{\gamma\gamma}}{\mathcal{P} w^i_{\gamma\gamma} + (1-\mathcal{P}) w^i_{\gamma j}}
\end{equation}
for the $i$-th bin. It can deviate significantly from the average purity $\mathcal{P}$ of the full
sample. In Fig.~\ref{fig:purity}, we show the binned purities for the mixed samples with several
choices of $\mathcal{R}$ compared to the purity determined by ATLAS with the $2\times 2$ sideband
\cite{Aad:2012tba} and the matrix method \cite{Aad:2011mh} and by CMS with a method described in
\cite{Chatrchyan:2014fsa}, respectively. We choose the same binning of the purity as is used in the
respective analysis.
\begin{figure}[tb]
\centering
\begin{subfigure}[c]{0.49\linewidth}
\centering
\begin{minipage}[t]{1\textwidth}
\includegraphics[width=0.95\textwidth]{figures/0BinnedPurity_NewMethod}
\end{minipage}
\caption{ATLAS spin-0\qquad\qquad\quad\null}
\label{subfig:Purity_0}
\end{subfigure}
\centering
\begin{subfigure}[c]{0.49\linewidth}
\centering
\begin{minipage}[t]{1\textwidth}
\vspace{1pt}\includegraphics[width=\textwidth]{figures/2BinnedPurity_NewMethod}
\end{minipage}
\caption{ATLAS spin-2\qquad\qquad\quad\null}
\label{subfig:Purity_2}
\end{subfigure}
\vspace{20pt}
\centering
\begin{subfigure}[b]{0.49\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/1BinnedPurity_NewMethod}
\caption{CMS EBEB\qquad\qquad\quad\null}
\label{subfig:Purity_1}
\end{subfigure}
\centering
\begin{subfigure}[b]{0.49\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/3BinnedPurity_NewMethod}
\caption{CMS EBEE\qquad\qquad\quad\null}
\label{subfig:Purity_3}
\end{subfigure}
\caption{Purity of the combined distribution as a function of $m_{\gamma\gamma}$ for several
choices of $\mathcal{R}$. The dashed lines show the central value for the purity as determined
by the experiments and shaded the areas show the corresponding error.}
\label{fig:purity}
\end{figure}\par
While the local purity is within the error band in most of the considered mass range (even for
$\mathcal{R}=1$), it does decrease for large invariant masses and our ansatz predicts a deviation
from the experimental value. Given the low statistics in this range, we consider this as a way to
falsify our proposal in the future rather than a contradiction with the currently available data.
\section{Statistical treatment}
\label{sec:stat-treatm}
The experimental analyses estimate the background shape by fitting a function $f(x)$ with
$x=m_{\gamma\gamma}/\sqrt{s}$ to the measured data. In the ATLAS spin-0 analysis, the following
ansatz is used:
\begin{align}
f(x)=N\left(1-x^{1/3}\right)^b x^{\sum_{j=0}^k a_j \left(\log x\right)^j}\label{ATLASfun}
\end{align}
where $k=0$ was chosen. For the CMS analyses as well as the ATLAS spin-2, we use
\begin{align}
f(x)=Nx^{a+b\log x}\,.\label{CMSfun}
\end{align}
Note, that the ATLAS spin-2 analysis uses a mixture of Monte Carlo (for the $p p \to \gamma\gamma$
background) and data driven distributions (for the $p p \to \gamma j$ and $p p \to jj$ background),
leading to similar results as the fit function approach. In the data driven method, the shape of the
different backgrounds is obtained by extracting the corresponding events from control samples and
fitting their distribution with a function. The relative contribution to the observed
$pp \to \gamma\gamma$ sample is extracted from the data between
$200\,\ensuremath{\,\text{GeV}}< m_{\gamma\gamma}<500\,\ensuremath{\,\text{GeV}}$. For more details on this method see
\cite{Aaboud:2016tru}. Given the small statistics in the large invariant mass bins this approach
roughly corresponds to our LO MG distribution.\par
In order to see how the significance of the 750\,GeV excess changes with our ansatz, we fit the
distribution $w_{\textrm{mix}}$, defined in Eq.~\eqref{eq:5}, once with $w_{\gamma j}^{\textrm{MG}}$
corresponding to $\mathcal R=0$, and then with $w_{\gamma j}^{}$ as estimated at NLO with showering
and hadronization, including fakes and finally with $\mathcal{R}$ as a free fit parameter (as well
as the appropriate fit function $f(x)$) to the measured data. As an additional template, one could
extend the fit function $f(x)$ by a modification similar to the one described in Eq.~\eqref{eq:1},
which we will however not do for the sake of simplicity. The fits are performed with two methods
which yield similar results.
Firstly, we maximize the likelihood
\begin{align}
L=\prod_{i=1}^{N_\textnormal{bins}} P_{N_\textnormal{e}^i}(N_\textnormal{m}^i)
\end{align}
where the product goes over all bins and $P_{N_\textnormal{e}}(N_\textnormal{m})$ is the Poisson probability to
measure $N_\textnormal{m}$ events when $N_\textnormal{e}$ events are expected.
Secondly, we minimize
\begin{align}
\chi^2=\sum_{i=1}^{N_\textnormal{bins}} (N_\textnormal{m}^i-N_\textnormal{e}^i)^2/N_\textnormal{e}^i
\end{align}
where we rebin the data such that each bin contains at least 10 events in order for the $\chi^2$
distribution to provide a reasonable description of the statistical uncertainties, see
e.g.~\cite{Cowan:2010js}. In both fit methods, the overflow of the experimental histograms is
treated as one single bin. The best-fit parameters determine the number of expected events
$N_\textnormal{e}$ in the signal region (SR).\par
Since we are mostly interested in the local significance of the 750\,GeV excess, the SR is chosen by
eye from the measured distribution with the aim to capture the excess. We obtain a $p$-value by
comparing the number of measured events in the SR $N_\textnormal{m}$ with $N_\textnormal{e}$ (more precisely:
calculating the Poisson probability to measure at least $N_\textnormal{m}$ events):
\begin{equation}
\label{eq:6}
p=\sum\limits_{n=N_\textnormal{m}}^\infty \frac{N_\textnormal{e}^n}{n!} e^{-N_\textnormal{e}}\,.
\end{equation}
Clearly this simple approach which does not use any signal modeling and relies on a discrete width
and position of the SR is far from perfect. Consequently, it is not surprising that the obtained
significances of the excess are smaller than the ones reported by the experiments, even when we use
the same fit functions. Instead of focusing on absolute values one should therefore rather consider
the \textit{reduction} of the significance that results from modifying the background. The results
of the fits are shown in Table \ref{tab:stat_results}.\par
\begin{table}[tb!]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}[c]{|r|r|c|c|c|c|c|c|c|}
\hline
\multirow{3}{*}{}
&Analysis & \mc{ATLAS spin-0} & ATLAS spin-2 & \mc{CMS EBEB} &
\mc{CMS EBEE} \\\hhline{|=========|}
\parbox[t]{2mm}{\multirow{3}{*}{\rotatebox[origin=c]{90}{meas.}}}
&SR & 730-770\,GeV & 720-760\,GeV & 720-780\,GeV & \mc{710-770\,GeV} & \mc{710-770\,GeV} \\\hhline{~~-------}
&$\int \mathcal{L}\,\textnormal{d}t$ & $3.2\,\text{fb}^{-1}$ & $15.4\,\text{fb}^{-1}$& $3.2\,\text{fb}^{-1}$ & $2.7\,\text{fb}^{-1}$& $12.9\,\text{fb}^{-1}$& $2.7\,\text{fb}^{-1}$& $12.9\,\text{fb}^{-1}$ \\
&$N_m$ & 15 & 33 & 40 & 12 & 24 & 21 & 53 \\\hhline{|=========|}
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{fitfunction}}}
&$-2\log L$ & 270 & - & 330 & 200 & - & 200 & - \\
&$\sigma$ & 3.4 & - & 2.9 & 1.9 & - & 1.7 & - \\\hhline{~--------}
&$\chi^2/n$ & 1.6 & 0.75 & 1.2 & 0.80 & 1.0 & 1.2 & 0.98 \\
&$\sigma$ & $3.1_{-0.2}^{+ 0.2}$ & $1.2^{+0.2}_{-0.2}$ &$2.9_{-0.3}^{+ 0.3}$
&$1.8_{-0.2}^{+ 0.3}$ & $-1.5^{+0.2}_{-0.3}$ &$1.6_{-0.3}^{+ 0.3}$ & $-1.3^{+0.2}_{-0.2}$ \\\hhline{|=========|}
\parbox[t]{2mm}{\multirow{13}{*}{\rotatebox[origin=c]{90}{distribution}}}
&$\mathcal{R}$: \hfill$-2\log L$ & 270 & 360 & 330 & 210 & 310 & 210 & 300 \\
&$\mathcal{R}$ & 0.86 & 0.11 & 0.97 & -0.048 & -0.7 & 0.24 & -0.12 \\
&$\sigma$ & 2.2 & 0.0 & 2.0 & 1.5 & -1.2 & 1.4 & -0.23 \\\hhline{~--------}
&MG: \hfill$\chi^2/n$ & 1.5 & 0.76 & 1.3 & 0.78 & 1.6 & 1.1 & 1.1 \\
&$\sigma$ &$3.0_{-0.0}^{+ 0.0}$ & $0.2^{+0.0}_{-0.0}$ & $3.4_{-0.1}^{+ 0.1}$
&$1.4_{-0.1}^{+0.1}$ & $-2.6^{+0.1}_{-0.1}$ &$1.9_{-0.2}^{+ 0.2}$ & $-0.86^{+0.14}_{-0.14}$ \\\hhline{~--------}
&NLO: \hfill $\chi^2/n$& 1.5 & 0.76 & 1.2 & 0.80 & 1.6 & 1.2 & 1.0 \\
&$\sigma$ &$3.0_{-0.0}^{+0.0}$ & $0.3^{+0.0}_{-0.0}$ &$3.2_{-0.1}^{+0.1}$ &
$1.4_{-0.1}^{+0.1}$ &$-2.6^{+0.1}_{-0.1}$& $1.7_{-0.2}^{+0.2}$ & $-1.2^{+0.1}_{-0.1}$ \\\hhline{~--------}
&NLO$\times$fakes: \hfill$\chi^2/n$& 1.5 & 1.4 & 1.1 & 0.92 & 2.0 & 1.2 & 1.2 \\
&$\sigma$ &$2.7_{-0.0}^{+0.0}$ &$-0.3^{+0.0}_{-0.0}$ & $2.9_{-0.1}^{+0.1}$ & $1.2_{-0.1}^{+0.1}$ &$-3.0^{+0.1}_{-0.1}$ &$1.4_{-0.2}^{+0.2}$&$-1.7^{+0.1}_{-0.1}$ \\\hhline{~--------}
&$\mathcal{R}$: \hfill $\chi^2/n $& 1.2 & 0.75 & 1.0 & 0.81 & 1.3 & 1.1 & 1.1 \\
&$\mathcal{R}$ &$1.2_{-0.5}^{+ 0.6}$ & $0.2^{+0.2}_{-0.2}$ &$1.1_{-0.4}^{+ 0.4}$
&$-0.15_{-0.39}^{+ 0.51}$ &$-0.6^{+0.2}_{-0.2}$ &$0.30_{-0.22}^{+ 0.29}$ &$-0.091^{+0.084}_{-0.074}$ \\
&$\sigma$ &$2.0_{-0.4}^{+ 0.4}$ &$-0.2^{+0.4}_{-0.4}$ &$1.9_{-0.5}^{+ 0.5}$ &$1.5_{-0.4}^{+ 0.4}$ & $-1.3^{+0.4}_{-0.4}$ &$1.2_{-0.4}^{+ 0.4}$ &$-0.40^{+0.42}_{-0.42}$ \\
&$p_\textnormal{F-test}$ & 0.021 & 0.20 & 0.0027 & 0.73 & 0.014 &0.19 & 0.30 \\\hline
\end{tabular}
}
\caption{Results of the fits to the data of all four analyses. In the first block from the top the
signal region is defined and the number of measured events in this region $N_m$ is given. The
results of a likelihood- and a $\chi^2$ fit of the fit function (value of the maximal likelihood
and minimal $\chi^2$, respectively, and the local significance of the $750\,\ensuremath{\,\text{GeV}}$ excess) are
given in the second block. Finally, the third block contains the results of a likelihood and
$\chi^2$ fit of the background distributions described in Section \ref{sec:setup} to the
data. When $\mathcal{R}$ was fitted its best-fit value and the corresponding local significance
of the excess are given, otherwise just the significance. In the last line the result of the
F-test, testing whether $\mathcal{R}$ should be used as fit parameter, is given. For the
minimized $\chi^2$ the parameter $n$ is the difference of number of bins and fit parameters. The
errors indicate the 1-$\sigma$ interval of the systematic uncertainty of the fit. Note that the
results of the spin-0 analysis with $3.2\,\text{fb}^{-1}$ are based on the analysis with looser photon
identification as described in~\cite{Aaboud:2016tru}.}
\label{tab:stat_results}
\end{table}
Finally, an F-test is performed to determine if the generalization of our mixed distribution with
$w_{\gamma j}^{\textrm{MG}}$ to the one with $w_{\gamma j}(\mathcal{R})$ given in Eq.~\eqref{eq:1}
is needed to describe the data. This test investigates the improvement of a fit when the fit
function is extended by an additional parameter. For this purpose, a test statistic
\begin{align}
F=\frac{(\chi^2_1-\chi^2_2)/(n_1-n_2)}{\chi^2_2/n_2}
\end{align}
is calculated, where $\chi^2_{1,2}$ are the minimized $\chi^2$'s of the two fit functions, $n_{1,2}$
are the numbers of bins (27 for the ATLAS spin-0) minus the number of input parameters (3 vs. 4 for
the fitting function and 1 vs. 2 for our distribution), and the subscripts refer to the two fit
functions with 2 signifying the extended function. Eventually the $p$-value is determined as
\begin{align}
p_\textnormal{F-test}=\int_F^{\infty} \digamma(x; n_1-n_2,n_2)\textnormal{d}x,
\end{align}
with $\digamma$ being the Fisher distribution. An additional fit parameter is warranted if
$p_\textnormal{F-test}<5\%$, see~\cite{Aaboud:2016tru}.\par
We find that for the ATLAS searches, the F-test suggests that $\mathcal{R}$ should be included as a
fitting parameter. The probability of an accidental improvement due to $\mathcal{R}>0$ is only
$2.1\%$ (ATLAS spin-0) and $0.27\%$ (ATLAS spin-2). On the other hand, the CMS categories do not
prefer a significant non-zero $\mathcal R$, see Table \ref{tab:stat_results}. Furthermore, as a
consistency check, we apply the F-test on the ATLAS fitting function for spin-0,
Eq.~\eqref{ATLASfun}, and find that adding a $k=1$ component to the function does not pass the
test. Hence, as mentioned in~\cite{Aaboud:2016tru} only the leading term of the function with $k=0$
is retained. The above in conjunction with the results collected in Table 2 suggest that it is
possible that the basis of functions used in Eq.~\eqref{ATLASfun} is not sufficient to accommodate
the deformation of the distribution proposed by us (or at least not the first term in the functional
form).\par
In the $\chi^2$ fit of the constructed distribution with $w_{\gamma j}^{MG}$ we find similar results
for the local significance as with the $\chi^2$ fit of the functional approach. However, in
particular in the two ATLAS analyses, using $\mathcal{R}$ as an additional fit parameter reduces the
local significance of the 750\,GeV excess by 1-1.5 units. The fact that the reduction is stronger in
the spin-2 analysis corroborates our working assumption that the background description deteriorates
in the forward region. This is further supported by the observation that in the CMS EBEB analysis,
which collects only events with both photon candidates in the central region, no reduction in
significance is observed and the best fit value for $\mathcal{R}$ is even slightly negative. Only in
the CMS EBEE analysis where one photon candidate is in the forward region the significance is
reduced by fitting $\mathcal{R}$, albeit less than in the ATLAS analyses.\par
Since $\mathcal{R}>0$ flattens the distribution one might worry that the reduction in the local
significance is obtained by overshooting the measured distribution in the high invariant mass
region. By verifying that both the minimal $\chi^2$ and the maximal likelihood hardly change between
the functional and the distribution fit we show that this is not the case.\par
As a final exercise, we try to obtain a ``combined'' significance from the analyses of the 2015 data
set. Clearly a proper statistical combination cannot be done, since we neglect correlations between
the various analyses and also fit for a single universal value of $\mathcal{R}$. Realistically,
$\mathcal R$ is expected to be somewhat different for the different analyses since they cover
different regions of phase space. Nevertheless, since the naive combination in Eq.~\eqref{sigs}
suffers from similar issues we set them aside and proceed as follows. We sum the $\chi^2$'s of the
analyses included in the combination and fit for a common $\mathcal{R}$ while keeping the
normalizations as separate variables. By combining the two CMS analyses we obtain $\sigma=2.4$ (1.9)
for $w_{\gamma j}^{\textrm{MG}}$ (with $w_{\gamma j}(\mathcal{R})$, best fit $\mathcal{R}=0.22$) and
$\sigma=1.9$ with $w_{\gamma j}^{\textnormal{NLO}\times\textnormal{fakes}}$. A combination of the ATLAS analyses is
impossible since they are not independent. However, we can combine each of them with the two CMS
analyses and obtain for ATLAS spin-0 combined with CMS $\sigma=3.6$ (2.6 with
$w_{\gamma j}(\mathcal{R})$, best fit $\mathcal{R}=0.46$; 3.1 with
$w_{\gamma j}^{\textnormal{NLO}\times\textnormal{fakes}}$) and for ATLAS spin-2 combined with CMS $\sigma=4.2$ (2.8
with $w_{\gamma j}(\mathcal{R})$, best fit $\mathcal{R}=0.53$; 3.4 with
$w_{\gamma j}^{\textnormal{NLO}\times\textnormal{fakes}}$), where the significance numbers before the brackets are
obtained for $w_{\gamma j}^{\textrm{MG}}$.
\section{The new energy frontier: searches beyond 1 TeV}
\label{sec:post-ichep-update}
Around ICHEP 2016, ATLAS and CMS updated their analyses, now based on 15.4\,$\text{fb}^{-1}$ and 12.9\,$\text{fb}^{-1}$,
respectively. In the updated ATLAS spin-0 analysis \cite{ATLAS:2016eeo} and the CMS EBEB and EBEE
analyses \cite{CMS:2016crm} the large excess around $750\,\ensuremath{\,\text{GeV}}$ vanished and no other significant
excesses were found. An update of the ATLAS spin-2 analysis has not been presented. While CMS
processed the data exactly as before, ATLAS made some adjustments, perhaps most importantly, using a
tighter photon isolation. We repeat the fits and the statistical treatment of the reported results
with the same methods as described above and report the results for the larger dataset in Table
\ref{tab:stat_results}. Note that there is a downwards fluctuation in the signal region in the full
CMS dataset which even leads to a slightly negative significance.\par
\begin{figure}[tb]
\centering
\begin{subfigure}[b]{0.45\linewidth}
\centering
\includegraphics[width=1\textwidth]{figures/Spin0_ATLAS_comparison}
\label{subfig:comp_ATLAS}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.45\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/Spin1_CMS_EBEB_comparison}
\label{subfig:comp_EBEB}
\end{subfigure}\\
\begin{subfigure}[b]{0.45\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/FunctionRatios}
\label{subfig:small_oldvsnew}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.45\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/Spin3_CMS_EBEE_comparison}
\label{subfig:com_EBEE}
\end{subfigure}
\caption{Upper plots and lower right plot: Comparison of the measured and fitted distributions
with the small and the full datasets. In addition, the pure digamma spectrum as obtained from
MCFM is shown. In the upper left plot the comparison is between the smaller Moriond 2016 dataset
with the old photon isolation method and the full ICHEP 2016 dataset. The distributions and
functions are normalized to have the same value at the low $m_{\gamma\gamma}$ end of the
histograms.\\
The lower left plot shows the ratios of the normalized fit functions $f_X$ fitted to the ATLAS
spin-0 data set $X$ with the year 2015 (2016) in the brackets indicating the old Moriond
(updated ICHEP) photon isolation criteria.}
\label{fig:comparison}
\end{figure}
\begin{figure}[tb]
\centering
\begin{subfigure}[b]{0.32\linewidth}
\centering
\includegraphics[width=1\textwidth]{figures/7_Compare_fits}
\label{subfig:fitcomp_ATLAS}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/5_Compare_fits}
\label{subfig:fitcomp_EBEB}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.32\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/6_Compare_fits}
\label{subfig:fitcomp_EBEE}
\end{subfigure}
\caption{Best fit point of the $\chi^2$ fit of the appropriate function to the Moriond 2016
dataset in red with the 1- and 2-$\sigma$ contours. The best fit point for the fit to the full
ICHEP 2016 dataset is shown in blue and in the left plot the best fit point for the ATLAS spin-0
$3.2\,\text{fb}^{-1}$ dataset with the new photon isolation is shown in green.}
\label{fig:fit_comparison}
\end{figure}
\begin{figure}[tb]
\centering
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/5ErrorBands}
\label{subfig:ErrorBands_EBEB}
\end{subfigure}\hfill
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=\textwidth]{figures/6ErrorBands}
\label{subfig:ErrorBands_EBEE}
\end{subfigure}\\
\begin{subfigure}[b]{0.5\linewidth}
\centering
\includegraphics[width=1\textwidth]{figures/4ErrorBands}
\label{subfig:ErrorBands_ATLAS}
\end{subfigure}
\caption{Plot of the normalized distributions with several choices of $\mathcal{R}$ and normalized
fit functions to the old and new datasets, all divided by the distribution for
$\mathcal{R}=0$. In the plot for the ATLAS spin-0 analysis the dashed lines show the results
obtained using the old dataset with the new photon isolation criteria.}
\label{fig:ErrorBands}
\end{figure}
Comparing the new fit functions to the ones based on the previous small datasets presented at
Moriond 2016 we find a steeper functional fit in all three analyses, see
Fig.~\ref{fig:comparison}. While the new best fit parameters are within one standard deviation for
the two CMS fits, the ones for the ATLAS fit deviate by almost two standard deviations after
marginalizing over the normalization, see Fig.~\ref{fig:fit_comparison}. This might, however, be an
effect of the changed photon isolation as the fit to the $3.2\,\text{fb}^{-1}$ dataset with the updated photon
identification also deviates by more than one standard deviation from the previous best fit point. A
better understanding of the effect of the fake photons could be obtained by investigating the result
of changing the isolation criteria with the full $15.4\,\text{fb}^{-1}$ dataset. The tighter isolation
criteria are also reflected in the better agreement between the fitted distributions and the MCFM
generated digamma spectrum.\par
In order to show the changes in the fits, the ratio of the normalized fit functions for the ATLAS
spin-0 analysis is shown in the lower left plot of Fig.~\ref{fig:comparison}. A direct comparison of
the data and the un-normalized fit functions, even for the fits to the two different $3.2\,\text{fb}^{-1}$
sets, is difficult since the binning of data has changed. The large change in the fit parameters is
reflected in the deviation of more than 5\,\% for the comparison of the fits to the two $3.2\,\text{fb}^{-1}$
datasets and the even greater deviation when comparing with the fit function to the full
$15.4\,\text{fb}^{-1}$ dataset. While in the previous signal region near 750\,GeV the change is of the order
of 10\,\% it is greater than 30\,\% near 1.6\,TeV. This shows that the actual shape of the digamma
spectrum at high invariant masses is hard to predict precisely by an extrapolation and is therefore
very much subject to systematic uncertainties.\par
Finally in Fig.~\ref{fig:ErrorBands} the ratios of several normalized distributions and fit
functions to the normalized distribution with $\mathcal{R}=0$ are shown. These include the
distributions with the best fit value for $\mathcal{R}$ based on the Moriond 2016 dataset and the
ICHEP 2016 dataset and also the NLO distributions and the fit functions to the old and new
datasets. In the case of the ATLAS spin-0 analysis also the distribution and fit function to the
2015 dataset with the new photon isolation is shown. By comparing the curves we find that a sizable
systematic uncertainty can be inferred from the differences between the fit functions.
\section{Conclusions}
\label{sec:conclusions}
This paper deals with a problem that often arises in searches for new physics at the energy
frontier. In this context the challenge is to look for a new resonance at the upper end of a
distribution where only limited knowledge on the SM background is available. As a case study we
focus on the 750\,GeV anomaly where we examine in particular the implications of the possibility
that the excess in the 2015 data set is not only due to a (malicious) statistical fluctuation but
also a result of a physical effect. We discuss possible issues with the background: how much
photon-jet contamination is still allowed in the region of interest? How could it affect the
significance of the excess?
We study these questions using currently available theoretical tools for computing the photon-jet
mass distributions and apply them to the small set of publicly available data. However, this
approach is limited by our ability to thoroughly disentangle the effects of the
$(p_T,\eta)$-dependent jet fake rate and the theoretical uncertainty of the shape of the photon-jets
background. We therefore choose to model these combined effects by an $m_{\gamma j}$ dependent
reweighting of the invariant mass distribution, keeping the overall purity within the quoted ranges.
We first study a physics-driven reweighting procedure: we convolve a mass dependent K-factor with a
rapidity and transverse momentum dependent photon fake rate for the jets. The K-factor is extracted
comparing the NLO leading-jet-photon to the LO quark-photon spectrum, and the phase space dependence
of the fake rate is estimated from the experimental literature~\cite{ATLAS:2011kuc}. Both correction
factors are approximate, based on incomplete information, and should be taken with a grain of
salt. Motivated by this result, we then consider a more phenomenological deformation of the
$p p \to \gamma j$ spectrum. It allows us to study the sensitivity of the significances on a single
continuous quantity $\mathcal{R}$ (see Eq.~\eqref{eq:1}) which parametrizes an effective
deformation.
To summarise our results for the 750\,GeV case study based on the 2015 data we focus on the simpler
effective ansatz where we find the following:
\begin{itemize}
\item For the ATLAS spin-0 analysis, the significance of the excess can be reduced by
$\Delta \sigma \simeq1.1 $ when comparing the fitting function defined in \eqref{ATLASfun} with
our best-fit to the $\mathcal R$-modified distribution. A comparable reduction is found for the
ATLAS spin-2 measurement. Here however, it is less straightforward to determine the reduction,
since the estimation of the background shape in our ansatz differs from that of the ATLAS
analysis, which is not reproducible since the required data is not publicly available. Strictly
speaking, $w_{\gamma j}^{\textrm{MG}}$ does therefore not correspond to the ATLAS approach but is
the best approximation we can get. Since, however ATLAS claims to find comparable results with the
corresponding fitting function defined in \eqref{CMSfun}, we can reduce the significance with the
$\mathcal{R}$-modified distribution with respect to the fitting function by
$\Delta \sigma \simeq 1.0$ as well as with respect to the distribution with
$w_{\gamma j}^{\textrm{MG}}$ by $\Delta\sigma\simeq 1.5$.
\item The effect is smaller for the CMS 13\,TeV analyses with $\Delta \sigma \simeq 0.3-0.4$,
depending on the category.
\item In a ``combined fit'' to independent ATLAS and CMS datasets, the significance can be reduced
by as much as $\Delta \sigma \approx 1.0 \,(1.4)$ for the ATLAS spin-0 (spin-2) combined with CMS.
\end{itemize}
The larger preference for an enhanced photon-jet contribution in the spin-2 sample could point to
its higher sensitivity to the large rapidity region where jet fakes are more difficult to
reject. Finally, an F-test shows that the ATLAS data support using a more complex distribution.
To summarise our results for the 750\,GeV case study based on the 2016 data we find the following:
\begin{itemize}
\item For the ATLAS spin-0 analysis, we find that the new data prefers $\mathcal{R}$ in the range
$0.2\pm 0.2$, eliminating the remaining significance of $1.2\,\sigma$ in the full data set. As for
the spin-2 case no data is currently available.
\item The updated CMS analyses based on $12.9\,\text{fb}^{-1}$ even have a downwards fluctuation with respect
to the fit function near 750\,GeV leading to negative significances. Correspondingly the best fit
values for $\mathcal{R}$ are negative and ameliorate the situation.
\end{itemize}
We emphasize that our simplified ansatz for the effective modification of the photon-jet background
with $\mathcal{R}$ is not meant to necessarily represent a new background source nor the exact shape
of the background contamination in the signal region. Rather its envelope (corresponding to the
shaded area in Fig.~\ref{fig:NormalizedDists}) is expected to reflect a possible combination of
higher order QCD contributions, fragmentation, isolation and detector effects, which are outside of
theoretical control and, in the high invariant mass signal region, also beyond direct experimental
probes with currently available data. We further note that a quark flavor tagged dijet sample might
provide a high-statistics measurement of the relevant photon fake rates (see for
instance~\cite{Chekanov:2004kz}).
We have also employed our analysis to compare the difference between the fitting functions used by
ATLAS (with the new isolation criteria) given the 2015 and 2016 data sets. The fitting functions
where extrapolated to invariant masses beyond the TeV region. In summary we have found that:
\begin{itemize}
\item A variation of about 30\% in the extrapolated background near $m_{\gamma\gamma}=1.6\,\ensuremath{\,\text{TeV}}$ is
obtained.
\end{itemize}\par
To conclude, we have extensively examined the status of LHC diphoton searches. We have compared the
analyses performed on both 2015 and 2016 datasets in order to scrutinize the current state of the
art measurements for systematic effects. Using our approach we have reevaluated the current
experimental sensitivity to beyond standard model physics, especially in the tails of the diphoton
invariant mass distributions, beyond the TeV range. We found that the extrapolation of background
shapes is subject to sizeable uncertainties, potentially affecting the significance of possible
future excesses near the edge of the measured distributions. Furthermore, our analysis motivates
further Monte Carlo studies of the dominant diphoton backgrounds, based on jet flavor tagging
algorithms. Knowledge of whether a jet is of ``quark" or ``gluon" origin would improve our
estimation for the jet-photon faking backgrounds to next to leading order QCD accuracy. It is
important to note that diphoton-based searches at even larger invariant masses, that are highly
motivated, are being performed at present and will continue to be an integral part of the LHC
experimental physics program at the high energy frontier.
\section*{Acknowledgments}
We would like to thank Rikkert Frederix for useful discussions. J.F.K.~would like to thank CERN for
hospitality while this work was being completed and acknowledges the financial support from the
Slovenian Research Agency (research core funding No. P1-0035). The work of GP is supported by
grants from the BSF, ISF and ERC and the Weizmann-UK Making Connections Programme. AW is supported
by the DFG cluster of excellence ``Origin and Structure of the Universe'' and the European
Commission (AMVA4NewPhysics, 2020-MSCA-ITN-2015).
\bibliographystyle{JHEP}
|
\section{Introduction}
A univariate linear Hawkes process is a simple point process $N$ whose (stochastic) intensity $\lambda$ at time $t$ is given by
\begin{equation*}
\lambda({t}):= \mu + \int_{-\infty}^{t-}h(t-s)N(ds) = \mu + \sum_{\tau_i<t}h(t-\tau_i), \label{dynamics}
\end{equation*}
where
$\tau_i$ are the occurrences of the points before time $t$, and $h(\cdot):[0, \infty) \rightarrow [0, \infty)$.
See Section~\ref{sec:hawkes} for accurate definitions, multivariate extensions, and related properties.
We use the notation $N({t}):=N(0,t]$ to denote the number of
points in the interval $(0,t]$. When $h \equiv 0$, the Hawkes process $N$ becomes a Poisson process with rate $\mu$.
In the literature, the parameter $\mu$ is called the \textit{baseline intensity}, and $h(\cdot)$ is called
the \textit{exciting function} or sometimes referred to as the \textit{kernel function}.
The linear Hawkes process was first introduced by A.G. Hawkes in 1971 \cite{Hawkes, Hawkes71II}. It exhibits both self--exciting (i.e., the occurrence
of an event increases the probabilities of future
events) and clustering properties. Hence it is very appealing in point process modeling and it has wide applications
in various domains, including neuroscience \cite{Johnson96, Pernice2012, Reynaud2013}, seismology \cite{Ogata1988}, genome analysis \cite{Gusto2005, Reynaud2010}, social network \cite{Blundell2012, Crane2008}, finance (see the recent survey paper \cite{Bacry2015} and the references therein) and others.
This paper focuses on stationary Hawkes processes and their applications in specific queueing systems. A Hawkes process is stationary if its distribution does not change under time shift. See Section~\ref{sec:hawkes} for accurate definitions. In this paper, we develop approximations of a stationary Hawkes process with a large baseline intensity $\mu$. Mathematically, under a mild assumption on the exciting function (Assumption~1), we
establish a functional central limit theorem (FCLT) for a sequence of univariate stationary Hawkes processes $N^{\mu}$ indexed by the baseline intensity $\mu$ which goes to infinity (see Theorem~\ref{thm:FCLT}). These Hawkes processes share a common fixed exciting function $h$. The limit process turns out to be a Gaussian process which is non--Markovian unless $h \equiv 0$. This limiting Gaussian process has stationary but dependent increments.
To illustrate the strength of the Gaussian approximation for the Hawkes processes, we study a specific queueing model with a stationary Hawkes traffic: an infinite--server queue with general service time distributions. The Hawkes process could be a potential traffic model especially for financial market data feeds\footnote{Market data feeds are typically composed of event messages that provide, in real time, the status of the market such as asset prices, reports of completed trades, and order activities. While some industry white paper \cite{Low} suggests that the market data traffic clearly exhibits clustering, we are not aware of academic studies or publicly available data on market data feeds.} for several reasons. First,
the Hawkes process
naturally extends the classical Poisson process. Second, stock order flows and the occurrence of financial market events are known to exhibit clustering (in time) and self--exciting features, e.g., trades trigger other trades \cite{Bacry2015, Bowsher2007, Cont2012, Fonseca2014, Hewlett2006}. The standard Poisson process
can not capture these features while the Hawkes process can adequately model such clustering and self--exciting behavior.
Third, the Hawkes process is a highly versatile and flexible model which can exhibit a broad range of correlation structure, depending on the specification of how past events affect the occurrence of current and future events (\cite{Bacry2015}). Finally, the Hawkes process is amenable to statistical inference (see, e.g., \cite{Bacry2015, Daley, Ozaki1979}).
Infinite--server queues are interesting in their own right since they naturally arise in the study of many applications such as electric power consumption and insurance mathematics \cite{Blanchet2014, Glynn2002}. In addition, as argued in \cite{whittIS2012}, infinite--server queues often serve as useful approximations
for multi--server queues which are classical models for large--scale service systems (e.g., server farms, call centers). In the financial context, such an infinite--server queue can serve as an approximate model for describing the market data feed sent from exchanges, processed by many parallel computer servers, and then delivered to consuming applications of end--users.
Since Hawkes processes are non--Markovian in general \red{and the inter-arrival times are correlated, it is challenging to analyze the performance of an infinite-server queue with Hawkes traffic and general service time distributions, either analytically or numerically.} Hence, we consider the regime that the baseline intensity $\mu$ of the Hawkes input is large. Such a regime could be relevant since the market data traffic is generated by many market participants and the market
data volumes are huge in practice (e.g., in the range of gigabytes per second). Relying on \cite{Kri1997}, we develop heavy--traffic approximations for the performance of such an infinite--server queue fed by a univariate stationary Hawkes process with a large baseline intensity $\mu$ (Proposition~\ref{thm:IS1}). The limiting queue length process is a Gaussian process. We compute its covariance function as well as its steady--state distribution explicitly, both of which depend on the distribution of service times as well as the detailed form of the covariance density of the Hawkes traffic (Proposition~\ref{prop:covX} and Corollary~\ref{prop:X-infty}). In the special case of exponential service time distributions, the limiting queue length process is an Ornstein-Uhlenbeck (OU) process driven by a Gaussian process. This Gaussian--driven OU process is non--Markovian in general. We illustrate through examples and numerical experiments that the Gaussian approximation for the steady--state queue length is effective.
We also extend our functional central limit theorem to multivariate stationary Hawkes processes (Theorem~\ref{thm:multi-FCLT}) and study infinite--server queues with multivariate Hawkes traffic and general service time distributions. Such a model can be viewed as a multi--class queueing model with correlated and mutually--exciting arrivals. We show that the limiting queue length process is a multivariate Gaussian process (Proposition~\ref{thm:IS-multi-gen}).
When the service times of each class of customers are independent exponentials, this limiting queue length process becomes a multi--dimensional Gaussian--driven OU process (Proposition~\ref{thm:IS-mutual}).
To summarize, our paper is the
first one that studies the large baseline intensity asymptotics for stationary Hawkes processes. Unlike the existing limit theorems for Hawkes process in the
literature, our proof relies on the immigration-birth representation of the
linear Hawkes processes \cite{HawkesII}, and the delicate analysis of the moments of the stationary Hawkes process. Our paper is also the first to study queues with stationary Hawkes traffic. We obtain new explicit results for the performance of infinite-server queueing systems which allows us to better understand the impact of self--exciting and mutually--exciting Hawkes traffic on the system performance.
\bigskip
\textbf{Related Literature.}
Two streams of research that are closely related to our work are Hawkes processes and infinite-server queues.
We now explain the difference between our study and the existing literature in these two areas.
\textit{Asymptotics of Hawkes processes.}
Note that most of the existing literature on limit theorems
for Hawkes processes are for {\it{large--time asymptotics}}, where one scales both time and space. See \cite{Bacry, Bordenave, ZhuMDP} for large--time asymptotics of linear Hawkes processes, \cite{Karabash, ZhuCIR} for large--time asymptotics for extensions of linear Hawkes processes,
\cite{Jaisson, JaissonII} for the nearly unstable case where $\Vert h\Vert_{L^{1}}\approx 1$, \cite{Zhang2015} for the generalized {Markovian} Hawkes processes (or affine point processes), and \cite{ZhuThesis} for large--time asymptotics of nonlinear Hawkes processes.
These large--time asymptotics are different from our large-$\mu$ asymptotics (no time-scaling is involved).
We will see later, see e.g. Theorem \ref{thm:FCLT},
that the time-space and intensity-space scalings are not equivalent.
For Poisson processes, these two scalings are equivalent and both lead to a Brownian limit.
For Hawkes processes, for the time-space scaling, we obtain the Brownian limit, see e.g. \cite{ZhuCLT}.
On the other hand, if we consider large baseline intensity $\mu$ and scale down the space, we get a non-Markovian Gaussian limit (Theorem \ref{thm:FCLT}). The primary reason is that the Hawkes process $N^{\mu}$ with a baseline intensity $\mu$, say $\mu$ is a positive integer,
can be expressed as partial sums of i.i.d. copies of a Hawkes process $N^{1}$ which has baseline intensity one (see Sections~2 and 3). Thus for the intensity-space scaling we consider,
the covariance structure of $N^1$ is still preserved in the limit, and
the covariance structure of $N^1$ does not coincide with that of a Brownian motion
since $N^1$ has dependent time increments, which leads to the non-Brownian Gaussian limit.
Other than the large--time asymptotics, limit theorems for non-stationary Markovian Hawkes processes with a large initial intensity have been established in our recent studies \cite{GZ, GZ2}.
Large--dimension asymptotics have been studied in \cite{Delattre,Chevallier,DF}, in which the authors studied the asymptotics
for the multivariate Hawkes process and its extensions
where the number of dimension goes to infinity, and obtained a mean--field limit.
\textit{Infinite-server queues.}
In the setting of infinite--server queues, our work complements the stream of research on heavy--traffic approximations of such queues, see, e.g., \cite{Eick93, Iglehart1965, Jamol2016, Pang2015, Pang2007, Pang2010, Reed2015, whitt2002} and the references therein. In these studies, the heavy--traffic limit of the arrival process is typically a Brownian motion or a deterministic time--changed Brownian motion. With Hawkes traffic, we obtain a non--Markovian limit but the Gaussian structure still allows us to obtain elegant formulas for transient and steady--state performance measures. From the traffic modeling perspective, we also mention that certain Poisson cluster processes have been used to model tele--traffic arrivals (see, e.g., \cite{Fasen2010, Fay2006, Hohn2003}).
The linear Hawkes processes which can be seen as Poisson cluster processes (see e.g. \cite{Bacry2015, Daley}) are not covered by these studies.
\bigskip
\textbf{Organization of this paper.}
The rest of the paper is organized as follows. In Section~\ref{sec:hawkes},
we formally introduce stationary linear Hawkes processes and review some of their properties.
In Section~\ref{sec:2}, we state the main result on the functional central limit theorem for univariate stationary Hawkes processes with large baseline intensity $\mu$ and describe the properties of the limiting Gaussian process.
In Section~\ref{sec:inf-server}, we develop heavy--traffic approximations for infinite--server queues with univariate Hawkes traffic. We also discuss in detail the special case when service times are exponentially distributed. In Section~\ref{sec:multi-hawkes}, we extend our results to multivariate stationary Hawkes processes and study infinite--server queues with multivariate Hawkes traffic.
The proofs of all the results are collected in the Appendix.
\section{Introduction to Stationary Hawkes processes} \label{sec:hawkes}
In this section, we formally introduce stationary linear Hawkes processes and review some of their properties.
\subsection{Definition and stationarity condition}
Let $N$ be a simple point process on $\mathbb{R}$, that is,
a family $\{N(C)\}_{C\in\mathcal{B}(\mathbb{R})}$ of random variables
with values in $\{0, 1, 2, \ldots, \}\cup\{\infty\}$ indexed
by the Borel $\sigma$-algebra $\mathcal{B}(\mathbb{R})$ of the real line $\mathbb{R}$,
where $N(C)=\sum_{n\in\mathbb{Z}}1_{C}(T_{n})$ and $(T_{n})_{n\in\mathbb{Z}}$
is a sequence of extended real-valued random variables so that almost surely $T_{0}\leq 0<T_{1}$, $T_{n}<T_{n+1}$
on $\{T_{n}<\infty\}\cap\{T_{n+1}>-\infty\}$ for every $n\in\mathbb{Z}$.
Let $\mathcal{F}_{t}=\sigma(N(C),C\in\mathcal{B}(\mathbb{R}),C\subset(-\infty,t]))$.
The process $\lambda(t)$ is called the $\mathcal{F}_{t}$-intensity of $N$ if for
all intervals $(a,b]$, we have
\begin{equation}\label{martingality}
\mathbb{E}[N((a,b])|\mathcal{F}_{a}]
=\mathbb{E}\left[\int_{a}^{b}\lambda(s)ds\Big|\mathcal{F}_{a}\right],
\qquad
\text{a.s.}
\end{equation}
The univariate linear Hawkes process with baseline intensity $\mu>0$ and exciting function $h:\mathbb{R}_{+}\rightarrow\mathbb{R}_{+}$
is a simple point process $N$ admitting the $\mathcal{F}_{t}$-intensity
\begin{equation}\label{intensity}
\lambda(t)=\mu+\int_{-\infty}^{t-}h(t-s)N(ds).
\end{equation}
Due to \eqref{intensity}, the univariate Hawkes process is sometimes also called the self--exciting point process in the literature.
A commonly used nontrivial example of the exciting function $h$ is an exponential function, i.e.,
$h(t)= \alpha e^{-\beta t}$ for $t \ge 0$, where $\alpha, \beta >0$. In this special case, the process $(\lambda(t), N(t))$
is Markovian, and the intensity process $\lambda(t)$ itself is also Markovian, see e.g. \cite{Errais}.
The power law function $h(t)=\frac{1}{(1+\delta t)^{\gamma}}$, where $\delta,\gamma>0$ is also
a popular choice for the exciting function in the literature, see e.g. \cite{Bacry2015}.
The multivariate Hawkes process extends the univariate Hawkes process to $k \ge 1$ dimensions as follows.
Let $\mathbb{N}:=(\mathbb{N}^{1} ,\ldots, \mathbb{N}^{k})$,
where $\mathbb{N}^{i}$ are simple point processes on $\mathbb{R}$ with no common points, and
for each $1\leq i\leq k$, $\mathbb{N}^{i}$ has the intensity:
\begin{equation}\label{TakeExpectation}
\lambda^{i}(t)=\mu_{i}+\sum_{j=1}^{k}\int_{-\infty}^{t-}h_{ij}(t-s) \mathbb{N}^{j}(ds),
\end{equation}
where $\mu_{i}>0$ and $h_{ij}(\cdot):\mathbb{R}_{+}\rightarrow\mathbb{R}_{+}$ for $1\leq i, j\leq k$. Due to \eqref{TakeExpectation}, the multivariate Hawkes process is sometimes also called the mutually--exciting point process in the literature.
To facilitate the presentation, we summarize below the key properties of the linear stationary Hawkes processes that will be used in the paper. Write $\Vert f\Vert_{L^{1}} = \int_{0}^{\infty} f(t)dt$
for a function $f:[0,\infty)\rightarrow[0,\infty)$.
\begin{enumerate}
\item [(a)]
(Stationarity).
For a simple point process $N$, stationarity of $N$ means its distribution does not change under time shift. More precisely, $N$ is stationary if
the process $\theta_{t}N$ has the same distribution as the process $N$ for any $t$, where
$\theta_{t}$ is a shift operator defined as $\theta_{t}N(C)=N(t+C)$ for every $C\in\mathcal{B}(\mathbb{R})$.
This directly implies that a stationary Hawkes process $N$ has stationary increments, and
the intensity process $\lambda(\cdot)$ is a stationary process where the distribution of $\lambda(t)$ does not depend on $t.$
Similarly, we say a multivariate point process $\mathbb{N}=(\mathbb{N}^{1} ,\ldots, \mathbb{N}^{k})$ is stationary, if
$(\theta_{t} \mathbb{N}^{1},\ldots,\theta_{t} \mathbb{N}^{k})$ has the same distribution as $(\mathbb{N}^{1} ,\ldots, \mathbb{N}^{k})$ for any $t$.
Under the assumption $\Vert h\Vert_{L^{1}}<1$, there is a unique stationary
version of the Hawkes process $N$ with the intensity \eqref{intensity}, see e.g. \cite{Bremaud}.
More generally, under the assumption that the spectral radius of the $k \times k$ matrix
$\mathbb{H}:= (\Vert h_{ij}\Vert_{L^{1}})_{1\leq i,j\leq k}$ is strictly less than $1$,
there is a unique stationary version of the multivariate Hawkes process $\mathbb{N}$
with the intensity \eqref{TakeExpectation}, see e.g. \cite{Bremaud}.
\item [(b)]
(Martingality). By the definition of the intensity in \eqref{martingality}, we have
for any simple point process $N$ with the intensity $\lambda$,
$N(t)-\int_{0}^{t}\lambda(s)ds$ is a martingale.
Moreover, its predictable quadratic variation is given by $\int_{0}^{t}\lambda(s)ds$
so that $\left(N(t)-\int_{0}^{t}\lambda(s)ds\right)^{2}-\int_{0}^{t}\lambda(s)ds$ is also
a martingale. We will apply this martingale property to univariate stationary Hawkes processes and the marginal processes of multivariate stationary Hawkes processes in the proofs of Theorem~\ref{thm:FCLT} and \ref{thm:multi-FCLT}.
\item [(c)]
(First-order mean). For stationary $k-$variate Hawkes processes,
by taking expectations on both hand sides of \eqref{TakeExpectation}
and by the martingale property \eqref{martingality}, we have for each $t,$
\begin{equation*}
\bar{\lambda}_{i}:=\mathbb{E}[\lambda^{i}(t)]
=\mu_{i}+\sum_{j=1}^{k}\int_{-\infty}^{t-}h_{ij}(t-s)\bar{\lambda}_{j}ds,
\end{equation*}
which implies that
\begin{equation}\label{BarLambdaDefn}
\bar{\lambda}=(\mathbb{I}-\mathbb{H})^{-1}\mu,
\end{equation}
where $\bar{\lambda}=(\bar{\lambda}_{i})_{1\leq i\leq k}$, $\mu=(\mu_{i})_{1\leq i\leq k}$ and $\mathbb{I}$ is the identity matrix.
\item [(d)]
(Covariance density and variance function). For a stationary $k-$variate Hawkes process $(\mathbb{N}^{1} ,\ldots, \mathbb{N}^{k})$,
the covariance density matrix $\Phi(\tau)=(\Phi_{ij}(\tau))_{1\leq i,j\leq k}$, where $\Phi_{ij}(\tau) := \mathbb{E} [d\mathbb{N}^i(t+\tau) d\mathbb{N}^j(t)]/(dt)^2 - \bar \lambda_i \bar \lambda_j$ which does not depend on $t$, is given as follows,
see e.g. \cite{Hawkes,Hawkes71II}.
For $\tau\geq 0$,
\begin{equation} \label{eq:cov-density}
\Phi(\tau)=h(\tau)\text{diag}(\bar{\lambda})+\int_{-\infty}^{\tau}h(\tau-v)\Phi(v)dv,
\end{equation}
and $\Phi_{ij}(-\tau)=\Phi_{ji}(\tau)$ for every $\tau>0$ and $1\leq i,j\leq k$,
where $\bar{\lambda}$ is defined in \eqref{BarLambdaDefn}.
Here $\text{diag}(\bar{\lambda})$ is the diagonal matrix with entries $\bar{\lambda}_{i}$'s on the diagonal,
and with slight abuse of notations, $h(t)=(h_{ij}(t))_{1\leq i,j\leq k}$.
The variance function for the stationary $k-$variate Hawkes process, $\mathbb{K}(t)=(K_{ij}(t))_{1\leq i,j\leq k}
:=\text{Var}(\mathbb{N}(t))=(\text{Cov}(\mathbb{N}^{i}(t),\mathbb{N}^{j}(t)))_{1\leq i,j\leq k}$ is given by
\begin{equation}\label{eq:var-function}
\mathbb{K}(t):=\text{diag}(\bar{\lambda})t+2\int_{0}^{t}\int_{0}^{t_{2}}\Phi(t_{2}-t_{1})dt_{1}dt_{2}.
\end{equation}
\item [(e)]
(Association).
Intuitively, since the Hawkes process has the self- and mutually-exciting properties,
there are positive correlations between counts across time intervals. To make this statement rigorous,
we will use the notion of \emph{association} from probability theory, see e.g. \cite{Evans}.
Let $\mathcal{X}$ and $\mathcal{Y}$ be complete and separable metric spaces,
with closed orders $\leq_{\mathcal{X}}$ and $\leq_{\mathcal{Y}}$. A map $f:\mathcal{X}\rightarrow\mathcal{Y}$
is non-decreasing if $x_{1}\leq_{\mathcal{X}}x_{2}$ implies $f(x_{1})\leq_{\mathcal{Y}}f(x_{2})$.
An $\mathcal{X}$-valued random variable $X$ is \emph{associated} if for each
pair of bounded, Borel measurable, non-decreasing functions $f,g:\mathcal{X}\rightarrow\mathbb{R}$,
we have $\text{Cov}(f(X),g(X))\geq 0$.
Let $\mathcal{S}$ be a locally compact, separable, metric space
and denote by $M(\mathcal{S})$ the space of Radon measures on $\mathcal{S}$
equipped with the vague topology with a partial ordering which is closed
by declaring that $\mu\leq\nu$ if $\mu(B)\leq\nu(B)$ for all Borel sets $B$.
A random measure is an $M(\mathcal{S})$-valued random variable.
The linear $k-$variate Hawkes process $\mathbb{N}=(\mathbb{N}^{1},\ldots,\mathbb{N}^{k})$
is equivalent to a marked linear Hawkes process $\mathbb{N}^{\dagger}$,
which is a random measure defined
on the space $\mathcal{S}=\mathbb{R}\times\{1,2,\ldots,k\}$,
via $\mathbb{N}^{\dagger}(C,i)=\mathbb{N}^{i}(C)$, for any Borel sets $C$ of $\mathbb{R}$
and $i\in\{1,2,\ldots,k\}$. The random measure $\mathbb{N}^{\dagger}$
is infinitely divisible since the $k-$variate linear Hawkes process $\mathbb{N}$ is a special case of the Poisson cluster process (see e.g. \cite{Daley, JHR}), which is infinitely divisible.
Theorem 1.1. in \cite{Evans} (which first appears in \cite{BW})
says any infinitely divisible random measure on $\mathcal{S}$ is associated.
The association property of Hawkes processes implies that
the covariance density \eqref{eq:cov-density} is non-negative and it will also
be used to show the finiteness of the moment generating function of
the stationary Hawkes process in the proof of Theorem~\ref{thm:multi-FCLT}.
\end{enumerate}
Throughout the paper, we will always assume that we are working with the stationary
version of a Hawkes process. More specifically, we will
make the following assumption on the exciting function of $k-$dimensional Hawkes processes which guarantees the existence of the stationary version. This assumption is satisfied in most applications of Hawkes processes, see e.g. \cite{Bacry2015, Hawkes, ZhuThesis} and the references therein.
\begin{assumption}\label{assump1}
For all $1\leq i,j\leq k$, the exciting function $h_{ij}$ is non-negative, locally bounded, and Riemann integrable.
In addition, the spectral radius of the $k \times k$ matrix
$\mathbb{H}:= (\Vert h_{ij}\Vert_{L^{1}})_{1\leq i,j\leq k}$ is strictly less than $1$.
\end{assumption}
\subsection{Immigration--birth representation}\label{ImmigrationSection}
In this section, we review the well--known immigration birth representation of linear Hawkes processes (see, e.g., \cite{HawkesII, JHR}) which is the key to the proof of our results.
For the univariate stationary Hawkes process with intensity dynamics \eqref{intensity},
we assume that immigrants arrive according to a homogeneous
Poisson process with constant rate $\mu$ on the real line $\mathbb{R}$.
Each immigrant would produce children and the number
of children has a Poisson distribution with mean $\Vert h\Vert_{L^{1}}$. Conditional on the number of
the children of an immigrant, the children are born independently, and each child is born
at a time with a probability density function $\frac{h(t)}{\Vert h\Vert_{L^{1}}}$.
In other words, children are born according to an inhomogeneous Poisson process
with intensity $h(\cdot)$. Each child would produce children
according to the same laws independent of other children. All the immigrants produce children independently.
The number of points of a linear Hawkes process on a time interval $(0,t]$ equals
the total number of immigrants and the descendants on the interval $(0,t]$.
Note that the immigration--birth representation holds similarly for the multivariate Hawkes process, see e.g. \cite{JHR}.
For a $k$--variate Hawkes process $(\mathbb{N}^{1},\ldots,\mathbb{N}^{k})$
with the intensity \eqref{TakeExpectation} for $\mathbb{N}^{i}$, where $1\leq i\leq k$,
we consider immigrants of $k$ types, and the type-$i$ immigrants arrive
according to a homogeneous Poisson process with intensity $\mu_{i}$,
and each type-$i$ immigrant produce children of type $j$ according to
an inhomogeneous Poisson process with intensity $h_{ji}(\cdot)$.
Each child of type $i$ would produce children of different types according
to the same laws independent of other children.
All the immigrants produce children independently.
The number of points $\mathbb{N}^{i}$ on a time interval $(0,t]$
equals to the total number of immigrants and the descendants of type $i$ on the interval $(0,t]$.
Also note that the immigration--birth representation does not require the stationarity of the Hawkes process,
or the monotonicity of the exciting function, see e.g. \cite{ZhuMDP}.
\section{FCLT for univariate stationary Hawkes processes} \label{sec:2}
In this section we develop approximations for a \textit{univariate} stationary Hawkes process with a large baseline intensity $\mu$.
Consider a univariate stationary Hawkes process $N^{\mu}$ with stochastic intensity in \eqref{intensity}. We write $N^{\mu}$
to emphasize that the baseline intensity of this Hawkes process is $\mu$. Our goal is to
establish a functional central limit theorem for a sequence of stationary Hawkes processes $N^{\mu}$ in the asymptotic regime $\mu \rightarrow \infty$.
Note that the exciting function is fixed, i.e., this sequence of Hawkes processes shares a common exciting function $h$ with $\Vert h\Vert_{L^{1}} <1$.
To facilitate the presentation, let us define
\begin{equation} \label{eq:var}
K(t):=\frac{t}{1-\Vert h\Vert_{L^{1}}}
+2\int_0^t \int_0^{t_2}\phi(t_{2}-t_{1})dt_{1}dt_{2},
\end{equation}
where $\phi:[0, \infty)\rightarrow [0, \infty)$ satisfies the integral equation:
\begin{equation} \label{eq:phi}
\phi(t)=\frac{h(t)}{1-\Vert h\Vert_{L^{1}}}+\int_{0}^{\infty}h(t+v)\phi(v)dv
+\int_{0}^{t}h(t-v)\phi(v)dv,
\end{equation}
and $\phi(-t) = \phi(t)$ for $t>0$.
The function $\phi$ and $K$ are just the covariance density and variance functions
for the univariate stationary Hawkes process with baseline intensity $1$, respectively. See Equations~\eqref{eq:cov-density} and \eqref{eq:var-function}.
Note that the covariance density $\phi$ is non-negative
since the linear Hawkes process is associated.
When $h\equiv 0$, the linear Hawkes process reduces to the Poisson process with independent increments
and thus $\phi\equiv 0$. On the other hand, when $\phi\equiv 0$, from \eqref{eq:phi},
it is clear that $h\equiv 0$. Hence, $\phi\equiv 0$ if and only if $h\equiv 0$.
We now present a result on the functional central limit theorem for such univariate stationary Hawkes processes. Write $(D([0,\infty),\mathbb{R}),J_{1})$ as the space of c\`{a}dl\`{a}g processes on $[0, \infty)$ that are equipped with Skorohod $J_1$ topology (see, e.g., Billingsley \cite{Billingsley}), and write $``\Rightarrow"$ for convergence in distribution. Recall from \eqref{BarLambdaDefn} that $\bar \lambda = \frac{\mu}{1 - \Vert h\Vert_{L^{1}}}.$
\begin{theorem} \label{thm:FCLT}
Under Assumption~\ref{assump1}, we have as $\mu\rightarrow\infty$,
\begin{equation*}
\frac{N^{\mu}(t)- \bar \lambda t}{\sqrt{\mu}}\Rightarrow G(t),
\end{equation*}
in $(D([0,\infty),\mathbb{R}),J_{1})$, where $G$ is a mean-zero almost surely continuous Gaussian process
with the covariance function, for $t \ge s$,
\begin{equation}\label{eq:cov-G}
\mbox{Cov}(G(t),G(s))=\int_{s}^{t}\int_{0}^{s}\phi(u-v)dvdu+K(s).
\end{equation}
\end{theorem}
The proof of this result is given in Appendix~\ref{sec:proof1}.
We now briefly explain the intuition behind this result.
Without loss of generality, we assume $\mu$ takes integer values. By the immigration--birth representation of Hawkes processes, one can deduce that for a stationary univariate Hawkes process $N^{\mu}$ with a baseline intensity $\mu$ and an exciting function $h$, we can decompose it as the sum of $\mu$ i.i.d stationary Hawkes processes, each having a baseline intensity one and an exciting function $h$. Then one expecte by central limit theorem type of arguments that $N^{\mu}$ will be asymptotically Gaussian when we send $\mu$ to infinity.
We next discuss the covariance function of $G$ in \eqref{eq:cov-G}.
In general, the covariance function of $G$ in \eqref{eq:cov-G} is semi-explicit and we can compute it by first numerically solving $\phi$ via the integral equation \eqref{eq:phi}.
In the special case when $h(t)= \alpha e^{-\beta t}$ where $\alpha< \beta$, the covariance function of $G$ is explicit. To see this, we first deduce from \eqref{eq:phi} that
\begin{equation*}
\phi(t)=\frac{\alpha e^{-\beta t}}{1- \frac{\alpha}{\beta}}+ \alpha e^{-\beta t} \cdot \int_{0}^{\infty}e^{-\beta v}\phi(v)dv
+\alpha e^{-\beta t} \cdot \int_{0}^{t} e^{\beta v}\phi(v)dv,
\end{equation*}
which yields that
\begin{equation}\label{eq:phi-exp}
\phi(t)=\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}e^{-(\beta-\alpha)t},\qquad t\ge 0.
\end{equation}
Plugging this into \eqref{eq:var}, we find that
\begin{align}\label{eq:Kt-exp}
\mbox{Var}(G(t)) = K(t)&=\frac{t}{1-\frac{\alpha}{\beta}}
+2\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}\int_{0}^{t}\int_{0}^{t_{2}}e^{-(\beta-\alpha)(t_{2}-t_{1})}dt_{1}dt_{2}
\nonumber\\
&=\frac{\beta^3 }{(\beta-\alpha)^3} t -\frac{\alpha\beta(2\beta-\alpha)}{(\beta-\alpha)^{4}}\left[1-e^{-(\beta-\alpha)t}\right],
\end{align}
and for $t \ge s$,
\begin{eqnarray*}\label{eq:cov-G-exponential}
\mbox{Cov}(G(t),G(s))&=&\int_{s}^{t}\int_{0}^{s}\phi(u-v)dvdu+K(s) \nonumber \\
&=& \frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{4}} \left(e^{(\alpha -\beta)s} - e^{(\alpha -\beta)t} \right) \cdot \left( e^{(\beta-\alpha)s} -1\right) +K(s)\nonumber \\
&=& \frac{\beta^3 }{(\beta-\alpha)^3} s + \frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{4}} \left( -1 - e^{(\alpha -\beta)(t-s)} + e^{(\alpha -\beta)t} + e^{(\alpha -\beta)s} \right). \nonumber \\
\end{eqnarray*}
In this special case, we notice that $K(\cdot)$, the variance function of $G$, is nonlinear in $t$ in general. This is very different from the case when $N^{\mu}$ is a Poisson process (i.e., $h \equiv 0$) where $G$ becomes a standard Brownian motion. In addition, we find from \eqref{eq:Kt-exp} that when $h$ is a single exponential function, the variance function
$K(\cdot)$ have the following properties: $K(\cdot)$ is Lipschitz continuous, convex, and asymptotically linear as $t \rightarrow \infty$.
For a general exciting function $h,$ we next summarize important properties of $K(t)=\mbox{Var}(G(t))$ defined in \eqref{eq:var} and $\phi(t)$ defined in \eqref{eq:phi} in the following result. These properties provide us a better understanding about the variance of the limit Gaussian process $G$.
\begin{proposition}\label{prop:Kt}
Under Assumption~1, the following hold:
\begin{itemize}
\item [(a)]
\begin{equation*}
\lim_{t\rightarrow\infty}\frac{K(t)}{t}=\frac{1}{(1-\Vert h\Vert_{L^{1}})^{3}}.
\end{equation*}
\item [(b)]
$\Vert\phi\Vert_{L^{1}}<\infty$, and the variance function $K(\cdot)$ is convex and Lipschitz continuous on $[0,\infty)$.
\item [(c)] If in addition $\int_{0}^{\infty}t^{2}h(t)dt<\infty$, then
\begin{align*}
&\lim_{t\rightarrow\infty}\left[K(t)-\frac{t}{(1-\Vert h\Vert_{L^{1}})^{3}}\right]
\\
&=\frac{1}{\pi(1-\Vert h\Vert_{L^{1}})^{3}}\int_{\mathbb{R}}\frac{1}{\omega^{2}}
\frac{(1-\Vert h\Vert_{L^{1}})^{2}-|1-\hat{h}(\omega)|^{2}}
{|1-\hat{h}(\omega)|^{2}}d\omega <0,
\nonumber
\end{align*}
where $\hat{h}$ is given by
\begin{eqnarray*}
\hat{h}(\omega) = \int_{0}^{\infty} e^{i\omega t} h(t) dt.
\end{eqnarray*}
\end{itemize}
\end{proposition}
The proof of this result is given in Appendix~\ref{sec:prop-Kt}.
Part (a) of this result is known in the literature, and we include it here mainly for completeness. The results in other parts appear to be new.
Having characterized the covariance and variance functions of $G$, we can now elaborate further properties of the Gaussian process $G$. We summarize them in the following result. The proof is given in Appendix~\ref{sec:G-prop}.
\begin{proposition}\label{prop:G-property}
Under Assumption~1,
the Gaussian process $G$ in Theorem~\ref{thm:FCLT} has stationary increments. In addition, the Gaussian process $G$ is not Markovian unless $h\equiv 0$. Furthermore,
the paths of $G$ are H\"{o}lder continuous of order $\gamma$ for every $\gamma<\frac{1}{2}$.
\end{proposition}
\begin{remark}
The increments of the Gaussian process $G$ are positively correlated and dependent in general. This is clear from \eqref{eq:cov-G} since for $s, \tau>0$
\begin{equation*}
\mbox{Cov}(G(s+\tau)- G(s),G(s))=\int_{s}^{s+\tau}\int_{0}^{s}\phi(u-v)dvdu,
\end{equation*}
which is nonzero and positive.
\end{remark}
\section{Infinite--server queues with self-exciting traffic} \label{sec:inf-server}
In this section we study infinite--server queues with high-volume self-exciting traffic, i.e., the arrival process is modeled by a univariate stationary Hawkes process. We establish limit theorems for such queues in Section~\ref{sec:limit}, characterize the limit process in Section~\ref{sec:property}, and discuss in detail the special case of exponential service time distributions in Section~\ref{sec:exp}.
\subsection{Limit theorems for $GI/\infty$ queues with self--exciting Hawkes traffic} \label{sec:limit}
In this section, we follow \cite{Kri1997} to establish the limit theorems for $GI/\infty$ queues with self--exciting Hawkes traffic.
We consider a sequence of infinite-server queueing models indexed by $\mu$ and
let $\mu \rightarrow \infty$. For each fixed $\mu$, the customers arrive to the $\mu-$th system according to a stationary univariate Hawkes process $N^{\mu}$ with a baseline intensity $\mu$ and an exciting function $h$. Hence, the average arrival rate is $\bar \lambda =\frac{\mu}{1-\Vert h\Vert_{L^{1}}}$. Write $Q^{\mu}(t)$ as the number of customers in the $\mu-$th system at time $t$.
We assume given an i.i.d. sequence of nonnegative random variables $\{\bar \eta_i: i \ge 1\}$ with a cumulative distribution function $F_0(x) = \mathbb{P}(\bar \eta_1 \le x)$ and another i.i.d. sequence of nonnegative random variables $\{\eta_i: i \ge 1\}$ with a cumulative distribution function $F(x) = \mathbb{P}(\eta_1 \le x)$. Assume $F_0(0)=F(0)=0$ for simplicity. The customers initially present in the infinite--server queueing system have remaining service times $\bar \eta_1, \ldots, \bar \eta_{Q^{\mu}(0)}$; the new arriving customers have service times $\eta_1, \eta_2, \ldots.$
All these service times, $Q^{\mu}(0)$, and the arrival process $N^{\mu}$ are assumed to be mutually independent. Then we have (see, e.g., \cite{Kri1997, Pang2010})
\begin{eqnarray*} \label{eq:Q-mu}
Q^{\mu}(t) = \sum_{i=1}^{Q^{\mu}(0)} 1_{\bar \eta_i >t} + \sum_{i=1}^{N^{\mu}(t)} 1_{\tau_i + \eta_i >t},
\end{eqnarray*}
where $\tau_i$ is the arrival time of the $i$-th new customer.
It follows from Theorem~3 in \cite{Kri1997} and our Theorem~\ref{thm:FCLT} that the following result holds. The proof is omitted.
\begin{proposition} \label{thm:IS1}
Suppose Assumption~1 holds.
Assume that for some constant $q_0$ and random variable $\xi,$
\begin{eqnarray} \label{eq:initial}
\sqrt{\mu} \left( \frac{Q^{\mu}(0)}{\mu} - q_0 \right) \Rightarrow \xi, \quad \text{as $\mu \rightarrow \infty.$}
\end{eqnarray}
Then the sequence of processes
$X^{\mu}$ defined by
\begin{eqnarray}
X^{\mu}(t) &=&{\sqrt{\mu}} \left( \frac{Q^{\mu}(t)}{\mu} - q_0 (1-F_0(t)) - \frac{1}{1-\Vert h\Vert_{L^{1}}} \cdot \int_{0}^t (1-F(t-u)) du \right), \label{eq:X-Q-1d}
\end{eqnarray}
as $\mu \rightarrow \infty,$
converges in distribution in $(D([0,\infty),\mathbb{R}),J_{1})$ to the process $X$ where
\begin{equation} \label{eq:X}
X(t) =(1- F_0(t)) \xi + \sqrt{q_0} \cdot W^0 (F_0(t)) + \theta(t) + \int_{0}^t (1- F(t-u)) dG(u) .
\end{equation}
Here, $W^0=\{W^0(x): x \in [0,1]\}$ is a Brownian bridge,
$G$ is the mean-zero Gaussian process given in Theorem~\ref{thm:FCLT}, $\theta$ is a mean-zero Gaussian process with covariance function given by
\begin{equation} \label{eq:cov-theta}
\mathbb{E}[ \theta(s) \theta(t)] = \frac{1}{1-\Vert h\Vert_{L^{1}}} \cdot \int_{0}^s F(s-u) (1-F(t-u)) du, \quad \text{$0 \le s \le t$.}
\end{equation}
The random elements
$\xi, W^0, G, \theta$ are mutually independent.
\end{proposition}
\begin{remark}
The integral $ \int_{0}^t (1- F(t-u)) dG(u) $ in \eqref{eq:X} is defined in a pathwise sense and is understood as the result of integration by parts. See Theorem~3 in \cite{Kri1997}.
In addition, it is known in the literature (see, e.g., \cite{Kri1997, Pang2010}) that one can represent the Gaussian process
$\theta$ as an integral with respect to a random field, that is,
\begin{equation*}
\theta(t) = - \int_{0}^t \int_0^{t} 1_{s +x \le t} dU\left(\frac{s}{1-\Vert h\Vert_{L^{1}}}, F(x) \right),
\end{equation*}
where the
Kiefer process $U(\cdot, \cdot)$ is a two-parameter continuous centered Gaussian process on $\mathbb{R}_+ \times [0,1]$ with covariance function
\begin{equation*}
\mathbb{E}[ U(s, x) U(t,y) ]= (s \wedge t ) (x \wedge y - xy).
\end{equation*}
\end{remark}
As $\xi$ is independent of the other three Gaussian processes $W^0, G, \theta$, so for given $\xi = x_0 \in \mathbb{R},$ we obtain that the limit process $X$ in Proposition~\ref{thm:IS1} is Gaussian. We next discuss the properties of this Gaussian limit $X$ with a given initial condition $X(0)=\xi= x_0 \in \mathbb{R}.$
\subsection{Properties of the Gaussian process $X$ in \eqref{eq:X}} \label{sec:property}
In this section, we characterize the Gaussian limit process $X$ in Proposition~\ref{thm:IS1} by computing the mean, covariance function, and long-term behavior of $X$ with a given initial condition $X(0)= x_0 \in \mathbb{R}.$
It is clear from Proposition~\ref{thm:IS1} that for each fixed $t \ge 0,$ the mean of $X(t)$ is given by:
\begin{equation*}\label{eq:EXt}
\mathbb{E}[X(t)| X(0) = x_0]=(1-F_{0}(t)) x_0.
\end{equation*}
To compute the covariance of $X$, we can obtain from Proposition~\ref{thm:IS1} that for
$t\geq s \ge 0$,
\begin{align}
\text{Cov}(X(t),X(s))
&=q_{0}\text{Cov}(W^{0}(F_{0}(t)),W^{0}(F_{0}(s)))+\text{Cov}(\theta(t),\theta(s))
\nonumber \\
&\qquad
+\text{Cov}\left(\int_{0}^{t}(1-F(t-u))dG(u),\int_{0}^{s}(1-F(s-v))dG(v)\right). \label{eq:covX}
\end{align}
By using the property of Brownian bridge, for $t\geq s$, we have
\begin{equation*}
q_{0}\text{Cov}(W^{0}(F_{0}(t)),W^{0}(F_{0}(s)))
=q_{0}F_{0}(s)(1-F_{0}(t)).
\end{equation*}
In addition, $\text{Cov}(\theta(t),\theta(s))$ is already given in \eqref{eq:cov-theta}. Hence,
it suffices to compute the last term in \eqref{eq:covX}.
We can directly compute that
\begin{align*}
&\text{Cov}\left(\int_{0}^{t}(1-F(t-u))dG(u),\int_{0}^{s}(1-F(s-v))dG(v)\right)
\nonumber \\
&=\mathbb{E}\left[\int_{0}^{t}(1-F(t-u))dG(u)\int_{0}^{s}(1-F(s-v))dG(v)\right]
\nonumber \\
&=\mathbb{E}\left[\int_{0}^{s}(1-F(t-u))dG(u)\int_{0}^{s}(1-F(s-v))dG(v)\right]
\nonumber \\
&\qquad\qquad\qquad
+\mathbb{E}\left[\int_{s}^{t}(1-F(t-u))dG(u)\int_{0}^{s}(1-F(s-v))dG(v)\right]
\nonumber
\\
&=\int_{0}^{s}(1-F(t-u))(1-F(s-u))dK(u)
\nonumber
\\
&\qquad\qquad\qquad\qquad
+\int_{0}^{s}\int_{s}^{t}(1-F(t-u))(1-F(s-v))\phi(v-u)dvdu
\nonumber
\\
&=\frac{1}{1-\Vert h\Vert_{L^{1}}}\int_{0}^{s}(1-F(t-u))(1-F(s-u))du
\nonumber
\\
&\qquad\qquad\qquad\qquad
+\int_{0}^{s}\int_{0}^{t}(1-F(t-u))(1-F(s-v))\phi(v-u)dudv,
\end{align*}
where we used \eqref{eq:var}.
Thus, we obtain the following result.
\begin{proposition}[Covariance function of the Gaussian process $X$ in \eqref{eq:X}] \label{prop:covX}
Given $X(0)= x_0 \in \mathbb{R}.$
For $0 \le s \le t,$ we have
\begin{align*}
\text{Cov}(X(s),X(t))
&=q_{0}F_{0}(s)(1-F_{0}(t)) + \frac{1}{1-\Vert h\Vert_{L^{1}}}\int_{0}^{s}(1-F(t-u))du
\\
&\qquad\qquad
+\int_{0}^{s}\int_{0}^{t}(1-F(t-u))(1-F(s-v))\phi(v-u)dudv,
\end{align*}
where $\phi$ is determined by the exciting function $h$ from Equation~\eqref{eq:phi}, and $\phi(x) = \phi(-x)$ for $x<0$.
\end{proposition}
An immediate observation from this result is that the covariance density of the traffic input Hawkes process, together with the service time distributions, leads to a delicate correlation structure of the limiting scaled queue length process $X$.
From Proposition~\ref{prop:covX}, we can immediately find that given $X(0)= x_0 \in \mathbb{R},$
\begin{align}
\text{Var}(X(t))
&=q_{0}F_{0}(t)(1-F_{0}(t))
+\frac{1}{1-\Vert h\Vert_{L^{1}}}\int_{0}^{t}(1-F(u))du
\nonumber \\
&\qquad\qquad
+\int_{0}^{t}\int_{0}^{t}(1-F(u))(1-F(v))\phi(v-u)dudv. \label{eq:varXt}
\end{align}
Note that $(1-F_0(t)) \xi$ converges to 0 almost surely as $t\rightarrow\infty$.
In view of \eqref{eq:X} and by letting $t\rightarrow\infty$ in \eqref{eq:varXt}, we get the following result about the long-term behavior of the limiting process $X$.
\begin{corollary}\label{prop:X-infty}
As $t \rightarrow \infty,$ the sequence of random variables $X(t)$ in \eqref{eq:X} converges in distribution to $X(\infty)$ which is a Gaussian random variable with mean zero and variance
\begin{align*}
\text{Var}(X(\infty))
=\frac{1}{1-\Vert h\Vert_{L^{1}}}\int_{0}^{\infty}(1-F(u))du
+\int_{0}^{\infty}\int_{0}^{\infty}(1-F(u))(1-F(v))\phi(v-u)dudv.
\end{align*}
\end{corollary}
\subsection{A special case: exponential service times} \label{sec:exp}
In this section, we discuss in detail the special case that service times of each customer are mutually independent and exponentially distributed. Without loss of generality, we consider service time distribution with mean one.
Then we have the following result.
\begin{proposition} \label{thm:IS}
Suppose Assumption~1 holds.
Assume \eqref{eq:initial} and
\begin{equation}\label{ass1}
q_0= \frac{1}{1-\Vert h\Vert_{L^{1}}}, \quad \text{and} \quad F(x)=F_0(x) =1- e^{-x}, \quad \text{$x \ge 0$.}
\end{equation}
Then as $\mu \rightarrow \infty,$
the sequence of processes
$X^{\mu}$ in \eqref{eq:X-Q-1d} converges in distribution to the process $X_e$ with continuous sample paths in $(D([0,\infty),\mathbb{R}),J_{1})$ and
\begin{equation*}\label{eq:Xe1}
X_e(t) =\xi \cdot e^{-t} + \frac{1}{\sqrt{1-\Vert h\Vert_{L^{1}}}}\cdot e^{-t} \cdot \int_{0}^t e^s dB(s) + e^{-t} \cdot \int_{0}^t e^s dG(s),
\end{equation*}
or equivalently,
\begin{equation} \label{eq:Xe}
X_e(t) =\xi - \int_{0}^t X_e(s) ds + \frac{1}{\sqrt{1-\Vert h\Vert_{L^{1}}}} \cdot B(t) +G(t),
\end{equation}
where $G$ is the mean-zero Gaussian process given in Theorem~\ref{thm:FCLT}, $B$ is a standard Brownian motion, and $\xi, G, B$ are mutually independent. In addition, the Gaussian process $X_e$ is non--Markovian unless $h \equiv 0$.
\end{proposition}
The proof of the weak convergence in this result immediately follows from Proposition~\ref{thm:IS1} and Part II of Theorem~3 in \cite{Kri1997}. The non-Markovian property of $X_e$ is also evident given the non-Markovian property of $G$ in Proposition~\ref{prop:G-property}. We omit the proof.
Note under the assumptions in \eqref{ass1}, one can readily verify from \eqref{eq:X-Q-1d} that
\begin{eqnarray} \label{eq:X-mu-t-exp}
X^{\mu}(t) ={\sqrt{\mu}} \left( \frac{Q^{\mu}(t)}{\mu} - q_0 \right) = \frac{1}{\sqrt{\mu}} \left(Q^{\mu}(t) - \frac{\mu}{1-\Vert h\Vert_{L^{1}}}\right) = \frac{1}{\sqrt{\mu}} \left(Q^{\mu}(t) - \bar \lambda \right).
\end{eqnarray}
In the classical case where the traffic is Poisson, i.e., $h \equiv 0$, it is well known that $G$ reduces to a standard Brownian motion, and the sequence $X^{\mu}$ converges in distribution to the limit process $X_e$ where $X_e$ is an Ornstein--Uhlenbeck (OU) diffusion process (driven by a Brownian motion) which is Markovian.
When the traffic model is a Hawkes process and the exciting function $h$ is nonzero, Equation~\eqref{eq:Xe} suggests that the limit process $X_e$ can be viewed as an Ornstein-Uhlenbeck (OU) process driven by the centered Gaussian process $Y$ where
\begin{equation*}
Y(t) := \frac{1}{\sqrt{1-\Vert h\Vert_{L^{1}}}} \cdot B(t) +G(t), \quad \text{for $t \ge 0.$}
\end{equation*}
We explore additional properties of the process $X_e$ in the next section.
\subsubsection{Properties of the Gaussian-driven OU process $X_e$}
From the results in Section~\ref{sec:property}, we can immediately obtain the mean, the covariance function and the long-term behavior of the Gaussian-driven OU process $X_e$.
\begin{proposition}\label{prop:Xe-cov}
Assume that $X_e(0)=x_0 \in \mathbb{R}$. Then we have
\begin{equation*}
\mathbb{E}[X_e(t)|X_e(0)=x_0]= x_0 \cdot e^{-t}, \quad \text{$t \ge 0$.}
\end{equation*}
In addition, for $0\leq s\leq t$, we have
\begin{equation*} \label{eq:cov_Xe}
\mbox{Cov}(X_e(s),X_e(t))
=e^{-s-t}\bigg[ (e^{2s}-1) \cdot \frac{1}{1 - \Vert h \Vert_{L^1}}
+\int_{0}^{t}\int_{0}^{s} e^{u+ v} \phi(u-v)dv du
\bigg],
\end{equation*}
where $\phi$ is determined by the exciting function $h$ from Equation~\eqref{eq:phi}.
Finally, as $t \rightarrow \infty,$ the sequence of random variables $X_e(t)$ in \eqref{eq:Xe} converges in distribution to $X_e(\infty)$ which is a Gaussian random variable with mean zero and variance
\begin{equation} \label{eq:var-Xe-infty}
\mbox{Var}(X_e(\infty))
=\int_{0}^{\infty}e^{-t}\phi(t)dt+ \frac{1}{1-\Vert h\Vert_{L^{1}}}.
\end{equation}
\end{proposition}
One can obtain an explicit formula for the covariance function and long-term limit of $X_e$ in the special case $h(t)=\alpha e^{-\beta t}$ where $0 \le \alpha< \beta$.
Recall when $h(t)=\alpha e^{-\beta t}$, we have $\phi$ given in \eqref{eq:phi-exp}.
Therefore, from Proposition~\ref{prop:Xe-cov} we can compute that for $0\leq s\leq t$:
\begin{align*}
&\mbox{Cov}(X_e(s),X_e(t)) \\
&=e^{-s-t}\bigg[(e^{2s}-1)\frac{\beta}{ \beta - \alpha}
+\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}
\int_{0}^{t}\int_{0}^{s}e^{(u+v)}e^{-(\beta-\alpha)|u-v|}dvdu\bigg]
\\
&=e^{-s-t}\bigg[(e^{2s}-1)\frac{\beta}{ \beta - \alpha}
+\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}
\frac{(e^{(1-\beta+\alpha)t}-e^{(1-\beta+\alpha)s})
(e^{(\beta-\alpha+1)s}-1)}{(1+\beta-\alpha)(1-\beta+\alpha)}
\\
&\qquad
+\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}
\frac{1}{1+\beta-\alpha}\left(\frac{e^{2s}-1}{2}-\frac{e^{(1-\beta+\alpha)s}-1}{1-\beta+\alpha}\right)
\\
&\qquad\qquad
+\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}
\frac{1}{1-\beta+\alpha}
\left(\frac{e^{2s}-e^{(1-\beta+\alpha)s}}{1+\beta-\alpha}
-\frac{e^{2s}-1}{2}\right)\bigg].
\end{align*}
In addition,
\begin{equation} \label{VAR1}
\mbox{Var}(X_e(\infty))
=\frac{\alpha\beta(2\beta-\alpha)}{2(\beta-\alpha)^{2}}\cdot \frac{1}{1+\beta-\alpha}+\frac{\beta}{ \beta - \alpha}.
\end{equation}
When $h$ is not a single exponential function, let us discuss how to compute the Laplace
transform of $\phi$ in \eqref{eq:var-Xe-infty} in general. It is proved in Lemma 11 of \cite{ZhuI}
that if $h$ is positive, continuous and integrable, $h$ can be approximated
by a sum of exponentials in both $L^{1}$ and $L^{\infty}$ norms,
that is there exist $\alpha_{i}\in\mathbb{R}$ and $\beta_{i}\in\mathbb{R}_{+}$ so that
$h_{n}(t):=\sum_{i=1}^{n}\alpha_{i}e^{-\beta_{i}t}\geq 0$ for every $t>0$
and $h_{n}\rightarrow h$ in both $L^{1}$ and $L^{\infty}$ norms.
For such a general exciting function $h$, let us write $h(t)=\sum_{i=1}^{\infty}\alpha_{i}e^{-\beta_{i}t}$.
Thus, for any $\omega>0$,
\begin{align*}
\int_{0}^{\infty}e^{-\omega t}\int_{0}^{\infty}h(t+v)\phi(v)dvdt
&=\int_{0}^{\infty}e^{-\omega t}\int_{0}^{\infty}\sum_{i=1}^{\infty}\alpha_{i}e^{-\beta_{i}(t+v)}\phi(v)dvdt
\\
&=\sum_{i=1}^{\infty}\frac{\alpha_{i}}{\beta_{i}+\omega}\tilde{\phi}(\beta_{i}),
\end{align*}
where $\tilde{g}(\omega):=\int_{0}^{\infty}e^{-\omega t}g(t)dt$ for $\omega>0$ and a given function $g$.
Thus, by taking Laplace transform on both sides of \eqref{eq:phi}, we get
\begin{equation}\label{phiLaplace}
\tilde {\phi}(\omega)=\frac{\tilde{h}(\omega)}{(1-\tilde{h}(\omega))(1-\Vert h\Vert_{L^{1}})}
+\frac{1}{1-\tilde{h}(\omega)}\sum_{i=1}^{\infty}\frac{\alpha_{i}}{\beta_{i}+\omega}\tilde{\phi}(\beta_{i}).
\end{equation}
By letting $\omega=\beta_{i}$, $i=1,2,\ldots$, we get
\begin{equation*}
\tilde{\phi}(\beta_{i})=\frac{\tilde {h}(\beta_{i})}{(1-\tilde{h}(\beta_{i}))(1-\Vert h\Vert_{L^{1}})}
+\frac{1}{1-\tilde{h}(\beta_{i})}\sum_{j=1}^{\infty}\frac{\alpha_{j}}{\beta_{j}+\beta_{i}}\tilde{\phi}(\beta_{j}).
\end{equation*}
Let $\mathbf{\tilde X}$ denote the vector
with $\mathbf{\tilde X}_{i}=\tilde{\phi}(\beta_{i})$, and $\mathbf{R}$ be the vector
with $\mathbf{R}_{i}=\frac{\tilde{h}(\beta_{i})}{(1-\tilde {h}(\beta_{i}))(1-\Vert h\Vert_{L^{1}})}$
and $\mathbf{M}$ be the matrix with entries $\mathbf{M}_{ij}=\frac{1}{1-\tilde{h}(\beta_{i})}\frac{\alpha_{j}}{\beta_{j}+\beta_{i}}$,
and finally $\mathbf{I}$ be the identity matrix. Thus, we have
\begin{equation} \label{eq:tildeX}
\mathbf{\tilde X}=\mathbf{R}+\mathbf{M}\mathbf{\tilde X},
\end{equation}
which implies that $\mathbf{\tilde X}=\mathbf{R}(\mathbf{I}-\mathbf{M})^{-1}$ provided that
$\mathbf{I}-\mathbf{M}$ is invertible, which holds
if for example the spectral radius of $\mathbf{M}$ is strictly less than $1$.
In practice, if we consider $h(t)=\sum_{i=1}^{d}\alpha_{i}e^{-\beta_{i}t}$, for some finite $d\in\mathbb{N}$,
where $\beta_{i}>0$, $\alpha_{i}\in\mathbb{R}$
and $h(t)\geq 0$ for every $t\geq 0$, then one can readily obtain $\mathbf{R}$ and $\mathbf{M}$, and hence
$\mathbf{\tilde X}$ and $\tilde{\phi}(\beta_{i})$ can be easily solved.
Once the values of $\tilde{\phi}(\beta_{i})$ are determined, so is the Laplace transform of $\phi$
given in \eqref{phiLaplace}. An example to illustrate this procedure will be provided in the next section (Example 2).
\subsubsection{Gaussian approximations and numerical experiments} \label{sec:Gaussian-approx}
Note that Proposition~\ref{thm:IS} and \eqref{eq:X-mu-t-exp} suggest that when the baseline intensity $\mu$ of the Hawkes arrival process is large, we can heuristically approximate the steady-state distribution of the number of customers $Q^{\mu}(\infty)$ in the $Hawkes/M/\infty$ queue as follows:
\begin{equation} \label{eq:approx-Gaussian}
Q^{\mu}(\infty) \approx \bar \lambda + \sqrt{\mu} X_e(\infty),
\end{equation}
where the random variable $\bar \lambda + \sqrt{\mu} X_e(\infty)$ follows a normal distribution with mean zero, and variance $\mu \cdot Var(X_e(\infty))$ where $Var(X_e(\infty))$ is given in \eqref{eq:var-Xe-infty}. A more precise statement of the approximation in \eqref{eq:approx-Gaussian} is
\begin{equation} \label{eq:approx-gau}
\mathbb{P}(Q^{\mu}(\infty) = i ) \approx \frac{1}{\sigma} f\left( \frac{i - \bar \lambda }{\sigma}\right) , \quad \text{for $i=0, 1, \ldots,$}
\end{equation}
where $\sigma:= \sqrt{\mu \cdot \text{Var}(X_e(\infty))} $, and $f$ is the probability density function of a standard normal distribution.
We now present numerical experiments to demonstrate that the Gaussian approximation in \eqref{eq:approx-gau} is effective by making comparisons with simulations of the $Hawkes/M/\infty$ queue. We consider two examples.
\textit{Example 1.} We first consider the Hawkes input process with a single exponential function:
\begin{equation*}\label{eq:h1}
h_1(t) = \frac{1}{2} e^{-t}.
\end{equation*}
It is clear that $\Vert h_1\Vert_{L^{1}} = \frac{1}{2}$. Suppose the stationary Hawkes input process has a baseline intensity $\mu$. Then we can infer from \eqref{VAR1} that the Gaussian random variable $\bar \lambda + \sqrt{\mu} X_e(\infty)$ has mean $\bar \lambda = 2 \mu$ and variance $\sigma^2 = \mu \cdot\text{Var}(X_e(\infty)) = 3 \mu$, where we have used \eqref{VAR1} to find that
$\text{Var}(X_e(\infty))=3.$
We now compare the Gaussian approximation in \eqref{eq:approx-gau} with simulations in Figure~\ref{fig1}. We observe that the approximation
agrees with the simulation results well, even for moderately large $\mu$ such as twenty.
\begin{figure}[htb]
\centering
\subfigure[$\mu =20$]{
\begin{minipage}[b]{0.44\textwidth}
\centering
\includegraphics[width=\textwidth]{1.pdf}
\end{minipage} }
\subfigure[$\mu = 100$]{
\label{fig:mini:subfig:1}
\begin{minipage}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{2.pdf}
\end{minipage}}
\caption{The steady-state distribution of the number of customers $Q^{\mu}(\infty)$ in the $Hawkes/M/\infty$ queue where the input is a stationary Hawkes process with a baseline intensity $\mu$ and an exciting function $h_1(t) = \frac{1}{2} e^{-t}$. The service time distribution is exponential with mean one. The Gaussian approximation in \eqref{eq:approx-gau} is compared with simulations.}
\label{fig1}
\end{figure}
\textit{Example 2:}
We next consider a Hawkes input process with an exciting function which is a sum of exponentials:
\begin{equation*} \label{eq:h2}
h_2(t) = \frac{1}{10} e^{-\frac{1}{4}t} + \frac{2}{5} e^{-4t}.
\end{equation*} It is also clear that $\Vert h_2\Vert_{L^{1}} = 0.5$. In this case, when the baseline intensity of the Hawkes process is $\mu$, we have the Gaussian random variable $\bar \lambda + \sqrt{\mu} X_e(\infty)$ has mean $\bar \lambda = 2 \mu$ and variance $\sigma^2 = \mu \cdot Var(X_e(\infty)) = 2.5246 \mu$.
To see this, we compute $Var(X_e(\infty))$ using \eqref{eq:var-Xe-infty} and Equations~\eqref{phiLaplace} and \eqref{eq:tildeX}.
Note that $\tilde h_2 (1) =\int_0^{\infty} e^{-t}h_2(t) dt = 0.16$, and hence
\begin{equation} \label{eq:t-phi1}
\int_0^{\infty} e^{-t} \phi(t) dt = \tilde \phi(1) = \frac{0.16}{0.84} \cdot 2 + \frac{0.08}{0.84} \cdot (\tilde \phi(0.25) + \tilde \phi(4)).
\end{equation}
One can readily verify from the expression of $h_2$ that
\begin{equation*}
\mathbf{R} = [0.8333, 0.1587] , \quad \text{and} \quad \mathbf{M} = \begin{bmatrix}
0.2833 & 0.1333 \\
0.0254 & 0.054 \\
\end{bmatrix}.
\end{equation*}
This yields
\begin{equation*}
\mathbf{\tilde X} = [\tilde \phi(0.25), \tilde \phi(4)]=\mathbf{R}(\mathbf{I}-\mathbf{M})^{-1}= [1.1745, 0.3333].
\end{equation*}
On combining \eqref{eq:t-phi1} and \eqref{eq:var-Xe-infty} we deduce that when the traffic is a Hawkes process with an exciting function $h_2$, then we have
\begin{equation*}
\mbox{Var}(X_e(\infty))
= \tilde \phi(1) + 2 = 2.5246.
\end{equation*}
\begin{figure}[htb]
\centering
\subfigure[$\mu =20$]{
\begin{minipage}[b]{0.45\textwidth}
\centering
\includegraphics[width=\textwidth]{3.pdf}
\end{minipage} }
\subfigure[$\mu = 100$]{
\label{fig:mini:subfig:1}
\begin{minipage}[b]{0.44\textwidth}
\centering
\includegraphics[width=\textwidth]{4.pdf}
\end{minipage}}
\caption{The steady-state distribution of the number of customers $Q^{\mu}(\infty)$ in the $Hawkes/M/\infty$ queue where the input is a stationary Hawkes process with a baseline intensity $\mu$ and an exciting function $h_2(t) =\frac{1}{10} e^{-\frac{t}{4}} + \frac{2}{5} e^{-4t}$. The service time distribution is exponential with mean one. The Gaussian approximation in \eqref{eq:approx-gau} is compared with simulations.}
\label{fig2}
\end{figure}
We next demonstrate in Figure~\ref{fig2} that the Gaussian approximation in \eqref{eq:approx-gau} is effective by comparing with simulations of the infinite--server queue with Hawkes input where the exciting function is $h_2$. We also observe that the Gaussian approximation for the steady--state customer number $Q^{\mu}(\infty)$ agrees with the simulation results very well.
\begin{comment}
Using the infinite-server approximation (see, e.g., \cite{whitt2002}), we can obtain a heuristic
rule for capacity dimensioning as follows: given a performance target the wait probability is $p>0$, we can deduce from \eqref{eq:X-Q} that
\[p = \mathbb{P}(\text{Delay}) \approx \mathbb{P}( Q \ge s) = \mathbb{P}\left( \frac{Q-m}{\sqrt{v}} \ge \frac{s-m}{\sqrt{v}} \right), \]
where $X(\infty)$ is a Gaussian random variable with mean zero and variance given in \eqref{eq:var-X-infty}. Hence, we deduce that one should set the capacity, the number of servers $S$, at
\begin{eqnarray} \label{eq:N}
S &=& \frac{\mu}{1-\Vert h\Vert_{L^{1}}} + q \cdot \sqrt{\mbox{Var}(X(\infty))} \cdot\sqrt{\mu} \nonumber \\
&=& \frac{\mu}{1-\Vert h\Vert_{L^{1}}} + q \cdot \sqrt{(1-\Vert h\Vert_{L^{1}}) \cdot \mbox{Var}(X(\infty))} \cdot\sqrt{\frac{\mu}{1-\Vert h\Vert_{L^{1}}}},
\end{eqnarray}
where $q$ is determined by the equation $p = 1 - \Phi(q)$, where $\Phi$ is the cumulative distribution function of a standard normal random variable. Here the quantity $\frac{\mu}{1-\Vert h\Vert_{L^{1}}}$ is the average arrival rate to the system, or equivalently, the offered load, since the mean service time is assumed to be one. The formula \eqref{eq:N} is known as ``square root staffing formula": the capacity level $S$ is given by the offered load plus an additional ``square root" capacity margin.
What is worthy to note is that in \eqref{eq:N}, we infer from \eqref{eq:var-X-infty} that
\begin{eqnarray*}
{( 1-\Vert h\Vert_{L^{1}} ) \cdot \mbox{Var}(X(\infty))} &=& \frac{1}{2}\frac{1}{(1-\Vert h\Vert_{L^{1}})^{2}}+\frac{1}{2}(1-\Vert h\Vert_{L^{1}}) + (1-\Vert h\Vert_{L^{1}}) \int_{0}^{\infty}e^{-t}\phi(t)dt \nonumber \\
&\geq& \frac{1}{2}\frac{1}{1-\Vert h\Vert_{L^{1}}}+\frac{1}{2}(1-\Vert h\Vert_{L^{1}})\geq 1.
\end{eqnarray*}
with equality achieved if and only if $h \equiv 0$, i.e., the Poisson arrivals. The insights we gain from this computation are twofold: (a) when the arrival is given by a stationary Hawkes process, to achieve the same steady--state delay probability in a many-server queueing system, one needs more capacities/servers compared with a Poisson process with the same average arrival rate; (b) The amount of additional capacities needed depends in a delicate way on the exciting function $h$ which characterizes the correlation in the arrivals.
\end{comment}
\section{Infinite-server queues with mutually--exciting traffic} \label{sec:multi-hawkes}
In this section we extend Theorem~\ref{thm:FCLT} to multivariate stationary Hawkes processes, and establish limit theorems for infinite--server queues with multivariate Hawkes traffic.
\subsection{FCLT for multivariate stationary Hawkes processes}
In this section, we establish an FCLT for multivariate stationary Hawkes processes.
We consider a $k$-dimensional stationary Hawkes process $\mathbb{N}^{(\mu)}=(N^{\mu,1},N^{\mu,2},\ldots,N^{\mu,k})$,
where $N^{\mu,i}$ has the intensity:
\begin{equation} \label{eq:intensity-multi}
\lambda^{\mu,i}(t)=\mu p_{i}+\sum_{j=1}^{k}\int_{-\infty}^{t-}h_{ij}(t-s)N^{\mu,j}(ds),
\end{equation}
where $p_{i} \ge 0$ for $1\leq i\leq k$, and the exciting kernel $h_{ij}(\cdot)$ satisfies Assumption~\ref{assump1}. With slight abuse of notations, we still use $\mu>0$ as a scaling parameter, and study the limit of $\mathbb{N}^{\mu}$ as we send $\mu \rightarrow \infty$. Note that for each fixed $\mu>0$, we can obtain from \eqref{BarLambdaDefn} that for each $t,$
\begin{equation*} \label{eq:lambda-a}
\mathbb{E}[\mathbb{N}^{(\mu)}(t)]= \bar{\lambda} t =\mu a t,
\end{equation*}
where $\bar \lambda=(\bar \lambda_i)_{1 \le i \le k}$ is the average arrival rate and the vector $a=(a_i)_{1 \le i \le k}$ is given by
\begin{equation} \label{eq:a}
a := \left( \mathbb{I} - \mathbb{H}\right)^{-1} \cdot p,
\end{equation}
with $p=(p_i)_{1 \le i \le k}$, and $\mathbb{H}= (\Vert h_{ij}\Vert_{L^{1}})_{1\leq i,j\leq k}$.
Similar as in the univariate case, we can infer from the immigration-birth representation of multivariate Hawkes processes, when $\mu$ is a positive integer,
that the multivariate stationary Hawkes process $\mathbb{N}^{(\mu)}$ can be written as the sum of i.i.d. copies of $\mathbb{N}^{(1)}$, see e.g. \cite{JHR}. The covariance density of $\mathbb{N}^{(1)}$, which we still use the notation $\Phi=(\Phi_{ij})_{1\leq i,j\leq k}$ as in \eqref{eq:cov-density}, is given by
\begin{equation}\label{cov-multi}
\Phi(\tau)=h(\tau)\text{diag}(a)+\int_{-\infty}^{\tau}h(\tau-v)\Phi(v)dv, \quad \text{for $\tau\geq 0$,}
\end{equation}
and $\Phi_{ij}(-\tau)=\Phi_{ji}(\tau)$ for every $\tau>0$ and $1\leq i,j\leq k$.
Here $\text{diag}(a)$ is the diagonal matrix with entries $a_{i}$'s on the diagonal,
and $h(t)=(h_{ij}(t))_{1\leq i,j\leq k}$. The variance function of $\mathbb{N}^{(1)}$, which we still use the notation
$\mathbb{K}(t)=(K_{ij}(t))_{1\leq i,j\leq k}$ as in \eqref{eq:var-function}, is given by
\begin{equation}\label{var-multi}
\mathbb{K}(t):=\text{diag}(a)t+2\int_{0}^{t}\int_{0}^{t_{2}}\Phi(t_{2}-t_{1})dt_{1}dt_{2}.
\end{equation}
Then we can obtain the following limit theorem for multivariate stationary Hawkes processes, which extends Theorem~\ref{thm:FCLT}. The proof is given in Appendix~\ref{sec:proof-multivariate}.
\begin{theorem}\label{thm:multi-FCLT}
Under Assumption~\ref{assump1}, we have as $\mu\rightarrow\infty$,
\begin{equation*}
\frac{\mathbb{N}^{(\mu)}(t)-\bar \lambda t}{\sqrt{\mu}}\Rightarrow\mathbb{G}(t),
\end{equation*}
in $(D([0,\infty),\mathbb{R}^{k}),J_{1})$, where $\mathbb{G}= (\mathbb{G}_i)_{1 \le i \le k}$ is a mean-zero almost surely continuous $k-$dimensional Gaussian process
with the covariance function
\begin{equation*}\label{eq:multi-cov-G}
\mbox{Cov}(\mathbb{G}(t),\mathbb{G}(s))=\int_{s}^{t}\int_{0}^{s}\Phi(u-v)dvdu+\mathbb{K}(s), \quad \text{for $t \ge s$,}
\end{equation*}
where $\Phi$ is given in \eqref{cov-multi} and $\mathbb{K}$ is given in \eqref{var-multi}.
\end{theorem}
\subsection{Limit theorem for $GI/\infty$ queues with multivariate Hawkes traffic}
In this section, we rely on Theorem~\ref{thm:multi-FCLT} to develop approximations for infinite--server queues with high-volume multivariate stationary Hawkes traffic.
Such a queueing model can be viewed as a multi-class queueing model with correlated arrivals as mutually--exciting Hawkes processes.
We first establish limit theorems for such queues.
Similar as in Section~\ref{sec:limit}, we consider a sequence of infinite-server queueing models indexed by $\mu \in \mathbb{R}_{+}$ and
let $\mu \rightarrow \infty$. For each fixed $\mu$, there are $k$ classes of customers arriving to the $\mu-$th system according to a stationary $k-$dimensional Hawkes process $\mathbb{N}^{(\mu)}$ with a baseline intensity vector $\mu \cdot {p}$ and an exciting kernel $(h_{ij})_{1\leq i,j\leq k}$. That is, the arrival process of customer class $i$ is the $i-$th component of the multivariate stationary Hawkes process $\mathbb{N}^{(\mu)}$.
In addition, for $1 \le i \le k$, each customer class $i$ may have different service requirements. For each $i$,
we assume given an i.i.d. sequence of nonnegative random variables $\{\bar \eta_{i,j}: j \ge 1\}$ with a cumulative distribution function $F_{i0}(x) = \mathbb{P}(\bar \eta_{i,j} \le x)$ and another i.i.d. sequence of nonnegative random variables $\{\eta_{i,j}: j \ge 1\}$ with a cumulative distribution function $F_i(x) = \mathbb{P}(\eta_{i,1} \le x)$. Assume $F_{i0}(0)=F_i(0)=0$ for all $i$ for simplicity. The customers of class $i$ initially present in the infinite--server queueing system have remaining service times $\bar \eta_{i,1}, \bar \eta_{i,2}, \ldots$; the new arriving customers of class $i$ have service times $\eta_{i,1}, \eta_{i,2}, \ldots.$
All these service times, the random initial numbers of customers of each class denoted by $\mathbb{Q}_1^{\mu}(0), \ldots, \mathbb{Q}_k^{\mu}(0)$, and the multivariate Hawkes arrival process $\mathbb{N}^{(\mu)}$ are assumed to be mutually independent.
Denote $\mathbb{Q}_i^{\mu}(t) $ as the number of customers of class $i$ in the $\mu-$th system at time $t$, and write the vector $\mathbb{Q}^{\mu}(t) := (\mathbb{Q}_i^{\mu}(t))_{1 \le i \le k}$.
Given Theorem~\ref{thm:multi-FCLT}, we can then obtain the following result. Recall the vector $a=(a_i)_{1 \le i \le k}$ given in \eqref{eq:a}.
\begin{proposition} \label{thm:IS-multi-gen}
Suppose Assumption~1 holds.
Assume that for some vector constant $q=(q_{10}, \ldots, q_{k0})$ and random vector $(\xi_1, \ldots, \xi_k),$
\begin{eqnarray} \label{eq:initial}
\sqrt{\mu} \left( \frac{\mathbb{Q}^{\mu}(0)}{\mu} - q \right) \Rightarrow (\xi_1, \ldots, \xi_k), \quad \text{as $\mu \rightarrow \infty.$}
\end{eqnarray}
Then the sequence of $k-$dimensional processes
$\mathbb{X}^{\mu}$ with its $i-$th component defined by
\begin{eqnarray}
\mathbb{X}_i^{\mu}(t) &=&{\sqrt{\mu}} \left( \frac{\mathbb{Q}_i^{\mu}(t)}{\mu} - q_{i0} (1-F_{i0}(t)) - a_i \cdot \int_{0}^t (1-F_i(t-u)) du \right), \label{eq:X-Q}
\end{eqnarray}
as $\mu \rightarrow \infty,$
converges in distribution in $(D([0,\infty),\mathbb{R}^k),J_{1})$ to the process $\mathbb{X}=(\mathbb{X}_i)_{1 \le i \le k}$ where
\begin{eqnarray*} \label{eq:X-multi-gen}
\mathbb{X}_i(t) &=&(1- F_{i0}(t)) \xi_i + \sqrt{q_{i0}} \cdot W^{i0} (F_{i0}(t)) + \int_{0}^t (1- F_i(t-u)) d\mathbb{G}_i(u) \nonumber \\
&& - \int_0^t \int_0^t 1_{s +x \le t} dU_i \left(a_i s, F_i(x) \right).
\end{eqnarray*}
Here, $W^{i0}=\{W^{i0}(x): x \in [0,1]\}$ is a Brownian bridge,
$\mathbb{G}_i$ is the $i-$th component of the $k-$dimensional Gaussian process $\mathbb{G}$ given in Theorem~\ref{thm:multi-FCLT}, $U_i$ is a
Kiefer process which is a two-parameter continuous centered Gaussian process on $\mathbb{R}_+ \times [0,1]$ with covariance function
\begin{equation*}
\mathbb{E}[ U_i(s, x) U_i(t,y) ]= (s \wedge t ) (x \wedge y - xy).
\end{equation*}
For all $1 \le i \le k$, all the random elements
$\xi_i, W^{i0}, U_i$ are mutually independent, and they are independent of $ \mathbb{G}$.
\end{proposition}
Note that in the above result, the weak convergence of the component process $\mathbb{X}_i^{\mu}$ to $\mathbb{X}_i$ for each fixed $i$ follows directly from our Theorem~\ref{thm:multi-FCLT} and Theorem~3 in \cite{Kri1997}. However, we still need to show the joint weak convergence of the sequence $(\mathbb{X}_1^{\mu}, \ldots, \mathbb{X}_k^{\mu})$ as $\mu \rightarrow \infty$. We provide a proof in Appendix~\ref{sec:proof-13}.
One can also readily see that the limit process $\mathbb{X}$ in Proposition~\ref{thm:IS-multi-gen} is a $k$-dimensional Gaussian process given its initial state. The covariance function of $\mathbb{X}$ can be computed in a similar manner as we have done in Section~\ref{sec:property} in the one--dimensional case. For notational simplicity and illustration purposes, we study this limiting Gaussian process in detail for the special case of exponential service times in the following section.
\subsection{An example: exponential service times}
In this section, we consider the special case that class $i$ customers have the mean
service requirement $1/r_i$ with $r_i>0$, and the service time distributions are independent exponentials for all $i=1, \ldots, k$. That is,
\begin{equation} \label{eq:exp-multi}
F_i(x)= F_{i0}(x) = 1- e^{-r_i x}, \quad \text{for $x \ge 0$, and for each $i=1, \ldots, k$.}
\end{equation}
Then we immediately obtain the following result from Proposition~\ref{thm:IS-multi-gen} and Part II of Theorem~3 in \cite{Kri1997}. The proof is omitted.
\begin{proposition} \label{thm:IS-mutual}
Suppose Assumption~1 and \eqref{eq:exp-multi} holds. Assume \eqref{eq:initial} with $q_{i0} = a_i/r_i$ for $i=1, \ldots, k$.
Then as $\mu \rightarrow \infty,$ the sequence of processes
$\mathbb{X}^{\mu}$ in \eqref{eq:X-Q} converges in distribution to the process $\mathbb{X} = (\mathbb{X}_1, \ldots, \mathbb{X}_k)$ with continuous sample paths in $(D([0,\infty),\mathbb{R}^k),J_{1})$ and for $i=1, \ldots, k$,
\begin{eqnarray} \label{eq:X-multi}
\mathbb{X}_i(t) = \xi_i - r_i \int_{0}^t \mathbb{X}_i (s) ds + \mathbb{G}_i(t) + \sqrt{a_i} \cdot \mathbb{B}_i (t),
\end{eqnarray}
where $\mathbb{G}= (\mathbb{G}_1, \ldots, \mathbb{G}_k) $ is the mean-zero Gaussian process given in Theorem~\ref{thm:multi-FCLT}, $\mathbb{B}= (\mathbb{B}_1, \ldots, \mathbb{B}_k) $ is a standard $k-$dimensional Brownian motion, and $\mathbb{X}(0)= (\xi_i)_{1 \le i \le k}$, $\mathbb{G}$ and $\mathbb{B}$ are mutually independent.
\end{proposition}
Proposition~\ref{thm:IS-mutual} suggests that given $\mathbb{X}(0) \in \mathbb{R}^k$, the limit process $\mathbb{X}$ can be viewed as a $k$-dimensional Gaussian--driven OU process.
We next provide a characterization of this multi--dimensional Gaussian--driven OU process $\mathbb{X}$ by computing its covariance function and steady--state distribution explicitly.
Given $\mathbb{X}(0) \in \mathbb{R}^k$, for $0 \le s \le t$, we write
\begin{equation*}
\text{Cov}(\mathbb{X}(t), \mathbb{X}(s) ) =(\text{Cov}(\mathbb{X}_{i}(t),\mathbb{X}_{j}(s)))_{1\leq i,j\leq k}.
\end{equation*}
To compute this covariance function, we note from \eqref{eq:X-multi} that for each $i,$
\begin{equation*}
\mathbb{X}_{i}(t)=\mathbb{X}_{i}(0)e^{-r_{i}t}
+\sqrt{a_{i}}e^{-r_{i}t}\int_{0}^{t}e^{r_{i}s}d\mathbb{B}_{i}(s)
+e^{-r_{i}t}\int_{0}^{t}e^{r_{i}s}d\mathbb{G}_{i}(s).
\end{equation*}
Therefore, we can compute that
\begin{align*}
\text{Cov}(\mathbb{X}_{i}(t),\mathbb{X}_{j}(s))
&=\text{Cov}\left(\sqrt{a_{i}}\int_{0}^{t}e^{r_{i}(u-t)}d\mathbb{B}_{i}(u),
\sqrt{a_{j}}\int_{0}^{s}e^{r_{j}(v-s)}d\mathbb{B}_{j}(v)\right)
\\
&\qquad
+\text{Cov}\left(\int_{0}^{t}e^{r_{i}(u-t)}d\mathbb{G}_{i}(u),
\int_{0}^{s}e^{r_{j}(v-s)}d\mathbb{G}_{j}(v)\right).
\end{align*}
We can compute that
\begin{align*}
&\text{Cov}\left(\sqrt{a_{i}}\int_{0}^{t}e^{r_{i}(u-t)}d\mathbb{B}_{i}(u),
\sqrt{a_{j}}\int_{0}^{s}e^{r_{j}(v-s)}d\mathbb{B}_{j}(v)\right)
\\
&=1_{i=j}\cdot a_{i}e^{-r_{i}(t+s)}\text{Var}\left(\int_{0}^{s}e^{r_{i}v}d\mathbb{B}_{i}(v)\right)
=1_{i=j}\cdot \frac{a_{i}}{2r_{i}}[e^{-r_{i}(t-s)}-e^{-r_{i}(t+s)}],
\end{align*}
and similar as in the univariate Hawkes process case,
\begin{align*}
&\text{Cov}\left(\int_{0}^{t}e^{r_{i}(u-t)}d\mathbb{G}_{i}(u),
\int_{0}^{s}e^{r_{j}(v-s)}d\mathbb{G}_{j}(v)\right)
\\
&=
\int_{0}^{s}e^{-r_{i}(t-u)}e^{-r_{j}(s-u)}dK_{ij}(u)
+\int_{0}^{s}\int_{s}^{t}e^{-r_{i}(t-u)}e^{-r_{j}(s-v)}\Phi_{ij}(v-u)dvdu
\\
&=1_{i=j}\cdot a_{i}\int_{0}^{s}e^{-r_{i}(t-u)}e^{-r_{i}(s-u)}du
+
2\int_{0}^{s}\int_{0}^{v}e^{-r_{i}(t-u)}e^{-r_{j}(s-v)}\Phi_{ij}(u-v)dvdu
\\
&\qquad\qquad
+\int_{0}^{s}\int_{s}^{t}e^{-r_{i}(t-u)}e^{-r_{j}(s-v)}\Phi_{ij}(v-u)dvdu
\\
&=1_{i=j}\cdot \frac{a_{i}}{2r_{i}}[e^{-r_{i}(t-s)}-e^{-r_{i}(t+s)}]
+\int_{0}^{s}\int_{0}^{t}e^{-r_{i}(t-u)}e^{-r_{j}(s-v)}\Phi_{ij}(v-u)dvdu,
\end{align*}
by using the definition of $K_{ij}(u)$ in \eqref{var-multi}.
Hence, we conclude that for $1 \le i,j \le k$ and $0 \le s \le t,$
\begin{align}\label{eq:cov-kd}
&\text{Cov}(\mathbb{X}_{i}(t),\mathbb{X}_{j}(s))
\\
&=1_{i=j}\cdot \frac{a_{i}}{r_{i}}[e^{-r_{i}(t-s)}-e^{-r_{i}(t+s)}]
+\int_{0}^{s}\int_{0}^{t}e^{-r_{i}(t-u)}e^{-r_{j}(s-v)}\Phi_{ij}(v-u)dvdu.
\nonumber
\end{align}
In addition, it readily follows from \eqref{eq:cov-kd} that
\begin{equation*}
\text{Cov}(\mathbb{X}_{i}(t),\mathbb{X}_{j}(t))
=1_{i=j}\cdot \frac{a_{i}}{r_{i}}[1-e^{-2r_{i}t}]
+\int_{0}^{t}\int_{0}^{t}e^{-r_{i}u}e^{-r_{j}v}\Phi_{ij}(v-u)dvdu,
\end{equation*}
where $\Phi_{ij}$ is given in \eqref{cov-multi}.
As $t \rightarrow \infty$, the sequence of random vectors $\mathbb{X}(t)$ converges in distribution to a limiting $k-$dimensional Gaussian random vector $\mathbb{X}(\infty)$ which has mean zero and covariance
\begin{equation*}
\text{Cov}(\mathbb{X}_{i}(\infty),\mathbb{X}_{j}(\infty))
=1_{i=j}\cdot \frac{a_{i}}{r_{i}}
+\int_{0}^{\infty}\int_{0}^{\infty}e^{-r_{i}u}e^{-r_{j}v}\Phi_{ij}(v-u)dvdu.
\end{equation*}
Hence, we have obtained the covariance function and the steady--state distribution of the multi--dimensional Gaussian--driven OU process $\mathbb{X}$ in \eqref{eq:X-multi}.
\section*{Acknowledgements}
We are grateful to two anonymous referees, and the Associate Editor
for very careful readings of the manuscript, and helpful suggestions,
that greatly improve the quality of the paper.
We also thank Jim Dai for helpful comments and Junfei Huang for many useful discussions.
Xuefeng Gao acknowledges support from Hong Kong RGC ECS Grant 24207015 and CUHK Direct Grants for Research with project codes 4055035 and 4055054. Lingjiong Zhu is grateful to the support from NSF Grant DMS-1613164.
\clearpage
|
\section{Introduction}
For recent years, extensive efforts have been paid to the distributed estimation and optimization problems motived by their wide applications in sensor networks \cite{WSN1,WSN2}, cognitive networks \cite{WSN3}, multi robots \cite{robot}, as well as in distributed learning \cite{Learning}. This paper studies a distributed optimization problem, where $n$ agents connected in a network collectively minimize a convex cost function $\sum_{i=1}^n f_i(x)$ subject to a convex set constraint $\bigcap_{i=1}^n \Omega_i$.
The local cost function of agent $i$ takes the form $f_i(x)=E[h_i(x,\vartheta_i)],$ where $\vartheta_i$ is a random variable. In such a problem, the cost function $f_i(x)$ is difficult to calculate, but samples of the cost function $h_i(x,\vartheta_i)$ may serve as estimates for its expectation. It is assumed that the local constraint set $\Omega_i$ of agent $i$ is closed and convex, and the communication relationship among the agents is described by a random network. Besides, communications are imperfect since there are noises in the channels through which agents exchange information.
There exist many papers considering the related problems. A unconstrained cooperative optimization problem is investigated in \cite{optimization0} and \cite{optimization1} over the deterministic and the random switching networks, respectively. A distributed stochastic subgradient projection algorithm is proposed in \cite{nedic} to solve
a constrained optimization problem, where all agents are subject to a common convex constraint set, and subgradients of local cost functions are corrupted by stochastic errors.
Effects of stochastic subgradient errors on the convergence of the algorithm over deterministic switching networks are investigated. A distributed asynchronous algorithm with two diminishing step sizes
is proposed in \cite{nediB2} to solve the distributed constrained stochastic optimization problem.
The estimates are shown to converge to a random point in the optimal set a.s., when constraint sets are compact, the global constraint set has a nonempty interior set, and cost functions are non-smooth but with bounded subgradients. A distributed algorithm based on dual subgradient averaging is proposed in \cite{Duchi}, where it is shown how do the network size and the spectral gap of the network influence convergence rates. Consensus-based distributed primal-dual subgradient methods are given in\cite{nediB1,zhu}, where \cite{zhu} solves a problem with the cost function being the sum of local cost functions and with global convex inequality constraints known to all agents, while \cite{nediB1} solves a problem with a coupled global cost function and with inequality constraints. A primal-dual algorithm with constant step size is proposed and its convergence is shown in \cite{Ling} for the deterministic unconstrained optimization problem over undirected connected graph with perfect communications. Besides, performance of the continuous time primal-dual algorithms is also investigated in \cite{Feijer,Liu}. Generally speaking, the above mentioned algorithms can be divided into three categories: \cite{optimization0,optimization1,nedic, nediB2} belong to the primal domain algorithms, \cite{Duchi} belongs to the dual domain algorithm, while \cite{ nediB1,Ling,zhu,Liu,Feijer} belong to the primal-dual domain algorithms.
In this paper, we propose a stochastic approximation based distributed primal-dual algorithm to solve the distributed constrained stochastic optimization problem. Since it is equivalent to a convex optimization problem with a linear equality constraint and a convex set constraint,
by incorporating the augmented Lagrange technique with the projection method,
a distributed primal-dual algorithm is derived. The algorithm is distributed as in an iteration each agent updates its estimate using the noisy observations for gradients of the local functions
and the noisy observations for both primal and dual variables of the neighboring agents.
Contributions of the paper are as follows. 1)
Stability and convergence of the algorithm are proved for the constrained problem. The communication graphs
are assumed to be independent identically distributed (i.i.d) with the mean graph being undirected and connected. Communication noises and gradient errors are assumed to be martingale difference sequences (mds). Convex sets are required to have smooth boundaries with the global constraint having at least one relative interior, and gradient functions are required to be Lipschitz continuous. Then with diminishing step-size, the estimates are shown to be bounded a.s. by using the convergence theorem for martingales, and to converge to the optimal solution set a.s. by use of the results for constrained stochastic approximation \cite{Kushner}. Compared with \cite{nediB2}, here gradients are only required to be Lipschitz continuous without boundedness assumption, constraint sets are not assumed to be compact, and the global constraint is only required to have at least one relative interior point. Different from \cite{ nediB1,Ling,zhu}, the stochastic optimization problem is investigated over random networks with imperfect communications. 2) Asymptotic properties are considered for the unconstrained problem. Through dimensionality reduction, asymptotic normality and efficiency of the algorithm are established. In comparison with \cite{Duchi}, we have shown the exact influence of random networks, imperfect communications, gradient errors and the structure of cost functions on the rate of convergence.
The organization of the paper is as follows. In Section \ref{sec:PS},
some preliminary information about graph theory and convex analysis is provided and the problem is formulated.
In Section \ref{sec:Algorithm}, the basic assumptions are introduced and a stochastic approximation based distributed primal-dual algorithm is designed. Convergence for the constrained problem is established in Section \ref{sec:Convergence},
while asymptotic properties for the unconstrained problem are given in Section \ref{sec:Normality}.
Numerical simulations are demonstrated in Section \ref{sec:Simulation} with some concluding remarks given in Section \ref{sec:Conclusion}.
\section{Preliminaries and Problem Statement} \label{sec:PS}
We first introduce some preliminary results about graph theory, convex functions and convex sets,
then formulate the distributed optimization problem.
\subsection{ Graph Theory }
Consider a network of $n$ agents. The communication relationship among
agents is described by a digraph $\mathcal{G } =(\mathcal{V }, \mathcal{E }_{\mathcal{G }},\mathcal{A}_{\mathcal{G }}\}$, where $\mathcal{V }=\{ 1,\cdots,n\}$ is the node set
with node $i$ representing agent $i$;
$\mathcal{E }_{\mathcal{G }} \subset \mathcal{V } \times \mathcal{V } $ is the edge set,
and $(j,i)\in\mathcal{E }_{\mathcal{G }}$ if and only if agent $i$ can get information from agent $j$;
$\mathcal{A}_{\mathcal{G }} =[a_{ij} ]\in \mathbb{R}^{n\times n} $ is the adjacency matrix of $\mathcal{G } $,
where $a_{ij} >0$ if $(j, i)\in \mathcal{E }_{\mathcal{G }} $, and $a_{ij}=0$, otherwise.
Here, we assume the self-edge $(i,i)$ is not allowed, i.e., $a_{ii}=0~\forall i \in \mathcal{V } . $
The Laplacian matrix of graph $\mathcal{G}$ is defined as
$\mathcal{L}_{\mathcal{G}}= \mathcal{D}_{\mathcal{G}}-\mathcal{A}_{\mathcal{G}}$ with $\mathcal{D}_{\mathcal{G}}=diag\{ \sum_{j=1}^n a_{1j}, \cdots, \sum_{j=1}^n a_{nj})$, where and hereafter
$diag\{D_1,\cdots, D_n\}$ denotes the block diagonal matrix with main diagonal blocks being square matrices $D_i,~i=1,\cdots,n,$ and with the off-diagonal blocks being zero matrices.
For a bidirectional graph $\mathcal{G } $,
$ (i,j)\in \mathcal{E }_{\mathcal{G }} $ if and only if $ (j,i)\in \mathcal{E }_{\mathcal{G }}$.
The graph $\mathcal{G } $ is undirected if $\mathcal{A}_{\mathcal{G }}$ is symmetric.
The undirected graph $\mathcal{G } $ is connected if for any pair $i,j\in \mathcal{V}$, there exists a sequence of nodes
$i_1, \cdots, i_{p } \in \mathcal{V}$ such that $ (i,i_1)\in \mathcal{E }_{\mathcal{G }} , $ $ (i_1,i_2)\in \mathcal{E }_{\mathcal{G }} $,
$\cdots$, $(i_{p},j) \in \mathcal{E }_{\mathcal{G }} $. For matrix $A=[a_{ij}] \in \mathbb{R}^{n \times n}$ with $a_{i j}\geq 0~ \forall i,j=1,\cdots,n$, denote by
$\mathcal{G}_{A}= \{\mathcal{V},\mathcal{E}_{\mathcal{G}_{A}}, \mathcal{A}_{\mathcal{G}_{A}}\}$ the digraph generated by $A$, where $\mathcal{V }=\{ 1,\cdots,n\}$, $ \mathcal{A}_{\mathcal{G}_{A}}= A$, and
$(j,i)\in\mathcal{E}_{\mathcal{G}_{A}}$ if $a_{ij} >0 $.
The following lemma presents some properties of the Laplacian matrix $\mathcal{L} $ corresponding to an
undirected graph $\mathcal{G} $.
\begin{lem} \label{lem1} \cite{graph}
The Laplacian matrix $\mathcal{ L} $ of an undirected graph $\mathcal{G} $ has the following properties:
i) $\mathcal{L}$ is symmetric and positive semi-definite;
ii) $\mathcal{ L}$ has a simple zero eigenvalue and the corresponding eigenvector is $\mathbf{1}$,
and all the other eigenvalues are positive if and only if $\mathcal{G} $ is connected,
where $\mathbf{1}$ denotes the vector with all entries equal to 1.
\end{lem}
\subsection{Gradient, Projection Operator and Normal Cone}
For a given function $f: \mathbb{R}^m \rightarrow [-\infty, \infty],$
denote its domain as $ \textrm{dom} (f) \triangleq(x \in \mathbb{R}^m: f(x) < \infty\}.$
Let $f(\cdot)$ be a convex function, and let $x \in \textrm{dom} (f)$.
For a smooth (differentiable) function $f(\cdot)$, denote by $\nabla f(x)$ and by $\nabla^2 f(x)$ the gradient and Hessian of $f(\cdot)$ at point $x$, respectively.
Then \begin{equation} \label{gradient}
f(y) \geq f(x) + \nabla f(x)^T (y-x) ~~ \forall y \in \textrm{dom} (f),
\end{equation}
where $x^T$ denotes the transpose of $x.$
For a nonempty closed convex set $\Omega \subset \mathbb{R}^m$ and a point $x \in \mathbb{R}^m$, we call the point in $\Omega$ that is closest to $x$ the projection of $x$ on $\Omega$ and denote it by $P_{\Omega} (x)$. $P_{\Omega} (x)$ contains only one element for any $x \in \mathbb{R}^m,$
and it satisfies the following non-expansive property \cite[Theorem 2.13]{opt}
\begin{equation}\label{pro}
\| P_{\Omega} (x)-P_{\Omega} (y) \| \leq \| x-y\| ~~ \forall x, y\in\mathbb{R}^m.
\end{equation}
Consider a convex closed set $\Omega \subset \mathbb{R}^m$ and a point $x \in \Omega$. Define
the normal cone to $\Omega$ at $x$ as $N_{\Omega}(x) \triangleq(v \in \mathbb{R}^m: \langle v, y-x \rangle \leq 0 ~~\forall y \in \Omega\}$. It is shown that \cite[Lemma 2.38]{opt}
\begin{equation}\label{normalcone}
N_{\Omega}(x) =\{ v \in \mathbb{R}^m: P_{\Omega}(x+v)=x)~~\forall x \in \Omega.
\end{equation}
A set $C$ is affine if it contains the lines that pass through any
pairs of points $x,y \in C$ with $x \ne y$. Let $\Omega \subset \mathbb{R}^m$ be a nonempty convex set.
We say that $x \in \mathbb{R}^m$ is a relative interior point of $\Omega $ if
$x \in \Omega$ and there exists an open sphere $S$ centered at $x$ such that
$S \cap \textrm{aff}(\Omega) \subset \Omega,$
where $\textrm{aff}(\Omega)$ is the intersection of all affine sets containing $\Omega$.
A pair of vectors $ x^{*} \in \Omega $ and $ z^{*} \in \Psi$ is called a saddle point
of the function $ \Phi(x, z)$ in $ \Omega \times \Psi$ if \begin{equation}
\Phi(x^{*}, z) \leq \Phi(x^{*}, z^{*}) \leq \Phi(x , z^{*}) ~~ \forall x \in \Omega,~ ~ \forall z \in \Psi. \nonumber
\end{equation}
These definitions can be found in \cite{Bertsekas}.
\subsection{Problem Statement}
Consider a network of $n$ agents. The objective of the network is to solve the following constrained optimization problem
\begin{equation}\label{problem1}
\begin{split}
& \textrm{minimize} ~~ f(x)=\sum_{i=1}^n f_i(x) ,\\
& \textrm{subject to} ~~ x \in \Omega_o= \bigcap_{i=1}^n \Omega_i,
\end{split}
\end{equation}
where $f_i(x): \mathbb{R}^m \rightarrow \mathbb{R}$ is the local
cost function of agent $i$, and $\Omega_i \subset \mathbb{R}^m$
is the local constraint set of agent $i$. Assume that $f_i(\cdot)$ is a smooth convex function on $\Omega_i$,
and $\Omega_i$ is a closed convex set only known to agent $i$.
Assume there exists at least one finite solution $x^*$ to the problem \eqref{problem1}.
For the problem \eqref{problem1}, denote by $f^{*}= \min_{ x \in \Omega_o} f(x)$ the optimal value,
and by $ \Omega_{o}^{*}=\{ x \in \Omega_o: f(x)=f^{*}) $ the optimal solution set.
Further, assume that for each $i \in \mathcal{V}$, the values of $f_i(\cdot)$ and $\nabla f_i(\cdot)$ are observed with noises. For example, $f_i(x)=E[h_i(x,\vartheta_i)],$ where $h_i:\mathbb{R}^m \times \Theta_i \rightarrow \mathbb{R}$ with $\vartheta_i$ being a random variable defined on $\Theta_i$, and the expectation $E[\cdot]$ is taken with respect to $\vartheta_i.$
In this case, one may only observe $\nabla h_i(x_i, \vartheta_i)$ for some given samples of $\vartheta_i$,
while the exact gradient $\nabla f_i(x_i)$ is difficult to calculate.
Let the communication relationship among agents at time $k$ be described by a directed graph
$\mathcal{G}_k =\{ \mathcal{V}, \mathcal{E}_{\mathcal{G}_k}, \mathcal{A}_{\mathcal{G}_k}\}$, where
$\mathcal{V }=\{ 1,\cdots,n\}$ is the node set, $\mathcal{E}_{\mathcal{G}_k}$ is the edge set, and $\mathcal{A}_{\mathcal{G}_k} =[ a_{ij,k} ]_{i,j=1}^n $ is the adjacency matrix.
Denote by $\mathcal{L}_k=[ l_{ij,k} ]_{i,j=1}^n$ the Laplacian matrix of digraph $ \mathcal{G}_k$.
Denote by $\mathcal{N}_{i,k}=\{j \in \mathcal{V}:(j,i) \in \mathcal{E}_{\mathcal{G}_k}\}$ the neighbors of
agent $i$ at time $k.$ Besides, neighboring agents exchange information through channels which may contain noises.
The noises may be introduced by quantization errors \cite{Xie,Rabbat}, or actively introduced to achieve differential privacy \cite{DCU1}.
\section{Primal-Dual Algorithm } \label{sec:Algorithm}
We now propose a distributed primal-dual algorithm to solve the distributed stochastic optimization problem, and list some conditions and preliminary lemmas to be used in the sequel.
\subsection{Algorithm Design }
Denote by $x_{i,k} \in \mathbb{R}^m$ the estimate for the optimal solution to problem \eqref{problem1} given by agent $i$ at time $k$, and by $\lambda_{i,k} \in \mathbb{R}^m$ the auxiliary variable of agent $i$. Hereafter,
we call $x_{i,k}$ and $\lambda_{i,k}$ the primal and dual variables for agent $i$ at time $k$. Agents exchange information in the following way:
if $(j,i) \in \mathcal{E}_{\mathcal{G}_k}$, then agent $i$ gets the noisy observations $\{x_{ij,k}, \lambda_{ij,k}\}$ of $\{x_{j,k}, \lambda_{j,k}\}$ given as follows:
\begin{equation}
\label{ }
\begin{split}
& x_{ij,k} = x_{j,k}+ \omega_{ij,k} \textrm{~ if }(j,i) \in \mathcal{E}_{\mathcal{G}_k}, \\&
\lambda_{ij,k} = \lambda_{j,k}+ \zeta_{ij,k} \textrm{~ if }(j,i) \in \mathcal{E}_{\mathcal{G}_k},
\end{split}
\end{equation}
where $\omega_{ij,k}$ and $\zeta_{ij,k}$ denote the communication noises.
The sequences $\{ x_{i,k}\}$ and $\{ \lambda_{i,k}\}$ are updated as follows:
\begin{equation}\label{algorithm1}
\begin{split}
& x_{i , k+1 }= P_{\Omega_i} \big( x_{i,k}- \gamma_k g_{i,k}
- \gamma_k \sum_{j \in \mathcal{N}_{i,k}} a_{ij,k}(\lambda_{i,k} -\lambda_{ij,k}) - \gamma_k \sum_{j \in \mathcal{N}_{i,k}} a_{ij,k}(x_{i,k} -x_{ij,k}) \big), \\
& \lambda_{i ,k+1}=\lambda_{i,k}+\gamma_k\sum_{j\in \mathcal{N}_{i,k}} a_{ij,k} (x_{i,k}-x_{ij ,k}),
\end{split}
\end{equation}
where $\gamma_k$ is the step size and $g_{i,k}$ denotes the noisy observation of $ \nabla f_i(x_{i,k}) $:
\begin{equation}
\label{ }
g_{i,k}= \nabla f_i(x_{i,k}) + v_{i,k},
\end{equation}
where $v_{i,k}$ is the observation noise.
Note that the algorithm \eqref{algorithm1} is distributed as in an iteration each agent updates its local estimates only using the local gradient observations and the noisy observations for primal and dual variables of its neighbors.
Set $ X_k \triangleq col\{ x_{1,k} , \cdots, x_{n,k}\}, $ $ \Lambda_k \triangleq col\{ \lambda_{1 ,k} \cdots, \lambda_{n ,k}\},$
and $ \nabla \widetilde{f} (X_k) \triangleq col\{ \nabla f_1(x_{1,k}) ,\cdots, \nabla f_n(x_{n, k}) \},$
where by $ col\{x_1,\cdots, x_n\} $ we mean $ (x_1^T, \cdots, x_n^T)^T$. Define
$v_k \triangleq col\{ v_{1,k} ,\cdots, v_{n,k}\}$, $\omega_k \triangleq col\{ \omega_{1,k} ,\cdots, \omega_{n,k}\}$
with $\omega_{i,k} \triangleq \sum_{j=1}^n a_{ij,k}\omega_{ij,k}$,
and $\zeta_k \triangleq col\{ \zeta_{1,k},\cdots, \zeta_{n,k}\}$ with $\zeta_{i,k}\triangleq \sum_{j=1}^n a_{ij,k}\zeta_{ij,k}$.
Then the algorithm \eqref{algorithm1} can be rewritten in the compact form as follows
\begin{equation}\label{compact0}
\begin{split}
& X_{k+1}= P_{\Omega} \big( X_k - \gamma_k \nabla \widetilde{f}(X_{k}) -\gamma_k( \mathcal{ L} _k \otimes\mathbf{I} _m) \big( \Lambda_k + X_k \big) + \gamma_k \big( \zeta_k +\omega_k-v_k \big) \big), \\
&\Lambda_{ k+1}=\Lambda_k+\gamma_k ( \mathcal{L}_k \otimes\mathbf{I} _m) X_k- \gamma_k \omega_k ,
\end{split}
\end{equation}
where
$\Omega=\prod_{i=1}^n \Omega_i$ denotes the Cartesian product, the symbol
$\otimes $ denotes the Kronecker product, and $\mathbf{I}_m$ denotes the identify matrix of size $m$.
\subsection{Assumptions}
We impose the following assumptions on the constraint sets and on the cost functions.
\begin{ass}\label{ass-set}
a) $ \Omega_o$ has at least one relative interior point.
b) There exists a constant $L_f>0$ such that for any $i \in \mathcal{V}$
\begin{equation}\label{Lipschitz}
\| \ \nabla f_i(x) - \nabla f_i(y)\| \leq L_f \| x-y \|~~ \forall x,y \in \Omega_i.
\end{equation}
c) For any $i \in \mathcal{V}, $ the set $\Omega_i$ is determined by $p_i$ inequalities:
$$ \Omega_i=\{ x\in \mathbb{R}^m: q_{ij} (x) \leq 0, ~ \forall j=1,\cdots, p_i\},$$
where $q_{ij}(\cdot),~j=1,\cdots, p_i$ are continuously differentiable real-valued functions on
$\mathbb{R}^m$.
Moreover, $\{ \nabla q_{ij}(x), ~ j \in A_i(x)\}$ are linearly independent, where $A_i(x)=\{ j: q_{ij}(x)=0\}$.
\end{ass}
\begin{rem}
The existence of the relative interior point will be used to guarantee that the primal and dual problems defined in Section \ref{sec:Algorithm}. C have the same optimal solution. The globally Lipschitz condition is used to guarantee the boundedness of the estimates. Assumption \ref{ass-set}-c indicates that
all local constraint sets have smooth boundaries. In fact, Assumption \ref{ass-set}-c corresponds to A4.3.2 in \cite{Kushner} but without compactness requirement. \end{rem}
The following conditions are imposed on the communication graphs and on the adjacency matrices.
\begin{ass} \label{ass-graph}(Mean graph is connected and undirected)
a) $(\mathcal{A}_{\mathcal{G}_k} )_{ k \geq 0}$
is an i.i.d sequence with expectation denoted by $\mathcal{\bar{A}} = E [ \mathcal{A}_{\mathcal{G}_k} ]$.
b) The graph $\mathcal{G}_{ \mathcal{\bar{A}}}$ generated by $\mathcal{\bar{A}} $ is undirected and connected.
c) There exists a constant $\eta>0$ such that
$$ E [ a_{ij,k} ^2] =\sigma_{ij}\leq \eta^2 ~~ \forall i,j \in \mathcal{V}.$$
d) $\mathcal{L}_k$ is independent of $\mathcal{F}_{k-1}$, where
\begin{equation}\label{def-algebra}
\mathcal{F}_{k} = \sigma\{X_0, \Lambda_0, \omega_{ij,t}, \zeta_{ij,t}, v_{i,t},
\mathcal{L}_t, 0 \leq t \leq k , 1\leq i,j \leq n\}.
\end{equation}
\end{ass}
\begin{rem}
Note that Assumption \ref{ass-graph} does not require the random graph
at any instance be undirected or strongly connected. It only requires the mean graph be
undirected and connected. The gossip-based communication protocol \cite{gossip}
and the broadcast-based communication \cite{broadcast}
both satisfy Assumption \ref{ass-graph} when the underling graph is bidirectional and strongly connected.
\end{rem}
Set
\begin{equation}\label{def-algebras}
\mathcal{F}_{k}'= \sigma\{ \mathcal{L}_{k+1 } , \mathcal{F}_{k}\}.
\end{equation}
Note that the adjacency matrix $\mathcal{A}_{\mathcal{G}_k} $ is uniquely defined by $ \mathcal{L}_{k } $ with
$a_{ij,k}=-l_{ij,k}~\forall i\neq j $ and $ a_{ii,k}=0$. Thus, the covariance of $ \mathcal{L}_{k } $
is finite by Assumption \ref{ass-graph}-c, $\mathcal{A}_{\mathcal{G}_k} $ is independent of $\mathcal{F}_{k-1}$ by Assumption \ref{ass-graph}-d, and $ \mathcal{L}_{k } $ is adapted to $\mathcal{F}_{k-1}' $ by its definition \eqref{def-algebras}.
The following conditions are imposed on the communication noises and gradient errors.
\begin{ass}\label{ass-noise2}
a) For any $i,j\in \mathcal{V},$ $\{\omega_{ij,k} ,\mathcal{F}_{k}'\}$ is an mds with
\begin{align}
& E[\omega_{ij,k} |\mathcal{F}_{k-1}']=\mathbf{0},~~ E[\| \omega_{ij,k} \|^{2 } |\mathcal{F}_{k-1}'] \leq \mu^2,\nonumber
\end{align}
and
\begin{align}
& E[\omega_{ij,k} \omega_{ij,k} ^T |\mathcal{F}_{k-1}'] \triangleq R_{\omega,ij} .\label{covW} \end{align}
b) For any $i,j\in \mathcal{V},$ $\{\zeta_{ij,k} ,\mathcal{F}_{k}'\}$ is an mds with
$$ E[\zeta_{ij,k} |\mathcal{F}_{k-1}']=\mathbf{0}~~E[\| \zeta_{ij,k} \|^{2 } |\mathcal{F}_{k-1}'] \leq \mu^2,$$
and
\begin{align} & E[ \zeta_{ij,k} \zeta_{ij,k} ^T |\mathcal{F}_{k-1}'] \triangleq R_{\zeta,ij}. \label{covZ}
\end{align}
c) For any $i\in \mathcal{V}$, $\{v_{i,k},\mathcal{F}_{k}'\}$ is an mds with
\begin{align}
& E[v_{i,k} | \mathcal{F}_{k-1}' ]=\mathbf{0}, ~~E[ \| v_{i,k} \|^{2 } | \mathcal{F}_{k-1}' ] \leq c_v(1+\|x_{i,k}\|^{2 }), \label{gradientnoise}
\\
& \lim_{k \rightarrow \infty} E[v_{i,k} v_{i,k} ^T |\mathcal{F}_{k-1}'] \triangleq R_{v,i}. \label{covV} \end{align}
\end{ass}
In Section \ref{sec:Convergence}, \eqref{covW}, \eqref{covZ} and \eqref{covV} are not needed.
The simplified version of Assumption \ref{ass-noise2} with \eqref{covW}, \eqref{covZ} and \eqref{covV} removed will be called Assumption \ref{ass-noise}.
\begin{ass} \label{ass-noise}
a)
Assumption \ref{ass-noise2}-a with \eqref{covW} removed.
b)
Assumption \ref{ass-noise2}-b with \eqref{covZ} removed.
c)
Assumption \ref{ass-noise2}-c with \eqref{covV} removed.
\end{ass}
\begin{rem}
The communication noises introduced by the probabilistic quantization \cite{Xie,Rabbat} is shown to be an i.i.d sequence with bounded second moments, and hence
satisfy Assumption \ref{ass-noise}-a and \ref{ass-noise}-b.
Assumption \ref{ass-noise2}-c holds true in many cases, for example, in the quadratic distributed stochastic optimization problem \eqref{filter1} discussed in Section \ref{sec:Simulation}.
\end{rem}
By Assumption \ref{ass-graph}-a, $\{\mathcal{L}_k \}_{ k \geq 0}$ is an i.i.d sequence. Set $\mathcal{\bar{L}}\triangleq E [\mathcal{L}_k ]$. Then $\mathcal{\bar{L}} $ is the Laplacian matrix of the undirected connected graph $\mathcal{G}_{ \mathcal{\bar{A}}}$.
Define
\begin{align}
& e_{1,k}\triangleq \big( ( \mathcal{\bar{L}}-\mathcal{ L}_k ) \otimes\mathbf{I} _m \big) ( \Lambda_k + X_k ), \label{def-e1}
\\&e_{2,k}\triangleq \zeta_k +\omega_k-v_k,\label{def-e2}
\\&e_{3,k}\triangleq \big(( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes\mathbf{I} _m \big) X_k- \omega_k. \label{def-e3}
\end{align}
Then \eqref{compact0} can be rewritten as:
\begin{equation}\label{compact}
\begin{split}
& X_{k+1}= P_{\Omega} \big( X_k - \gamma_k \nabla \widetilde{f}(X_{k}) -\gamma_k ( \mathcal{\bar{L}} \otimes\mathbf{I} _m) \big( \Lambda_k + X_k \big) + \gamma_k \big(e_{1,k}+e_{2,k}\big) \big), \\
&\Lambda_{k+1}=\Lambda_k+\gamma_k ( \mathcal{\bar{L}} \otimes\mathbf{I} _m) X_k+ \gamma_k e_{3,k} .
\end{split}
\end{equation}
We impose the following condition on the step size $\{\gamma_k\}$.
\begin{ass}\label{ass-stepsize}
$$\gamma_k >0, ~ \sum_{k=1}^{\infty} \gamma_k =\infty,\textrm{ and }\sum_{k=1}^{\infty} \gamma_k^2 < \infty .$$
\end{ass}
\subsection{Preliminary Lemmas}
We now give some preliminary results about the formulated distributed optimization problem.
\begin{lem} \label{lem2}
The problem \eqref{problem1} is equivalent to the following constrained optimization problem
\begin{equation}\label{problem2}
\begin{split}
& \textrm{minimize} ~~\widetilde{f}(X)\stackrel{\Delta}{=} \sum_{i=1}^n f_i(x_i) ,\\
& \textrm{subject to} ~~ ( \mathcal{\bar{L}} \otimes\mathbf{I} _m ) X=\mathbf{0}, ~~ X \in \Omega ,
\end{split}
\end{equation}
where $X= col\{x_1,\cdots, x_n\} $.\end{lem}
The result can be easily derived since $( \mathcal{\bar{L}} \otimes\mathbf{I} _m ) X=\mathbf{0}$
if and only if $x_i=x_j~\forall i,j\in \mathcal{V}.$
Define $ \Phi(X,\Lambda)\triangleq \widetilde{f}(X)+\Lambda^T ( \mathcal{\bar{L}} \otimes\mathbf{I} _m ) X$ as the Lagrange function, where $\Lambda \in \mathbb{R}^{mn}$ is the Lagrange multiplier. Then the problem \eqref{problem2} can be rewritten as $ \inf\limits_{X \in \Omega} \sup\limits_{\Lambda \in \mathbb{R}^{ mn} } \Phi(X, \Lambda)$, while the dual problem is defined as follows
\begin{equation}\label{pdual}
\sup\limits_{\Lambda \in \mathbb{R}^{ mn} } \inf\limits_{X \in \Omega} \Phi(X, \Lambda).
\end{equation}
\begin{lem} \label{lem3} Assume Assumption \ref{ass-set}-a and Assumption \ref{ass-graph}-b hold. Then $ \Phi(X, \Lambda)$ has at least one saddle point in $ \Omega \times \mathbb{R}^{mn}$.
A pair $(X^{*}, \Lambda^{*}) \in \Omega \times \mathbb{R}^{mn}$ is
the primal-dual solution to the problems \eqref{problem2} and \eqref{pdual}
if and only if $(X^{*}, \Lambda^{*}) $ is a saddle point
of $ \Phi(X, \Lambda)$ in $ \Omega \times \mathbb{R}^{mn}$.
\end{lem}
{\bf Proof}:
Assumption \ref{ass-set}-a implies that there exists a relative interior $\bar{X} $ of set
$\Omega$ such that $(\mathcal{L} \otimes \mathbf{I}_m) \bar{X}=0$. Since $f^{*}$ is finite, by \cite[Proposition 5.3.3]{Bertsekas} we know that
\begin{equation}\label{minmax}
\inf_{X \in \Omega} \sup_{\Lambda \in \mathbb{R}^{ mn} } \Phi(X, \Lambda)=\sup_{\Lambda \in \mathbb{R}^{ mn} } \inf_{X \in \Omega} \Phi(X, \Lambda),
\end{equation}and there exists at least one dual optimal solution.
Since the minimax equality \eqref{minmax} holds, by \cite[Proposition 3.4.1]{Bertsekas}
$X^{*}$ is the primal optimal solution and $\Lambda^{*}$ is the dual optimal solution if and only if
$(X^{*}, \Lambda^{*})$ is a saddle point of $ \Phi(X,\Lambda)$ on $\Omega \times \mathbb{R}^{mn}$.
Since there exists at least one primal and dual optimal solution pair, we conclude that $ \Phi(X, \Lambda)$ has at least one saddle point in $ \Omega \times \mathbb{R}^{mn}$.
This completes the proof. \hfill $\blacksquare$
\section{Convergence Theorems} \label{sec:Convergence}
In this section, we analyze stability and convergence of the algorithm \eqref{algorithm1}.
For notational simplicity, we assume $m=1$ in this section. This does not influence the convergence analysis for the general case $m \geq 1$.
\subsection{Stability Analysis}
\begin{thm}\label{thm2} (Stability) Let $\{ x_{i,k}\}$ and $\{ \lambda_{i,k}\}$ be produced by the algorithm \eqref{algorithm1} with any initial values $x_{i,0},~ \lambda_{i,0}$.
Let Assumptions \ref{ass-set}-a, \ref{ass-set}-b, \ref{ass-graph}, \ref{ass-noise}, and \ref{ass-stepsize} hold. Then $ \| X_k -X^{*}\|^2+ \| \Lambda_k-\Lambda^{*} \|^2 $ converges a.s.,
where $(X^{*}, \Lambda^{*})$ is a saddle point of $ \Phi(X, \Lambda)$ in $ \Omega \times \mathbb{R}^{mn}$.
\end{thm}
This theorem establishes that the sequences $\{ X_k\}$ and $\{\Lambda_k\}$ are bounded a.s.,
and the distance between the pair $(X_k, \Lambda_k)$ and the saddle point $(X^{*}, \Lambda^{*})$ converges a.s.
Before proving the theorem, we first give some preliminary lemmas.
The following lemma establishes properties of noise sequences $\{e_{1,k})$, $\{e_{2,k}\}$ and $\{e_{3,k}\}$
defined in \eqref{def-e1}, \eqref{def-e2}, and \eqref{def-e3}, respectively.
\begin{lem} \label{lemma1}
Let Assumptions \ref{ass-graph}-a, \ref{ass-graph}-c, \ref{ass-graph}-d, and
\ref{ass-noise} hold. Then the following assertions take place a.s. \begin{align}
&E[ e_{1,k} | \mathcal{F}_{k-1}]=\mathbf{0},~~E[ \| e_{1,k}\|^2 | \mathcal{F}_{k-1}]
\leq C_{01} \big\| \Lambda_k+ X_k \big\|^2,\label{noise1} \\&
E[ e_{2,k} | \mathcal{F}_{k-1}]=\mathbf{0}, ~~E[ \| e_{2,k} \|^2 | \mathcal{F}_{k-1} ] =C_{02} +3c_v \|X_k\|^2,\label{noise2} \\&
E[ e_{3,k} | \mathcal{F}_{k-1}]=\mathbf{0}, ~~E [ \| e_{3,k} \|^2 | \mathcal{F}_{k-1}]\leq C_{01} \| X_k \|^2 + C_{03}, \label{noise3}
\end{align}
where $C_{01}=E [\| \mathcal{L}_k -\mathcal{\bar{L}}\|^2]$, $C_{02}=3c_v n+ 6n^3 \mu^2 \eta^2$,
and $C_{03}=n^3\mu^2 \eta^2$.
\end{lem}
\textbf{Proof:} By Assumption \ref{ass-graph}-d we have
\begin{equation}\label{conditional-Laplacian}
E \big[ \mathcal{\bar{L}}-\mathcal{ L} _k
| \mathcal{F}_{k-1} \big] = \mathcal{\bar{L}}- E [ \mathcal{ L}_k ]= \mathbf{0}.
\end{equation}
Since $X_k $ and $\Lambda_k$ are adapted to $\mathcal{F}_{k-1} $
by \eqref{algorithm1} \eqref{def-algebra}, from \eqref{def-e1} \eqref{conditional-Laplacian} it follows that
\begin{align}
&E[ e_{1,k} | \mathcal{F}_{k-1}]= E \big[ \mathcal{\bar{L}}-\mathcal{ L} _k
| \mathcal{F}_{k-1} \big] ( \Lambda_k+ X_k )=\mathbf{0}, \nonumber \end{align}
and
\begin{equation}
\begin{split}
E[ \| e_{1,k}\|^2 | \mathcal{F}_{k-1}] & \leq E \big[ \| \mathcal{\bar{L}}-\mathcal{ L} _k \|^2
| \mathcal{F}_{k-1} \big] \cdot \| \Lambda_k+ X_k \|^2\\&
=E [\| \mathcal{L}_k -\mathcal{\bar{L}}\|^2] \cdot \| \Lambda_k+ X_k \|^2. \nonumber
\end{split}\end{equation}
Therefore, \eqref{noise1} holds.
Since $a_{ij,k}$ is adapted to $\mathcal{F}_{k-1}'$ by \eqref{def-algebras},
from Assumption \ref{ass-noise}-a it follows that for any $i\in \mathcal{V}$
\begin{equation} \label{equ0}
\begin{array}{lll}
& E[\omega_{i,k} | \mathcal{F}_{k-1}'] = \sum_{j=1}^n a_{ij,k} E [ \omega_{ij,k} | \mathcal{F}_{k-1}'] =\mathbf{0}.
\end{array} \end{equation} Similarly, by Assumption \ref{ass-noise}-b it is shown that
\begin{align}\label{estimate0}
& E[\zeta_{i,k} | \mathcal{F}_{k-1}'] = \mathbf{0} ~ ~\forall i\in \mathcal{V}. \end{align}
Then from \eqref{equ0} \eqref{estimate0} and Assumption \ref{ass-noise}-c, by \eqref{def-e2} we derive
\begin{align}
& E[ e_{2,k} | \mathcal{F}_{k-1}']=E[ \omega_k | \mathcal{F}_{k-1}']+E[ \zeta_k | \mathcal{F}_{k-1}']+E[v_k | \mathcal{F}_{k-1}']= \mathbf{0}.
\label{equ3}
\end{align}
Since $\mathcal{F}_{k-1} \subset \mathcal{F}_{k-1}^{'}$, by \eqref{equ0} \eqref{equ3} we see
\begin{align}
& E[ \omega_{k} | \mathcal{F}_{k-1}]= E \big[ E[ \omega_{k} | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big]= \mathbf{0}~a.s., \label{equ4}\\&E[ e_{2,k} | \mathcal{F}_{k-1}]= E \big[ E[ e_{2,k} | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big]= \mathbf{0}~a.s. \nonumber
\end{align}
Since $a_{ij,k}$ is adapted to $\mathcal{F}_{k-1}'$ by \eqref{def-algebras}, from Assumption \ref{ass-noise}-a it follows that
\begin{equation}
\begin{split}
&E[ \| a_{ij,k} \omega_{ij,k} \|^2 | \mathcal{F}_{k-1}'] \leq a_{ij,k}^2 E[ \| \omega_{ij,k} \|^2 | \mathcal{F}_{k-1}'] \leq a_{ij,k}^2 \mu^2. \nonumber
\end{split}
\end{equation}
Since $a_{ij,k}$ is
independent of $\mathcal{F}_{k-1} $ by Assumption \ref{ass-graph}-d, from $\mathcal{F}_{k-1} \subset \mathcal{F}_{k-1}^{'}$ by Assumption \ref{ass-graph}-c we obtain
\begin{equation}
\begin{split}
E[ \| a_{ij,k} \omega_{ij,k} \|^2 | \mathcal{F}_{k-1}] &= E \big[ E[ \| a_{ij,k} \omega_{ij,k} \|^2 | \mathcal{F}_{k-1}'] \big| \mathcal{F}_{k-1} \big]
\\& \leq \mu^2 E [ a_{ij,k} ^2 ] \leq \mu^2 \eta^2 ~~ \forall i \in \mathcal{V}~a.s.\nonumber
\end{split}
\end{equation}
Then by the conditional Minkowski inequality $
\big(E[\| \sum_{i=1}^kX_i \|^2 \big | \mathcal{F}] \big)^{\frac{1}{2}} \leq \sum_{i=1}^k \big(E[\| X_i \|^2 \big | \mathcal{F}] \big)^{\frac{1}{2}} $, and by $\omega_{i,k}= \sum_{j=1}^n a_{ij,k} \omega_{ij,k} $ we derive
\begin{equation}
\begin{split}
\big(E[ \| \omega_{i,k} \|^2 \big | \mathcal{F}_{k-1} ] \big)^{\frac{1}{2}} & \leq \sum_{j=1}^n \big( E[ \| a_{ij,k} \omega_{ij,k} \|^2 | \mathcal{F}_{k-1}] \big)^{\frac{1}{2}} \leq n \mu \eta ~~ \forall i \in \mathcal{V}~a.s. \nonumber
\end{split}
\end{equation}
Similarly, by Assumption \ref{ass-noise}-b we derive
$$ \big( E[\| \zeta_{i,k} \|^2 | \mathcal{F}_{k-1} ] \big)^{\frac{1}{2}} \leq n \mu \eta~~ \forall i \in \mathcal{V}~~a.s.$$
Then by the definitions of $\omega_k $ and $\zeta_k$ we conclude that
\begin{equation}\label{n1}
\begin{split}
&E [\| \omega_k \|^2 | \mathcal{F}_{k-1} ] = \sum_{i=1}^n E [\| \omega_{i,k} \|^2 | \mathcal{F}_{k-1} ] \leq n^3\mu^2 \eta^2~a.s. ,
\\&E [\| \zeta_k \|^2 | \mathcal{F}_{k-1}] =\sum_{i=1}^n E [\| \zeta_{i,k} \|^2 | \mathcal{F}_{k-1} ]\leq n^3\mu^2 \eta^2~a.s.
\end{split}
\end{equation}
By \eqref{gradientnoise} we have
\begin{equation}
E [\| v_k \|^2 | \mathcal{F}_{k-1}']= \sum_{i=1}^n E [ \| v_{i,k} \|^2 | \mathcal{F}_{k-1}'] \leq c_v(n+ \|X_k\|^2) .\nonumber
\end{equation}
Then by noticing that $\mathcal{F}_{k-1} \subset \mathcal{F}_{k-1}^{'}$ and $X_k$ is adapted to $\mathcal{F}_{k-1}$ we have
\begin{equation} \label{61}
\begin{split}
E [\| v_k \|^2 | \mathcal{F}_{k-1}]= E\big[ E [\| v_k \|^2 | \mathcal{F}_{k-1}'] \big|\mathcal{F}_{k-1}\big] \leq c_v(n+ \|X_k\|^2) ~~a.s. \end{split}
\end{equation}
Thus, by \eqref{def-e2} from \eqref{n1} \eqref{61} we obtain
\begin{equation}
\begin{split}
E[ \| e_{2,k} \|^2 | \mathcal{F}_{k-1} ] &= 3( E [\| \omega_k \|^2 | \mathcal{F}_{k-1} ]+E [\| \zeta_k\|^2 | \mathcal{F}_{k-1} ]+E [\| v_k \|^2 | \mathcal{F}_{k-1} ] ) \\ &\leq 6n^3 \mu^2 \eta^2 +3c_v(n+ \|X_k\|^2) ~a.s. \nonumber
\end{split}
\end{equation}
Hence \eqref{noise2} holds.
We now consider properties of the noise sequence $\{e_{3,k})$ defined in \eqref{def-e3}.
Since $X_k $ is adapted to $\mathcal{F}_{k-1} $, by \eqref{conditional-Laplacian} \eqref{equ4}
we have
\begin{align}
E[ e_{3,k}| \mathcal{F}_{k-1}] = E [ \mathcal{\bar{L}}-\mathcal{ L}_k
| \mathcal{F}_{k-1} ] X_k -E[\omega_k|\mathcal{F}_{k-1}]=\mathbf{0}~a.s. \nonumber
\end{align}
Since $X_k , \mathcal{L}_k$ are adapted to $ \mathcal{F}_{k-1}'$
and $\mathcal{F}_{k-1} \subset\mathcal{F}_{k-1}'$, by \eqref{equ0} we derive
\begin{equation}\label{01}
\begin{split}
& E [ \omega_k^T ( \mathcal{L}_k -\mathcal{\bar{L}}) X_k | \mathcal{F}_{k-1}] =
E \Big [ E \big [ \omega_k^T | \mathcal{F}_{k-1}' \big] ( \mathcal{L}_k -\mathcal{\bar{L}}) X_k \big | \mathcal{F}_{k-1}\Big] =0~a.s.\end{split}
\end{equation}
Hence by \eqref{n1} and Assumption \ref{ass-graph}-d we conclude that
\begin{equation}\label{01}
\begin{split}
E [ \| e_{3,k} \|^2 | \mathcal{F}_{k-1}]&= E[ \|( \mathcal{L}_k -\mathcal{\bar{L}}) X_k \|^2 |\mathcal{F}_{k-1}] + E [\| \omega_k \|^2 | \mathcal{F}_{k-1} ]
+ 2 E [ \omega_k^T ( \mathcal{L}_k -\mathcal{\bar{L}}) X_k | \mathcal{F}_{k-1}] \\& \leq E [\| \mathcal{L}_k -\mathcal{\bar{L}}\|^2] \| X_k \|^2 + n^3\mu^2 \eta^2~a.s.\end{split}
\end{equation}
Therefore, \eqref{noise3} holds.
\hfill $\blacksquare$
\begin{lem} \label{lemma2}
Let Assumptions \ref{ass-graph}-a, \ref{ass-graph}-c, \ref{ass-graph}-d and \ref{ass-noise} hold. Then for any $X \in \Omega$ and $\Lambda \in \mathbb{R}^{mn}$
\begin{equation} \label{result1}
\begin{split}
E [ \| X_{k+1}- X \| ^2 | \mathcal{F}_{k-1} ] & \leq \| X_k-X \|^2 +\gamma_k ^2 \| \nabla \widetilde{f}(X_{k}) +\mathcal{\bar{L}} \big( \Lambda_k + X_k\big) \|^2 \\
& +2 \gamma_k \big(\Phi(X,\Lambda_k) - \Phi(X_k, \Lambda_k) \big)
-2\gamma_k \big ( X_k-X \big)^T \mathcal{\bar{L}} X_k \\&
+ C_{01}\gamma_k^2 \| \Lambda(k) + X_k\|^2 + 3c_v \gamma_k^2\| X_k\|^2+ C_{02}\gamma_k^2 ~~a.s.,
\end{split}
\end{equation}
and
\begin{equation} \label{result2}
\begin{split}
E [ \| \Lambda_{k+1}- \Lambda \| ^2 | \mathcal{F}_{k-1}] &\leq \| \Lambda_{k} -\Lambda \|^2 +\gamma_k ^2\| \mathcal{\bar{L}}X_k \|^2 + C_{03}\gamma_k ^2 \\&+2\gamma_k \big( \Phi(X_k,\Lambda_k) -\Phi(X_k, \Lambda )\big) + C_{01} \gamma_k ^2 \| X_k \|^2~~a.s.
\end{split}
\end{equation}
\end{lem}
{\bf Proof:}
By using the non-expansive property \eqref{pro} of the projection operator, from \eqref{compact} we obtain
\begin{equation}\label{inequ1}
\begin{split}
\| X_{k+1}- X \| ^2 & \leq \big \| X_k - \gamma_k \nabla \widetilde{f} (X_k) -\gamma_k\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) -X + \gamma_k \big(e_{1,k}+e_{2,k}\big) \big \|^2 \\
& \leq I_0(k) + \gamma_k ^2 I_1(k) + 2 \gamma_k I_2(k) ~~ \forall X \in \Omega,
\end{split}
\end{equation}
where $I_0(k)= \| X_k - \gamma_k \nabla \widetilde{f} (X_k) -\gamma_k\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) -X \|^2$,
$I_1(k)= \| e_{1,k}+e_{2,k} \|^2$, $I_2(k)=\big(e_{1,k}+e_{2,k}\big)^T
\Big( X_k - \gamma_k \nabla \widetilde{f} (X_k) -\gamma_k\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) -X\Big) $.
Since $e_{1,k}$ is adapted to $\mathcal{F}_{k-1}'$ by \eqref{def-algebras} \eqref{def-e1},
by $\mathcal{F}_{k} \subset \mathcal{F}_{k}'$ and \eqref{equ3} we see that
\begin{equation}\label{equ5}
\begin{split}& E [ e_{1,k}^T e_{2,k}| \mathcal{F}_{k-1}]= E \big[ E [ e_{1,k}^T e_{2,k} | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big ] = E \big[ e_{1,k}^T E [ e_{2,k} | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big ]=0~a.s.
\end{split}
\end{equation}
Thus, from here by \eqref{noise1} and \eqref{noise2} we derive
\begin{equation}\label{compon1}
\begin{split}
E [I_1(k) |\mathcal{F}_{k-1}] &= E [ \| e_{1,k}\| ^2 |\mathcal{F}_{k-1}]+E [ \| e_{2,k}\| ^2 |\mathcal{F}_{k-1}] +
2E [e_{1,k} ^T e_{2,k} |\mathcal{F}_{k-1}]
\\& \leq C_{01} \| \Lambda_k + X_k\|^2 + C_{02} + 3c_v \| X_k\|^2~a.s.
\end{split}
\end{equation}
Since $X_k,\Lambda_k$ are adapted to $\mathcal{F}_{k-1}$, by \eqref{noise1} \eqref{noise2} we derive
\begin{equation}\label{compon2}
\begin{split}
&E [I_2(k) |\mathcal{F}_{k-1}] = E \big[ e_{1,k}+e_{1,k} |\mathcal{F}_{k-1} \big]^T \big( X_k - \gamma_k \nabla \widetilde{f}(X_k) -\gamma_k\mathcal{\bar{L}} ( \Lambda_k + X_k ) -X\big)=0~a.s.
\end{split}
\end{equation}
Since $I_0(k)$ is adapted to $\mathcal{F}_{k-1}$, combining \eqref{inequ1}, \eqref{compon1}, \eqref{compon2} we obtain
\begin{equation} \label{bound01}
\begin{split}
E [ \| X_{k+1}- X \| ^2 | \mathcal{F}_{k-1}] & \leq I_0(k) +C_{01} \gamma_k ^2\| \Lambda_k + X_k\|^2
+3 c_v \gamma_k^2\| X_k\|^2+ C_{02} \gamma_k ^2 ~~a.s .
\end{split}
\end{equation}
Note that
\begin{equation}\label{bound11}
\begin{split}
I_0(k) & = \| X_k - \gamma_k \nabla \widetilde{f}(X_k)
-\gamma_k\mathcal{\bar{L}} \big( \Lambda_k+ X_k \big) -X \|^2\\
& \leq \| X_k-X \|^2+ \gamma_k ^2 \| \nabla \widetilde{f}(X_k) +\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) \|^2\\
& - 2 \gamma_k\big ( X_k-X \big)^T\Big( \nabla \widetilde{f}(X_k) +\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) \Big).
\end{split}
\end{equation}
Since $\Phi(X, \Lambda_k) $ is convex in $X\in \Omega$, by \eqref{gradient}
we derive
$$ \Phi(X,\Lambda_k) \geq \Phi(X_k, \Lambda_k)+ (X-X_k)^T \big ( \nabla \widetilde{f}(X_k)+ \mathcal{\bar{L}} \Lambda_k \big),$$
and hence
$$ -(X_k-X) ^T \big ( \nabla \widetilde{f}(X_k)+ \mathcal{\bar{L}} \Lambda_k\big) \leq \Phi(X,\Lambda_k) - \Phi(X_k, \Lambda_k).$$
Then by \eqref{bound11} we conclude that
\begin{equation}
\begin{split}
I_0(k) & \leq \| X_k-X \|^2+ \gamma_k ^2 \| \nabla \widetilde{f}(X_k) +\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) \|^2\\
& +2 \gamma_k \big(\Phi(X,\Lambda_k) - \Phi(X_k, \Lambda_k) \big)
-2\gamma_k \big ( X_k -X \big)^T \mathcal{\bar{L}} X_k , \nonumber
\end{split}
\end{equation}
which incorporating with \eqref{bound01} yields \eqref{result1}.
For any $\Lambda \in \mathbb{R}^n$
\begin{equation}\label{inequ2}
\begin{split}
\| \Lambda_{k+1}- \Lambda \| ^2& = \| \Lambda_k+\gamma_k \mathcal{ \bar{L}} X_k-\Lambda +\gamma_ke_{3,k} \|^2 \\ & = I_3(k) +\gamma_k ^2\| e_{3,k} \|^2 +2\gamma_k e_{3,k}^T\big(\Lambda_{k}+\gamma_k \mathcal{ \bar{L}} X_k -\Lambda\big),\end{split}
\end{equation}
where $I_3(k)=\| \Lambda_k+\gamma_k \mathcal{ \bar{L}} X_k -\Lambda \|^2 .$
Since $X_k$ and $\Lambda_k$ are adapted to $\mathcal{F}_{k-1}$,
from \eqref{noise3} we see
\begin{equation}
\begin{split}
& E[ e_{3,k}^T(\Lambda_{k}+\gamma_k \mathcal{ \bar{L}} X_k -\Lambda) | \mathcal{F}_{k-1}] =
E[ e_{3,k}^T | \mathcal{F}_{k-1}] (\Lambda_{k}+\gamma_k \mathcal{ \bar{L}} X_k -\Lambda)=0~a.s. \nonumber
\end{split}
\end{equation}
Noticing that $I_3(k)$ is adapted to $ \mathcal{F}_{k-1}$, from here by \eqref{noise3} \eqref{inequ2} we obtain
\begin{equation}\label{bound02}
\begin{split}
E[ \| \Lambda_{k+1}- \Lambda \| ^2 | \mathcal{F}_{k-1}] \leq I_3(k)
+ C_{01} \gamma_k ^2 \| X_k \|^2 + C_{03}\gamma_k ^2~~a.s .
\end{split}
\end{equation}
By the definition of $\Phi(X, \Lambda) $, we derive
$$ \Phi(X_k, \Lambda_k) =\Phi(X_k, \Lambda )+ (\Lambda_k-\Lambda )^T \mathcal{\bar{L}} X_k$$
and hence
\begin{equation}
\begin{split}
I_3(k) &=
\| \Lambda_{k} -\Lambda \|^2 +\| \gamma_k \mathcal{\bar{L}}X_k \|^2+2\gamma_k(\Lambda_k -\Lambda )^T \mathcal{\bar{L}}X_k \\& =\| \Lambda_{k} -\Lambda \|^2 +\| \gamma_k \mathcal{\bar{L}}X_k \|^2 +2\gamma_k \big( \Phi(X_k ,\Lambda_k) -\Phi(X_k , \Lambda )\big), \nonumber
\end{split}
\end{equation}
which incorporating with \eqref{bound02} yields \eqref{result2}.
\hfill $\blacksquare$
\textbf{Proof of Theorem \ref{thm2}:}
Summing up both sides of \eqref{result1} and \eqref{result2}, and by replacing $(X, \Lambda)$ with
$(X^{*}, \Lambda^{*})$ we obtain
\begin{equation} \label{sum1}
\begin{split}
& E [ \| X_{k+1}- X^{*} \| ^2 | \mathcal{F}_{k-1} ] +E [ \| \Lambda_{k+1}- \Lambda^{*} \| ^2 | \mathcal{F}_{k-1}]
\\ &\leq \| X_k -X^{*} \|^2+ \| \Lambda_{k} -\Lambda^{*} \|^2 + \gamma_k ^2 \| \nabla \widetilde{f}(X_k) +\mathcal{\bar{L}} \big( \Lambda_k + X_k \big) \|^2
\\& +2 \gamma_k \big(\Phi(X^{*},\Lambda_k) - \Phi(X_k, \Lambda^{*}) \big)
-2\gamma_k \big ( X_k-X^{*} \big)^T \mathcal{\bar{L}} X_k \\&
+ C_{01} \gamma_k^2\| \Lambda_k + X_k\|^2 +\gamma_k^2\| \mathcal{\bar{L}}X_k \|^2 +
\gamma_k ^2 ( C_{01}+3c_v) \| X_k \|^2 + ( C_{02}+ C_{03}) \gamma_k ^2 ~a.s.
\end{split}
\end{equation}
Since $(X^{*}, \Lambda^{*})$ is a saddle point for $\Phi(X,\Lambda)$, by Lemma \ref{lem3} $X^{*}$ is the optimal solution to the problem \eqref{problem2}. Then from Lemma \ref{lem2} it follows that
\begin{equation}\label{consensus}
\mathcal{\bar{L}} X^{*}=\mathbf{0},~\textrm{and }~\mathcal{\bar{L}} X_k= \mathcal{\bar{L}} ( X_k -X^{*}),
\end{equation}
and hence
\begin{equation}
\begin{split}
& \nabla \widetilde{f}(X_k) + \mathcal{\bar{L}} ( \Lambda_k+X_k ) = \nabla \widetilde{f}(X_k) - \nabla \widetilde{f}(X^{*}) +
\mathcal{\bar{L}} ( \Lambda_k-\Lambda^{*} )+ \mathcal{\bar{L}}( X_k-X^{*})
+ \mathcal{\bar{L}}\Lambda^{*} + \nabla \widetilde{f}(X^{*}).\nonumber
\end{split}
\end{equation}
Then by \eqref{Lipschitz} we obtain
\begin{equation} \label{b1}
\begin{split}
&\| \nabla \widetilde{f}(X_k) + \mathcal{\bar{L}} ( \Lambda_k+X_k ) \|^2 \\&\leq
4 \big( \| \nabla \widetilde{f}(X_k) - \nabla \widetilde{f} (X^{*})\| ^2+ \| \mathcal{\bar{L}} ( \Lambda_k -\Lambda^{*} ) || ^2 + \| \mathcal{\bar{L}} ( X_k-X^{*} ) || ^2 + \|\mathcal{\bar{L}}\Lambda^{*} +\nabla \widetilde{f}(X^{*}) \|^2 \big)
\\& \leq 4c_1 \| \Lambda_k -\Lambda^{*}\| ^2+ (4c_1+ 4L_f^2) \|X_k-X^{*}\|^2+ c_2,
\end{split}
\end{equation}
where $c_1= \| \mathcal{\bar{L} } \|^2$, $c_2= 4\|\mathcal{\bar{L}}\Lambda^{*} +\nabla \widetilde{f}(X^{*}) \|^2 $.
From \eqref{consensus} we derive
\begin{align}
& \|\mathcal{\bar{L}} X(k) \|^2 \leq c_1 \| X_k- X^{*} \|^2.\label{b2} \end{align}
Note that $ \| \Lambda_k+ X_k\|^2 \leq 3( \| \Lambda_k -\Lambda^{*}\|^2 +\| X_k -X^{*}\|^2+\| \Lambda^{*} + X^{*}\|^2)$ and $\| X_k \|^2 \leq 2( \| X_k-X^{*} \|^2+ \| X^{*} \|^2 ).$
Then by \eqref{sum1}, \eqref{b1} and \eqref{b2} we derive
\begin{equation} \label{sum2}
\begin{split}
& E [ \| X_{k+1}- X^{*} \| ^2 | \mathcal{F}_{k-1} ] +E [ \| \Lambda_{k+1}- \Lambda^{*} \| ^2 | \mathcal{F}_{k-1}]
\\ &\leq \big( 1+ (5 c_1+ 5 C_{01}+ 4 L_f^2+6c_v) \gamma_k^2 \big)\| X_k -X^{*} \|^2 \\&+ \big( 1+(4c_1+ 3C_{01} ) \gamma_k^2 \big)\| \Lambda_{k} -\Lambda^{*} \|^2 +\gamma_k ^2 \big( c_2+C_{02}+ C_{03} +3C_{01} \| \Lambda^{*} +X^{*}\|^2\\&+ 2(3 c_v+ C_{01} )\| X^{* }\|^2)
+2 \gamma_k \big(\Phi(X^{*},\Lambda_k ) - \Phi(X_k, \Lambda^{*}) \big) -2 \gamma_k \big ( X_k-X^{*} \big)^T \mathcal{\bar{L}} X_k ~a.s.
\end{split}
\end{equation}
Since $\mathcal{\bar{L} } $
is the Laplacian matrix of some connected undirected graph by Assumption \ref{ass-graph}-b,
from \eqref{consensus} and Lemma \ref{lem1} it follows that \begin{equation}\label{b5}
\begin{split}
& \big ( X_k-X^{*} \big)^T \mathcal{\bar{L}} X_k =\big ( X_k-X^{*} \big)^T \mathcal{\bar{L}} \big( X_k - X^{*}\big) \geq 0.
\end{split}
\end{equation}
Noticing $X_k \in \Omega$, by definition of the saddle point we see
\begin{equation}
\Phi(X^{*} , \Lambda_k) \leq \Phi(X^{*}, \Lambda^{*}) \leq \Phi(X_k , \Lambda^{*}) ~~ \forall k \geq 0. \nonumber
\end{equation}
Then by setting $V_{k}= \| X_k-X^{*} \|^2 + \| \Lambda_{k} -\Lambda^{*} \|^2$, from \eqref{sum2} \eqref{b5} we derive
\begin{equation} \begin{split}
& E\big [ V_{k+1} | \mathcal{F}_{k-1} \big ] \leq ( 1 + C_{11}\gamma_k^2 ) V_k+ C_{12}\gamma_k ^2~~a.s. ,\nonumber \end{split}
\end{equation}
where $C_{11}=5 c_1+ 5 C_{01}+ 4 L_f^2+ 6c_v, $ and $C_{12}=c_2+C_{02}+ C_{03} +3C_{01} \| \Lambda^{*} +X^{*}\|^2+ 2(3c_v+ C_{01} ) \| X^{* }\|^2 $.
Consequently, by Assumption \ref{ass-stepsize} and Lemma \ref{martingale} in Appendix we conclude that $\| X_{k}- X^{*} \| ^2 + \| \Lambda_{k}- \Lambda^{*} \| ^2$ converges a.s. \hfill $\blacksquare$
\subsection{Consensus and Consistency}
The following theorem shows that the estimates given by all agents reach a consensus belonging to the optimal solution set of problem \eqref{problem1}.
\begin{thm}\label{thm1} Let $\{ x_{i,k}\}$ and $\{ \lambda_{i,k}\}$ be produced by the algorithm \eqref{algorithm1}
with any initial values $x_{i,0},~ \lambda_{i,0}$.
Let Assumptions \ref{ass-set}, \ref{ass-graph}, \ref{ass-noise}, and \ref{ass-stepsize} hold. Then
i) (Consensus) $\lim\limits_{k \rightarrow \infty} (x_{i,k} -x_{j,k})=0~~\forall i ,j \in \mathcal{V}~~a.s.$
ii) (Consistency)
\begin{equation}\label{mainresults}
\lim_{k \rightarrow \infty} d(x_{i,k} , \Omega_o^{*})=0 ~~\forall i \in \mathcal{V} ~~a.s.,
\end{equation}
where $d(X,A)=\inf\limits_{\theta \in A} \| \theta-X\|. $
Moreover, if $f(\cdot)$ has a unique optimal solution $x^{*}$, then
\begin{equation}\label{mainresults2}
\lim_{k \rightarrow \infty} x_{i,k} =x^* ~~\forall i \in \mathcal{V} ~~a.s.
\end{equation}
\end{thm}
\textbf{Proof:} By setting
\begin{equation}\label{ }
\begin{split}
& \theta=\begin{pmatrix}
X \\
\Lambda
\end{pmatrix},~~e_k=\begin{pmatrix}
e_{1,k}+e_{2,k} \\
e_{3,k}
\end{pmatrix} , ~~\Phi= \Omega \times \mathbb{R}^{mn}, \\
& g(\theta)= \begin{pmatrix}
- \nabla \widetilde{f}(X) -\mathcal{\bar{L}}\big(X+\Lambda \big) \\
\mathcal{\bar{L}} X
\end{pmatrix} , \nonumber
\end{split}
\end{equation}
we can rewrite \eqref{compact} in the form of algorithm \eqref{constrained} with $Y_k=g(\theta_k)+e_k$.
We intend to use Lemma \ref{CSA} in Appendix to prove the theorem. Thus, we have to verify B1-B4.
Since $X_k,~\Lambda_k$ are bounded a.s., from \eqref{noise1}, \eqref{noise2}, \eqref{noise3} we conclude that
\begin{equation}
\begin{split}
&E[ \| e_{k}\|^2 ] \leq 2E[ \| e_{1,k}\|^2 ] +2E[ \| e_{2,k}\|^2 ] +E[ \| e_{3,k}\|^2 ] < \infty . \nonumber
\end{split}
\end{equation}
Thus, $E [ \|Y_k\|^2] =2E [ \|g(\theta_k)||^2] +2E[ \| e_{k}\|^2] < \infty , $ and hence B1 holds.
From \eqref{noise1} \eqref{noise2} \eqref{noise3} it follows that B2 holds.
By the definition of $g(\theta)$, from Assumption \ref{ass-set}-b it is seen that B3 holds.
By Theorem \ref{thm2} we conclude that $\theta_k$ is bounded a.s., and hence B4 holds.
In summary, we have validated B1-B4. Then by Lemma \ref{CSA} we conclude that
$(X_k, \Lambda_k) $ converge a.s. to some limit set of the following projected ODE in $\Omega \times \mathbb{R}^{mn}$:
\begin{equation} \label{UDI}
\begin{split}
& \dot{X}(t)= -\nabla \widetilde{f}(X(t)) -\mathcal{\bar{L}}\big(X(t)+\Lambda(t) \big)-Z(t) , ~Z(t) \in N_{\Omega}(X(t)),\\&
\dot{ \Lambda}(t)=\mathcal{\bar{L}} X(t),
\end{split}
\end{equation}
where $Z(\cdot) $ is the projection or constraint term, the minimum force needed to keep $X(\cdot)$ in $\Phi.$
Define $V(X,\Lambda) =\| X-X^{*}\|^2+ \| \Lambda-\Lambda^{*}\|^2$. By \eqref{UDI} we derive
\begin{equation}
\begin{array}{ll}
\dot{V}(X,\Lambda)
&=( X-X^{*})^T \dot{X}+(\Lambda-\Lambda^{*})^T \dot{\Lambda}
\\&=-( X-X^{*})^T \widetilde{f}(X) -( X-X^{*})^T\mathcal{\bar{L}}\big(X +\Lambda \big) -( X-X^{*})^T Z
+( \Lambda-\Lambda^{*})^T \mathcal{\bar{L}} X. \nonumber \end{array}
\end{equation}
Since $Z(t) \in N_{\Omega}(X(t))$, by the definition of normal cone we derive
$( X-X^{*})^T Z \geq 0.$ Since $ \mathcal{\bar{L}} $ is symmetric, by \eqref{consensus} we derive
\begin{equation} \label{ineqs}
\begin{array}{ll}
\dot{V}(X,\Lambda) & \leq -( X-X^{*})^T \widetilde{f}(X)- X ^T\mathcal{\bar{L}} X -X ^T\mathcal{\bar{L}} \Lambda +
\Lambda^T \mathcal{\bar{L}} X -(\Lambda^{*})^T \mathcal{\bar{L}} (X-X^{*}) \\& \leq -( X-X^{*})^T \big( \widetilde{f}(X)+\mathcal{\bar{L}} \Lambda^{*}\big)- X^T\mathcal{\bar{L}} X \\& \leq \Phi(X^{*}, \Lambda^{*}) -\Phi(X, \Lambda^{*}) -X^T\mathcal{\bar{L}} X,
\end{array}
\end{equation}
where in the last inequality we have used $\Phi(X^{*},\Lambda^{*}) \geq \Phi(X, \Lambda^{*})+ (X^{*}-X )^T \big ( \nabla \widetilde{f}(X )+ \mathcal{\bar{L}} \Lambda^{*} \big)$ since $\Phi(X, \Lambda^{*})$ is convex with respect to $X.$ Noting that $\mathcal{\bar{L}} $ is positive semi-definite, by the definition of saddle point we derive
\begin{equation}\label{dot}
\begin{array}{ll}
\dot{V}(X,\Lambda) & \leq0. \nonumber \end{array}
\end{equation}
By the LaSalle invariant theorem \cite{nonlinear}, the trajectories produced by \eqref{UDI} converge to the largest invariant set contained in the set $S=\{(X,\Lambda)\in \Omega \times \mathbb{R}^{mn}: \dot{V}(X,\Lambda) =0\}.$
By \eqref{ineqs} it is clear that $S=\{(X,\Lambda)\in \Omega \times \mathbb{R}^{mn}:X^T\mathcal{\bar{L}} X=0,
~\Phi(X^{*}, \Lambda^{*}) -\Phi(X, \Lambda^{*}) =0\}.$
If $X^T\mathcal{\bar{L}} X=0$, then by noticing that $\mathcal{\bar{L}}$ is the Laplacian matrix of an undirected connected graph, from Lemma \ref{lem1} we have $X= \mathbf{1} \otimes x$ for some $x \in \mathbb{R}^m$.
Since $(X^{*}, \Lambda^{*})$ is a saddle point of $\Phi(X,\Lambda)$, $X^{*}$ is an optimal solution to the problem \eqref{problem2} by Lemma \ref{lem3}. Thus, $\Phi(X^{*}, \Lambda^{*})=\widetilde{f}(X^{*})+(\Lambda^{*})^T \mathcal{\bar{L}} X^{*}=\widetilde{f}(X^{*})=f^{*}.$ If $ \Phi(X, \Lambda^{*})-\Phi(X^{*}, \Lambda^{*}) =0$, then from $X= \mathbf{1} \otimes x$ and $X\in \Omega$ we conclude that $f(x)=f^*,~ x\in \Omega_{o}.$
Thus, $x$ is also an optimal solution to problem \eqref{problem1}, and hence
$S=\{(X,\Lambda):X= \mathbf{1} \otimes x, ~ x \in \Omega_o^{*}, ~\Lambda \in \mathbb{R}^{mn}\}.$ Therefore, $(X_k, \Lambda_k) $ converges to the largest invariant set in set $S$. Consequently, the estimates given by all agents finally reach consensus, and hence \eqref{mainresults} holds.
Furthermore, if $\Omega_o^{*}=\{x^*\}$, then by \eqref{mainresults} we derive \eqref{mainresults2}.
The proof is completed.
\hfill $\blacksquare$
\section{Asymptotic Properties}\label{sec:Normality}
In this section, we establish asymptotic properties of the distributed primal-dual algorithm \eqref{algorithm1} when there is no constraint, i.e., $\Omega_i =\mathbb{R}^m~~\forall i \in \mathcal{V}$.
\subsection{Dimensionality Reduction }
We now introduce a linear transformation to the algorithm \eqref{compact}.
Note that $\mathcal{\bar{L}}$ is the Laplacian matrix of an undirected connected graph by Assumption \ref{ass-graph}-b. Then by Lemma \ref{lem1}
$\mathcal{\bar{L}}$ has a simple zero eigenvalue while all other eigenvalues are positive.
Thus, there exists an orthogonal matrix $\mathcal{V}=(\mathcal{V}_1~\mathcal{V}_2)$, where $\mathcal{V}_2=\frac{\mathbf{1}}{\sqrt{n}}$ and each column of $ \mathcal{V}_1 \in \mathbb{R}^{n \times (n-1)}$ is an eigenvector corresponding to some positive eigenvalue of $\mathcal{\bar{L}}$, such that \begin{equation}\label{diagnoal}
\mathcal{V}^T \mathcal{ \bar{L}} \mathcal{V}=\begin{pmatrix}
\mathcal{ S} & \mathbf{ 0 } \\
\mathbf{0} & 0
\end{pmatrix}
\end{equation}
where $\mathcal{S} =diag\{\kappa_2, \cdots, \kappa_n\} \in \mathbb{R}^{(n-1) \times (n-1)}$ with $\kappa_i,i=2,\cdots,n$ being positive eigenvalues of $\mathcal{\bar{L}}$.
By multiplying both sides of \eqref{diagnoal} from left with $\mathcal{V}$,
it follows that \begin{align}\label{left}
\mathcal{ \bar{L}} \mathcal{V}=(\mathcal{V}_1~\mathcal{V}_2)\begin{pmatrix}
\mathcal{ S} & \mathbf{ 0 } \\
\mathbf{0} & 0
\end{pmatrix} =(\mathcal{V}_1\mathcal{S}~\mathbf{0} ).
\end{align}
Similarly, by multiplying both sides of \eqref{diagnoal} from right with $\mathcal{V}^T$,
we obtain \begin{align}\label{right}
\mathcal{V}^T\mathcal{\bar{L}}=\begin{pmatrix}
\mathcal{ S} & \mathbf{ 0 } \\
\mathbf{0} & 0
\end{pmatrix} \begin{pmatrix}
\mathcal{V}_1^T \\
\mathcal{ V}_2^T
\end{pmatrix}= \begin{pmatrix}
\mathcal{S} \mathcal{V}_1^T \\
\mathbf{0}
\end{pmatrix}.
\end{align}
Let $(X^{*}, \Lambda^{*}) $ be the primal-dual solution pair of the problems \eqref{problem2} and \eqref{pdual} when $ \Omega_i= \mathbb{R}^{m}~\forall i \in \mathcal{V}$. Then by Lemma \ref{lem3}, $(X^{*}, \Lambda^{*}) $ satisfies
$$\Phi(X^{*} , \Lambda) \leq \Phi(X^{*}, \Lambda^{*}) \leq \Phi(X , \Lambda^{*})~~\forall X, \Lambda \in \mathbb{R}^{mn},$$
and hence \begin{equation}\label{optimalcd}
\nabla \widetilde{f}(X^*)+ ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \Lambda^*=\mathbf{0} ,~~( \mathcal{ \bar{L}} \otimes \mathbf{I}_m)X^*=\mathbf{0} .
\end{equation}
The first equality in \eqref{optimalcd} is the optimality condition for $\min\limits_X \widetilde{f}(X)+X^T ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \Lambda^* $, where the minimum is attained at $X^*$.
Therefore, from \eqref{compact} \eqref{optimalcd} and $\Omega=\mathbb{R}^{mn}$ it follows that
\begin{align} X_{k+1}-X^{*}&= X_k -X^{*}- \gamma_k \big( \nabla \widetilde{f}(X_{k}) - \nabla \widetilde{f}(X^*) \big) \nonumber \\& -\gamma_k ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \big( \Lambda_k-\Lambda^* + X_k -X^*\big) + \gamma_k \big(e_{1,k}+e_{2,k}\big) , \label{primal1} \\
\Lambda_{k+1}-\Lambda^*&=\Lambda_k-\Lambda^* +\gamma_k ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) (X_k-X^*)+ \gamma_k e_{3,k} . \label{dual1}
\end{align}
Define $$\widetilde{\Lambda}_{1,k}\triangleq (\mathcal{V}_1^T \otimes \mathbf{I}_m) (\Lambda_k-\Lambda^*),~\widetilde{\Lambda}_{2,k}\triangleq(\mathcal{V}_2^T \otimes \mathbf{I}_m)(\Lambda_k-\Lambda^*).$$
Then by multiplying both sides of \eqref{dual1} with $\mathcal{V}^T \otimes \mathbf{I}_m$ from left, by the rule of Kronecker product
\begin{equation}\label{pkrp}
(A \otimes B)(C \otimes D)=(A \otimes C)(B \otimes D)
\end{equation}
and by \eqref{right} we obtain
\begin{equation}
\begin{split}
\begin{pmatrix}
\widetilde{\Lambda}_{1,k+1} \\
\widetilde{\Lambda}_{2,k+1}
\end{pmatrix}
= \begin{pmatrix}
\widetilde{\Lambda}_{1,k} \\
\widetilde{\Lambda}_{2,k}
\end{pmatrix} +\gamma_k \begin{pmatrix}
\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m \\
\mathbf{0}
\end{pmatrix} \widetilde{X}_k + \gamma_k \begin{pmatrix}
\mathcal{V}_1^T \otimes \mathbf{I}_m \\
\mathcal{V}_2^T \otimes \mathbf{I}_m
\end{pmatrix} e_{3,k} . \nonumber
\end{split}
\end{equation}
Hence
\begin{equation}\label{dual2}
\widetilde{\Lambda}_{1,k+1} = \widetilde{\Lambda}_{1,k} +\gamma_k (\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X}_k+ \gamma_k (\mathcal{V}_1^T \otimes \mathbf{I}_m) e_{3,k}.
\end{equation}
Since $\mathcal{V}\mathcal{V}^T=\mathbf{I}_n$, by \eqref{left} and by \eqref{pkrp} we derive
\begin{equation}
\begin{split}
( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \big( \Lambda_k-\Lambda^*) &= ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) (\mathcal{V}\otimes \mathbf{I}_m) ( \mathcal{V}^T \otimes \mathbf{I}_m) \big( \Lambda_k-\Lambda^*) \\&=(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m ~\mathbf{0} )\begin{pmatrix}
\widetilde{\Lambda}_{1,k} \\
\widetilde{\Lambda}_{2,k}
\end{pmatrix} =( \mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m) \widetilde{\Lambda}_{1,k} . \nonumber
\end{split}
\end{equation}
Then by setting $\widetilde{X}_k \triangleq X_k-X^*,$ from \eqref{primal1} we derive
\begin{equation} \label{primal2}
\begin{split}\widetilde{ X}_{k+1} = \widetilde{ X}_k& - \gamma_k \big( \nabla \widetilde{f}(\widetilde{X}_{k}+X^*) - \nabla \widetilde{f}(X^*) \big) \\& -\gamma_k ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}_k-\gamma_k (\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m) \widetilde{\Lambda}_{1,k} + \gamma_k \big(e_{1,k}+e_{2,k}\big).
\end{split}
\end{equation}
\subsection{Asymptotic Normality and Efficiency}
To investigate the asymptotic properties of the algorithm \eqref{dual2}\eqref{primal2}, we need the following conditions.
\begin{ass}\label{ass-stepsize1} $\gamma_k = \frac{1}{k^\nu}$ with $\nu \in (\frac{2}{3},1).$
\end{ass}
\begin{ass}\label{ass-function}
a) $f(\cdot)$ is strictly convex and the unique optimal solution is $x^*$.
b) The Hessian matrix of $f_i(\cdot) $ at point $x^*$ is $H_i$, and $ \sum_{i=1}^n H_i$ is positive definite.
c) There exists a constant $c>0$ such that $\| \nabla f_i(x)-\nabla f_i(x^*)-H_i(x-x^*)\| \leq c \|x-x^*\|^2~\forall i\in \mathcal{V}.$
\end{ass}
\begin{rem}
By Assumption \ref{ass-function}-b, the Hessian matrix $\nabla^2f(x^*) $ is positive definite.
If in addition, for any $i\in \mathcal{V}$, the Hessian matrix function $ \nabla^2 f_i(\cdot) $ is globally Lipschitz, then by \cite[Lemma 1.2.4]{Nesterov} we derive
$$\| \nabla f_i(y)-\nabla f_i(x)- \nabla^2 f_i(x)(y-x)\| \leq \frac{M}{2} \|y-x\|^2,$$
where $M>0$ is a constant. Hence Assumption \ref{ass-function}-c holds.
\end{rem}
\begin{ass} \label{ass-independency}
a) For any $i \neq j \in \mathcal{V}$, $v_{i,k} $ and $v_{j,k}$ are conditionally independent given $\mathcal{F}_{k-1}'.$
b) For any $(i_1,j_1)\neq (i_2,j_2)$ with $i_1,i_2,j_1, j_2 \in \mathcal{V}$,
$\omega_{i_1j_1,k}$ and $ \omega_{i_2j_2,k}$ are conditionally independent given $\mathcal{F}_{k-1}'$,
$\zeta_{i_1j_1,k}$ and $\zeta_{i_2j_2,k}$ are conditionally independent given $\mathcal{F}_{k-1}'.$
c) For any $i , j \in \mathcal{V}$, $v_{i,k}$, $\omega_{ij,k}$, and $\zeta_{ij,k}$ are conditionally independent given $\mathcal{F}_{k-1}'.$
d) For any $i\in \mathcal{V}$, $v_{i,k}$ and $\mathcal{L}_k$ are conditionally independent given $\mathcal{F}_{k-1}.$
\end{ass}
Define \begin{equation}\label{symbol}
\begin{split} & R_{\omega,i} \triangleq \sum_{j=1}^n \sigma_{ij}R_{\omega,ij}, \quad R_{\zeta,i} \triangleq \sum_{j=1}^n \sigma_{ij}R_{\zeta,ij} ,~~R_v \triangleq diag \big \{R_{v,1},\cdots,R_{v,n} \big) ,\\ &
R_{\omega} \triangleq diag \big \{R_{\omega,1} ,\cdots,R_{\omega,n} \big\}, ~~R_{\zeta} \triangleq diag \big \{R_{\zeta,1} ,\cdots,R_{\zeta,n} \big \}, \\& S_1\triangleq E[\big(( \mathcal{ L} _k - \mathcal{ \bar{L}} )\mathcal{V}_1S^{-1} \mathcal{V}_1^T \otimes \mathbf{I}_m \big)\nabla \widetilde{f}(X^*) \nabla \widetilde{f}(X^*)^T\big(\mathcal{V}_1S^{-1} \mathcal{V}_1^T( \mathcal{ L} _k - \mathcal{ \bar{L}} )^T \otimes \mathbf{I}_m \big)],\\&\mathcal{H}=diag\{H_1,\cdots, H_n\},~~ S_2=R_{v}+R_{\omega}+R_{\zeta}.
\end{split}
\end{equation}
\begin{thm}\label{thm3} (Asymptotic Normality) Set $\Omega_i=\mathbf{R}^m~\forall i \in \mathcal{V}.$
Let Assumptions \ref{ass-set}-b, \ref{ass-graph}, \ref{ass-noise2}, \ref{ass-stepsize1}, \ref{ass-function}, and \ref{ass-independency} hold. Then
$ \theta_k=col\{ \widetilde{X}_k , \widetilde{ \Lambda}_{1,k}\}$ is asymptotically normal:
$$ \theta_k /\sqrt{\gamma_k} \xlongrightarrow [k \rightarrow \infty]{d} N(\mathbf{0},\Sigma), $$
where $\Sigma=\int_{0}^{\infty} e^{Ft}\Sigma_1 e^{F^Tt} dt,$
\begin{equation}\label{hessian}
F\triangleq - \begin{pmatrix}
( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) +\mathcal{H} & \mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m \\
-\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m & 0
\end{pmatrix} ,
\end{equation} and
$$ \Sigma_1=\begin{pmatrix}
& S_1+S_2 & -R_{\omega} (\mathcal{V}_1 \otimes \mathbf{I}_m )\\
& - (\mathcal{V}_1^T\otimes \mathbf{I}_m) R_{\omega} &(\mathcal{V}_1^T \otimes \mathbf{I}_m ) R_{\omega} ( \mathcal{V}_1 \otimes \mathbf{I}_m )
\end{pmatrix}.$$
\end{thm}
\begin{thm}\label{thm4} Set $\Omega_i=\mathbf{R}^m~\forall i \in \mathcal{V}.$
Let Assumptions \ref{ass-set}-b, \ref{ass-graph}, \ref{ass-noise2}, \ref{ass-stepsize1}, \ref{ass-function}, and \ref{ass-independency} hold.
Define $ \bar{\theta}_n=\frac{1}{n}\sum_{k=1}^n\theta_k.$
Then $\{\bar{\theta}_k\}$ is asymptotically efficient:
$$ \sqrt{k} \bar{\theta}_k \xlongrightarrow [k \rightarrow \infty]{d} N\big(\mathbf{0}, F^{-1}\Sigma_1(F^{-1})^T \big). $$
\end{thm}
\subsection{Proof of Theorems \ref{thm3} and \ref{thm4}}
Before proving the results, we give some lemmas to be used in the proof of Theorem \ref{thm3}.
\begin{lem} \label{Hurwitz}\cite[Lemma 2]{Sayed_TSP_2015} Let a block matrix $F$ have the following form
$$F=-\begin{pmatrix}
X &Y^T \\
-Y & \mathbf{0}
\end{pmatrix},$$
and let $X \in \mathbb{R}^{p \times p}$ be positive definite and $Y \in \mathbb{R}^{p \times q}$ be of full row rank.
Then the matrix $F$ is Hurwitz.
\end{lem}
\begin{lem} \label{lem-hessian} Let Assumption \ref{ass-graph}-b and Assumption \ref{ass-function}-b hold. Then $F$ defined by \eqref{hessian} is Hurwitz.
\end{lem}
{\bf Proof:}
Since $H_i~\forall i\in \mathcal{V}$ are Hessian matrices of convex functions, $\mathcal{H} $ is semi-positive definite.
The matrix $ \mathcal{ \bar{L}} $ is semi-positive definite since it is a Laplacian matrix of an undirected graph.
Therefore, a nonzero vector $x \in \mathbb{R}^{mn}$ satisfies
$x^T(\mathcal{ \bar{L}} \otimes \mathbf{I}_m+\mathcal{H}) x=0$ if and only if
\begin{equation}\label{iif}
x^T ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) x=0,~x^T\mathcal{H} x=0 .
\end{equation}
Since $ \mathcal{ \bar{L}} $ is the Laplacian matrix of an undirected connected graph, by Lemma \ref{lem1} a nonzero vector $x \in \mathbb{R}^{mn}$ satisfies $x^T ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) x=0$ if and only if $x=\mathbf{1} \otimes u~\forall u \neq \mathbf{0} \in \mathbb{R}^m . $
However, by Assumption \ref{ass-function}-b $$(\mathbf{1} \otimes u)^T \mathcal{H} (\mathbf{1} \otimes u)= u^T (\sum_{i=1}^n H_i )u>0~~\forall u \neq \mathbf{0} . $$
Therefore, the two equalities in \eqref{iif} do not hold simultaneously.
Thus, $ ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) +\mathcal{H}$ is positive definite.
Note that $\mathcal{S}\mathcal{V}_1^T$ is of full row rank. Then, by Lemma \ref{Hurwitz} we see that
$F$ defined by \eqref{hessian} is Hurwitz.
\hfill $\blacksquare$
\textbf{ Proof of Theorem \ref{thm3}:}
Set
\begin{equation}\label{gtheta}
\begin{split}
& \theta\triangleq \begin{pmatrix}
\widetilde{X} \\
\widetilde{ \Lambda}_1
\end{pmatrix},~~e_k \triangleq \begin{pmatrix}
e_{1,k}+e_{2,k} \\
( \mathcal{V}_1^T\otimes \mathbf{I}_m) e_{3,k}
\end{pmatrix} , \\
& g(\theta)\triangleq-\begin{pmatrix}
& g_1(\theta) \\
& g_2(\theta)
\end{pmatrix} = -\begin{pmatrix}
\ \nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*) + ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}+(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m) \widetilde{\Lambda}_{1} \\
-(\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X} \end{pmatrix} .
\end{split}
\end{equation}
Then we can rewrite \eqref{dual2} \eqref{primal2} as
$$\theta_{k+1}=\theta_k+\gamma_kY_k,$$
where $Y_k=g(\theta_k)+e_k$.
We want to apply Lemma \ref{Stable1} i). For this, we have to validate conditions C0-C3.
\textbf{Step 1:} We first show C0.
By Assumption \ref{ass-set}-b, from \cite[Theorem 2.1.5]{Nesterov} it follows that
\begin{equation}\label{equc}
\langle x-y , \nabla f_i(x) - \nabla f_i(y)\rangle \geq \frac{1}{L_f } \| \nabla f_i(x)- \nabla f_i(y) \|^2 ~~\forall x,y \in \mathbb{R}^m.
\end{equation}
Set $V_1(\theta) \triangleq \frac{1}{2}(\|\widetilde{X}\|^2+\|\widetilde{\Lambda}_1\|^2).$
Then by \eqref{gtheta} \eqref{equc} we obtain \begin{equation}\label{Lyv1}
\begin{split}
&\nabla V_1(\theta)^T g(\theta)\\&=- \widetilde{X}^T\big( \ \nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*) + ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}+(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m)\widetilde{\Lambda}_{1} \big)+ \widetilde{\Lambda}_1^T (\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X} \\&=- \widetilde{X}^T ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}-\alpha\widetilde{X}^T\big( \ \nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*)\big)\\&
\leq - \widetilde{X}^T ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}- \frac{1}{L_f } \| \ \nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*)\|^2\\&
\leq - \widetilde{X}^T ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}.
\end{split}
\end{equation}
Set $V_2(\theta) \triangleq \widetilde{f}(\widetilde{X}+X^*)-\widetilde{f}(X^*)- \widetilde{X}^T \nabla \widetilde{f}(X^*) +
\frac{1}{2}\big( \widetilde{X}^T( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X} \big)+ \widetilde{X}^T(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m)\widetilde{\Lambda}_{1}.$ Then
\begin{equation}\label{Lyv2}
\begin{split}
&\nabla V_2(\theta)^T g(\theta)=- \|g_1(\theta)\|^2+ \| (\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X}\|^2.
\end{split}
\end{equation}
By \eqref{right} we have
$$ (\mathcal{V}^T\mathcal{\bar{L}} \otimes \mathbf{I}_m) \widetilde{X} = col\{
(\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X}, \mathbf{0}\}.$$
Then by $\mathcal{V}\mathcal{V}^T=\mathbf{I}_n$ and the properties of Kronecker products \eqref{pkrp} we derive
$$ \| (\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X}\|^2 =\| (\mathcal{V}^T\mathcal{\bar{L}} \otimes \mathbf{I}_m) \widetilde{X} \|^2= \widetilde{X} ^T(\mathcal{\bar{L}}^2\otimes \mathbf{I}_m) \widetilde{X} .$$
Hence by \eqref{Lyv2} we derive
\begin{equation}\label{Lyv3}
\begin{split}
&\nabla V_2(\theta)^T g(\theta)=- \|g_1(\theta)\|^2+ \widetilde{X} ^T(\mathcal{\bar{L}}^2\otimes \mathbf{I}_m) \widetilde{X} .
\end{split}
\end{equation}
Set $V(\theta)\triangleq V_1(\theta)+\alpha V_2(\theta) $ with $0< \alpha<\frac{1}{ \kappa^* }$, where $\kappa^*= \max\limits_{i=2,\cdots,n}\kappa_i $.
Then by \eqref{Lyv1} and \eqref{Lyv3} we derive
\begin{equation}\label{Lyv4}
\begin{split}
&\nabla V(\theta)^T g(\theta)=- \widetilde{X} ^T\big((\mathcal{\bar{L}}-\alpha\mathcal{\bar{L}}^2)\otimes \mathbf{I}_m\big) \widetilde{X} -\alpha \|g_1(\theta)\|^2 .
\end{split}
\end{equation}
Since $\mathcal{V}^T\mathcal{L} \mathcal{V}=diag\{0,\kappa_2, \cdots, \kappa_n\},$ we have $\mathcal{V}^T\mathcal{\bar{L}}^2 \mathcal{V}=diag\{0,\kappa_2^2, \cdots, \kappa_n^2\} $.
Then all possible distinct eigenvalues of $\mathcal{\bar{L}}-\alpha\mathcal{\bar{L}}^2$ are $0, $ and $ \kappa_i-\alpha\kappa_i^2, i=2,\cdots, n$. By $ 0< \alpha \leq \frac{1}{\kappa_n} $ we derive $\alpha\kappa_i \leq 1~~ \forall i=1,\cdots, n$, and hence
$ \kappa_i-\alpha\kappa_i^2= \kappa_i(1- \alpha\kappa_i)\geq 0~ ~\forall i=1,\cdots, n.$
Thus for any $\alpha$ with $0< \alpha<\frac{1}{ \kappa^* }$, the matrix $\mathcal{L} - \alpha\mathcal{L}^2$ is positive semi-definite. Then by \eqref{Lyv4} we have
\begin{equation}\label{Lyv5}
\nabla V(\theta)^T g(\theta) \leq 0~~\forall \theta \in \mathbb{R}^{(2n-1) m}.
\end{equation}
The equality holds if and only if $ \widetilde{X} ^T\big((\mathcal{\bar{L}}-\alpha\mathcal{\bar{L}}^2)\otimes \mathbf{I}_m\big) \widetilde{X} =0,~g_1(\theta)=\mathbf{0}.$
Since the matrix $\mathcal{L} - \alpha\mathcal{L}^2$ is positive semi-definite, the equality
$ \widetilde{X} ^T\big((\mathcal{\bar{L}}-\alpha\mathcal{\bar{L}}^2)\otimes \mathbf{I}_m\big) \widetilde{X} =0$ implies that $\widetilde{X}=\mathbf{1} \otimes \tilde{x}$. Then by multiplying both sides of
\begin{equation}\label{equ1} \nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*) + ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}+(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m) \widetilde{\Lambda}_{1} =\mathbf{0}
\end{equation} from left with $\mathbf{1}^T\otimes \mathbf{I}_m $, from $\mathbf{1}^T\mathcal{V}_1=0$ and $\mathbf{1}^T\mathcal{\bar{L}}=0$ by \eqref{pkrp} it follows that $$\nabla f(x^*+\tilde{x})-\nabla f(x^*)=\mathbf{0}.$$
Since $f(\cdot)$ is strictly convex with $x^*$ being the unique optimal solution,
$\nabla f(x^*+\tilde{x})= \nabla f(x^*)=\mathbf{0},$ and hence, $\tilde{x}=\mathbf{0}$. Then from \eqref{equ1} we see $(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m ) \widetilde{\Lambda}_{1} =\mathbf{0} $, and hence $(\mathcal{V}_1^T\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m ) \widetilde{\Lambda}_{1} =\mathbf{0} $.
By noticing that $\mathcal{V}_1^T\mathcal{V}_1=\mathbf{I}_{n-1}$ and $S$ is a diagonal matrix with positive diagonal entries, we obtain $\widetilde{\Lambda}_{1} =\mathbf{0} $. Consequently, $\nabla V(\theta)^T g(\theta) =0$ only if $ \theta=\mathbf{0}$. Therefore, by \eqref{Lyv5} we derive C0.
\textbf{Step 2: } We now verify C1. We use Lemma \ref{Stable0} to prove $ \lim\limits_{k\rightarrow \infty} \theta_k =\mathbf{0} ~a.s.$
Note that $X_k,\Lambda_k$ are bounded with probability one by Theorem \ref{thm2}.
Then by the definition of $\theta_k $ we know that C1' holds.
From Lemma \ref{lemma1} and Assumption \ref{ass-stepsize1}, by the convergence theorem for mds \cite{YSChow} $$\sum_{k=1}^{\infty} \gamma_k e_{k}<\infty~a.s.,$$
and hence C2' holds.
By the definition of $g(\theta)$ it is seen that C3' holds.
Since it has already been proven in Step 1 that C0 holds, by Lemma \ref{Stable0} we obtain C1.
\textbf{Step 3:} We now verify C2.
Define $\varepsilon_k=e_kI_{[\|\theta_k \|\leq \epsilon]},~ \nu_k=e_kI_{[\|\theta_k \|> \epsilon]}$, where $\epsilon>0$ is a constant.
By noting that $ \lim\limits_{k\rightarrow \infty} \theta_k =\mathbf{0} ~a.s.$,
there exists $k_0$ possibly depending on samples such that
\begin{equation}\label{unib}
\|\theta_k \| \leq \epsilon~~\forall k \geq k_0~~a.s.
\end{equation}Thus,
$\nu_k=\mathbf{0}~\forall k \geq k_0~a.s.,$
and hence \eqref{nuk} holds.
Since $\theta_k$ is adapted to $ \mathcal{F}_{k-1}$, from \eqref{noise1}, \eqref{noise2}, \eqref{noise3}
and by $e_k$ defined in \eqref{gtheta} we derive
\begin{equation}\label{csm}
E[\varepsilon_k | \mathcal{F}_{k-1}]=E[e_k | \mathcal{F}_{k-1}]I_{[\|\theta_k \|\leq \epsilon]}=\mathbf{0}~a.s.
\end{equation}
Since $(\mathcal{ L} _k \otimes \mathbf{I}_m ) X^*=\mathbf{0}$, by \eqref{def-e1} we derive
\begin{equation}
\begin{split}
e_{1,k} &= \big( (\mathcal{\bar{L}}-\mathcal{ L} _k)\otimes \mathbf{I}_m \big)
\big( \Lambda_k-\Lambda^* \big) +\big( (\mathcal{\bar{L}}-\mathcal{ L} _k)\otimes \mathbf{I}_m \big)\widetilde{X}_k +\big( (\mathcal{\bar{L}}-\mathcal{ L} _k )\otimes \mathbf{I}_m \big) \Lambda^*.\nonumber
\end{split}
\end{equation}
By noticing that $ \mathcal{V}_1\mathcal{V}_1^T =\mathbf{I}_n-\frac{\mathbf{1}\mathbf{1}^T}{n}$
and $\mathcal{\bar{L}}\mathbf{1}= \mathcal{L}_k \mathbf{1}=\mathbf{0} $, we derive
\begin{equation}\label{mult1}
\begin{split}
\mathcal{\bar{L}}-\mathcal{ L} _k =( \mathcal{\bar{L}}-\mathcal{ L} _k )\mathcal{V}_1\mathcal{V}_1^T.
\end{split}
\end{equation}
Then by $\widetilde{\Lambda}_{1,k}=(\mathcal{V}_1^T \otimes \mathbf{I}_m)(\Lambda_k-\Lambda^*)$ we see that
\begin{equation}\label{lambda1}
\begin{split}
& (\mathcal{\bar{L}}-\mathcal{ L} _k )\otimes \mathbf{I}_m
( \Lambda_k-\Lambda^* ) = ( \mathcal{\bar{L}}-\mathcal{ L} _k )\mathcal{V}_1 \otimes \mathbf{I}_m \widetilde{\Lambda}_{1,k}.
\end{split}
\end{equation}
Then by multiplying both sides of the first equality in \eqref{optimalcd} with $\mathcal{V}^T \otimes \mathbf{I}_m $ from left, and by \eqref{pkrp} \eqref{right} we obtain
\begin{equation}
\label{ }
\begin{split}
& -(\mathcal{V}^T \otimes \mathbf{I}_m)\nabla \widetilde{f}(X^*) = (\mathcal{V}^T\otimes \mathbf{I}_m) ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \Lambda^* = \big(\mathcal{V}^T \mathcal{ \bar{L}} \otimes \mathbf{I}_m\big) \Lambda^* =col\{
( \mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m)\Lambda^*,
\mathbf{0 } \}.
\nonumber
\end{split}
\end{equation}
So, $ (\mathcal{S} \mathcal{V}_1^T\otimes \mathbf{I}_m) \Lambda^* =-(\mathcal{V}_1^T\otimes \mathbf{I}_m)\nabla \widetilde{f}(X^*).$
Then by \eqref{mult1} we obtain
\begin{equation}\label{mult2}
\begin{split}
& \big( (\mathcal{\bar{L}}-\mathcal{ L} _k)\otimes \mathbf{I}_m \big) \Lambda^* =\big(( \mathcal{\bar{L}}-\mathcal{ L} _k )\mathcal{V}_1\mathcal{V}_1^T\otimes \mathbf{I}_m \big ) \Lambda^*
\\&=\big(( \mathcal{\bar{L}}-\mathcal{ L} _k )\mathcal{V}_1S^{-1} S \mathcal{V}_1^T \otimes \mathbf{I}_m\big) \Lambda^* =\big(( \mathcal{ L} _k - \mathcal{ \bar{L}} ) \mathcal{V}_1S^{-1} \mathcal{V}_1^T\otimes \mathbf{I}_m\big)\nabla \widetilde{f}(X^*). \nonumber
\end{split}
\end{equation}
Thus, by \eqref{lambda1} we obtain
\begin{equation}\label{e1}
\begin{split}
e_{1,k}&= \big( ( \mathcal{\bar{L}}-\mathcal{ L} _k )\mathcal{V}_1 \otimes \mathbf{I}_m)\big) \widetilde{\Lambda}_{1,k} +\big( (\mathcal{\bar{L}}-\mathcal{ L} _k )\otimes \mathbf{I}_m \big)\widetilde{X}_k \\&+\big(( \mathcal{ L} _k - \mathcal{ \bar{L}} ) \mathcal{V}_1S^{-1} \mathcal{V}_1^T\otimes \mathbf{I}_m\big)\nabla \widetilde{f}(X^*).
\end{split}
\end{equation}
Note that $X_k $ and $ \Lambda_k$ are adapted to $\mathcal{F}_{k-1} $, and that $\mathcal{L}_k$
is independent of $\mathcal{F}_{k-1}$ by Assumption \ref{ass-graph}-d. Then by \eqref{e1} we derive
\begin{equation}\label{noise-cond1}
\begin{split}
E[ \| e_{1,k}\|^2 | \mathcal{F}_{k-1}] \leq 3C_{01}( \| \widetilde{ X}_k\|^2+ \|\widetilde{\Lambda}_{1,k}\|^2+C_{04} )~~a.s., \end{split}\end{equation}
where $C_{04}=\|(\mathcal{V}_1S^{-1} \mathcal{V}_1^T \otimes \mathbf{I}_m)\nabla \widetilde{f}(X^*)\|^2.$
Since $X_k=X^*+\widetilde{X}_k$, by \eqref{noise2}\eqref{noise3} we derive \begin{align}
& E[\| e_{2,k}\|^{2} | \mathcal{F}_{k-1}] \leq C_{02} +6c_v (\|X^*\|^2+ \| \widetilde{ X}_k\|^2)~~a.s., \label{noise-cond2} \\&
E[\| e_{3,k}\|^{2} | \mathcal{F}_{k-1}]\leq 2C_{01}(\|X^*\|^2+ \| \widetilde{ X}_k\|^2) + C_{03}~~a.s. \label{noise-cond3}
\end{align}
Since $\theta_k$ is adapted to $\mathcal{F}_{k-1}$, by \eqref{noise-cond1} \eqref{noise-cond2} \eqref{noise-cond3} we know that there exists a constant $K>0$ such that \begin{equation}\label{condnoise}
E[\| \varepsilon_k\|^{2} | \mathcal{F}_{k-1} ]=E[\| e_{k}\|^{2}| \mathcal{F}_{k-1} ] I_{[ \| \theta_k \|\leq \epsilon]} \leq K~~\forall k\geq 1 ~~a.s.
\end{equation}
Consequently, by \eqref{csm} and \eqref{condnoise} we know that \eqref{A-1} holds.
By the Chebyshev's inequality from \eqref{condnoise} we have
$$\mathbb{P}(\|\varepsilon_k \| > a) \leq \frac{E[\| \varepsilon_k \|^2]}{a^2} \leq \frac{K}{a^2}~~\forall k \geq 1.$$
Then by the Schwarz inequality from \eqref{condnoise} we derive
\begin{equation}
\begin{split} E[\|\varepsilon_k I_{[\|\varepsilon_k \| > a]} \|] & \leq (E[\|\varepsilon_k \|^2] )^{\frac{1}{2}}(E[I_{[\varepsilon_k \| > a]}^2] )^{\frac{1}{2}} \leq \sqrt{K} \sqrt{\mathbb{P}(\|\varepsilon_k \| > a)} \leq \frac{K}{a} ~~\forall k \geq 1.
\nonumber
\end{split}
\end{equation}
Therefore,
$\lim\limits_{a \rightarrow \infty} \sup_k E[\|\varepsilon_k I_{[\|\varepsilon_k \| > a]} \|]=0,$ and hence \eqref{A-3} holds.
Note that
\begin{equation}\label{defe}
\begin{split}
e_ke_k^T&=\begin{pmatrix}
e_{1,k}+e_{2,k} \\
(\mathcal{V}_1^T \otimes \mathbf{I}_m ) e_{3,k}
\end{pmatrix} ( e_{1,k}^T+e_{2,k} ^T, e_{3,k}^T (\mathcal{V}_1 \otimes \mathbf{I}_m ) \\
&= \begin{pmatrix}
& (e_{1,k}+e_{2,k} ) (e_{1,k}+e_{2,k} ) ^T & (e_{1,k}+e_{2,k} ) e_{3,k}^T( \mathcal{V}_1 \otimes \mathbf{I}_m ) \\
& (\mathcal{V}_1^T \otimes \mathbf{I}_m ) e_{3,k} (e_{1,k}+e_{2,k} )^T & (\mathcal{V}_1^T \otimes \mathbf{I}_m ) e_{3,k} e_{3,k}^T ( \mathcal{V}_1 \otimes \mathbf{I}_m )\end{pmatrix},
\end{split}
\end{equation}
and that $ \lim\limits_{k\rightarrow \infty} \widetilde{ X}_k =\mathbf{0}~ a.s. , \lim\limits_{k\rightarrow \infty} \widetilde{ \Lambda}_{1,k} =\mathbf{0} ~ a.s. $, and $\widetilde{ X}_k,\widetilde{ \Lambda}_{1,k}$ are adapted to $\mathcal{F}_{k-1}$. Then by Assumption \ref{ass-graph}-d, from \eqref{unib} \eqref{e1} and the definition of $S_1$ given in \eqref{symbol} we derive
\begin{equation}\label{cond-e1}
\begin{split}
&E[ e_{1,k} e_{1,k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} S_1~a.s . \end{split}
\end{equation}
By Assumptions \ref{ass-noise2}-a and \ref{ass-independency}-b we derive \begin{equation} \label{ind1}
\begin{split}
E[\omega_{i_1j_1,k}\omega_{i_2j_2,k}^T|\mathcal{F}_{k-1}']&=E[ \omega_{i_1j_1,k} |\mathcal{F}_{k-1}']E[ \omega_{i_2j_2,k}^T|\mathcal{F}_{k-1}']=\mathbf{0}~~\forall (i_1,j_1) \neq (i_2,j_2). \end{split}
\end{equation}
Thus, noticing that $a_{ij,k}~\forall i,j \in\mathcal{V}$ are adapted to $\mathcal{F}_{k-1}'$ we obtain
\begin{equation}\label{ }
\begin{split}
&E[ \omega_{i_1,k}\omega_{i_2,k}^T | \mathcal{F}_{k-1}'] = \sum_{j_1,j_2=1}^n a_{i_1j_1,k}a_{i_2j_2,k} E[\omega_{i_1j_1,k} \omega_{i_2j_2,k}^T| \mathcal{F}_{k-1}']
=\mathbf{0}~~\forall i_1 \neq i_2. \nonumber \end{split}
\end{equation}
Then by $\mathcal{F}_k \subset\mathcal{F}_k'$ we have
\begin{equation}\label{corre-cond-w}
\begin{split}
&E[ \omega_{i_1,k}\omega_{i_2,k}^T | \mathcal{F}_{k-1}]= E\big[ E[ \omega_{i_1,k}\omega_{i_2,k}^T | \mathcal{F}_{k-1}'] \big| \mathcal{F}_{k-1} \big]
=\mathbf{0}~~\forall i_1 \neq i_2. \end{split}
\end{equation}
By \eqref{ind1} and Assumption \ref{ass-noise2}-a we obtain
\begin{equation}\label{ }
\begin{split}
E[ \omega_{i,k}\omega_{i,k}^T | \mathcal{F}_{k-1}'] &= \sum_{j_1,j_2=1}^n a_{ij_1,k}a_{ij_2,k} E[\omega_{ij_1,k} \omega_{ij_2,k}^T| \mathcal{F}_{k-1}'] = \sum_{j=1}^n a_{ij,k}^2 E[\omega_{ij,k} \omega_{ij,k}^T| \mathcal{F}_{k-1}'] = \sum_{j=1}^n a_{ij,k}^2 R_{\omega,ij}. \nonumber \end{split}
\end{equation}
Then by Assumptions \ref{ass-graph}-c and \ref{ass-graph}-d, from $\mathcal{F}_k \subset\mathcal{F}_k'$ it follows that
\begin{equation}\label{cond-w}
\begin{split}
E[ \omega_{i,k}\omega_{i,k}^T | \mathcal{F}_{k-1}]&= E\big[ E[ \omega_{i,k}\omega_{i,k}^T | \mathcal{F}_{k-1}'] \big| \mathcal{F}_{k-1} \big] \\&=E[ \sum_{j=1}^n a_{ij,k}^2 R_{\omega,ij} | \mathcal{F}_{k-1}]
= \sum_{j=1}^n E[ a_{ij,k}^2 ]R_{\omega,ij}= R_{\omega,i} ~~a.s.\end{split}
\end{equation}
Similarly,
\begin{equation}\label{ }
\begin{split}
&E[ \zeta_{i_1,k}\zeta_{i_2,k}^T | \mathcal{F}_{k-1}]= \mathbf{0}~~\forall i_1 \neq i_2~~a.s,
~~E[ \zeta_{i,k}\zeta_{i,k}^T | \mathcal{F}_{k-1}]= R_{\zeta,i}~~a.s. \nonumber \end{split}
\end{equation}
From here, by \eqref{corre-cond-w} and \eqref{cond-w} we conclude that
\begin{equation}\label{cond-wz}
\begin{split}
&E[ \omega_k \omega_k^T | \mathcal{F}_{k-1}] = diag \big \{R_{\omega,1} ,\cdots,R_{\omega,N} \big\} =R_{\omega}~~a.s., \\& E[\zeta_k \zeta_k^T | \mathcal{F}_{k-1}] = diag \big \{R_{\zeta,1} ,\cdots,R_{\zeta,N} \big\}=R_{\zeta}~~a.s.
\end{split}
\end{equation}
By Assumptions \ref{ass-noise2} and \ref{ass-independency}-c, similar to \eqref{cond-wz} we
can show that
\begin{equation}\label{cond-wzv}
\begin{split}
&E[ \omega_k\zeta_k^T | \mathcal{F}_{k-1}] = \mathbf{0}~~a.s., ~~ E[ \omega_kv_k^T | \mathcal{F}_{k-1}] = \mathbf{0}~~a.s., ~~E[ \zeta_k v_k^T | \mathcal{F}_{k-1}] = \mathbf{0}~~a.s. \end{split}
\end{equation}
From Assumptions \ref{ass-noise2}-c and \ref{ass-independency}-a it follows that
$$E[v_{i,k}v_{j,k}^T|\mathcal{F}_{k-1}']=E[v_{i,k} |\mathcal{F}_{k-1}']E[ v_{j,k}^T|\mathcal{F}_{k-1}']=\mathbf{0}~~\forall i \neq j,$$
and hence by Assumption \ref{ass-noise2}-c we obtain
\begin{equation}\label{ }
\begin{split}
E[ v_k v_k^T | \mathcal{F}_{k-1}'] &= diag \big \{E[ v_{1,k} v_{1,k}^T | \mathcal{F}_{k-1}'] ,\cdots,E[ v_{n,k} v_{n,k}^T | \mathcal{F}_{k-1}'] \big) \\&
\xlongrightarrow [k \rightarrow \infty]{} diag \big \{R_{v,1},\cdots,R_{v,n} \big) = R_v~a.s. \nonumber
\end{split}
\end{equation}
Noting that $v_{k}$ and $\mathcal{L}_k$ are conditionally independent given $\mathcal{F}_{k-1}$ by Assumption \ref{ass-independency}-d, by \cite[Corollary 7.3.2]{YSChow} we have
$$E[ v_k v_k^T | \mathcal{F}_{k-1}'] =E[ v_k v_k^T | \mathcal{F}_{k-1},\mathcal{L}_k]=E[ v_k v_k^T | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} R_v~~a.s.$$
For $e_{2,k} $ defined by \eqref{def-e2} , by \eqref{cond-wz} and \eqref{cond-wzv} we derive
\begin{equation}\label{ }
\begin{split}
E[ e_{2,k} e_{2,k}^T | \mathcal{F}_{k-1}]&= E[ v_kv_k^T +\omega_k\omega_k^T+\zeta_k\zeta_k^T | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} R_v+R_{\omega}+R_{\zeta}= S_2~~a.s. \nonumber \end{split}
\end{equation}
Thus by noticing that $\theta_k$ is adapted to $\mathcal{F}_{k-1}' $, from \eqref{unib} we derive
\begin{equation}\label{cond-e2}
\begin{split}
&E[ e_{2,k} e_{2,k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] =E[ e_{2,k} e_{2,k}^T | \mathcal{F}_{k-1}] I_{[ \| \theta_k \|\leq \epsilon]}\xlongrightarrow [k \rightarrow \infty]{} S_2~~a.s. \end{split}
\end{equation}
Since $\mathcal{F}_{k} \subset \mathcal{F}_{k}'$ and $e_{1,k},\theta_k$ are adapted to $\mathcal{F}_{k-1}'$,
by \eqref{equ3} we obtain \begin{equation} \begin{split} &E [ e_{1,k} e_{2,k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}]= E \big[ E [ e_{1,k} e_{2,k} ^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big ] \\&= E \big[ e_{1,k} I_{[ \| \theta_k \|\leq \epsilon]} E [ e_{2,k}^T | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big ]=\mathbf{0}~~a.s. ,\nonumber
\end{split}
\end{equation}
which incorporating with \eqref{cond-e1} \eqref{cond-e2} yields \begin{equation}\label{cov12}
\begin{split}
&E[ (e_{1,k}+e_{2,k})( e_{1,k}+ e_{2,k})^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}]
\xlongrightarrow [k \rightarrow \infty]{} S_1+S_2~~a.s. \end{split}
\end{equation}
By \eqref{equ0} we see that
\begin{equation}\label{mdsw}
E[\omega_{k} | \mathcal{F}_{k-1}'] =\mathbf{0} .
\end{equation}Hence, noticing that $\mathcal{F}_{k} \subset \mathcal{F}_{k}'$ and that $e_{1,k}I_{[ \| \theta_k \|\leq \epsilon]}$ is adapted to $\mathcal{F}_{k-1}'$, we obtain
\begin{equation}\label{cond1w}
\begin{split}
E [ e_{1,k} I_{[ \| \theta_k \|\leq \epsilon]} \omega_k^T| \mathcal{F}_{k-1}] &=E\big[E [ e_{1,k} I_{[ \| \theta_k \|\leq \epsilon]} \omega_k^T| \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1}\big]\\&
=E\big[ e_{1,k} I_{[ \| \theta_k \|\leq \epsilon]} E [ \omega_k^T| \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1}\big]=\mathbf{0}~~a.s.
\end{split}
\end{equation}
Note that
\begin{equation}\label{tildex}
\begin{split}
( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_mX_k =( \mathcal{L}_k -\mathcal{\bar{L}})\otimes \mathbf{I}_m\widetilde{X}_k,
\end{split}
\end{equation} and that $\widetilde{ X}_k $ and $ \widetilde{ \Lambda}_{1,k}$ are adapted to $\mathcal{F}_{k-1}$.
Then from \eqref{def-e3} \eqref{e1} \eqref{cond1w}, by Assumption \ref{ass-graph}-c and \ref{ass-graph}-d we derive
\begin{equation}\label{cond23}
\begin{split}
&E[ e_{1,k} e_{3,k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] =E[ e_{1,k} \big( ( \mathcal{L}_k -\mathcal{\bar{L}})\otimes \mathbf{I}_m\widetilde{X}_k \big)^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} \mathbf{0}~~a.s., \end{split}
\end{equation}
where the limit takes place because $ \lim\limits_{k\rightarrow \infty}\widetilde{ X}_k =\mathbf{0}~a.s.$, and $ \lim\limits_{k\rightarrow \infty} \widetilde{ \Lambda}_{1,k} =\mathbf{0} ~a.s.$
Since $\mathcal{L}_k, X_k $ are adapted to $\mathcal{F}_{k-1}'$, by \eqref{def-e3}\eqref{equ3} we derive
\begin{equation} \begin{split} E [ e_{2,k} e_{3,k}^T| \mathcal{F}_{k-1}'] & = E [ e_{2,k} \big(( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m X_k \big)^T | \mathcal{F}_{k-1}' ]- E [ e_{2,k} \omega_k^T | \mathcal{F}_{k-1}' ] \\&
= E [ e_{2,k} | \mathcal{F}_{k-1}' ] \big(( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m X_k \big)^T- E [ e_{2,k} \omega_k^T | \mathcal{F}_{k-1}' ] = - E [ e_{2,k} \omega_k^T | \mathcal{F}_{k-1}' ] . \nonumber
\end{split}
\end{equation}
Then by $\mathcal{F}_{k} \subset \mathcal{F}_{k}'$ we conclude that
\begin{equation} \begin{split} E [ e_{2,k} e_{3,k}^T| \mathcal{F}_{k-1}]&= E \big[ E [ e_{2,k} e_{3,k} ^T | \mathcal{F}_{k-1}'] \big | \mathcal{F}_{k-1} \big ] \\&= -E \big[ E [ e_{2,k} \omega_k^T | \mathcal{F}_{k-1}' ] | \mathcal{F}_{k-1} \big ]=-E [ e_{2,k} \omega_k^T | \mathcal{F}_{k-1}]~~a.s. \nonumber
\end{split}
\end{equation}
Noticing $ e_{2,k}$ defined by \eqref{def-e2}, by \eqref{cond-wz} and \eqref{cond-wzv} we derive
$$E [ e_{2,k} e_{3,k}^T| \mathcal{F}_{k-1}]=-E [ \omega_k \omega_k^T | \mathcal{F}_{k-1}]=-R_{\omega}~~a.s.$$
Since $\theta_k$ is adapted to $\mathcal{F}_{k-1}$, by \eqref{unib} we obtain
\begin{equation} \begin{split}
& \lim_{k \rightarrow \infty}E [ e_{2,k} e_{3,k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}]
= \lim_{k \rightarrow \infty}E [ e_{2,k} e_{3,k}^T | \mathcal{F}_{k-1}]I_{[ \| \theta_k \|\leq \epsilon]}
=-R_{\omega} \lim_{k \rightarrow \infty}I_{[ \| \theta_k \|\leq \epsilon]} =-R_{\omega}~~a.s., \nonumber \end{split}
\end{equation}
which incorporating with \eqref{cond23} yields
\begin{equation}\label{cov123}
\begin{split}
&E[ (e_{1,k}+e_{2,k} ) e_{3,k}^T ( \mathcal{V}_1 \otimes \mathbf{I}_m )I_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} -R_{\omega} ( \mathcal{V}_1 \otimes \mathbf{I}_m )~~a.s. \end{split}
\end{equation}
Since $\mathcal{L}_k$ and $\widetilde{X}_k$ are adapted to $\mathcal{F}_{k-1}'$, by \eqref{mdsw} and $\mathcal{F}_{k} \subset \mathcal{F}_{k}'$ we obtain
\begin{equation} \label{condwx}
\begin{split}
& E [ \omega_k \big(( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m\widetilde{X}_k \big)^T| \mathcal{F}_{k-1}] \\&=
E\big[ E [ \omega_k \big(( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m\widetilde{X}_k \big)^T| \mathcal{F}_{k-1}'] \big| \mathcal{F}_{k-1}\big]\\&= E\big[ E [ \omega_k | \mathcal{F}_{k-1}'] \big(( \mathcal{L}_k -\mathcal{\bar{L}})\otimes \mathbf{I}_m\widetilde{X}_k \big)^T\big| \mathcal{F}_{k-1}\big]=\mathbf{0}~~a.s.
\end{split}
\end{equation}
By definition of $e_{3,k}$ and \eqref{tildex} we see
\begin{equation}\label{cov3}
\begin{split}
e_{3,k} e_{3,k}^T &=-( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m\widetilde{X}_k \omega_k^T-\omega_k \big(( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m\widetilde{X}_k \big)^T \\
&+\big( ( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m \big)\widetilde{X}_k \widetilde{X}_k^T \big(( \mathcal{L}_k -\mathcal{\bar{L}}) ^T \otimes \mathbf{I}_m\big) +\omega_k\omega_k^T. \end{split}
\end{equation}
Since $ \lim\limits_{k\rightarrow \infty} \widetilde{ X}_k =\mathbf{0}~a.s.,$ and $\widetilde{ X}_k$ is adapted to $\mathcal{F}_{k-1},$ by Assumption \ref{ass-graph}-c and \ref{ass-graph}-d we obtain
\begin{equation}\label{ }
\begin{split}
& \| E\big[ ( \mathcal{L}_k -\mathcal{\bar{L}}) \otimes \mathbf{I}_m \widetilde{X}_k \widetilde{X}_k^T ( \mathcal{L}_k -\mathcal{\bar{L}}) ^T \otimes \mathbf{I}_m \big| \mathcal{F}_{k-1}\big]\|
\leq E[\| \mathcal{L}_k -\mathcal{\bar{L}}\|^2] \| \widetilde{ X}_k \|^2\xlongrightarrow [k \rightarrow \infty]{} 0 ~~a.s.\nonumber \end{split}
\end{equation}
Then by \eqref{cond-wz} \eqref{condwx} \eqref{cov3} we obtain
\begin{equation}\label{ }
\begin{split}
&E[ e_{3,k} e_{3,k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} R_{\omega}~~a.s.,\nonumber \end{split}
\end{equation}
which incorporating with \eqref{defe}, \eqref{cov12} and \eqref{cov123} yields
\begin{equation}\label{ }
\begin{split}
&E[ e_{k} e_{k}^TI_{[ \| \theta_k \|\leq \epsilon]} | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} \Sigma_1~~a.s.\nonumber \end{split}
\end{equation}
Hence by the definition of $ \varepsilon_k $ we obtain
\begin{equation}\label{tidle2}
E[ \varepsilon_k \varepsilon_k^T | \mathcal{F}_{k-1}] \xlongrightarrow [k \rightarrow \infty]{} \Sigma_1~~a.s.
\end{equation}
By \eqref{condnoise} we derive
$$ E\big[\sup_k E[\| \varepsilon_k\|^{2} | \mathcal{F}_{k-1} ]\big] \leq K.$$
Then by the Lebesgue dominated convergence theorem \cite[Corroloary 4.2.3]{ YSChow} and by \eqref{tidle2} we have
\begin{equation}
\begin{split}
\lim_{k \rightarrow \infty}E\big[E[ \varepsilon_k \varepsilon_k^T | \mathcal{F}_{k-1}] \big]=E\big[ \lim_{k \rightarrow \infty} E[ \varepsilon_k \varepsilon_k^T | \mathcal{F}_{k-1}] \big]= \Sigma_1. \nonumber
\end{split}
\end{equation}
Thus, $\lim\limits_{k \rightarrow \infty}E[ \varepsilon_k \varepsilon_k^T ]=\lim\limits_{k \rightarrow \infty}E\big[E[ \varepsilon_k \varepsilon_k^T | \mathcal{F}_{k-1}] \big]=\Sigma_1,$ and hence \eqref{A-2} holds.
So, C2 has been verified.
Step 4: It remains to check C3. By \eqref{gtheta} and \eqref{hessian} we derive
\begin{equation}
\begin{split}
g(\theta)-F\theta &= \begin{pmatrix}
\big( ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) +\mathcal{H}\big) \widetilde{X} + \mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m \widetilde{\Lambda}_1\\
-\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m \widetilde{X}
\end{pmatrix} \\& -\begin{pmatrix}
\nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*) + ( \mathcal{ \bar{L}} \otimes \mathbf{I}_m) \widetilde{X}+(\mathcal{V}_1\mathcal{S} \otimes \mathbf{I}_m) \widetilde{\Lambda}_{1} \\
-(\mathcal{S} \mathcal{V}_1^T \otimes \mathbf{I}_m) \widetilde{X} \end{pmatrix}
\\&=- \begin{pmatrix}
& \nabla \widetilde{f}(\widetilde{X}+X^*) - \nabla \widetilde{f}(X^*) - \mathcal{H} \widetilde{X} \\
& \mathbf{0}
\end{pmatrix} . \nonumber
\end{split}
\end{equation}
Then by Assumption \ref{ass-function}-c we obtain
$$\| g(\theta)-F\theta\|^2\leq c \|\widetilde{X}\|^2\leq c \|\theta\|^2,$$
and hence C3 holds.
In summary, we have verified C0-C3. Then by Lemma \ref{Stable1} i) the assertion of the theorem follows.
\hfill $\blacksquare$
\textbf{ Proof of Theorem \ref{thm4}:}
Since it is shown in the proof of Theorem \ref{thm3} that C0-C3 hold, by Lemma \ref{Stable1} ii) we immediately derive the assertion. \hfill $\blacksquare$
\section{Numerical Examples} \label{sec:Simulation}
In this section, we do simulations for the distributed parameter estimation problem considered in \cite{Sayed_TSP_2015}. We aim at estimating the unknown $m $-dimensional vector
$x^*$ based on the data gathered by $n$ spatially distributed sensors in the network.
Each agent $i=1,\cdots,n$ at time $k$ has access to its real scalar measurement $ d_{i,k}$
given by the following linear time-varying model
\begin{equation}\label{model}
d_{i,k}=u_{i,k}x^* +\nu_{i,k}, \nonumber
\end{equation}
where $u_{i,k} \in \mathbb{R}^{1 \times m}$ is the regression vector accessible to agent $i$,
and $\nu_{i,k} $ is the local observation noise of agent $i$.
Assume that $\{u_{i,k}\}$ and $\{\nu_{i,k}\}$ are mutually independent iid Gaussian sequences with distributions $N(\mathbf{0}, R_{u,i})$ and $N(0,\sigma_{i,\nu}^2)$, respectively. Besides, we allow some covariance matrices nonpositive definite, but require $\sum_{i=1}^n R_{u,i}$ be positive definite.
This parameter estimation problem is modeled as solving the following distributed stochastic optimization problem
\begin{equation}\label{filter1}
\min_{x} ~ f(x)=\sum_{i=1}^n f_i(x)\stackrel{\Delta}{=} E [ \parallel d_{i,k}-u_{i,k}x\parallel ^2].
\end{equation}
So, $f_i(x)=(x-x^*)^T R_{u,i}(x-x^*)+\sigma_{i,\nu},^2$ and $\nabla f_i(x)= R_{u,i}(x-x^*).$
Therefore, $x^*$ is the unique optimal solution to \eqref{filter1} when $\sum_{i=1}^n R_{u,i}$ is positive definite.
Let $m=3$. Set $x^*=(1,2,3)$, and $$R_{u,1}=\begin{pmatrix}
1 & 0 & 0 \\
0 & 1&0 \\
0 & 0&0
\end{pmatrix},~R_{u,2}=\begin{pmatrix}
0 & 0 & 0 \\
0 & 1&0 \\
0 & 0&1
\end{pmatrix} ,R_{u,2}=\begin{pmatrix}
1 & 0 & 0 \\
0 & 0&0 \\
0 & 0&1
\end{pmatrix},~ \sigma_{i,\nu}=\sqrt{ 0.1}~\forall i \in \mathcal{V}.$$
Set $n=3$ with the underling undirected graph being fully connected.
At any time $k\geq 0$, with equal probability $\frac{1}{3}$ for each edge, we randomly choose one edge from the graph. Set $a_{ij,k}=a_{ji,k}=1$ when the edge between $i$ and $j$ is chosen. For any $i,j\in \mathcal{V}$,
let the communication noises $\{\omega_{ij,k}\}$
and $\{\zeta_{ij,k}\}$ be mutually independent iid Gaussian sequences $N(0,0.1\mathbf{I}_3)$.
\begin{figure}
\centering
\includegraphics[width=5in]{one.png}
\caption{ Estimates $x_{1,k}$, histographs and limit distributions for
$(x_{1,k}-x^*)/\sqrt{\gamma_k} $ at $k=1000$} \label{GOne}
\end{figure}
Set $\gamma_k=\frac{1}{k^{0.75}}$.
By using $u_{i,k}$ and $d_{i,k}$ observed at time $k$, the noisy observation of the gradient $\nabla f_i(x_{i,k}) $ is constructed as $g_{i,k}=u_{i,k}^Tu_{i,k}x_{i,k}-d_{i,k}u_{i,k}^T$.
Let $\{ x_{i,k}\}$ and $\{ \lambda_{i,k}\}$ be produced by the algorithm \eqref{algorithm1} with initial values $x_{i,0}=\mathbf{0},~ \lambda_{i,0}=\mathbf{0}$.
Since $v_{i,k}=(u_{i,k}^Tu_{i,k}- R_{u,i})(x_{i,k}-x^*)-\nu_{i,k}u_{i,k}^T,$ it is seen that
$\{v_{i,k}\}$ satisfies Assumption \ref{ass-noise}-c
with $c_v= max\{ E[\|u_{i,k}^Tu_{i,k}- R_{u,i}\|^2], \sigma_{i,v}^2\|R_{u,i}\|\} $. Then
$\lim\limits_{k \rightarrow \infty} x_{i,k}=x^*$ by Theorem \ref{thm1},
and hence $\lim\limits_{k \rightarrow \infty} E[v_{i,k} v_{i,k}^T | \mathcal{F}_{k-1}']=\sigma_{i,v}^2 R_{u,i}$.
As a result, the gradient observation noise for the distributed parameter estimation problem satisfies
Assumption \ref{ass-noise2}-c.
The algorithm \eqref{algorithm1} is calculated for $1000$ independent samples with $k \leq 1000$.
For $i=1,2,3,$ the estimates $x_{i,k}$ and the histographs for each component of
$(x_{i,k}-x^*)/\sqrt{\gamma_k}$ at time $1000$ are shown in Figs. \ref{GOne}, \ref{GTwo}, and \ref{GThree}, respectively. We use the normal distribution to fit the $1000$ independent samples
for each component of $(x_{i,k}-x^*)/\sqrt{\gamma_k},~i=1,2,3$ with $k=1000$. It is shown that the
data are fitted with the normal distribution by the Kolmogrov-Smirnov test with the significance level $\alpha=0.05.$
Fig. \ref{GOne}-a demonstrates estimates given by agent $1$ for components of $x^*=(1,2,3)$, where the real lines denote true values, while the dashed lines are their estimates. The estimation errors
$(x_{i,k}-x^*)/\sqrt{\gamma_k}$ are presented in Figs. \ref{GOne}-b, \ref{GOne}-c, \ref{GOne}-d, where the histographs are given by errors of 1000 samples at time $k=1000$, which are fitted by Gaussian densities.
Figs. \ref{GTwo} and \ref{GThree} are for agents 2 and 3, respectively.
\begin{figure}
\centering
\includegraphics[width=5in]{two.png}
\caption{ Estimates $x_{2,k}$, histographs and limit distributions for $(x_{2,k}-x^*)/\sqrt{\gamma_k} $ at $k=1000$} \label{GTwo}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=5in]{three.png}
\caption{ Estimates $x_{3,k}$, histographs and limit distributions for $(x_{3,k}-x^*)/\sqrt{\gamma_k} $ at $k=1000$} \label{GThree}
\end{figure}
\section{Conclusions}\label{sec:Conclusion}
In this work, a stochastic approximation based distributed primal-dual algorithm is proposed to solve the distributed constrained stochastic optimization problem over random networks with imperfect communications.
The local estimates derived at all agents all shown to a.s. reach a consensus belonging to the optimal solution set.
Besides, we established conditions for the unconstrained problem, under which the
asymptotic normality and asymptotic efficiency of the proposed algorithm are established.
The influence on the convergence rate of the
conditional covariance matrices of communication noises and gradient errors,
properties of the cost function like gradients and Hessian matrices at the optimal point, as well as the random graphs and its mean graph is demonstrated in the paper.
\appendices
\numberwithin{equation}{section}
\section{Some Results on Stochastic Approximation}
To ease reading, some results on non-neagetive super-martingales \cite{supermds} and some information from stochastic approximation \cite{Chen2002} \cite{Kushner} are cited below.
\begin{lem} \cite{supermds} \label{martingale}
Let $(\Omega,\mathcal{F},\mathbb{P})$ be a probability space and $\mathcal{F}_0\subset \mathcal{F}_1 \subset\cdots $
be a sequence of sub-$\sigma$-algebras of $\mathcal{F}$. Let $\{d_k\}$ and
$\{w_k\}$ be nonnegative $\mathcal{F}_{k}$-measurable random variables such that
$$ E[d_{k+1}| \mathcal{F}_{k}] \leq (1+\alpha_k) d_k + w_k ,$$
where $\alpha_k \geq 0$ and $\sum_{k=1}^{\infty}\alpha_k <\infty$.
If $\sum_{k=1}^{\infty} w_k < \infty~a.s.$, then $\{d_k\}$ converges a.s. \end{lem}
We now introduce the convergence results for the constrained stochastic approximation algorithm \cite{Kushner}. Consider the following recursion
\begin{equation}
\label{constrained}
\theta_{k+1}=P_{\Phi}(\theta_k+\gamma_kY_k),
\end{equation}
where $\Phi \in \mathbb{R}^m$ is a convex constraint set.
We list the conditions to be used.
B1: $\sup_k E[ \| Y_k\|^2]<\infty~~a.s.$
B2: There is a function $g(\cdot)$ such that
$$ E_k[Y_k]=E[Y_k|\theta_0,Y_i, i<k]=g(\theta_k)~~a.s.$$
B3: $g(\cdot)$ is continuous.
B4: $\theta_k$ is bounded a.s.
\begin{lem} \cite[Theorem 5.2.3] {Kushner} \label{CSA}
Let $\{\theta_k\}$ be generated by \eqref{constrained}.
Assume that the convex set $\Phi$ satisfies the same condition as Assumption \ref{ass-set}-c imposed on $\Omega_i$. Let B1-B4, and Assumption \ref{ass-stepsize} hold. Then $\theta_k$ converges a.s. to the limit set of the following projected ODE \cite{Kushner} in $\Phi$:
$$\dot{\theta}=g(\theta)+z ,~~z(t)\in -N_{\Phi} (\theta(t)),$$
where $z(\cdot) $ is the projection or constraint term, the minimum force needed to keep $\theta(\cdot)$ in $\Phi.$
\end{lem}
We introduce asymptotic properties of the sequence $\{\theta_k\}$ generated by the following recursion:
\begin{equation}\label{recursion}
\theta_{k+1}=\theta_k+\gamma_k g( \theta_k)+\gamma_ke_k.
\end{equation}
We need the following conditions.
C0 There exists a continuously differentiable function $v(\cdot) $ such that
$$g(x)^T \nabla v(x)<0~~\forall x \neq \mathbf{0}.$$
C1' $ \theta_k$ is bounded a.s.
C1 $\lim\limits_{k \rightarrow \infty} \theta_k=\mathbf{0}~~a.s.$
C2' $\sum_{k=1}^{\infty} \gamma_k e_{k}<\infty~a.s.$
C2 The noise sequence $\{e_{k}\}$ can be decomposed into two parts $e_{k}= \varepsilon_k+\nu_k$ such that
\begin{equation}\label{nuk}
\nu_k=o(\sqrt{\gamma_k})~~a.s.,
\end{equation} and $\{\varepsilon_k,\mathcal{F}_k\}$ is an mds satisfying conditions:
\begin{align}
& E[\varepsilon_k| \mathcal{F}_{k-1}]=\mathbf{0},~~\sup_k E[\|\varepsilon_k\|^2| \mathcal{F}_{k-1}] \leq \sigma
\textrm{~with } \sigma \textrm{~being a constant}, \label{A-1}
\\& \lim_{k \rightarrow \infty} E[\varepsilon_k\varepsilon_k^T| \mathcal{F}_{k-1}]= \lim_{k \rightarrow \infty} E[\varepsilon_k\varepsilon_k^T]=S_0~~a.s.,\label{A-2}
\\& \lim_{a \rightarrow \infty} \sup_k E[\|\varepsilon_k\|^2I_{[\|\varepsilon_k\|>a]}]=0. \label{A-3}
\end{align}
C3' $g(\cdot)$ is measurable and locally bounded.
C3 $g(\cdot)$ is measurable and locally bounded. As $\theta \rightarrow \mathbf{0},$
$$\| g(\theta)-F\theta\|\leq c\|\theta\|^2,$$
where $c>0$ and $F$ is stable.
\begin{lem} \label{Stable0}\cite[Theorem 2.2.1]{Chen2002} Let $\{\theta_k\}$ by generated by \eqref{recursion} with an arbitrary initial value $\theta_0$. Let Assumption \ref{ass-stepsize1}, and C0, C1', C2', and C3' hold. Then
$$\lim\limits_{k \rightarrow \infty} \theta_k=\mathbf{0}~~a.s.$$
\end{lem}
\begin{lem} \label{Stable1} Let $\{\theta_k\}$ by generated by \eqref{recursion}. Let Assumption \ref{ass-stepsize1}, and C0, C1, C2, and C3 hold. Then
i) $\frac{1}{\sqrt{\gamma_k}} \theta_k$ is asymptotically normal:
$$\frac{1}{\sqrt{\gamma_k}} \theta_k \xlongrightarrow [k \rightarrow \infty] {d}N(\textbf{0},S),$$
where $S=\int_{0}^{\infty} e^{Ft}S_0 e^{F^Tt} dt$;
ii) $\bar{\theta}_k$ is asymptotically efficient:
$$ \sqrt{k} \bar{\theta}_k \xlongrightarrow [k \rightarrow \infty] {d}N(\textbf{0},S),$$
where $S=F^{-1}S^0 (F^{-1})^T,$ and $\bar{\theta}_k=\frac{1}{k}\sum_{p=1}^k \theta_p $.
\end{lem}
\begin{rem}
Lemma \ref{Stable1} i) is \cite[Theorem 3.3.2]{Chen2002} for the case: $r=0, \beta=1, \alpha=0, x^0=\mathbf{0}$. Since
the noise sequence $\{e_{k}\}$ satisfies C2, by \cite[Remarks 3.4.1 and 3.4.2]{Chen2002} it is seen that A3.4.3 in \cite{Chen2002} holds. Then by \cite[Theorem 3.4.2]{Chen2002} with $ \beta=1,~x^0=\mathbf{0}$ the assertion of
Lemma \ref{Stable1} ii) follows.
\end{rem}
|
\section{Introduction}
Characterizing the molecular gas content of distant galaxies is
essential in order to understand the evolution of the cosmic star
formation rate density \citep{madau14}, and the build--up of stellar
mass \citep{bell03} throughout cosmic time \citep{carilli13}. A
unique way to fully characterize the molecular gas content in galaxies
in the early universe is through spectral line scans in well--studied
cosmological deep fields. In comparison to targeted observations of
individual galaxies, spectral scans have the advantage that molecular
gas reservoirs can be characterized without pre--selection through
other information (e.g., stellar mass, star--formation rate). Such
spectral line scans can also potentially reveal the presence of
gas--rich `dark' galaxies, i.e., galaxies that are invisible in the
optical wavebands, and that would not be selected as targets to search
for molecular gas emission \citep[e.g.,][]{walter12}. In a sense,
spectral line scans follow the spirit of the original {\it HST} deep
fields \citep[e.g.,][]{williams96,beckwith06}, as essentially no prior
knowledge/selection based on galaxy properties enters the choice of
field.
As the main constituent of the molecular gas in galaxies, molecular
hydrogen (H$_2$), is too weak to be detected, the next most abundant
tracer is typically used to measure the molecular gas content:
$^{12}$CO (hereafter: CO). Although this molecule is 10$^4$ times less
abundant, the line can be detected in various environments. As a
consequence, this molecule has been used at low and high redshift to
measure gas masses and kinematics. The CO line emission is observed in
various rotational transitions in galaxies \citep[e.g.,][]{carilli13}.
The rotational ground--state (J=1--0) of CO is at 115.271\,GHz, and
the higher rotational states (J$>$1) are approximately equally spaced
by that frequency\footnote{In reality, the spacing changes slightly as
the dipole moment changes for the higher transitions as a result of
centrifugal forces.}. The amount of high--J emission depends on the
{\em a priori} unknown excitation of the molecular gas. Nevertheless,
full frequency scans in the lowest frequency ALMA bands cover CO
emission at essentially all redshifts (see Fig.~\ref{fig_z_range}).
We here present the rationale for and the observational description of
ASPECS: The ALMA SPECtroscopic Survey in the {\it Hubble} Ultra--Deep
Field (UDF). This paper is structured as follows: Sec.~2 summarizes
our field choices, as well as the observations and data reduction. In
Sec.~3 we describe our methodology to identify line candidates in our
data cubes, and present the continuum maps of both the band~3 and
band~6 observations. In Sec.~4 we compare our findings to simple
expectations based on previous multi--wavelength analysis of the
galaxies in the field. We present our summary in Sec.~5.
A number of accompanying papers build on the data presented in this
paper (hereafter: {\it Paper~I}). In {\it Paper~II} (Aravena et
al.~2016a) we analyse the continuum information (mostly based on the
band~6 observations); in {\it Paper III} (Decarli et al.~2016a) we
discuss the implications for CO luminosity functions and the redshift
evolution of the cosmic molecular gas density; in {\it Paper~IV}
(Decarli et al.~2016b) we examine the properties of those galaxies in
the UDF that show bright CO emission; in {\it Paper~V} (Aravena et
al.~2016b) we search for [C\,{\sc ii}]{} emitters; in {\it Paper~VI} (Bouwens
et al.~2016) we investigate where high--redshift galaxies from ASPECS
lie in relation to known IRX--$\beta$ and IRX--stellar mass
relationships, and finally, in {\it Paper~VII} (Carilli et al.\ 2016)
we describe implications on intensity mapping experiments. Throughout
the paper we assume a standard cosmology with $H_0=70$ km s$^{-1}$
Mpc$^{-1}$, $\Omega_{\rm m}=0.3$ and $\Omega_{\Lambda}=0.7$, broadly
in agreement with the most recent {\em Planck} measurements
\citep{planck15}. Where required, we refer to the AB photometric
system (Oke \& Gunn 1973) for the magnitude definitions and to
Chabrier (2003) for the stellar initial mass function.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{z_range.eps}\\
\caption{CO and [C\,{\sc ii}]{} redshift coverage of our molecular line scans at 1mm and 3mm. See Table \ref{tab_z_range} for the exact redshift ranges of each transition. The 1mm+3mm synergy provides continuous CO redshift coverage at virtually any redshift, with only a tiny gap at $0.6309<z<0.6950$. The [C\,{\sc ii}]{} emission line is covered in the redshift range 6$<$z$<$8 and is discussed in {\it Paper~V}}
\label{fig_z_range}
\end{figure}
\begin{figure*}
\includegraphics[width=0.99\columnwidth]{fig_freq3_coverage.eps}
\includegraphics[width=0.99\columnwidth]{fig_freq1_coverage.eps}\\
\caption{RMS noise as a function of frequency in the 3mm ({\em left}) and 1mm ({\em right}) scans. At 3mm, each channel is 19.5\,MHz wide (five of the native channels), corresponding to 70\,km\,s$^{-1}${} at 84 GHz, and 51\,km\,s$^{-1}${} at 115 GHz. The original frequency settings (A-H) are labeled in the bottom panel, together with the frequency blocks (a-k) used in the data reduction. At 1mm, the channels are 31.3\.MHz wide (four of the native channels), corresponding to 44\,km\,s$^{-1}${} at 212\,GHz, and to 34\,km\,s$^{-1}${} at 272\,GHz. To first order, we reach uniform sensitivity as a function of frequency in both bands. The increase in noise towards high frequencies ($>$113\,GHz) in band 3 is due to the atmosphere (O$_2$).}
\label{fig_noise}
\end{figure*}
\begin{table}
\caption{\rm Lines and corresponding redshift ranges covered in the molecular line scans. For the 3mm data, comoving volume and volume--weighted average redshifts are computed within the primary beam, accounting for its frequency dependence. For the 1mm data, the area is fixed ($3700$ arcsec$^{2}$, as set by the size of the final mosaic).} \label{tab_z_range}
\begin{center}
\begin{tabular}{cccccc}
\hline
Transition & $\nu_0$ & $z_{\rm min}$ & $z_{\rm max}$ & $\langle z \rangle$ & Volume \\
& [GHz] & & & & [Mpc$^3$] \\
(1) & (2) & (3) & (4) & (5) & (6) \\
\hline
\multicolumn{6}{c}{band 3: 3mm (84.176--114.928 GHz)}\\
CO(1-0) & 115.271 & 0.0030 & 0.3694 & 0.2801 & 89 \\
CO(2-1) & 230.538 & 1.0059 & 1.7387 & 1.4277 & 1920 \\
CO(3-2) & 345.796 & 2.0088 & 3.1080 & 2.6129 & 3363 \\
CO(4-3) & 461.041 & 3.0115 & 4.4771 & 3.8030 & 4149 \\
CO(5-4) & 576.268 & 4.0142 & 5.8460 & 4.9933 & 4571 \\
CO(6-5) & 691.473 & 5.0166 & 7.2146 & 6.1843 & 4809 \\
CO(7-6) & 806.652 & 6.0188 & 8.5829 & 7.3750 & 4935 \\
\hline
[C\,{\sc i}]{}$_{1-0}$ & 492.161 & 3.2823 & 4.8468 & 4.1242 & 4287 \\
[C\,{\sc i}]{}$_{2-1}$ & 809.342 & 6.0422 & 8.6148 & 7.4031 & 4936 \\
\hline
\multicolumn{6}{c}{band 6: 1mm (212.032--272.001 GHz)}\\
CO(2-1) & 230.538 & 0.0000 & 0.0873 & 0.0656 & 1.4 \\
CO(3-2) & 345.796 & 0.2713 & 0.6309 & 0.4858 & 314 \\
CO(4-3) & 461.041 & 0.6950 & 1.1744 & 0.9543 & 1028 \\
CO(5-4) & 576.268 & 1.1186 & 1.7178 & 1.4297 & 1759 \\
CO(6-5) & 691.473 & 1.5422 & 2.2612 & 1.9078 & 2376 \\
CO(7-6) & 806.652 & 1.9656 & 2.8044 & 2.3859 & 2864 \\
\hline
[C\,{\sc i}]{}$_{1-0}$ & 492.161 & 0.8094 & 1.3212 & 1.0828 & 1233 \\
[C\,{\sc i}]{}$_{2-1}$ & 809.342 & 1.9755 & 2.8171 & 2.3973 & 2875 \\
[C\,{\sc ii}]{}$_{3/2-1/2}$ &1900.548 & 5.9873 & 7.9635 & 6.9408 & 4431 \\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Observations and data reduction}\label{sec_observations}
\subsection{Choice of Frequencies}
Given the unknown excitation of the molecular gas in a given
high--redshift galaxy, when inferring H$_2$ masses, it is advantageous
to observe the CO emission in the lowest rotational state possible to
minimize excitation corrections, modulo the impact of the Cosmic
Microwave Background (da Cunha et al.\ 2013b). With ALMA, the lowest
frequencies are accessible in band~3, which therefore is the primary
band of choice. An important complement are line scan observations in
band~6, as the combination of both bands results in the following: (i)
other than one small gap in redshift, there is essentially complete
redshift coverage at all redshifts (see Fig.~\ref{fig_z_range} and
Tab.~\ref{tab_z_range}), (ii) the CO excitation (or limits on it) can
be immediately constrained through the detection of multiple
rotational transitions, (iii) deep continuum maps in the respective
observing bands are available `for free', and (iv), the
highest--redshift sources at $6<z<8$ can be probed through [C\,{\sc ii}]{}
emission.
Band~7 (275--373\,GHz) observations may be preferred when one is
interested only in the continuum flux densities of the galaxies but
such observations would only recover very high J (J$>$6) transitions
at $z>2$, which may not be highly excited in main sequence galaxies
\citep{daddi15}. Also, the field of view is smaller than in band~6,
necessitating more extensive mosaicing. The bandwidth of band~7
($\sim100$\,GHz) requires more than 13 frequency tunings (each with a
bandwidth of 8\,GHz). For all of these reasons, band~6 is preferred
over band~7 to complement the band~3 observations.
We obtained full frequency scans in both ALMA band~3 and band~6. In
band~3 this implied 5 frequency setups, labelled A--E in
Fig.~\ref{fig_noise}. Both the upper and lower sideband cover
3.75\,GHz, with a gap of $\sim$8\,GHz. For that reason, the central
range in band 3 was covered twice, resulting in observations with
lower noise in that frequency window. Such an overlap region did not
result from the setup of the band~6 frequency scan, as the gap between
the upper and lower sideband in band~6 is 12\,GHz (see right panel in
Fig.~\ref{fig_noise}).
panels of Fig.~\ref{fig_noise} shows the resulting noise as a function
of frequency. The noise increase in band~3 towards the higher
frequencies is due the atmospheric oxygen line significantly
increasing the system temperatures above $>$113\,GHz. As a consequence
of the higher frequency, the noise in band~6 observations was
significantly higher (and less well--behaved due to skylines etc) than
in band~3.
\subsection{Choice of field}
In principle such molecular line scan observations could be obtained
at (almost) any position in the sky that is not affected by foreground
emission (either our Galaxy, or other nearby galaxies). However, the
analysis and interpretation of the detected galaxies is greatly
facilitated if a field is chosen for which multi--wavelength
observations already exist. It also should be a field that is easily
accessible to ALMA. The {\it Hubble} Ultra Deep Field
\citep[UDF,][]{beckwith06} is the cosmological field with the deepest
observations in all important wavebands, with 18,000 catalogued
galaxies \citep{coe06}. The UDF is situated in the 30$'$ Extended
{\it Chandra} Deep Field South \citep[ECDFS][]{lehmer05} / GOODS-South
\citep{giavalisco04} / CANDELS \citep{grogin11,koekemoer11} region, so
the large--scale structure around this field is well quantified.
The goal of the ALMA frequency scan was to reach a sensitivity such
that the predicted `knee' of the CO luminosity function could be
reached at $z\sim2$ \citep[e.g.,][]{sargent14}. Given that multiple
frequency settings were needed to cover both band~3 and band~6, and
given the limited amount of time available in ALMA cycle~2, this
implied that only the area corresponding to one pointing in band~3
could be covered by our observations. This $\sim1'$ region was covered
with a 7--point mosaic in band 6 (see Fig.~\ref{fig_pointings}). Our
pointing was chosen to lie in the deepest part of the UDF, the
so--called UDF12 (Ellis et al.\ 2013) or eXtremely Deep Field
\citep[XDF,][]{illingworth13} (hereafter: XDF), and included the
highest number of $z$--drop galaxy candidates, i.e. galaxies at $z>6$,
that could be detectable in [C\,{\sc ii}]{} emission. The field also comprises
a significant overlap with the deepest MUSE observations of the UDF
(Bacon et al., in prep.). The region covered by our observations
comprise $\sim$10\% of the total area of the UDF (corresponding to a
co-moving survey volume of ~18,000 Mpc$^3$ out to $z\sim8$) and
harbors roughly $\sim$1500 optical/NIR--selected galaxies. In
Fig.~\ref{fig_pointings} we also present the star formation rates and
stellar masses of all galaxies covered by our observations, based on
fitting of the galaxies' spectral energy distribution
(Sec.~\ref{sec_magphys}).
\subsection{Choice of Array Configuration}
ALMA has been designed to reach high (sub--arcsec) angular resolution.
However, to be sensitive to the full molecular gas reservoir in a
galaxy, observations in a compact array configuration are essential to
ensure that no extended CO emission is missed by the
interferometer. Note that this is not related to the `missing short
spacing' problem\footnote{The missing short spacing problem means that
the interferometer is `blind' to spatial scales above a certain
size. Given the likely clumpiness of high--redshift galaxies, missing
short spacings should typically not be a concern in high--redshift
galaxy observations.}. For instance, in observations with extended
ALMA array configurations the synthesized beam will end up being
smaller than the typical size of a high--redshift galaxy. As a result,
the amount of emission per beam is only a fraction of the total
emission of the galaxy, while the noise does not change. In the case
of low S/N detections, this will result in the non-detection of a
source, whereas the emission would be detected by a compact
configuration. Our observations were taken in the C34-2 and C34-1
configurations, resulting in beam sizes of $3.6''\times 2.1''$
(band~3) and $1.7''\times 0.9''$ (band~6), i.e. well matched to the
expected sizes of the galaxies under consideration.
\begin{figure*}
\begin{center}
\includegraphics[width=0.99\columnwidth]{fig_pointings.eps}
\includegraphics[width=0.99\columnwidth]{fig_ms_zbins.eps}\\
\end{center}
\caption{{\it Left:} The FWHM of the primary beam (i.e. areal coverage) of our 3mm (orange) and 1mm (cyan) observations, overlaid on a three--color {\it HST} F435W / F775W / F105W image of the field from the XDF survey \citep{illingworth13}. The circles show the primary beam of each pointing at the central frequencies of the two scans. {\it Right:} Star formation rates vs. stellar masses for the galaxies in the target field, derived from MAGPHYS fitting (described in Sec.~\ref{sec_magphys}). The four panels show galaxies in different redshift ranges.}
\label{fig_pointings}
\end{figure*}
\subsection{Observations}
The project consists of two spectral scans, one at 3mm (band~3) covering the
frequency range 84--115 GHz and one at 1mm (band~6) covering the
frequency range 212--272 GHz (see Figs.~\ref{fig_z_range} and~\ref{fig_noise}).
The time allocated for both projects amounts to a total of $\sim$40\,hours including overheads
(split approximately 50--50 between band~3 and band~6).
The 3mm observations (ALMA Project ID: 2013.1.00146.S) were carried
out between July $1^{\rm st}$, 2014 and January $6^{\rm th}$,
2015. The 3mm scan consisted of a single pointing (RA=03:32:37.90
Dec=--27:46:25.0, J2000.0) and 5 frequency settings (see
Fig.~\ref{fig_noise}). Each setting had $4\times1.875$ GHz spectral
windows (two in the upper side band, and two in the lower side band),
and was observed in three execution blocks. The native channel width
is $3.9025$\,MHz, or $\sim$12\,km\,s$^{-1}$ at $\sim$100\,GHz.
Observations were carried out in a relatively compact (C34-2) array
configuration with 29--41 antennas, with baselines ranging between 12
and 612\,m. The quasar J0348--2749 was observed in the majority of the
execution blocks as phase and amplitude calibrator, while Uranus and
the quasars J0334--4010 and J0334--4008 were used as flux and bandpass
calibrators. Data were calibrated and imaged with the {\it Common
Astronomy Software Applications} package (CASA) version 4.2.2 of the
ALMA pipeline.
To combine the different setups we adopted the following approach:
1) For each execution block, we split out
cubes in frequency ranges as shown in Fig.~\ref{fig_noise} (a-k). 2)
The frequency ranges marked with the letters e-g, i.e., where upper
and lower side band observations from different frequency settings
overlap, were re-sampled using the CASA task \textsf{ms.cvel}. 3) We
then combined all the available data for each individual frequency range (a-k)
using the CASA task \textsf{concat}. 4) Upper and lower side band data
come with different weighting scales, although the data quality is
comparable. We therefore ran \textsf{statwt} in order to homogenize
the weighting system in the concatenated data. 5) We combined all the
frequency ranges using the task \textsf{concat} again.
We imaged the 3mm cube after averaging over two and five native
channels (7.8\,MHz and 19.5\,MHz respectively) using natural
weighting. The 19.5\,MHz channels correspond to 70\,km\,s$^{-1}${} at 84 GHz,
and 51\,km\,s$^{-1}${} at 115 GHz. We created a band~3 continuum map as well
(see discussion in Sec.~\ref{sec_continuum}). The corresponding
primary beams of the ALMA antennas are $75''$ at 84 GHz and $55''$ at
115 GHz. The restored synthesized beam size is $3.5''\times 2.0''$
(FWHM) with PA=84$^\circ$. We thus adopted a pixel scale of
$0.5''$\,pixel$^{-1}$, and an image size of $90''\times90''$. A
primary beam correction has been applied for all quantitative
analysis. The final data set covers the frequency range
84.176--114.928 GHz, and reaches an rms of
$0.1-0.25$\,mJy\,beam$^{-1}$ per 19.5\,MHz channel (see
Fig.~\ref{fig_noise}). For comparison, the PdBI spectral scan at 3mm
in the {\it Hubble} Deep Field North (Decarli et al.\ 2014, Walter et
al.\ 2012, 2014) reached a sensitivity of $\sim0.3$\,mJy\,beam$^{-1}$
per 90\,km\,s$^{-1}${} channel, or $\sim0.4$\,mJy\,beam$^{-1}$ at the sampling
adopted here. Therefore these ALMA observations are a factor 3--4
deeper at $\nu<113$\,GHz than the previous 100\,hour (on--source)
effort with PdBI (Decarli et al.\ 2014).
The 1mm observations (ALMA Project ID: 2013.1.00718.S) were carried
out between December $12^{\rm th}$, 2014 and April $21^{\rm st}$,
2015. In order to cover a similar area as the 3mm pointing, a 7--point
mosaic was observed, centred on the same coordinates as for the 3mm
observations (see Fig.~\ref{fig_pointings}). For each pointing
position, eight frequency settings were needed to cover the entire
band (see Fig.~\ref{fig_noise}), resulting in continuous coverage from
212--272\,GHz. In this case, there was no overlap between different
spectral windows of various frequency tunings (see
Fig.~\ref{fig_noise}). Observations were carried out in the most
compact available array configuration (C34-1) with 30--34
antennas. Baselines ranged between 12 and 350\,k$\lambda$. The quasar
J0348--2749 was adopted as phase and amplitude calibrator, while
Uranus and the quasar J0334--4008 acted as flux and bandpass
calibrators. The cube was imaged in spectral samplings of 4, 8 and 12
native channels, corresponding to $15.6$\,MHz, $31.2$\,MHz, and
$46.8$\,MHz respectively, as well as in a continuum image. The
$31.2$\,MHz sampling corresponds to 44\,km\,s$^{-1}${} at 212\,GHz and to
34\,km\,s$^{-1}${} at 272\,GHz. We adopted natural weighting, yielding a
synthesized beam of $1.5''\times 1.0''$ with PA=-79$^\circ$. We
adopted a pixel scale of $0.3''$ per pixel. The final mosaic covers a
region of approximately $75''\times 70''$ to the half--sensitivity
point.
\section{Line search}\label{sec_search}
The data reduction resulted in two data cubes, one in band~3 and
one in band~6, as well as continuum maps, which we discuss later (Sec.~\ref{sec_continuum}).
We here describe our methodology to search for line
emitting sources in these cubes.
\subsection{The blind line search}
For our blind search of line candidates, we developed an IRAF--based
routine, \textsf{findclumps}, which operates directly on the imaged
data cubes. The script performs floating averages of a number of
channels, computes the rms of the averaged maps, and searches for
peaks exceeding a certain S/N threshold using the IRAF task
\textsf{daofind}. The position, frequency, and S/N of the recovered
candidates is saved. As input, we used the $7.8$\,MHz and $31.2$\,MHz
sampling for the 3\,mm and 1\,mm cubes respectively. Since the
significance of a line detection is maximized when averaging over a
frequency range comparable with the actual width of the line, we ran
our search over 3, 5, 7, and 9--channel windows, i.e., kernel line
widths of $\sim$50--300\,km\,s$^{-1}${} (an inspection using larger
line--widths did not result in additional detections -- this is also
supported by our completeness test, see below).
The list of line candidates identified by this procedure is then
trimmed in order to keep only candidates that lie within $\sqrt{2}
\times$ the primary beam radius at 3mm (= $53''$ at 84\,GHz, $39''$ at
115\,GHz), equivalent to a response of $\sim$30\% and within a fixed
radius of $30.9''$ at 1mm (given that the latter is a mosaic).
The floating-average approach and the use of different windows of
spectral sampling allow us to avoid missing candidates because of a
priori choices in terms of spectral bins. However, our candidate lists
are subject to multiplicity both spatially and spectrally. Moreover,
the 1mm search is bound to pick up bright continuum sources as
potential line candidates. We therefore masked a posteriori the line
candidates associated with the two brightest 1mm continuum sources
(see Sec.~\ref{sec_continuum} and {\it Paper II}). We consider as
duplicates line candidates that are offset by less than one
synthesized beam ($\sim\!2.5''$ at 3mm, $\sim\!1.5''$ at 1mm) and that
appear in consecutive channels in the floating average.
When assessing the reliability of our line candidates, we need to keep two
separate issues in mind (`fidelity' and `completeness', which we discuss
in Secs.~\ref{sec_fidelity} and~\ref{sec_completeness}).
\subsubsection{Fidelity}\label{sec_fidelity}
First, is a given line detection significant? This question is harder
to address in practice than one would naively think: the S/N of a
single detection will be a function of the width of the line, and the
noise in the cubes is not Gaussian. The best way to address this
question is to perform two independent searches: (a) for positive
emission; these candidates would correspond to both real astrophysical
sources and noise peaks, (b) for negative emission; these candidates
would only correspond to non-astrophysical sources\footnote{An
interesting hypothesis is that at least some of the negative sources
are in fact real absorption systems due to absorption against the
CMB. However, our checks revealed that none of the significant
negative sources are either associated with a galaxy visible in the
UDF, nor with a strong continuum emission. We conclude that the
negative sources revealed by our search are physically
implausible.}. These latter sources can be used to define a term that
we refer to as {\em fidelity}, i.e. we can statistically subtract the
unphysical `negative' lines from the physical `positive' ones.
We thus assess the degree of fidelity in our line search by running
the same search over the positive and negative peaks. The basic
assumption is that, given the interferometric nature of our data set,
and that we do not expect to detect absorption features against very
high-S/N continuum emission, all the `negative' line candidates will
be noise peaks, while the `positive' line candidates will be a mixture
of noise peaks and genuine lines. The search for negative peaks is
performed in the exact same way as the one for positive emission. By
comparing the results of these two searches, we can quantify the
fidelity of our search at a given line candidate significance as
follows: \begin{equation}\label{eq_fidelity} {\rm fidelity}(S/N) = 1 -
\frac{N_{\rm neg}(S/N)}{N_{\rm pos}(S/N)} \end{equation} where $N_{\rm
pos}(S/N)$ and $N_{\rm neg}(S/N)$ are the number of positive and
negative line candidates with a given $S/N$, respectively. This
definition is such that, if the number of negative candidates at a
given $S/N$ is comparable to the number of positive candidates, then
the fidelity is null; if it is negligible, then the fidelity is close
to 100\%. For the analysis of our blind search, we request a fidelity
level of 60\% or higher. This threshold was chosen so that at the
lowest accepted significance, more than half of the `positive' line
candidates are real. We determine the signal-to-noise ($S/N$) ratio
computed by \textsf{findclumps} as follows: For each floating-averaged
channel, we compute the map rms (which will constitute the `noise'
term) and we take the peak pixel value at the position of a line
candidate as `signal'. We emphasize that, since the averaging window
is not optimized to match the actual width of a line candidate (also
this approach assumes spatially-unresolved line emission), this
definition of $S/N$ is by construction conservative. The $S/N$ values
of each line candidate are reported in Tab.~\ref{tab_lines}. In
Fig.~\ref{fig_fidelity} we show how the fidelity of our line search
changes as a function of the line $S/N$. It is convenient to have an
analytical description of the fidelity dependence on $S/N$. While not
physically motivated, the following error function provides a good
description of the observed trend, with the following
parameterization: \begin{equation}\label{eq_fidelity2} {\rm
fidelity}(S/N) = \frac{1}{2}\ {\rm
erf}\left(\frac{S/N-C}{\sigma}\right)+0.5 \end{equation} where $C_{\rm
3mm}$=5.1, $C_{\rm 1mm}$=5.0, $\sigma_{\rm 3mm}$=0.4, $\sigma_{\rm
1mm}$=0.8. This implies that we reach 60, 80, and 95\% fidelity levels
at $S/N$=$5.17$, $5.34$, and $5.57$ at 3mm, and at $S/N$=$5.15$,
$5.50$, and $5.97$ at 1mm. We will use this equation to assess the
fidelity for our individual line detections.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_purity.eps}\\
\caption{The {\em fidelity} in our line search, plotted as a red histogram as a function of the line $S/N$ of the individually detected candidates. The fidelity is defined as in eq.~\ref{eq_fidelity}. The number of candidates as a function of $S/N$ is also shown. We model the fidelity dependence on $S/N$ as an error function (solid black line). The search reaches 60\% fidelity at $S/N\sim5.2$ both at 1mm and 3mm, although the latter shows a sharper increase of fidelity with $S/N$. We choose a fidelity level of $>$60\% for the sources that enter our analysis, implying that, at the lowest significance, out of a sample of 10 candidates, 6 are likely real, and 4 sources at similar $S/N$ were also detected with negative signal.}
\label{fig_fidelity}
\end{figure}
\subsubsection{Completeness}\label{sec_completeness}
The second question concerns our ability to extract faint sources from
our data cubes ({\em completeness}). We address this by inserting
artificial line sources of various strengths and widths in our data
cubes, then calculating our ability to recover them in our line
search. This is a standard way of deriving the completeness of sources
in the low S/N regime.
We assess the level of {\em completeness} in our blind line search by
adding 2500 artificial line sources to the data cube,
re--running our line searching algorithm, and comparing the number of
recovered sources with the input catalog. The line candidates are
assumed to have a gaussian profile along the spectral axis, and the
shape of the synthetic beam in the maps. The line spatial positions,
peak frequencies, peak flux densities and widths are randomly
generated with uniform distributions as follows: RA, Dec and the line
peak frequency are required to be homogeneously distributed within the
cubes. The line peak flux density range between 0.22--1.00\,mJy (at
3mm) and between 0.5--2.0\,mJy (at 1mm), where the fainter side is set
to roughly match the 1-$\sigma$ typical limit of each channel. The
line widths span the range 50--500\,km\,s$^{-1}${}. In
Fig.~\ref{fig_completeness} we show how the completeness of our line
search is a function of the input width and peak flux density of the
lines. At 3mm, the completeness is $>50$\% for peak flux
densities $F_\nu^{\rm line}>0.45$\,mJy, and for line widths $\Delta
v>100$\,km\,s$^{-1}${}. We also observe a minor dependence of the completeness
on the frequency due to the decreasing sensitivity towards the high
frequency end of the scan (see Fig.~\ref{fig_noise}). The line search
in the 1mm mosaic shows a completeness $>50$\% for peak flux
densities $>$0.8\,mJy and widths $>100$\,km\,s$^{-1}${}. These
completeness corrections will be used extensively in {\it Paper~III}.
\begin{table*}
\caption{{\rm Catalogue of the line candidates identified in our analysis. (1) Line ID. (2-3) Right ascension and declination (J2000). (4) Central frequency and uncertainty, based on Gaussian fit. (5) Velocity integrated flux and uncertainty. (6) Line Full Width at Half Maximum, as derived from a Gaussian fit. (7) signal-to-noise as measured by the line searching algorithm. (8) Spatially coincident optical/NIR counterpart? (9) Comments on line identification.}} \label{tab_lines}
\begin{center}
\begin{tabular}{ccccccccl}
\hline
ID & RA & Dec & Frequency & Flux & FWHM & $S/N$ & Opt/NIR & Comments \\
ASPECS... & (J2000.0) & (J2000.0) & [GHz] & [Jy\,km\,s$^{-1}$] & [km\,s$^{-1}$] & & c.part? & \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) \\
\hline
\multicolumn{9}{l}{\bf ~3mm (band 3)} \\
3mm.1 & 03:32:38.52 & -27:46:34.5 & $ 97.567_{-0.003}^{+0.003}$ & $0.72\pm0.03$ & $500_{-30}^{+30}$ & 19.91 & Y & J=3; J=7,8 also detected \\
3mm.2 & 03:32:39.81 & -27:46:11.6 & $ 90.443_{-0.003}^{+0.003}$ & $0.44\pm0.08$ & $540_{-30}^{+30}$ & 12.80 & Y & J=2; J=5 tentatively detected. Confirmed\\
\multicolumn{8}{c}{ } & by opt. spectroscopy \\
3mm.3 & 03:32:35.55 & -27:46:25.7 & $ 96.772_{-0.003}^{+0.003}$ & $0.13\pm0.01$ & $ 57_{-30}^{+30}$ & 9.48 & Y & J=2 is ruled out by optical spectroscopy \\
3mm.4 & 03:32:40.64 & -27:46:02.5 & $ 91.453_{-0.003}^{+0.003}$ & $0.23\pm0.03$ & $ 73_{-30}^{+30}$ & 5.86 & N & lack of counterpart suggests J$>$2 \\
3mm.5 & 03:32:35.48 & -27:46:26.5 & $110.431_{-0.003}^{+0.003}$ & $0.18\pm0.02$ & $ 82_{-25}^{+25}$ & 5.42 & Y & J=2 confirmed by optical spectroscopy \\
3mm.6 & 03:32:35.64 & -27:45:57.6 & $ 99.265_{-0.003}^{+0.003}$ & $0.23\pm0.02$ & $160_{-30}^{+30}$ & 5.40 & N & lack of counterpart suggests J$>$2 \\
3mm.7 & 03:32:39.26 & -27:45:58.8 & $100.699_{-0.003}^{+0.003}$ & $0.08\pm0.01$ & $ 60_{-30}^{+25}$ & 5.40 & N & lack of counterpart suggests J$>$2 \\
3mm.8 & 03:32:40.68 & -27:46:12.1 & $101.130_{-0.003}^{+0.003}$ & $0.19\pm0.01$ & $100_{-30}^{+25}$ & 5.30 & N & no match with nearby galaxy; J$>$2 \\
3mm.9 & 03:32:36.01 & -27:46:47.9 & $ 98.082_{-0.003}^{+0.003}$ & $0.09\pm0.01$ & $ 64_{-30}^{+30}$ & 5.28 & N & lack of counterpart suggests J$>$2 \\
3mm.10 & 03:32:35.66 & -27:45:56.8 & $102.587_{-0.003}^{+0.003}$ & $0.24\pm0.02$ & $120_{-25}^{+25}$ & 5.18 & Y & J=3 ($z$=$2.37$) would match $z_{\rm grism}=2.33$\\
\hline
\multicolumn{9}{l}{\bf ~1mm (band 6)} \\
1mm.1 & 03:32:38.54 & -27:46:34.5 & $227.617_{-0.003}^{+0.003}$ & $0.79\pm0.04$ & $463_{-10}^{+80}$ & 18.28 & Y & J=7 \\
1mm.2 & 03:32:38.54 & -27:46:34.5 & $260.027_{-0.059}^{+0.003}$ & $1.10\pm0.05$ & $478_{-70}^{+11}$ & 16.46 & Y & J=8 \\
1mm.3 & 03:32:38.54 & -27:46:31.3 & $225.181_{-0.003}^{+0.003}$ & $0.22\pm0.02$ & $101_{-18}^{+18}$ & 5.87 & Y & J=3 would imply $z$=0.54, and $z_{\rm grism}=0.59$ \\
1mm.4 & 03:32:37.36 & -27:46:10.0 & $258.333_{-0.003}^{+0.016}$ & $0.27\pm0.02$ & $150_{-20}^{+20}$ & 5.62 & N & if [C\,{\sc ii}]{}, tentative CO(6-5) detection is reported. \\
\multicolumn{8}{c}{ } & Possibly lensed by foreground Elliptical?\\
1mm.5 & 03:32:38.59 & -27:46:55.0 & $265.320_{-0.031}^{+0.003}$ & $0.72\pm0.03$ & $211_{-10}^{+37}$ & 5.47 & N & lack of other lines suggests J=4\\
1mm.6 & 03:32:36.58 & -27:46:50.1 & $222.553_{-0.003}^{+0.003}$ & $0.56\pm0.02$ & $302_{-40}^{+12}$ & 5.45 & Y & J=4 yields $z$=$1.07$, J=5 yields $z$=$1.59$, \\
\multicolumn{8}{c}{ } & J=6 yields $z$=$2.11$, tentative second line for J=4 or J=6 \\
1mm.7 & 03:32:37.91 & -27:46:57.0 & $257.042_{-0.003}^{+0.003}$ & $1.78\pm0.03$ & $179_{-11}^{+11}$ & 5.43 & N & lack of other lines suggests J=4 \\
1mm.8 & 03:32:37.68 & -27:46:52.6 & $222.224_{-0.003}^{+0.022}$ & $0.39\pm0.02$ & $210_{-12}^{+30}$ & 5.33 & N & lack of counterpart excludes J=2,3; lack of second \\
\multicolumn{8}{c}{ } & line exclude CO. [C\,{\sc ii}]{}? \\
1mm.9 & 03:32:36.14 & -27:46:37.0 & $249.085_{-0.003}^{+0.016}$ & $0.34\pm0.02$ & $150_{-20}^{+20}$ & 5.19 & N & J=4; lack of counterparts excludes J$<$4, and lack\\
\multicolumn{8}{c}{ } & of other lines excludes J$>$4 \\
1mm.10 & 03:32:37.08 & -27:46:19.9 & $237.133_{-0.003}^{+0.003}$ & $0.49\pm0.04$ & $281_{-12}^{+48}$ & 5.18 & N & J=4 or 6 due to lack of counterparts and other\\
\multicolumn{8}{c}{ } & lines. J=4 favoured because of excitation\\
1mm.11 & 03:32:37.71 & -27:46:41.0 & $223.067_{-0.025}^{+0.003}$ & $0.27\pm0.02$ & $169_{-12}^{+35}$ & 5.16 & N & lack of other lines suggests J=3\\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{figure*}
\includegraphics[width=0.99\columnwidth]{fig_completeness_Fv_3mm.eps}
\includegraphics[width=0.99\columnwidth]{fig_completeness_Fv_1mm.eps}\\
\caption{Completeness assessment of our line search. In each diagram, each circle represents an artificially injected line candidate. Filled symbols highlight the candidates that we recover in our analysis. The histograms show the marginalization along the y- and x-axis respectively, showing the level of completeness (i.e., the fraction of input line candidates that our script successfully identifies) as a function of the line width ($\Delta v$) and peak flux density ($F_\nu^{\rm line}$), respectively. The 3mm case is shown on the left, the 1mm one is on the right.}
\label{fig_completeness}
\end{figure*}
\subsection{Line candidates}
\subsubsection{Properties}
For our subsequent analysis we consider only those sources that
have a {\em fidelity} of greater than 60\% and where the extracted
line is detected at $>$2$\sigma$ in consecutive channels (width:
$\sim$25\,km\,s$^{-1}${} at 3\,mm, $\sim$40\,km\,s$^{-1}${} at 1\,mm). Our blind
search resulted in 10 line candidates from the 3\,mm search, and 11
line candidates from the 1\,mm search (see Tab.~\ref{tab_lines} and
the figures in the Appendix). Given our requirement on the {\em
fidelity} in our search, we expect that $<$4 out of these line
candidates are spurious in each band. We show the candidates, sorted
by S/N of the line emission, in Figs.~\ref{fig_ps_3mm_a} (band 3)
and~\ref{fig_ps_1mm_a} (band 6). In each case, the left panel shows an
{\it HST} color composite, and the middle panel shows the {\it HST}
image in greyscale, and the CO line candidates in contours. The right
panel shows the spectrum extracted at the position of the line
candidate. The basic parameters of the candidate lines (RA, Dec,
frequency, integrated flux, line width and $S/N$) are summarized in
Tab.~\ref{tab_lines}.
\subsubsection{Optical/NIR counterparts}
We have searched for optical/NIR counterparts by matching the
positions of the sources in the multi--wavelength catalogs (Sec.~2.2)
with our line candidates. Whether a specific CO line candidate has a
counterpart or not is summarized in column~8 of Tab.~\ref{tab_lines}
(see also Figs.~\ref{fig_ps_3mm_a} and~\ref{fig_ps_1mm_a}). The lines
that show an optical/NIR counterpart with matching redshift are
discussed in detail in {\it Paper~IV}.
\subsubsection{Redshift determination}
Given the (almost) equi--distant spacing of the rotational transitions
of CO, it is not straightforward to assign a unique redshift to each
candidate in a number of cases.
\paragraph{Multiple CO lines?}
For certain redshifts, more than one CO transition is covered by our
band~3 and band~6 scans. We use this information to constrain the
redshift of some of the candidate. Likewise, in other cases a certain
redshift solution can be ruled out if other detectable lines are not
detected. This information is given in the `comments' column of
Tab.~\ref{tab_lines}.
\paragraph{Optical/NIR spectroscopic redshifts:}
In some cases, spectroscopic redshifts are available for the
optical/NIR counterparts, either through longslit spectroscopy
\citep{lefevre05,kurk13,skelton14,morris15}, or {\it HST} grism
observations \citep{morris15,momcheva16}. We also record this
information in the `comments' column of Tab.~\ref{tab_lines}.
\paragraph{Lack of optical/NIR counterparts:}
In a number of cases, no optical/NIR counterpart of the line candidate
is visible in the {\it HST} image. This is can be due to the fact
that the source is spurious. But if the candidate was real, and
assuming that there is no signficant reddening by dust, then the
exquisite depth of the available optical/NIR observations (in
particular the {\it HST}/WFC3 IR images and the Spitzer/IRAC images)
can place constraints on the stellar mass of galaxies as a
function of redshift. Our MAGPHYS fits (see Sec.~\ref{sec_magphys}) of
the available photometry suggest that a galaxy securely detected in
H-band (1.6$\mu$m) at $>50$\,nJy (corresponding to a secure,
$>$10-$\sigma$ detection in a few bands) has a stellar mass of
$>$4$\times10^6$\,M$_\odot${}, $>$2$\times10^7$\,M$_\odot${}, and
$>$10$^8$\,M$_\odot${} at $z$=0.5, 1.0, and 2.0, respectively. Because of
the combination of low molecular gas content, and likely
elevated $\alpha_{\rm CO}$ values \citep{bolatto13}, we do not expect
to detect CO in galaxies with $M_*\ll 10^9$\,M$_\odot${}. Therefore
can use the lack of an optical counterpart to set constraints on the
redshift of the candidate. In particular, we assume that line
candidates selected in band~3 and lacking an optical/NIR
counterpart are at $z>2$ (i.e., the line is identified as CO(3-2) or a
higher-J transition). In the case of band~6 candidates, we give
priority to the constraints from the multiple line
(non--)detection. The `lack of counterpart' argument is chosen only to
rule out the lowest--$z$ scenarios (J$<$4, corresponding to
$z<0.695$). This additional constraint on the line candidates is also
given in the `comments' column of Tab.~\ref{tab_lines}.
The total CO flux of all line candidates is 2.55\,Jy\,km\,s$^{-1}$,
whereas the total flux of the candidates that have no optical/NIR
counterpart is 0.83\,Jy\,km\,s$^{-1}$ (from Tab.~\ref{tab_lines}),
i.e. $\sim$33\% of the total. As some of the line candidates that
do not show an optical/IR counterpart are likely spurious, and
considering that the brightest CO detections with optical/NIR
counterparts dominate the total emission, the flux fraction of real
objects without optical/NIR counterpart is likely lower.
\begin{figure*}
\includegraphics[width=0.99\columnwidth]{fig_3mm_cont.eps}
\includegraphics[width=0.99\columnwidth]{fig_1mm_cont.eps}\\
\caption{Continuum images at 3mm ({\em left}) and 1mm ({\em right}). In both panels, we plot a contour at the 3$\sigma$ level, where 1-$\sigma$ is 3.8\,$\mu$Jy\,beam$^{-1}$ in the 3mm observations and 12.7\,$\mu$Jy\,beam$^{-1}$ in the 1mm observations. Both images have been primary-beam corrected. Note that at 3\,mm, only one source is clearly detected at $S/N>3$. The 1mm continuum map is extensively discussed in {\it Paper II}.}
\label{fig_continuum}
\end{figure*}
\subsection{Other CO-- and [C\,{\sc ii}]{}--detected galaxies}
This paper describes our blind search results. An alternative approach
to finding line emission in the galaxies covered by our observations
is to search the cubes at the position of optical galaxies that have
accurate spectroscopic redshifts. Such additional information
(position and redshift) could in principle help to identify plausible
CO or [C\,{\sc ii}]{} emission lines at lower significance than those revealed
by the automatic search. We have performed such a search, which has
resulted in the detection of 3 additional galaxies that are
tentatively detected in CO emission. These detections are presented
and discussed in {\it Paper~IV}. We note that the inclusion of these 3
galaxies would not change the statistical analysis based on the much
larger sample presented here. Similarly, in {\it Paper~V} we
investigate the presence of [C\,{\sc ii}]{} emission in galaxies for which a
photometric redshift from SED fitting or the detection of a clear
drop-out in the $z$ band suggests redshifts $z>6$.
\subsection{Continuum emission}\label{sec_continuum}
The frequency scans can be used to obtain very high--sensitivity maps
of the continuum, by collapsing the two data cubes along the frequency
axis, after removing the few channels that contain significant line
emission. The resulting continuum maps with noise levels at their
center of 3.8\,$\mu$Jy\,beam$^{-1}$ (band~3) and
12.7\,$\mu$Jy\,beam$^{-1}$ (band 6) are shown in
Fig.~\ref{fig_continuum} and will be discussed in detail in {\it
Paper~II}.
\section{Comparison with expectations}\label{sec_magphys}
We present a detailed comparison of the evolution of the CO luminosity
functions, and the resulting cosmic density of molecular hydrogen in
{\it Paper~III}. As a sanity check, we here briefly compare the number
of CO--detected galaxies with previous expectations based on a
multi--wavelength analysis of the galaxies in the UDF.
For each galaxy in the UDF, \citet{dacunha13} estimated stellar
masses, SFRs, IR luminosities, and expected CO and [C\,{\sc ii}]{} fluxes and
luminosities by fitting the optical/NIR photometry provided by
\citet{coe06}, using the MAGPHYS spectral energy fitting code
\citep{dacunha08}. We show the resulting star formation rates, and
stellar masses, in four redshift bins in the right hand panel of
Fig.~\ref{fig_pointings}. Note that typical selections of main
sequence galaxies for CO follow--up usually target stellar masses
M$_{\rm star}>10^{10}$\,M$_\odot$ and star formation rates
SFR$>50$\,M$_\odot$\,yr$^{-1}$ (e.g. Daddi et al. 2008, Tacconi et
al. 2008, Genzel et al.\ 2008, Tacconi et al.\ 2012, Daddi et al.\
2015, Genzel et al.\ 2015). I.e. this selection would target galaxies
in the top right part of each diagram, as the UDF contains many
galaxies that are much less massive / star forming.
In Fig.~\ref{fig_magphys} we show the expected numbers of line
detections in the 3\,mm and 1\,mm bands, respectively. In this plot,
the expected number of lines from \citet{dacunha13}, originally
computed for the entire $3'\times3'$UDF, has been scaled to the areal
coverage of our survey. In \citet{dacunha13}, two extreme CO
excitation cases were considered in order to transform predicated
CO(1--0) luminosities into higher--J line luminosities: the
low--excitation case of the global Milky Way disk, and the
high--excitation case of the nucleus of the local starburst galaxy M82
\citep{weiss07}. For each line flux plotted on the abscissa, this
range of excitation conditions is indicated by the grey region on the
ordinate.
In this figure, we compare to our observations, which are plotted as
red--shaded regions. For each flux bin on the abscissa, the number
counts with the Poissonian error bars are shown on the ordinate. For
this back--of--the envelope calculation, we do not correct our
measurements for completeness or fidelity (this is done in detail in
{\it Paper~III}). A number of things need to be kept in mind in this
comparison: the total number of detected sources is low, which results
in large uncertainties in the measurements on the ordinate. At 1\,mm,
the data in the highest flux bin (around 1\,mJy\,km\,s$^{-1}$) is
significantly higher than the predictions. Note however that
measurement includes the two high--J CO detections of ASPECS~1mm.1/2,
a galaxy that was not included in the UDF catalog on which the
predictions by \citet{dacunha13} were based. Larger areas are
required to see if there is indeed an excess of high--J CO emission
present. Overall, we conclude that within the large uncertainties,
there is reasonable agreement between the observations and previous
expectations. This is discussed in detail in {\it Paper~III}.
\begin{figure}
\includegraphics[width=0.99\columnwidth]{fig_magphys_pred.eps}\\
\caption{Comparison between the MAGPHYS-based predictions of line fluxes from \citet{dacunha13}, in grey, and
the flux distribution of the line candidates actually observed in our survey (red boxes). The numbers from \citet{dacunha13}
are computed over the whole UDF, and scaled down to match the same area coverage of our survey. We consider here only the transitions
that we cover in our scan (see Tab.~\ref{tab_z_range}). The lower and upper sides of the shaded grey area refer the cases of
Milky Way- and M82-like CO excitation. In the case of our ALMA constraints, the vertical size of the boxes show the Poissonian
uncertainties in the number of lines detected in a certain flux range. Our ALMA constraints are not corrected for the fidelity
and completeness of our line search. The number of detected lines is in general agreement with the expectations, in particular if one keeps in mind that ASPECS~1mm.1/2, whose high--J CO emission dominates the highest flux bin at 1mm, was not included in the \citet{dacunha13} study.}
\label{fig_magphys}
\end{figure}
\section{Summary}
We present the rationale for and the observational description of
ASPECS, our complete band~3 and band~6 spectral line scan with ALMA of
the {\it Hubble} Ultra--Deep Field (UDF). This field was chosen
because it has the deepest multi--wavelength data available, it will
remain a key cosmological deep field in the future (in particular in
the era of JWST) and is easily observable by ALMA. We discuss our
survey design of the full frequency scans in band~3 (84--115\,GHz) and
band~6 (212--272\,GHz) and report the relevant parameters of our final
dataset. Critically, ALMA allows us to reach approximately uniform
depth (line sensitivity: $\sim L'_{\rm CO}\sim 2 \times
10^{9}$\,K~km\,s$^{-1}$\,pc$^2$) across a broad range of redshifts.
The spectral line scans cover the different rotational transitions of
the CO molecule at different redshifts, leading to essentially full
redshift coverage. We present a customized algorithm to identify line
candidates in our data. This algorithm takes varying linewidths of the
possible emission lines into account. We assess the {\rm fidelity} of
our line search by comparing the number of positive candidates to the
respective number of negative candidates, the latter being unphysical.
We also calculate the {\em completeness} of our search, by quantifying
our ability to recover artificial sources in our data. We present CO
spectra and {\it HST} postage stamps of the most signficant
detections. Based on whether multiple CO lines are detected, and
whether optical spectroscopic (either slit or grism) redshifts as well
as optical/NIR counterparts exist, we give constraints on the most
likely line identification of our candidates.
Out of the 10 line candidates (3mm band) reported in our search
(Tab.~\ref{tab_lines}), we expect $<$4 candidates to be spurious,
given our statistical analysis. There are a number of line candidates
at positions where no optical/NIR counterpart is present. The total CO
flux of these candidates is less than 33\% of the total flux of all
candidates, i.e. candidate sources without counterparts only
contribute a small fraction of the total measured flux in the targeted
field. We also present continuum maps of both the band~3 and band~6
observations. The observed flux distribution of the line candidates is
in general agreement with the empirical expectations by
\citet{dacunha13} based on SED modeling of the optical/NIR emission of
galaxies in the UDF.
The data presented in this paper ({\it Paper~I}) form the basis of a
number of dedicated studies presented in subsequent papers:
\noindent $\bullet$ In {\it Paper~II} (Aravena et al.\ 2016a) we
present 1.2\,mm continuum number counts, dust properties of individual
galaxies, and demonstrate that our observations
recover the cosmic infrared background at the wavelengths considered.
\noindent $\bullet$ In {\it Paper~III} (Decarli et al.\ 2016a) we
discuss the implications for CO luminosity functions and the resulting
constraints on the gas density history of the Universe. Based on our
data we show that there is a sharp decrease (by a factor of $\sim$5)
in the cosmic molecular gas density from redshift $\sim$ 3 to 0.
\noindent $\bullet$ In {\it Paper~IV} (Decarli et al.\ 2016b) we
examine the properties of those galaxies in the UDF that show bright
CO emission, and discuss these also in the context of the bright
optical galaxies that are not detected in CO.
\noindent $\bullet$ In {\it Paper~V} (Aravena et al.\ 2016b) we search
for [C\,{\sc ii}]{} emitters in previously reported Lyman--break galaxies at
6$<$z$<$8.
\noindent $\bullet$ In {\it Paper VI} (Bouwens et al.\ 2016) we
investigate where high--redshift galaxies from ASPECS lie in relation
to known IRX--$\beta$ and IRX--stellar mass relationships, concluding
that less dust continuum emission is detected in z$>$2.5 than expected
(unless high dust temperatures, T$\sim$50\,K, are assumed).
\noindent $\bullet$ Finally, in {\it Paper VII} (Carilli et al.\ 2016)
we discuss implications on CO intensity mapping experiments, and
contributions towards the emission from the cosmic microwave
background.
The data presented here demonstrate the unique power of ALMA spectral
scans in well--studied cosmological deep fields. The current size of
the survey is admittedly small, limited by the amount of time
available in ALMA `early science'. More substantial spectral scan
surveys with ALMA of the full UDF (and beyond) will become feasible
once ALMA is fully operational.
\section*{Acknowledgments}
We thank the referee for a constructive report that helped improve the
presentation of the data. FW, IRS, and RJI acknowledge support through
ERC grants COSMIC--DAWN, DUSTYGAL, and COSMICISM, respectively. MA
acknowledges partial support from FONDECYT through grant 1140099. DR
acknowledges support from the National Science Foundation under grant
number AST-\#1614213 to Cornell University. FEB and LI acknowledge
Conicyt grants Basal-CATA PFB--06/2007 and Anilo ACT1417. FEB also
acknowledge support from FONDECYT Regular 1141218 (FEB), and the
Ministry of Economy, Development, and Tourism's Millennium Science
Initiative through grant IC120009, awarded to The Millennium Institute
of Astrophysics, MAS. IRS also acknowledges support from STFC
(ST/L00075X/1) and a Royal Society / Wolfson Merit award. Support for
RD and BM was provided by the DFG priority program 1573 `The physics
of the interstellar medium'. AK and FB acknowledge support by the
Collaborative Research Council 956, sub-project A1, funded by the
Deutsche Forschungsgemeinschaft (DFG). PI acknowledges Conict grants
Basal-CATA PFB--06/2007 and Anilo ACT1417. RJA was supported by
FONDECYT grant number 1151408. This paper makes use of the following
ALMA data: 2013.1.00146.S and 2013.1.00718.S. ALMA is a partnership of
ESO (representing its member states), NSF (USA) and NINS (Japan),
together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic
of Korea), in cooperation with the Republic of Chile. The Joint ALMA
Observatory is operated by ESO, AUI/NRAO and NAOJ. The 3mm-part of
ALMA project had been supported by the German ARC.
|
\section{Introduction}
What enables latent variable discovery is the particular probabilistic dependencies between variables, will typically be representable only by a proper subset of the possible causal models over those variables, and therefore provide evidence in favour of those models and against all the remaining models, as can be
seen in the Bayes factor. Some dependency structures between observed variables will provide evidence favoring latent variable models over fully observed models,\footnote{I.e., models all of whose variables are measured or observed.} because they can explain the dependencies better than any fully observed model. We call such structures
``triggers'' and did a systematically search for them. The result is a clutch of triggers, many of which have not been reported before to our knowledge. These triggers can be
used as a data preprocessing analysis by the main discovery algorithm.
\subsection{Latent Variable Discovery}
Latent variable modeling has a long tradition in causal discovery, beginning
with Spearman's work~\cite{Spear1904} on intelligence
testing. Factor analysis and related methods can be used to posit
latent variables and measure their hypothetical effects. They do not
provide clear means of deciding whether or not latent variables are present
in the first place, however, and in consequence there has been some
controversy about that status of exploratory versus confirmatory
factor analysis. In this regard, causal
discovery methods in AI have the advantage.
One way in which discovery algorithms may find evidence confirmatory
of a latent variable model is in the greater simplicity of such a
model relative to any fully observed model that can represent the data
adequately, as Friedman pointed out using the example in Figure~\ref{fig:friedman}~\cite{friedman1997learning}.
\begin{figure}[H]
\begin{center}
\subfigure{
\label{(a)}
\includegraphics[width=0.16\textwidth]{no_latent_model.jpg}
} \qquad \qquad
\subfigure{
\label{(b)}
\includegraphics[width=0.16\textwidth]{with_latent_model.jpg}
}
\end{center}
\caption{An illustration of how introducing a latent variable $H$ can
simplify a model~\cite{friedman1997learning}.}
\label{fig:friedman}
\end{figure}
Another advantage for latent variable models is that they can better encode the actual dependencies and independencies in the data. For
example, Figure~\ref{fig:hidden} demonstrates a latent variable model of four
observed variables and one latent variable. If the data support the
independencies $W ~\raisebox{-0.3ex}{\rotatebox{90}{\ensuremath{\models}}}~\{Y,Z\}$ and $Z ~\raisebox{-0.3ex}{\rotatebox{90}{\ensuremath{\models}}}~\{W,X\}$, it is
impossible to construct a network in the observed variables alone that
reflects both of these independencies while also reflecting the
dependencies implied by the d-connections in the latent variable model. It is
this kind of structure which can allow us to infer the existence of
latent variables, i.e., one which constitutes a trigger for latent
variable discovery.
\begin{figure}[H]
\centering
\includegraphics[width=0.25\textwidth]{oneLatentFourObserved.jpg}
\caption{One causal structure with four observed variables and one
latent variable $H$.}
\label{fig:hidden}
\end{figure}
\section{Searching Triggers for Latent Variables}
In this paper, latent variables are typically considered only in scenarios where they
are common causes which having two or more children. As Friedman~\cite{friedman1997learning} points out, a latent variable as a leaf or as a root with only one child would marginalize out without
affecting the distribution over the remaining variables. So too would
a latent variable that mediates only one parent and one child. For simplicity, we also only search for triggers for isolated latent variables rather than multiple latent variables.
We start by enumerating all possible fully observed DAGs in a given number of variables (this step is already super exponential~\cite{robinson1977counting}!. Then it generates all possible d-separating evidence sets. For example, for the four variables $W, X, Y$ and $Z$,
there are eleven evidence sets:\footnote{Note that sets of three or more evidence variables leave nothing left over to separate.} \[ \phi, \{W\}, \{X\}, \{Y\}, \{Z\},
\{WX\},\] \[\{WY\}, \{WZ\}, \{XY\}, \\ \{XZ\}, \{YZ\}\]
For each fully observed DAG it produces the corresponding dependencies for each
evidence set using the d-separation criterion (i.e., for the four
variables $W, X, Y$ and $Z$, the search produces eleven dependency
matrices). Next, it generates all possible single hidden-variable
models whose latent variable is a common cause of two observed variables. It
then generates all the dependencies between observed variables
for each latent variable model, conditioned upon each evidence
set. The set of dependencies of a latent variable model is a {\bf
trigger} if and only if these dependency sets cannot be matched by
any fully observed DAG in terms of d-separation.\footnote{In this search, labels (variable
names) are ignored, of course, since all that matters are the
dependency structures.}
We ran our search for 3, 4 and 5 observed variables. Any structures
with isolated nodes are not be included. As Table~\ref{tab:number} shows, for three observed variables, we find no trigger, meaning the set of dependencies implied by all possible hidden models can also be found in one or more fully observed models. There are two triggers for four observed variables, the corresponding DAGs are shown in Table~\ref{tab:4var},
together with the corresponding latent variable models. For five
observed variables, we find 57 triggers (see Appendix~\ref{app:triggers}).
\begin{table} [H]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\pbox{10cm}{Observed variables} & \quad \small{DAGs} \quad & \quad \pbox{20cm}{Connected DAGs} \quad & \quad \small{Triggers} \quad \\
\hline
3 & 6 & 4 &0 \\
4 & 31 & 24 &2 \\
5 & 302 & 268 &57 \\
\hline
\end{tabular}
\end{center}
\caption{Number of triggers found}
\label{tab:number}
\end{table}
\begin{table} [H]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.18\textwidth, height=15mm]{hidden_model1.jpg} &
\includegraphics[width=0.18\textwidth, height=15mm]{hidden_model2.jpg} \\
\end{tabular}
\end{center}
\caption{The DAGs of the two triggers found for four observed variables.}
\label{tab:4var}
\end{table}
All the dependency structures in the observed variables revealed as
triggers will be better explained with latent variables than
without. While it is not necessary to take triggers into account
explicitly in latent variable discovery, since random structural
mutations combined with standard metrics may well find them, they can
be used to advantage in the discovery process, by focusing it, making
it more efficient and more likely to find the right structure.
\section{Learning Triggers with Causal Discovery Algorithms}
The most popular causal discovery programs in
general, come from the Carnegie Mellon group and are incorporated into
TETRAD, namely FCI and PC~\cite{tetrad}. Hence, they are the natural
foil against which to compare anything we might produce. Our ultimate
goal is to incorporate latent variable discovery
into a metric-based program. As that is a longer project, here we
report experimental results using an ad hoc arrangement of running the
trigger program as a filter to ordinary PC (Trigger-PC) and comparing
that with the unaltered FCI and PC. Our experimental procedure, briefly, was:
\begin{enumerate}
\item Generate random networks of a given number of variables, with three categories of dependency: weak, medium and strong.
\item Generate artificial data sets using these networks.
\item Optimize the alpha level of the FCI and PC programs using the above.
\item Experimentally test and compare FCI and PC.
\end{enumerate}
The FCI and PC algorithms do not generally return a single DAG, but a
hybrid graph~\cite{spirtes2000causation}. An arc between two nodes may be undirected '\textemdash' or bi-directional '$\leftrightarrow$', which indicates the presence of a latent common cause. Additionally, the graph produced by FCI may contain
'o\textemdash o' or 'o$\rightarrow$'. The circle represents an
unknown relationship, which means it is not known whether or not an
arrowhead occurs at that end of the arc~\cite{spirtes2000causation}. So, in order to measure how close the models learned by FCI and PC are to the true model, we developed a
special version of the edit distance between graphs (see the Appendix~\ref{editdistanc}).
\subsection{Step one: generate networks with different dependency strengths.}
Genetic algorithms (GAs)~\cite{Russel2003} are commonly applied as a
search algorithm based on an artificial selection process that simulates
biological evolution. Here we used a GA algorithm to find good representative, but
random, graphs with the three levels of desired dependency strengths
between variables: strong, medium and weak. The idea is to test the learning algorithms across different degrees of difficulty in recovering arcs (easy, medium and difficult, respectively). Mutual information~\cite{pearl1988probabilistic} is used to assess the
strengths of individual arcs in networks.
To make the learning process more efficient, we set the arities for
all nodes in a network to be the same, either two or three. We
randomly initialized all variables' CPT parameters for each individual
graph and used a whole population of 100 individuals. The GA was run
100 generations. We ran the GA for each configuration (number of nodes
and arities) three times, the first two to obtain networks with the
strongest and weakest dependencies between parents and their children
and the third time to obtain networks closest to the average of those
two degrees of strength.
\subsection{Step two: generate artificial datasets.}
All networks with different arc strength levels were used to generate
artificial datasets with sample sizes of 100, 1000 and 10000. We used Netica API~\cite{neticaAPI} to generate random cases. The default
sampling method is called ``Forward Sampling"~\cite{neticasampling}
which is what we used.
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\pbox{10cm}{Number of \\ observed variables} & \quad \pbox{2cm}{Structure type} \quad & \quad \pbox{3cm}{Number of \\ structures} \quad & \pbox{5cm}{Total number \\ of simulated \\ datasets} \quad \\
\hline
4 &Trigger & 2 & 36 \\
4 &DAG & 24 & 432 \\
5 &Trigger & 57 & 1026 \\
5 &DAG & 268 & 4824 \\
\hline
\end{tabular}
\end{center}
\caption{Number of simulated datasets}
\label{tab:simulateddatano}
\end{table}
As Table~\ref{tab:simulateddatano} shows, there are a relatively large
number of simulated datasets. This is due to the different state
numbers, arc strength levels and data sizes. For example, there are 57
trigger structures for 5 observed variables, so there are $57 \times
2 \times 3 \times 3 = 1026$ simulated datasets.
\subsection{Step three: optimize alpha to obtain the shortest edit distance from true models}
FCI and PC both rely on statistical significance tests to decide whether an arc exists between two variables and on its orientation. They have a default alpha level (of 0.05), but the authors have in the past criticized experimental work
using the default and recommended instead optimizing the alpha level
for the task at hand, so here we do that. The idea is to give the
performance of FCI and PC the benefit of any possible doubt. Given the
artificial data sets generated, we can use FCI and PC with different
values of alpha to learn networks and compare the results to the
models used to generate those data sets. We then used our version of
edit distance between the learned and generating models (see Appendix~\ref{editdistanc}) to find the optimal alpha levels for both algorithms.
We first tried simulated annealing to search for an optimal alpha, but
in the end simply generated sufficiently many random values from the
uniform distribution over the range of [0.0, 0.5]. We evaluated alpha
values for the datasets with 2 states and 3 states separately.
As shown in the following graphs, the average edit distances
between the learned and true models approximate a parabola with a
minimum around 0.1 to 0.2. The results below are specific to
the exact datasets and networks we developed for this experimental
work.
\noindent
1) FCI algorithm
Results for FCI were broadly similar. In summary, the optimal alphas found for the above cases (in the same
order) were: 0.12206, 0.19397, 0.20862 and 0.12627.
\begin{itemize}
\item Number of observed variables: 4 \\
Datasets: 2 state DAG structure simulated dataset \\
Number of datasets: 24*9 = 216
\begin{figure} [h]
\centering
\includegraphics[width=0.4\textwidth]{fci4var2state_alpha.png}
\label{fig:fci4var2state}
\end{figure}
Results: \\
Minimum average edit distance: 14.85185 \\
Maximum average edit distance: 15.82870 \\
Mean average edit distance: 15.37168 \\
Best Alpha: 0.12206 \\
95\% confidence level: 0.03242 \\
95\% confidence interval: (15.37168-0.03242, 15.37168+0.03242) \\
\item Number of observed variables: 4 \\
Datasets: 3 state DAG structure simulated dataset \\
Number of datasets: 24*9 = 216
\begin{figure} [H]
\centering
\includegraphics[width=0.4\textwidth]{fci4var3state_alpha.png}
\label{fig:fci4var3state}
\end{figure}
Results: \\
Minimum average edit distance: 11.42593 \\
Maximum average edit distance: 13.43981 \\
Mean average edit distance: 12.09355 \\
Best Alpha: 0.19397 \\
95\% confidence level: 0.07259 \\
95\% confidence interval: (12.09355-0.07259, 12.09355+0.07259) \\
\item Number of observed variables: 5 \\
Datasets: 2 state DAG structure simulated dataset \\
Number of datasets: 268*9 = 2412
\begin{figure} [H]
\centering
\includegraphics[width=0.4\textwidth]{fci5var2state_alpha.png}
\label{fig:fci5var2state}
\end{figure}
Results: \\
Minimum average edit distance: 23.72844 \\
Maximum average edit distance: 24.98466 \\
Mean average edit distance: 24.08980 \\
Best Alpha: 0.20862 \\
95\% confidence level: 0.03880 \\
95\% confidence interval: (24.08980-0.03880, 24.08980-0.03880) \\
\item Number of observed variables: 5 \\
Datasets: 3 state DAG structure simulated dataset \\
Number of datasets: 268*9 = 2412
\begin{figure} [H]
\centering
\includegraphics[width=0.4\textwidth]{fci5var3state_alpha.png}
\label{fig:fci5var3state}
\end{figure}
Results: \\
Minimum average edit distance: 18.66501 \\
Maximum average edit distance: 20.89221 \\
Mean average edit distance: 19.32648 \\
Best Alpha: 0.12627 \\
95\% confidence level: 0.08906 \\
95\% confidence interval: (19.32648-0.08906, 19.32648-0.08906) \\
\end{itemize}
\noindent
2) PC algorithm
Results for PC were quite similar. In
summary, the optimal alphas found for the above cases (in the same
order) were: 0.12268, 0.20160, 0.20676 and 0.13636.
\subsection{Step four: compare the learned models with true model.}
Finally, we were ready to test FCI and PC on the artificial datasets
of trigger (i.e., latent variable) and DAG structures. Artificial
datasets generated by trigger structures were used to determine True
Positive (TP) and False Negative (FN) results (i.e., finding the real
latent and missing the real latent, respectively), while the datasets
of (fully observed) DAG structures were used for False Positive (FP)
and True Negative (TN) results. Assume the latent variable in every
trigger structure is the parent of node A and B, we used the following
definitions:
\begin{itemize}
\item TP: The learned model has a bi-directional arc between A and B.
\item FN: The learned model lacks a bi-directional arc between A and B.
\item TN: The learned model has no bi-directional arcs.
\item FP: The learned model has one or more bi-directional arcs.
\end{itemize}
We tested the FCI and PC algorithms on different datasets with their
corresponding optimized alphas. We do not report confidence intervals
or significance tests between different algorithms under different
conditions, since the cumulative results over 6,318 datasets suffices
to tell the comparative story.
The following tables show the confusion matrix summing over all
datasets (see Appendix~\ref{app:conmatrix} for more detailed results):
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
& \multicolumn{2}{|c|}{FCI} & \multicolumn{2}{|c|}{PC}\\
\hline
& \quad Latent &\quad No Latent &\quad Latent &\quad No Latent\\
\hline
Positive &235 &981 &226 &819 \\
\hline
Negative &827 &4275 &836 &4437 \\
\hline
\end{tabular}
\end{center}
\label{tab:confusionmatrix}
\end{table}
With corresponding optimal alpha, the FCI's predictive accuracy was
0.71 (rounding off), its precision 0.19, its recall 0.22 and its false
positive rate 0.19. The predictive accuracy for PC was 0.74, the
precision was 0.22, its recall 0.21 and the false positive rate
was 0.16.
We also did the same tests using the default alpha of 0.05. The
results are shown as follow (see Appendix~\ref{app:conmatrix} for more detailed
results):
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
& \multicolumn{2}{|c|}{FCI} & \multicolumn{2}{|c|}{PC}\\
\hline
& \quad Latent &\quad No Latent &\quad Latent &\quad No Latent\\
\hline
Positive &211 &767 &205 &615 \\
\hline
Negative &851 &4489 &857 &4641 \\
\hline
\end{tabular}
\end{center}
\label{tab:confusionmatrix2}
\end{table}
With alpha of 0.05, the FCI's predictive accuracy was 0.74, its
precision 0.22, its recall 0.19 and its false positive rate 0.15.
PC's predictive accuracy was 0.77, the precision was 0.25, its
recall 0.19 and the false positive rate was 0.12.
As we can see from the results, the performance of FCI and PC are
quite similar. Neither are finding the majority of latent variables
actually there, but both are at least showing moderate false positive
rates. Arguably, false positives are a worse offence than false
negatives, since false negatives leave the causal discovery process no
worse off than an algorithm that ignores latents, whereas a false
positive will positively mislead the causal discovery process.
\section{Applying Triggers in Causal Discovery}
\subsection{An extension of PC algorithm (Trigger-PC)}
We implemented triggers as a data filter into PC, yielding Trigger-PC, and see how well it would work. If Trigger-PC finds a trigger pattern in the data, then it returns that trigger
structure, otherwise it returns whatever structure the PC algorithm
returns, while replacing any incorrect bi-directed arcs by undirected
arcs. So the Trigger-PC algorithm (see Algorithm~\ref{alg:Trigger-PC})
gives us a more specific latent model structure and, as we shall see,
has fewer false positives.
\begin{algorithm}[ht]
\caption{Trigger-PC Algorithm}
\label{alg:Trigger-PC}
\begin{algorithmic}[1]
\State Let $D$ be the test dataset;
\State Let $D\_labels$ be the variable labels in $D$;
\State Let $Triggers$ be the triggers given the number of variables in $D$;
\State Perform conditional significant tests to get the dependency pattern $P$ in $D$;
\State Let $matchTrigger = false$;
\State Let $T$ be an empty DAG;
\State Let $label\_assignments$ be all possible label assignments of $D\_label$;
\For{each $trigger$ in $Triggers$}
\State Let $t$ be the unlabeled DAG represented by $trigger$;
\For{each $label\_assignment$ in $label\_assignments$}
\State Assign $label\_assignment$ to $t$, yield $t\_labeled$;
\State Generate dependency pattern of $t\_labeled$, yield $t\_pattern$;
\If{$P$ matches $t\_pattern$}
\State $matchTrigger = true$;
\State $T = t\_labeled$;
\State break;
\EndIf
\EndFor
\If{$matchTrigger = true$}
\State break;
\EndIf
\EndFor
\If{$matchTrigger = true$}
\State Output $T$;
\EndIf
\If{$matchTrigger = false$}
\State Run PC Algorithm with input dataset $D$;
\State Let $PC\_result$ be the result structure produced by PC Algorithm;
\If{there is any bi-directed arcs in $PC\_result$}
\State Replace all bi-directed arcs by undirected arcs, yield $PC\_result*$;
\State Output $PC\_result*$;
\EndIf
\EndIf
\end{algorithmic}
\end{algorithm}
We tested our Trigger-PC algorithm with the alpha optimized for the PC
algorithm. The resultant confusion matrix was (see Appendix~\ref{app:conmatrix} for more details):
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
& \multicolumn{2}{|c|}{trigger-PC}\\
\hline
& \quad Latent &\quad No Latent \\
\hline
Positive &30 &3 \\
\hline
Negative &1032 &5253 \\
\hline
\end{tabular}
\end{center}
\label{tab:triggerpc-confusionmatrix}
\end{table}
Trigger-PC's predictive accuracy was 0.84, its precision 0.91, its recall 0.03 and the false positive rate 0.0006. We can see that Trigger-PC is finding far fewer latents than either PC or FCI, but when it asserts their existence we can have much greater confidence in the claim. As we indicated above, avoiding false positives, while having at least some true positives, appears to be the more important goal in latent variable discovery.
As before, we also tried the default 0.05 alpha in Trigger-PC, with the results (see mode details in Appendix~\ref{app:conmatrix}):
\begin{table}[H]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
& \multicolumn{2}{|c|}{trigger-PC}\\
\hline
& \quad Latent &\quad No Latent \\
\hline
Positive &35 &4 \\
\hline
Negative &1027 &5252 \\
\hline
\end{tabular}
\end{center}
\label{tab:triggerpc-confusionmatrix2}
\end{table}
And, again, these results are only slightly different.
With alpha of 0.05, the trigger-PC's predictive accuracy was 0.84, its precision 0.90, its recall 0.03 and the false positive rate 0.0008.
\section{Conclusion}
We have presented the first systematic search algorithm to discover
and report latent variable triggers: conditional probability
structures that are better explained by latent variable models than by
any DAG constructed from the observed variables alone. For simplicity
and efficiency, we have limited this to looking for single latent
variables at a time, although that restriction can be removed. We
have also applied this latent discovery algorithm directly in an
existing causal discovery algorithm and compared the results to
existing algorithms which discover latents using different
methods. The results are certainly different and arguably superior.
Our next step will be to implement this approach within a metric-based
causal discovery program.
\bibliographystyle{ieeetran}
\section*{Module \arabic{modulenum}:~{#2}}
\addtocounter{modulenum}{1}
}
\renewcommand{\floatpagefraction}{0.7}
\newcommand{\cs}[1]{\mbox{c}_{#1}}
\newcommand{\sn}[1]{\mbox{s}_{#1}}
\newcommand{\threevec}[3]{
\left[
\begin{array}{c} #1 \\ #2 \\ #3 \end{array}
\right]
}
\newcommand{\hvec}[4]{
\left[
\begin{array}{c} #1 \\ #2 \\ #3 \\ #4 \end{array}
\right]
}
\newcommand{\qvec}[1]{
\left[
\begin{array}{c} 0 \\ #1 \end{array}
\right]
}
\newcommand{\newspecial}[1]{
\ifnum\includefigs>0
\special{#1}
\fi
}
\newcommand{\slidehead}[2][2ex]{%
\slideheading{#2}\vspace{#1}
}
\newcommand{\sh}[1]{\slideheading{#1}}
\newcommand{\subhead}[3]{\vspace{#1}\begin{center}{\em #3}\end{center}\vspace{#2}}
\def\makeslideheading#1{%
\begin{center}\Large\bf #1\end{center}}
\newtheorem{prop}{Proposition}
\newtheorem{defn}{Definition}
\newtheorem{result}{Result}
\newtheorem{conject}{Conjecture}
\newtheorem{rulex}{Rule}
\newtheorem{princ}{Principle}
\newenvironment{prf}{\rm}{}
\newcommand{\standardpagesetup}{
\addtolength{\textheight}{55mm}
\addtolength{\textwidth}{25mm}
\addtolength{\voffset}{-31mm}
\addtolength{\marginparwidth}{30mm}
\oddsidemargin = 0 mm
\topmargin = 0 mm
\parskip = 2 mm
\textfloatsep = 0 mm
\intextsep = 0 mm
}
\newcommand{\pdfpagesetup}{
\addtolength{\textheight}{40mm}
\addtolength{\textwidth}{25mm}
\addtolength{\voffset}{-31mm}
\addtolength{\marginparwidth}{30mm}
\oddsidemargin = 0 mm
\topmargin = 20 mm
\parindent = 6 mm
\textfloatsep = 0 mm
\intextsep = 0 mm
}
\newcommand{\afour}{
\paperheight 297mm
\paperwidth 210mm
\textheight 247mm
\textwidth 160mm
\oddsidemargin -5mm
\topmargin -15mm
}
|
\section{Introduction}\label{intro}
Harmonic oscillator (HO) is a successful model regarding its quantization, both canonical and geometrical one, being an example of solvability.
The theoretical study of HO occupy a central place in quantum mechanics, one can find a range of its applications in the real world of atomic and molecular systems.
The potential function is the key element, that defines the evolution of a system.
Only some functions have proven to be mathematically reliable and fewer meet the conditions to describe a physical system.
Exactly solvable potentials, in quantum mechanics, have been proposed starting with the early mid of the last century and continue to be an interesting topic.
Beside hydrogen atom, the other fundamental quantum mechanic example of solvable system, there are other solvable physical systems, described by potentials, known as: the Morse potential \cite{2}, the Eckart potential \cite{3}, the Rosen-Morse potential \cite{4}, the trigonometric and hyperbolic Pöschl-Teller potentials \cite{5}, the Manning-Rosen potential \cite{6}, the Woods-Saxon potential \cite{7}, the Scarf potential \cite{8} and the pseudo-Gaussian potential \cite{pgo}.
In this paper, we introduce a class of basic functions which define solvable potentials, for certain physical systems, that can be reduced to the HO potential by some elementary transformations and, as a matter of fact, admits geometric quantization (GQ) procedure.
The study is made upon the geometrical properties of phase space of physical systems from the perspective of both the Hamiltonian functions, this is the foliation perspective, and the flow of the vector fields, this is the dynamical system perspective.
Mainly, critical points of the system, known as singularities, are investigated in order to find the conditions of integrability and thereby the dynamics of the system.
Our reasoning is based on famous results of both Morse lemma, regarding the study of critical points of Hamiltonian function, and Eliasson theorem, which gives the integrability conditions.
We conclude by a proposition that gives the sufficient conditions needed by a physical system in order to be integrable in HO sense, this means to apply GQ procedure (given by Simms \cite{Simm}) used to the quantization of energy states of HO.
In our approach, we make use of GQ because, as it is known, it is a mathematically rigorous global generalization of the canonical quantization (CQ) technique.
As a matter of fact GQ is suited for our approach, which is a generalization for an entire class, because there is no need for a specific physical system to be pointed out.
Having this in mind, we start with a general form for potential, namely $V$, showing which are the requirements that have to be accomplished by it in order to become the HO potential.
The classical systems are described using the technique of symplectic manifolds, which provides the adequate Hamiltonian formulation of autonomous mechanics (the time-dependent systems are not considered in this work).
When GQ, based on symplectic manifolds, takes into account the phase space geometry of the classical physical system and aims to find a quantum counterpart, one can say that GQ follows the so called Dirac program \cite{AM}.
Methods of GQ have been applied with great success to the theory of representation of Lie groups \cite{Kon}, however, its usefulness in applications to quantum theory has been rather limited.
At least in this terms the cornerstones of quantum mechanics, the harmonic oscillator and the hydrogen atom, GQ gives results in agreement with those of CQ \cite{K,AK,S,Simm,Kumm,Cord}.
The geometric quantization achievements are still under those accomplished by canonical quantization. In this view, every new example contributes to the edification of GQ.
Being accepted at this level, GQ can further show its usefulness when one needs to take into account the topological and geometrical structure of the classical phase space, for instance in the case of Gravitation \cite{Gr}.
Particular systems that can be reduced to the quantization scheme of HO have already been reported, we notice here the nonisotropic harmonic oscillator \cite{MM}, the rigid body with a single rotor about the third principal axis and an internal torque \cite{Puta}, the Kepler problem obtained from HO using a Segre map \cite{GS} and the symmetric rigid body \cite{I}.
In quantum mechanics finding new solvable potentials is a question of central interest and a subject of long time studies. We know only a handful of classical potentials for which we can determine explicitly all the bound states in terms of elementary transcendental functions and to write explicitly their energies in terms of quantum numbers.
We hope that contribution of this work to GQ theory illustrates the general theory with new specific examples and with the same extent brings out the properties required by the potential of certain physical systems in order to be integrable by admitting GQ procedure.
In quantum physics the Schr\" odinger picture (or representation) is a formulation where the physical states evolve in time. Opposite, we can think the states remain constant and the operators (observables) evolve in time. Physically this is the difference between active and passive viewpoint. In GQ the representation is given by polarization. As it is usual, we call Schr\" odinger representation that one were the polarization is spanned by the vector field determined by position coordinate. The dynamics of the system is given by Schr\" odinger equation (SE), a second order partial differential equation, which describes the evolution of the states. The potential function is part of the differential operator, acting on Hilbert space, called Hamiltonian.
An extensive study on Schr\" odinger equations reducible to other hypergeometric or rescaled confluent equation can be consulted in ref. \cite{fluegge, AHP, milson}.
The rest of this paper is organized as follows:
In the first section we present some aspects of both GQ and CQ focusing on what it is considered an achievement in each of them. The GQ steps are mentioned along with Schr\" odinger representation as it comes naturally from phase space of the physical system.
In the second section, the behavior of physical system is studied through its critical points, also known as stationary points, which define how states evolve. The study is made from the perspective of both potential function (passive mode) and Hamiltonian vector field (dynamic mode), thus we split this section in two subsections. In the former, with the help of Morse lemma, we seek mathematical aspects of a class of potentials that can be reduced to harmonic oscillator one.
In the latter, making use of Eliasson theorem, we seek mathematical conditions that have to be fulfilled by Hamiltonian vector field such that the system is integrable in HO sense. In the end of the latter subsection we give a result that states about the conditions to be fulfilled by the physical system to be integrable with the momentum map isomorphic with the HO momentum map, from both perspective.
The next section is dedicated to geometric quantization of HO in holomorphic representation using a particular K\"ahler polarization. We emphasize the importance of choosing the correct Hamiltonian acting on holomorphic sections as a correspondent of energy function of the system. Here one should consider the commutation relation of complex coordinates, as it comes from quantum mechanic in the energy or H-representation.
In the last section we introduce two physical systems that illustrate the reasoning exposed throughout the paper. We show that the potentials belong to the introduced class of functions and therefore GQ prescription exposed in the previous section can be applied.
Finally, we want to emphasize the importance of the conditions, developed here, by presenting an example of a physical system that has the potential local isomorphic with HO, but does not admit GQ having no isomorphism between momentum maps.
The paper ends with conclusions.
\section{About geometric quantization.}
An achievement in quantum mechanics is to build the spectrum and the corresponding eigenfunctions for energy operator, using Schr\" odinger's equation with some given boundary conditions. Once this is done, we can say that we have an integrable physical system, further on this subject is exposed in \cite{Fordy}.
In geometric quantization procedure, for a given physical system, the quantization rules are used to obtain a Hilbert space of states and a set of operators acting on Hilbert space, representing quantum observables.
Formally, we can say that the GQ procedure involves three major steps: prequantization, polarization, and metaplectic correction.
Prequantization gives a preliminary Hilbert space and a complete, but reducible, representation of the classical observables.
Prequantization is not enough to get the correct energy spectrum of physical systems, thus some additional structures where introduced to obtain a quantization of a symplectic manifold. One of these is a polarization, and this generally leads to a rather severe technical complications.
The polarization is needed to reduce the pre-Hilbert space, this is made by preserving the complete commuting set of observables.
The complication occurs due to lack of natural measure on the space of quantum states in most cases, further more when such measure exists sometimes GQ is still not completely correct regarding the spectrum of energies.
In this case it is needed to modify the quantization scheme to what is known as half-form or metaplectic quantization. At this stage GQ becomes successful, but the reverse of the medal is it becomes a little bit complicated and unwieldy.
The introduction of a metaplectic structure provides the measure in terms of which the quantum Hilbert space inner product is defined.
Physical systems admit the Schr\" odinger representation, which comes naturally from phase space, by introducing the cotangent bundle $M =T^*Q$, with canonical basis $\{q_i,p_j\}$ and standard symplectic form $\Omega = dq_i \wedge dp_i$.
The polarization $P$ is taken to be the span of the vertical vector fields $\left\{ \frac{\partial}{\partial p_i}\right\}_{i=1\ldots n}$.
The polarized sections $\psi$ are the sections for which $\frac{\partial\psi}{\partial p_i} =0$, i.e. those which are constant along the fibers. In this way the dimension of pre-Hilbert space is reduced, the sections depend on half of variables i.e. $\psi=\psi(q^i)$ and form the Hilbert space.
The corresponding quantum operators for the momenta respectively position are written as:
\beq \label{oppq}
\Op_{p_j} = -i\hbar \derpar{}{q^j} \ ; \
\Op_{q^j} = q^j.
\eeq
Physics sets Hamilton's function to be the most significant observable, having a kinetic and potential part, $H=K+V$. It defines the energy and describes the motion of system.
The dynamics is generated by a smooth potential $V(q)$ acting on Hilbert space, described with the help of equation of motion:
\begin{prop} \label{ham}
The Hamiltonian function $H = p^2+V(q)$, up to some factors and constants, produces a vector field, $X_{H}$, such that the dynamics is given by Hamilton's equations $i_{X_{H}}\Omega=-dH$.
\end{prop}
\noindent The classical observable $H$ gives rise to the Hamiltonian vector field $X_{H}$, the flow of $X_{H}$ is called the Hamilton flow.
\section{Critical points.}
\subsection{Critical points of Hamiltonian function.}\label{cph}
It is important for the dynamics of the system to know how critical points of the Hamiltonian function are managed.
We are focusing around the critical points of the potential, which tell us about the change in shape of the potential. This information can be used to make a qualitative prediction about the spectra of Hamiltonian operator.
Our reasoning is to show, by Morse lemma, the existence of a diffeomorphism in a neighborhood of a nondegenerate singular point of a smooth function that takes the given function to its quadratic part. In this way, we can reduce our system to one having quadratic potential.
With other words, if the potential admits critical points, by Morse lemma, the topological changes of manifold are put into one-to-one correspondence with these points.
Let us see how this things work, we begin with some needful definitions, briefly presented, but an exhaustive presentation can be found in \cite{Arnold-cp}.
\begin{definition} \hfill
\begin{itemize}
\item The critical points of a smooth function are the points where the differential vanishes.
\item A critical point is nondegenerate if the second differential (Hessian) is a nondegenerate quadratic form.
\item The index $\lambda$
of nondegenerate critical point is the dimension of maximal subspace on which the Hessian is negative definite.
\item A smooth function is called Morse function if all critical points are nondegenerate.
\end{itemize}
\end{definition}
\begin{lemma}[Morse]
Let $f$ be a Morse function and $p\in M$ a critical point of manifold $M$. There are local coordinates $\{x_1,\ldots, x_n\}$, with $p=(0,\ldots,0)$ such that $f(p\prime)= f(p) -x^2_1,\ldots, -x^2_\lambda + x^2_{\lambda+1},\ldots,+x^2_n$ for every point $ p\prime \in U$ a small neighborhood, and $\lambda$ is the index of $f$ at $p$.
\end{lemma}
What we can say, as an immediate result of Morse lemma, for our considered physical system is that:
\begin{corol} \label{morsepot}
Let potential $V(q)$ be a Morse function, then in some neighborhood of a critical point the potential can be represented in the Morse formulation:
\beq
V(q)= \pm q^2_1 \pm \ldots \pm q^2_n
\eeq
this is made using canonical coordinates.
\end{corol}
It is easy to see that for a Morse function the critical points are isolated. Hence, the sets of critical points is a 0 - dimensional manifold.
In physics the Morse function critical points are called stationary points, which can be stable or unstable. Of course, for physical reasons we are interested in the stable ones.
\begin{definition}
A stable stationary point is a critical point of a Morse function having the Hessian positive definite.
\end{definition}
With this we can state that the dynamics of physical system is locally generated by potential having the following form:
\begin{prop} \label{potHO}
Let potential $V(q)$ be a Morse function, then in some neighborhood of a stable stationary point the potential can be represented in the Morse formulation:
\beq
V(q)= q^2_1 + \ldots + q^2_n .
\eeq
\end{prop}
We have seen that by studying the potential, we determine the conditions needed to write it locally as its quadratic part, the global dynamics have to be determined for each particular physical system.
\subsection{Critical points of Hamiltonian flow}
The study of the singularities of Hamiltonian systems can be made by using two different approaches: one can study the Hamiltonian functions themselves, this is the foliation perspective (as we presented in subsection \ref{cph}), or one can analyze the flow of the vector fields, this is the dynamical viewpoint and follows in this section.
\begin{definition}
\begin{itemize}
\item We say $X_{H}$ is complete if it generates a global flow on manifold.
\item A function is called proper function if it has the property $f^{-1}(compact)=(compact)$.
\end{itemize}
\end{definition}
\begin{lemma}[Gordon, \cite{Gordon}]
Let $X$ be a ${\cal C}^1$ vector field on a manifold $M$ of class ${\cal C}^1$. Then $X$ is complete if there exist a ${\cal C}^1$-function $E$, a proper ${\cal C}^0$-function $f$, and constants $\alpha,\, \beta$, such that for all points in manifold $x\in M$
\begin{enumerate}
\item $| X(E(x)) | \leq\alpha |E(x)| $
\item $| f(x) | \leq\beta |E(x)| $
\end{enumerate}
\end{lemma}
The above result, in the case of symplectic manifold, can be rewritten as follows:
\begin{prop}[\cite{puta-pp} ]
Let $(M, \Omega)$ be a smooth $2n$-dimensional manifold and $H\in \cal C^\infty$ the Hamilton's function on manifold. If $H$ is proper and bounded below then $X_{H}$ is complete.
\end{prop}
\begin{definition}
The flow of a Hamiltonian $H$ on a $2n$-dimensional symplectic manifold $(M, \Omega)$ is said to be integrable, or completely integrable if there exist $n$ everywhere independent integrals $\{f_1=H, f_2, \ldots , f_n \}$ of the flow which are in involution.
\end{definition}
\begin{obs} \label{pb}
The algebra of classical observables, including Hamiltonians, comes naturally endowed with the Poisson bracket: $\{\cdot, H\}=\Omega(\cdot , X_{H})$.
\end{obs}
The evolution of a function $f$ under the flow of $X_{H}$ is given by the equation $\dot f=\{H,f\}$.
An integral of the Hamiltonian $H$ is a function which is invariant under the flow of $X_{H}$, i.e. a function $f$ such that $\{H, f\}= 0$.
The functions $\{f_1,\ldots f_n\}$ are denoted by $F= (f_1,\ldots f_n)$, which is usually called the moment map.
\begin{rem}
On a $2n$-dimensional symplectic manifold, a completely integrable system is a moment map function.
\end{rem}
The moment map describe and takes all characteristics of the system. The singularity becomes a feature of the moment map. Accordingly, we seek systems with moment map isomorphic with that of HO by studying the singularities.
Singularities corresponding to fixed critical points of relative equilibrium of the system are one of the main characteristics of the dynamics.
The Hessian in these points is a non-degenerate quadratic form.
In accordance with the linear classification of Cartan subalgebras of $sp(2n,\mathbb{R})$, any such Cartan subalgebra has a basis build with three type of blocks: two uni-dimensional ones (the elliptic and the real hyperbolic) and a two-dimensional one called focus-focus \cite{Will}.
\begin{rem}\label{ff}
There are only two types of non-degenerate singularities for integrable systems in dimension 2: hyperbolic (when the Hessian is indefinite) or elliptic (when the Hessian is positive or negative definite)
\end{rem}
Here is a major result, given by Eliasson \cite{Eliasson}, that gives a connection between integrable system and critical points:
\begin{teor}[Eliasson]
The non-degenerate critical points of a completely integrable system are linearizable.
\end{teor}
This problem of symplectic linearization, closely related to the spirit of Morse lemma, was solved successfully by Vey and Colin de Verdiere \cite{Vey}.
\begin{teor}[Vey]
Let $F : (M^2,\Omega) \to \mathbb{R}$ be a function and let p be a non-degenerate singular point of $F$. Let $Q$ be the quadratic form corresponding to the Hessian of $F$ at $p$.
Then there exists a local diffeomorphism from a neighborhood $Z$ of $p$ to a neighborhood of $0 \in \mathbb{R}^2$ taking $F$ to a function $\phi(Q)$. If the hessian $Q$ is positive definite the germ of the function $\phi$ characterizes the pair $(F, \,\Omega)$. If $Q$ is not definite then the jet at the point $p$ of the function $\phi$ characterizes the pair $(F, \,\Omega)$.
\end{teor}
\begin{rem} \label{vecHO}
As a consequence of this theorem, after putting $Q$ in coordinates form ($\Omega = \d x \wedge d y$) we can assume from now on that the foliation in a neighborhood of a singular point $p$ corresponding to $0$ is given by the vector field:
\begin{itemize}
\item $Y = - y\frac{\partial}{\partial x} + x \frac{\partial}{\partial y}$ when $ Q = x^2+y^2$ p is elliptic
\item $Y = x\frac{\partial}{\partial x} - y \frac{\partial}{\partial y}$ when $ Q = xy$ p is hyperbolic
\end{itemize}
\end{rem}
\noindent It is usual to call these ($x,y$) coordinates - Eliasson coordinates.
The following result gives the conditions that have to be performed by a physical system such that it can be reduced to HO. This means an isomorphism of momentum maps, it relates the potential to Hamiltonian flow and its critical points of of each of two systems. If the isomorphism is established the GQ can be applied to it using set of rules to be followed in GQ of HO.
We choose to give this result in two dimensions to avoid the signature discussions of Hessian. This means that, in accordance with the remark \ref{ff}, we do not consider the focus-focus case.
Two dimensions in phase space $(q, p)$ represent a system with one degree of freedom in physical space, but the reasoning, of the following theorem, can be extended to the n-dimensional case $(q^n, p^n)$, leaving out the focus-focus case that is not relevant in applications in physics.
\begin{teor} \label{HOcond}
Let be a physical system acting on a 2-dimensional manifold, observed in Eliasson coordinates and described by its moment map. The system is integrable having the moment map isomorphic with HO moment map if at least one of the following conditions are fulfilled:
\renewcommand{\theenumi}{\roman{enumi}
\begin{enumerate}
\item The Hamiltonian is proper and bounded below.
\item The Hamiltonian vector field is complete and has elliptic critical point.
\end{enumerate}
\end{teor}
\begin{dem}
\renewcommand{\theenumi}{\roman{enumi}
In Eliasson coordinates, $(q, p)$ we have:
\begin{enumerate}
\item The Hamiltonian is proper and bounded below implies there exists a stationary point such that the Hessian is positive definite, i.e. we have a stable stationary point. This Hamiltonian is a Morse function so according to prop. (\ref{potHO}), it takes the form $H= \Op_{p_j}^2 + \Op_{q^j}^2 $, which is the HO Hamiltonian, so the system is HO integrable.
\item The Hamiltonian vector field is complete and has one or more elliptic critical points, according to rem. (\ref{vecHO}) it takes the form $H= \Op_{p_j}^2 + \Op_{q^j}^2 $, so the system is HO integrable.
\end{enumerate}
\end{dem}
\begin{obs}
Theorem \ref{HOcond} gathers together the foliation and flow of the vector fields perspective.
\end{obs}
\begin{obs}
In quantum mechanics, one typically describes a physical system with a C*-algebra of physical observables. The isomorphism in theor. \ref{HOcond} is meant as a $C^*$-algebra isomorphism, the unbounded operators are considered in the Weyl form \cite{Strocc}, in order to satisfy the canonical commutation relations.
\end{obs}
We are going to extend the above result through the following remark:
\begin{rem}\label{remmea}
In the hypothesis of theorem \ref{HOcond}, the equivalence of following conditions holds:
\begin{enumerate}
\renewcommand{\theenumi}{{\em\roman{enumi}}
\item The system is integrable having the moment map isomorphic with HO moment map.
\item The Hamiltonian is proper and bounded below.
\item The Hamiltonian vector field is complete and has elliptic critical points.
\end{enumerate}
\end{rem}
\section{Harmonic oscillator quantization in Bargmann-Fock representation.}\label{BFR}
The representation is determined by the choice of polarization. The Schr\" odinger representation is given by a polarization spanned by the vector field determined by position coordinate. In the momentum representation the polarization is spanned by the vector field of momentum coordinate. The kernel between these representations is given the by Fourier transform.
The holomorphic or Bargmann-Fock representation is obtained when we use a particular K\"ahler polarization, in which we have to make a coordinate change, from the standard phase space coordinates $(p,q)$, by introducing the complex ones $\{ z_j, \bar z_j \}$ where $z_j := p_j+iq^j$. Quantum mechanic names this as energy or H-representation.
CQ method applies in both Schr\" odinger and Bargmann-Fock representation, while GQ method applies just in the last and its main features are presented below.
In this representation the symplectic form becomes:
$$
\Omega=\frac{i}{2} \d \bar z_j \wedge \d z_j,
$$
the structure $(\Tan^*Q,\Omega ,{\cal J})$ is a K\"ahler manifold,
where the complex structure is given by:
$$
{\cal J}\left(\derpar{}{p_j} \right) = \left(\derpar{}{q^j}\right)
\quad , \quad
{\cal J}\left(\derpar{}{q^j}\right) = -\left(\derpar{}{p_j}\right).
$$
Now, we can consider the polarization $\P$, spanned by $\left\{ \derpar{}{\bar z_j}\right\}$
and as a symplectic potentials we can chose
\beq \label{spot}
\Theta = \frac{i}{4}(\bar z_j \d z_j - z_j \bar z_j)
\eeq
or the adapted one
$$
\theta = \frac{i}{2}\bar z_j\d z_j .
$$
The polarized sections are the holomorphic sections of the complex line bundle
$\Tan^*Q \times \Complex$
and using the symplectic potential (\ref{spot}), the quantum operators acting on these sections are:
\beq
\Op_{z_j}=-2\hbar\derpar{}{\bar z_j}+\frac{z_j}{2}
\quad ; \quad
\Op_{\bar z_j} \equiv \Op_{z_j}^+=2\hbar\derpar{}{z_j}+\frac{\bar z_j}{2}.
\label{opers}
\eeq
In physics this approach is considered more elegant reinforcing Dirac
notation, which depends upon the arguments of linear algebra. The raising and lowering operators (\ref{opers}),
or ladder operators, are the predecessors of the creation and annihilation operators used in the quantum mechanical description of interacting photons. The ladder operators obey commutation rules $\left [ \Op_{z_j},\, \Op_{\bar z_j}\right]=1$.
In this representation, a relevant role is played by the so-called number operator:
$$
\Op_{\bar z_j z_j} = 2\hbar \left( z_j\derpar{}{z_j}-\bar z_j \derpar{}{\bar z_j} \right) .
$$
The polarized sections are:
$$
\psi (z_j,\bar z_j) = F(z) e^{-\frac{z_j \bar z_j}{4\hbar}},
$$
which are holomorphic sections on $\Complex$, and the inner product is given by
$$
\langle \psi_1 \mid \psi_2\rangle =\left( \frac{1}{2\pi \hbar}\right)^n
\int_M F_1(z)\bar F_2(z) e^{-\frac{z_j\bar z_j}{2\hbar}}\Lambda_\Omega
$$
Let us see the action of operators on polarized sections, which are the eigenstates of the system:
\begin{eqnarray}
\Op_{z_j} |\psi \rangle &=& z_j \psi = z_j F(z)e^{-\frac{z_j\bar z_j}{4\hbar}}
\\
\Op_{z_j}^+|\psi \rangle &=& 2\hbar\derpar{\psi}{z_j}+\frac{\bar z_j}{2}\psi
= 2\hbar\derpar{F}{z_j}e^{-\frac{z_j\bar z_j}{4\hbar}}
\\
\Op_{\bar z_j z_j}|\psi \rangle &=& 2 \hbar\left( z_j\derpar{\psi}{z_j} -\bar
z_j\derpar{\psi}{\bar z_j}\right) = \hbar z_j \derpar{F}{z_j}
\end{eqnarray}
The eigenfunctions $F(z)$ are homogeneous polynomials of degree $n$, the action of number operator will reveals the twice of state number:
$$
Spec \left( \Op_{\bar z_j z_j}\right) = 2n.
$$
To obtain the correct spectrum, one should consider the proper Hamiltonian. There is a confusion to take the number operator instead of Hamilton operator and after introducing the complex coordinates, to write down the Hamiltonian as $\Op_H = \frac{1}{2}\Op_{\bar z_j z_j} $, which leads us to the incorrect physical spectrum. In these conditions a correct choice should be:
$$
\Op_H = \frac{1}{2}\left( \Op_{\bar z_j z_j} + \Op_{z_j \bar z_j} \right)
$$
and using the commutation relations $\left[ \Op_{ z_j} ,\, \Op_{\bar z_j} \right]=1$ we get:
\begin{equation}\label{opzHO}
\Op_H = \frac{1}{2} \left( \Op_{\bar z_j z_j} + 1 \right).
\end{equation}
one should get the correct spectrum:
\begin{equation}\label{sHO}
Spec \left( \Op_{\bar z_j z_j}\right) = n+\frac{1}{2}.
\end{equation}
In the following we use this representation using the standard CQ notations:
\begin{itemize}
\item ladder up/down operators will be denoted $a:=\Op_{z_j}$ and $a^+:= \Op_{\bar z_j}$.
\item the number operator $\rm N:= \Op_{\bar z_j z_j} $.
\item the Hamiltonian $ H:=\Op_H$.
\end{itemize}
\section{Examples and conclusions}
\subsection{Illustrative Examples }
Let us consider the Schr\"odinger equation:
\beq
\left(- \partial_q^2+V(q)-E\right)\phi(q)=0,\label{schro3}
\eeq
the
parameter $E$ is the energy.
This equation is the eigenvalue problem for the Hamilton operator.
In accordance with proposition \ref{ham}, to describe the HO states the Hamiltonian vector field have to be generated by:
\beq H:=- \partial_q^2+\omega^2 q^2, \label{schro}
\eeq
where the momentum operator is written in its differential form (\ref{oppq}), we use natural units $\hbar=c=1$ and the parameter $\omega$ is proportional with the angular velocity.
Affine transformations, as translations $q\mapsto aq +b$ with
$a\neq0,$ $b$ constants, preserve the class of potentials.
In this manner all equations
\beq\label{THO}
\left(-\partial_q^2+\omega^2 q^2+\rho
q+\lambda\right)\phi(q)=0,
\eeq
given by parameters $\rho,\,\lambda$, named translated harmonic oscillator (THO) equations are in fact HO equations.
All elementary transformations as multiplication by a function or/and a change of variables, not depending on
the equation parameters, will reduce the THO equation (\ref{THO}) to the HO equation (\ref{schro}).
An exception is made when $\omega=0$ case in which equation (\ref{THO}) becomes the Airy equation, which can be reduced to a special hypergeometric equation by a transformation.
\begin{example}\label{ex1}
Consider the following class of potentials:
\begin{eqnarray}
\label{pot1}
V_{\omega,\lambda}(q) =
\omega^2\left(\frac{3q}{2}\right)^{\frac23}+
\lambda\left(\frac{2}{3q}\right)^{\frac23}-\frac{5}{36}\frac{1}{q^2}.
\end{eqnarray}
Let us show that Schr\"odinger equations (\ref{schro3}), having this potential, with the following coordinate transformation
$z=\left(\frac{3q}{2}\right)^{\frac23}$
will lead us to the Schr\"odinger equations (\ref{THO}).
It is easy to verify that with the transformation operator: $U = z^\frac{1}{4}$ we obtain successively:
\begin{eqnarray*}&&
U^{-1}\Op_H U\\
&=&z^{\frac14-1}\left(-\partial_z^2+\omega^2 z^2+\rho
z+\lambda\right)z^{-\frac14}\\
&=&-\frac1z\partial_z^2+\frac{1}{2z^2}\partial_z+\omega^2
z+\rho+\frac{\lambda}{z}
-
\left(\frac{1}{4}+\frac{1}{4^2}\right)\frac{1}{z^3}\\
&=&-\partial_q^2+\omega^2\left(\frac{3q}{2}\right)^{\frac23}+\rho+
\lambda\left(\frac{2}{3q}\right)^{\frac23}-\left(\frac{1}{4}+\frac{1}{4^2}\right)
\left(\frac{2}{3q}\right)^2. \end{eqnarray*}
We notice this is the starting potential rescaled with the energy $-\rho$.
The physical system described by eq. (\ref{schro3}) with the potential (\ref{pot1}) admits GQ steps used for HO and
according with the standard procedure, presented in section \ref{BFR}, the spectra of energy operator is given by (\ref{sHO}), but it is shifted with $-\rho$.
$$
\Op_H = \frac{1}{2} \left( \Op_{\bar z_j z_j} + 1 \right) + \rho.
$$
One can verify that the both conditions of prop. \ref{HOcond} are fulfilled, this is Hamiltonian is proper and bounded below and Hamiltonian vector field has elliptic critical points.
A closer look to the potential reveals that its first two terms are a Lennard-Jones' type potential \cite{LJ} plus the repulsive inverse square potential.
\end{example}
\begin{example}\label{ex2}
Consider the following class of potentials:
\begin{eqnarray*}
V_{\rho,\lambda}(q) =
\frac{\rho}{(2q)^{\frac12}}+
\frac{\lambda}{2q}-\frac{3}{16}\frac{1}{q^2}
.\end{eqnarray*}
In the same manner, with suitable coordinate transformation:
$z=(2r)^{\frac12},$ we show th equivalence of the Hamilton operators of equations (\ref{schro3}) (\ref{THO}), with the transformation operator: $U= z^{-\frac{1}{2}}$ successively we get:
\begin{eqnarray*}&&
U^{-1}\Op_H U\\
&=&z^{\frac12-2}\left(-\partial_z^2+\omega^2 z^2+\rho
z+\lambda\right)z^{-\frac12}\\
&=&-\frac{1}{z^2}\partial_z^2+\frac{1}{z^3}\partial_z+\omega^2
+\frac{\rho}{z}+\frac{\lambda}{z^2}
-\left(\frac{1}{2}+\frac{1}{2^2}\right)\frac{1}{z^4}\\
&=&-\partial_q^2+\omega^2+\frac{\rho}{(2q)^{\frac12}}+\frac{\lambda}{2q}-
\left(\frac{1}{2}+\frac{1}{2^2}\right)\frac{1}{2^2q^2}
. \end{eqnarray*}
What we observe now is that the rescaling factor of the energy $-\omega^2$. The GQ will follow according with the standard procedure with the spectra of energy operator $\Op_H $, given by (\ref{sHO}), shifted with $-\omega^2$.
\end{example}
\begin{example}\label{ex3}
Consider the following class of potentials:
\begin{equation}\label{pot}
V_{\lambda,\mu}^s(q)= (\lambda+\sum_{k=0}^s C_k q^{2k})\,\exp(-\mu q^2),
\end{equation}
where the coefficients have the following expression:
$
C_k=\frac{(\lambda+k)\mu^{k}}{k!}\,.
$
with $\lambda$, $\mu$ real parameters and $s$ a positive integer.
For each fixed parameters, it can be observed that in a vicinity of origin the potential behaves like HO potential, with a good approximation we can write:
$$
V_{\lambda,\mu}^s(r)=\lambda+\mu^2 q^2 - {\cal O} (q^{2s}).
$$
and discard the last term, which is a progressive summation of powers o $q^{2s}$ with grater than $s>1$. However, the potential do not accomplish the property of being proper neither the Hamiltonian vector field is complete albeit having elliptic critical points. In quantum mechanics it is not possible to approximate a potential in any condition, it will not preserve the particle probability of localization. One can make approximations upon potential, just if the system is in interaction and therefore apply perturbation theory. In a previous work \cite{pgo} we have proven that the system is integrable, but not in the HO sense. We found that the energy spectra of Hamiltonian is is different from that of HO.
\end{example}
\subsection{Conclusions}
In this paper we investigate the mathematical properties of the potential as part of the Hamiltonian, which describe the physical systems, things looked less in the approaches of physics. We aimed to find criteria for potential function in order to get an integrable system according to both CQ and GQ as well.
The dynamics of the system are given by the stationary points of Hamiltonian, in this view we considered stationary points of great interest in our investigation.
Thus, it have been shown that in a neighborhood of stable stationary point the potentials, via Morse lemma, can be always locally represented as its quadratic part. This condition is not a sufficient one, as example \ref{ex3} shows.
Further, describing the physical system by the associate momentum map we defined its reduction to HO by the existence of an isomorphism between their momentum maps.
With the help of Eliasson's theorem, we give a result, by theorem \ref{HOcond}, which states the conditions for momentum map to be isomorphic with momentum map of HO. These conditions are found to refer to the property of Hamiltonian and Hamiltonian vector field, accordingly with their stationary points.
In accordance with physical laws of steady states just elliptic stationary points have to be considered. The global conditions apply to each potential individually, each integrable potential function describing a real physical system seems to accomplish the conditions of giving a global action. Exception seems to be made for periodically-repeating potential. In this case the potential domain can be, mathematically, restricted to one stationary point. This is ensured by Bloch\textsc{\char13}s theorem, which states that: {\em the system is solvable on each repeating domain}. The energy eigenfunctions have a basis consisting entirely of Bloch wave energy eigenstates. This decomposition of potential on smaller domains are gained by Eliasson's theorem as well, which states that stationary points are linearizable.
Based on this reasoning, two physical systems were presented, namely example \ref{ex1} and example \ref{ex2}. An isomorphism with HO momentum map were found and the reduced systems were quantified using GQ as was introduced in section \ref{BFR}.
In example \ref{ex3} a system being very similar with HO if the behavior of Hamiltonian is considered in a vicinity of the origin. This example emphasize the importance of the condition that the Hamiltonian to be proper. Without this condition even if the system is integrable, see \cite{pgo}, it is no momentum map isomorphism to the HO. To apply GQ there is need another approach than that from section \ref{BFR}. We conclude that no approximations can be made upon a potential.
There are some some misinterpretations about the possibility to calculate with less errors at least first energy level in the case of Taylor approximation to HO of Hamiltonian in the neighborhood of a stationary point.
|
\section{Supplemental Material}
\subsection{Experimental timing}
Details of a typical experimental sequence are given in
Fig.~\ref{Reichsoellner_supmat_Fig1}. The sequence starts with the loading of
the two BECs into the optical lattice and creating a Cs MI, followed by the
transport of the Rb BEC, the Rb-Cs pair formation, the production and the
subsequent detection of molecules. We denote the intensity in the dipole
trapping beams by $I_\text{y,Rb}$, $I_\text{y,Cs}$, and $I_\text{x}$ for the
beams along the $y$~direction for the inital Rb and Cs dipole traps and for the
beam joining the two samples along the transport $x$~direction, respectively
(see Fig.~1(a) of the main article). In the course of the sequence, the lattice
depth $V^\text{Cs} = 2.6 V^\text{Rb}$, the Rb trap position $x_\text{Rb}$ along
the transport direction, the intensities $I_\text{y,Rb}$, $I_\text{y,Cs}$, and
$I_\text{x}$, the magnetic offset field $B$, and the magnetic field gradient
$\left| \nabla B \right|$ are ramped. Note that the Cs atoms see a $1.08$ times
larger trap frequency than the Rb atoms for a given dipole trap laser power.
Initially, the Rb BEC and the Cs BEC are spatially separated along the
$x$ direction by $\approx 100$~$\mu$m with $\nu_{\text{x,Rb}}=38$~Hz,
$\nu_{\text{x,Cs}} = 10$~Hz, $\nu_{\text{y,Rb}}=14$~Hz,
$\nu_{\text{y,Cs}}=15$~Hz, $\nu_{\text{z,Rb}} = 39$~Hz, and
$\nu_{\text{z,Cs}} = 18$~Hz. Here, $\nu_{\text{x,Rb}}$ denotes the trap
frequency for the Rb trap in $x$~direction, and analogously for the other trap
frequencies. The initial distance of the two BECs is large enough to avoid
spilling from one sample into the other. In the course of the transport, while
the confinement of the two samples is controlled by the intensities
$I_\text{y,Rb}$, $I_\text{y,Cs}$, and $I_\text{x}$, it is strongly modified
along the $x$~direction when the underlying dipole traps start to overlap. Beam
steering to move the Rb trap along the $x$ and the $z$~directions is achieved
by a two-axes translational piezo flexure stage onto which we have attached the
fiber tip of the fiber that delivers the light for the Rb trap beam propagating
in the $y$~direction. Beam steering is done with $\mu$m precision as verified
by in-situ absorption images. The initial gradient
$\left| \nabla B \right| = 31.1$~$\text{G}/\text{cm}$ levitates the Cs sample
against gravity, but slightly overlevitates the Rb sample.
Upon loading both ensembles into the 3D optical lattice with a lattice spacing
of $\lambda/2=532.25$~nm we drive the SF-to-MI phase transition for the Cs
sample while leaving the Rb sample superfluid. For this, we exponentially
increase the lattice depth to
$V^\text{Cs} = V_\text{mix}^\text{Cs} = 20$~$E_\text{rec}^\text{Cs}$
($V_\text{mix}^\text{Rb} = 7.7$~$E_\text{rec}^\text{Rb}$). At the same the time
underlying Cs dipole trap is stiffened to assure that we create a clean
one-atom MI shell for Cs.
Subsequently, the Rb sample is transported through the lattice towards and onto
the Cs sample within typically $1500$~ms by moving the underlying Rb dipole
trap along the $x$~direction. The gradient $\left| \nabla B \right|$ is ramped
to $30.1$~$\text{G}/\text{cm}$ in the course of the first $500$~ms of the
transport process for optimum levitation of Rb. We slightly adjust the vertical
position of the Rb trap in the course of the transport for optimal overlap with
the Cs sample.
After $998$~ms of the transport, i.e. $502$~ms before the Rb sample has reached
its final position, the offset field $B$ is ramped within $2$~ms to the zero
crossing for $a_\text{RbCs}$ at $354.95$~G. Simultaneously
$\left| \nabla B \right|$ is adjusted to $25.9$~$\text{G}/\text{cm}$ to
compensate for the change in the magnetic moment of Rb as $B$ is changed.
Within the last $500$~ms of the transport the underlying Cs dipole trap beam
along the $y$~direction is adiabatically turned off to avoid its influence on
the final phase of the Rb transport.
\begin{figure}[tbp]
\includegraphics[width=\columnwidth]{Reichsoellner_supmat_Fig1}
\caption{Typical timing sequence for the experimental cycle
with typical values for the Rb trap position $x_\text{Rb}$ along the transport
direction, the beam intensities $I_\text{y,Rb}$, $I_\text{y,Cs}$, and
$I_\text{x}$, the lattice depth $V^\text{Cs} = 2.6 V^\text{Rb}$, the magnetic
offset field $B$, and the magnetic field gradient $\left| \nabla B \right|$ as
a function of time. The various experimental stages (A) lattice loading, (B) Rb
transport, (C) Rb localization, (D) molecule formation, (E) Stern-Gerlach
separation, and (F) dissociation and detection are indicated in the lower part
of the diagram. The horizontal axis is not to scale.}
\label{Reichsoellner_supmat_Fig1}
\end{figure}
As soon as the Rb transport is finished we increase the lattice depth
$V^\text{Cs}$ to $V_\text{fin}^\text{Cs} = 36$~$E_\text{rec}^\text{Cs}$
($V_\text{fin}^\text{Rb}=13.8$~$E_\text{rec}^\text{Rb}$) to drive the Rb sample
into the MI regime and to form Rb-Cs atom pairs in the lattice. By
adiabatically sweeping $B$ within $2$~ms over the pole of the Feshbach
resonance at $355.74$~G to $348.4$~G we associate the paired atoms to weakly
bound molecules (see the Zeeman diagram in
Fig.~\ref{Reichsoellner_supmat_Fig2}). Subsequently, within $1$~ms, we jump $B$
to $316.0$~G, where an avoided crossing is located that allows us to transfer
the molecules into the second state by ramping $B$ in $2$~ms to $314.0$~G (see
Fig.~\ref{Reichsoellner_supmat_Fig2} and the discussion in the next section).
This final molecular state possesses a magnetic moment that significantly
differs from the one of the unbound atoms, allowing Stern-Gerlach separation of
molecules and atoms. We switch off the dipole trap and lattice potentials
abruptly and spatially separate the molecules from nonassociated atoms within
$3$~ms time of flight. For detection we reverse the Feshbach association ramp
to dissociate the molecules back to atoms. These atoms and also the atoms that
were not subject to molecule formation are subsequently detected by standard
absorption imaging.
\begin{figure}[tbp]
\includegraphics[width=\columnwidth]{Reichsoellner_supmat_Fig2}
\caption{Zeeman diagram for the molecular states with
$M_\text{tot}=4$ just below the ground-state two-atom
$(f_\text{Rb},f_\text{Cs})=(1,3)$ threshold. Energies as a function of $B$ are
given relative to the field-dependent dissociation threshold (horizonal dashed
line). The red arrow marks the position of the Feshbach-resonance zero crossing
at which Rb-Cs atom pairs are produced. The magneto-association path is
indicated by the green arrows and the thick green line, first following state
$|a\rangle$ and then ending in state $|b\rangle$ as $B$ is lowered. The inset
zooms into the resonance region, showing the zero-crossing position at
$354.95$~G (red arrow), an overlapping molecular state $|e\rangle$ hitting
threshold at $353.57$~G (black, nearly vertical line) and the pole of the
Feshbach resonance at $352.74$~G where state $|a\rangle$ hits threshold.}
\label{Reichsoellner_supmat_Fig2}
\end{figure}
\subsection{Feshbach states}
In the following we adopt the labelling
$|n(f_\text{Rb},f_\text{Cs})L(m_{f_\text{Rb}},m_{f_\text{Cs}})M_\text{tot}\rangle$
of the Feshbach states used in Ref.~\cite{Takekoshi2012gad2}, where a broader
overview on the Rb-Cs Feshbach states and resonances is given. The number $n$
indicates the vibrational level as counted from threshold (i.e.\ $n=-1$ refers
to the least bound vibrational level, $n=-2$ is the second least bound
vibrational level, etc.)\ and the quantum numbers $f$ and $m_f$ label the total
atomic angular momentum and its projection onto the $B$-field axis,
respectively. The quantum number $L$ (with $s$ for $L=0$, $p$ for $L=1$, $d$
for $L=2$ etc.)\ denotes the molecular rotational angular momentum and
$M_\text{tot}=m_{f_\text{Rb}}+m_{f_\text{Cs}}+M_L$ is the sum of all projected
angular momenta. Since here $M_\text{tot}=4$ for all relevant states we omit
$M_\text{tot}$.
For the mixing process and molecule creation we make use of a $2.21$-G-wide
Rb-Cs Feshbach resonance, whose pole is located at $352.74$~G. The background
interspecies scattering length in the vicinity of this resonance is
$a_\text{RbCs}=654$~$a_0$ \cite{Ruth2}. The Rb and Cs samples are initially
overlapped at the zero crossing for $a_\text{RbCs}$ at $354.95$~G, indicated by
the red arrow in Fig.~\ref{Reichsoellner_supmat_Fig2}. From there $B$ is ramped
to the pole of the resonance, jumping nonadiabatically over a very narrow Feshbach
resonance (with a width of $1.2$~mG) located at $353.57$~G caused by the
molecular state $|e\rangle=|-6(2,4)d(0,2)\rangle$ \cite{Ruth2} (see inset in
Fig.~\ref{Reichsoellner_supmat_Fig2}). The atoms then enter a Feshbach state
that initially is a combination of state $|a\rangle=|-1(2,4)s(1,3)\rangle$ and
$|d\rangle=|-6(2,4)s(0,4)\rangle$ (green curve in the inset to
Fig.~\ref{Reichsoellner_supmat_Fig2}). Upon further lowering $B$ the state
becomes $|a\rangle$. In order to allow the separation of molecules from unbound
atoms via the Stern-Gerlach technique, an anti-crossing of molecular states at
$315$~G is used to adiabatically transfer the molecules from state $|a\rangle$
to $|b\rangle$. The latter one is together with state $|c\rangle$ a
superposition of the molecular states $|-6(2,4)d(0,3)\rangle$ and
$|-6(2,4)s(1,3)\rangle$. The exact composition of the molecular states is of
importance for future work when the molecules are to be transferred by STIRAP
from their weakly-bound state to their rovibronic ground state.
\subsection{Convolution modeling}
We characterize the overlap of the two atomic samples, as presented in Fig.~2
of the main article, by using a simple model that assumes two spheres of
homogeneous density with radii $r_1$ and $r_2$. The convolution $C(d)$ of two
spheres as a function of the distance $d$ between their centers is given by
\begin{equation*}
C(d) = \frac{\pi}{12d}(r_1+r_2-d)^2 \left[ d^2 + 2d(r_1+r_2) -3(r_1-r_2)^2 \right]
\end{equation*}
for partial overlap $\left| r_1 - r_2 \right| < d < (r_1+r_2)$. We allow for an
asymmetric offset $N_\text{RbCs}^\text{offset}$ to account for the trailing
tail that is observed in the experiment when convolving the samples along the
transport $x$~direction. Setting $x_\text{Cs}=0$ for the center of the Cs
sample, the full fit function to the number of RbCs molecules $N_\text{RbCs}$
then reads
\begin{widetext}
\begin{equation*}
N_\text{RbCs}(x_\text{Rb}) = \left\{\begin{array}{lrcccr}
0, &&&x_\text{Rb}&\leq&-(r_1+r_2) \\ N_\text{RbCs}^\text{max}\frac{C(-x_\text{Rb})}{C\left(\left|r_1-r_2\right|\right)} & -(r_1+r_2)&<&x_\text{Rb}&<&-\left|r_1-r_2\right| \\
N_\text{RbCs}^\text{max}, & -\left|r_1-r_2\right|&\leq&x_\text{Rb}&\leq&\left|r_1-r_2\right| \\ (N_\text{RbCs}^\text{max}-N_\text{RbCs}^\text{offset})\frac{C(x_\text{Rb})}{C\left(\left|r_1-r_2\right|\right)} + N_\text{RbCs}^\text{offset}, & \left|r_1-r_2\right|&<&x_\text{Rb}&<&(r_1+r_2) \\
N_\text{RbCs}^\text{offset}, & (r_1+r_2)&\leq&x_\text{Rb}&& \\
\end{array}\right.
\label{equation2015112701}
\end{equation*}
\end{widetext}
For $r_1 \neq r_2$ the function has a characteristic volcanolike profile,
where $N_\text{RbCs}^\text{max}$ defines the height the plateau when one sphere
is fully enclosed by the other. The fit cannot attribute $r_1$ and $r_2$ to the
individual species, however for a sufficiently large difference in the atom
number we presume that the smaller radius belongs to the smaller sample (here
Cs). Even though the atom samples possess some ellipticity due to the shape of
their trapping potential and even though their densities are inhomogeneous, the
fit should yield acceptable estimates for samples' radii along the direction of
the convolution measurement. When compared to a simple calculation that assumes
the formation of defect-free single-shell MIs the sample radii that result from
the convolution-measurement fits are roughly $1$~$\mu$m larger. This deviation
is probably the result of a reduced density in the outer region of the MI as a
result of nonzero temperatures. Note that the convolution measurement shows
that we control the relative positioning of the two atom samples along the two
steering axes to a much higher precision than the resolution of our imaging
setup. It in particular allows us to calibrate a roughly $5$-$\mu$m chromatic
offset between the absorption images of the two species.
\subsection{Filling fraction and entropy}
We normalize the number of RbCs molecules $N_\text{RbCs}$ by the number
$N_\text{Cs}^\text{BEC}$ of atoms in the Cs BEC and thereby obtain a lower
bound for the filling fraction of the RbCs molecules in the optical lattice
when we assume a filling fraction of unity for the Cs MI right after lattice
loading and full coverage of the Cs sample by a homogeneous density of Rb
atoms. The first assumption is well fulfilled for the center region of the Cs
MI (see section Cs MI characterization). The second assumption, however, is not
necessarily fulfilled and can locally lead to an underestimation of the filling
fraction. The Rb sample is elongated along the $y$~axis, therefore a
significantly higher Rb atom number is required to fully cover the nearly
spherical Cs MI. Although the initial Rb BEC is about twice as large as the Cs
BEC, this requirement is not necessarily met, since $20$\%--$50$\% of the Rb
atoms are stuck in the lattice during transport. Furthermore, the Rb density is
not homogeneous, especially since we are not yet able to drive the Rb sample
deep into the MI regime. We note that the elongation of the Rb sample is
actually necessary to provide sufficient overlap since we have no handle on the
relative position of the atom clouds along the $y$~axis and a small offset
exists due to the fact that the focus of the Cs trapping beam is axially shifted
from the intersection with the trapping beam that defines the transport axis.
Several aspects limit the filling fraction of the RbCs molecules: The finite
lifetime of the Cs MI at high magnetic fields (see section Cs MI
characterization) as well as losses during the overlapping process, which we
attribute to interspecies three-body recombination, create vacancies in the Cs
MI. Finite coverage of the Cs MI by the Rb sample and variations in the Rb
density from $n_\text{Rb}=1$ limit the overall pair formation efficiency. The
loss of Cs atoms in absence of the Rb sample is about $15$\%. The number of Rb
atoms that are lost during the mixing process (excluding atoms that are stuck
on the way to the Cs sample during the transport) is about $1.6(1)$ times the
number of additionally lost Cs atoms in presence of Rb. It therefore appears
that this loss is dominated by the Rb-Rb-Cs three-body loss mechanism, further
reducing the number of Cs atoms by $35$\% of the Cs BEC size. We observe an
enhancement of the interspecies loss for slower transport velocities, which can
be compensated by bringing the lattice depth $V_\text{mix}$ closer to the
critical value at which the superflow of Rb breaks down. This is not surprising
considering that both the critical transport velocity and the three-body loss
rate scale with the tunneling time in the optical lattice. In the end, about
$60$\% of the remaining Cs atoms ($30$\% of $N_\text{Cs}^\text{BEC}$) are
paired with Rb and successfully form RbCs molecules.
For calculation of the molecules' entropy we assume that lattice sites occupied
by unassociated atoms can be emptied, e.g. by a blow-away technique
\cite{Thalhammer2006llf2,Danzl2010auh2}, without affecting the molecular sample.
The probability for an empty site is then $1-p$, where $p$ is the molecular
filling fraction, and the entropy per molecule is \cite{Budker2008apa2}
\begin{equation*}
s \approx \frac{k_\text{B}}{p} \left( p \ln \left( p \right) + \left( 1-p \right) \ln \left( 1-p \right) \right).
\end{equation*}
Taking $p=30\%$ as a lower bound gives $ s = 2$~$k_\text{B}$. This value compares
well with the one reported in Ref.~\cite{Moses2015coa2}.
\subsection{On-site interspecies interaction}
For moderate values of the interspecies scattering length $a_\text{AB}$ the
Hubbard on-site interaction between atoms A and B is given by
\begin{equation*}
U_\text{AB}=\frac{4 \pi \hbar^2}{2\mu_\text{AB}} a_\text{AB} \int{w_\text{A}^{*}(\bm{r}) w_\text{B}^{*}(\bm{r}) w_\text{B}(\bm{r}) w_\text{A}(\bm{r}) d^3r.}
\end{equation*}
Here, $\mu_\text{AB} = m_\text{A}m_\text{B}/(m_\text{A}+m_\text{B})$ is the
reduced mass and $w_{\text{A},\text{B}}$ are the Wannier functions for A and B,
respectively. Independent of the depth of the lattice we find that
$U_\text{RbCs} = - U_\text{RbRb}$ for $a_\text{RbCs} = -82$~$a_0$.
\subsection{Cs MI characterization}
At low values for $B$ (i.e. $B=21$~G, where we form the Cs BEC) the Cs sample
in the optical lattice is stable for many seconds irrespective of the lattice
depth. Increasing $B$ to high values (necessary to access the Rb-Cs Feshbach
resonance and to tune the interspecies interaction) however changes the Cs
intraspecies scattering length $a_\text{CsCs}$ dramatically. For $B=354.95$~G,
where we perform the mixing procedure, $a_\text{CsCs}$ reaches about
$2500$~$a_0$ \cite{Berninger2013fsf2}. We emphasize that in this regime it is of
crucial importance to drive the system deeply into the MI state to protect the
Cs MI from enhanced Cs-Cs-Cs three-body loss.
We analyze the stability of the Cs MI at high values for $B$ (see Fig.~4(b) of
the main article) by preparing first a Cs MI at $B=21$~G for a lattice depth
$V^\text{Cs}$. We increase $B$ within $2$~ms to $B=354.95$~G and then hold the
sample for a variable hold time $\tau$ before lowering $B$ again back to
$B=21$~G within another $2$~ms. In the end the lattice is ramped down
adiabatically and the BEC fraction of the Cs sample is determined. We find that
the BEC fraction decays exponentially with $\tau$ and determine the
$1/e$ lifetime $t_\text{MI}$ as a function of $V^\text{Cs}$. The result is
plotted in Fig.~4(b) of the main article. We note that the decay of the Cs atom
number due to Cs-Cs-Cs three-body losses is in general slower than the decay of
the BEC fraction. To maintain a stable Cs MI state over the course of the
$540$~ms that the atoms spend at high values for $B$ during the mixing process
(see section Experimental timing) the mixing lattice depth
$V_\text{mix}^\text{Cs}$ has to be at least $20$~$E_\text{rec}^\text{Cs}$
(diamond in Fig.~4(b)). Below that value a significant reduction in the RbCs
molecule production efficiency is observed (see Fig.~4(a)).
We check that we produce a relatively clean single-atom MI shell by using a
technique described in Ref.~\cite{Meinert2013qqi2}. The lattice depth along the
vertical axis is lowered to $10$~$E_\text{rec}^\text{Cs}$ to allow the atoms to
tunnel and a potential tilt along this axis is applied adiabatically. When the
tilt per lattice site approaches the Cs on-site interaction $U_\text{CsCs}$ a
quantum phase transition is driven to a density-wave ordered state where every
other lattice site along the tilt axis is occupied by a Cs-Cs atom pair. The Cs
pairs are associated to weakly bound Cs$_2$ molecules by means of a magnetic Cs
Feshbach resonance to determine the number of Cs pairs that were formed this
way. We find that about $80$\% of the entire Cs sample forms pairs, comparable
to the results presented in Ref.~\cite{Meinert2013qqi2}. Such a high pair
formation efficiency can only be achieved in a low-defect MI shell.
\subsection{Critical transport velocity}
The lattice depth $V_\text{mix}$ that is used for mixing the two samples sets a
critical transport velocity $v_\text{crit}$ for transport of Rb
\cite{Altman2005sit2}. Beyond $v_\text{crit}$ the superflow becomes chaotic in
regions with average filling $n_\text{Rb} = 1$, triggering superflow breakdown
for the entire Rb sample. The Rb sample gets stuck close to its initial
position and cannot follow the movement of the underlying Rb dipole trap.
In the vicinity of the SF-to-MI transition, for a 3D lattice, the critical
quasimomentum that corresponds to the critical transport velocity
$v_\text{crit}$ is given by \cite{Altman2005sit2}
\begin{equation*}
q_\text{crit} = \frac{{\hbar}k_\text{rec}}{\pi} \text{Re}\!\left[ \sqrt{2(1-u/u_\text{c})} \right].
\end{equation*}
Here, ${\hbar}k_\text{rec} = h/ \lambda$ is the recoil momentum set by the
lattice light, and the interaction strength is $u = U_\text{RbRb}/J_\text{Rb}$,
with the Hubbard interaction parameter $U_\text{RbRb}$ and the tunneling matrix
element $J_\text{Rb}$ for Rb. The critical value for $u$, which marks the
SF-to-MI quantum-phase transition \cite{Jaksch1998cba2}, is
$u_\text{c} \approx 34.8$, which is reached for
$V_\text{mix}^\text{Rb} \approx 13.8$~$E_\text{rec}^\text{Rb}$. The group
velocity with which the Rb sample moves through the lattice is defined by the
dispersion relation of the lowest lattice band $E(q)$ as
$v_\text{g}(q) = \partial E(q)/\partial q$ and allows us to calculate the
critical transport velocity $v_\text{crit} = v_\text{g}(q_\text{crit})$
(dash-dotted line in Fig.~4(c) of the main article).
The value for the critical quasimomentum $q_\text{crit}$ predicted in
Ref.~\cite{Altman2005sit2} was experimentally tested in Ref.~\cite{Mun2007pdf2}
with a Rb sample in a sinusoidally moving 3D optical lattice and good agreement
between theory and experiment was found. In our experiment we perform linear
transport by means of the underlying dipole trap that moves at constant speed
in the presence of a stationary 3D lattice. The breakdown in superfluid
transport is reflected by an abrupt loss of pair formation efficiency as shown
in Fig.~4(a) of the main article. Concomitantly, we find from in-situ
absorption images that the Rb sample experiences sudden inhibition of
transport. The critical transport velocity $v_\text{crit}$ as determined from
our data is much lower by at least a factor $2$ than the one predicted by
theory. Interestingly, the data are fit well by shifting the theoretical values
by $3.5$~$E_\text{rec}^\text{Rb}$ towards shallower lattice depths (dashed line
in Fig.~4(c)). The discrepancy between our experiment and theory merits further
investigations.
\subsection{Magnetic field calibration}
The magnetic field $B$ is calibrated by microwave spectroscopy. To assure that
we work under the same conditions as in the molecule formation experiment we
use the sequence for molecule creation in an optical lattice as described above
(see also Fig.~\ref{Reichsoellner_supmat_Fig1}) but load only a Cs sample,
create a Cs one-atom MI, and substitute the molecule creation procedure by a
$20$-ms microwave pulse. A microwave antenna, powered by a programmable
microwave source with $5$~W output, irradiates the Cs atoms and drives the
$\pi$ transition of the lowest Cs hyperfine states
$|f_\text{Cs},m_{f_\text{Cs}}\rangle=|3,3\rangle$ to $|4,3\rangle$. We measure
the number of Cs atoms in state $|3,3\rangle$ and $|4,3\rangle$ by means of
Stern-Gerlach separation as we scan the microwave frequency. We extract the
resonance position from Gaussian fits to the data and calculate the
corresponding value for $B$ via the Breit-Rabi formula \cite{Steck2003cdl2}. The
values for $B$ can then be converted to values for $a_\text{RbCs}$ \cite{Ruth2}.
In this way we are able to determine the zero-crossing of the interspecies
scattering length $a_\text{RbCs}$ at $354.95$~G with an accuracy of $17$~mG. We
note that the pole of the interspecies Feshbach resonance together with its
adjacent zero crossing leave significant signatures in a different set of
experiments in which we probe the interference contrast of the superfluid Rb
sample as we scan across the Feshbach resonance. In particular, at a certain
value for $B$ we observe a drastic loss of interference contrast. This we
associate with the pole of the resonance. The position of the pole agrees with
what we expect from our $B$-field calibration.
Magnetic field stability is crucial during the mixing process when
$B \approx 355$~G. The most dominant magnetic noise contributions are at
multiples of the line frequency $50$~Hz. A feed-forward technique allows us to
suppress Fourier components at $50$~Hz, $100$~Hz, $150$~Hz, and $250$~Hz. For
this, a pick-up coil senses their phase and amplitude. During the mixing
procedure these Fourier components are then added to $B$ with a $\pi$ phase
shift synchronized to line. We thereby suppress the total magnetic field noise
down to $50$~mG rms, corresponding to about $15$-$a_0$ rms uncertainty for the
interspecies scattering length in the vicinity of its zero crossing. The
magnetic field gradient $\left| \nabla B \right| =25.9$~$\text{G}/\text{cm}$
causes an additional shift of $16$~$a_0$ across the typical $18$-$\mu$m
atom-sample diameter.
|
\section{Introduction}
Equilibrium is a central concept of statistical mechanics. In Boltzmannian statistical mechanics (BSM) equilibrium is standardly associated with the largest macro-region, where macro-regions are parts of the accessible phase space consisting of micro-states that are the supervenience base for the same macro-state. In two recent papers we argue that the standard picture lacks a foundation and ought to be replaced by an alternative approach (Werndl and Frigg 2015a, 2015b). We develop this approach in detail under the assumption that the underlying micro-dynamics is deterministic. In this paper we give up this assumption and generalise our approach to systems with a stochastic micro-dynamics.\\
In Section \ref{Rethought} we introduce the main pillars of our programme. In Section \ref{Stochastic} we present stochastic systems. In Section \ref{S-Equilibrium} we carry over our key concepts from the deterministic to the stochastic context and formulate the main theorems, which we prove in the Appendix. In Section \ref{Example} we illustrate our claims with the example of the lattice gas, an important and widely used model in physics. In Section \ref{Conclusion} we summarise our results and add some concluding remarks.
\section{Boltzmannian Equilibrium Rethought}\label{Rethought}
In this section we briefly present the new definition of equilibrium we proposed in previous work (Werndl and Frigg 2015a, 2015b). Consider a system consisting of $n$ particles in an isolated and bounded container. The system's micro-state is a point $x$ in its $6n$-dimensional state space $\Gamma$. The system's dynamics is given by a deterministic time evolution $\phi_{t}$, where $\phi_{t}(x)$ is the state into which $x \in \Gamma$ evolves after $t$ time steps. The system's energy is preserved and so the system's motion is confined to the energy hypersurface $\Gamma_{E}$. The hypersurface is equipped with a sigma algebra $\Sigma_{E}$ and a normalised measure $\mu_{E}$ which is invariant under $\phi_{t}$. Taken together these elements constitute the measure-preserving deterministic dynamical system $(\Gamma_{E},\Sigma_{E},\mu_{E},\phi_{t})$.\\
From a macroscopic point of view the system can be characterised by a set $\{v_{1}, ..., v_{k}\}$ of macro-variables ($k \in \field{N}$). The $ v_{i}$ are functions on $\Gamma_{E}$ that assume values in the range $\field{V}_{i}$, and capital letters $V_{i}$ denote the values of $v_{i}$. A particular set of values $\{V_{1}, ..., V_{k}\}$ defines a \emph{macro-state} $M_{V_{1}, \ldots, V_{k}}$. A set of macro-states is complete iff (if and only if) it contains all states a system can be in. In Boltzmannian statistical mechanics macro-states supervene on micro-states and hence every macro-state $M$ is associated with a macro-region $\Gamma_{M}$ consisting of all $x\in\Gamma_{E}$ for which the system is in $M$. For a complete set of macro-states the $\Gamma_{M}$ form a partition of $\Gamma_{E}$.\newline
The equilibrium macro-state is $M_{eq}$ and its macro-region is $\Gamma_{M_{eq}}$. A crucial aspect of the standard presentation of BSM is that $\Gamma_{M_{eq}}$ is the largest macro-region. The notion of the `largest macro-region' can be interpreted in two ways. The first takes `largest' to mean that the equilibrium macro-region takes up a large part of $\Gamma_{E}$. We say that $\Gamma_{M_{eq}}$ is \emph{$\beta$-dominant} iff $\mu_{E}(\Gamma_{M_{eq}}) \geq\beta$ for a particular $\beta\in (\frac{1}{2},1]$. If $\Gamma_{M_{eq}}$ is $\beta$-dominant, then it is in fact also $\beta'$-dominant for all $\beta'$ in $(1/2, \, \beta)$. The second reading takes `largest' to mean `larger than any other macro-region'. We say that $\Gamma_{M_{eq}}$ is \emph{$\delta$-prevalent} iff $\min_{M \neq M_{eq}} [\mu_{E}(\Gamma_{M_{eq}}) -\mu_{E}(\Gamma_{M})]\geq\delta$ for a particular real number $\delta > 0$. This implies that if $\Gamma_{M_{eq}}$ is $\delta$-prevalent, then it is also $\delta'$-prevalent for all $\delta'$ in $(0, \, \delta)$. We do not adjudicate between these different definitions - either meaning of `large' can be used to define equilibrium. It ought to be pointed out, however, that they are not equivalent: whenever an equilibrium macro-region is $\beta$-dominant, there exists a range of values for $\delta$ so that the macro-region is also $\delta$-prevalent for these values, but the converse fails.\\
The question now is: why is the equilibrium state $\beta$-dominant or $\delta$-prevalent? A justification ought to be as close as possible to thermodynamics. In thermodynamics a system is in equilibrium when change has come to a halt and all thermodynamic variables assume constant values (cf. Reiss 1996, 3). This would suggest a definition of equilibrium according to which every initial condition lies on a trajectory for which $\{v_{1}, ..., v_{k}\}$ eventually assume constant values. Yet this is unattainable for two reasons. First, the values of the $v_{i}$ will never cease to fluctuate due to Poincar\'{e} recurrence. This, however, is no cause for concern. Experimental results show that systems exhibit fluctuations away from equilibrium (Wang et al.\ 2002), and hence the strict thermodynamic notion of equilibrium is actually \emph{unphysical}. Second, in dynamical systems we cannot expect \textit{every} initial condition to approach equilibrium (see, for instance, Callender 2001). For these reasons we define equilibrium as the macro-state in which trajectories starting in most initial conditions spend most of their time.\\
To make this idea precise we introduce the long-run fraction of time a system spends in a region $A\in \Sigma_{E}$ when the system starts in micro-state $x$ at time $t=0$:
\begin{eqnarray}\label{LF}
LF_{A}(x)&=&\lim_{t\rightarrow\infty}\frac{1}{t}\int_{0}^{t}1_{A}(\phi_{\tau}(x))d\tau\,\,\textnormal{for continuous time, i.e.}\,\,t\in\field{R},\,\,\\
LF_{A}(x)&=&\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1}1_{A}(\phi_{\tau}(x))\,\,\textnormal{for discrete time, i.e.}\,\,t\in\field{Z},\nonumber
\end{eqnarray}
\noindent where $1_{A}(x)$ is the characteristic function of $A$, i.e.\ $1_{A}(x)=1$ for $x\in A$ and $0$ otherwise.\\
The notion `most of their time' is beset with the same ambiguity as the `largest macro-state'. On the first reading most of the time means more than half of the total time. This leads to the following formal definition of equilibrium:
\begin{quote}
\textit{BSM $\alpha$-$\varepsilon$-Equilibrium.} Consider an isolated system $S$ whose macro-states are specified in terms of the macro-variables $\{v_{1}, ..., v_{k}\}$ and which, at the micro level, is a measure-preserving deterministic dynamical system $(\Gamma_{E},\Sigma_{E},\mu_{E},\phi_{t})$. Let $\alpha$ be a real number in $(0.5, 1]$, and let $1 \gg \varepsilon \ge 0$ be a very small real number. If there is a macro-state $M_{V_{1}^{*}, ..., V_{k}^{*}}$ satisfying the following condition, then it is the $\alpha$-$\varepsilon$-equilibrium state of $S$: There exists a set $Y\subseteq \Gamma_{E}$ such that $\mu_{E}(Y)\geq 1-\varepsilon$, and all initial states $x\in Y$ satisfy
\begin{equation}
LF_{\Gamma_{M_{V_{1}^{*}, ..., V_{l}^{*}}}}\!(x) \, \geq \, \alpha.
\end{equation}
We then write $M_{\alpha\textnormal{-}\varepsilon\textnormal{-}eq}\, := \, M_{V_{1}^{*}, ..., V_{k}^{*}}$.
\end{quote}
\noindent An obvious question concerns the value of $\alpha$. Often the assumption seems to be that $\alpha$ is close to one. This is reasonable but not the only possible choice. For our purposes nothing hangs on a particular choice of $\alpha$ and so we leave it open what the best choice would be.\\
On the second reading `most of the time' means that the system spends more time in the equilibrium macro-state than in any other macro-state. This idea can be rendered precise as follows:
\begin{quote}
\textit{BSM $\gamma$-$\varepsilon$-Equilibrium.}
Consider an isolated system $S$ whose macro-states are specified in terms of the macro-variables $\{v_{1}, ..., v_{k}\}$ and which, at the micro level, is a measure-preserving deterministic dynamical system $(\Gamma_{E},\Sigma_{E},\mu_{E},\phi_{t})$. Let $\gamma$ be a real number in $(0, 1]$ and let $1 \gg \varepsilon \ge 0$ be a very small real number so that $\gamma>\varepsilon$. If there is a macro-state $M_{V_{1}^{*}, ..., V_{l}^{*}}$ satisfying the following condition, then it is the $\gamma$-$\varepsilon$-equilibrium state of $S$: There exists a set $Y\subseteq \Gamma_{E}$ such that $\mu_{E}(Y)\geq 1-\varepsilon$ and for all initial conditions $x\in Y$:
\begin{equation}
LF_{\Gamma_{M_{V_{1}^{*}, ..., V_{l}^{*}}}}\!(x) \, \geq \,
LF_{\Gamma_{M}}\!(x)+\gamma
\end{equation}
\noindent for all macro-states $M \neq M_{V_{1}^{*}, ..., V_{l}^{*}}$. We then write $M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}\, := \, M_{V_{1}^{*}, ..., V_{k}^{*}}$.
\end{quote}
\noindent As above, nothing in what we say about equilibrium depends on the particular value of the parameter $\gamma$ and so we leave it open what the best choice would be.\\
We contend that these two definitions provide the relevant notions of equilibrium in BSM. But the definitions remain silent about the size of equilibrium macro-regions, and they do not in any obvious way imply anything about seize. These regions being extremely small would be compatible with the definitions. That these macro-regions have the right size is a result established in the following two theorems:
\begin{quote}
\emph{Deterministic Dominance Theorem}: If $M_{\alpha\textnormal{-}\varepsilon\textnormal{-}eq}$ is an $\alpha$-$\varepsilon$-equilibrium of system $S$, then $\mu_{E}(\Gamma_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-}eq}}) \geq \beta$ for $\beta=\alpha(1-\varepsilon)$.\footnote{We assume that $\varepsilon$ is small enough so that $\alpha(1-\varepsilon)> \frac{1}{2}$.}
\end{quote}
\begin{quote}
\emph{Deterministic Prevalence Theorem}: If $M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}$ is a $\gamma$-$\varepsilon$-equilibrium of system $S$, then $\mu_{E}(\Gamma_{M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}}) \geq \mu_{E}(\Gamma_{M})+\gamma-\varepsilon$ for all macro-states $M\neq M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}$.
\end{quote}
Both theorems are completely general in that no dynamical assumptions are made.\footnote{We assume that the dynamics is stationary, i.e. that $\phi_{t}$ does not depend on time explicitly. This, however, is not a substantive assumption in the current context because standard systems in statistical mechanics such as gases and crystals are stationary.} Thus the theorems also apply to strongly interacting systems. It is worth highlighting that the theorems make the conditional claim that if an equilibrium exits, then it is large in the relevant sense. Some systems have equilibria and for these the theorem holds. For instance the baker's gas (a gas consisting of $N$ copies of the baker's transformation) has an equilibrium in the requisite sense and the relevant macro-region is large (see Lavis (2005) for a discussion of the baker's gas). Other systems don't have equilibria, and for these the antecedent of the conditional is not satisfied. If, for instance, the dynamics is given by the identity function, no approach to equilibrium takes place.\\
There are many systems in statistical mechanics where a stochastic dynamics is considered. Important examples include the Ising model, the lattice gas, the six vertex model and the eight vertex model (cf.\ Baxter 1982; Lavis and Bell 1999). Hence the above definitions and results do not apply to them and so the question arises whether they can be carried over to stochastic systems. We now introduce stochastic systems and then show that such a generalisation is possible.
\section{Stochastic Processes}\label{Stochastic}
In order to introduce stochastic processes, we first need to define random variables. Intuitively, a random variable $Z$ gives the outcome of a probabilistic experiment, where the distribution $P\{Z\in A\}$ specifies the probability that the outcome will be in a certain range $A$. Formally, a \textit{random variable} is a measurable function $Z:\Omega\rightarrow\bar{X}$, where $(\Omega,\Sigma_{\Omega},\nu)$ is a probability space and $(\bar{X},\Sigma_{\bar{X}})$ measurable space. $\bar{X}$ is the set of possible outcomes and is therefore referred to as the \textit{outcome space}. The \textit{probability measure} $P\{Z\in A\}=\nu(Z^{-1}(A))$ for all $A\in\Sigma_{\bar{X}}$ is called the \textit{distribution} of $Z$. \\
A stochastic process consists of a string of the kind of probabilistic experiments that are described by a random variable. Formally, a \emph{stochastic process} $\{Z_{t}\}$, $t\in\field{R}$ (continuous time) or $\field{Z}$ (discrete time), is a family of random variables $Z_{t}$, which are defined on the same probability space $(\Omega,\Sigma_{\Omega},\nu)$ and take values in the same measurable space $(\bar{X},\Sigma_{\bar{X}})$ such that $Z(t,\omega)=Z_{t}(\omega)$ is jointly measurable in $(t,\omega)$. Intuitively speaking, each $\omega$ encodes an entire possible history (present and future) of a stochastic process, and thus $\Omega$ is the set of all possible histories the stochastic process (we illustrate this idea with a simple example below). A \textit{realisation} is a possible path of the stochastic process. That is, it is a function $r_{\omega}:\field{R}\rightarrow\bar{X}$, $r_{\omega}(t)=Z(t,\omega)$, for $\omega\in\Omega$ arbitrary (cf. ~Doob 1953,~4--46). The difference between $\omega$ and $r_{\omega}$ is simply that while $r_{\omega}$ gives a possible path of the stochastic process in terms of sequences of elements of $\bar{X}$, $\omega$ just \textit{encodes} such a possible history.\\
If the random variable does not depend explicitly on time (if, for instance, the outcome does not depend on when you toss a coin), then we have a stationary stochastic process and in what follows all the stochastic processes we will be working with will be assumed to be stationary. Formally: A stochastic process $\{Z_{t};\, t\in \field{Z}\}$ is \emph{stationary} iff the distributions of the multi-dimensional random variable $(Z_{t_{1}+h},\ldots, Z_{t_{n}+h})$ is the same as the one of $(Z_{t_{1}},\ldots, Z_{t_{n}})$ for all $t_{1},\ldots,t_{n}\in \field{R}$ or $\field{Z}$, $n\in\field{N}$, and all $h\in \field{Z}$ or $\field{R}$ (—\textit{ibid.}).\\
Let us now give an example, namely the discrete stochastic process that describes a bi-infinite series of coin tosses of a fair coin with probability $p_{H}=1/2$ (`Heads') and $p_{T}=1/2$ (`Tails'), $p_{H\cup T}=1$ and $p_{\emptyset}=0$. In this case $\bar{X}=\{H,T\}$ and $\Sigma_{\bar{X}}$ is the power set of $\bar{X}$. $\Omega$ is the set encoding all possible histories of the stochastic process. That is, $\Omega$ is defined as the set of all sequences $\omega=(\ldots\omega_{-1}\omega_{0}\omega_{1}\ldots)$ with
$\omega_{i}\in\bar{X}$ corresponding to one of the possible
outcomes of the $i$-th trial in a doubly infinite sequence of
trials. $\Sigma_{\Omega}$ is the $\sigma$-algebra generated by
the cylinder-sets
\begin{equation}\label{cylinder}
C^{G_{1}...G_{n}}_{i_{1}...i_{n}}\!\!=\!\!\{\omega\in
\Omega\,|\,\omega_{i_{1}}\!\!\in\!\!G_{1},\!\ldots\!,\omega_{i_{n}}\!\!\in\!\!G_{n},
G_{j}\in\Sigma_{\bar{X}}\!,\,i_{j}\!\in\!\field{Z},\,i_{1}\!<\!\ldots\!<\!i_{n},\,1\!\leq j\!\leq\!n\}.
\end{equation}
Since the outcomes are independent, these sets have probability
$\bar{\nu}(C^{G_{1}...G_{n}}_{i_{1}...i_{n}}):=p_{G_{1}} \times \ldots \times p_{G_{n}}$. Let $\nu$ be defined as the unique extension of $\bar{\nu}$ to a measure on $\Sigma_{\Omega}$. Finally, define $Z_{t}(\omega):=\omega_{t}$ (the $t$-th coordinate of $\omega$). Then $Z_{t}(\omega)$ gives us the outcome of the coin toss at time $t$, $P\{Z_{t}=H\}=\nu(Z_{t}^{-1}(\{H\}))=1/2$ and $P\{Z_{t}=T\}=\nu(Z_{t}^{-1}(\{T\}))=1/2$ for any $t$. Hence $\{Z_{t}\}$ is the stochastic process describing an infinite series of tosses of a fair coin, and it is also clear that this process is stationary.\footnote{Here we can also illustrate the difference between an $\omega$ and a realisation $r(\omega)$. We could, for instance, also use `0' and `1' to encode the path of a stochastic process (where `0' encodes the outcome Heads and `1' encodes the outcome Tails). Then $\Omega$ would consist of sequences such as $\omega=(\ldots,0,1,0,1,\ldots)$, but $r(\omega)=(\ldots H,T,H,T,\ldots)$. More radically, we could also use a real number $\omega\in[0,1]$ to encode a sequence of 0s and 1s (via its binary development) and thus a sequence of outcomes of tossing a coin.}\\
\section{Equilibrium for Stochastic Processes}\label{S-Equilibrium}
Let us now return to BSM as introduced in the previous section. In the context of stochastic processes $\bar{X}$ plays the role of $\Gamma_{E}$ as giving the set of possible outcomes of the system. $Z_{t}(\omega)$ is the stochastic equivalent of $\phi_{t}(x)$ in that it gives the state of the system at time $t$. More specifically, the dynamics is determined by the probability measure $\nu$, from which transition probabilities (such as\ $P\{Z_{t}=H\,\,|$ given that $Z_{t-1}=T\}$) can be derived. These are the stochastic equivalent of $\phi_{t}(x)$ because they specify how the system evolves over time. Realisations are the stochastic equivalent of trajectories in the deterministic case in that they describe possible evolutions of the system. The probability measure $P$ defined on $\bar{X}$ is the stochastic equivalent of $\mu_{E}$ because it gives the probability of certain outcomes. Finally, the condition of stationarity is the stochastic analogue of the condition that $\mu_{E}$ is invariant in the deterministic case.\\
The macro characterisation of the system does not change, and so we consider again the macro-variables $\{v_{1}, ..., v_{k}\}$. The mathematical expression of supervenience is that the $v_{i}$ are functions on $\bar{X}$. That is, $v_{i}:\bar{X}\rightarrow \field{V}_{i}$. As above, a particular set of values $\{V_{1}, ..., V_{k}\}$ defines a \emph{macro-state} $M_{V_{1}, \ldots, V_{r}}$, and a complete set of macro-states contains all states as system can be in. Again, every macro-state $M$ is associated with a \textit{macro-region} $\bar{X}_{M}$ consisting of all $\bar{x}\in \bar{X}$ for which the system is in $M$. The definitions of prevalence and dominance carry over to the current context unchanged. That is, a macro-region $\bar{X}_{M_{eq}}$ is \emph{prevalent} iff $P\{\bar{X}_{M_{eq}}\}>P\{\bar{X}_{M}\}+\gamma$ for some $\gamma \in (0, 1]$ for all $M\neq M_{eq}$, and $\bar{X}_{M_{eq}}$ is $\beta$-dominant
iff $P\{\bar{X}_{M_{eq}}\}\geq \beta$ for some $\beta\in (\frac{1}{2},1]$.\\
The aim now is to carry over the above definitions of equilibrium from the deterministic to the stochastic context. To this end we first have to introduce the notion of the long-run fraction of time a realisation spends in a region $A\in \Sigma_{\bar{X}}$:
\begin{eqnarray}
LF_{A}(\omega)&=&\lim_{t\rightarrow\infty}\frac{1}{t}\int_{0}^{t}1_{A}(Z_{\tau}(\omega))d\tau\,\,\textnormal{for continuous time, i.e.}\,\,t\in\field{R},\,\,\\
LF_{A}(\omega)&=&\lim_{t\rightarrow\infty}\frac{1}{t}\sum_{\tau=0}^{t-1}1_{A}(Z_{\tau}(\omega))\,\,\textnormal{for discrete time, i.e.}\,\,t\in\field{Z}.
\end{eqnarray}
\noindent We are now in a position to state the stochastic definitions of equilibrium:
\begin{quote}
\textit{Stochastic $\alpha$-$\varepsilon$-Equilibrium.} Consider an isolated system $S$ whose macro-states are specified in terms of the macro-variables $\{v_{1}, ..., v_{k}\}$ and which, at the micro level, is a stationary stochastic process $\{Z_{t}\}$. Let $\alpha$ be a real number in $(0.5, 1]$, and let $1 \gg \varepsilon \ge 0$ be a very small real number. If there is a macro-state $M_{V_{1}^{*}, ..., V_{k}^{*}}$ satisfying the following condition, then it is the stochastic $\alpha$-$\varepsilon$-equilibrium state of $S$: There exists a set $\Omega^{*} \subseteq \Omega$ such that $\nu(\Omega^{*})\geq 1-\varepsilon$, and for all $\omega\in\Omega^{*}$:
\begin{equation}
LF_{\bar{X}_{M_{V_{1}^{*}, ..., V_{k}^{*}}}}(\omega) \geq \alpha.
\end{equation}
We then write $M_{\alpha\textnormal{-}\varepsilon\textnormal{-}eq}\, := \, M_{V_{1}^{*}, ..., V_{k}^{*}}$.
\end{quote}
\noindent The definition of the $\gamma$-$\varepsilon$-equilibrium is now straightforward:
\begin{quote}
\textit{Stochastic $\gamma$-$\varepsilon$-Equilibrium.} Consider an isolated system $S$ whose macro-states are specified in terms of the macro-variables $\{v_{1}, ..., v_{k}\}$ and which, at the micro level, is a stationary stochastic process $\{Z_{t}\}$. Let $\gamma$ be a real number in $(0, 1]$, and let $1 \gg \varepsilon \ge 0$ be a very small real number so that $\varepsilon<\gamma$. If there is a macro-state $M_{V_{1}^{*}, ..., V_{k}^{*}}$ satisfying the following condition, then it is the stochastic $\alpha$-$\varepsilon$-equilibrium state of $S$: There exists a set $\Omega^{*} \subseteq \Omega$ such that $\nu(\Omega^{*})\geq 1-\varepsilon$, and all $\omega\in\Omega^{*}$ satisfy
\begin{equation}
LF_{\bar{X}_{M_{V_{1}^{*}, ..., V_{k}^{*}}}}(\omega)\geq LF_{\bar{X}_{M}}(\omega)+\gamma
\end{equation}
\noindent for all $M\neq M_{V_{1}^{*}, ..., V_{k}^{*}}$. We then write $M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}\, := \, M_{V_{1}^{*}, ..., V_{k}^{*}}$.
\end{quote}
The core result of this paper is that the two central theorems of the deterministic case, the Dominance Theorem and the Prevalence Theorem, have stochastic analogues. We now state the theorems and give the proof in the Appendix.
\begin{quote}
\emph{Stochastic Dominance Theorem}: If $M_{\alpha\textnormal{-}\varepsilon\textnormal{-}eq}$ is a stochastic $\alpha$-$\varepsilon$-equilibrium of system $S$, then $P\{\bar{X}_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-}eq}}\} \geq \beta$ for $\beta=\alpha(1-\varepsilon)$.\footnote{We assume that $\varepsilon$ is small enough so that $\alpha(1-\varepsilon)> \frac{1}{2}$.}
\end{quote}
\begin{quote}
\emph{Stochastic Prevalence Theorem}: If $M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}$ is a stochastic $\gamma$-$\varepsilon$-equilibrium of system $S$, then $P\{\bar{X}_{M_{\varepsilon\textnormal{-}eq}}\} \geq P\{\bar{X}_{M}\}+\gamma-\varepsilon$ for all macro-states $M$ with $M\neq M_{\gamma\textnormal{-}\varepsilon\textnormal{-}eq}$.
\end{quote}
As in the deterministic case, both theorems are completely general in that no dynamical assumptions are made and hence the theorems apply to stochastic process with any dynamics.\footnote{We assume that the dynamics is stationary, but, as in the deterministic case, this is not a substantive assumption because standard stochastic systems in statistical mechanics are stationary.} As in the deterministic case it is worth noting that the theorems make the conditional claim that if an equilibrium exits, then it is large in the relevant sense. There are processes that do not have an equilibrium. For instance, consider the stochastic process of throwing a fair die (with six sides). Suppose that the macro-variable of concern is whether the die shows an even number (2, 4, 6) or an odd number (1, 3, 5). Then there will be no equilibrium because for almost any realisation half of the time the dice will show an even number and half of the time they will show an odd number.\\
\section{Example: The Lattice Gas}\label{Example}
We now illustrate the definitions and theorems of the previous section with the lattice gas. The lattice gas is a popular model not only of gases (as its name would suggest), but in fact also of liquids and solids.\footnote{See Baxter (1982) and Cipra (1987) for more details about the lattice gas.} The lattice gas models a fluid in the sense that flows are represented by particles moving from site to site, and because the system is in contact with an energy and particle reservoir, particles can also be created and annihilated. More specifically, consider a lattice with $N\in\field{N}$ sites. Each lattice site can either be occupied by a particle or be empty. This is formalised by associating with every lattice site $i$ a variable $s_{i}$, which takes the value 1 if the site is occupied and 0 if the site is empty. Thus the micro-state of the lattice is a vector $s=(s_1,\ldots,s_N)$, specifying which sites are occupied and which ones are empty. Hence the system's $\bar{X}$ consists of the $2^N$ possible arrangements of different numbers of particles on the $N$ sites, and $\Sigma_{\bar{X}}$ is the power set of $X$. Now the elements of $\Omega$ encode the history, present and future of the stochastic process in all its details (for the lattice gas discrete time steps are considered). That is, $\Omega$ consists of all bi-infinite sequences $\omega=(\ldots\omega_{-1}\omega_{0}\omega_{1}\ldots)$ where the $i$-th coordinate $\omega_{i}$ is an arbitrary vector $s$. $\Sigma_{\Omega}$ is the $\sigma$-algebra generated by cylinder sets that are described in Equation \ref{cylinder} if we replace the $G$s by $B$s. Finally, $Z_{t}(\omega):=\omega_{t}$ (the $t$-th coordinate of $\omega$).\\
The probability measure $\nu$ depends on the exact stochastic dynamics of the system. Many different kinds of stationary stochastic dynamics are considered for the lattice gas model (cf. Baxter 1982; Cipra 1988). At this point it is not necessary to commit to any specific stochastic dynamics. It suffices to say that a stochastic dynamics will determine the measure assigned to the cylinder sets $\bar{\nu}(C^{B_{1}...B_{n}}_{i_{1}...i_{n}})$ and its unique extension $\nu$. What we need to mention, however, is that the potential energy and the grand-canonical probability distribution will constrain the dynamics. The simplest still somewhat realistic potential is the so-called square-well potential, where only nearest neighbour interactions are taken into account. The underlying idea is that there cannot be two particles on the same site, that particles are attracted when they are close to each other and that no interaction takes place when they are far apart.
\begin{eqnarray}\label{NNI}
\phi(i,j) & = & \left\{
\begin{array}{ll}
\infty & \mbox{if $i=j$} \\
-\xi & \mbox{if $i,j$ are nearest neighbours}\\
0 & \mbox{otherwise}
\end{array}
\right\},
\end{eqnarray}
\noindent where $i$ and $j$ denote sites of the lattice and $\xi>0$. The total potential energy of the system is given by $E(s)=\sum_{i,j}\phi(i,j)s_{i}s_j$, where the sum is over all pairs of sites on the lattice (with $i\neq j$).\\
The probability measure of a set of micro-states $A$ is given by the grand-canonical probability distribution $P\{A\}$. This distribution depends on the effective chemical potential $\mu_{c}$ (one can think of the chemical potential as a measure for how accepting the system is of new particles, or for how much work one has to do to add a new particle to the system):
\begin{equation}\label{probLG}
\frac{\sum_{s\in A}\exp[\frac{\xi}{4kT}(\sum_{\textnormal{all}\,\,i,j}(2s_{i}-1)(2s_{j}-1)+
(2\mu_{c}+q\xi)\sum_{all\,\,i}(2s_{i}-1)+N(\frac{1}{2}q\xi+2\mu_{c}))]}{
\sum_{\textnormal{all}\,\,s}\exp[\frac{\xi}{4kT}(\sum_{\textnormal{all}\,\, i,j}(2s_{i}-1)(2s_{j}-1)+(2\mu_{c}+q\xi)\sum_{all\,\, i}(2s_{i}-1)+N(\frac{1}{2}q\xi+2\mu_{c}))]},
\end{equation}
where $k$ is the Boltzmann constant, $T$ is the temperature and $q$ is the number of nearest-neighbours. For any stochastic dynamics that satisfies the constraints that the potential energy is given by Equation (\ref{NNI}) and that the probabilities are specified by the grand-canonical partition function, $\{Z_{t}\}$ is a stochastic process describing the lattice gas.\\
The \emph{macro-states} usually considered are defined by the average particle density per site over the entire system: $M^{LG}_j=j/N$ where $j$ denotes the total number of molecules $s_{1}+s_2\ldots+s_N$. The \emph{macro-regions} $\bar{X}_{M^{LG}_j}$ are defined as the set of micro-states for which the system is in macro-state $M^{LG}_j$. \\
The behaviour of the lattice gas depends on the values of the various parameters. For the purpose of illustrating our ideas, we will consider two kinds of behaviour (corresponding to ranges of parameter values). First, consider a sufficiently large $\mu_{c}$ (that is, when the system readily accepts new particles). In this case, under the usual stochastic dynamics considered, the system will spend most of the time in the macro-state where all sites are occupied, i.e.\ in $M^{LG}_N$, for almost all initial states (in a measure-theoretic sense) (cf. \ Baxter 1982). For this reason $M^{LG}_N$ is a $\gamma$-$0$ equilibrium. Thus, by the Stochastic Prevalence Theorem, $M^{LG}_N$ is $\gamma$-prevalent.\footnote{Note that it is also clear from Equation~(\ref{probLG}) that for sufficiently large $\mu_{c}$, $M^{LG}_N$ corresponds to the largest macro-region.}\\
Second consider a sufficiently small negative-valued $\mu_{c}$ (in which case the system tends to annihilate particles). Then, under the usual stochastic dynamics considered, the system will spend most of the time in the macro-state where all sites are empty, i.e.\ in $M^{LG}_0$, or almost all initial states (cf.\ Baxter 1982). Therefore, $M^{LG}_0$ is a $\gamma$-$0$-equilibrium. Thus, by the Stochastic Dominance Theorem, $M^{LG}_N$ is $\gamma$-prevalent.\footnote{Again, this is clear from Equation~(\ref{probLG}).}\\
To conclude, \emph{the lattice gas represents an important physical system that has equilibria in our sense}. Let us end with a few remarks on why this system is physically important. First, it provides a good model of \emph{condensation} and the \emph{liquid-gas transition}.\footnote{Mathematically speaking, the lattice gas is equivalent to the \emph{Ising model}. The Ising model is one of the best developed and most widely studied models in physics and is discussed in nearly every modern textbook on statistical mechanics. In particular, the lattice gas on a square lattice with $\mu_{C}=-\xi/8$ is equivalent to the two-dimensional Ising model with no external field, which is famous for being one of the very few exactly solved models that display \emph{phase transitions} (Baxter 1982).} A theory of condensation was developed based on the lattice model, which was shown to qualitatively reproduce the main features of condensation and was empirically confirmed for many cases (Kierlik et al.\ 2002; Young and Lee 1952). To give an example, De Ribaupierre and Manchester (1974) found that the lattice gas provides a satisfactory model of condensation for a hydrogen in palladium system. Pan et al.\ (1998) found that the lattice gas gives a fair description of the liquid-gas transition in excited nuclear systems formed as a result of a heavy ion collisions. Finally, the lattice gas also models \emph{melting and freezing} phenomena well (see Kikuchi and Cahn 1980). For instance, Clarke et al.\ (1979) found that the lattice gas model provides a good description of melting for graphite intercalated with caesium.\\
\section{Conclusion}\label{Conclusion}
We presented stochastic formulations of the notions of an $\alpha$-$\varepsilon$-equilibrium and a $\gamma$-$\varepsilon$-equilibrium, and we have formulated and proven stochastic equivalents of the Dominance Theorem and the Prevalence Theorem. This completes the transfer of the basic notions of our framework from the deterministic to the stochastic context. There is, however, an important disanalogy between the two contexts as far as the existence of an equilibrium state is concerned. In the deterministic context we were able to prove an existence theorem (Werndl and Frigg 2015, pp. 26-29). There is no straightforward generalisation of this theorem to the stochastic context. This is because the conditions that need to hold for an equilibrium to exist in the existence theorem are conditions on the ergodic components. However, stochastic processes do not have such ergodic components. It is true that the deterministic representation of a stochastic process (cf. the Appendix for a definition) has ergodic components. However, these are components of $\Omega$ and not of $\bar{X}$, and an existence theorem would need to be about $\bar{X}$. Thus, as far as we see, the ergodic components of $\Omega$ are not useful to characterise the circumstances under which equilibria exist for stochastic processes. Hence there is an open question about when a stochastic equilibrium exists.
\pagebreak
\section{Appendix}
\subsection{Proof of the Stochastic Dominance Theorem}\label{Appendix1}
First of all, let us show that a stationary stochastic process $\{Z_{t}\}$ can be represented by a measure-preserving deterministic system $(X,\Sigma_{X},\mu_{X},T_{t})$. Let $X$ be the set of all possible realisations, i.e., functions $x(\tau)$ from $\field{R}$ or $\field{Z}$ to $\bar{X}$. Let $\Sigma_{X}$ be the $\sigma$-algebra generated by the cylinder-sets
\begin{equation}\label{cylinder}
C^{A_{1}...A_{n}}_{i_{1}...i_{n}}\!\!=\!\!\{x\!\in\! X\,|\,x(i_{1})\!\!\in\!\! A_{1},...,x(i_{n})\!\!\in\!\! A_{n}, A_{j}\!\in\! \Sigma_{\bar{X}},i_{j}\!\in\!\field{R}\,\,\textnormal{or}\,\, \field{Z},\,i_{1}\!\!<...<\!\!i_{n},1\!\leq j\!\leq n\}.
\end{equation}
Let $\mu_{X}$ be the unique probability measure arising by assigning to each $C^{A_{1}...A_{n}}_{i_{1}...i_{n}}$ the probability $P\{Z_{i_{1}}\in A_{1},\ldots,Z_{i_{n}}\in A_{n}\}$. The evolution functions shift a realisation $t$ times to the left, i.e., $T_{t}(x(\tau))=x(\tau+t)$. The $T_{t}$ are invariant under the dynamics because $\{Z_{t}\}$ is stationary. $(X,\Sigma_{X},\mu_{X},T_{t})$ is a measure-preserving deterministic system called the \emph{deterministic representation} (cf.~Doob 1953,~621--622; Werndl 2009, 2011).\\
Let $W=\{x(\tau)\in X\,\,|\,\,x(\tau)=Z_{\tau}(\omega)$ for all $\tau$ for a $\omega\in\Omega^{*}\}$. Note that $\mu_{X}(W)\geq1-\varepsilon$. Consider first the case of an $\alpha$-$\varepsilon$-equilibrium $M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}$. Then it follows that for all $x\in W$:
\begin{equation}\label{wuzi}
LF_{X_{Q_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}}}}(x)\geq \alpha,
\end{equation}
where $Q_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}}=\{x\in X\,\,|\,\,x(0)\in \bar{X}_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}}\}$.\\
Hence $Q_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}}$ is an $\alpha$-$\varepsilon$-equilibrium of $(X,\Sigma_{X},\mu_{X},T_{t})$. It follows from the (deterministic) Dominance Theorem (Frigg and Werndl 2015a) that $\mu_{X}(Q_{M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}})> \alpha(1-\varepsilon)$, which immediately implies that $P\{M_{\alpha\textnormal{-}\varepsilon\textnormal{-eq}}\}> \alpha(1-\varepsilon)$.\\
\subsection{Proof of the Stochastic Prevalence Theorem}\label{Appendix}
The proof proceeds in the same fashion as the previous one. That is, consider again the measure-preserving deterministic system $(X,\Sigma_{X},\mu_{X},T_{t})$ that represents the stationary stochastic process $\{Z_{t}\}$. Suppose that $M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}$ is an $\gamma$-$\varepsilon$-equilibrium. \\
As before, let $W=\{x(\tau)\in X\,\,|\,\,x(\tau)=Z_{\tau}(\omega)$ for all $\tau$ for a $\omega\in\Omega^{*}\}$. Note that $\mu_{X}(W)\geq 1-\varepsilon$.\\
Then for all $x\in W$ and all $M\neq M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}$ it holds that
\begin{equation}LF_{X_{Q_{M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}}}}(x)\geq LF_{X_{Q_{M}}}+\gamma-\varepsilon,
\end{equation}
where
$Q_{M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}}=\{x\in X\,\,|\,\,x(0)\in \bar{X}_{M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}}\}$ and
$Q_{M}=\{x\in X\,\,|\,\,x(0)\in \bar{X}_M\}$.
Hence $Q_{M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}}$ is an $\gamma$-$\varepsilon$-equilibrium of $(X,\Sigma_{X},\mu_{X},T_{t})$. \\
It follows from the (deterministic) Prevalence Theorem (cf.\ Werndl and Frigg 2015a) that
$\mu_{X}(Q_{M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}})\geq\mu_{X}(Q_{M})+\gamma-\varepsilon$ for all $M\neq M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}$. This immediately implies that $P\{M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}\}\geq
P\{M\}+\gamma-\varepsilon$ for all $M\neq M_{\gamma\textnormal{-}\varepsilon\textnormal{-eq}}$.
\section*{References}
\noindent Baxter, Rodney. 1982. \emph{Exactly Solved Models in Statistical Mechanics}. San Diego: Academic Press Limited.\\
\noindent Callender, Craig. 2001. Taking Thermodynamics Too Seriously. \emph{Studies in History and Philosophy of Modern Physics} 32: 539-553.\\
\noindent Cipra, Barry A. (1987). An Introduction to the Ising Model. \textit{American Mathematical Monthly} 94: 937-954.\\
\noindent Clarke, R., N. Caswell, and S. A. Solin. 1979. Melting and Staging in Graphite Intercalated with Cesium. \emph{Physical Review Letters} 42: 61-64. doi:http://dx.doi.org/\\\-10.1103/PhysRevLett.42.61.\\
\noindent de Ribaupierre, Y., and F. D. Manchester. 1974. Experimental study of the critical-point behaviour of the hydrogen in palladium system. I. Lattice gas aspects. \emph{Journal of Physics C: Solid State Physics} 7: 2126-2139. doi:10.1088/0022-3719/7/12/006.\\
\noindent Doob, Joseph L. 1953. \emph{Stochastic Processes}. New York: John Wiley \& Sons.\\
\noindent Frigg, Roman. 2008. A Field Guide to Recent Work on the Foundations of Statistical Mechanics. In \emph{The Ashgate Companion to Contemporary Philosophy of Physics}, ed. Dean Rickles, 99-196. London: Ashgate.\\
\noindent Kierlik, E., P. A. Monson, M. L. Rosinberg, and G. Tarjus. 2002. Adsorption hysteresis and capillary condensation in disordered porous solids: a density functional study. \emph{Journal of Physics: Condensed Matter} 14: 9295-9315. doi:10.1088/0953-8984/14/40/319.\\
\noindent Kikuchi, Ryoichi, and John W. Cahn. 1980. Grain-boundary melting transition in a two-dimensional lattice-gas model. \emph{Physical Review B} 21: 1893-1897.\\
\noindent Lavis, David. 2005. Boltzmann and Gibbs: An attempted reconciliation. {\it Studies in History and Philosophy of Modern Physics} 36: 245-273.\\
\noindent Lavis, David, and George M. Bell. 1999. \emph{Statistical Mechanics of Lattice Systems: Volume 1: Closed-Form and Exact Solutions}. Berlin: Springer.\\
\noindent Jicai, Pan, Subal Das Gupta, and Martin Grant. 1998. First-Order Phase Transition in Intermediate-Energy Heavy Ion Collisions. \emph{Physical Review Letter} 80: 1182-1885.\\
\noindent Reiss, Howard. 1996. \emph{Methods of Thermodynamics}. Mineaola/NY: Dover.\\
\noindent Wang, Genmiao, Edith M. Sevinck, Emil Mittag, Debra J. Searles, and Denis J.\ Evans. 2002. Experimental Demonstration of Violations of the Second Law of Thermodynamics for Small Systems and Short Time Scales. \emph{Physical Review Letters} 89: 050601.\\
\noindent Werndl, Charlotte. 2009. Are Deterministic Descriptions and Indeterministic Descriptions Observationally Equivalent?. \emph{Studies in History and Philosophy of Modern Physics} 40: 232-242.\\
\noindent Werndl, Charlotte. 2011. On the Observational Equivalence of Continuous-Time Deterministic and Indeterministic Descriptions. \emph{European Journal for the Philosophy of Science} 1: 193-225.\\
\noindent Werndl, Charlotte, and Roman Frigg. 2015a. Reconceptualising Equilibrium in Boltzmannian Statistical Mechanics and Characterising its Existence. \emph{Studies in History and Philosophy of Modern Physics} 44 : 470-479.\\
\noindent Werndl, Charlotte, and Roman Frigg. 2015b. Rethinking Boltzmannian Equilibrium. \emph{Philosophy of Science} 82: 1224-1235.\\
\noindent Yang, Chen Ning, and Tsung-Dao Lee. 1952. Statistical Theory of Equations of State and Phase Transitions. I. Theory of Condensation. \emph{Physical Review} 87: 404-409.
\end{document}
|
\section{Introduction\label{sec:intro}}
The electronic structure and quantum dynamics of a system
can be modelled using several approaches based on, e.g., wave function
methods~\cite{landau1977quantum}, Green's
functions~\cite{stefanucci2013nonequilibrium,nolting2009fundamentals,marini_yambo:_2009}, density
matrix theory~\cite{rand_lectures_2010,boyd_nonlinear_2008} or density
functional theory (DFT)~\cite{martin2004electronic,marques2012fundamentals}.
In practice, to model larger and larger electronic systems,
high-performance computing facilities along with optimized algorithms are continually
developed. To improve the scaling of the algorithm, hybrid approaches have been devised
to break down the computational complexity of composite systems which
include a small subsystem --- still amenable of
a fully quantum-mechanical treatment --- and a larger environment --- which
is dealt with a lower level of approximation, most often classical.
Examples of such composites include solvated
molecules~\cite{cossi_ab_1996,cossi_new_2002,tomasi_quantum_2005},
protein-ligand
interactions~\cite{murphy_mixed_2000,hensen_combined_2004,grater_protein/ligand_2005}
and semiconductor-metal nanoparticle
hybrids~\cite{lin_plasmonic_2013,singh_enhancement_2013,zhang_semiconductor-metal_2006,artuso_strongly_2010,artuso_hybrid_2012,paspalakis_control_2013,kosionis_optical_2013,cheng_coherent_2007,li_optical_2012,singh_enhancement_2013}. In all these
cases, we are more interested in the dynamics of the smaller subsystem and we
look at the environment as a source of unavoidable perturbations.
Such hybrid approaches rely on the possibility to
separate the composite system into two or more components whose dynamics are
solved using different levels of approximation and to treat the residual
interaction between the subsystems in an appropriate way. For example, a continuum
solvation model (such as the polarizable continuum model) may be used in the
solvated molecule problem where the molecule is treated using quantum mechanics
(QM)
and the solvent treated as a dielectric
continuum, the interaction being electrostatic in
nature~\cite{cossi_ab_1996,cossi_new_2002,tomasi_quantum_2005,corni_equation_2015}. Various
quantum mechanics/molecular mechanics (QM/MM) approaches have also been applied to
model the protein-ligand interaction. In these cases, the ligand is treated using
QM while the protein environment via MM and
the potentials associated with the protein's molecular
make-up is approximated by means of classical force fields~\cite{murphy_mixed_2000,hensen_combined_2004,grater_protein/ligand_2005}.
Hybrid methods have also been applied to model the coupling between molecules and metal
nanoparticles (MNPs) upon optical excitation. For small MNPs, the composite system
can still be treated fully quantum mechanically~\cite{zuloaga_quantum_2009}. For
larger MNPs, classical electrodynamics is employed to model the MNP dynamics
whereas a quantum description of the molecule is retained. In this case,
the interaction between the MNP and the molecule is modeled through an
effective electromagnetic coupling. These hybrid approaches make use of numerical methods such
as the finite-difference time domain (FDTD)
to solve the classical electrodynamics problem --- namely,
the Maxwell's equations --- while the dynamics of the
molecular electrons are solved by means of time-dependent
DFT. The overall dynamics are made self-consistent
by including the electromagnetic field generated by the MNP into the molecular evolution
and \emph{vice versa}~\cite{chen_classical_2010,coomar_near-field:_2011,gao_communication:_2013,sakko_dynamical_2014}.
In this work we propose an alternative, simpler and much less computationally expensive method
that avoids the solution of Maxwell's equations when the near--field effects in the electromagnetic
coupling between the MNP and the quantum system (e.g., a molecule or a quantum dot) are negligible.
To this end, we shall present a generalized model for treating the time-dependent interaction
between a quantum system (QS) and classical system (CS) coupled through an
electromagnetic field. The interaction is considered in the dipole-dipole
approximation
within the quasi-static limit. The dynamics of the QS are described via the
density matrix master equation involving an effective field which depends on the
time-dependent dipole moment of the CS. We note here that whilst we employ density
matrix theory for the quantum dynamics, the method is general and can be
applied to any time-dependent quantum mechanical approach such as those
mentioned in the opening paragraph.
The CS is modeled using classical electrodynamics in the linear response
regime where the time-dependent dipole moment is reproduced by the
introduction of a set of auxiliary degrees of freedom.
These degrees of freedom enter into a set of projected equations
of motion (PEOM) and are constrained by modeling the frequency-dependent
polarizability of the CS.
As a testbed, we consider the hybrid system consisting of a semiconducting
quantum dot (SQD) and MNP. In particular, the SQD is treated as an abstract
two-level QS while the MNP is modelled as a gold nanosphere. This system has been
studied
extensively~\cite{zhang_semiconductor-metal_2006,artuso_strongly_2010,artuso_hybrid_2012,paspalakis_control_2013,kosionis_optical_2013,cheng_coherent_2007,li_optical_2012,singh_enhancement_2013}
because it can be solved analytically by means of the rotating wave approximation
(RWA). For continuous wave excitation, we show that this analytical benchmark for the
energy absorption is correctly retrieved by the proposed hybrid approach.
Pulsed excitations are also examined and agreement between the proposed method and the
RWA approximation is shown if picosecond pulses are used. However, for a femtosecond pulse
the RWA breaks down and an approach like the proposed method must be preferred.
The paper is organised as follows: In Section~\ref{sec:methods}, we
describe the dipole-dipole interaction between the QS and CS and derive the
PEOM method for treating the time-dependent
dipole moment of the CS. The method is applied to a simple SQD-MNP system in
Section~\ref{sec:testbed1} and the results for energy absorption rates and
population inversion are compared with those from semi-analytical
approximations. Finally, the conclusions are presented in
Section~\ref{sec:conclusions}.
\section{\label{sec:methods}The PEOM Method}
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{qs-cs.pdf}
\caption{Schematic diagram showing the dipole-dipole
interaction between a QS and a CS,
separated by a distance $R$.
When an external field, $\veext$, is applied, a dipole, $\vpqs$, is
induced in the QS thus generating a field. The CS thus experiences this
dipole field in addition to the external field, and we denote the total
field felt by the CS as $\vecs$. Similarly, due to $\vecs$, a dipole field is generated
in the CS which is in turn felt by the QS in addition to $\veext$, and we
denote the total field felt by the QS as $\veqs$. In this way, the QS
and CS dynamics are coupled through the external field.}
\label{fig:two_particles}
\end{figure}
We consider a QS and a CS
separated by a distance, $R$. An external field, $\veext(t)$, is applied inducing
a dipole-dipole interaction
between the two systems (see \figref{fig:two_particles}). To simplify the
notation, we assume that the QS and CS are isotropic media, though the method
can be easily generalized to the anisotropic case. We write
$\veext(t)\equiv\eext(t)\ve$ and denote the unit vector pointing along the
line separating the centers of the particles as $\hat n$. The fields felt by
the QS and CS are then,
respectively~\cite{jackson_classical_1962,schmitt_preparation_1999,zhang_semiconductor-metal_2006},
\begin{subequations}
\noeqref{eq:eqs,eq:ecs}
\begin{align}
\veqs(t) &= \eext(t)\ve + \frac{\pcs(t)}{\epsb R^3}\g \ , \label{eq:eqs}\\
\vecs(t) &= \eext(t)\ve + \frac{\pqs(t)}{\epsb R^3}\g \ , \label{eq:ecs}
\end{align}
\label{eq:fields1}
\end{subequations}
where $\pcs(t)$ ($\pqs(t)$) is the total dipole moment of
the CS (QS). $\epsb$ is the dielectric constant of the background medium and
\begin{equation}
\g=3\hat n\left( \ve \cdot \hat n \right) - \ve \ .
\label{eq:g_def}
\end{equation}
The form of the fields here assumes only a dipole interaction. This is valid if
$R$ is sufficiently large, but can be generalised to take into account
higher multipole interactions as shown in Ref.~\onlinecite{Yan2008a}.
For demonstration, we shall presently use the density matrix approach to describe the
quantum system, although the method is applicable to any time-dependent
model. In the density matrix formalism, the dipole moment of the QS is given by
\begin{equation}
\noeqref{eq:pqs}
\vpqs(t) = \tr{\bm\rho(t)\vmu} \ ,
\label{eq:pqs}
\end{equation}
where $\tr{\cdots}$ is the matrix trace operator, $\vmu$ is the dipole moment operator matrix and $\bm\rho$ is the QS
density matrix which evolves in time due to the field, $\veqs(t)$, via
the following master equation,
\begin{equation}
\dot{\bm\rho} = -\frac{i}{\hbar} \left[\bm H_0, \bm \rho\right] +
\frac{i}{\hbar} \veqs(t)\cdot \left[\vmu,\bm\rho\right] +
\eta\left(\bm\rho\right) \ .
\label{eq:rho_eom}
\end{equation}
In Eq.~\eqref{eq:rho_eom}, $\bm H_0$ is the Hamiltonian of the unperturbed QS
and the interaction energy with the time-dependent field $\veqs(t)$ is treated within the electric
dipole approximation (second term). $\eta\left(\bm\rho\right)$ is an
additional function which can be used to model phenomenological effects not
included in the Hamiltonian such as non-radiative decay (see Ref.~\onlinecite{boyd_nonlinear_2008}, for example).
We assume that the CS has a frequency dependent polarizability $\alpha(\omega)$
which is known, e.g. by experiment or \textit{ab-initio} calculations. Its
dipole moment can then be described (in the linear response regime)
via~\footnote{Gaussian units are assumed throughout the paper.}
\begin{equation}
\vpcs(\omega) = \epsb \alpha(\omega)\vecs(\omega) \ .
\label{eq:pcs_omega}
\end{equation}
In the time domain, the dipole moment
is written in terms of the response function $\alpha(t)$,
\begin{equation}
\vpcs(t) = \epsb \int_{-\infty}^{t}\alpha(t-t')\vecs(t')dt' \ .
\label{eq:P_2}
\end{equation}
Using a coupled iterative technique, we could then solve Eq.~\eqref{eq:rho_eom}
numerically to obtain the time-dependent response of the QS to
the effective field $\veqs(t)$.
However, computing the integral in Eq.~\eqref{eq:P_2}
at each time-step in the solution is cumbersome and the values of
$\vecs(t)$ and $\alpha(t)$ for each time-step must be held in memory which may not be
feasible for long simulations.
This leads to the main component of the PEOM method, an alternative to
calculating Eq.~\eqref{eq:P_2} directly by following a time-convolutionless
scheme inspired by Ref.~\onlinecite{stella_generalized_2014}.
We introduce $N$ complex auxiliary degrees of freedom, $\{s_k(t)\}$ for
$k=1,2,\ldots,N$, which satisfy the following EOMs
\begin{equation}\label{eq:s_EOM}
\dot{\vs}_k = -\left( \gamma_k + \imagn\omega_k \right)\vs_k + \imagn\epsb
\vecs(t)\
,
\end{equation}
and assume that $\vpcs(t)$ can be written as
\begin{equation}\label{eq:P_2_approx}
\vpcs(t) = \sum_{k=1}^{N} c_k \real{\vs_k(t)} \ ,
\end{equation}
so that the memory-dependent integral in Eq.~\eqref{eq:P_2} is replaced with an
expansion over the functions $\vs_k$ found by solving the differential
equations in Eq.~\eqref{eq:s_EOM}. As these differential equations no longer
contain a time-convolution (i.e., they are ``memoryless''), they can be
efficiently integrated by using standard iterative algorithms, e.g., the
Runge-Kutta fourth-order method. All that
is required is to find suitable values for the (real) parameters $\{c_k,
\gamma_k, \omega_k\}$ in Eq.~\eqref{eq:s_EOM}.
The formal solution of Eq.~\eqref{eq:s_EOM} is
\begin{equation}
\vs_k(t) = \epsb \int_{-\infty}^{t}\imagn\ e^{-\left( \gamma_k + \imagn\omega_k
\right)\left( t-t' \right)} \vecs (t') dt' \ .
\label{eq:s_k}
\end{equation}
Substituting the real part of Eq.~\eqref{eq:s_k} into Eq.~\eqref{eq:P_2_approx} and
rearranging yields
\begin{align}
\vpcs(t) = \epsb
&\int_{-\infty}^{t}
\left(
\sum_{k=1}^{N} c_k e^{-\gamma_k (t-t')}\sin[\omega_k(t-t')]
\right) \notag \\
& \times
\left( \vecs(t') \right) dt' \ ,
\end{align}
and comparing with Eq.~\eqref{eq:P_2} we see that
\begin{equation}
\alpha(t) =
\sum_{k=1}^{N} c_k e^{-\gamma_k t}\sin(\omega_k t) \ .
\label{eq:chi_t}
\end{equation}
Then taking the Fourier transform of
Eq.~\eqref{eq:chi_t} (using the causality condition) gives
\begin{equation}
\alpha(\omega) = \sum_{k=1}^{N}\frac{c_k}{2}
\left[ \frac{1}{\omega+\omega_k +\imagn\gamma_k}
-\frac{1}{ \omega-\omega_k +\imagn\gamma_k}\right] \ .
\label{eq:chi_approx}
\end{equation}
Hence, the parameters $\{c_k,\gamma_k,\omega_k\}$ may be found by fitting
the frequency-dependent polarizability of the CS (which is known)
to the fitting functions on the RHS of Eq.~\eqref{eq:chi_approx}
(e.g. using the least squares method as done in this work).
\section{The Quantum Dot-Metal Nanoparticle System\label{sec:testbed1}}
To test the PEOM method proposed in
\secref{sec:methods}, we use a semiconducting
quantum dot (SQD) as the QS and a metal nanoparticle (MNP) as the
CS since this hybrid system has been widely
studied~\cite{zhang_semiconductor-metal_2006,artuso_strongly_2010,artuso_hybrid_2012,paspalakis_control_2013,kosionis_optical_2013,cheng_coherent_2007,li_optical_2012,singh_enhancement_2013}
and some properties can be obtained analytically.
In particular, we look at the energy absorption rate (EAR) and
population inversion, which are associated with continuous and pulsed wave
excitation respectively.
From Eq.~\eqref{eq:fields1}, when an external field $\eext$ is applied, the
fields felt by the SQD and MNP respectively are
\begin{align}
\esqd &= \eext + g\frac{\pmnp}{\epsb R^3} \ , \label{eq:esqd}\\
\emnp &= \eext + g\frac{\psqd}{\epsb R^3} \ .
\end{align}
We have taken the external field to be polarized along the line connecting the
centers of the particles, allowing us to drop the vector notation and set $g=2$
(see Eq.~\eqref{eq:g_def}).
The SQD is
treated as a 2-level atomic system giving rise to a $2\times 2$ density matrix
with elements that can be written
as~\cite{boyd_nonlinear_2008,zhang_semiconductor-metal_2006}
\begin{equation}
\begin{cases}
\dot{\Delta} &= -\frac{4\mut}{\hbar}\esqd(t)\imag{\rho_{21}} -
\Gamma_{11}(\Delta - 1) \\
\dot\rho_{21} &= -\left( \Gamma_{21}+\imagn\omega_0
\right)\rho_{21} + \imagn\frac{\mut}{\hbar}\esqd(t) \Delta
\label{eq:eoms}
\end{cases} \ ,
\end{equation}
where $\Delta(t)=\rho_{11}(t)-\rho_{22}(t)$ is the population difference between the
ground and excited states with frequency difference $\omega_0$ which is known
as the exciton frequency. $\Gamma_{11}$ and $\Gamma_{21}$ are the population
decay and dephasing rates of the system respectively. The SQD is assumed to be
a
dielectric sphere with dielectric constant $\epss$ and so it has a screened dipole matrix element
$\mut=\mu_{21}/\epseff$ where $\mu_{21}$ is the bare dipole matrix element
and $\epseff=\frac{2\epsb+\epss}{3\epsb}$.~\cite{batygin_problems_1978}
For comparison with previous literature, the MNP is taken to be a gold sphere
of radius $a$ and its polarizability is approximated by the Clausius-Mossotti
formula,
\begin{equation}
\alphamnp(\omega) = a^3 \frac{\epsm(\omega)-\epsb}{
\epsm(\omega) + 2\epsb} \ ,
\label{eq:chi_sphere}
\end{equation}
where $\epsm(\omega)$ is the frequency-dependent dielectric function of the
bulk metal~\cite{landau_electrodynamics_1984} (we use the analytical model for
bulk gold as given by Etchegoin et al.~\cite{etchegoin_analytic_2006}). Note
that there are no fundamental reasons for using an analytical expression for
the polarizability. For example, $\alphamnp(\omega)$ may instead be extracted from
experimental data or computed using a first-principles approach.
We take the SQD system parameters from
Ref.~\onlinecite{zhang_semiconductor-metal_2006}.
The dielectric
constant is taken to be $\epss=6$ with transition dipole moment $\mu=0.65e$~nm and
exciton energy $\hbar\omega_0=2.5$~eV close to the plasmon peak of the gold MNP.
The decay and dephasing times are given by
$\Gamma_{11}^{-1}=0.8$~ns and $\Gamma_{21}^{-1}=0.3$~ns. We assume the
background medium is a vacuum so that $\epsb=1$.
\subsection{Energy Absorption Rate}
We first look at the EAR of the hybrid system which is a steady-state property,
found by considering the response to the following field,
\begin{equation}
\eext(t) = E_0 \cos(\omega_L t) .
\label{eq:plane_wave}
\end{equation}
In this case, Eq.~\eqref{eq:eoms} can be solved analytically within the RWA as
shown in, e.g.,
Refs~\onlinecite{zhang_semiconductor-metal_2006,artuso_strongly_2010,artuso_hybrid_2012}.
In the RWA, the off-diagonal density matrix elements are first separated into slowly and quickly
oscillating components,
\begin{subequations}
\noeqref{eq:rho21,eq:rho12}
\begin{align}
\rho_{21}(t) &= \bar\rho_{21}(t)e^{-\imagn\omega_L t} \ , \label{eq:rho21} \\
\rho_{12}(t) &= \bar\rho_{12}(t)e^{\imagn\omega_L t} \ , \label{eq:rho12}
\end{align}
\label{eq:rho_bar}
\end{subequations}
where $\bar\rho_{21}(t)$ and $\bar\rho_{12}(t)$ are assumed to vary on a much larger
timescale than $2\pi/\omega_L$. The RWA assumes
that $\omega_L\approx\omega_0$, neglecting terms oscillating at frequencies far
from $\omega_0$, so that the following modified EOMS can be obtained
from Eq.~\eqref{eq:eoms},
\begin{equation}
\begin{cases}
\dot\Delta &= 4\imag{\left( \frac{\omegaeff}{2}+G\bar\rho_{21} \right)\bar\rho_{12}}
+ \Gamma_{11}\left( 1-\Delta \right) \\
\dot{\bar\rho}_{21} &= \left[ \imagn\left( \omega_L-\omega_0 +G\Delta
\right)-\Gamma_{21} \right]\bar\rho_{21} + \imagn\frac{\omegaeff}{2}\Delta
\label{eq:2ls_eoms}
\end{cases}
\ ,
\end{equation}
where
\begin{subequations}
\noeqref{eq:omegaeff1,eq:G}
\begin{align}
\omegaeff &= \Omega_0 \left[ 1 +
\frac{g}{R^3}\alphamnp(\omega_L) \right] \ ,\label{eq:omegaeff1}\\
G &= \frac{g^2\mut^2 }{\hbar\epsb R^6}\alphamnp(\omega_L) \ , \label{eq:G}
\end{align}
\label{eq:omegaeff}
\end{subequations}
with $\Omega_0 = \mut E_0/\hbar$ being the Rabi frequency of the
isolated SQD.
The EAR of the
SQD is defined as~\cite{artuso_strongly_2010}
\begin{equation}
\qsqd = \frac{1}{2}\hbar\omega_0\Gamma_{11}\left( 1-\Deltass \right)\ ,
\label{eq:qsqd}
\end{equation}
where $\Deltass$ is the value of $\Delta(t)$ when a steady-state has been
reached, while the EAR of the MNP is~\cite{artuso_strongly_2010}
\begin{equation}
\qmnp = \left\langle\int j\cdot \emnpin\ dV \right\rangle \ ,
\label{eq:qmnp}
\end{equation}
where $j = \frac{\partial}{\partial t} (\pmnp(t)/V)$ is the current density in
the MNP which has volume $V$ and $\emnpin$ is the field inside the MNP. Within
the RWA, it can be shown that $\qmnp$ depends on $\rhoss$, the steady state
value of $\bar\rho_{21}(t)$~\cite{artuso_strongly_2010}. The total EAR of the system is then
$Q=\qmnp+\qsqd$. An analytical solution for $\Deltass$ and
$\rhoss$ can be obtained by setting the L.H.S. of
Eq.~\eqref{eq:2ls_eoms} equal to zero (see
Ref.~\onlinecite{zhang_semiconductor-metal_2006} for example).
As an alternative to the above RWA solution, we numerically solve the original
EOMs in Eq.~\eqref{eq:eoms} using the PEOM method where $\pmnp(t)$, which appears
in the expression for $\esqd(t)$, is approximated using
Eq.~\eqref{eq:s_EOM} and Eq.~\eqref{eq:P_2_approx}. The fitting parameters,
$\{c_k,\gamma_k,\omega_k\}$, are obtained from a least-squares fit of
$\alphamnp(\omega)$ to the model in Eq.~\eqref{eq:chi_approx} over the range 0-20~eV
which required $N=12$ fitting functions for sufficient accuracy.
\figref{fig:Q_low_compare} shows the total EAR for the hybrid system
as a function of the laser detuning (field intensity $I_0$=1~\wcm) for various separation distances of the two
particles. We can see the expected quenching of $Q$ and the red-shift of the
hybrid exciton energy as the particles are brought together as described in
Ref.~\onlinecite{zhang_semiconductor-metal_2006}. The analytical RWA
solutions are shown in
solid lines while the crosses are the results taken from the PEOM
method. In this case, we see perfect agreement between the two methods due
to the validity of the RWA for the case of a sinusoidal external field with
frequency very close to resonance with the SQD exciton frequency.
We now turn our attention to short-pulse excitation to
demonstrate a case where the RWA cannot be used.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{Q_low_compare.pdf}
\caption{Energy absorption rate, $Q$, for a SQD-MNP system with
separation distances $R=13$~nm (green), 15~nm (red), 20~nm (blue) and 80~nm
(black).
The solid
lines represent the steady-state analytical solution in the RWA while the crosses are
the results from the PEOM method.}
\label{fig:Q_low_compare}
\end{figure}
\subsection{Population Inversion and Breakdown of the RWA}
Population inversion occurs when the SQD is
excited from the ground state to the excited state so that $\rho_{11}=0$,
$\rho_{22}=1$ and is associated with short laser pulses (see, for example,
Ref.~\onlinecite{stievater_rabi_2001}).
We consider a
pulsed external field given by
\begin{equation}
\eext(t) = E_0 f(t) \cos(\omega_L t) \ ,
\end{equation}
where $f(t)$ is a dimensionless pulse envelope. The pulse area for an isolated
SQD is defined as
\begin{equation}
\theta = \Omega_0\int_{-\infty}^{\infty} f(t) dt \ ,
\label{eq:pulse_area}
\end{equation}
and it is known that population inversion occurs at the end
of the pulse for $\theta=(2n+1)\pi$
($n=0,1,2,\ldots$)~\cite{wang_decoherence_2005}.
We shall use a hyperbolic secant envelope defined by
\begin{equation}
f(t) = \sech\left( \frac{t-\tau_0}{\tau_p} \right) \ ,
\label{eq:sech}
\end{equation}
where $\tau_0$ is the center of the pulse and $\tau_p$ characterizes the pulse
width. We choose the central frequency, $\omega_L$, to be resonant with the
exciton frequency, i.e.
$\hbar\omega_L=\hbar\omega_0=2.5$~eV, and we describe the pulse shape in terms of the number of
cycles, $n$, by defining the pulse duration as
\begin{equation}
T = \frac{4\pi}{\omega_L}n \ ,
\label{eq:pulse_length}
\end{equation}
and choosing
\begin{equation}
\tau_0 = T/2 \ , \quad \tau_p = T/30 \ .
\label{eq:tau_0_tau_p}
\end{equation}
In this way we ensure the maximum amplitude, $E_0$, is achieved at the center
of the pulse and that the external field is sufficiently close to zero at $t=0$
and $t=T$ for the values of $E_0$ considered here.
For the sech pulse in Eq.~\eqref{eq:sech}, it can easily be shown from
Eq.~\eqref{eq:pulse_area} that for an isolated SQD, $\theta=\pi\Omega_0\tau_p$
and then a pulse of given duration can be described in terms of the pulse area by
choosing the following field amplitude,
\begin{equation}
E_0 = \frac{\hbar\theta}{\pi\mut\tau_p} \ .
\label{eq:theta_sech}
\end{equation}
In Ref.~\onlinecite{paspalakis_control_2013} it was shown that the pulse area
for an SQD when coupled to the MNP may be written approximately as $\theta =
\pi|\omegaeff|\tau_p$ so that Eq.~\eqref{eq:theta_sech} becomes
\begin{equation}
E_0 = \hbar\theta \left( \pi\mut\tau_p \left|1 +
\frac{g}{R^3}\alphamnp(\omega_L)\right| \right)^{-1} \ .
\label{eq:theta_sech_R}
\end{equation}
In particular, it was stated that for short pulses ($\tau_p\sim0.1$~ps)
with amplitude given by Eq.~\eqref{eq:theta_sech_R}, the resulting dynamics
should be independent of $R$ as the influence of the parameter $G$
(see Eq.~\eqref{eq:G}) becomes weaker.
In previous studies relating to pulsed excitations in SQD-MNP systems, the
time-scales have generally been limited to relatively long pulses. For example in
Ref.~\onlinecite{sadeghi_coherent_2010,sadeghi_inhibition_2009}, the external
field is switched on over tens of nanoseconds while in
Ref.~\onlinecite{paspalakis_control_2013,anton_plasmonic_2012}, picosecond
pulses are used. In such cases, the population dynamics can be found by solving
the RWA EOMs in Eq.~\eqref{eq:2ls_eoms} but replacing $\omegaeff$ with the
time-dependent form,
\begin{equation}
\Omega(t) = f(t) \omegaeff \ ,
\label{eq:omegaeff_t}
\end{equation}
giving
\begin{equation}
\begin{cases}
\dot\Delta &= 4\imag{\left( \frac{\Omega(t)}{2}+G\bar\rho_{21} \right)\bar\rho_{12}}
+ \Gamma_{11}\left( 1-\Delta \right) \\
\dot{\bar\rho}_{21} &= \left[ \imagn\left( \omega_L-\omega_0 +G\Delta
\right)-\Gamma_{21} \right]\bar\rho_{21} + \imagn
\frac{\Omega(t)}{2}\Delta
\label{eq:2ls_eoms_pulse}
\end{cases}
\ .
\end{equation}
The use of the RWA and slowly-varying envelope approximations respectively
imply that solutions to~\eqref{eq:2ls_eoms_pulse} are only valid if
${\bar\rho}_{21}(t)$ and $f(t)$ vary much more slowly than $2\pi/\omega_L$.
Recalling that $\hbar\omega_L=2.5$~eV, we therefore require the pulse duration
to be much greater than $\sim2$~fs.
Indeed, it is known that the RWA is not reliable for
ultrashort (femto- and subfemto-second)
pulses.~\cite{meier_coherent_2007,PhysRevA.52.3082,yang_ultrafast_2015} This
is demonstrated in
Fig.~\ref{fig:rwa_compare} where
we compare the solution of the original EOMs in
Eq.~\eqref{eq:eoms} with those of the modified RWA EOMs in
Eq.~\eqref{eq:2ls_eoms_pulse}, showing the excited state population dynamics
for an isolated SQD ($R\rightarrow\infty$) interacting with a picosecond
and femtosecond pulse of area
$5\pi$ (according to Eq.~\eqref{eq:theta_sech}).
Fig.~\ref{fig:rwa_compare}~(a) shows $\rho_{22}(t)$ for a 1000-cycle pulse
($\tau_p\approx 0.11$~ps) and we can see that the RWA in this case provides an
adequate description of the dynamics, with population inversion occurring at
the end of the pulse as expected for a $5\pi$ pulse. The inset shows a
magnified region in
which we can see the effect of the RWA neglecting the quickly oscillating
terms: however, in the picosecond time-scale, these effects have negligible
influence on the overall dynamics.
Fig.~\ref{fig:rwa_compare}~(b) shows $\rho_{22}(t)$ for a 10-cycle pulse
($\tau_p\approx 1.1$~fs) where the pulse duration is of a comparable time-scale
to $2\pi/\omega_L$. In this case, we can see that the quickly oscillating terms
neglected in the RWA solution have a more significant effect on the overall
dynamics: importantly, complete population inversion is not achieved at the end
of the pulse, and there is a much more oscillatory behaviour.
\begin{figure}[ht]
\centering
\begin{minipage}{0.9\columnwidth}
\centering
\includegraphics[width=1.0\columnwidth]{rwa_compare_ps.pdf}
\vspace{0.2cm}
\end{minipage}
\begin{minipage}{0.9\columnwidth}
\centering
\includegraphics[width=1.0\columnwidth]{rwa_compare_fs.pdf}
\end{minipage}
\caption{Excited state population dynamics for an isolated SQD
($R\rightarrow\infty$) interacting with a sech pulse of area $5\pi$
according to Eq.~\protect\eqref{eq:theta_sech}. The solid red line is the
numerical solution to the original EOMs in Eq.~\protect\eqref{eq:eoms}
while the dashed blue line is the solution to the modified RWA EOMs
in Eq.~\protect\eqref{eq:2ls_eoms_pulse}. (a) Dynamics for a 1000-cycle
pulse corresponding to $\tau_p\approx 0.11$~ps. (b) Dynamics for a
10-cycle pulse corresponding to $\tau_p\approx 1.1$~fs.}
\label{fig:rwa_compare}
\end{figure}
In Ref.~\onlinecite{yang_ultrafast_2015}, a numerical solution to
Eq.~\eqref{eq:eoms} for pulsed excitation in SQD-MNP systems beyond the RWA is proposed. In
deriving Eq.~\eqref{eq:2ls_eoms_pulse}, $\esqd(t)$ is expressed by separating
out the positive and negative frequency parts as
\begin{equation}
\esqd(t) \approx \frac{\hbar}{\mut}\left[ \left( \frac{\Omega(t)}{2}e^{-\imagn\omega_L t} +
G\rho_{21}(t) \right) + \text{c.c.} \right] .
\label{eq:esqd_t}
\end{equation}
Instead of invoking the usual RWA to arrive at
Eq.~\eqref{eq:2ls_eoms_pulse}, Yang et al. numerically solve
Eq.~\eqref{eq:eoms} using as $\esqd(t)$ the field in Eq.~\eqref{eq:esqd_t} (we
shall call this method the effective field method). However, in
deriving Eq.~\eqref{eq:esqd_t}, one must first separate out the slowly oscillating
components of the off-diagonal density matrix elements as in
Eq.~\eqref{eq:rho_bar} (see e.g. Ref.~\cite{artuso_strongly_2010}) and the
slowly-varying envelope approximation must also be used. Thus, while the
quickly oscillating terms are included, improving over the RWA, the pulse
duration must still be longer than $2\pi/\omega_L$. We shall presently
demonstrate how these assumptions render this approach unreliable for few-cycle
pulses when the interparticle distances are small.
\onecolumngrid
\begin{figure}[ht]
\begin{minipage}{\columnwidth}
\includegraphics[width=0.45\columnwidth]{rwa.pdf}
\hspace{0.25cm}
\includegraphics[width=0.45\columnwidth]{esqdt.pdf}
\vspace{0.25cm}
\end{minipage}
\begin{minipage}{\columnwidth}
\includegraphics[width=0.45\columnwidth]{peom.pdf}
\hspace{0.25cm}
\includegraphics[width=0.45\columnwidth]{peom_N1.pdf}
\end{minipage}
\caption{Excited state population dynamics for an SQD-MNP system interacting
with a 10-cycle sech pulse for various interparticle distances, $R$. The
field amplitude is chosen to give a $5\pi$ pulse area for each value
of $R$ according to Eq.~\protect\eqref{eq:theta_sech_R}.
(a) Solution to the modified EOMS in Eq.~\protect\eqref{eq:2ls_eoms_pulse} under the
RWA and assuming a slowly-varying pulse envelope.
(b) Solution to the original EOMs in Eq.~\protect\eqref{eq:eoms} beyond the RWA using
the effective field in Eq.~\protect\eqref{eq:esqd_t} which assumes a slowly-varying
pulse envelope.
(c) Solution to Eq.~\protect\eqref{eq:eoms} using the PEOM method where the auxiliary
parameters are obtained by fitting $\alphamnp(\omega)$ in
Eq.~\protect\eqref{eq:chi_sphere} accurately over the range 0--10~eV using $N=21$
fitting functions.
(d) Same as (c) but where $\alphamnp(\omega)$ is fitted over a small range
close to $\hbar\omega_0$ (2.455--2.545~eV) using $N=1$ fitting functions.
Insets: real (blue dashed) and imaginary (black dashed) parts of the fitted
polarizability, $\alphamnp(\omega)/a^3$ (exact function shown in
grey).}
\label{fig:rwa_peom_compare}
\end{figure}
\twocolumngrid
In Fig.~\ref{fig:rwa_peom_compare}, we compare the solutions for $\rho_{22}(t)$
based on the RWA, effective field method and the PEOM method. In each
case, the SQD-MNP system interacts with a 10-cycle sech pulse of area $5\pi$
with $R$-dependent amplitude given by Eq.~\eqref{eq:theta_sech_R} and the
excited state population dynamics are shown for various interparticle
distances.
In Fig.~\ref{fig:rwa_peom_compare}~(a), the modified
RWA EOMS in Eq.~\eqref{eq:2ls_eoms} are solved and we can see that complete population
inversion occurs at the end of the pulse and the dynamics are identical for
each $R$ as expected from Ref.~\onlinecite{paspalakis_control_2013} and
Eq.~\eqref{eq:theta_sech_R}.
In
Fig.~\ref{fig:rwa_peom_compare}~(b), the original EOMS in Eq.~\eqref{eq:eoms}
are solved beyond the RWA by taking $\esqd(t)$ of the form in
Eq.~\eqref{eq:esqd_t} similar to the calculations performed in
Ref.~\onlinecite{yang_ultrafast_2015}. In this case we see that the dynamics
are almost identical to the isolated SQD as shown in
Fig.~\ref{fig:rwa_compare}~(b) where the original EOMs are solved exactly.
At difference with the RWA solution in Fig.~\ref{fig:rwa_peom_compare}~(a),
complete population inversion does not occur as a consequence of the
RWA-breakdown.
On the other hand, the dynamics remain independent of $R$ as predicted by
Ref.~\onlinecite{paspalakis_control_2013}. We note at this point that
Ref.~\onlinecite{paspalakis_control_2013,yang_ultrafast_2015} employ a multipole description for the MNP response while our calculations use the simpler
dipole model. However, we have compared results using the same multipole
approximation and noticed no difference due to the short time-scales involved
here.
In Fig.~\ref{fig:rwa_peom_compare}~(c), we solve the original EOMs in
Eq.~\eqref{eq:eoms} using the PEOM method. We obtain the auxiliary parameters
describing the MNP dipole moment by fitting $\alphamnp(\omega)$ in
Eq.~\eqref{eq:chi_sphere} to the functions in Eq.~\eqref{eq:chi_approx}. This
is achieved by a least-squares fit over the range 0--10~eV using $N=21$ fitting
functions to gain a fit of sufficient accuracy (see inset).
We see that for large $R$ ($R=80$~nm),
$\rho_{22}(t)$ resembles the results in (b).
However, as the interparticle distance decreases, the dynamics change
considerably, with larger effect towards the end of the pulse. For each $R$, the dynamics are
similar up to around 18~fs by which point the pulse is almost over (see
Fig.~\ref{fig:pmnpt}~(b)). After this point, the population for each $R$
reaches the same maximum value (around 0.95), but at different times: $\sim23$~fs for $R=80$~nm,
$\sim21$~fs for $R=20$~nm and $\sim20$~fs for $R=13$~nm. The population then
decreases more steeply as $R$ decreases, reaching as low as 0.6 for $R=13$~nm.
We ascribe the different results obtained with the PEOM in
Fig.~\ref{fig:rwa_peom_compare}~(c) and the effective field method in
Fig.~\ref{fig:rwa_peom_compare}~(b) to the fact that femtosecond pulses ($\sim
10$ cycles) excite
a broad range of frequencies: in particular, much broader than the
sub-picosecond pulses ($\sim 100$ cycles) for which the effective field
method~\cite{yang_ultrafast_2015} was originally developed.
As stated earlier, in writing
Eq.~\eqref{eq:esqd_t}, $f(t)$ must be slowly varying and the off-diagonal
density matrix elements must also first be separated into slowly and quickly
oscillating components. These approximations indirectly force the MNP to
respond only as $\alphamnp(\omega_L)$ (i.e. only at the driving frequency), as
apparent in the definitions of $\omegaeff$ and $G$ in Eq.~\eqref{eq:omegaeff}.
However, the femtosecond pulse has a large bandwidth ($>1$~eV), thus exciting a
broad range of frequencies in the MNP response. Moreover, $\alphamnp(\omega)$
changes significantly over this range close to $\hbar\omega_0=2.5$~eV due to the
formation of the plasmon peak and thus one would expect the resulting
time-dependent dipole moment, $\pmnp(t)$, (and therefore $\esqd(t)$) to be
modified compared with that for long pulses of smaller bandwidths. In
Fig.~\ref{fig:pmnpt}, $\pmnp(t)$ is shown for the $R=13$~nm cases in
Fig.~\ref{fig:rwa_peom_compare} (b) and (c). We can see that in the effective
field method, the MNP responds in phase with the external field,
while in the PEOM method the dipole moment continues to propagate well after
the pulse is over, thus contributing to the decline in population of the SQD.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{pmnpt.pdf}
\caption{
Time evolution of the MNP dipole moment, $\pmnp(t)$, for the
$R=13$~nm cases in Fig.~\ref{fig:rwa_peom_compare}~(b) (solid blue) and (c)
(solid black) where the effective field method and PEOM method are used
respectively. The corresponding external field is shown in dashed red.}
\label{fig:pmnpt}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{change_length.pdf}
\caption{
Value of $\rho_{22}(t)$ at the end of the pulse
for various pulse durations. The interparticle distance is $R=13$~nm and the
pulse area is chosen to be $5\pi$ according to Eq.~\protect\eqref{eq:theta_sech_R}.}
\label{fig:change_length}
\end{figure}
We have stated that in the effective field method, the MNP responds only at the
driving frequency, $\omega_L$ (i.e. the polarizability is effectively constant,
$\alphamnp(\omega)\approx\alphamnp(\omega_L)$). This approximation is valid for
monochromatic waves (e.g. Eq.~\eqref{eq:plane_wave}) and for long pulses. On
the other hand, for very short pulses the frequency-dependence of the
polarizability is important due to the larger bandwidth. The PEOM method
overcomes this limitation as $\alphamnp(\omega)$ can be fitted over an
arbitrary frequency range (as in Fig.~\ref{fig:rwa_peom_compare}~(c)) so that
the relevant frequencies can be included in the dynamics. The constant
polarizability of the effective field method can be simulated within the PEOM
method by choosing a single, broad function ($N=1$) which agrees with
$\alphamnp(\omega)$ at $\omega_L$ and is approximately constant over the pulse
bandwidth region (see inset of Fig.~\ref{fig:rwa_peom_compare}~(d)).
Fig.~\ref{fig:rwa_peom_compare}~(d) then shows that the effective field results
from Fig.~\ref{fig:rwa_peom_compare}~(c) are indeed recovered.
Fig.~\ref{fig:change_length} summarizes the results and shows the range of
pulse durations for which the different approximations are valid. The three
different methods agree for $n>100$ cycles ($\tau_p>10$~fs). The effective field method
correctly describes the fall in final population as the pulse duration approaches
10 cycles due to the breakdown of the RWA, but we can see that when the full
response of the MNP is taken into consideration in the PEOM method, the effect
is much more enhanced.
Overall, the results of this section show that when examing the response of
ultrashort pulses (fewer than $\sim 60$ cycles), one cannot rely on the RWA or
on the assumption that the MNP polarizability ($\alphamnp$) responds only at
the driving frequency, $\omega_L$. Therefore, one should consider more advanced
approaches. The PEOM methods is a valid alternative as it is not bound by such
approximations yet still its computational cost and complexity is similar to
that of, e.g., the RWA or effective field method.
\section{Conclusions\label{sec:conclusions}}
We have described a transferable hybrid approach to the electron dynamics of
a quantum system dipolarly coupled to a larger environment that can be treated
classically. This hybrid approach is based on a robust
projected equations of motion (PEOM) formalism.
The capabilities of the proposed hybrid approach have been demonstrated for the widely studied
case of a semiconductor quantum dot (SQD) coupled to a metallic nanoparticle (MNP).
The SQD has been modelled as a two-level system, while a semi-empirical model of the
MNP susceptibility has been used.
We have validated this hybrid approach against both analytical and
semi-analytical benchmarks of the SQD-MNP response to
picosecond laser pulses, i.e., longer than $2\pi/\omega_0$, where
$\hbar \omega_0$ is the SQD energy gap.
This is the regime of validity of the rotating wave approximation (RWA).
However, the validity of the PEOM does not rely on either the RWA
or improvements on it (e.g., the effective field method\cite{yang_ultrafast_2015})
and we have also modeled the response to femtosecond laser pulses.
In this regime, we have shown that the response of the SQD-MNP
is strongly affected by the details of the MNP susceptibility.
By artificially ``blurring'' the details of the MNP susceptibility,
the results of the hybrid approach
match the prediction of the effective field method.
To this extent, the proposed hybrid approach is inherently more accurate than
the other methods which rely on the RWA and improvements on it.
Beyond the validation for a two-level system, the
PEOM formalism can be used for systems
with an arbitrary number of levels
and is independent from the theoretical framework used to model the quantum system, e.g., the SQD.
In this work, we have used a density matrix approach, but the PEOM can be easily formulated
within a time-dependent density-functional theory framework,
or the recently devised real-time approach to the Bethe-Salpeter
equation~\cite{attaccalite_real-time_2011} formalism.
The proposed hybrid approach shares similarities with other hybrid methods
\cite{coomar_near-field:_2011,gao_communication:_2013,sakko_dynamical_2014}
and, in principle, can be also coupled to a finite-difference time-domain (FDTD)
description of the electromagnetic field.
On the other hand, an accurate FDTD model is less crucial if the SQD and MNP are
sufficiently far apart. Moreover, the simpler dipolar coupling used in this work
is still popular~\cite{Li:16,terzis_2016,yang_ultrafast_2015} and
the PEOM formalism provides a necessary improvement as attention turns towards ultrafast
phenomena.
When both the near-field response and the electromagnetic
scattering can be safely neglected,
the proposed hybrid method provides a
computationally less expensive alternative to those more accurate approaches
which include an FDTD model of the electromagnetic field.
This hybrid approach is also easier to integrate
into existing electronic structure codes, including
codes which employ periodic-boundary conditions. This is
particularly relevant for modelling extended
quantum systems (e.g., two-dimensional
semiconductors) coupled to MNPs.~\cite{eda_two-dimensional_2013}
\begin{acknowledgments}
RM acknowledges financial support from the UK Engineering and Physical
Sciences Research Council.
\end{acknowledgments}
|
\section{Introduction}
Much progress in order to understand the AdS/CFT correspondence has been achieved since it was formulated \cite{Maldacena:1997re, Aharony:1999ti}.
An important tool is the study of small perurbations.
In the CFT side they correspond to deformations of the SYM action.
Small perturbations in the supergravity(SUGRA) side correspond to small fluctuations of the classical SUGRA fields around their "vacuum" values in
the maximally symmetric background solution, $AdS_5\times S^5$. There was extensive study of small fluctuations \cite{Kim:1985ez}. However, even at the linearized level, the research was mainly focused
on unitary representations \cite{Gunaydin:1984fk}. We feel that non-unitary representations are also important and have not been sufficiently studied in this context.
In particular, there are finite-dimensional representations
\cite{Mikhailov:2011af}. To the best of our knowledge, the classification of finite-dimensional representations is missing.
The beta-deformation is probably the simplest example of a finite-dimensional representation, it
is very well studied on the SUGRA side \cite{Lunin:2005jy}, \cite{Frolov:2005dj}, \cite{Grassi:2006tj}, \cite{Bedoya:2010qz}.
On the field theory side, it is a particular case
of the deformation studied by Leigh and Strassler \cite{Leigh:1995ep}.
It preserves $\mathcal{N}=1$ SUSY and depends on two complex parameters.
However, the full supermultiplet has never been studied on the field theory side.
And we fill this gap with the present paper.
Moreover, we develop a new method, which is hopefully useful also
for other non-unitary representations.
Namely, we consider the deformation of the SYM action, the one considered in \cite{Leigh:1995ep}. It is
evaluated on the tensor product of singleton representations. We will use the oscillator representation of the singleton representation
of $psl(4|4, \bf{R})$
\cite{Bars:1982ep}; for a review of the superconformal algebra see \cite{Beisert:2010kp}.
The $\mathcal{N}=4$ SYM admits a formulation in twistor space \cite{Witten:2003nn}, see also \cite{Boels:2006ir}.
A vector in the singleton representation can be understood as a wave function of the free
field on ${\bf{R}}^{2,2}$ supported on the $\alpha$-plane.
This method allows us to reduce the problem to finding the
fininite-dimensional invariant subspace in the tensor
product of three singleton representations. Basically, in order to classify deformations of the SYM action, it is enough to
evaluate them on tensor products of the spaces of solutions of free equations, which are dual to singletons.
\vspace{.3cm}
{\bf{Plan of the paper}}
\;We make a review of $\mathcal{N}=4$ SYM theory and $psl(4|4,\bf{R})$ superalgebra. Namely, we focus on the oscillator representation of
$psl(4|4,\bf{R})$, which we call
oscillator picture. We also review the relation between the oscillator picture and supertwistors. We discuss how to write single trace operators on both pictures in Section ~\ref{II}. Then in Section
~\ref{III} we present the structure of the supermultiplet of $\beta-$deformation on the field theory side. Using the oscillator
picture we show the conformal invariance of $\mathcal{O}[B]$ in Appendix ~\ref{a}.
\section{Marginal Deformations in $\mathcal{N}=4$ SYM theory and $psl(4|4,{\bf{R}})$}
\label{II}
${\mathcal{N}=4}$ SYM is a field theory with the maximal amount of supersymmetry in four dimensions and is unique \cite{Grimm:1977xp}, for an extended
review see Chapter 3 of \cite{D'Hoker:2002aw}. The action of this field theory is
%
%
\begin{eqnarray}
\label{N=4SYM}
S_4
&=&-\int d^4x\big(\frac{1}{4}F^{(a)}_{\mu\nu}F^{\mu\nu(a)}+\frac{1}{4}D_{\mu}\phi^{AB(a)}D^{\mu}\phi_{AB}^{(a)}+i\bar\psi^{(a)}_A\bar\sigma^{\mu}D_{\mu}\psi^{(a)A} +\frac{\sqrt{2}}{2}g_{YM}f_{abc}\psi^{\alpha A(a)}\phi_{AB}^{(b)}\psi_{\alpha}^{B(c)} \nonumber \\
&+&\frac{g^2_{YM}}{16}f_{abc}f_{ade}\phi^{AB(b)}\phi^{CD(c)}\phi_{AB}^{(d)}\phi_{CD}^{(e)} -\frac{\sqrt{2}}{2}g_{YM}f_{abc}\bar\psi^{(a)}_{\dot\beta A}\phi^{AB(b)}\bar\psi^{\dot\beta (c)}_{B} \big),
\end{eqnarray}
where capital letters run from $1$ to $4$.
$\mathcal{N}=4$ SYM admits a formulation in terms of $\mathcal{N}=1$ language, one of the four fermions is combined with the gauge field
to form the vector superfield. The remaining fermions are combined with three complex scalar, to form three chiral superfield.
The superpotential $\mathcal{W}$ is given by
\begin{equation}
\label{W}
\mathcal{W}=\frac{g_{YM}\sqrt{2}}{3!}\epsilon_{ijk}f_{abc}\Phi^i_a\Phi^j_b\Phi^k_c,
\end{equation}
where $i,j,k$ indices running from $1$ to $3$, $\Phi^i$ denote the three chiral superfields and $a,b,c$ denote gauge group SU(N) indices.
In this work we are interested in study deformations of this field theory. To begin with, let us start with the deformation studied by Leigh and Strassler
~\cite{Leigh:1995ep}. The superpotential ~\eqref{W} is deformed to
\begin{eqnarray}
\mathcal{W}\mapsto \mathcal{W}+\frac{1}{6}h_{ijk}d_{abc}\Phi^i_a\Phi^j_b\Phi^k_c,
\end{eqnarray}
where $d_{abc}=\Tr(T_a\{T_b,T_c\})$. The action at linear order in $h$ is
\begin{eqnarray}
S=S^{\text{free}}_4-\frac{1}{4}\int d^4x B^{[MN]}_{(IJ)}d_{abc}\psi^{(a)I}\psi^{(b)J}\phi_{MN}^{(c)}+\text{h.c.} \phantom{-},
\end{eqnarray}
the deformation is conformal invariant at classical level for all values of $h_{ijk}$. $S^{\text{free}}_4$ denotes the action of $\mathcal{N}=4$ SYM in free field theory. Here the coupling $B^{[MN]}_{(IJ)}$ obeys
the tracelessness condition ~\cite{Bedoya:2010qz}
\begin{eqnarray}
B^{[MN]}_{(IN)}&=&0,
\end{eqnarray}
where $B^{[mn]}_{(ij)}=h_{ijl}\epsilon^{lmn}$, with latin indices running from $1$ to $3$. We will see that this condition is crucial for matching
with the field theory side.%
\subsection{The superalgebra $psl(4|4, \bf{R})$ }
The symmetry group of the action~\eqref{N=4SYM} is $PSU(2,2|4)$. Local operators of $\mathcal{N}=4$ SYM are arranged in multiplets of the algebra
$psl(4|4, \bf{R})$ \cite{Dobrev:1985qv}.
The generators of the algebra $psl(4|4,\bf{R})$ are the $sl(2)\times sl(2)$ rotations $l_{\beta}^{\;\alpha},\; \dot l_{\dot\beta}^{\;\dot\alpha } $, the $sl(4)$ rotations, the translation $p_{\alpha\dot\alpha}$, the
conformal generator $k^{\dot\alpha\alpha}$, %
the supersymmetry generators $q_{A\alpha}$, $\bar q^{A}_{\dot\alpha}$, the superconformal generators $s^{A \alpha}$, $\bar s^{\dot\alpha}_{A}$ as well as
dilatation generator $D$.
Using $\mathcal{N}=1$ language to write down the action $S_4$, it
breaks the original ${\mathcal{R}}$ symmetry into
$SU(3)\times U(1)$ .
\subsubsection{Oscillator representation of $psl(4|4, \bf{R})$ superalgebra}%
\label{oscrepres}%
The oscillator method was developed in order to construct unitary irreducible representations of non-compact groups \cite{Bars:1982ep}
in terms of its maximal compact subgroup.
Namely, the generators of the $gl(4|4, \bf{R})$ superalgebra can be represented in terms of two sets of bosonic oscillators $(a^{\alpha}, \;a^\dagger_{\alpha})$, $(b^{\dot\alpha}, b^\dagger_{\dot\alpha})$ with $\alpha, \;\dot \alpha=1,2$ and one set of fermionic oscillator
$(c_{A}, \;c^{\dagger A})$ with $A=1,2,3,4$. The non-vanishing relations of commutation ar
\begin{eqnarray}
[a^{\alpha}, a^\dagger_{\beta}]=\delta^{\alpha}_{\phantom{-}\beta}, \phantom{--} [b^{\dot\alpha},b^\dagger_{\dot\beta}]&=&\delta^{\dot\alpha}_{\phantom{-}\dot\beta} \phantom{--} \{c_{A},c^{\dagger B}\}=\delta_A^{\phantom{-}B}.
\end{eqnarray}
The $gl(4|4, \bf{R})$ generators can be written as products of two oscillators, the supercharges, translations and conformal transformations are
\begin{eqnarray}
\label{off-diagonal-generators}
q_{A\alpha}&\;=\;&c_{A}a^\dagger_{\alpha}, \;\; \phantom{--} s^{A\alpha}\;=\;c^{A\dagger}a^{\alpha}, \nonumber \\% \nonumber
\bar q^{A}_{\dot\alpha}&=&c^{A\dagger}b^{\dagger}_{\dot\alpha}, \;\; \phantom{--}\bar{s}^{\dot\alpha}_{ A}\;=\;b^{\dot\alpha}c_{A}, \nonumber \\
p_{\alpha\dot\alpha}&=&a^\dagger_{\alpha}b^\dagger_{\dot\alpha}, \;\; \phantom{--} k^{\dot\alpha\alpha}\;=\;b^{\dot\alpha}a^{\alpha}.
\end{eqnarray}
The dilatation generator reads as follows
%
%
\begin{align}
D\;=\; 1+\frac{1}{2}a^\dagger_{\gamma}a^{\gamma}+\frac{1}{2}b^\dagger_{\dot\gamma}b^{\dot\gamma},
\end{align}
and the $sl(2)\times sl(2)$ and $sl(4)$ rotation generators are
\begin{eqnarray}
\label{compact-generators}
l_{\beta}^{\;\;\alpha}&=&a^\dagger_{\beta}a^{\alpha}-\frac{1}{2}\delta_{\beta}^{\;\alpha}a^{\dagger}_{\gamma}a^{\gamma}, \nonumber \\
\bar l_{\dot\beta}^{\;\;\dot\alpha}&=&b^\dagger_{\dot\beta}b^{\dot \alpha}-\frac{1}{2}\delta^{\;\dot\alpha}_{\dot\beta}b^\dagger_{\dot\gamma}b^{\dot\gamma}, \nonumber \\
r^{A}_{\;\; B}&=&c^{\dagger A}c_B-\frac{1}{4}\delta^{A}_{\;B}c^{\dagger D}c_{D}.
\end{eqnarray}
Also there are two $gl(1)$ generators%
\begin{eqnarray}
\label{gl-1 generators}
C&=&\;1-\frac{1}{2}a^\dagger_{\gamma}a^{\gamma}+\frac{1}{2}b^\dagger_{\dot\gamma}b^{\dot\gamma} -\frac{1}{2}c^{\dagger D}c_D \nonumber \\
B&=&\;-1+\frac{1}{2}a^\dagger_{\gamma}a^{\gamma}-\frac{1}{2}b^\dagger_{\dot\gamma}b^{\dot\gamma} ,
\end{eqnarray}
which are the central charge and outer derivation, respectively.
All fields in $\mathcal{N}=4$ SYM are uncharged with respect to the central charge $C$, therefore can be dropped. This procedure leads to $sl(4|4, \bf{R})$. The
generator $B$ does not appear in commutators in $sl(4|4,\bf{R})$ and can be projected out, giving the algebra $psl(4|4, \bf{R})$.
\subsection{$\mathcal{N}=4$ SYM fields in terms of oscillators and supertwistors}
Here we give a brief description of how we can write down solutions of free SYM theory in both oscillator representation and supertwistors. Also we present an equivalence between oscillator and supertwistor pictures. We finish this section writing composite operators in both pictures.
To write down the free field components of $\mathcal{N}=4$ SYM in the oscillator picture.
Let be $|0\rangle$ an invariant non-physical vacuum under $psl(4|4,\bf{R})$, it is annhilated by $c_A$, $b^{\dot\alpha}$ and $a^{\alpha}$ \cite{Gunaydin:1984fk}.
Let us define scalar fields as
\begin{eqnarray}
\label{scalar}
\phi^{AB}\;=\;c^{\dagger A}c^{\dagger B}|0\rangle
\end{eqnarray}
the another fields without derivatives can be obtained by applying the SUSY's generators $q_{A\alpha}$ and $\bar q^{A\dot\alpha}$ on ~\eqref{scalar}. They read as follows
\begin{eqnarray}
\label{fermions}
\psi^D_{\alpha}\sim a^\dagger_{\alpha}c^{\dagger D}|0\rangle \phantom{--} , \phantom{--}\bar\psi_{M\dot\alpha}\sim \epsilon_{MNBD}b^{\dagger}_{\dot\alpha}c^{\dagger N}c^{\dagger B}c^{\dagger D}|0\rangle,
\end{eqnarray}
and the self-dual and antiself-dual of the field strength are given by
\begin{eqnarray}
\label{sd-asd}
f_{\alpha\beta}&\sim&a^{\dagger}_{\alpha}a^{\dagger}_{\beta}|0\rangle \phantom{--} \text{and} \phantom{--} \bar f_{\dot\alpha\dot\beta}\sim \epsilon_{ABDE}b^{\dagger}_{\dot\alpha}b^{\dagger}_{\dot\beta}c^{A\dagger}c^{B\dagger}c^{D\dagger}c^{E\dagger}|0\rangle,
\end{eqnarray}
where the tildes means up to a proportionality constant.%
\subsubsection{From oscillators representation to supertwistors}%
The set of oscillators introduced above can be written in terms of supertwistors variables. %
Supertwistors $\mathcal{Z}$ parametrize the space $\bf{RP}^{3|4}$ \cite{Witten:2003nn}
\begin{align}
\mathcal{Z}\;=\;
\left(
\begin{array}{c}
\lambda_{\alpha}\\
\mu^{\dot\alpha}\\
\psi^A
\end{array}
\right),
\end{align}
where $\lambda_{\alpha}$ and $\mu^{\dot\alpha}$ are two-components bosons and $\Psi^A$ is a four-components fermions. A twistor $(\lambda_{\alpha},\mu^{\dot\alpha})$
defines a two-dimensional isotropic subspaces of ${\bf{R}}^{2,2}$
\begin{align}
\mu^{\dot\alpha}+\lambda_{\alpha}x^{\dot\alpha\alpha}\;=\; 0,
\end{align}
which are called $\alpha-$planes. In the supersymmetric case, $\psi^A$ define also a plane in $\theta$ space
\begin{align}
\psi^A +\theta^{A\alpha}\lambda_{\alpha}\;=\;0,
\end{align}
where $x^{\dot\alpha\alpha}$ and $\theta^{A\alpha}$ are coordinates in superMinkowski spacetime
$\mathbb{\bf M}^{4|8}$.
The oscillators are related to the the variables $\lambda, \, \mu$ and $\psi$, say the twistor picture, by %
\begin{align}
a^{\dagger}_{\rho}\;\rightarrow\; \lambda_{\rho} , \;\;b^{\dot\alpha }\rightarrow\mu^{\dot\alpha}, \; \;\text{and}\;\; c^{\dagger A}\;\rightarrow\;\psi^A,
\end{align}
and so on. The generators of $psl(4|4, \bf{R})$ can be written in
terms of supertwistor variables leading to first-order differential operators ~\cite{Witten:2003nn}.
On-shell $\mathcal{N}=4 $ SYM fields can be described by a scalar superfield $\Phi(\lambda,\mu,\eta)$ \cite{Berkovits:2009by}:
\begin{align}
\Phi(\mathcal{Z})\;=\;\tilde f+\eta_A\tilde \psi^A+\frac{1}{2!}\eta_A\eta_B\tilde \phi^{AB}+\frac{1}{3!}\eta_{A}\eta_{B}\eta_{C}\epsilon^{ABCD}\tilde {\bar\psi}_D+\eta_{1}\eta_{2}\eta_{3}\eta_{4}\tilde {\bar f}.
\end{align}
Notice that $\tilde f$ and $\tilde{ \bar f}$ are independent of $\eta$ and that all the fields are in the twistor picture. The scalar $\Phi(\mathcal{Z})$ in twistor languague ~\cite{Adamo:2011cb,Cachazo:2004kj} is written as
\begin{align}
\label{onshell-twistor}
\Phi(\mathcal{Z})\;=\;2\pi i\int_{\mathbb{C}}\frac{ds}{s}e^{s(\mu^{\dot\alpha}\bar p_{\dot\alpha}+\eta_A\xi^A)}\bar\delta^2(s\lambda_{\alpha}-p_{\alpha}),
\end{align}%
see Appendix C of ~\cite{Adamo:2011cb} for a derivation of~\eqref{onshell-twistor}
In order to recover the spacetime dependence, we need to Penrose transform $\tilde f,\; \tilde \psi^A,\; \tilde \phi^{AB}, \; \tilde{\bar \psi}$ and $\tilde{\bar f}$.
For a review see appendix A of ~\cite{Witten:2003nn}.%
The scalar field~\eqref{scalar} is given by %
\begin{align}
\phi^{AB}(x)\;=\;\int \frac{\lambda^{\alpha}d\lambda_{\alpha}}{2\pi i}\left.\frac{\partial}{\partial \eta_A}\frac{\partial}{\partial \eta_B}\Phi(\mathcal Z)\right\vert_{\eta=0},
\end{align}
roughly speaking, we can state that the non-physical vacuum $|0\rangle$ can be related to~\eqref{onshell-twistor} up to Penrose transform. The another free fields~\eqref{fermions} and~\eqref{sd-asd} are given as follows
\vspace{.3cm}
\begin{small}
\begin{center}
\begin{tabular}{| m{4em}| m{13em}| m{21em}|
\hline \vspace{0.3cm}
Fields & Oscillators & Stwistors\\
\hline \vspace{0.9cm}
$\Psi^{D}_{\alpha}(x)$ &$\mathcal{R}a^\dagger_{\alpha} c^{\dagger D}|0\rangle$ & $\int \frac{\lambda^{\alpha'}d\lambda_{\alpha'}}{2\pi i}\lambda_{\alpha}\left.\frac{\partial}{\partial \eta_D}\Phi(\mathcal Z) \right\vert_{\eta=0}$ \\%\nonumber
\hline \vspace{0.9cm}
$\bar\psi_{M \dot\alpha}(x)$ &$\epsilon_{MNBD}\mathcal{R}b^{\dagger}_{\dot\alpha}c^{\dagger N}c^{\dagger B}c^{\dagger D}|0\rangle$ & $\epsilon_{MNBD}\int \frac{\lambda^{\alpha'}d\lambda_{\alpha'}}{2\pi i}\frac{\partial}{\partial \mu^{\dot\alpha}}\left.\frac{\partial}{\partial \eta_N}\frac{\partial}{\partial \eta_B}\frac{\partial}{\partial \eta_D}\Phi(\mathcal{Z})\right\vert_{\eta=0}$ \\%\nonumber
\hline \vspace{0.9cm}
$f_{\alpha\beta}(x)$ & $\mathcal{R}a^{\dagger}_{\alpha}a^{\dagger}_{\beta}|0\rangle$ & $\int \frac{\lambda^{\alpha'}d\lambda_{\alpha'}}{2\pi i}\lambda_{\alpha}\lambda_{\beta}\left.\Phi({\mathcal{Z}})\right\vert_{\eta=0}$ \\
\hline \vspace{0.9cm}
$\bar f_{\dot\alpha\dot\beta}(x)$& $\epsilon_{ABDE}\mathcal{R}b^{\dagger}_{\dot\alpha}b^{\dagger}_{\dot\beta}c^{A\dagger}c^{B\dagger}c^{D\dagger}c^{E\dagger}|0\rangle$& $\epsilon_{ABDE}\int \frac{\lambda^{\alpha'}d\lambda_{\alpha'}}{2\pi i}\frac{\partial}{\partial \mu^{\dot\alpha}}\frac{\partial}{\partial \mu^{\dot\beta}}\left.\frac{\partial}{\partial \eta_A}\frac{\partial}{\partial \eta_B}\frac{\partial}{\partial \eta_D}\frac{\partial}{\partial \eta_E}\Phi(\mathcal{Z})\right\vert_{\eta=0}$ \\
\hline
\end{tabular}
\end{center}
\end{small}
\vspace{0.3cm}
where $\mathcal{R}=e^{ip\cdot x}$ is the translation operator.
\subsubsection{Single trace operators}%
However in gauge theories with gauge group $SU(N)$, e.g. $\mathcal{N}=4$ SYM. We need the product of these fields leading to composite operators. In this work we are interested in
single trace composite operators.
In this section, we explain how to write down single trace operators in terms of oscillators. We set up the following notation.
\begin{enumerate}
\item If $\psi_1$ and $\psi_2$ are fermions then $\psi_1\bullet \psi_2 = {1\over 2} (\psi_1\otimes \psi_2 - \psi_2\otimes \psi_1)$, \footnote{$\bullet$ means the super-symmetrized tensor product.}
\item if $\psi_1$ and $\phi_2$ are bosons then $\phi_1\bullet \phi_2 = {1\over 2} (\phi_1\otimes \phi_2 + \phi_2\otimes \phi_2)$, and
\item if $\phi$ is boson and $\psi$ is fermion then $\phi\bullet\psi = {1\over 2} (\phi\otimes\psi + \psi \otimes\phi)$.
\end{enumerate}
At linear order in $h $, say $\mathcal{O}[B]$. The multiplet
in the oscillator picture is %
\begin{eqnarray}
\label{B-multiplet}
\mathcal{O}[B]&=& \int d^4x \prod_{j=1}^3 \mathcal{R}^{(j)} B^{[MN]}_{(IJ)} \Psi^{(IJ)}_{[MN]} ,
\end{eqnarray}%
where $\Psi^{(IJ)}_{[MN]}$ is given by %
\begin{equation}
\label{Psi}
\Psi^{(IJ)}_{[MN]} = \;a^{\dagger}_{[1} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle.
\end{equation}%
The index $j$ denotes in which singleton representation acts
the $\mathcal{R}^{(j)}$ operator.
~\eqref{B-multiplet} is invariant under conformal transformation. In Appendix ~\ref{a} we show this fact using the oscillator representation.%
\newpage
In the {\bf{twistor picture}}~\eqref{B-multiplet} is given by
\begin{eqnarray}
\label{B-yukawa-twistor}
\mathcal{O}[B]\;=\; \int d^4x\; \prod^3_{i=1} \int \frac{\lambda^{\alpha'}_{(i)}d\lambda_{\alpha'(i)}}{2\pi i}B^{[MN]}_{(IJ)}\Psi^{(IJ)}_{[MN]},
\end{eqnarray}
where $\Psi^{(IJ)}_{[MN]}$ reads as
\begin{align}
\label{Psi-twistor}
\Psi^{(IJ)}_{[MN]} \;=\;\epsilon_{MNAB}\epsilon^{\alpha\beta}\lambda_{\beta(1)}\left.\frac{\partial}{\partial \eta_{I(1)}}\Phi^{(1)}(\mathcal Z)\right\vert_{\eta=0}\;\lambda_{\alpha(2)}\left.\frac{\partial}{\partial \eta_{J(2)}}\Phi^{(2)}(\mathcal Z)\right\vert_{\eta=0}\;\left.\frac{\partial}{\partial \eta_{A(3)}}\frac{\partial}{\partial \eta_{B(3)}}\Phi^{(3)}(\mathcal Z)\right\vert_{\eta=0}.
\end{align}
We see that~\eqref{B-yukawa-twistor} is finite on-shell. We can evaluate it on the product of three off-shell
fields and the integral is convergent \footnote{We thank Prof. Andrei Mikhailov for making clear this point.}. This means that~\eqref{B-yukawa-twistor}\
defines an element of the dual space to the tensor product
of three singleton representations.%
\section{The structure of the supermultiplet}
\label{III}
In this section we describe the full supermultiplet of $\beta-$deformation in the field theory side. Instead of working in both oscillator picture and/or twistor picture, we just work out in the oscillator picture
\subsection{Descendants multiplets from $\mathcal{O}[B]$ }
It is known that $(c^{\dagger I}c_{M}\wedge c^{\dagger J}c_{N})_0$, which is $(g\wedge g)_0$ corresponds to $\Psi_{[MN]}^{(IJ)}$ ~\cite{Bedoya:2010qz}.
Below we list in a table the set of deformations arising from the multiplet $\mathcal{O}[B]$.
\begin{table}[htb]
\centering
\caption{Descendants from $\mathcal{O}[B]$
\label{table-1}
\begin{tabular}{ | m{7em} | m{12em}|
\hline \vspace{0.3cm}
$(g\wedge g)_0$ & Field theory side\\
\hline \vspace{0.3cm}
$c^{\dagger I}c_{M}\wedge c^{\dagger J}c_{N} $ & $\Psi_{[MN]}^{(IJ)}$ \\%\nonumber
\hline \vspace{0.3cm}
$a^{\dagger}_{\rho}c_{M}\wedge c^{\dagger J} c_{N}$ & $(q_{A\rho})\Psi_{[MN]}^{(IJ)}$ \\%\nonumber
\hline \vspace{0.3cm}
$a^{\dagger}_{\rho}c_{M}\wedge a^\dagger_{\varrho}c_{N}$ & $(q_{B\varrho})(q_{A\rho})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$0$ & $(q_{[D|\varpi|})(q_{B|\varrho|})(q_{A]\rho})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$b^\dagger_{\dot\alpha}c^{\dagger I}\wedge \;c^{\dagger J}c_{N} $ & $(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$b^\dagger_{\dot\alpha}c^{\dagger I}\wedge \; b^\dagger_{\dot\beta}c^{\dagger J}$ & $(\bar q^{B}_{\dot\beta})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$0$ & $(\bar q^{[D}_{\dot\rho})(\bar q^{B}_{\dot\beta})(\bar q^{A]}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$b^{\dot\alpha}c_{M}\wedge c^{\dagger J}c_N$ & $(\bar s^{\dot\alpha}_{A})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$b^{\dot\alpha}c_{M}\wedge b^{\dot\beta}c_{N}$ & $(\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$0$ & $(\bar s^{\dot\rho}_{D})(\bar s^{\dot\beta}_{B})( s^{\dot\alpha}_{A})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$a^{\alpha}c^{\dagger I}\wedge c^{\dagger J}c_N$ & $( s^{\alpha A})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$a^{\alpha}c^{\dagger I}\wedge a^{\beta} c^{\dagger J}$ & $( s^{\beta B})( s^{\alpha A})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$0$ & $( s^{\rho D})(s^{\beta B})(s^{\alpha A})\Psi_{[MN]}^{(IJ)}$\\
\hline
\end{tabular}
\end{table}
On the left hand side we listed the representations of $psl(4,4|\bf{R})$ namely $(g\wedge g)_0$.
The AdS/CFT correspondence implies that those representations are related to the field theory by a certain intertwining
operator. In the field theory side
representations of the algebra comes from the subspace in the tensor product of three singletons.
\subsubsection{Action of $q$ and $\bar q$ on $\mathcal{O}[B]$}
Here and the following subsection we describe the supermultiplet of $\beta-$deformation in the field theory side.
From the supersymmetry algebra, we know that $q_{A\rho}$ and $\bar q^A_{\dot\rho}$ commute with $P_{\mu}$. They act just on $\Psi^{(IJ)}_{[MN]}$.
To begin with, $q_{A\rho}\Psi^{(IJ)}_{[MN]}$ is %
\begin{align}
\label{qpsi}
q_{A\rho}\Psi^{(IJ)}_{[MN]}\;=\; (c_A a^{\dagger}_{\rho}) \Psi^{(IJ)}_{[MN]} \;\equiv\;&
2 \delta_A^{(I}\;a^{\dagger}_{\rho} a^{\dagger}_{[1} |0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle \nonumber
\\
& + 2 \;a^{\dagger}_{[1} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{AMNL} a^{\dagger}_{\rho} c^{\dagger L} |0\rangle,
\end{align}%
where the right hand side (rhs) is
%
%
%
\begin{eqnarray}
2 \delta_A^{(I}\;f_{\rho[1}\;\psi_{2]}^{J)}\phi_{MN}+2\epsilon_{AMNL}\psi^{(I}_{[1}\psi^{J)}_{2]}\psi^L_{\rho},
\end{eqnarray}
in terms of oscillators. To be more clear, what we get is
\begin{eqnarray}
\label{O-Arho}
\mathcal{O}_{A\rho}&\equiv& \;\int d^4x B^{[MN]}_{(IJ)}\prod_{j=1}^3 \mathcal{R}^{(j)}q_{A\rho}\Psi^{(IJ)}_{[MN]} \nonumber \\
& =& \int d^4xB^{[MN]}_{(IJ)} ( \delta_A^{I}\;\epsilon^{\alpha\beta}f_{\rho\beta}\;\psi_{\alpha}^{J}\phi_{MN}+\epsilon^{\alpha\beta}\epsilon_{AMNL}\psi^{I}_{\beta}\psi^{J}_{\alpha}\psi^L_{\rho}).
\end{eqnarray}
The result of acting with a second SUSY generator $q_{B\varrho}$ on ~\eqref{qpsi} is
\begin{align}
\label{qqpsi}
(c_B a^{\dagger}_{\varrho})(c_A a^{\dagger}_{\rho}) \Psi^{(IJ)}_{[MN]} \;=\;&
2 \delta_A^{(I}\;a^{\dagger}_{\rho} a^{\dagger}_{[1} |0\rangle \bullet a^{\dagger}_{2]} a^{\dagger}_{\varrho}\; \delta^{J)}_{B} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle
\nonumber \\ &-4 \delta_A^{(I}\;a^{\dagger}_{\rho} a^{\dagger}_{[1} |0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{BMNL} a^{\dagger}_{\varrho}c^{\dagger L} |0\rangle \nonumber
\\
& + 4 \delta^{(I}_{B}\;a^{\dagger}_{[1} a^{\dagger}_{\varrho}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{AMNL} a^{\dagger}_{\rho} c^{\dagger L} |0\rangle
\nonumber \\ &+ 2 \;a^{\dagger}_{[1} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{ABMN} a^{\dagger}_{\rho} a^{\dagger}_{\varrho} |0\rangle,
\end{align}%
it leads us to the following
\begin{eqnarray}
\label{O-BvarrhoArho}
\mathcal{O}_{(\varrho B),(\rho A)} &\equiv& \int d^4x B^{[MN]}_{(IJ)}\prod_{j=1}^3 \mathcal{R}^{(j)} (c_B a^{\dagger}_{\varrho})(c_A a^{\dagger}_{\rho}) \Psi^{(IJ)}_{[MN]} \nonumber \\
&=&\int d^4x B^{[MN]}_{(IJ)}\epsilon^{\alpha\beta}\big[\delta^{(I}_{A}\delta^{J)}_{B}f_{\rho\beta}f_{\alpha \varrho}\phi_{MN}-2 \epsilon_{BMNL}\delta^{(I}_{A}f_{\rho\beta}\psi_{\alpha}^{J)}\psi_{\varrho}^{L}\nonumber \\
&+&2\epsilon_{AMNL}\delta^{(I}_{B}f_{\varrho\beta}\psi^{J)}_{\alpha}\psi^{L}_{\rho}+\epsilon_{ABMN}\psi_{\beta}^{(I}\psi_{\alpha}^{J)}f_{\rho\varrho}\big]
\end{eqnarray}
where the terms in brackets come from~\eqref{qqpsi}.%
\vspace{.3cm}
\underline{Symmetry of $qq\Psi$}%
Antisymmetrization in $A$ and $B$ on~\eqref{qqpsi} gives us%
\begin{align}
\epsilon^{PQAB}(c_B a^{\dagger}_{\varrho})(c_A a^{\dagger}_{\rho}) \Psi^{(IJ)}_{[MN]}\;=\;0+\text{mod}(\Psi^{(IJ)}_{[MJ]}), \nonumber
\end{align}
this means that
\begin{align}
q_{[A|(\rho|} q_{|B]|\varrho)} \Psi^{(IJ)}_{[MN]}\; =\; 0+\text{traces} .
\end{align}
The above result implies that
\begin{eqnarray}
q_{[D|\varpi|}q_{B|\varrho|}q_{A]\rho}\Psi^{(IJ)}_{[MN]}&=&0.
\end{eqnarray}
Therefore we conclude that there are just two descendants~\eqref{O-Arho} and~\eqref{O-BvarrhoArho} obtained from~\eqref{B-multiplet} by applying once and twice $q$, respectively.
\vspace{.3cm}
\underline{Acting with $\bar q$ on $\Psi^{(IJ)}_{[MN]}$}
The action of $\bar q^A_{\dot\alpha}$ and $\bar q^B_{\dot\beta}\bar q^A_{\dot\alpha}$ on $\Psi^{(IJ)}_{[MN]}$ reads as follow
\begin{align}
\label{barqpsi}
(c^{\dagger A} b^{\dagger}_{\dot\alpha}) \Psi^{(IJ)}_{[MN]} \;=\;&
2 \;b^{\dagger}_{\dot\alpha} a^{\dagger}_{[1}c^{\dagger A}c^{\dagger I} |0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle \nonumber
\\
& + \;a^{\dagger}_{[1} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} b^{\dagger}_{\dot\alpha} c^{\dagger A}c^{\dagger K}c^{\dagger L} |0\rangle,
\end{align
an
\begin{align}
\label{barqbarqpsi}
(c^{\dagger B} b^{\dagger}_{\dot\beta})(c^{\dagger A} b^{\dagger}_{\dot\alpha}) \Psi^{(IJ)}_{[MN]} \;=\;&
2 \;b^{\dagger}_{\dot\beta}b^{\dagger}_{\dot\alpha} a^{\dagger}_{[1}c^{\dagger B}c^{\dagger A}c^{\dagger (I} |0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle \nonumber
\\
&-2 \;b^{\dagger}_{\dot\alpha} a^{\dagger}_{[1}c^{\dagger A}c^{\dagger (I} |0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} b^{\dagger}_{\dot\beta} c^{\dagger B}c^{\dagger K} c^{\dagger L} |0\rangle \nonumber
\\
& + \;2a^{\dagger}_{[1}b^{\dagger}_{\dot\beta}c^{\dagger B} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} b^{\dagger}_{\dot\alpha} c^{\dagger A}c^{\dagger K}c^{\dagger L} |0\rangle \nonumber
\\
&+a^{\dagger}_{[1} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL}b^{\dagger}_{\dot\beta} b^{\dagger}_{\dot\alpha} c^{\dagger B} c^{\dagger A}c^{\dagger K}c^{\dagger L} |0\rangle,
\end{align}
where we have used the momentum conservation and on-shell condition to get~\eqref{barqbarqpsi}. Those results lead us to
\begin{eqnarray}
\label{O-dotalpha*A}
\mathcal{O}^{A}_{\dot\alpha}&\equiv&\int d^4x \prod^3_{i=1} \mathcal{R}^{(i)} B^{[MN]}_{(IJ)}(c^{\dagger A} b^{\dagger}_{\dot\alpha}) \Psi^{(IJ)}_{[MN]} \nonumber \\
&=&\int d^4x B^{[MN]}_{(IJ)}\epsilon^{\alpha\beta} [\partial_{\beta\dot\alpha}\phi^{AI}\psi_{\alpha}^J\phi_{MN}+\frac{1}{2}\psi^{I}_{\beta}\psi^{J}_{\alpha}\delta^{[PA]}_{MN}\bar \psi_{ P \dot\alpha}],
\end{eqnarray}
and
\begin{eqnarray}
\label{O-dotbetadotalpha*BA}
\mathcal{O}^{BA}_{\dot\beta\dot\alpha}&\equiv& \int d^4x \prod^3_{i=1} \mathcal{R}^{(i)} B^{[MN]}_{(IJ)} (c^{\dagger B} b^{\dagger}_{\dot\beta})(c^{\dagger A} b^{\dagger}_{\dot\alpha}) \Psi^{(IJ)}_{[MN]} \nonumber \\
&=& \int d^4x B^{[MN]}_{(IJ)}\epsilon^{\alpha\beta}\big[\epsilon^{ABPI}\partial_{\beta(\dot\alpha}\bar\psi_{\dot\beta)P}\psi^{J}_{\alpha}\phi_{MN}-\partial_{\beta\dot\alpha|}\phi^{AI}\psi^{J}_{\alpha}\delta^{PB}_{MN}\bar\psi_{\dot\beta P} \nonumber \\
&+& \partial_{\beta\dot\beta|}\phi^{BI}\psi^{J}_{\alpha}\delta^{PA}_{MN}\bar\psi_{\dot\alpha P}+\frac{1}{2}\psi^{I}_{\beta}\psi^{J}_{\alpha}\delta^{AB}_{MN}\bar f_{(\dot\beta\dot\alpha)}\big],
\end{eqnarray}
where the terms in brackets of~\eqref{O-dotalpha*A} and~\eqref{O-dotbetadotalpha*BA} come from~\eqref{barqpsi} and~\eqref{barqbarqpsi}, respectively.
Acting with three times $\bar q$ on $\Psi$, {\it i.e.} $(c^{\dagger D} b^{\dagger}_{\dot\rho})(c^{\dagger B} b^{\dagger}_{\dot\beta})(c^{\dagger A} b^{\dagger}_{\dot\alpha}) \Psi^{(IJ)}_{[MN]}$, following an antisymmetrization in $A$, $B$ and $D$, we got
\begin{align}
(c^{\dagger [D} b^{\dagger}_{(\dot\rho})(c^{\dagger B} b^{\dagger}_{\dot\beta})(c^{\dagger A]} b^{\dagger}_{\dot\alpha)}) \Psi^{(IJ)}_{[MN]} \;=\;0,
\end{align}
this implies that the only descendants obtained from $\mathcal{O}[B]$ after applying once and twice $\bar q $ are~\eqref{O-dotalpha*A} and\eqref{O-dotbetadotalpha*BA}.
These computations were almost straightforward since $q$ and $\bar q$ commute with $\mathcal{R}^{(j)}$. This will not happen with $s^{A\alpha }$ and $\bar s^{\dot\alpha}_{A}$, since the commutator of them with $\mathcal{R}^{(j)}$ is non-zero
\subsubsection{Action of $s$ and $\bar s$ onto $\mathcal{O}[B]$
The commutator of $\bar s$ with $p_{\mu}$ is proportional to $q$, it implies the following resul
\begin{align}
\label{[s,R^j]}
[\bar s^{\dot\alpha}_{A}, \mathcal{R}^{(j)}]=-\frac{i}{2}\mathcal{R}^{(j)}x^{\dot\alpha\omega}a^{\dagger}_{\omega}c_{A},
\end{align}
where $x^{\dot\alpha\omega}$ is the contraction $\bar\sigma_{\mu}^{ \dot\alpha\omega}x^{\mu}$. Therefore the action of $\bar s^{\dot\alpha}_{A}$ on $\mathcal{O}[B]$ is %
\begin{align}
\label{barsO}
\tilde{\mathcal{O}}^{\dot\alpha}_{A}\;\equiv\;(b^{\dot\alpha}c_{A})\mathcal{O}[B] \;=\;& -\frac{i}{2}\int d^4x \prod_{j=1}^3 \mathcal{R}^{(j)} B^{[MN]}_{(IJ)}x^{\dot\alpha\omega}(a^\dagger_{\omega}c_A)\Psi^{(IJ)}_{[MN]},
\end{align}
where the rhs is
\begin{align}
-i\int d^4x B^{[MN]}_{(IJ)}x^{\dot\alpha\omega}\epsilon^{\alpha\beta}\{ \delta_A^{I}\;f_{\omega\beta}\;\psi_{\alpha}^{J}\phi_{MN}+\epsilon_{AMNL}\psi^{I}_{\beta}\psi^{J}_{\alpha}\psi^L_{\omega}\},
\end{align}
in terms of oscillators.
The result of acting $\bar s$ twice on $\mathcal{O}[B]$ is
\begin{align}
\label{barsbarsO}
\tilde{\mathcal{O}}^{\dot\beta\dot\alpha}_{BA}\;\equiv\;(b^{\dot\beta}c_{B}) (b^{\dot\alpha}c_{A})\mathcal{O}[B] \;=\;&-\frac{1}{4} \int d^4x \prod_{j=1}^3 \mathcal{R}^{(j)} B^{[MN]}_{(IJ)}x^{\dot\alpha\omega}x^{\dot\beta\rho}(a^\dagger_{\rho}c_B)(a^\dagger_{\omega}c_A)\Psi^{(IJ)}_{[MN]},
\end{align}
where the rhs is
\begin{align}
-\int d^4x B^{[MN]}_{(IJ)}x^{\dot\alpha\omega}x^{\dot\beta\rho}\epsilon^{\alpha\beta}\{\delta^{I}_{A}\delta^{J}_{B}f_{\omega\beta}f_{\alpha\rho}\phi_{MN}-2 \epsilon_{BMNL}\delta^{I}_{A}f_{\omega\beta}\psi_{\alpha}^{J}\psi_{\rho}^{L} \nonumber \\
\;+\;2\epsilon_{AMNL}\delta^{I}_{B}f_{\rho\beta}\psi^{J}_{\alpha}\psi^{L}_{\omega}+\epsilon_{ABMN}\psi_{\beta}^{I}\psi_{\alpha}^{J}f_{\omega\rho}\}.
\end{align}
\vspace{.3cm}
\underline{Symmetry of $\bar s \bar s \mathcal{O}[B]$
Antisymmetrization in $A$ and $B$ on ~\eqref{barsbarsO} lead us to
\begin{align}
\bar s^{(\dot\beta}_{[B}\bar s^{\dot\alpha)}_{A]}\mathcal{O}[B]=0,
\end{align}
it implies that
\begin{align}
\bar s^{(\dot\varrho}_{[D} \bar s^{\dot\beta}_{B}\bar s^{\dot\alpha)}_{A]}\mathcal{O}[B]=0 .
\end{align}
Therefore, we conclude that the only descendants obtained from ~\eqref{B-multiplet} after applying once a twice the generator $\bar s $ are ~\eqref{barsO} and ~\eqref{barsbarsO}, respectively
\vspace{.3cm}
\underline{Acting $s$ on $\mathcal{O}[B]$}
The action of the superconformal generator $s^{A\alpha}$ on $\mathcal{O}[B]$ reads as follow
\begin{eqnarray}
\label{sO}
\tilde{\mathcal{O}}^{\alpha A}&\equiv& (a^{\alpha}c^{\dagger A})\mathcal{O}[B] \nonumber \\
&= & \int d^4x \prod^{3}_{j=1}\mathcal{R}^{(j)}B^{[MN]}_{(IJ)}\Big( -ix^{\dot\omega\alpha}a^{\dagger}_{[1} b^\dagger_{\dot\omega}c^{\dagger A} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle \nonumber
\\
&-&\frac{i}{2} a^{\dagger}_{[1} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} x^{\dot\omega\alpha} b^\dagger_{\dot\omega} c^{\dagger A} c^{\dagger K} c^{\dagger L} |0\rangle \nonumber
\\
&+& 2 \delta^{\alpha}_{\;[1}\;c^{\dagger A} c^{\dagger (I}|0\rangle \bullet a^{\dagger}_{2]} c^{\dagger J)} |0\rangle \bullet \epsilon_{MNKL} c^{\dagger K} c^{\dagger L} |0\rangle
\Big),
\end{eqnarray}
where $x^{\dot\omega\alpha}$ comes from the commutator of $s^{A\alpha}$ with $\mathcal{R}^{(j)}$. The rhs of Eqn. ~\eqref{sO} is just %
\begin{align}
\int d^4x B^{[MN]}_{(IJ)}\epsilon^{\beta\rho}\{\delta^{\alpha}_{\;\rho}\phi^{AI}\psi^{J}_{\beta}\phi_{MN}-\frac{i}{2}x^{\dot\omega\alpha}(\partial_{\rho\dot\omega}\phi^{AI}\psi_{\beta}^{J}\phi_{MN}+\frac{1}{2}\psi^{I}_{\beta}\psi^{J}_{\rho}\delta_{MN}^{PA}\bar \psi_{P\dot\omega})\},
\end{align}
in the oscillator picture
$\tilde{\mathcal{O}}^{\alpha A}$~\eqref{sO} is a descendant obtained from ~\eqref{B-multiplet} and also $s^{B\varrho}s^{A\alpha}\mathcal{O}[B]$, say
$\tilde{\mathcal{O}}^{(B\varrho),(A\alpha)}$, is. However, the structure of $\tilde{\mathcal{O}}^{(B\varrho),(A\alpha)}$ is not showed explicitly here.They
are just the two descendants from~\eqref{B-multiplet}, since it can be proved that
\begin{align}
s^{[D|\varpi|}s^{B|\varrho|}s^{A]\alpha}\mathcal{O}[B]\;=\;0.
\end{align}
Those results described the structure of the supermultiplet of deformation on the field theory side and they were listed in the Table ~\ref{table-1}.
However we state that still there are another descendant multiplets which can be obtained from ~\eqref{B-multiplet}. Bellow, we list more results .
\begin{small}
\label{llll}
\begin{center}
\begin{tabular}{ | m{12em} | m{12em}|
\hline \vspace{0.3cm}
$(g\wedge g)_0$ & Field theory side\\
\hline \vspace{.3cm}
$a^{\dagger}_{\varrho}b^\dagger_{\dot\alpha}\wedge c^{\dagger J}c_N-b^\dagger_{\dot\alpha}c^{\dagger I}\wedge a^{\dagger}_{\varrho}c_N$ & $(q_{B\varrho})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$\; \; \; \; \; \;b^\dagger_{\dot\alpha}a^{\dagger}_{[\varrho}\wedge a^\dagger_{\varpi]}c_N$ &$(q_{C\varpi})(q_{B\varrho})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.4cm}
$\;\;\;\;\;\;\;0$ & $(q_{D\vartheta})(q_{C\varpi})(q_{B\varrho})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$a^{\dagger}_{\varrho}b^\dagger_{\dot\alpha}\wedge b^\dagger_{\dot\beta}c^{\dagger J}-b^{\dagger}_{\dot\alpha}c^{\dagger I}\wedge a^\dagger_{\varrho} b^\dagger_{\dot\beta}$ & $(q_{D\varrho})(\bar q^{B}_{\dot\beta})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$ \\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;b^\dagger_{\dot\alpha}a^{\dagger}_{[\varrho}\wedge a^\dagger_{\varpi]} b^\dagger_{\dot\beta}$ & $(q_{E\varpi})(q_{D\varrho})(\bar q^{B}_{\dot\beta})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;0$ & $q\;q\;q (\bar q^{B}_{\dot\beta})(\bar q^{A}_{\dot\alpha})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$a^{\alpha}b^{\dot\alpha}\wedge c_N c^{\dagger J}-b^{\dot\alpha}c_M\wedge a^{\alpha} c^{\dagger J}$& $(s^{\alpha B})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$ \\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;b^{\dot\alpha}a^{[\alpha}\wedge a^{\beta]} c^{\dagger J}$& $(s^{\beta D})(s^{\alpha B})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$ \\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;\;0$ & $(s^{\rho E})(s^{\beta D})(s^{\alpha B})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$ \\
\hline \vspace{0.3cm}
$a^{\alpha}b^{\dot\alpha}\wedge b^{\dot\beta}c_N-b^{\dot\alpha}c_M\wedge a^{\alpha}b^{\dot\beta}$ & $(s^{\alpha D})(\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$ \\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;b^{\dot\alpha}a^{[\alpha}\wedge a^{\beta]} b^{\dot\beta}$ & $(s^{\beta E})(s^{\alpha D})(\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;0$& $sss(\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$ b_{\dot\beta}^{\dagger} b^{\dot\alpha}\wedge c^{\dagger J}c_{N}-b^{\dot\alpha}c_M\wedge b^\dagger_{\dot\beta}c^{\dagger J}$ & $(\bar q^{B}_{\dot\beta})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$ \\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;b_{[\dot\beta}^{\dagger} b^{\dot\alpha}\wedge b^\dagger_{\dot\rho]}c^{\dagger J}$ & $(\bar q^{D}_{\dot\rho})(\bar q^{B}_{\dot\beta})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;0$ & $(\bar q^{E}_{\dot\varrho})(\bar q^{D}_{\dot\rho})(\bar q^{B}_{\dot\beta})(\bar s^{\dot\alpha}_{A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$b^\dagger_{\dot\rho}b^{\dot\alpha}\wedge b^{\dot\beta}c_N-b^{\dot\alpha}c_M\wedge b^\dagger_{\dot\rho}b^{\dot\beta}$& $(\bar q^{D}_{\dot\rho})(\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;b^\dagger_{[\dot\rho}b^{\dot\alpha}\wedge b^\dagger_{\dot\varrho]}b^{\dot\beta}$& $(\bar q^{E}_{\dot\varrho})(\bar q^{D}_{\dot\rho})(\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}\
$\;\;\;\;\;\;0$ & $\bar q \bar q \bar q (\bar s^{\dot\beta}_{B})(\bar s^{\dot\alpha}_{A})\Psi_{[MN]}^{(IJ)}$\\
\hline \vspace{0.3cm}
$a^\dagger_{\beta}a^{\alpha}\wedge c^{\dagger J}c_N -a^{\alpha}c^{\dagger I}\wedge a^\dagger_{\beta}c_N $ & $(q_{B\beta})( s^{\alpha A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;a^\dagger_{[\beta}a^{\alpha}\wedge a^\dagger_{\rho]}c_N$ & $(q_{D\rho})(q_{B\beta})( s^{\alpha A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;0$ & $(q_{E\varrho})(q_{D\rho})(q_{B\beta})( s^{\alpha A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$a^\dagger_{\rho}a^{\alpha}\wedge a^{\beta} c^{\dagger J}-a^{\alpha}c^{\dagger I}\wedge a^\dagger_{\rho}a^{\beta} $ & $(q_{D\rho})( s^{\beta B})( s^{\alpha A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;a^\dagger_{[\rho}a^{\alpha}\wedge a^{\dagger}_{\varrho]}a^{\beta}$ & $(q_{E\varrho})(q_{D\rho})( s^{\beta B})( s^{\alpha A})\Psi^{(IJ)}_{[MN]}$\\
\hline \vspace{0.3cm}
$\;\;\;\;\;\;0$ & $qqq( s^{\beta B})( s^{\alpha A})\Psi^{(IJ)}_{[MN]}$\\
\hline
\end{tabular}
\end{center}
\end{small}
The results listed in this table are possible, too. For instance, let us choose
\begin{align}
sss(\bar s^{\dot\beta}_B)(\bar s^{\dot\alpha}_A)\Psi^{(IJ)}_{[MN]},
\end{align}
using relation of commutations. It can be written as
\begin{align}
\label{sssbarsbarspsi-discussion}
ss(\bar s^{\dot\beta}_B)(\bar s^{\dot\alpha}_A)s^{\star\circ}\Psi^{(IJ)}_{[MN]}-\delta^{\star}_A ss(\bar s^{\dot\beta}_B) k^{\dot\alpha\circ}\Psi^{(IJ)}_{[MN]}+\delta^{\star}_Bssk^{\dot\beta\circ}(\bar s^{\dot\alpha}_A)\Psi^{(IJ)}_{[MN]} ,
\end{align}
let us recall that the complete expresion in the field theory side is~\eqref{B-multiplet}, for a brievity let us say $\Psi^{(IJ)}_{[MN]}$.
In Appendix \ref{a} we check that~\eqref{B-multiplet} is invariant under the generator of conformal transformations $k^{\dot\alpha\alpha }$. Moreover the generators $k$ and $\bar s$ commutes. Therefore the last two terms in~\eqref{sssbarsbarspsi-discussion}
vanish. With this procedure we have the following equality
\begin{align}
sss(\bar s^{\dot\beta}_B)(\bar s^{\dot\alpha}_A)\Psi^{(IJ)}_{[MN]}\;=\;ss(\bar s^{\dot\beta}_B)(\bar s^{\dot\alpha}_A)s^{\star\circ}\Psi^{(IJ)}_{[MN]}\;=\; (\bar s^{\dot\beta}_B)(\bar s^{\dot\alpha}_A)[sss\Psi^{(IJ)}_{[MN]}]\;=\;0,
\end{align}
where we have used the result listed in table \ref{table-1}.
With the same reasoning we get additional representations of the $psl(4|4,\bf{R})$ algebra, in subspaces of the tensor product of
three singleton representations
\vspace{.7cm}
{\bf{Acknowledgments}}
I am gratefull to Prof. Andrei Mikhailov for suggesting the problem and for his advice. The current work
was supported by the
Brazilian CNPq scholarship and in part by FAPESP grant 2014/18634-9.
|
\section{Introduction}\label{s_intro}
It is now well established that super massive black holes (SMBHs) are
present at the centres of galaxies with bulges. The observational
correlations found between the black-hole mass and various parameters
of the host galaxy, e.g., velocity dispersion, bulge mass, and bulge
luminosity, have led to the idea that the formation and evolution of
early type galaxies and their nuclear SMBHs are tightly linked
\citep[see e.g.][and references therein]{Kormendy2013}.
\n5419 is a luminous elliptical galaxy ($M_V = -23.1$). Such bright
early type galaxies (brighter than $M_V \sim -21$) often have a low
density `core' with a typical size of a few tens or hundreds of
parsecs. Inside the core or break radius, $r_b$, the light profile is
much shallower than the inward extrapolation of the outer S\'ersic
profile \citep[e.g.][]{Lauer1995,Lauer2005,Graham2003a,Rusli2013b}.
Moreover, core galaxies differ from fainter ellipticals in their
isophotal shapes and degree of rotational support: core ellipticals
have boxy instead of disky isophotes and are supported by anisotropic
velocity dispersions rather than rotational stellar motions
\citep[e.g.][]{Nieto1991a,Kormendy1996a,Faber1997,Lauer2012a}. These
morphological and kinematic distinctions between core galaxies and
fainter ellipticals suggest that the processes involved in their
formation are different.
The most plausible mechanism for core formation is black hole scouring
that occurs in non-dissipative mergers of galaxies as a result
of the dynamics of their central SMBHs: the SMBHs spiral into the
centre of the merger via dynamical friction, ultimately forming a
black hole binary \citep[][]{Begelman1980}. As the binary shrinks, it
ejects stars on intersecting orbits, creating a low-density core with
a mass deficit of the order of or a few times larger than the mass of
the binary \citep[see e.g.][and references
therein]{Milosavljevic2001a,Merritt2006a,Merritt2013a}. The black hole
binary model can explain the correlations between core structure,
black hole masses, mass deficits and the observed orbital structure in
core galaxies \citep[e.g.][]{Faber1997,Milosavljevic2001a,
Kormendy2009a, Hopkins2009a,Thomas2014,Thomas2016}. Dissipation-driven core
formation has been reported from numerical $N$-body simulations that
include the dynamical effects of AGN feedback, but these results have
not been tested yet in detail against the wealth of observations
available for core elliptical galaxies.
Core scouring implies the formation of black hole binaries. Their
subsequent evolution can follow different paths. In particular, if
the binary separation decreases enough and the SMBHs manage to
coalesce, the merged black hole could be ejected from the galaxy
centre by anisotropic emission of gravitational waves \citep[e.g.][]{Begelman1980}.
As the probability that the remnant black hole recoils at a velocity
exceeding the escape velocity of a large elliptical galaxy is low
\citep[e.g.][]{Lousto2012a}, it will most likely remain bound to the
galaxy on a radial orbit. On the other hand, if the binary decay
stalls, subsequent mergers may bring a third SMBH or second black hole
binary to the centre \citep[e.g.][]{Valtonen1996}. The interactions of
this newly formed multiple SMBH system will eventually displace one or
more of the SMBHs from the nucleus. In both of these scenarios,
off-centred SMBHs are expected to be found in the centre of bright
elliptical galaxies. The observational evidence of binary or recoiled
SMBHs that supports these theoretical predictions is still scarce,
usually involving pairs of SMBHs at kiloparsec separations
\citep[e.g.][]{Komossa2003a,Ballo2004,Bianchi2008,Koss2011,Fu2011a,McGurk2015,Comerford2015}. On
smaller scales, mostly indirect evidence has been reported, with only
one secure case known so far \citep[CSO 0402$+$379,
][]{Rodriguez2006}.
In this paper we study \n5419, the dominant galaxy of the poor cluster
Abell~753. This large elliptical galaxy was first classified as a
core by \citet{Lauer2005}, who modelled its inner $\sim 10$~arcsec
surface brightness profile derived from a {\it HST}/WFPC2 F555W
image.
Based on a Nuker-law \citep{Lauer1995} fit to \n5419 profile, they
derived a break radius $r_b = 2.38$~arcsec (650~pc), making this relatively
low-surface brightness core the largest among the 42 objects
classified as core galaxies in their sample. Moreover, the {\it HST}
image presented by \citet[][]{Lauer2005} revealed the presence of a
double nucleus at the centre of the galaxy, with a projected
separation of a few tens of parsecs (see also \citealt{Capetti2005b,Lena2014}; this work).
Additionally, radio observations
\citep[][]{Goss1987,Subrahmanyan2003} and the detection of hard X-rays
\citep[][]{Balmaverde2006} indicate that the centre of \n5419 hosts a
low-luminosity AGN (LLAGN). All this makes \n5419 an interesting case for studying
the interplay between core galaxies and their SMBHs.
We observed \n5419 as part of our black hole survey, consisting of 30 galaxies observed with \S\ at the Very Large Telescope (VLT, \citealt{Nowak2007, Nowak2008, Nowak2010, Rusli2011, Rusli2013a, Rusli2013b, Saglia2016}; Bender et al. in preparation; Erwin et al. in preparation; Thomas et al. in preparation).
Here we report the results obtained for \n5419. The \S\ data of \n5419 and additional long-slit spectroscopy from the Southern African Large Telescope are presented in Section~\ref{s_data}. In Sections~\ref{s_photo} and \ref{s_kin} we present the results from the imaging and spectroscopy of \n5419. Section~\ref{s_dyn} deals with the dynamical modelling of the galaxy. A discussion about the double nucleus of \n5419 can be found in Section~\ref{s_nucleus} and our final conclusions in Section~\ref{s_summary}.
We assume a distance to \n5419 of 56.2~Mpc, derived from the radial velocity corrected for Local Group infall onto Virgo taken from Hyperleda, $v_{\rm vir}=4047$\kms, and a value of $H_0 = 72~{\rm km~s^{-1}~Mpc^{-1}}$. At this distance, 1~arcsec corresponds to 273~pc.
\section{Observations and data reduction}\label{s_data}
\subsection{\S\ IFS data}\label{s_SINFdata}
Adaptive-optics-assisted near-infrared (NIR) observations of \n5419 were obtained with the \S\ integral field spectrograph at the 8~m VLT UT4 on 2009 May 20$-$24. A field of view (FOV) of $\sim 3 \times 3$~arcsec ($820 \times 820$~pc) was covered at a spatial sampling of $0.05 \times 0.10$~arcsec$^2$~pixel$^{-1}$ in the \Kb\ (1.95$-$2.45\micron) with a spectral resolution of ${\rm R}\sim 5000$.
The observations were performed according to a standard object-sky-object strategy, followed by the observation of a point spread function (PSF) star to assess the adaptive optics (AO) performance and characterize the PSF \citep[see e.g.][]{Rusli2013a,Mazzalay2013a}.
Individual 10~min exposures dithered by a few spatial pixels resulted in a total on-source exposure time of 250~min.
During the observations, the AO was operated in the laser guide star mode. This improves the spatial resolution to $\approx 0.2$~arcsec (55~pc), as given by the full width at half maximum (FWHM) of the PSF stars associated with the observations.
The \S\ data were reduced using {\sc esorex} \citep{Modigliani2007}, including all the standard reduction steps, i.e. bias subtraction, flat-fielding, bad pixel removal, detector distortion and wavelength calibration, sky subtraction \citep{Davies2007a}, reconstruction of the object data cubes and telluric correction. The data cubes were re-sampled to a spatial scale of $0.05 \times 0.05$~arcsec$^2$~pixel$^{-1}$. 25 data cubes were combined to produce one single final cube. The lower panel of Fig.~\ref{f_spectra} shows an example spectrum in the region of the NaI and CO absorption lines, obtained from the combination of the entire \S\ FOV. No emission lines are observed in our data.
\subsection{Long-slit SALT data}\label{s_SALTdata}
In addition to the NIR data for the centre, optical spectra along the position angles\footnote{Position angles are measured from north to east.}
PA=78\deg\ (hereafter MJ axis) and PA=168\deg\ (hereafter MN axis)
were obtained at the 10~m Southern African Large Telescope (SALT). The
observations utilized the Robert Stobie Spectrograph (RSS) in long
slit mode (slit width 1.25~arcsec) and grating PG2300 covering the
wavelength range 4800$-$5700~\AA. The plate scale is
0.14~arcsec~pix$^{-1}$ and the spectral resolution, measured from the
arc lamps, is ${\rm FWHM} \sim 40$\kms. Three 850 second exposures
were obtained on 2013 Mar 22 for PA 78\deg\ and two 1100 second
exposures for PA 168\deg\ were made on 2013 Apr 30. The bias-subtracted and flattened images provided by the SALT pipeline were wavelength calibrated in two dimensions by fitting ThAr and CuAr
comparison spectra with fifth order polynomials employing the
long-slit menu in {\sc iraf}. Flux calibrations used the G93-48
standard in {\sc iraf}.
The upper panel of Fig.~\ref{f_spectra} shows an example SALT spectrum in the observed frame, obtained by summing up the signal over three pixels (0.42~arcsec) around the continuum maximum, and continuum normalized.
\begin{figure}
\centering
\includegraphics[width=1.\columnwidth]{plots/fig1.eps}
\caption {Optical (SALT, upper panel) and NIR (\S, lower panel) continuum-normalized spectrum of \n5419, in the observed frame. The labels indicate the principal lines observed in the galaxy spectra.}
\label{f_spectra}
\end{figure}
\section{The light profile: a depleted core with a double nucleus}\label{s_photo}
\begin{figure*}
\centering
\includegraphics[width=1.\textwidth]{plots/fig2.eps}
\caption {Close-up of the inner $1.5 \times 1.5$~arcsec ($410 \times 410$~pc) of \n5419.
Left panels: {\it HST}/WFPC2 {\it F555W} image and \Kb\ VLT/\S\ collapsed cube and isophotes. The images are centred at the position of the continuum maximum; the spatial orientation is indicated in the upper panel. N1 and N2 denote the central point source and the off-centre nucleus, respectively. The ellipses on the upper-right corners of the images indicate the Gaussian FWHM of the respective PSFs.
Central and right panels show the corresponding PSF profiles and point-source subtracted images.}
\label{f_ims}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{plots/fig3.eps}
\caption{Surface brightness profile (circles), best-fitting Core-S\'ersic model (red-solid line) and residuals. The blue- and red-dashed curves correspond to the $HST$ surface brightness profile of \n5419 before the point source subtraction and the rescaled $HST$ PSF profile, respectively (see Section~\ref{ss_core} for details).}
\label{f_corefit}
\end{figure}
To study the light distribution of \n5419 we use {\it HST}/WFPC2 F555W\footnote{proposal ID $=$ 6587, PI $=$ D. Richstone} and 3.6\micron\ {\it Spitzer}
IRAC1\footnote{program ID $=$ 30318, PI $=$ G. Fazio} images retrieved
from the archives, as well as our \S\ collapsed cube.
Overall, \n5419 shows a smooth light distribution, with no signs of obvious distortions that
could indicate a recent merger or the presence of dust.
The {\it HST} image shows what appears to be a double nucleus at the centre of the galaxy (upper-left panel of Fig.~\ref{f_ims}), first seen by \citet{Lauer2005} and \citet{Capetti2005b}.
The brighter of the two nuclei, N1, is located at the galaxy's photocentre\footnote{A low-significance displacement of 0.6~pixel (0.02~arcsec) between the central point-source and the galaxy's photocentre of \n5419 is reported by \citet{Lena2014}.}.
It is unresolved in the {\it HST} image and is likely associated with the LLAGN of \n5419.
The second nucleus, N2, is seen in the form of a bright blob, off-centred by approximately 0.25~arcsec (70~pc) towards the south, almost aligned with the semi-minor axis of the galaxy. The off-centre nucleus has a Gaussian FWHM of $\approx 0.15$~arcsec, slightly larger than the {\it HST} PSF ($0.07$~arcsec).
Additionally, a `bridge' of weak emission is observed between the two nuclei, suggesting that they are physically related and that their proximity is not merely a projection effect.
\subsection{The depleted stellar core in \n5419}\label{ss_core}
The surface brightness profile of \n5419 was constructed by fitting ellipses to
the {\it HST} and {\it Spitzer} images. For this we employed the task ellipse of the
{\sc stsdas} package of {\sc iraf}.
As a first step, we used a rescaled $HST$ PSF to subtract the central point source N1 from the $HST$ image.
The scaling factor was computed iteratively to make the slope of the light profile inside 0.2~arcsec consistent with the one around 1~arcsec.
Once N1 was subtracted (upper-right panel of Fig.~\ref{f_ims}), we derived the final $HST$ profile masking N2, as well as several globular clusters.
Finally, the {\it HST} and {\it Spitzer} profiles were matched by determining the scaling and sky
value that minimize the magnitude square differences between the two
profiles in the 5$-$18~arcsec range. Since the galaxy fills the entire
{\it HST} FOV, the sky background was estimated from the larger
{\it Spitzer} image. Fig.~\ref{f_corefit} shows the
final surface brightness profile, with and without point source, calibrated to Cousins $R$-band using
the aperture photometry reported by \citet{Poulain1994}. The scaled $HST$ PSF is also included.
The remaining light profile of \n5419 without the two compact central
sources is well described by a Core-S\'ersic function
\citep[e.g.][]{Graham2003a, Trujillo2004a, Rusli2013b}. This is shown in Fig.~\ref{f_corefit},
where we plot the surface brightness profile together with the corresponding Core-S\'ersic fit
and its residuals. The best-fitting model has a S\'ersic index $n=7.2$,
a projected half-light radius $r_e=110.8$~arcsec, a core break radius
$r_b=1.58$~arcsec, a surface brightness at the break radius of
$\mu_b=17.09$~mag~arcsec$^2$, a flat inner slope $\gamma=0.09$ and a
transition parameter $\alpha=3.22$.
Note that the Core-S\'ersic fit
provides a much better description of \n5419's light profile than
a simple S\'ersic model, and clearly identifies this galaxy as
a core elliptical (e.g. \citealt{Rusli2013b}).
Given these parameters, we can estimate the light deficit in the core from
the luminosity difference between the best-fitting Core-S\'ersic model
and its S\'ersic component. We follow \citet{Rusli2013b} and obtain a
$R$-band luminosity deficit of $(3.78\pm0.15) \times 10^9 $~\Ls\
(assuming an extinction correction of $A_R = 0.193$) . This
corresponds to $M_{R}=-19.5$~mag or $M_{V}=-18.8$~mag (with $V-R=0.67$ from \citealt[][]{Poulain1994}).
\subsection{The double nucleus as seen by \S}
\label{ss_doublenuc}
\begin{figure}
\centering
\includegraphics[width=.8\columnwidth]{plots/fig4.eps}
\caption {$V-K$ colour maps derived form the {\it HST} and \S\ images before (upper panel) and after (lower panel) the subtraction of the central point source. The contours correspond to the {\it HST} isophotes (see Fig.~\ref{f_ims}). North is right, east is up.}
\label{f_color}
\end{figure}
The double nucleus mentioned above has only been detected in the optical
with {\it HST} thus far. Although they are not resolved, these two nuclei are also seen in our \S\ data.
The lower-left panel of Fig.~\ref{f_ims} shows a zoom into the central $r\sim 0.5$~arcsec (135~pc) of the \S\ image
of \n5419 centred at the continuum maximum. The \Kb\ isophotes are overlaid. Taking
into account the differences in spatial resolution between the \HST\
and \S\ images, it is clear that the elongation of the isophotes near
the centre is the result of the presence of the secondary nucleus.
We used these images to derive the $V-K$ colour of the different
components in the centre of \n5419. We calibrated the \HST\ image to
the \Vb\ in the Johnson system using the transformation from
\citet{Holtzman1995a} and calibrated the \S\ collapsed cube to the
\Kb\ using a 2MASS image.
We degraded the spatial resolution and pixel scale of the \HST\ image to match
those of our \S\ data and created a $V-K$ colour map.
The upper panel of Fig.~\ref{f_color} shows a close-up around N1 and N2 of
the resulting $V-K$ colour map.
The contours correspond to the \HST\ isophotes.
The central point source stands out from the rest of the galaxy by its bluer colour.
While the galaxy colour, measured in regions further away from N1 and N2, varies from $\approx 2.99-3.02$,
N1 has a colour $V-K \approx 2.8$.
However, this value does not represent that of N1 alone, since it also includes light from the galaxy.
Similarly as for the \HST\ image, we subtracted the central point source from the \S\ image (lower-right panel of Fig.~\ref{f_ims}) and derived the intrinsic $V-K$ colour of N1 from the integrated
magnitudes of the \HST\ and \S\ rescaled PSFs. Unlike in the optical, N1 is not prominent in the \Kb\
(it accounts for only $\sim 5$~percent of the light coming from the inner $r=0.15$~arcsec),
resulting in a relative blue colour of $V-K =1.13$.
The lower-panel of Fig.~\ref{f_color} shows again the $V-K$ colour map but this time with
the central point source subtracted.
A decrease of the $V-K$ colour can be clearly seen around the region where the secondary nucleus is located.
To obtain its intrinsic colour we isolated the light from N2 by subtracting
from the N1-subtracted \HST\ and \S\ images a galaxy model constructed from the ellipse-fitting results and measured the magnitudes in an aperture of $r=0.15$~arcsec ($40$~pc) centred at N2. The absolute magnitudes of N2 in the $V$- and $K$-bands are
${M_V}=-11.93$ and ${M_K}=-13.61$, corresponding to luminosities of
${L_V}=4.9 \times 10^6$~\Ls\ and ${L_K}=5.7 \times 10^6$~\Ls. The
intrinsic colour of N2 is thus $V-K=1.68$.
While the colour of \n5419 is typical of an old early type galaxy \citep[e.g.][]{Bower1992}, the off-centre
nucleus would correspond to a younger stellar population.
A summary of the measured luminosities and colours of the central point source and the off-centre nucleus are given in Table~\ref{t_color}. All values have been corrected for Galactic extinction ($A_V=0.199$ and $A_K=0.022$ from NED).
\begin{table}
\caption{Luminosities and colour of the central point source N1 and the off-centre nucleus N2 measured inside apertures of 0.15~arcsec in radius. The $V-K$ colour of the galaxy is also included for comparison.}
\label{t_color}
\centering
\begin{tabular}{cccc}
\hline
& $L_V$ [\Ls] & $L_K$ [\Ls] & $V-K$ \\
\hline
N1 & $1.3 \times 10^7$ & $9.2 \times 10^6$ & 1.13\\
N2 & $4.9 \times 10^6$ & $5.7 \times 10^6$ & 1.68\\
Galaxy & -- & -- & $\sim 3$\\
\hline
\end{tabular}
\end{table}
\subsection{Stellar mass of the off-centre nucleus}\label{ss_massnuc}
Assuming that the luminosity of N2 is entirely of stellar origin, we use the
single stellar population (SSP) models of
\citet{Maraston1998,Maraston2005} to derive some of its basic
properties. For a Salpeter initial mass function (IMF), the colour of N2
corresponds to a metal-poor, $\sim 2$~Gyr old stellar population with a \Vb\
mass-to-light ratio M/L$_{V} \simeq 1$. This M/L$_{V}$ ratio together
with the derived luminosity implies a stellar mass for the off-centre
nucleus of about $5 \times 10^6$~\Ms\ inside a region of $\approx 40$~pc
of radius.
Our stellar mass estimate lies in the range of extremely massive
clusters and low-mass dwarf galaxies and ultra-compact dwarfs
\citep[e.g.][]{Walcher2005,Erwin2015a}. Given the blue colour and
young age, it is unlikely that N2 is a massive globular
cluster. Globular clusters in early-type galaxies are typically old
\citep[ages $\gtrsim 10$~Gyr; e.g.][]{Hempel2007a,Chies-Santos2011},
although younger clusters ($\sim 2-8$~Gyr) have been claimed in a few
cases \citep[e.g.][]{Goudfrooij2001,Puzia2002,Hempel2003}. Even an
intermediate-age cluster is unlikely, since they usually have a
$V-K>2$. We can further compare the \Vb\ magnitude of N2 with
those of the other clusters observed in the \HST\ image of
\n5419. Their magnitude does not vary much from cluster to cluster,
with a mean value of $M_V= -10.45 \pm 0.60$~mag estimated from ten
clusters. These are relatively bright compared to, for example, the
ones found in the Milky Way or M31. Still, N2 is about an order
of magnitude brighter.
Therefore it is unlikely that the off-centre nucleus is a globular cluster.
It is possible that some fraction of the light we see is
due to AGN emission, and that the colour of the stellar population in
the secondary nucleus is in fact redder than assumed here. This would
imply a higher mass-to-light ratio. However, even in the most extreme
cases, the stellar mass would only increase by a factor of five.
\section{Kinematics and line strength measurements}\label{s_kin}
In order to obtain the kinematic information of \n5419 we used two
different datasets: \S\ integral field spectroscopic (IFS) data covering the inner regions of the
galaxy with high spatial resolution and long-slit SALT data obtained
along the MJ and MN axis providing information out to $\sim
15$~arcsec. In the following sections we describe the extraction of
the kinematics from these datasets and the results. Particular care
was taken in estimating the contribution from non-stellar emission to
the NIR continuum, which could bias the derived kinematics.
Additionally, in Section~\ref{s_indices} we measure optical line
indices to study the stellar populations.
\subsection{\S\ kinematics: a high dispersion region near the centre}\label{s_SINFkin}
\begin{figure*}
\centering
\includegraphics[width=1.\textwidth]{plots/fig5.eps}
\caption {Kinematic maps of \n5419 derived from the \S\ data. The spatial orientation is indicated in the left panel; the major axis of the galaxy is aligned with the $y$-axis.}
\label{f_sinf_kin}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{plots/fig6.eps}
\caption {Velocity and velocity dispersion maps of the inner region of
\n5419 derived from the \S\ data. North is right and east is up. Overplotted
are the isophotes of the {\it HST} image (see Fig.~\ref{f_ims}). The dashed circle and white ellipse in the right panel indicate the binning size and the Gaussian FWHM of the \S\ PSF, respectively.}
\label{f_sinf_kin_zoom}
\end{figure}
Non-parametric line-of-sight velocity distributions (LOSVDs) were derived using a maximum penalized likelihood technique \citep[MPL,][]{Gebhardt2000a}. We apply the approach described in detail by \citet{Nowak2007,Nowak2008}. In brief, a set of late type stellar template spectra (observed with \S\ in the same configuration as the galaxy) were convolved with the LOSVDs in order to match the continuum-subtracted galaxy spectra in the region of the first two CO bandheads [$^{12}$CO(2$-$0) and $^{12}$CO(3$-$1)]. Only stars with CO equivalent widths ($W_{\rm CO}$) similar to the galaxy were used in the fitting to minimize template mismatch.
In order to have a relatively high SNR (at least $\sim 40$), the individual spectra of the galaxy were combined using a radial and angular binning scheme following \citet{Gebhardt2003}.
The uncertainties in the derived LOSVDs were estimated from a set of 100 Monte Carlo simulations of the galaxy spectra. These were created by adding different amounts of noise to the best fitting stellar template convolved with the derived LOSVD.
While the non-parametric LOSVDs are used in the dynamical modelling
(see Section~\ref{s_dyn}), it is illustrative to express the LOSVD in
Gauss-Hermite moments. Fig.~\ref{f_sinf_kin} shows the velocity,
velocity dispersion, $h_3$ and $h_4$ maps derived from the
parametrization of the LOSVDs by Gauss-Hermite expansions up to order
four. The terms $h_3$ and $h_4$ quantify the asymmetric and symmetric deviations
of the LOSVDs from a Gaussian function, respectively.
The \S\ kinematic maps show that \n5419 is dispersion dominated,
with a velocity dispersion around $\sim 350$\kms\ and rotational motions
with an amplitude of no more than $\sim 50$\kms. $h_4$ is predominantly
negative over the entire \S\ FOV.
The velocity dispersion map of Fig.~\ref{f_sinf_kin} shows a hint of
an increase in $\sigma$ in the innermost bins ($r \lessapprox
0.4$~arcsec), almost reaching 450\kms. To explore this in more detail
we relaxed the constraint on the minimum required SNR defining our
binning scheme and derived the kinematics for the inner $r\sim
0.5$~arcsec again. In order to preserve as much of spatial
information as possible and, at the same time, ensure a SNR high
enough to obtain meaningful kinematic parameters, we integrated the
spectra over a circular aperture of $r=2$~pixels at each spatial
position. Fig.~\ref{f_sinf_kin_zoom} shows the velocity and velocity
dispersion maps of the inner $r \sim 0.5$~arcsec obtained in this
way. Since the SNR of the spectrum at each pixel is relatively low, the
resulting maps are rather noisy. However, there is an extended region
with a high velocity dispersion at the centre. It is elongated in the
N$-$S direction and seems associated to the two nuclei in the centre,
approximately following the isophotes derived from the {\it HST} image
(see Section~\ref{s_photo}).
We note that the extent of this region is much larger than the \S\ PSF shown by the
white ellipse. Unlike in the dispersion map, there is no
particular pattern in the rotational velocity in the innermost 100~pc
of \n5419.
\subsection{Is the non-stellar continuum affecting our kinematic measurements?}\label{s_WCO}
\n5419 is known to harbour a LLAGN at its centre.
Many early-type galaxies contain compact, high-brightness temperature radio cores
associated with LLAGNs and, in a few cases, parsec-scale jets are also
observed \citep[see e.g.][and references therein]{Ho2008a}.
While there is no evidence of a well-defined jet on scales of arcseconds or
higher in \n5419, this galaxy does have a compact radio core \citep[$r
< 0.7$~arcsec; e.g.][]{Goss1987,Subrahmanyan2003} similar to those typically found in LLAGNs.
It is possible, then, that the off-centre nuclear structure is
a jet similar to the optical jet observed in M87.
Since we are interested in the kinematics of the stars, which is sensitive
to the equivalent width of the CO lines, it is important
to determine how much this quantity is altered by the presence of a
non-stellar continuum. For a stellar population that contains
late-type stars, it has been shown that the $W_{\rm CO}$ is rather
independent of the star formation history and age
\citep[e.g.][]{Davies2007b}; any additional contribution to the NIR
from a non-stellar continuum will dilute the $W_{\rm CO}$.
In order to estimate the contribution from non-stellar emission to the
NIR continuum, we measured the equivalent width of the CO(2$-$0) line
at 2.29\micron\ in the unbinned \S\ spectra. Due to the low SNR of the individual
spectra, the resulting $W_{\rm CO}$ map is rather noisy. However, it does not show
any particular global pattern or any sign of a gradient towards either of the two nuclei,
suggesting that the non-stellar emission coming from the centre is not
enough to significantly alter the equivalent width of the CO.
The $W_{\rm CO}$ is consistent with a constant value of 15.7~\AA\ over the entire
\S\ FOV, with a standard deviation of 2.2~\AA. We assume that this is the
intrinsic $W_{\rm CO}$ of \n5419.
We can further compare this value with the ones measured from higher SNR spectra obtained by integrating our data in apertures of $r=0.15$~arcsec centred at the position of the central point source N1 and the off-centre nucleus N2, $W_{\rm CO}=14.5 \pm 1.5$ and $15.3 \pm 1.0$, respectively.
The maximum change in equivalent width is measured at N1, with a decrease of less than 10 percent. This would be expected if the non-stellar continuum contributes about 10 percent of the light in the CO region.
Note that these numbers are consistent with the fractional light associated with N1, as estimated from the PSF subtraction of the \S\ collapsed cube (see Section~\ref{s_photo}).
Therefore, even if all the extra light at the galaxy's unresolved photocentre
comes from an AGN, its contribution to the total light in the centre would be
very small (at most 10 percent) and the change in the CO equivalent width would
be negligible. We checked that a 10 percent AGN contribution to the CO region
has
no effects on the derived stellar kinematics.
In summary, the amount of any non-stellar light in the centre of \n5419 is not enough to affect our kinematic measurements. Moreover, since the extended high-$\sigma$ region at $r<0.35$~arcsec ($\approx 100$~pc) is not an artefact related to uncorrected continuum emission, it is unlikely that the off-centre source is an optical jet. Additional evidence against the jet scenario comes from the comparison of the colour of N2 with that of the well-studied optical jet in M87. The mean $V-K$ value of M87's jet is close to 1 \citep{Zeilinger1993,Stiavelli1997}, much bluer than the value of 1.68 we obtained for N2 (Section~\ref{ss_doublenuc}).
\subsection{Long-slit kinematics: a counter rotating core}\label{s_SALTkin}
To recover the full LOSVD from the SALT data we used the Fourier
correlation method (FCQ, \citealt{Bender1990a}). While this method has the
advantage of minimizing template mismatch, since it operates in
Fourier space, the masking of spectral regions is
problematic. Therefore, to avoid masking the spectral regions of the
gap between the SALT detector chips, we derived the kinematics using
only the 5135$-$5431\AA\ range, where the strongest absorption features
are observed (e.g. Mgb, Fe5270). The SALT spectra were binned along
the spatial direction to achieve a minimum SNR of 30 per \AA\ at all
radii. We used a 10~Gyr old synthetic stellar spectrum as kinematic
template \citep{Vazdekis2010}. The uncertainties in the kinematic
parameters were estimated from Monte Carlo simulations of synthetic
spectra with artificial noise, based on the best-fitting set of
parameters for each spectrum. Our stellar kinematic measurements
along the MJ and MN axes, are shown in Fig.~\ref{f_salt_kin}.
The kinematics derived from the SALT data in the inner $r \sim
2$~arcsec is consistent with what is seen with \S: \n5419 shows little
rotation and a velocity dispersion of $\sim 350$\kms. In fact, while
the overall rotation amplitude is low, the long-slit data reveal that
the outer parts of the galaxy ($r>5$~arcsec) are rotating in the
opposite direction as the centre. The transition region, where the
sign of the angular momentum flips, is at $r \sim 5-6$~arcsec or
roughly four times the core radius $r_b$ (see Section~\ref{ss_core}).
\begin{figure}
\centering
\includegraphics[width=1.\columnwidth]{plots/fig7.eps}
\caption {Stellar kinematics along the major (MJ) and minor (MN) axes derived from the
SALT data. Positive radii indicate projected distances E and S from
the centre for the MJ and MN axes, respectively.}
\label{f_salt_kin}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.77\textwidth]{plots/fig8.eps}
\caption {Panels (a)$-$(c): measured Lick indices profiles along the major
(circles) and minor (squares) axes. The lines correspond to the best-fit single
stellar population models. Panel (d): $\chi^2$ for each line index (the
green dot-dashed line corresponds to Mgb, the red dotted line to Fe and
the blue dashed line to \Hb) and the total $\chi^2$ (black solid line).
Panels (e)$-$(i): age, metallicity, overabundance, colours and M/L in different
bands as a function of radius. The dotted lines show the 3$\sigma$ errors.}
\label{f_line_index}
\end{figure*}
\subsection{Line indices and stellar population analysis}\label{s_indices}
The wavelength range covered by SALT allows us to measure \Hb, Mgb5175, Fe5270 and Fe5335 line strength indices. These lines are known to be useful to constrain the age (particularly \Hb) and the chemical composition of the stellar population. Other lines in the SALT range that are commonly used to constrain stellar population parameters are Fe5015 and Fe5406. However, these lines are observed at the edges of the gaps between the detector chips (see Fig.~\ref{f_spectra}) and were not included in the following analysis.
The Mg, Fe and \Hb\ line strength indices were derived for the binned spectra along the MJ and MN axis using the same synthetic star spectrum as for the kinematics. Its resolution was degraded to match that of the Lick system, and the galaxy's velocity dispersions measured in the previous section were also taken into account. We used the band definitions of \citet{Worthey1994a}.
In the following we indicate the average Iron index with $\langle {\rm Fe} \rangle = ({\rm Fe5270} + {\rm Fe5335})/2$.
To asses how well our measurements agree with the Lick system, we measured the line strength in several SSP model spectra from \citet{Vazdekis1999} multiplied by the SALT instrumental response and compared them with their Lick indices values. The largest discrepancy was found for the Fe5335 index, in which case we found a difference of 0.76~\AA\ between our measurement and the value given by \citet{Vazdekis1999}. The measured indices were corrected by these systemic offsets, and are shown in Fig.~\ref{f_line_index}.
We study the stellar population using the SSP models of the Lick line indices of \citet{Maraston1998,Maraston2005} with a Kroupa IMF. These models cover ages of up 15~Gyr,
metallicities [Z/H] from $-2.25$ to $+0.67$ and overabundances [$\alpha$/Fe] from $-0.3$ to $+0.5$ \citep[][]{Thomas2003a}.
We follow the procedure of
\citealt{Saglia2010a} (see their Section~4.1). At each radius, the age, metallicity and overabundance were derived by fitting the SSP models to the three line indices.
Fig.~\ref{f_line_index} shows the resulting age, metallicity and overabundance derived from the line indices. We also include colours and M/L ratios for different bands corresponding to the best-fit SSP models.
The models fit well the measured indices except for the case of \Hb, which is
too low to be reproduced by the models, especially at larger radii.
As a result, the age of the stellar population is found to be the maximum age explored by the models, 15~Gyr (panel {\it e} of Fig.~\ref{f_line_index}).
This is usually regarded as an effect of ionized gas emission, which partly fills the hydrogen absorption lines, leading to a weaken measurement of the \Hb\ index and, therefore, an overestimation of the age of the stellar population. We do not see any evidence of gas emission in our spectra. However, \citet{Macchetto1996} reported the detection of weak H$\alpha + $\NII\ emission in the inner few arcsecond ($r\lesssim 5$~arcsec) of \n5419.
Overall, the models point to an old, metal-rich, slightly $\alpha$/Fe-overabundant galaxy. The inferred $B-V$, $U-B$ and $V-R$ colours agree
reasonable well with those reported from broad-band photometry
\citep[$B-V=1.08$, $U-B=0.68$ and $V-R=0.67$ for
$r=11.4$~arcsec;][]{Poulain1994}, while the $V-K$ colour is similar to, though slightly redder than (possibly due to the overestimation of the stellar age) the value of $\sim3$ measured from the $HST$
and \S\ data (Section~\ref{s_photo}). The mass-to-light ratios are
around 9, 5 and 4~\Ms/\Ls\ for the $B-$, $V-$ and $R$-band,
respectively (assuming a Kroupa IMF).
\section{Dynamical modelling}\label{s_dyn}
\begin{figure*}
\centering
\includegraphics[width=0.78\textwidth]{plots/fig9.eps}
\caption {The $\chi^2$ curves (solid-red line) and model likelihoods (grey;
arbitrarily scaled) versus black hole mass $\mbh$ (left) and stellar
mass-to-light ratio $\ml$ (right; extinction corrected). The
one-dimensional model statistics were computed following
\citet{McConnell2011b}. The point with error bars indicates the best-fit
model and the 68 percent confidence interval.}
\label{f_chi2}
\end{figure*}
For the dynamical analysis of the \S\ and SALT spectra we assumed that
\n5419 is axisymmetric. The \S\ data consists of 29 symmetric
spatial bins in each quadrant. We first computed a folded kinematic
data set by averaging the respective LOSVDs $\cal L$ of the four
quadrants. The LOSVDs measured via MPL were sampled at 23 bins in the
line-of-sight velocity $v_\mathrm{los}$. Each velocity channel was
averaged individually, i.e. for a given $i \in [1,23]$ we averaged the
measured ${\cal L}_k (v_{\mathrm{los},i})$ ($k=1 \ldots 4$) of the
four quadrants, weighting each ${\cal L}_k$ by its error (and taking
into account that $v$ changes sign when crossing the galaxy's minor axis). We used these
folded SINFONI kinematics together with the long-slit data from SALT
as input for our dynamical models. The information contained in the wings
of the LOSVDs is important for the estimation of the BH mass
(e.g., \citealt{Nowak2010}, see also Appendix~\ref{s_appendix}).
As discussed in Section~\ref{s_SINFkin}, there is an extended region of
high velocity dispersion in the innermost $\approx 100$~pc of \n5419
($\approx 0.35$~arcsec). It seems
morphologically connected to the double nucleus in the centre. As the
exact nature of this double nucleus is not clear at the moment (see
Section~\ref{s_nucleus}), we decided to omit all kinematical data points
inside $r<0.35$~arcsec. Accordingly, we used the photometric galaxy
model discussed in Section~\ref{ss_core} as constraint for the light
distribution in the orbit model. In this way, our model is independent of whether
the enhanced velocity dispersion near the centre is a
feature of the global stellar population of \n5419 or whether
it comes from a separate stellar component with a different orbital structure
(associated with the double nuclear structure).
We constructed equilibrium models for \n5419 using our implementation of
Schwarzschild's orbit superposition method
\citep{Schwarzschild1979,Richstone1988,Gebhardt2003,Thomas2004,Thomas2005}. Schwarzschild
models are very flexible and do not require any a priori assumptions
upon the anisotropy of the stellar velocities. In brief, the model
construction requires four steps:
\begin{itemize}
\item Deprojection of the observed surface-brightness distribution to obtain the three dimensional intrinsic stellar luminosity density
$j$ \citep{Magorrian1999}.
\item Setup of a trial mass distribution:
\begin{equation}
\rho = \ml \times j + \rho_\mathrm{DM} + \frac{\mbh}{4 \pi r^2} \times \delta(r),
\end{equation}
assuming a stellar mass-to-light ratio $\ml$, a black hole mass \Mbh\ and a parametrised dark-matter (DM) halo density $\rho_\mathrm{DM}$ given by
\begin{equation}
\rho_\mathrm{DM} = \frac{v_c^2}{4\pi G}~\frac{3r_c^2+r^2}{(r_c^2+r^2)^2},
\end{equation}
where $r_c$ is the core radius inside which the density slope of the DM is constant and $v_c$ is the asymptotic circular velocity of the DM. $\delta(r)$ denotes Dirac's delta function.
\item Computation of a representative library of time-averaged stellar orbits.
\item Solving for the set of orbital weights or occupation numbers, respectively, that minimises the $\chi^2$ difference between the model and the observed LOSVDs.
\end{itemize}
These four steps are repeated, varying independently the free
parameters of the model: $\ml$, \Mbh\ and the parameters of the dark
matter halo. The best-fitting \Mbh\ and $\ml$ and their errors are
derived from a one dimensional likelihood as in
\citet{McConnell2011b}. We used $\sim 27,000$ orbits to sample the
space of integrals of motion for each trial $\rho$.
For the dynamical fits we used 125 constraints from the light distribution
and
759
from the kinematic observations in the radial range $0.33 \le r \le
16.6$~arcsec. As noted above, we avoided the innermost LOSVDs because of the asymmetric
structure visible in the \S\ maps and in the \HST\ image.
Our best-fit model has $\mbh = 7.24^{+2.74}_{-1.91} \times 10^9$~\Ms\
and an extinction-corrected \Rb\ mass-to-light ratio of $\ml = 5.37^{+1.86}_{-1.42}$ (see
Fig.~\ref{f_chi2}). Models without black hole and/or without DM halo are
ruled out by a $\Delta\chi^2 > 20$ and 40, respectively.
The evidence for the BH comes from the high velocity wings of the LOSVDs
inside the BH's sphere of influence and the constraints on the orbit
distribution at larger radii \citep[][see also
Appendix~\ref{s_appendix}]{Nowak2010,Rusli2013a}.
From the cumulative stellar mass distribution, we found a sphere-of-influence
radius of \n5419's black hole $-$ i.e. the radius at which the enclosed mass in stars equals the black hole mass $-$ of $r_\mathrm{SOI}
= 1.4$~arcsec. Thus, even when leaving out the very central data
points, our measurements resolve the central black hole's sphere of influence by more than a
factor of 4.
\begin{figure*}
\centering
\includegraphics[width=0.77\textwidth]{plots/fig10.eps}
\caption {Kinematic data and the best-fit model. Points with error
bars show the SINFONI (red circles) and SALT (dark-red squares) folded kinematic data $-$ from
top-left to bottom-right: rotation velocity $v$, dispersion
$\sigma$, and Gauss-Hermite parameters $h_3$ and $h_4$. Within each
panel, data points are separated by the position angle ($\Delta
\mathrm{PA}_\mathrm{maj}$) relative to the galaxy's major axis. The grey solid
lines indicate the best-fit dynamical model. Data points inside
$r<0.35$~arcsec were omitted in the fit (see Section~\ref{s_dyn}). The
dotted lines show the inward extrapolation of the dynamical model
down to the smallest measured radius. Note that for the dynamical
modelling we use the full LOSVDs (Section~\ref{s_dyn}), which is essential for
constraining the BH mass (see discussion in Appendix~\ref{s_appendix}).}
\label{f_dynmodel}
\end{figure*}
Fig.~\ref{f_dynmodel} shows the kinematics of the best-fit model
together with the (folded) data points. It also shows the innermost
data points that we omitted in the fit and the inward extrapolation of
the model into these regions. The model does not follow the strong
increase in $\sigma$ inside $0.35$~arcsec. This is not surprising. The
model is based on the fitted isophotes with the two nuclei in the
centre masked or subtracted, respectively. The flat central core in
the remaining light profile provides very little light in the centre
that could cause a strong gradient. The flat $\sigma$ distribution of
the model near the centre, despite the presence of a massive black hole implied
by the fit, together with the morphological connection of the
high-$\sigma$ region and the double nucleus suggests that the high
dispersion near the centre comes from the extra light of one or both
of the two nuclei. Without higher spatial resolution data, we can not improve
our model in the innermost regions.
\subsection{SMBH and core properties}\label{s_BH-Core}
Since \n5419 is a core galaxy, we can investigate how well its
SMBH correlates with the core properties. The strongest correlation
found by \citet{Rusli2013b} was that between SMBH mass and core radius.
The core radius of $r_b=1.58$~arcsec ($\approx 430$~pc) that we obtained
from the fitting of a Core-S\'ersic function to the surface brightness
profile implies a black hole mass of $\mbh \approx 6.4 \times 10^9 \,
\msun$, in good agreement with our measurement. In other similar core
galaxies a tight correlation between $r_{\rm SOI}$ and $r_b$ has been found,
with $r_{\rm SOI} = r_b$ \citep{Thomas2016}. The $r_\mathrm{SOI}
= 1.4$~arcsec and $r_b=1.58$~arcsec of \n5419 are consistent with this
relation.
With the mass-to-light ratio given by the dynamical models and the
light deficit estimated in Section~\ref{s_photo} ($\Delta L_R = 3.78
\times 10^9 $~\Ls), we can also compute the stellar mass deficit in the
core of \n5419. This is simply given by $M_{\rm def} = \ml \times
\Delta L = 2.0 \times 10^{10}$~\Ms. The ratio between the mass deficit
in the core and the black hole mass, $M_{\rm def}/\mbh=2.8$, is well
within the range typically derived from observations and consistent
with the theoretical predictions from simulations of single dry
mergers \citep[see][and references therein]{Rusli2013b}.
\section{The double nucleus in \n5419 and high dispersion regions}\label{s_nucleus}
In Section~\ref{s_photo} we showed that the second nucleus observed
previously in {\it HST} images is also seen in our \S\ data. This
structure, located at 0.25~arcsec ($\sim 70$~pc) from the unresolved
central source, is relatively blue compared to the rest of the galaxy;
we obtained $V-K=1.68$ for the off-centre nucleus and $\sim 3$ for the
galaxy. Our \S\ velocity dispersion map indicates a central region of
high velocity dispersion ($\sigma \gtrsim 400$\kms) that includes both
nuclei and has a similar shape as the {\it HST} isophotes at
$r \sim 0.35$~arcsec (Fig.~\ref{f_sinf_kin_zoom}). In
Section~\ref{s_WCO} we ruled out the possibility that the steep
central $\sigma$ gradient connected with the two nuclei could be
explained as an artefact related to non-stellar continuum emission
associated with the LLAGN of \n5419. Thus, the high velocity
dispersion near the centre likely has a dynamical origin. We modelled
the galaxy assuming that the stars follow a Core-S\'ersic surface
brightness distribution, as indicated by the photometry at
$r\gtrsim 0.4$~arcsec (Sections~\ref{s_photo} and \ref{s_dyn}). The
inward extrapolation of our best-fit dynamical model does not,
however, explain the excess central velocity dispersion as the result
of the massive central black hole and the stellar orbits detected in
the (cored) main body of \n5419. All this suggest that the high dispersion
comes from a distinct stellar subsystem associated with the extra light
in the double-nucleus region which we have omitted from the models.
The colour difference between the two nuclear light peaks in \n5419
makes it unlikely that the eccentric, elongated high-velocity
region is an asymmetric disk around a single black hole \citep[similar
to, though much larger, than the nuclear disk in M31;
e.g.][]{Tremaine1995,Bender2005}. Instead we assume that the
high-$\sigma$ stars originally formed a bound, undisturbed system
around one of the nuclei and that the present elongated structure is
the result of the two nuclei being in dynamical interaction. We can get an
order-of-magnitude mass estimate for the original component from
$\sigma^2={\rm G}M_{\star}/r_{\star}$, where we assume that
$r_{\star}$ is some 10~pc (upper limit from the larger of the two
nuclear structures, N2; cf. Section~\ref{s_photo}). Together with the observed
$\sigma = 400$\kms\, this points to a mass of about $M \sim 10^9\,\msun$,
with an uncertainty probably as large as a factor of $\sim 10$.
The light emitted by the off-centre nucleus N2 only accounts for a
stellar mass of $\la 10^7$~\Ms\ (see Section~\ref{ss_massnuc}). If the
compact nuclear structure was originally associated with N2, then it
immediately follows that N2 hosts a SMBH with a mass of the order of
$10^9$~\Ms. Because of the lack of any gradient in the line-of-sight velocity
of the high-$\sigma$ stars, however, it seems more plausible that they
were originally connected with the massive black hole at the galaxy's
photocentre (i.e., at rest relative to the main body of \n5419). This
would imply that a fraction of the spatially unresolved emission of N1
may still come from a very compact stellar component, in addition to
the galaxy's AGN. The fact that the interaction with N2 could (partly)
dissolve this highly bound structure implies that N1 and N2 have
roughly similar masses. In conclusion then, irrespective of whether
the high-$\sigma$ stars originally belonged to N1 or N2,
the data presented here strongly suggest that there are two SMBHs of similar
mass (to within a factor of $\sim 10$) in the nucleus of \n5419, separated
by a distance of only $\sim 70$~pc.
If the centre of \n5419 really hosts two SMBHs, then N2 is probably
the remain of a galaxy nucleus, which is also suggested by the fact
that its colour and brightness are atypical for a globular cluster
(Section~\ref{ss_massnuc}). The light profile of \n5419 doesn't show
obvious distortions that would indicate an ongoing merger, yet the
kinematically decoupled core of \n5419 provides evidence for a
(probably minor) merger in the galaxy's latest evolutionary
phase. Since these mergers can last several Gyr (e.g. \citealt{Boylan-Kolchin2008}),
we may be witnessing the last phase of such a minor merger.
Note, however, that the lack of a systemic radial velocity difference
between N1 and N2 implies a rather straight relative motion parallel
to the plane of the sky, while on larger scales the observed counter rotation requires
that at least some of the merger's orbital velocity was perpendicular
to the plane of the sky.
If \n5419 is not in a late merger phase, could the two black holes
have formed a stable, bound binary long ago? This is unlikely, since (for
two circularly orbiting similar SMBHs) one would expect both black
holes to be displaced from the centre of mass, i.e. the photocentre of
the galaxy. Furthermore, the rather straight tidal feature seen between N1 and N2
and the lack of a gradient in the line-of-sight velocity
would imply that the binary orbit has to be of low angular momentum and we must be
seeing this binary very close to face-on, which is statistically
unlikely.
A more exotic explanation for two black holes not being in a close
orbit would be that we see a SMBH which was ejected from the centre on
a low angular momentum orbit in a past multiple merger event and is
now coming back from large radii. While this scenario is
intrinsically unlikely too, the fast fly-by of N2 would
account for a number of observational aspects of \n5419. Firstly, it
would explain why we only see a straight tidal feature between N2 and
the central SMBH. Secondly, it would explain the fact that we cannot
detect signs of an ongoing merger any more.
In summary, the increased velocity dispersion at $r \la
0.35$~arcsec strongly suggests that \n5419 hosts two SMBHs in close
proximity near the centre. However, we cannot reach a definitive
picture for the origin of the perturber N2. The stellar kinematics
at $r \ga 0.35$~arcsec are consistent with a black hole mass of
$\mbh = 7.24^{+2.74}_{-1.91} \times 10^9$~\Ms\ located at the
galaxy's photocentre (Section~\ref{s_dyn}). That this $\mbh$ is only
slightly above the black hole mass we had expected if the galaxy
were similar to other massive core ellipticals without a secondary
nucleus suggests that the systematic uncertainties on $\mbh$ related
to N2 are small. Detailed numerical simulations and observations at
a higher spatial resolution are required to put stronger constraints
on the innermost mass distribution and on the origin of the two
nuclei.
\section{Summary and conclusions}\label{s_summary}
We have presented high-resolution \Kb\ \S/VLT IFS and optical
long-slit spectroscopic observations of the galaxy \n5419. The
kinematics derived from these data show that \n5419 is a
dispersion-dominated galaxy. The rotational velocity does not exceed
50\kms; although a clear rotational pattern is observed, revealing a
counter-rotating core in the inner few arcseconds. The velocity
dispersion is about 350\kms and almost constant over the low-surface
brightness core. However, inside $0.35$~arcsec ($\approx 100$~pc), where the galaxy hosts
a double nucleus, it increases reaching values of 420-430\kms.
We use orbit-based dynamical models to model the stellar kinematics
outside the double nucleus in the centre. From this analysis we derive a
\Mbh\ of $7.24^{+2.74}_{-1.91} \times 10^9$~\Ms. This mass is consistent
with the large core radius ($r_b=1.58$~arcsec or $\approx 430$~pc) obtained
by fitting a Core-S\'ersic function to the surface brightness profile,
given the known correlation between core radius and SMBH mass
\citep{Rusli2013b}.
The \Rb\ mass-to-light ratio derived from the dynamical modelling,
$\ml = 5.37$, is about a factor $1.1-1.4$ larger than the one derived
from the stellar population analysis in the inner few arcseconds, for
which we assumed a Kroupa IMF. The dynamical $\ml$ thus lies between
a Kroupa and Salpeter IMF.
We have also discussed different scenarios in which the observed
properties of the double nucleus of \n5419 can be explained. While the
nature of the double nucleus in \n5419 is certainly puzzling, our
observations suggest that this galaxy might host two SMBHs in close
proximity. If this is the case, \n5419 is a promising target to study
the interaction between SMBHs at the centre of galaxies, and their
formation and evolution. More clues could be obtained from
milliarcsecond radio imaging and higher spatial-resolution spectroscopic
data, that would allow us to study in a resolved manner the structures
seen in the central 0.5~arcsec (135~pc) of \n5419.
\section*{Acknowledgments}
Some observations used in this paper were obtained from the South African Large Telescope (SALT) under proposals 2012-1-DC-003 and 2012-2-DC-001.
\bibliographystyle{mnras}
\newcommand{\noop}[1]{}
|
\section{Free analysis}
\section{Proof of the homomorphic Tomiyama's theorem}
\label{proofhomtom}
We now prove our analog of Tomiyama's theorem for homomorphic conditional expectations.
\begin{proof}
Our proof follows Tomiyama's original method in \cite{tom57}.
Suppose $E$ is a homomorphic conditional expectation.
Without loss of generality, assume all $C^*$-algebras involved are weakly closed. (That is, we can extend everything with the Stinespring theorem.) It is sufficient to show that for any projection $e$ in $\hat{B}$
we have that $E(em) = E(e)E(m).$
Let $e$ be a projection in $\hat{B}.$ Let $x$ be a
positive element of $M.$
Note that
$$E(exe) \leq E(e\|x\| e) = \|x\|E(e).$$
So, since $E(e)$ is a projection by the homomorphic property, $$E(exe) = E(e)E(exe)E(e).$$
Now with a general element $m \in M,$
$$ 0 \leq \left( \begin{smallmatrix} 1 & 0 \\ 0 & 1- E(e)\end{smallmatrix} \right) E\left( \begin{smallmatrix} 1 & me
\\ em^* & emm^*e \end{smallmatrix} \right) \left( \begin{smallmatrix} 1 & 0 \\ 0 & 1- E(e)\end{smallmatrix} \right)
= \left( \begin{smallmatrix} 1 & E(m^*e)(1-E(e)) \\ (1-E(e))E(em) & 0\end{smallmatrix} \right), $$
and so $$(1-E(e))E(em) = 0.$$
Thus, $E(em) = E(e)E(em) =
E(e)E(em) + E(e)E((1-e)m)= E(e)E(m)$
and we are done.
\end{proof}
\section{Introduction}
Classically, R. Nevanlinna proved the following result.
\begin{theorem}[Nevanlinna \cite{nev22}]
Let $\Pi$ denote the upper half plane.
Let $f: \Pi \rightarrow \mathbb{C}.$
The function $f$ is analytic, maps $\Pi$ to $\cc{\Pi}$
and satisfies
$$\liminf_{s\rightarrow \infty} sf(sz) = -z^{-1},$$
for all $z\in \Pi,$
if and only if
there exists a probability measure $\mu$ on $\mathbb{R}$
such that
$$f(z) = \int_\mathbb{R} \frac{1}{t - z} \dd\mu(t).$$
\end{theorem}
Thus, functions with positive imaginary part satisfying good asymptotics are parametrized
by probability measures on the real line.
The quantity
$$f(z) = \int_\mathbb{R} \frac{1}{t - z} \dd\mu(t),$$
occurring in Nevanlinna's theorem is often referred to as
the \dfn{Cauchy transform}.
Recent work by Anshelevich and Williams \cite{anw14, will13, 2015williams}
has explored the connection between distribution and function theory in free probability in terms of the noncommutative Cauchy transform and the related $R$-transform.
The Cauchy transform and the $R$-transform have served as a vibrant part of free probability, which is evidenced by the large amount of recent work on the subject.
We resolve the correspondence between Cauchy transforms and the class of functions on the upper half plane in the noncommutative context of operator-valued free probability and
free analysis.
\subsection{The noncommutative context}
Let $B$ be a $C^*$-algebra.
The \dfn{matrix universe over $B$}, denoted $\MU{B},$ is the set of square matrices over $B,$ that is
$$\MU{B} = \bigcup^{\infty}_{n=1} M_n(B).$$
Next, the \dfn{upper half plane over $B$,} denoted $\UHP{B},$ is given by
$$\UHP{B} = \{X \in \MU{B} | \hspace{2pt} \text{Im } X > 0\}.$$
Here, we say a self-adjoint operator $A > 0$ if its spectrum is contained in the positive reals and $A \geq 0$ if $A$ has spectrum contained in the non-negative reals.
Similarly, the \dfn{closed upper half plane over $B$,} denoted $\CUHP{B},$ is
$$\CUHP{B} = \{X \in \MU{B} | \hspace{2pt} \text{Im } X \geq 0\}.$$
For any $\mathcal{D} \subset \MU{B_1},$ a \dfn{free function}
$f: \mathcal{D} \rightarrow \MU{B_2}$ is graded and respects intertwining maps. That is, $f$ takes an $n \times n$ matrix over
$B_1$ to an $n \times n$ matrix over $B_2$, and
if
$\Gamma X = Y \Gamma$ for some rectangular matrix $\Gamma$ of scalars, then
$\Gamma f(X) = f(Y) \Gamma.$
We denote the set of free functions for $\mathcal{D}$ to $\mathcal{R}$
by $\Free{\mathcal{D}}{\mathcal{R}}.$ (For more elaborate exposition regarding free analysis, see e.g. the comprehensive presentation in \cite{vvw12}.)
In this noncommutative context, a \dfn{free Pick function} is just a free function $f: \UHP{B_1} \rightarrow \CUHP{B_2}.$
Given:
\begin{enumerate}
\item A $C^*$-algebra $B$,
\item A von Neumann algebra $M$ unitally containing $B,$
\item An unbounded self-adjoint operator
$A$ affiliated to $M,$ that is, an operator so that each of its spectral projections are contained in
$M,$
\item A noncommutative conditional expectation
$E: M \rightarrow B,$ that is, $E$ is a completely positive unital map satisfying $E(b_1mb_2) = b_1E(m)b_2$
for all $b_1, b_2 \in B$ and
$m \in M$,
\end{enumerate}
we define the \dfn{noncommutative Cauchy transform} of $A$
to be the free function
$f: \UHP{B} \rightarrow \UHP{B}$ given by the equation
$$f(Z) = \tidr{E}\left(\left(\tir{A}- Z\right)^{-1}\right),$$
where $\mathrm{id}$ denotes the identity map on matrices.
We have adopted a \emph{vertical tensor notation} to save space: $\tensor{A}{B}$
represents the same object as $A \otimes B$.
The obvious analogue of Nevanlinna's theorem would be that
any free function $f: \UHP{B} \rightarrow \UHP{B}$ satisfying
$$\lim_{
\begin{matrix}
s\rightarrow +\infty \\
s \in \mathbb{R}
\end{matrix}
} sf(sZ) = -Z^{-1}$$
for all $Z \in \UHP{B}$
would be given by a noncommutative Cauchy transform arising from some $M, E$ and $A$ which could be constructed from $f.$
\emph{The obvious analogue of Nevanlinna's theorem is shown to be
false in Subsection \ref{false}, and thus the ability to reconstruct an algebra, a conditional expectation and an unbounded operator from
a free function $f: \UHP{B} \rightarrow \UHP{B}$ is resolved in the negative.}
However, in an expanded ``homomorphic'' notion of conditional expectation, we show that self maps of the noncommutative upper half plane satisfying good asymptotic conditions are parametrized by Cauchy transforms.
\subsection{Main result}
\begin{definition}
Let $B$, $M$ be $C^*$-algebras.
Let $\hat{B}$ be a unital subalgebra of $M.$
We define a \dfn{homomorphic conditional expectation}
to be a completely positive unital map
$E: M \rightarrow B$
such that $E|_{\hat{B}}$ is a homomorphism.
\end{definition}
The name homomorphic conditional expectation is justified by the following analogue of Tomiyama's theorem \cite{tom57}.
\begin{proposition}[Homomorphic Tomiyama's theorem]
If $E: \hat{B} \rightarrow M$ is a homomorphic conditional expectation over $B,$ then
for all $b_1, b_2 \in \hat{B},$ $$E(b_1mb_2) = E(b_1)E(m)E(b_2).$$
\end{proposition}
We prove the above proposition in Section \ref{proofhomtom}
\begin{definition}
Let $B$, $\hat{B}$ be $C^*$-algebras.
We define a \dfn{symmetric dilation} to be a
completely positive map
$\psi: B \rightarrow \hat{B}$
so that there exists a $*$-homomorphism
$E: \hat{B} \rightarrow B$
such that $E \circ \psi$ is the identity.
\end{definition}
Our main result is as follows.
\begin{theorem}\label{mainresulthomcond}
Let $f: \UHP{B} \rightarrow \CUHP{B}$ be a free function.
The following are equivalent
\begin{enumerate}
\item For all $Z \in \UHP{B},$
$$\lim_{
\begin{matrix}
s\rightarrow +\infty \\
s \in \mathbb{R}
\end{matrix}
} sf(sZ) = -Z^{-1}.$$
\item There exists:
\begin{enumerate}
\item A von Neumann algebra $M,$
\item A unital subalgebra of $\hat{B} \subseteq M,$
\item An unbounded self-adjoint operator $A$ affiliated to $M,$
\item A homomorphic conditional expectation
$E: M \rightarrow B,$
\item A symmetric dilation
$\psi: B \rightarrow \hat{B}$
such that $E \circ \psi$ is the identity,
\end{enumerate}
so that the function $f$ can be written as
$$f(Z) = \tensor{E}{\mathrm{id}_n}\left[
\left(\tensor{A}{I_n}-
\tensor{\psi}{\mathrm{id}_n}(Z)\right)^{-1}\right].
$$
\end{enumerate}
\end{theorem}
We note that Williams showed that the above theorem holds when $E$ is a conditional expectation and $\psi$ is an identity map if we assume additionally that $f$ has some large analytic continuation at infinity corresponding to the classical compactly supported case \cite{will13}. Our result also generalizes previous results in \cite[Section 5]{pastd14}. In the language of this paper, the representations established in the earlier setting held for $B = \mathbb{C}^m.$
We emphatically take the viewpoint that homomorphic conditional expectations are what makes the Nevanlinna theorem work for noncommutative Cauchy transforms-- we leave to the reader whether or not they generate any deeply interesting analogue of operator-valued free probability.
However, we view that our results suggest that either
${\bf(1)}$ free function theory is an incomplete method for understanding free probability or ${\bf(2)}$ that theorems in free probability should extend somewhat trivially to
``homomorphic'' operator valued free probability.
\subsection{Failure of the main result in the usual free probabilistic case}\label{false}
We note that we cannot always reduce to the case where
the symmetric dilation
$\psi$ is the identity map and $E$ is a \emph{bona fide} conditional expectation.
Take $B = \mathbb{C}^2.$
Define $\psi(z_1,z_2) = (z_1,z_2,\frac{1}{2}(z_1+z_2)).$
Define $E(w_1,w_2,w_3) = (w_1,w_2).$
So, we have that $\hat{B} = \mathbb{C}^3.$
Now define $A$ acting on $\mathbb{C}^3$ to satisfy
$A (w_1,w_2,w_3) = A(w_1, w_3, w_2).$
Consider
$$f(Z) = \tensor{E}{\mathrm{id}_n}\left[
\left(\tensor{A}{I_n}-
\tensor{\psi}{\mathrm{id}_n}(Z)\right)^{-1}\right].
$$
One can show that for $(z_1, z_2)\in \mathbb{C}^2$,
$$f(z_1, z_2) = (-z_1^{-1}, -z_2^{-1}(1-2(z_1+z_2)^{-1}z_2^{-1})).$$ Now we observe that
$$f(z_1, z_2) = (-z_1^{-1}, 0) + \sum_k (0, -z_2^{-1}[2(z_1+z_2)z_2]^{-k}).$$
If we could choose $\tilde{E}$ a conditional expectation and $\tilde{\psi}$ to be the identity, the homogeneous terms in the above expansion would be polynomials in $z_1^{-1}$ and $z_2^{-1}$ but, evidently, they are not.
\section{Proof of the main result}
We now prove our main theorem, Theorem \ref{mainresulthomcond}.
The \dfn{ball over $B$,} denoted $\BallB{B},$ is the set of contractive matrices over $B,$ that is,
$$\BallB{B} = \{X \in \MU{B} | \hspace{2pt} \|X\|<1\}.$$
Similarly, the \dfn{right half plane over $B$,} denoted $\RHPB{B},$ is
$$\RHPB{B} = \{X \in \MU{B} | \hspace{2pt} \text{Re } X \geq 0\}.$$
In \cite{ppt16}, the following was proved.
\begin{theorem}[\cite{ppt16}] \label{finalcor}
Let $h: \BallB{B_1} \rightarrow \RHPB{B_2}.$
Then there exists:
\begin{enumerate}
\item A $C^*$-algebra $M$ unitally containing $B_1,$
\item A completely positive linear (not necessarily unital) map $R: M \rightarrow B_2,$
\item A unitary $U \in M,$
\item A bounded self-adjoint operator $T,$
\end{enumerate}
such that
\beq \label{herglotzfinal}
h(X) =\tensor{ iT}{I_n} + \tensor{R}{\mathrm{id}_n}
\left[\left( I + \tensor{U}{I_n}X\right)\left(I - \tensor{U}{I_n}X\right)^{-1}\right].
\eeq
\end{theorem}
We note that although the statement in \cite[Corollary 3.6]{ppt16} assumes an exactness hypothesis on $B_1$, recent advances in Agler model theory by Ball, Marx and Vinnikov in the preprint \cite[Corollary 3.2]{BMV16} give the full result by \cite[Lemma 3.3]{ppt16}.
We use Theorem \ref{finalcor} to show the following Nevanlinna representation via a Hilbert space geometric derivation.
\begin{theorem}
Let $f: \UHP{B_1} \rightarrow \CUHP{B_2}$ be a free function.
The following are equivalent
\begin{enumerate}
\item
$$\liminf_{
\begin{matrix}
s\rightarrow +\infty \\
s \in \mathbb{R}
\end{matrix}
} |isf(is)| < \infty.$$
\item There exists:
\begin{enumerate}
\item A von Neumann algebra $M,$
\item An unbounded self-adjoint operator $A$ affiliated to $M,$
\item A completely positive unital map
$\psi: B_1 \rightarrow M,$
\item A completely positive map
$R: M \rightarrow B_2,$
\end{enumerate}
so that the function $f$ can be written as
$$f(Z) = \tensor{R}{\mathrm{id}_n}\left[
\left(\tensor{A}{I_n}-
\tensor{\psi}{\mathrm{id}_n}(Z)\right)^{-1}\right].
$$
\end{enumerate}
\end{theorem}
\begin{proof}
We adopt the technique used in the proof of a general Nevanlinna types theorem as in \cite{aty13,pastd14}.
Let $f$ be as in the statement of the Theorem. By concretely realizing Theorem \ref{finalcor}, we can instantiate a Herglotz function $h$ which satisfies
$ih((Z + i)^{-1} (Z - i))=f(Z) - T$ for some self-adjoint $T$. By Theorem \ref{finalcor}, $h$ can be written concretely as
$$h(\Lambda) = \tir{V^\ast} \left(\tensor{L}{I} - \Lambda\right)^{-1} \left(\tensor{L}{I} + \Lambda\right) \tir{V}.$$
(Here we have concretely written $R(x) = V^\ast x V$ and used a resolvent of the form $(L - X)^{-1}(L+X)$ instead of $(1-UX)^{-1}(1+UX)$ to agree with \cite{aty13, ag90}. However, $L$ is still a unitary. If fact, the algebra will show that $L = U^\ast.$)
Let
$$f(Z)-\tir{T} = i\tir{V^\ast}\left(\tensor{L}{I} - {(Z + i)^{-1} (Z - i)}\right)^{-1} \left(\tensor{L}{I} + {(Z + i)^{-1} (Z - i)}\right)\tir{V}.$$
One can show as an elementary exercise in the spectral theorem that every vector of the form $Vw$ is in the domain of the normal inverse
$(1-L)^{-1}.$ Notably, this reduces to an exercise in measure theory and manipulation of classical Herglotz integrals.
\begin{lemma}
Any vector of the form $Vw$ is in the domain of the normal inverse of $(1-L)^{-1}.$ Namely, the range of $V$
is perpendicular to the kernel of $1-L.$
\end{lemma}
\begin{proof}
Consider our function
$$f(Z)-\tir{T} = i\tir{V^\ast}\left(\tensor{L}{I} - {(Z + i)^{-1} (Z - i)}\right)^{-1} \left(\tensor{L}{I} + {(Z + i)^{-1} (Z - i)}\right)\tir{V}.$$
Evaluate at $Z = is.$
$$f(is)-T = iV^\ast\left(L - {(is + i)^{-1} (is - i)}\right)^{-1} \left( L+ {(is + i)^{-1} (is - i)}\right)V.$$
So, since $L$ is unitary and thus normal, evaluating $w^*(f(is)-T)w$ gives, via the the spectral theorem,
\begin{align*}
w^*(f(is)-T)w& = iw^*V^\ast\left(L - {(is + i)^{-1} (is - i)}\right)^{-1} \left( L+ {(is + i)^{-1} (is - i)}\right)Vw
\\&= i\int_{\mathbb{T}}
\frac{\omega + {(is + i)^{-1} (is - i)}}
{\omega - {(is + i)^{-1} (is - i)}}
d\mu_{Vw}(\omega)
\\& = i\int_{\mathbb{T}}
\frac{\omega(s+1) + (s - 1)}
{\omega(s+1) - (s-1)}
d \mu_{Vw}(\omega).
\end{align*}
Note that the condition
$$\liminf_{
\begin{matrix}
s\rightarrow +\infty \\
s \in \mathbb{R}
\end{matrix}
} |isf(is)| < \infty.$$
implies \emph{a fortiori} that
$$\liminf_{
\begin{matrix}
s\rightarrow +\infty \\
s \in \mathbb{R}
\end{matrix}
} s\text{Im }f(is) < \infty.$$
So, consider
\begin{align*}s\text{Im }w^*f(is)w & = s\text{Im }w^*(f(is)-T)w
\\&= s\text{Im }i\int_{\mathbb{T}}
\frac{\omega(s+1) + (s - 1)}
{\omega(s+1) - (s-1)}
d\mu_{Vw}(\omega)
\\&= \int_{\mathbb{T}}
\frac{s^2}
{s^2 + 1 - (s^2 - 1)\text{Re }\omega}
d\mu_{Vw}(\omega)
.\end{align*}
As $s$ goes to infinity,
noting that the integrand is monotone increasing in $s$, by monotone convergence theorem
$$\int_{\mathbb{T}}
\frac{1}
{1 - \text{Re }\omega}
d\mu_{Vw}(\omega)
= \liminf_{s\to\infty} s\text{Im }w^*f(is)w < \infty.$$
Since
$$\int_{\mathbb{T}}
\frac{1}
{1 - \text{Re }\omega}
d\mu_{Vw}(\omega) = \int_{\mathbb{T}}
\frac{2}
{|1 - \omega|^2}
d \mu_{Vw}(\omega),$$
we are done, because $Vw$ is the domain of $f(L)$
if and only if $|f|^2$ is integrable with respect to $d\mu_{Vw}.$
\end{proof}
Straightforward algebra gives
\begin{align*}
f(Z)-\tir{T} &= i\tir{V^\ast}\left(\tensor{L}{I} - {(Z + i)^{-1} (Z - i)}\right)^{-1} \left(\tensor{L}{I} + {(Z + i)^{-1} (Z - i)}\right)\tir{V} \\
&= i\tir{V^\ast}\left({(Z+i)} \tensor{L}{I} - {(Z - i)} \right)^{-1} \left({(Z + i)} \tensor{L}{I} + {(Z - i)}\right)\tir{V} \\
&= i\tir{V^\ast}\left({Z} \tensor{L - I}{I} + i\tensor{L+I}{I}\right)^{-1} \left( {Z} \tensor{L+I}{I} - i \tensor{L - I}{I} \right)\tir{V}.
\end{align*}
Decompose $L$ into blocks acting on $\ker 1 - L$ and $\ker(1-L)^\perp$ as $$L = \left[ \begin{smallmatrix} 1 & 0 \\ 0 & L_0 \end{smallmatrix} \right]$$ so that $\ker 1 - L_0$ is trivial. Multiply through on the left by $I = \left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1- L_0)^{-1} \end{smallmatrix} \right] \left[ \begin{smallmatrix} 1 & 0 \\ 0 & 1- L_0 \end{smallmatrix} \right]$.
We get
\begin{align*}
&i\tir{V^\ast}
\tir{\left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1- L_0)^{-1} \end{smallmatrix} \right]}
\tir{\left[ \begin{smallmatrix} 1 & 0 \\ 0 & 1 - L_0 \end{smallmatrix} \right]}
\left(Z\tir{ (L- I)} +\tir{ i(L +I)}\right)^{-1} \left(Z\tir{(L+I)} -\tir{ i (L - i)}\right)
\tir{V} \\
&=i\tir{V^\ast} \tir{\left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1- L_0)^{-1} \end{smallmatrix} \right]}
\left(Z \tir{(L - I)}\tir{ \left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1 - L_0)^{-1} \end{smallmatrix} \right]} + \tir{i (L + I)}\tir{\left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1 - L_0)^{-1} \end{smallmatrix} \right]}\right)^{-1} \\ &\hspace{1in} \times \left(Z\tir{(L + I)} -\tir{ i (L - i)}\right)\tir{V}\\
&=i\tir{V^\ast}
\tir{\left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1- L_0)^{-1}\end{smallmatrix} \right]}
\left(Z \tir{\left[ \begin{smallmatrix} 0 & 0 \\ 0 & -1 \end{smallmatrix} \right]} + \tir{i \left[ \begin{smallmatrix} 2 & 0 \\ 0 & (1 + L_0) (1 - L_0)^{-1} \end{smallmatrix} \right]}\right)^{-1} \left(Z\tir{(L+I)} -\tir{ i (L - I)}\right) \tir{V}\\
&= i\tir{V^\ast}\tir{\left[ \begin{smallmatrix} 1 & 0 \\ 0 & (1- L_0)^{-1} \end{smallmatrix} \right]}
\left(Z \tir{\left[ \begin{smallmatrix} 0 & 0 \\ 0 & -1 \end{smallmatrix} \right] }+ \tir{i \left[ \begin{smallmatrix} 2 & 0 \\ 0 & (1 + L_0) (1 - L_0)^{-1} \end{smallmatrix} \right]}\right)^{-1} \\ &\hspace{1in} \times \left(Z\tir{\left[ \begin{smallmatrix} 2 & 0 \\ 0 & L_0 + 1\end{smallmatrix} \right]} - \tir{ i \left[ \begin{smallmatrix} 0 & 0 \\ 0 & L_0 - I \end{smallmatrix} \right]}\right)\tir{V}
\end{align*}
The operator $$A = i \frac{1 + L_0}{1 - L_0}$$ is a densely defined self-adjoint unbounded operator since $L_0$ has no kernel
Since we are only interested in $V^\ast M(Z) V$, the upper triangular form of
$$\left(Z \tir{\left[ \begin{smallmatrix} 0 & 0 \\ 0 & -1 \end{smallmatrix} \right]} + \tir{i \left[ \begin{smallmatrix} 2 & 0 \\ 0 & (1 + L_0) (1 - L_0)^{-1} \end{smallmatrix} \right]}\right)^{-1} $$ and the structure of $V$, namely that $V$ is perpendicular to the kernel of $1-L$ , gives that the relevant operator is the (2,2) block. Then compress $Z$ to $$\tidr{\psi}(Z) = Z_\psi=\tir{P}Z\tir{P^*}$$
where $P$ is the projection onto the perp of the kernel of
$I-L$.
Then our resolvent has the form
\begin{align*}
&f(Z) - \tir{T} = \tir{V^\ast (I - L_0)^{-1}}
\left(\tir{A} - Z_\psi\right)^{-1}
\left(iZ_\psi\tir{(L_0 + I)} + \tir{(L_0 - I)}\right)\tir{V} \\
&=\tir{V^\ast (I - L_0)^{-1}}
\left(\tir{A} - Z_\psi\right)^{-1}
\left(iZ_\psi\tir{(L_0 + I)(I - L_0)^{-1}} + I \right)\tir{(I - L_0) V}\\
&= \tir{V^\ast (I- L_0)^{-1}} \left(\tir{A} - Z_\psi\right)
^{-1} \left(Z_\psi \tir{A} + I\right) \tir{(I - L_0) V}\\
&=\tir{V^\ast (I - L_0)^{-1}} \left(\tir{A} - Z_\psi\right)^{-1} \left(Z_\psi \tir{A} - \tir{A^2} + \tir{A^2} + I\right)\tir{ (I - L_0) V} \\
&= \tir{V^\ast (I - L_0)^{-1}} \left(\tir{A} - Z_\psi\right)^{-1}
\left[\left(Z_\psi - \tir{A})\tir{A} + (\tir{A^2} + I\right)\right] \tir{(I - L_0) V}\\
&= \tir{V^\ast (I - L_0)^{-1} A (I - L_0)V} + \tir{V(I - L_0)^{-1}} \left(\tir{A} - Z_\psi\right)^{-1} \left(\tir{A^2} + I\right) \tir{(I - L_0)V} \\
&= \tir{V^\ast AV} + \tir{V^\ast (I - L_0)^{-1} }\left(\tir{A} - Z_\psi\right)^{-1} \left(\tir{A^2} + I\right) \tir{(I - L_0)V}\\
&= \tir{V^\ast AV} +\tir{V^\ast (I - L_0)^{-1} }\left(\tir{A} - Z_\psi\right)^{-1} \tir{(I - L_0^\ast)^{-1}V}.
\end{align*}
Now, the asymptotic condition implies that the constant terms must vanish, so
$$f(Z) = \tir{V^\ast (I - L)^{-1} }\left(\tir{A} - Z_\psi\right)^{-1} \tir{(I - L^\ast)^{-1}V}.$$
Defining a new $R(x) = V^\ast (I - L)^{-1} x (I - L^\ast)^{-1} V $ and $\psi$ to be as above,
we are done.
\black
\end{proof}
The main result Theorem \ref{mainresulthomcond} now follows by noting that
$$E(-\psi(Z)^{-1}) = \lim_{
\begin{matrix}
s\rightarrow +\infty \\
s \in \mathbb{R}
\end{matrix}
} sf(sZ) = -Z^{-1}.$$
So we see that
$$E(\psi(Z)^{-1}) = Z^{-1}.$$
One can show that
$$E(\psi(H_1)\ldots\psi(H_k)) = H_1 \ldots H_k,$$
by taking $Z = I_{k+1} - H,$ where $H$ has $H_1, \ldots, H_k$
on the upper diagonal.
\begin{lemma}
$$E(\psi(H_1)\ldots\psi(H_k)) = H_1 \ldots H_k.$$
\end{lemma}
\begin{proof}
Note
$$(I - H)^{-1} = \sum^{\infty}_{i=0} H^i,$$
and
$$\tidr{E}\left(\tidr{\psi}(I - H)^{-1}\right)
= \sum^{\infty}_{i=0} \tidr{E}\left(\left[\tidr{\psi}(H)\right]^i\right).$$
So we obtain that
$\tidr{E}\left(\left[\tidr{\psi}(H)\right]^k\right) = H^k.$
Evaluating at
$$H = \left( \begin{smallmatrix} 0 & H_1 \\
& \ddots & \ddots &\\
& & 0 & H_k \\
& & & 0 \end{smallmatrix} \right)$$
and looking at the block $(1, k+1)$ entry
gives the claim.
\end{proof}
Now, we obtain the necessary homomorphic properties by
letting $\hat{B}$ be the algebra generated
the range of $\psi, $ so we are done.
|
\section{Introduction}
Over the last decade, one of the main challenges in theoretical physics, and particularly in cosmology, refers to identifying the mysterious nature of the two dominant components that, according to observations and most of the theoretical frameworks, compose the universe, namely: the so-called dark energy and dark matter. While the latter is likely behaving nowadays as a pressureless fluid, the former should have an effective negative pressure in order to explain the accelerated expansion of the universe. Most of the dark energy models assume an extra and unknown field which would be the responsible for the accelerated expansion, but other realistic proposals include infrared modifications of General Relativity (for some reviews, see \cite{Nojiri:2010wj}). In any case, the main problem arises because of the large number of models, either extra fields or modified gravity, capable of explaining the observational data and leading to similar statistical evidence.
However, searching for models that include other effects may provide an alternative way of exploring the nature of the dark sector to, hopefully, allow for discriminating among the different theoretical models. In this sense, some proposals introduce the possibility of an interaction between dark matter and dark energy, which may reveal new features of both components (for a recent review see \cite{Wang:2016lxa}). Indeed, several interactions have been suggested, where basically the full Lagrangian contains a particular interaction term, including non-minimally coupled theories \cite{Koivisto:2005nr}. A different and more phenomenological way of exploring such a possibility goes directly through the field equations, where an interacting term, usually dubbed $Q$, is included in the continuity equations such that the total energy is conserved but a flow of energy exists between the two components, assuming that both behave like perfect fluids \cite{Zimdahl:2001ar}-\cite{Pereira:2008at}. Note that, despite the fact that these models are constructed from a phenomenological viewpoint, they can also be obtained from a variational principle \cite{Skordis:2015yra}. Moreover, a major motivation to explore dark couplings, besides the search for new physics, lies on the possibility of solving the coincidence problem, since some suitable interaction terms lead to attractor solutions with an order one ratio of dark matter and dark energy \cite{Zimdahl:2001ar,Boehmer:2008av}. The existence of such scenarios has been explored in different frameworks, from holographic dark energy to periodic universes or future singularities \cite{Nojiri:2005sr,Nojiri:2005sx}. An important issue that these models face is the potential presence of instabilities at early cosmological times \cite{He:2008si}. In any case, it is interesting to notice that some cosmological observations point out to the viability of some of these interactions \cite{Pereira:2008at}.
The required negative pressure for the dark energy component has led to exploring models with an effective equation of state (EoS) that violates the null energy condition (NEC), $\rho+p>0$, usually called {\it phantom} \cite{phantom}. Violations of the NEC can be easily achieved with extra fields that can arise from modified gravity theories or high energy physics. However, violating the NEC is not harmless and, in fact, phantom models may give rise to divergences in some cosmological parameters occurring at a finite time, thus potentially inducing physical singularities that jeopardize the regular structure of the spacetime. Moreover, in many situations, these divergences additionally signal the presence of pathologies in the perturbations. In fact, a phantom cosmological evolution supported by minimally coupled scalar fields necessarily leads to the appearance of laplacian and/or gradient instabilities in the perturbations. On the other hand, the so-called {\it Big Rip}, which is one of the most extensively studied future singularities, is characterized by a divergent scale factor at a finite time, called the {\it Rip} time, and this makes every binding structure in the universe eventually break apart. Other future singularities analyzed in the literature lead to regular spacetimes in the sense of geodesic completeness, but still may give rise to arbitrarily large tidal forces so that the passage of physical structures through the singularity is not free of peril. It is worth mentioning that the standard lore relies on quantum effects to tame all these divergences, since quantum corrections should become important when the Planck scale is reached \cite{Nojiri:2004ip}.
In addition to the theoretical appealing of these scenarios, it is interesting to note that an effective EoS for dark energy $w<-1$ is allowed by observational data (and, from some sources, even favoured \cite{Lazkoz:2006gp}) and, thus, a universe with a future singularity might be a plausible scenario for our universe. In this sense, the classification provided in Ref.~\cite{Nojiri:2005sx} and updated in Ref.~\cite{Fernandez-Jambrina:2014sga}, shows how each divergence, usually appearing in the scale factor and its derivatives, affects the universe expansion and its structures. Moreover, there are some alternative non-singular scenarios (understood as the absence of divergences in finite time) that may lead to the break of some structures, as the {\it Little Rip} \cite{Frampton:2011sp}, {\it Pseudo-Rip} \cite{PseudoRip} and {\it Little Sibling} \cite{Mariam}.
In this paper, we present an analysis of future singularities within the framework of interacting dark energy-dark matter models. The appearance of different singular cosmological scenarios in interacting models with variable cosmological constant and exotic quintessence fields has been analyzed in \cite{Chimento}. Here we will describe the interactions in an effective way through the continuity equations and we find that every singularity found so far in the literature can be mapped into a singularity of the interaction term $Q$, and so we dub it {\it $Q$-singularity}. This means that the flow of energy between the dark matter and dark energy components diverges at a finite time, inducing a divergence in the scale factor and/or its derivatives and leading to one of the future singularities analyzed in the literature. Furthermore, a new future singularity is analyzed, where the $Q$-singularity provides a divergence in the derivative of the EoS for dark energy. We then study some specific interacting terms and EoS for dark energy, analyzing those cases where the interaction diverges.
The paper is organized as follows: in section \ref{sec:II} a review on cosmological singularities is given. Section \ref{sec:III} is devoted to the introduction of the $Q$-singularities, where several ansatz are considered and some models reconstructed. In section \ref{sec:IV} some interacting terms, previously analyzed in the literature, are considered, where for some particular EoS for dark energy a $Q$-singularity occurs. Finally, section \ref{sec:V} contains the conclusions of the paper.
\section{Future singularities in Cosmology} \label{sec:II}
Assuming a homogeneous and isotropic universe at large scales, in compliance with the cosmological principle, the line element is given by Friedman-Lema\^itre-Robertson-Walker (FLRW) metric
\begin{equation} \label{eq:FLRW}
ds^2=-dt^2+a(t)^2\left(dx^2+dy^2+dz^2\right),
\end{equation}
where $a(t)$ is the scale factor and we have assumed spatially flat sections. For the matter sector we shall consider a perfect fluid, whose energy-momentum tensor is given by
\begin{equation} \label{eq:pf}
T_{\mu\nu}=(\rho+p) u_{\mu} u_{\nu} + p g_{\mu\nu}\ ,
\end{equation}
where $u_{\mu}$ is a normalized timelike vector, $u_{\mu}u^{\mu}=-1$, and $\rho$ and $p$ are the energy density and pressure of the fluid, respectively. Under these assumptions, the Einstein equations, $G_{\mu\nu}(g)=\kappa^2 T_{\mu\nu}$, where $\kappa^2=8\pi G$ with $G$ the Newton's constant, yield the equations
\begin{equation} \label{eq:Fried}
H^2=\frac{\kappa^2 }{3} \rho \quad{\rm and}\quad \dot{H}=-\frac{\kappa^2}{2}(\rho+p)\ .
\end{equation}
These equations determine the background evolution of the Hubble parameter, $H \equiv \frac{\dot{a}}{a}$ (a dot denotes a time derivative), once the matter content (\ref{eq:pf}) is fully specified. The above equations can be combined to obtain the continuity equation $\dot{\rho}+3H(\rho+p)=0$, which can be trivially solved for a constant EoS, $w \equiv \frac{p}{\rho}$, as $\rho(t) \propto a^{-3(1+w)}$. Inserting this result into the Friedman equations (\ref{eq:Fried}) yields the well known result
\begin{equation} \label{eq:Hasol}
H=\frac{2}{3(1+w)(t-t_s)} \Rightarrow a(t) \propto (t-t_s)^{\frac{2}{3(1+w)}} \ ,
\end{equation}
where $t_s$ is some constant. Then one just needs to provide a specific EoS $w$ and Eq.(\ref{eq:Hasol}) gives a full solution. This immediately brings forward a potential problem: if $w<-1$ the above solution leads to an expanding universe such that at $t=t_s$ both the scale factor $a(t)$ and the energy density $\rho$ diverge. This is indeed the tip of a broader problem, namely: depending on the properties of the matter under consideration, some cosmological quantities may diverge at a finite time $t=t_s$.
In a curved space-time the trouble with singularities is very subtle and a broad literature deals with this problem from different perspectives. In this sense, the most rigorous and well accepted criterion about the nature of spacetime singularities relies on the concept of \emph{geodesic completeness}, namely, whether any null and time-like geodesic can be extended to arbitrarily large values of its affine parameter or not \cite{Geodesics,Wald}. In this sense, a number of theorems have been established to determine the conditions upon which a given spacetime contains a singularity \cite{Theorems}. These theorems give a precise mathematical formulation to the physically intuitive idea that, being null geodesics attached to the transmission of information and time-like geodesics to the free falling of idealized physical observers, in a well behaved spacetime nothing should suddenly cease to exist or emerge from nowhere.
Let us review the analysis of the geodesic equation when assuming a flat FLRW spacetime (\ref{eq:FLRW}). A geodesic curve $\gamma^{\mu}=x^{\mu}(\lambda)$, where $\lambda$ is the affine parameter, satisfies, in a coordinate system, the following equation \cite{Chandra,Wald}
\begin{equation} \label{eq:geoeq}
\frac{{\rm d}^2 x^\mu}{{\rm d} \lambda^2}+\Gamma^\mu_{\alpha\beta}\frac{{\rm d} x^\alpha}{{\rm d} \lambda}\frac{{\rm d} x^\beta}{{\rm d} \lambda}=0 \ ,
\end{equation}
where $\Gamma^\mu_{\alpha\beta}$ are the Christoffel symbols of the spacetime metric $g_{\mu\nu}$. This way, the geodesic equations (\ref{eq:geoeq}) become
\begin{eqnarray}
\frac{{\rm d}^2 t}{{\rm d} \lambda^2}+Ha^2\delta_{ij}\frac{{\rm d} x^i}{{\rm d} \lambda}\frac{{\rm d} x^j}{{\rm d} \lambda}=0\, ,\label{tgeodesic}\\
\frac{{\rm d}^2 x^i}{{\rm d} \lambda^2}+2H\frac{{\rm d} x^i}{{\rm d} \lambda}\frac{{\rm d} t}{{\rm d} \lambda}=0 \label{xgeodesic}.
\end{eqnarray}
By using $H=\frac{\dot{a}}{a}=\frac{1}{a}\frac{{\rm d} a/{\rm d} \lambda}{{\rm d} t/{\rm d}\lambda}$, Eq. (\ref{xgeodesic}) can be rewritten as follows
\begin{equation}
\frac{{\rm d}}{{\rm d} \lambda}\left(a^2\frac{{\rm d} x^i}{{\rm d} \lambda}\right)=0 \ ,
\end{equation}
which gives
\begin{equation}
\frac{{\rm d} x^i}{{\rm d} \lambda}=\frac{u^i_0}{a^2} \ ,
\end{equation}
where $u^i_0$ are integration constants. Then, using the above result, Eq. (\ref{tgeodesic}) yields
\begin{equation}
\left(\frac{{\rm d} t}{{\rm d}\lambda}\right)^2=\frac{\vert\vec{u}_0\vert^2}{a^2}+C_0 \ ,
\end{equation}
where $C_0$ is another integration constant. We thus see that the geodesics will be regular (with a well defined tangent vector) as long as the scale factor remains regular. Hence, if the scale factor does not diverge and is non-vanishing (so the metric is regular), the 4-velocities of the geodesics remain regular and the spacetime will be said to be non-singular. If the scale factor diverges at some point, then the geodesics stop there and cannot go through it. As we have discussed above, it is important to notice that the geodesics are insensitive to divergences in the expansion rate $H$ or its derivatives if they do not correspond to a singular behavior of the scale factor. This will be the case of the types II, III and IV singularities below, where the scale factor remains finite while all the divergences only appear in its derivatives.
When any geodesic path cannot be indefinitely extended, one would be interested in understanding the underlying reason for that. Taking into account that in many spacetimes their geodesically incomplete character comes alongside the divergence of (some) curvature scalars, one might blame the presence of infinitely large tidal forces for the existence of incomplete paths. Therefore, a framework has been developed to determine the impact of tidal forces upon physical (extended) observers \cite{TK}, establishing the criteria of \emph{strong} singularities if the body is unavoidable destroyed as it crosses the divergent region, and \emph{weak} in case it could retain its identity, i.e., its finite extended nature. There are two broadly used criteria (known as Tipler and Krolak criteria) to classify singularities as weak or strong according to the convergence of the following integrals:
\begin{eqnarray}
T(u)&\equiv&\int_0^\lambda {\rm d}\lambda' \int_0^{\lambda'} {\rm d} \lambda''R_{ij}u^i u^j,\\
K(u)&\equiv&\int_0^\lambda {\rm d}{\lambda'} ''R_{ij}u^i u^j.
\end{eqnarray}
where $u^i$ is the 4-velocity of the geodesic towards the singularity and $R_{ij}$ the components of the Ricci tensor. From these expressions it is clear that a spacetime containing a divergence in (some of) the curvature scalars can still be regular according to the above criteria.
Equipped with the two tools described above (geodesic completeness and weak/strong singularities), a number of future singularities have been found and studied in detail in the literature:
\begin{itemize}
\item Type I (``Big Rip singularity"): For $t \rightarrow t_s$, $a \rightarrow \infty$, $\rho \rightarrow \infty$ and $ \vert p \vert \rightarrow \infty$. This case yields incomplete null and time-like geodesics \cite{TI}. Thus it represents a genuine space-time singularity.
\item Type II (``Sudden singularity"): For $t \rightarrow t_s$, $a \rightarrow a_s$, $\rho \rightarrow \rho_s$ and $\vert p \vert \rightarrow \infty$. Geodesics are complete and observers are not necessarily crushed (weak singularity \cite{TII,barrowIV}).
\item Type III (``Big Freeze singularity"): For $t \rightarrow t_s$, $a \rightarrow a_s$, $\rho \rightarrow \infty$ and $ \vert p \vert \rightarrow \infty$. Geodesically complete solutions, which can be either strong or weak \cite{TIII}
\item Type IV (``Generalized Sudden singularity"): For $t \rightarrow t_s$, $a \rightarrow a_s$, $\rho \rightarrow \rho_s$ and $ \vert p \vert \rightarrow p_s$ but second and higher derivatives of the Hubble parameter $H$ diverge. Geodesics are complete and the singularity is weak \cite{barrowIV,TIV}.
\item Type V (``$w$-singularity"): For $t \rightarrow t_s$, $a \rightarrow \infty$, $\rho \rightarrow 0$ and $ \vert p \vert \rightarrow 0$ but the equation of state $w \rightarrow \infty$. These singularities are weak as well \cite{TV}.
\end{itemize}
This list summarizes the current knowledge on cosmological singularities in the literature, where the divergent quantities are identified. In this classification it is common to implicitly assume that the usual Friedman equations hold, so that the singularities originate from some exotic properties of the matter sector (e.g., violation of the NEC). However, it is worth pointing out that some scenarios might lead to the appearance of cosmological future singularities for non-exotic matter fields, e.g dust or radiation fluids. This is the case, for instance, in some Born-Infeld inspired theories \cite{singularBI}, $f(R)$ gravities \cite{Nojiri:2003ft} or modified gravity theories formulated in generalized Weyl geometries \cite{Jimenez:2016opp}.
\section{Q-singularities} \label{sec:III}
In this section we shall further specify the setup discussed in the previous section and consider that the matter sector comprises non-relativistic dark matter, with $w_{m}=0$, and dark energy, with $w_{DE} \equiv p_{DE}/\rho_{DE} \neq 0$ a certain function of time. These two components will be assumed to interact according to \cite{Kodama}
\begin{equation}
\nabla_{\mu} {T^{\mu(m)}}_{\nu}=Q_{\nu} \hspace{0.1cm};\hspace{0.1cm} \nabla_{\mu} {T^{\mu(DE)}}_{\nu}=-Q_{\nu} \ ,
\label{Inteq}
\end{equation}
where the $4$-vector $Q_{\nu}$ governs the stress-energy transfer between the two dark components. The choice of $Q_\nu$ will determine the specific model under consideration. A natural choice is to assume that this vector lies within the space spanned by the 4-velocities of dark matter and dark energy. While this is important at the level of the perturbations, the fact that dark energy and dark matter are usually assumed to share a common rest frame on large scales makes it irrelevant for the homogeneous evolution\footnote{In models of moving dark energy \cite{MDE}, the two dark components can have a relative motion even at the background level so that more general interactions could be envisioned. We will not consider these scenarios here and, in any case, if dark energy interacts with dark matter, they are expected to have a common rest frame at large scales.}. In other words, dark matter and dark energy have the same background 4-velocity $u_\nu$ and, thus, the interaction term simply becomes $Q_{\nu}=Q u_{\nu}$. With these considerations in mind, the Eqs.(\ref{Inteq}) read
\begin{eqnarray}
\dot{\rho}_m+3H\rho_m&=&Q(t)\ , \label{Eqbis1} \\
\dot{\rho}_{DE}+3H(1+w_{DE})\rho_{DE}&=&-Q(t) \label{Eqbis3} \ .
\end{eqnarray}
and which is nothing but the field equations of the matter sector. In these equations, $Q(t)$ accounts for the energy exchange rate between the two dark sectors, so that $Q(t)>0$ implies a transfer of energy from the dark matter sector to the dark energy one and the other way around for $Q(t)<0$. Although we have added an interaction between the two dark components, the modification is such that the gravitational field equations remain unchanged
\begin{eqnarray}
H^2&=&\frac{\kappa^2}{3}\left(\rho_m+\rho_{DE}\right)\label{FLRWeq2} \ , \\
\dot{H}&=&-\frac{\kappa^2}{2} \left[\rho_m+\rho_{DE}+p_{DE}\right] \label{FLRWeq3} \ .
\end{eqnarray}
These two equations are consistent with the Bianchi identities as well as with the modified continuity equations. We can combine the two gravitational equations to express the dark matter and dark energy densities in terms of the Hubble expansion rate and the dark energy equation of state as
\begin{eqnarray}
\rho_{DE}&=&-\frac{1}{w_{DE}\kappa^2} \left(3H^2 + 2\dot{H} \right) \\
\rho_{m}&=&\frac{1}{w_{DE}\kappa^2} \left[3(1+w_{DE})H^2 + 2\dot{H} \right].
\end{eqnarray}
We can now combine these equations with the matter ones (\ref{Eqbis1}), (\ref{Eqbis3}) so that we finally obtain an expression for $Q(t)$ as
\begin{eqnarray}\label{Eqbis2}
Q(t) &=& \frac{1}{\kappa^2 w_{DE}} \left[9(1+w_{DE})H^3+6(2+w_{DE})H\dot{H} \right. \nonumber \\
&+&2\ddot{H}- \left. \frac{\dot{w}_{DE}}{w_{DE}} \left( 3 H^2 + 2 \dot{H}\right) \right].
\end{eqnarray}
Hence, by assuming a particular cosmological evolution $H=H(t)$, the corresponding interacting term $Q(t)$ can be obtained.
In the above expression for $Q(t)$ we can see that a future singularity implying a divergence in $H$, $\dot{H}$ or $\ddot{H}$ will typically induce a divergence in $Q$. Moreover, also divergences in the equation of state will also lead to a divergent interaction. As a novel feature, we additionally find that a divergence in $\dot{w}_{DE}$ gives rise to a singular interaction as well. Since the background field equations do not involve the derivatives of the equation of state parameter, this type of divergences is expected to be harmless for the homogeneous evolution, but the perturbations might be sensitive to them since the adiabatic sound speed in a barotropic fluid is given by
\begin{equation}
c_s^2=\frac{\dot{p}}{\dot{\rho}}=w\left(1+\frac{\rho\dot{w}}{\dot{\rho}w}\right),
\end{equation}
and, therefore, a divergence in $\dot{w}_{DE}$ could induce a divergence in the sound speed of dark energy. Here we are mainly interested in how the different types of singularities can be mapped into a divergence of $Q$ so that we will not explore any further these potential physical implications, which, in addition, would require a full covariant formulation to study the perturbations.
We will end our general treatment by giving an alternative relation between the interaction term and the Hubble expansion rate that will complement Eq.(\ref{Eqbis2}). For that, we first note that the continuity equations (\ref{Eqbis1}) and (\ref{Eqbis3}) can be solved for the energy densities given arbitrary expansion rate, dark energy equation of state and interaction term. The solutions can then be expressed as
\begin{eqnarray}
\rho_m&=&\bar{\rho}_m\left(1+\int\frac{Q}{\bar{\rho}_m}{\rm d} t\right)\\
\rho_{DE}&=&\bar{\rho}_{DE}\left(1-\int\frac{Q}{\bar{\rho}_{DE}}{\rm d} t\right),
\end{eqnarray}
where $\bar{\rho}_m$ and $\bar{\rho}_{DE}$ denote the homogeneous standard solutions in non-interacting models, i.e.,
\begin{eqnarray}
\bar{\rho}_m&=&\rho_m^0a^{-3}\\
\bar{\rho}_{DE}&=&\rho_{DE}^0\int e^{-3\int H(1+w_{DE}){\rm d} t}{\rm d} t,
\end{eqnarray}
with $\rho_m^0$ and $\rho_{DE}^0$ integration constants. We can then insert these solutions into the Friedman equation (\ref{FLRWeq2}) to obtain:
\begin{equation}
H^2-\frac{\kappa^2}{3}\big(\bar{\rho}_m+\bar{\rho}_{DE}\big)=
\frac{\kappa^2}{3}\left(\bar{\rho}_m\int\frac{Q}{\bar{\rho}_m}{\rm d} t-\bar{\rho}_{DE}\int\frac{Q}{\bar{\rho}_{DE}}{\rm d} t\right).
\label{eqFbis1}
\end{equation}
Analogously, the second gravitational equation (\ref{FLRWeq3}) can be expressed as
\begin{eqnarray}
&\dot{H}&+\frac{\kappa^2}{2}\big[\bar{\rho}_m+(1+w_{DE})\bar{\rho}_{DE}\big]\label{eqFbis2}\\
&=&-\frac{\kappa^2}{2}\left[\bar{\rho}_m\int\frac{Q}{\bar{\rho}_m}{\rm d} t-(1+w_{DE})\bar{\rho}_{DE}\int\frac{Q}{\bar{\rho}_{DE}}{\rm d} t\right].\nonumber
\end{eqnarray}
In these expressions we have explicitly separated the usual gravitational equations in the absence of interactions (LHS) from the modifications coming from the interactions (RHS). We can easily see that a divergence in $Q$ can give rise to divergences in $H$ or $\dot{H}$, but one can have a singular interaction $Q$ while $H$ and $\dot{H}$ remain finite. This is so because $Q$ only enters the above expressions inside the integrals, which can improve the smoothness of $Q$. In order to obtain Eq. (\ref{Eqbis2}) we had to take derivatives of the equations and this can introduce additional divergences that might lack physical relevance (like the singularities originating from $\dot{w}$ that are expected to have effects only at the perturbations level). However, expressions (\ref{eqFbis1}) and (\ref{eqFbis2}) are obtained after solving the continuity equations and, thus, divergences appearing there have a more direct physical relevance.
In this work we are interested in knowing under which circumstances the interacting term yields future singularities on the cosmological background evolution. To this end we shall split our analysis into constant and time-dependent EoS for the dark energy component in the following.
\subsection{$w_{DE}=$constant}
This choice removes the $\dot{w}_{DE}$ contribution in Eq.(\ref{Eqbis2}). Let us parameterize the Hubble factor as
\begin{equation} \label{eq:paramH}
H(t)=A+\frac{2}{3t}+B(t_s-t)^{\alpha} \ ,
\end{equation}
where $A$ and $B$ are some constants and $\alpha \neq 0$ controls the type of singularity at the time $t_s$. Note that such parametrization is assumed in such a way that the matter dominated epoch is recovered asymptotically while at late-times the cosmological constant term should dominate together with the last term in (\ref{eq:paramH}), which denotes deviations from $\Lambda$CDM model. For the appropriate value of $\alpha$, the last term will eventually dominate when approaching the singularity, if it occurs. From the parametrization (\ref{eq:paramH}) one obtains the dominant term for the scale factor $a(t)$ as
\begin{equation}
a(t)=a_0t^{2/3} e^{\left[At+B(t_s-t)^{\alpha} \left(\frac{t-t_s}{1+\alpha} \right)\right] } \ ,
\end{equation}
if $\alpha \neq -1$ and
\begin{equation}
a(t)=a_0t^{2/3} \frac{e^{At}}{t_s-t} \ ,
\end{equation}
if $\alpha =-1$. In these expressions $a_0$ is an integration constant. We can now expand $Q(t)$ as given in (\ref{Eqbis2}) around $t=t_s$ with the parametrization (\ref{eq:paramH}) to obtain
\begin{equation}
Q(t) \simeq (t-t_s)^{\alpha-2} \ ,
\end{equation}
which diverges for $\alpha<2$. We will proceed by splitting the analysis in different subcases for $\alpha$ to identify the type of future singularity and its relation to the divergence of $Q(t)$.
\begin{itemize}
\item $\alpha=-1$: At $t=t_s$ one has
\begin{eqnarray}
a(t) &\simeq& a_0 \frac{t^{2/3} e^{At}}{t_s-t} \rightarrow \infty \\
H(t) &\simeq& \frac{B}{t_s-t} \rightarrow \infty \ .
\end{eqnarray}
According to the classification introduced in section \ref{sec:II}, the divergence of both the scale factor and the Hubble parameter is distinctive of a \emph{Big Rip} singularity, which prevents the completeness of geodesics. Note that it is the divergence of $H$ that induces the one of $Q(t)$ at $t=t_s$.
\item $-1<\alpha<0$: At $t=t_s$ one has
\begin{eqnarray}
a(t) &\simeq& a_0 t^{2/3} e^{At_s}=a_s \\
H(t) &\simeq& \frac{B}{(t_s-t)^{\vert x \vert}} \rightarrow \infty \ .
\end{eqnarray}
According to the classification introduced in Sec.\ref{sec:II}, this is a type III \emph{Big Freeze} singularity.
\item $0<\alpha<1$: At $t=t_s$ we have
\begin{eqnarray}
a(t) &\simeq& t^{2/3} e^{At}=a_s \\
H(t) &\simeq& A \ ,
\end{eqnarray}
and thus both are finite, corresponding to an asymptotically de Sitter space. However, the first derivative of the Hubble parameter, $\dot{H} \simeq \frac{B}{(t_s-t)^{\vert \alpha-1 \vert}}$, diverges at $t=t_s$, which is the term in (\ref{Eqbis2}) related to the divergence of $Q(t=t_s) \rightarrow \infty$. This is a \emph{Sudden} singularity.
\item $1<\alpha<2$: At $t=t_s$ one has
\begin{eqnarray}
a(t) &\simeq& t^{2/3}e^{At} =a_s \\
H(t) &\simeq& A \\
\dot{H} &\simeq& -\frac{2}{3t^2} \ ,
\end{eqnarray}
and thus all of them finite. However, second time derivatives of $H$ appearing in Eq.(\ref{Eqbis2}) do diverge, $\ddot{H} \simeq \frac{B\alpha (\alpha-1)}{(t_s-t)^{\vert \alpha \vert}} \rightarrow \infty$. This is a \emph{Generalized Sudden} singularity.
\end{itemize}
It is also worth pointing out that in those cases with $\alpha <-1$, at $t=t_s$ one has
\begin{eqnarray}
a(t) &\simeq& a_0 \frac{t^{2/3}}{(t_s-t)^{\vert 1+\alpha \vert}} e^{-\frac{B}{\vert 1 + \alpha \vert}} \rightarrow 0 \\
H(t) &\simeq& \frac{B}{(t_s-t)^{\vert \alpha \vert}} \rightarrow \infty \ ,
\end{eqnarray}
which is a Big Bang type singularity.
The analysis above shows that the four classes of future singularities introduced by Nojiri, Odintsov and Tsujikawa in Ref.\cite{Nojiri:2005sx} can be understood to be just particular cases of singularities of the function $Q(t)$ at $t=t_s$, where each type of divergence in the scale/Hubble factor or its derivatives comes from a different degree of divergence of $Q(t)$. In this sense, if $\alpha \geq 2$ the function $Q(t)$ is finite and no future singularity emerges.
\subsection{$w_{DE} \equiv w_{DE}(t)$}
Next we will extend our analysis to a non-constant dark energy EoS. Let us parameterize it around $t=t_s$ as
\begin{equation} \label{eq:omegaeos}
w_{DE}\simeq w_{s}+(t_s-t)^{\beta} \ ,
\end{equation}
with $w_{s}$ some constant. A glance at Eq.(\ref{Eqbis2}), in combination with Eq.(\ref{eq:omegaeos}), tells us that if the parameter $\beta<1$ the term $\dot{w}_{DE}/w_{DE}^2$ will diverge. There are indeed several cases to be analyzed separately:
\begin{itemize}
\item If $-1<\beta<0$ then we have
\begin{eqnarray}
w_{DE}(t) &\simeq& \frac{1}{(t_s-t)^{\vert \beta \vert}} \rightarrow \infty \\
Q(t) &\simeq& \frac{\beta}{(t_s-t)^{\vert 1-\beta \vert} \left(w_s+(t_s-t)^{\beta} \right)^2} \rightarrow \infty \ .
\end{eqnarray}
Thus in this case we have a divergence in the function $w_{DE}$, which induces the one of $Q(t)$. These are $w_{DE}$-singularities, which correspond to the type V singularities discussed in Sec.\ref{sec:II}.
\item If $0<\beta<1$ then at $t=t_s$ one has
\begin{eqnarray}
w_{DE}(t) &\rightarrow& w_s \\
\dot{w}_{DE} &\simeq& \frac{1}{(t_s-t)^{\vert \beta -1 \vert}} \rightarrow \infty \\
Q(t) &\rightarrow& \infty \ .
\end{eqnarray}
This is a $\dot{w}_{DE}$-singularity, i.e., it is this contribution in (\ref{Eqbis2}) the responsible of inducing a divergence in $Q(t)$ (while $w$ remains finite). As this scenario is not included into the five types of future singularities discussed in Sec.\ref{sec:II}, we call them \emph{type VI singularities}. Assuming finiteness of both $a(t)$ and $H(t)$ in this case (otherwise one would end up into one of the type I-IV singularities above), it is easy to see that the corresponding spacetimes are geodesically complete due to the finiteness of the geodesic equations (\ref{tgeodesic}) and (\ref{xgeodesic}) (recall the discussion of Sec.\ref{sec:II}) and, likewise their $w_{DE}$-singularities partners, these type-VI singularities are weak. As discussed above, although this type of singularity does not represent any divergence for the homogeneous cosmological evolution, it might induce divergences in the perturbations through the adiabatic sound speed of dark energy. Finally, it would remain to check if this type of singularity can happen for a realistic physical model.
\item If $\beta<-1$ then the term $\dot{w}_{DE}/w_{DE}^2$ yields a finite contribution to $Q(t)$ in Eq.(\ref{Eqbis2}). In such a case one would need to consider the behaviour of the other terms in Eq.(\ref{Eqbis2}), thus obtaining again the types I-IV of singularities.
\end{itemize}
From the analysis above it is clear that the known five types of future singularities (and the new type VI found here) can be seen just as particular cases of the divergence of the energy flow between the dark components, that is, {\it $Q$-singularities}. In the next section we shall review some particular interacting models and discuss their relation with $Q$-singularities.
\section{Dark energy/dark matter couplings revisited} \label{sec:IV}
In this section we will consider some specific models for interacting dark energy and dark matter where we can explicitly show the appearance of different types of future cosmological singularities induced by the interacting terms.
\subsection{$Q=\zeta H \rho_{DE}$} \label{sec:IVA}
Over the last years, couplings between dark energy and dark matter have been widely explored in the literature where the effective $Q$-term is taken to be proportional to the energy density, either the dark matter or the dark energy or a combination of both, i.e. (see e.g. \cite{Boehmer:2008av,Chimento:2007yt,Gavela,Salvatelli})
\begin{equation}
Q=\zeta H \rho\ ,
\label{4.1}
\end{equation}
where $\zeta$ is a constant to be determined by the observations and $\rho$ can be an arbitrary combination of the dark matter and dark energy energy densities. Then, let us explore the possibility of the existence of a $Q$-singularity when considering this type of interactions.
\begin{figure*}[htbp]
\centering
\includegraphics[width=9.5cm]{phasemapBigRip.eps}\hspace{0.5cm}
\includegraphics[width=6.5cm]{plotBigRip.eps}
\caption{In this figure we plot the phase map (left panel) and numerical solutions (right panels) of the equation (\ref{4.3}) corresponding to the interaction $Q=\zeta H\rho_{DE}$ and with the parameters set to $\zeta=-1$ and $w_{DE}=-0.98$. We also show the solution for a matter dominated universe (green line) and the asymptotic solution with $\dot{H}=\lambda^{-1} H^2$ (blue) discussed in the main text. The black dot denotes the unstable Minkowski critical point. The red trajectory corresponds to the solution with the scale factor and Hubble expansion rate depicted in the right panel (where we have normalized to their initial values). This solution transits from matter domination to dark energy domination eventually approaching a Big Rip singular solution even though dark energy is not phantom, thus confirming the analytical findings discussed in the main text.}\label{plotBigRip}
\end{figure*}
Firstly, we will focus on the case when the interacting term is proportional to the dark energy density $Q=\zeta H \rho_{DE}$. In such a case, by combining the FLRW equations (\ref{eq:Fried}), the interacting term $Q$ can be expressed as follows:
\begin{equation}
Q=\zeta H\frac{3H^2+2\dot{H}}{\tilde{\kappa}^2}\
\label{4.2}
\end{equation}
where $\tilde{\kappa}^2=-w_{DE}\kappa^2$. From this expression is clear that a possible $Q$-singularity can only induce a {\it Big Rip} or a {\it Sudden} singularity, depending on the magnitude that diverges in (\ref{4.2}), $H$ and/or $\dot{H}$. It is important to note that these are only potential singularities that might occur. In particular, Big Rip singularities involve a divergence of the scale factor as well, and this cannot be directly inferred from the expression for $Q$, but we need to look at the corresponding solutions. We can reduce the cosmological evolution within this scenario to a second order differential equation for the Hubble expansion rate by combining expressions (\ref{Eqbis2}) and (\ref{4.2}) to obtain:
\begin{eqnarray}
&&2\ddot{H}+9(1+w_{DE})H^3+6(2+w_{DE})H\dot{H} \nonumber \\
&-&\frac{\dot{w}_{DE}}{w_{DE}} \left( 3 H^2 + 2 \dot{H}\right)+\zeta H\left(3H^2+2\dot{H}\right)=0\ .
\label{4.3}
\end{eqnarray}
From this equation we can analyse the behaviour of $H(t)$ and, thus, study the presence of future singularities. The remaining cosmological quantities, i.e., the energy densities, are algebraically related to the solutions of the above equation and, therefore, the cosmological evolution is completely determined by (\ref{4.3}). For simplicity, from now on we will assume constant equation of state for dark energy $\dot{w}_{DE}=0$. It is easy to see that the only critical point in that case is the Minkowski solution with $H=0$, which is unstable (see Fig.\ref{plotBigRip}). Moreover, although the equation is non-linear, it is easy to obtain exact solutions by taking advantage of its time rescaling invariance. This motivates to look for solutions with $\dot{H}=\lambda^{-1} H^2 $ with $\lambda$ some dimensionless parameter. These solutions lead to the usual cosmological evolution given by
\begin{equation}
H=\frac{\lambda}{t_s-t}\quad\Rightarrow\quad a(t)\propto \vert t_s-t\vert^{-\lambda}\ ,
\end{equation}
with $t_s$ some reference time. When inserting this ansatz into (\ref{4.3}) we obtain the following equation for $\lambda$:
\begin{equation}
\big(3+2\lambda^{-1}\big)\Big[3(1+w_{DE})+\zeta+2\lambda^{-1}\Big]=0
\end{equation}
with two branches of solutions, namely $\lambda=-2/3$ and $\lambda=-2/(3(1+w_{DE})+\zeta)$. The first branch corresponds to a matter dominated universe (which is unstable in the presence of the dark energy component), while the second branch corresponds to a universe where either the dark energy or the interaction term dominates. We see that, in that branch, the effective equation of state is given by $w_{\rm eff}=w_{DE}+\zeta/3$. Thus, an interaction term with $\zeta<-3(1+w_{DE})$ (transfer from dark matter to dark energy) can induce an effective phantom behaviour leading to a type I Big Rip singularity even if dark energy satisfies the null energy condition. This can be easily understood from the dark energy conservation equation which, with the interacting term under consideration can be written as
\begin{equation}
\dot{\rho}_{DE}+3H\left(1+w_{DE}+\frac13 \zeta\right)\rho_{DE}=0
\end{equation}
where we see that dark energy acquires the aforementioned effective equation of state determining the fate of the cosmological evolution. In Fig. \ref{plotBigRip} we show the phase map corresponding to the cosmological evolution of this model and a particular singular solution where our analytical results are also numerically confirmed.
Although the above results have been obtained for an interaction fully determined by the dark energy component, our findings are completely general for the type of interactions under consideration. In order to show that, let us now consider the more general interaction term given by a linear combination of $\rho_{DE}$ and $\rho_m$, i.e., $Q=H\Big(\zeta_1 \rho_{DE}+\zeta_2\rho_m\Big)$. By proceeding analogously, we find the following expression for $Q$
\begin{equation}
Q=\frac{H}{\tilde{\kappa}^2}\Big[3\Big(\zeta_1-(1+w_{DE})\zeta_2\Big)H^2+2(\zeta_1-\zeta_2)\dot{H}\Big] \ .
\end{equation}
For this general case we will also have solutions of the form $\dot{H}=\lambda^{-1}H^2$, where $\lambda$ is now given by
\begin{align}
\lambda^{-1}_{\pm}=&-\frac14\Big[3(2+w_{DE})+\zeta_1-\zeta_2\nonumber\\
&\pm\sqrt{(3w_{DE}+\zeta_1)^2+\zeta_2(6w_{DE}-2\zeta_1+\zeta_2)}\Big]\ .
\end{align}
We can see from this expression that we will also have future Big Rip singularities very much like in the previous case analyzed in detail. In this case the effective equation of state is given by $w_{\rm eff}=-(1+\frac{2}{3\lambda})$ and has a more complicated dependence on the model parameters, but it is easy to see that it can also lead to an effective phantom behaviour even if dark energy satisfies the null energy condition. In the limit of small couplings $\vert\zeta_{1,2}\vert\ll 1$ we obtain
\begin{eqnarray}
w_{\rm eff}^+&\simeq& w_{DE}+\frac13\zeta_1\nonumber\\
w_{\rm eff}^-&\simeq& -\frac13\zeta_2 \ ,
\end{eqnarray}
for each corresponding branch. We see that the second branch can never lead to a Big Rip for small couplings $\vert\zeta_{1,2}\vert \ll 1$ because the effective equation of state is $w_{\rm eff}^-\sim{\mathcal O}(\zeta_2)$. The first branch however, can give a Big Rip singularity for a non-phantom dark energy component if its equation of state is close to -1 and the interaction satisfies $\zeta_1<-3(1+w_{DE})$. Thus, if the interactions are small, only a negative $\zeta_1$ can induce an effective phantom behaviour. However, if the interactions are allowed to be larger, also a wide range of values of $\zeta_2$ and positive values of $\zeta_1$ can lead to future Big Rip singularities. In Fig. \ref{generalinteraction} we show the region in the parameter space where there is an effective phantom behaviour in the general case.
\begin{figure}[!h]
\centering
\includegraphics[width=8cm]{Generalinteraction.eps}
\caption{In this plot we show the region in the parameter space where an interaction of the form $Q=H\Big(\zeta_1 \rho_{DE}+\zeta_2\rho_m\Big)$ can lead to a future Big Rip singularity characterized by an effective phantom behaviour even if dark energy satisfies the null energy condition (we have taken $w_{DE}=-0.98$).}\label{generalinteraction}
\end{figure}
\begin{figure*}[!ht]
\includegraphics[width=9.2cm]{phasemapSudden2.eps}\hspace{0.5cm}
\includegraphics[width=6.5cm]{plotSudden.eps}\\
\vspace{1cm}
\includegraphics[width=9.2cm]{phasemapFreeze.eps}\hspace{0.5cm}
\includegraphics[width=6.5cm]{plotFreeze.eps}
\caption{In this plot we show the cosmological evolution for the interacting model $Q=\Gamma \rho_{DE}^n$. We have taken $w_{DE}=-0.98$ and $\Gamma=-1$ to show the presence of future singularities for non-phantom dark energy. In the upper panels we have taken $n=3$, which corresponds to the case exhibiting a future Sudden singularity with the asymptotic solution $H\simeq C(t_s-t)^p+H_s$ as discussed in the main text. In the upper left panel we show the phase map and indicate the matter dominated universe (green line), the critical points corresponding to Minkowski (black) and de Sitter (blue), both of which are unstable, and two trajectories whose numerical solutions are also shown in the upper right panels. We can see that most solutions asymptotically give a finite value of $H$ while $\dot{H}$ diverges, going either to $+\infty$ (dashed curves) or $-\infty$ (solid curves). These numerical solutions are in agreement and confirm the analytical results discussed in the main text. Notice also that there are solutions that remain close to the matter dominated universe but eventually evolve towards the singularity. This shows that there can indeed be a transition from matter domination to dark energy domination with a future singularity. Similarly, in the lower left panel we show the phase map for the case with $n=5/3$ that corresponds to a future Big Freeze singularity where the Hubble expansion rate (and its derivatives) diverges while the scale factor remains finite, as can bee explicitly seen in the numerical solutions given in the lower right panels, again confirming our analytical results.}\label{phasemapSudden}
\end{figure*}
\subsection{$Q=\Gamma\rho_{DE}^n$}
In the previous subsection we have shown that the most widely used interaction terms in the literature can easily induce a future Big Rip singularity even if dark energy satisfies the null energy condition. In order to show the appearance of other types of singularities within our framework, we will consider a slightly modified version of the interaction term given by
\begin{equation}
Q=\Gamma\rho_{DE}^n \ ,
\end{equation}
with $n$ a dimensionless constant and $\Gamma$ a parameter with dimension $(\rm mass)^{5-4n}$ controlling the strength of the interaction term. Again, the interaction can be expressed in terms of the Hubble expansion rate as
\begin{equation}
Q=\Gamma\left(\frac{3H^2+2\dot{H}}{\tilde{\kappa}^2}\right)^n .
\end{equation}
With this expression we can again obtain a differential equation for $H$ that will determine the cosmological evolution given by
\begin{eqnarray}
2\ddot{H}+6(2+w_{DE})H\dot{H}+9(1+w_{DE})H^3\nonumber\\+\tilde{\kappa}^{2(1-n)}\Gamma\left(3H^2+2\dot{H}\right)^n=0 \ ,
\end{eqnarray}
where we have taken $\dot{w}_{DE}=0$ for simplicity again. It is interesting to notice that this system has de Sitter critical points (in addition to the Minkowski critical point) determined by
\begin{eqnarray}
H_{\rm dS}=\left[-\frac{9 (1+w_{DE})}{3^n \tilde{\kappa}^{1-n}\Gamma}\right]^\frac{1}{2n-3} \ ,
\end{eqnarray}
which exists for dark energy models with $w_{DE}\neq-1$. Since we are interested in obtaining additional future singularities (other than Big Rip) we will now look for solutions where $\vert\dot{H}\vert\gg H^2$. Furthermore, we seek for solutions driven by the interaction term so that we will also assume that $\vert H\vert\ll\vert \Gamma (\dot{H}/\tilde{\kappa}^2)^{n-1}\vert$ so that the above equation reduces to
\begin{equation}
\ddot{H}+\mu\dot{H}^n\simeq0 \ ,
\label{Eq:asymp}
\end{equation}
with
\begin{equation}
\mu\equiv \Gamma\left(\frac{\tilde{\kappa}^2}{2}\right)^{1-n} \ .
\end{equation}
Notice that the above equation is invariant under a constant shift of $H$ and this is important to keep $H$ finite. In fact, the above equation can be easily solved to give the following asymptotic solution of the original equation:
\begin{equation}
H\simeq C(t_s-t)^p+H_s \ ,
\end{equation}
with $t_s$ and $H_s$ integration constants and
\begin{equation}
p=\frac{n-2}{n-1}, \quad\quad C=\frac{(1-n)^p}{n-2}\mu^{p-1} \ .
\end{equation}
This shows that the considered interaction term can induce singularities where $H$ remains constant while its derivative diverges. For instance, if we take $n=3$ we find solutions of the form $H\simeq C\sqrt{t_s-t}+H_s$ which give $H(t_s)=H_s$ but $\dot{H}\to \infty$ as $t\to t_s$, i.e., a type II or Sudden Singularity. This behaviour will be general for values of $n$ leading to $0<p<1$. This type of solutions are explicitly shown in Fig. \ref{phasemapSudden} (upper panels), where we can indeed confirm the analytical asymptotic behaviour for the solutions. Furthermore, these interactions also allow to find solutions of a Big Freeze or type III singularities by imposing, for instance, $p=-1/2$, which is achieved for $n=5/3$. In that case $H$ diverges as $H\simeq C(t_s-t)^{-1/2}$ at the singularity (as well as its derivatives), but the scale factor approaches the singularity as $a\simeq a_s e^{2C\sqrt{t_s-t}}$ and, thus, it remains finite. This behaviour will be typical for values of $n$ giving $-1<p<0$. Our analytical findings can be confirmed in Fig. \ref{phasemapSudden} (upper panels) from the phase map and explicit numerical solutions.
\section{Conclusions} \label{sec:V}
In this work we have revisited the so-called interacting dark energy models, where a coupling to dark matter is assumed, and we have established a general relation between such interactions and future cosmological singularities. We have considered the usual interaction terms at the level of the continuity equations so that the total energy is automatically conserved but a flow between both dark components exists. Within this scenario we have found that every future cosmological singularity taking place at a finite time can be directly mapped into a singularity of the interaction term, which we have dubbed as $Q$-singularity. This means that the energy flow diverges at finite time, naturally inducing one of the future singularities studied in the literature so far. Furthermore, our framework has allowed us to find a novel type of singularity characterized by a divergence in the time-derivative of the equation of state parameter of dark energy. Although this singularity is expected to not be relevant for the background evolution, it might signal the presence of divergences in the sound speed of the perturbations. We have exemplified these relations by considering parameterizations of the Hubble expansion rate and the dark energy equation of state.
As specific realizations of our general framework, we have also investigated the potential occurrence of the future singularities and their relation to $Q$-singularities when assuming some specific interacting terms given in terms of the energy density of dark energy. For usual interacting terms proportional to the energy density of dark energy, we have shown that the interaction with dark matter can induce a future Big Rip singularity even if the dark energy component does not violate the null energy condition, i.e., the interacting term can induce an effective phantom behavior for dark energy even if $w_{DE}>-1$. We have also considered interactions given in terms of an arbitrary power of the dark energy density and we have found that this interaction lead to other types of singularities such as Sudden and Big Freeze singularities. We have rigorously shown this by performing an analytical and a phase map analysis, whose results have also been confirmed by means of numerical solutions.
As a main result of our study we can conclude that interacting dark energy models provide a promising and very suitable framework to study cosmologies with future singularities, since they all can be accommodated in appropriate interaction terms, i.e., as $Q$-singularities.
\section*{Acknowledgments}
J.B.J. acknowledges the financial support of A*MIDEX project (No.~ANR-11-IDEX-0001-02), funded by the Investissements dAvenir French Government program, managed by the French National Research Agency (ANR); MINECO (Spain) projects FIS2014-52837-P and Consolider-Ingenio MULTIDARK
CSD2009-00064. D.R.G. is funded by the Funda\c{c}\~ao para a Ci\^encia e a Tecnologia (FCT, Portugal) postdoctoral fellowship No.~SFRH/BPD/102958/2014. D.S.G. is funded by the Juan de la Cierva program (Spain) No.~IJCI-2014-21733 and FCT (Portugal) No.~SFRH/BPD/95939/2013. V.S. is funded by the Polish National Science Center Grant DEC-2012/06/A/ST2/00395. The authors also acknowledge support from the FCT research grant UID/FIS/04434/2013. This article is based upon work from COST Action CA15117, supported by COST (European Cooperation in Science and Technology).
|
\section{Introduction}
The Chain Pair Simplification problem under the discrete \frechet\ distance\ is an interesting problem dealing with how well two polygonal curves can be simplified in relation to each other with a minimum number of points retained. Under the Hausdorff distance, the problem is \textbf{NP}-complete~\cite{Bereg:2008:SPC}. Recently, CPS-3F was shown to be in \textbf{P} with an algorithm that runs in $\mathcal{O}(m^2n^2\min\{m,n\})$ where $m$ and $n$ are the number of nodes on the two polygonal chains \cite{Fan:2015:WADS}.
Here, we exploit a trick from \cite{Wylie:2013:TCBB} where there are only two possible paths to simplify the curve, and we then modify this to allow for arbitrarily long sequences of simplification. This can be thought of as a partition problem between the two chains, however, any actual mapping requires pseudopolynomial time. The reason for this research was merely to see if this could be done. It is not easy to create two curves with this property, but this document gives a method for building pairs of polygonal curves that can only be simplified in specified ways. Further, the paths can be arbitrarily long before the curves come together again, and as many of these paths can be linked together as desired. This may be useful in proving other properties about the problem, such as in \cite{Wylie:2013:TCBB}.
For background information on the problems, the discrete \frechet\ distance, etc., please refer to the references.
\section{Weighted Chain Pair Simplification} \label{sec:weighted}
We first cover a general version of CPS-3F introduced in \cite{Fan:2015:WADS}. The weighted CPS problem allows for arbitrary weights on the vertices, and we want to find a solution such that neither of the two simplified chains have vertices whose weights sum to greater than a given value. The problem is formally defined as follows.
\begin{definition}[Weighted Chain Pair Simplification] \hfill \\
\noindent \textbf{Instance:} Given a pair of 3D chains $A$ and $B$, with lengths $m$ and $n$,
respectively, an integer $K$, three real numbers $\delta_1,\delta_2,\delta_3>0$,
and a weight function $C:\{a_1,\dots, a_m,b_1,\dots, b_n\} \rightarrow \mathbb{R}^+$. \\
\textbf{Problem:} Does there exist a pair of chains $A'$,$B'$ where $C(A'),C(B') \leq K$,
such that the vertices of $A'$,$B'$ are from $A,B$ respectively, where
$d_F(A,A') \leq \delta_1$, $d_F(B,B') \leq \delta_2$, and $d_F(A',B') \leq \delta_3$?
\end{definition}
\begin{theorem}
The weighted chain pair simplification problem under the discrete \frechet\ distance\ is weakly \textbf{NP}-complete \cite{Fan:2015:WADS}.
\end{theorem}
\begin{figure}[t!]
\centerline{\includegraphics[width=.2\textwidth]{images/general_reduction.pdf}}
\caption{The reduction for the weighted CPS-3F problem.}
\label{fig:wcps}
\end{figure}
The idea for this proof is a reduction from the Set Partition problem as seen in Figure \ref{fig:wcps}. Given a set of positive integers $S=\{s_1,\dots,s_n\}$, find two sets $P_1,P_2 \subset S$ such that $P_1 \cap P_2 = \emptyset$, $P_1 \cup P_2 = S$, and the sum of the numbers in $P_1$ equals the sum of the numbers in $P_2$. This is a weakly \textbf{NP}-complete\ special case of the classic subset-sum problem.
The reduction builds two curves with weights reflecting the values in $S$. We think of the two curves as the subsets of the partition of $S$. Although the problem requires positive weights, we also allow zero weights in our reduction for clarity. It is easy to modify the reduction to have positive weights and reside in a single dimension. Increasing all the weights by one gives $w(a_{2i-1})=w(b_{2i})=s_i+1$ and $w(a_{2i})=w(b_{2i-1})=1$. This adds one for every element of $S$ to both chains, so with $K=|S| + \mathfrak{S}/2$ the reduction is the same.
\section{Chain Pair Simplification (CPS-3F)} \label{sec:cps}
We now turn our attention to CPS-3F, which is the special case of the general weighted CPS problem where every vertex has a weight equal to one. The construction is similar to the weighted version but more involved. The formal definition of the problem is as follows.
\begin{definition}[Chain Pair Simplification]\hfill \\
\textbf{Instance:} Given a pair of polygonal chains $A$ and $B$, with lengths $O(m),O(n)$
respectively, an integer $K$, and three real numbers $\delta_1,\delta_2,\delta_3>0$.\\
\textbf{Problem:} Does there exist a pair of chains $A'$,$B'$ each of at most $K$ vertices
such that the vertices of $A'$,$B'$ are from $A,B$, respectively, where
$d_F(A,A') \leq \delta_1$,$d_F(B,B') \leq \delta_2$, and $d_F(A',B') \leq \delta_3$?
\end{definition}
In order to explain the reduction, it is first necessary to give some notation.
Given two polygonal chains $A=\langle a_1,\dots,a_m\rangle$, and $B=\langle b_1,\dots,b_n\rangle$, and constraints $\delta_1,\delta_2,\delta_3 \in \mathbb{R}^+$,
we define rectangles which will be used for our construction. First, let $\mathcal{D}=\{(a_i,b_j) |$ $a_i \in A,$ $b_j \in B$ and $\mathit{d}(a_i,b_j) \leq \delta_3\}$. This is the set of all pairs of nodes between the two chains that are at a distance of at most $\delta_3$ from each other.
\begin{figure}[h!]
\centerline{\includegraphics[width=.58\textwidth]{images/convert_graph2.pdf}}
\caption{The rectangle $r_{i,j}$ constructed
from subchains of $A,B$ where $\mathit{d}(a_i,b_j)\leq \delta_3$. Here $S_A(a_i,\delta_1)$ contains
the vertices $a_{i-1}$ to $a_{i+2}$, and $S_B(b_j,\delta_2)$ contains
the vertices $b_{j-1}$ to $b_{j+1}$. Thus, $r_{i,j}$ is defined by the min and max node indices
in each subchain.}
\label{fig:convert}
\end{figure}
We define $S_X(x_i, \delta)$ as the maximal continuous subchain containing $x_i$ on the polygonal chain $X$ such that all the vertices on this subchain are contained in the sphere centered at $x_i$ and with radius $\delta$. Let $\min$ and $\max$ refer to the minimum or maximum indexed element within $S_X(x_i,\delta)$. Now let $r_{i,j}$ be the rectangle defined as $\langle \min(S_A(a_i,\delta_1)),$ $\max(S_A(a_i,\delta_1)),$ $\min(S_B(b_j,\delta_2)),$ $\max(S_B(b_j,\delta_2)) \rangle $ such that $(a_i,b_j) \in \mathcal{D}$. A rectangle $r_{i,j}$ covers all the cells in a grid that are analogous to vertices in $S_A(a_i,\delta_1) \cup S_B(b_j,\delta_2)$ as shown in Figure \ref{fig:convert}.
\begin{figure}[t!]
\begin{center}
\subfigure[]{\label{fig:gadgeta}\includegraphics[width=.38\textwidth,keepaspectratio]{images/gadget.png}}
\hspace*{.5cm}
\subfigure[]{\label{fig:gadgetrects}\includegraphics[width=.35\textwidth,keepaspectratio]{images/cps_gadget.pdf}} \\
\end{center}
\caption{(a) The chains listed in Table \ref{tab:nodes}. The dotted lines connect pairs of vertices $(a_i,b_j)$ between the two
chains where $\mathit{d}(a_i,b_j) \leq \delta_3$. (b) The relationship in rectangles with the dots representing
the pairs between the chains where $\mathit{d}(a_i,b_j) \leq \delta_3 = 0.9$ which create the rectangles when $\delta_1=\delta_2=1.1$.}
\label{fig:cps_gadget}
\end{figure}
Marking just the pairs of nodes in the grid $A \times B$ that are within $\delta_3$ of each other (the pairs in $\mathcal{D}$) is equivalent to a discrete version of the often-used free-space diagram for the continuous Fr\'echet\ distance \cite{Alt:1995:JCOMPS}. The discrete version was similarly used in \cite{Avraham:2013:CORR}. Here, we have extended this idea to incorporate the possibility of simplifying $A$ and $B$ as well.
\begin{table}[h]
\renewcommand{\arraystretch}{1.0}
\centering
\begin{tabular}{|c | c | c | c |}
\hline
$A$ & ($x, y, z$) & $B$ & ($x, y, z$) \\ \hline
$a_1$ & (0.70, 0.00, 0.60) & $b_1$ & (0.70, 0.00, 1.40) \\
$a_2$ & (1.00, 0.00, 0.00) & $b_2$ & (1.00, 0.00, 2.00) \\
$a_3$ & (1.28, $-$0.56, 0.90) & $b_3$ & (0.96, 0.88, 1.55) \\
$a_4$ & (1.78, $-$0.49, 1.20) & $b_4$ & (1.64, 0.49, 0.80) \\
$a_5$ & (1.60, $-$0.60, 2.00) & $b_5$ & (1.57, 0.42, 0.10) \\
$a_6$ & (2.06, $-$0.35, 2.80) & $b_6$ & (1.92, 0.49, $-$0.70) \\
$a_7$ & (2.41, 0.00, 2.50) & $b_7$ & (2.41, 0.00, $-$0.50) \\
$a_8$ & (1.74, 0.46, 1.80) & $b_8$ & (1.78, $-$0.21, 0.20) \\
$a_9$ & (2.24, 0.46, 1.85) & $b_9$ & (2.13, $-$0.42, 1.20) \\
$a_{10}$& (2.70, 0.00, 0.00) & $b_{10}$ & (2.70, 0.00, 2.00) \\
$a_{11}$& (2.70, 0.00, 0.60) & $b_{11}$ & (2.70, 0.00, 1.50) \\
\hline
\end{tabular}
\caption{The nodes in the chains for $A$ and $B$.}
\label{tab:nodes}
\end{table}
Similar to WCPS-3F, we show a reduction from the set partition problem, but note that the reduction is pseudopolynomial, and exponential in the size of the input. Given the set $S=\{s_1,s_2,\dots, s_n\}$ where $s_i \in \mathbb{Z}^+$ $\forall$ $0 \leq i \leq n$, we want to find two sets $P_1 \subset S,P_2 \subset S$ such that $P_1 \cap P_2 = \emptyset$, $P_1 \cup P_2 = S$, and $\sum^{|P_1|}_{j=1}P_{1_j} = \sum^{|P_2|}_{j=1}P_{2_j}$.
The basis of the reduction is the relationship between the two chains listed in Table \ref{tab:nodes} and shown in Figure \ref{fig:gadgeta}. In the figure the two curves are shown with dotted lines between the pairs of vertices within $\delta_3=0.9$ of each other. Based on our definition for rectangles, these chains correspond to the ten rectangles shown in Figure \ref{fig:gadgetrects}. Based on $\delta_1 = \delta_2 = 1.1$ and $\delta_3 = 0.9$, the pairs of nodes between the $A,B$ are $(a_1, b_1)$, $(a_2, b_5)$, $(a_2, b_8)$, $(a_4, b_9)$, $(a_5, b_2)$, $(a_7, b_{10})$, $(a_8, b_2)$, $(a_9, b_4)$, $(a_{10}, b_7)$, and $(a_{11}, b_{11})$. Every node $h_i$ on either chain is also within $\delta_1=\delta_2=$1.1 of both its neighbors ($x_{i-1}, x_{i+1}$), but the distance is greater than 1.1 for the other nodes on the chain. The only exception is the first and last node which is within 1.1 of two nodes. Note that if $(a_1,b_1)$ and $(a_{11},b_{11})$ are moved such that they are not within $\delta_3$ of each other, then the configuration in Figure \ref{fig:gadgetrects} would be missing the top left and bottom right rectangles. These are only needed when beginning or ending a chain. This will be clear in the reduction.
The configuration in Figure \ref{fig:cps_gadget} works by having only two possible simplifications. Ignoring the starting and ending squares, one path yields chains (for $A',B'$) of lengths (3,4) while the other yields lengths of (4,3). We can think of this as both having 3 added to them, and then choosing which chain to add the extra node to. Now, if we stack them along the diagonal (again without the first and last rectangles). This would mean that the two gadgets would connect such that the only two paths would give simplified chains of (8,6) or (6,8). In this way, we can create chains such that only two possible simplifications exist with any chosen difference between the two.
In order for this to work, we need another chain to add on to the one defined in Table \ref{tab:nodes} such that the last three nodes in each chain are replaced by the first three nodes in the new chain, but the chain must also maintain all the relationships. Chains $A$ and $B$ were designed so that the beginning pair ($a_2, b_2$) align with the end pair ($a_{10}, b_{10}$) and each chain can be copied and ``attached'' to the end of the previous chain by a translation in the positive $x$ direction. By combining the new chains, $a_2$ replaces the node as $a_{10}$ and equivalently for $B$.
Let $H_{i\dots j}$ represent the consecutive subchain of a polygonal curve $H$ from node $h_i$ through $h_j$.
Then let $T^{(j,k)}$ denote the sequence $T$ translated by $1.7j$ units in the positive \textbf{X} direction and $10k$ in the positive \textbf{Z} direction, e.g., for all $t_i \in T^{(2,3)}$ the coordinates are $t_i = (x_i + 3.40,$ $y_i,$ $z_i + 30)$. The value $1.70$ is the distance between $a_2$ and $a_{10}$ and between $b_2$ and $b_{10}$, which are both aligned along the \textbf{X}-axis.
\begin{figure}[t!]
\begin{center}
\subfigure[]{\label{fig:gadget3a}\includegraphics[width=.5\textwidth,keepaspectratio]{images/gadget_3.pdf}}
\subfigure[]{\label{fig:gadget3b}\includegraphics[width=.47\textwidth,keepaspectratio]{images/cps_gadget_3.pdf}} \\
\end{center}
\caption{(a) A sequence of three repeated gadgets with the beginning and end nodes.
(b) The resulting rectangles from figure (a). There are only two paths
since the paths are monotonically non-decreasing. The possible simplified
chains are of lengths (9,12) or (12,9). This is equivalent to $s_i=3$ with 9 extra vertices
in each chain.}
\label{fig:cps_gadget3}
\end{figure}
Now constructing a new chain to represent $s_i=3$, we denote it as $X = \langle A^{(1,0)}_{1\dots 9},$ $A^{(2,0)}_{2\dots 9},$ $A^{(3,0)}_{2\dots 11} \rangle$, and $Y = \langle B^{(1,0)}_{1\dots 9},$ $B^{(2,0)}_{2\dots 9},$ $B^{(3,0)}_{2\dots 11} \rangle$.
The two polygonal chains are shown in Figure \ref{fig:gadget3a}, and the rectangle representation of their relationship is shown in Figure \ref{fig:gadget3b}. The only possible simplifications are highlighted in \ref{fig:gadget3b}, and they yield chains ($X',Y'$) of lengths (9,12) or (12,9).
The general reduction algorithm is as follows.
Given $S=\{s_1,s_2,\dots, s_n\}$ from the set partition problem.
For all $s_i$ where $s_i \in S$, create
$X_{s_i}=\langle A^{(1,i)}_{1}$, $A^{(1,i)}_{2\dots 9}$, $A^{(2,i)}_{2\dots 9}, \dots, A^{(s_{i}-1,i)}_{2\dots 9}, A^{(s_i,i)}_{2\dots 9}$, $A^{(s_i,i)}_{10}, A^{(s_i,i)}_{11}\rangle$.
You can think of each sequence from two to nine as a ``one''.
Thus, if $s_i=5$, you will need the starting node, five sequences from two to nine, and then the ending nodes.
Since ten and two represent the same node, ten is omitted except at the end when no more sequences are added.
We also create a $Y_{s_i}$ in the same manner using subsequences of $B$.
Finally, let $X=\langle X_{s_1}, \dots, X_{s_n} \rangle$ and $Y=\langle Y_{s_1}, \dots, Y_{s_n} \rangle$.
Now we can state that for an instance of the set partition problem with set $S$, $S$ can be evenly partitioned if and only if $X$ and $Y$
can be simplified via CPS-3F such that $\delta_1=\delta_2=1.1$, $\delta_3=0.9$, and $K= 2|S| + 3\mathfrak{S} + \mathfrak{S}/2$.
Here, $\mathfrak{S}$ is the sum of the elements of $S$. Note that since we create $t$ gadgets for each number $t$, the reduction is polynomial in the numeric value of the input, and exponential in the length of the input numbers to the set partition problem.
\section{Conclusion} \label{sec:conclusion}
In this paper we covered an interesting construction of two polygons as input for the CPS-3F problem and a pseudopolynomial time reduction from set partition to CPS-3F as an example of its use. This gadget may prove useful in constructing specific arrangements for other applications or may simply be an enjoyable excursion into part of the character of the chain pair simplification problem.
\nocite{Wylie:2013:TCBB}
\nocite{Fan:2015:WADS}
|
\section{Introduction}
Quantum models can be used to describe superfluids \cite{LoMo93}, quantum semiconductors \cite{FeZhou93}, weakly interacting Bose gases \cite{Grant73} and quantum trajectories of Bohmian mechanics \cite{Wyatt05}. They have attracted considerable attention in the last decades due, for example, to the development of nanotechnology applications.
In this paper, we consider the barotropic compressible quantum Navier-Stokes equations, which has been derived in \cite{BrMe2010}, under some assumptions, using a Chapman-Enskog expansion in Wigner equation. In particular, we are interested in the existence of global weak solutions together with the associated semi-classical limit. The quantum Navier-Stokes equation that we are considering read as:
\begin{equation}
\begin{split}
\label{eq_system}
&\rho_t+{\rm div}(\rho u)=0,
\\&(\rho u)_t+{\rm div}(\rho u\otimes u)+\nabla\rho^{\gamma}-2 {\rm div}(\sqrt{\nu\rho}\mathbb{S}_\nu+\sqrt{\kappa \rho}\mathbb{S}_\kappa)={\sqrt{\rho}} f + \sqrt{\kappa} {\rm div}({\sqrt{\rho}} \mathbb M),
\end{split}
\end{equation}
where
\begin{equation}\label{eq_formal}
\sqrt{\nu\rho}\mathbb{S}_\nu= \rho\mathbb{D} u,\qquad {\rm div}(\sqrt{\kappa \rho}\mathbb{S}_\kappa)=\kappa\rho\nabla\left(\frac{\Delta\sqrt{\rho}}{\sqrt{\rho}}\right),
\end{equation}
and with initial data
\begin{equation}
\label{initial data}
\rho(0,x)=\rho_0(x),\;\;\;(\rho u)(0,x)=(\rho_0 u_0)(x)\quad\text{ in } \O,
\end{equation}
where $\rho$ is the density, $\gamma>1$, $ u\otimes u$ is the matrix with components $ u_i u_j,$ $\mathbb{D} u=\frac{1}{2}\left(\nabla u+\nabla u^T\right)$ is the symetric part of the velocity gradient,
and $\O=\mathbb{T}^d$ is the $d-$dimensional torus, here $d=2$ or 3. The vector valued function $f$, and the matrix valued function $\mathbb M$ are source terms.
The relation (\ref{eq_formal}) between the stress tensors and the solution $(\sqrt{\rho}, {\sqrt{\rho}} u)$ will be proved in the following form. For the quantic part, it will be showed that
\begin{equation}\label{eq_quantic}
2\sqrt{\kappa\rho}\mathbb{S}_\kappa=2\kappa \left(\sqrt{\rho}\left(\nabla^2\sqrt{\rho}-4(\nabla\rho^{1/4}\otimes\nabla\rho^{1/4})\right)\right).
\end{equation}
For the viscous term, the matrix valued function $\mathbb{S}_\nu$ is the symmetric part of a matrix valued function ${\mathbb T}_\nu$, where
\begin{equation}\label{eq_viscous}
\sqrt{\nu\rho}{\mathbb T}_\nu=\nu\nabla(\rho u)-2\nu\sqrt{\rho} u\cdot\nabla\sqrt{\rho}.
\end{equation}
Whenever, $\rho$ is regular and away from zero, the quantic part of (\ref{eq_formal}) is equivalent to (\ref{eq_quantic}), and the matrix function ${\mathbb T}_\nu$ is formally $\sqrt{\nu\rho}\nabla u$. However, the a priori estimates do not allow to define $1/{\sqrt{\rho}}$ and $\nabla u$.
\vskip0.3cm
The energy of the system is given by
$$
E({\sqrt{\rho}},{\sqrt{\rho}} u) =\int_{\O} \left(\rho\frac{|u|^2}{2}+\frac{\rho^\gamma}{\gamma-1}+2\kappa|\nabla{\sqrt{\rho}}|^2\right)\,dx,
$$
with dissipation of entropy (in the case without source term)
$$
\mathcal{D}_E(\mathbb{S}_\nu)=2\int_{\O}|\mathbb{S}_\nu|^2\,dx,
$$
which is formally:
$$
2\nu\int_{\O}\rho|\mathbb{D}u|^2\,dx.
$$
In \cite{BrDe2003,BrDeLi2003}, Bresch and Desjardins introduced a new entropy of the system, now known as the BD entropy:
$$
\mathcal{E}_{BD}(\rho,u)=\int_\O \left(\rho \frac{|u+\nu\nabla\ln\rho|^2}{2}+\frac{\rho^\gamma}{\gamma-1}+2\kappa|\nabla{\sqrt{\rho}}|^2\right)\,dx,
$$
with associated dissipation (again without source term)
$$
\mathcal{D}_{BD} (\rho,u)=\nu \int_\O \left(\frac{4}{\gamma}|\nabla \rho^{\gamma/2}|^2+\kappa\rho|\nabla^2\ln\rho|^2+2\rho |\mathbb{A} u|^2\right)\,dx,
$$
where $\mathbb{A}$ is the antisymmetric part of the matrix $\nabla u$.
The function $\ln \rho$ is not controled by the a priori estimates. But we can use two other quantities, ${\mathcal E}_0$ and $\mathcal{D}_\mathcal{E}^0$, which are defined as follows:
\begin{eqnarray*}
&& {\mathcal E}_0({\sqrt{\rho}}, {\sqrt{\rho}} u)=\int_{\O}\left(\rho\frac{|u|^2}{2}+(2\kappa+4\nu^2)|\nabla{\sqrt{\rho}}|^2+\rho^\gamma\right)\,dx,\\
&&\mathcal{D}_\mathcal{E}^0({\sqrt{\rho}},{\sqrt{\rho}} u)=\int_{\O}\left(\nu |\nabla \rho^{\gamma/2}|^2+\nu \kappa\left(|\nabla\rho^{1/4}|^4+|\nabla^2{\sqrt{\rho}}|^2\right)+|{\mathbb T}_\nu|^2 \right)\,dx.
\end{eqnarray*}
From J\"ungel \cite{Jue10} the functions ${\mathcal E}_0$ and $\mathcal{D}_\mathcal{E}^0$, are equivalent, respectively, to $E({\sqrt{\rho}},{\sqrt{\rho}} u)+\mathcal{E}_{BD}({\sqrt{\rho}},{\sqrt{\rho}} u)$ and $\mathcal{D}_E({\sqrt{\rho}},{\sqrt{\rho}} u,\S)+\mathcal{D}_{BD}(\rho, u)$, whenever each term can be defined, and $\S$ is the symmetric part of ${\mathbb T}_\nu$ with ${\mathbb T}_\nu=\sqrt{\nu\rho}\nabla u$. Namely there exists a universal constant $C_*$ such that for any such $(\rho, u)$:
\begin{eqnarray*}
&&\frac{1}{C_*}(E({\sqrt{\rho}}, {\sqrt{\rho}} u)+\mathcal{E}_{BD}(\rho,u))\leq {\mathcal E}_0({\sqrt{\rho}},{\sqrt{\rho}} u)\leq C_*(E({\sqrt{\rho}}, {\sqrt{\rho}} u)+\mathcal{E}_{BD}(\rho,u)),\\
&&\frac{1}{C_*}(\mathcal{D}_E(\S)+\mathcal{D}_{BD}(\rho,u))\leq \mathcal{D}_\mathcal{E}^0({\sqrt{\rho}},{\sqrt{\rho}} u,{\mathbb T}_\nu)\leq C_*(De(\S)+\mathcal{D}_{BD}(\rho,u)).
\end{eqnarray*}
\vskip0.3cm
The aim of this paper is to construct weak solutions for the system (\ref{eq_system}) using the a priori estimates provided by the energy and BD entropy inequalities. The main idea is to introduce a slightly stronger notion of weak solution that we call renormalized solutions. They are defined in the following way. For any function ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$, there exists two measures $\overline{R}_{\varphi}, R_{{\varphi}}\in\mathcal{M}({\mathbb R}^+\times \O)$ such that the following is verified in the sense of distribution:
\begin{equation}\label{eq_renormalise}
\begin{split}
&\partial_t(\rho{\varphi}(u))+{\rm div}(\rho u {\varphi}(u))+{\varphi}'(u)\cdot\nabla\rho^\gamma\\
&\qquad\qquad -2{\rm div}({\sqrt{\rho}}{\varphi}'(u)(\sqrt{\nu}\S+\sqrt{\kappa}{\mathbb S}_\kappa))\\
&\qquad ={\sqrt{\rho}}{\varphi}'(u)\cdot f+\sqrt{\kappa}{\rm div}({\sqrt{\rho}}{\varphi}'(u)\mathbb M)+R_{\varphi},
\end{split}
\end{equation}
with ${\mathbb S}_\kappa$ verifies (\ref{eq_quantic}), and $\S$ is the symmetric part of ${\mathbb T}_\nu$ such that for every $i,j,k$ between 1 and $d$:
\begin{equation}\label{eq_viscous_renormalise}
\sqrt{\nu\rho}{\varphi}_i'(u)[{\mathbb T}_\nu]_{jk}=\nu \partial_j(\rho{\varphi}'_i(u)u_k)-2\In{\nu}{\sqrt{\rho}} u_k{\varphi}'_i(u)\partial_j{\sqrt{\rho}}+ \overline{R}_{\varphi},
\end{equation}
and
$$
\|R_{\varphi}\|_{\mathcal{M}({\mathbb R}^+\times\O)}+\|\overline{R}_{\varphi}\|_{\mathcal{M}({\mathbb R}^+\times\O)}\leq C\|{\varphi}''\|_{L^\infty}.
$$
The precise definition of weak solutions and renormalized weak solutions is laid out in definitions \ref{def_weak} and \ref{def_renormalise}.
Note that, taking a sequence of function ${\varphi}_n$ such that ${\varphi}_n(y)$ converges to $y_i$, but $\|{\varphi}''\|_{L^\infty}$ converges to 0, we can retrieve formally the equation (\ref{eq_system}). We will show that it is actually true that any renormalized weak solution is, indeed, a weak solution.
\vskip0.3cm
For every
$$
f\in \left[L^1\dom
\right]^d,\qquad \mathbb M\in\left[L^2{(\R^+;L^2(\O))}\right]^{d\times d},
$$
and every $(\sqrt{\rho_0}, \sqrt{\rho_0} u_0)$ such that $ {\mathcal E}_0(\sqrt{\rho_0}, \sqrt{\rho_0} u_0)$ is bounded,
we define
\begin{equation}
\Mc={\mathcal E}_0(\sqrt{\rho_0}, \sqrt{\rho_0} u)+\|f\|_{L^1{(\R^+;L^2(\O))}}+\|\mathbb M\|_{L^2{(\R^+;L^2(\O))}}.
\end{equation}
The main theorem of the paper is the following.
\begin{theo} \label{theo_main}
\begin{enumerate}
\item There exists a universal constant $\overline {C}_*>0$ such that the following is true.
Let $\sqrt{\rho_0}, \sqrt{\rho_0} u_0, f, \mathbb M$ be such that $\Mc$ is bounded.
Then, for any $\kappa\geq 0$, there exists a renormalized solution $({\sqrt{\rho}},{\sqrt{\rho}} u)$ of (\ref{eq_system}) with
\begin{eqnarray*}
&&{\mathcal E}_0({\sqrt{\rho}}(t), {\sqrt{\rho}}(t) u(t))\leq \Mc,\qquad\mathrm{for}\ \ t>0,\\
&& \int_0^\infty\mathcal{D}_\mathcal{E}^0({\sqrt{\rho}}(t), {\sqrt{\rho}}(t) u(t))\,dt\leq \Mc.
\end{eqnarray*}
Moreover, for every ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$,
\begin{eqnarray*}
&&\qquad \|R_{\varphi}\|_{\mathcal{M}({\mathbb R}^+\times\O)}+\|\overline{R}_{\varphi}\|_{\mathcal{M}({\mathbb R}^+\times\O)}\leq \overline {C}_* \|{\varphi}''\|_{L^\infty}\Mc.
\end{eqnarray*}
Moreover, $\rho\in C^0({\mathbb R}^+;L^p(\O))$ for $1\leq p<\sup(3,\gamma)$, and $\rho u\in C^0({\mathbb R}^+; L^{3/2}(\O)-weak)\cap C^0({\mathbb R}^+; L^{\frac{2\gamma}{\gamma+1}}(\O)-weak)$.
\item Any renormalized solution of \eqref{eq_system} is a weak solution of \eqref{eq_system} with the same initial value.
\item Consider any sequences $\kappa_n\geq0$, converging to $\kappa\geq0$, $\nu_n>0$ converging to $\nu>0$, $(\sqrt{\rho_{0,n}},\sqrt{\rho_{0,n}} u_{0,n}, f_n,\mathbb M_n)$ such that $M_0(\sqrt{\rho_{0,n}},\sqrt{\rho_{0,n}}u_{0,n} ,f_n,\M_n)$ is uniformly bounded, and an associated weak renormalized solution $(\sqrt{\rho_n}, \sqrt{\rho_n} u_n)$ to \eqref{eq_system}. Then, there exists a subsequence (still denoted with $n$), and $({\sqrt{\rho}}, {\sqrt{\rho}} u)$ renormalized solution to \eqref{eq_system} with initial value $(\sqrt{\rho_0}, \sqrt{\rho_0}u_0)$ and Planck constant $\kappa$, such that
$\rho_n$ converges to $\rho$ in $ C^0({\mathbb R}^+;L^p(\O))$ for $1\leq p<\sup(3,\gamma)$, and $\rho_n u_n$ converges to $\rho u$ in $C^0({\mathbb R}^+; L^{3/2}(\O)-weak)\cap C^0({\mathbb R}^+; L^{\frac{2\gamma}{\gamma+1}}(\O)-weak)$. The function ${{\mathbb T}_\nu}_{,n}$ converges weakly in $L^2({\mathbb R}^+\times\O)$ to ${\mathbb T}_\nu$. Moreover, for every function ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$, $\sqrt{\rho_n}{\varphi}(u_n)$ converges strongly in $L_{\mathrm{loc}}^p({\mathbb R}^+\times\O)$ to ${\sqrt{\rho}} {\varphi}(u)$ for $1\leq p< 6$.
\end{enumerate}
\end{theo}
Note that all the results hold for any values of $\kappa$, including the Navier-Stokes case $\kappa=0$. For $\kappa>0$ we can have from the a priori estimates, better controls on the solutions, and convergence in stronger norms. The stability part of the result includes the follwoing case of the semi-classical limits $0<\kappa_n\to 0$.
\begin{coro} The semi-classical limit.
Consider $(\sqrt{\rho_{0}},\sqrt{\rho_{0}} u_0, f,\mathbb M)$ such that the quantity $\Mc$ is bounded, and consider an associated weak renormalized solution $(\sqrt{\rho_\kappa}, \sqrt{\rho_\kappa} u_\kappa)$ to the quantum Navier-Stokes equations \eqref{eq_system} with $\kappa>0$.Then, there exists a subsequence (still denoted with $\kappa$), and $({\sqrt{\rho}}, {\sqrt{\rho}} u)$ renormalized solution to the Navier-Stokes equations (\eqref{eq_system} with $\kappa=0$) with same initial value such that
$\rho_\kappa$ converges to $\rho$ in $ C^0({\mathbb R}^+;L^p(\O))$ for $1\leq p<\sup(3,\gamma)$, and $\rho_\kappa u_\kappa$ converges to $\rho u$ in $C^0({\mathbb R}^+; L^{3/2}(\O)-weak)\cap C^0({\mathbb R}^+; L^{\frac{2\gamma}{\gamma+1}}(\O)-weak)$. The function ${{\mathbb T}_\nu}_\kappa$ converges weakly in $L^2({\mathbb R}^+\times\O)$ to ${\mathbb T}_\nu$. Moreover, for every function ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$, $\sqrt{\rho_\kappa}{\varphi}(u_\kappa)$ converges strongly in $L_{\mathrm{loc}}^p({\mathbb R}^+\times\O)$ to ${\sqrt{\rho}} {\varphi}(u)$ for $1\leq p< 6$.
\end{coro}
For this problem, the a priori estimates include control on the gradient of some density quantities. This provides compactness on both the density $\rho$ and the momentum $\rho u$. The difficulty is due to the fact that we have only a control of $\rho u^2$ in $L^\infty({\mathbb R}^+,L^1(\O))$.
This cannot prevent concentration phenomena in the construction of solutions in the term $\rho u\otimes u$. When $\kappa=0$, the same problem arises for the term ${\sqrt{\rho}} u\nabla{\sqrt{\rho}}$, since the only a priori estimates available on both ${\sqrt{\rho}} u$ and $\nabla {\sqrt{\rho}}$ are in $L^\infty({\mathbb R}^+,L^2(\O))$.
\vskip0.3cm
This problem can be avoided by the introduction of additional terms as drag forces or cold pressure, as proposed by Bresch and Desjardins \cite{BrDe2007}. In the case of drag forces, the system is
\begin{equation}
\begin{split}
\label{eq_system_r}
&\rho_t+{\rm div}(\rho u)=0,
\\&(\rho u)_t+{\rm div}(\rho u\otimes u)+\nabla\rho^{\gamma}-2 {\rm div}(\sqrt{\nu\rho}\mathbb{S}_\nu+\sqrt{\kappa \rho}\mathbb{S}_\kappa)\\
&\qquad\qquad\qquad=-r_0u-r_1\rho |u|^2u+{\sqrt{\rho}} f + \sqrt{\kappa} {\rm div}({\sqrt{\rho}} \mathbb M),
\end{split}
\end{equation}
still endowed with \eqref{eq_viscous} and \eqref{eq_quantic}. The solutions of this kind of augmented systems can be explicitly constructed via a Faedo-Galerkin method. This was first performed by Zatorska in \cite{zatorska2012} for the classical case ($\kappa=0$) with chemical reactions and where the drag forces were replaced by cold pressure terms as $\nabla (\rho^{-N})$, for $N$ big enough. In the quantum case solutions have been constructed in \cite{GiLaVi2015} in the case with cold pressure, and in \cite{VasseurYu2015} in the case of drag forces as \eqref{eq_system_r}.
\vskip0.3cm
When considering the system without additional terms (as drag forces or cold pressure), it has been shown in \cite{MelletVasseur2007} that, in the classical case $\kappa=0$, solutions of \eqref{eq_system} verify formally that
\begin{equation}\label{eq_log}
\int_{\O} \rho |u|^2\ln(1+|u|^2)\,dx
\end{equation}
is uniformly bounded in time, provided that the bound is valid at $t=0$. It was also shown that a sequence of solutions, such that this quantity is uniformly bounded at $t=0$ (together with the energy and BD entropy), will converge, up to a subsequence, to a solution of the Navier-Stokes equation.
\vskip0.3cm
The standard way to construct weak solutions of systems verifying the weak stability is to construct solutions to an approximated problem (as the Faedo Galerkin method) for which the a priori estimates are still uniformly valid. Usually, those solutions are smooth so that every formal computation is actually true. However, in this context, the approximated problem needs to be compatible with the usual energy, the BD entropy, and the additional mathematical inequality \eqref{eq_log}. Only recently, such an approximated problem has been found \cite{LiXin2015} in dimension two, and in dimension three with unphysical stress tensors. Moreover, the regularity of the associated solutions is limited (not $C^\infty$). For solutions constructed via a Faedo-Galerkin method, the energy and BD entropy estimates can be verified at the approximated level, since $u$ is an admissible test function. This is not the case for the mathematical inequality \eqref{eq_log}. In \cite{VasseurYu2015-2}, the construction of solutions to \eqref{eq_system} with $\kappa=0$ was obtained following a different strategy. It was shown that limits of solutions to \eqref{eq_system_r} when $\kappa$ converges to 0 verify \eqref{eq_log}, even if it is not verified for $\kappa>0$. The idea was to show that the quantity
$$
\int_\O \rho{\varphi}(|u|)\,dx
$$
are uniformly bounded for smooth and bounded functions ${\varphi}$. This allowed to recover \eqref{eq_log} for $\kappa=0$ thanks to a sequence of approximations ${\varphi}_n$ of the function $y\to y^2\ln(1+y^2)$.
\vskip0.3cm
Spririto and Antonelli showed in \cite{AntSpi2016} that, formally, an estimate on \eqref{eq_log} can still be obtained on solutions to \eqref{eq_system} for a range of $\kappa$ close (or bigger) than $\nu$. But this estimate cannot be used for the semi-classical limit.
\vskip0.3cm
The notion of renormalized solutions is inspired from \cite{VasseurYu2015-2}. However, this notion is not used to recover an estimate on \eqref{eq_log}, which is known to be not verified for some range of $\kappa$. The idea is that we can obtain the stability of renormalized solutions, since the notion avoids the problem of concentration. Consider, for instance, the term
$$
\rho_n u_n {\varphi}(u_n).
$$
Since $\rho_n$ and $\rho_n u_n$ are compact in $L^p$, for a $p>1$, we can show that, up to a subsequence, $\rho_n$ converges almost everywhere to a function $\rho$, and $u_n$ converges almost everywhere on $\{(t,x) \ | \ \ \rho(t,x)>0\}$ to a function $u$. Hence $\rho_n u_n {\varphi}(u_n)$ converges almost everywhere to $\rho u{\varphi}(u)$. The function ${\varphi}$ prevent concentration, and so $\rho_n u_n {\varphi}(u_n)$ converges strongly to $\rho u{\varphi}(u)$.
\vskip0.3cm
The challenge is then to show that the renormalized solutions, are indeed, weak solutions in the general sense (see Definition \ref{def_weak}).
It is obtained by considering a sequence of bounded functions ${\varphi}_n$, uniformly dominated by $y\to|y|$, and converging almost everywhere to $y \to y_i$, for a fixed direction $i$. This provides the momentum equation for $\rho u_i$ at the limit $n\to \infty$. The key point, is that, while performing this limit, the functions $\rho, u$ are fixed. Considering, for example, the term
$$
\rho u {\varphi}_n(u).
$$
The function ${\varphi}_n(u)$ converges almost everywhere to $u_i$. And, thanks to the Lebesgue's dominated convergence Theorem, $\rho u{\varphi}_n(u)$ converges in $L^1$ to $\rho u u_i$. Note that the boundedness of $\rho u u_i$ in $L^1$ is enough for this procedure. Choosing the sequence of ${\varphi}_n$ such that $\|{\varphi}_n''\|_{L^\infty}$ converges to 0, we show that the extra terms $R_{{\varphi}_n}$ and $\overline{R}_{{\varphi}_n}$ converge to 0 when $n$ converges to $\infty$.
\vskip0.3cm
The main difference with \cite{VasseurYu2015-2} is that we do not need to reconstruct the energy inequality nor the control on \eqref{eq_log} via the sequence of functions ${\varphi}_n$. Hence, we do not need an explicit form of the terms involving second derivatives of ${\varphi}$ in the definition of renormalized solutions. Those terms (for which we do not have stability) are dumped in the extra terms $R_{\varphi}$ and $\overline{R}_{\varphi}$.
\section{Preliminary results and main ideas}
We are first working on the System \eqref{eq_system_r} with drag forces. The definitions will be valid for all the range of parameter, $r_0\geq0, r_1\geq0, \kappa\geq0, \nu>0$. The energy and the BD entropy on solutions to \eqref{eq_system_r} provide controls on
\begin{eqnarray*}
&& {\mathcal E}_r({\sqrt{\rho}}, {\sqrt{\rho}} u)=\int_{\O}\left(\rho\frac{|u|^2}{2}+(2\kappa+4\nu^2)|\nabla{\sqrt{\rho}}|^2+\rho^\gamma+r_0(\rho-\ln\rho)\right)\,dx,\\
&&\mathcal{D}_\mathcal{E}^r({\sqrt{\rho}},{\sqrt{\rho}} u)=\\
&&\qquad\qquad\int_{\O}\left(\nu |\nabla \rho^{\gamma/2}|^2+\nu \kappa\left(|\nabla\rho^{1/4}|^4+|\nabla^2{\sqrt{\rho}}|^2\right)+|{\mathbb T}_\nu|^2 +r_0 |u|^2+r_1\rho |u|^4\right)\,dx.
\end{eqnarray*}
From these quantities, we can obtain the following a priori estimates. For the sake of completeness we show how to obtain them in the appendix.
\begin{equation}
\begin{split}
\label{eq_a_priori}
&{\sqrt{\rho}}\in L^\infty{(\R^+;L^2(\O))}, \qquad \nabla{\sqrt{\rho}}\in L^\infty{(\R^+;L^2(\O))}, \qquad \nabla \rho^{\gamma/2}\in L^2{(\R^+;L^2(\O))}\\
&{\sqrt{\rho}} u \in L^\infty{(\R^+;L^2(\O))}, \qquad {\mathbb T}_\nu\in L^2{(\R^+;L^2(\O))}, \qquad \sqrt{\kappa}\nabla^2{\sqrt{\rho}}\in L^2{(\R^+;L^2(\O))},\\
&\kappa^{1/4}\nabla \rho^{1/4}\in L^4{(\R^+;L^4(\O))}, \qquad r_1^{1/4}\rho^{1/4} u\in L^4{(\R^+;L^4(\O))},\\
&r_0^{1/2} u\in L^2{(\R^+;L^2(\O))}, \qquad r_0\ln\rho\in L^\infty{(\R^+;L^1(\O))}.
\end{split}
\end{equation}
Note that those a priori estimates are not sufficient to define $\nabla u$ as a function. The statement that ${\sqrt{\rho}} \nabla u$ is bounded in $L^2$ means that there exists a function ${\mathbb T}_\nu\in L^2{(\R^+;L^2(\O))}$ such that:
$$
\sqrt{\nu}{\sqrt{\rho}} {\mathbb T}_\nu={\rm div}(\rho u)-{\sqrt{\rho}} u\cdot \nabla{\sqrt{\rho}},
$$
which is, formally, $\rho\nabla u$.
The definition of weak solutions and renormalized weak solutions for the system \eqref{eq_system_r} are as follows.
\begin{defi}\label{def_weak}
We say that $({\sqrt{\rho}},{\sqrt{\rho}} u)$ is a weak solution to \eqref{eq_system_r}, if it verifies the a priori estimates \eqref{eq_a_priori}, and
for any function $\psi\in C^\infty_c(\R^+\times\O)$:
\begin{eqnarray*}
&&\int_0^\infty\int_\O \left(\rho \psi_t +\rho u\cdot \nabla\psi \right)dx\, dt=0,\\
&&\int_0^\infty\int_\O \left(\rho u \psi_t +(\rho u\otimes u-{2}\sqrt{\nu\rho} \S-2\sqrt{\kappa\rho}{\mathbb S}_\kappa-\sqrt{\kappa \rho} \mathbb M)\cdot \nabla\psi+{F} \psi \right)dx\, dt=0,
\end{eqnarray*}
with $\S$ the symmetric part of ${\mathbb T}_\nu$ verifying \eqref{eq_viscous}, ${\mathbb S}_\kappa$ verifying \eqref{eq_quantic}, and
\begin{equation}\label{eq_F}
F=-2\rho^{\gamma/2}\nabla\rho^{\gamma/2}-r_0 u-r_1\rho|u|^2u +{\sqrt{\rho}} f,
\end{equation}
and for any $\overline{\psi}\in C^\infty_c({\mathbb R})$:
\begin{eqnarray*}
&&\lim_{t\to0}\int_\O \rho(t,x)\overline{\psi}(x)\,dx=\int_\O \rho_0(x)\overline{\psi}(x)\,dx,\\
&&\lim_{t\to0}\int_\O \rho(t,x)u(t,x)\overline{\psi}(x)\,dx=\int_\O \rho_0(x)u_0(x)\overline{\psi}(x)\,dx.
\end{eqnarray*}
\end{defi}
\begin{defi}\label{def_renormalise}
We say that $({\sqrt{\rho}},{\sqrt{\rho}} u)$ is a renormalized weak solution to (\ref{eq_system_r}), if it verifies the a priori estimates (\ref{eq_a_priori}), and
for any function ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$, there exists two measures $R_{{\varphi}}, \overline{R}_{\varphi}\in \mathcal{M}({\mathbb R}^+\times\O)$, with
$$
\|R_{{\varphi}}\|_{ \mathcal{M}({\mathbb R}^+\times\O)}+ \|\overline{R}_{{\varphi}}\|_{ \mathcal{M}({\mathbb R}^+\times\O)}\leq C \|{\varphi}''\|_{L^\infty({\mathbb R})},
$$
where the constant $C$ depends only on the solution $({\sqrt{\rho}},{\sqrt{\rho}} u)$, and
for any function $\psi\in C^\infty_c(\R^+\times\O)$,
\begin{eqnarray*}
&&\int_0^\infty\int_\O \left(\rho \psi_t +\rho u\cdot \nabla\psi \right)dx\, dt=0,\\
&&\int_0^\infty\int_\O \left(\rho {\varphi}(u) \psi_t +\left(\rho {\varphi}(u) u-({2}\sqrt{\nu\rho} \S{+2}\sqrt{\kappa\rho}{\mathbb S}_\kappa{+}\sqrt{\kappa \rho} \mathbb M) {\varphi}'(u)\right)\cdot \nabla\psi\right.\\
&&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\left.+{F}\cdot {\varphi}'(u) \psi \right)dx\, dt=\left \langle R_{{\varphi}}, \psi\right\rangle,
\end{eqnarray*}
with $\S$ the symmetric part of ${\mathbb T}_\nu$ verifying \eqref{eq_viscous_renormalise}, ${\mathbb S}_\kappa$ verifying \eqref{eq_quantic}, {$f$ given by \eqref{eq_F},}
and for any $\overline{\psi}\in C^\infty_c({\mathbb R})$:
\begin{eqnarray*}
&&\lim_{t\to0}\int_\O \rho(t,x)\overline{\psi}(x)\,dx=\int_\O \rho_0(x)\overline{\psi}(x)\,dx,\\
&&\lim_{t\to0}\int_\O \rho(t,x)u(t,x)\overline{\psi}(x)\,dx=\int_\O \rho_0(x)u_0(x)\overline{\psi}(x)\,dx.
\end{eqnarray*}
\end{defi}
Let us define
$$
M_r({\sqrt{\rho}}_0,\sqrt{\rho_{0}} u_0, f,\M)={\mathcal E}_r(\sqrt{\rho_0}, \sqrt{\rho_0} u_0)+\|f\|_{L^1{(\R^+;L^2(\O))}}+\|\mathbb M\|_{L^2{(\R^+;L^2(\O))}}.
$$
The main theorem proved in this paper is the following.
\begin{theo} \label{theo_df}
\begin{enumerate}
\item There exists a universal constant $\overline {C}_*>0$ such that the following is true.
Let $\sqrt{\rho_0}, \sqrt{\rho_0} u_0, f, \mathbb M$ be such that $M_r({\sqrt{\rho}}_0,\sqrt{\rho_{0}} u_0, f,\M)$ is bounded.
Then, for any $\kappa\geq 0$, $r_0\geq0$, $r_1\geq0$, there exists a renormalized solution $({\sqrt{\rho}},{\sqrt{\rho}} u)$ of (\ref{eq_system_r}) with
\begin{eqnarray*}
&&{\mathcal E}_r({\sqrt{\rho}}(t), {\sqrt{\rho}}(t) u(t))\leq M_r({\sqrt{\rho}}_0,\sqrt{\rho_{0}} u_0, f,\M),\qquad\mathrm{for}\ \ t>0,\\
&& \int_0^\infty\mathcal{D}_\mathcal{E}^r({\sqrt{\rho}}(t), {\sqrt{\rho}}(t) u(t))\,dt\leq M_r({\sqrt{\rho}}_0,\sqrt{\rho_{0}} u_0, f,\M).
\end{eqnarray*}
Moreover, for every ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$,
\begin{eqnarray*}
&&\qquad \|R_{\varphi}\|_{\mathcal{M}({\mathbb R}^+\times\O)}+\|\overline{R}_{\varphi}\|_{\mathcal{M}({\mathbb R}^+\times\O)}\leq \overline {C}_* \|{\varphi}''\|_{L^\infty}\Mc.
\end{eqnarray*}
Moreover, $\rho\in C^0({\mathbb R}^+;L^p(\O))$ for $1\leq p<\sup(3,\gamma)$, and $\rho u\in C^0({\mathbb R}^+; L^{3/2}(\O)-weak)\cap C^0({\mathbb R}^+; L^{\frac{2\gamma}{\gamma+1}}(\O)-weak)$.
\item Any renormalized solution of \eqref{eq_system_r} is a weak solution of \eqref{eq_system_r} with the same initial value.
\item If $r_0>0$, $r_1>0$, and $\kappa>0$, then any weak solution to \eqref{eq_system_r} is also a renormalized solution to \eqref{eq_system_r} with the same initial value.
\item Consider any sequences $\kappa_n\geq0$, $r_{0,n}\geq0$, $r_{1,n}\geq0$, $\nu_n>0$, converging respectively to $\kappa\geq0$, $r_0\geq0$, $r_1\geq0$, and $\nu>0$, $(\sqrt{\rho_{0,n}}),\sqrt{\rho_{0,n}} u_{0,n}, f_n,\mathbb M_n)$ such that $M_{r_{n}}({\sqrt{\rho}}_{0,n},\sqrt{\rho_{0,n}} ,f_n,\M_n)$ is uniformly bounded, and an associated weak renormalized solution $(\sqrt{\rho_n}, \sqrt{\rho_n} u_n)$ to \eqref{eq_system_r}. Then, there exists a subsequence (still denoted with $n$), and $({\sqrt{\rho}}, {\sqrt{\rho}} u)$ renormalized solution to \eqref{eq_system_r} with initial value $(\sqrt{\rho_0}, \sqrt{\rho_0}u_0)$ Planck constant $\kappa$, and drag forces coefficients $r_0,r_1$ such that
$\rho_n$ converges to $\rho$ in $ C^0({\mathbb R}^+;L^p_\mathrm{loc}(\O))$ for $1\leq p<\sup(3,\gamma)$, and $\rho_n u_n$ converges to $\rho u$ in $C^0({\mathbb R}^+; L^{3/2}(\O)-weak)\cap C^0({\mathbb R}^+; L^{\frac{2\gamma}{\gamma+1}}(\O)-weak)$. The function ${{\mathbb T}_\nu}_{,n}$ converges weakly in $L^2({\mathbb R}^+\times\O)$ to ${\mathbb T}_\nu$. Moreover, for every function ${\varphi}\in W^{2,\infty}({\mathbb R}^d)$, $\sqrt{\rho_n}{\varphi}(u_n)$ converges strongly in $L_{\mathrm{loc}}^p({\mathbb R}^+\times\O)$ to ${\sqrt{\rho}} {\varphi}(u)$ for $1\leq p< 10\gamma/3$.
\end{enumerate}
\end{theo}
\begin{rem}
We can actually show in (1) of the previous theorem that there is one solution verifying
$$
\overline{R}_{{\varphi},i,k,j}=\sum_{l=1}^d {\sqrt{\rho}} u_j {\varphi}''_{i,l}(u){{\mathbb T}_\nu}_{k,l}.
$$
But this is not needed to show that a renormalized solution is a weak solution. We cannot do the same for the term $R_{\varphi}$.
\end{rem}
Note that Theorem \ref{theo_df} together with \cite{VasseurYu2015} implies Theorem \ref{theo_main}. Indeed, \cite{VasseurYu2015} provides the construction of weak solutions to \eqref{eq_system_r} with positive $r_0, r_1,\kappa$. Part (3) in Theorem \ref{theo_df} insures that this solution is actually a renormalized solution. Considering sequences $r_{0,n}\geq0$ and $r_{1,n}\geq0$ both converging to 0, part (4) of Theorem \ref{theo_df} provides at the limit a renormalized solution to \eqref{eq_system}.
\section{From weak solutions to renormalized solutions in the presence of drag forces}
This section is dedicated to the proof of part (3) of Theorem \ref{theo_df}. In the whole section we will assume that
$\kappa>0$, $r_0>0$, and $r_1>0$, and we will consider a fixed weak solution $({\sqrt{\rho}}, {\sqrt{\rho}} u)$ as in Definition \eqref{def_weak}. Let us define $\overline{g}_\varepsilon$ for any function $g$ as
$$
\overline{g}_\varepsilon(t,x)=\eta_\varepsilon\ast g(t,x), \qquad t>{\varepsilon},
$$
where
$$
\eta_{\varepsilon}(t,x)=\frac{1}{{\varepsilon}^{d+1}}\eta_1(t/{\varepsilon},x/{\varepsilon}),
$$
with $\eta_1$ a smooth nonnegative even function compactly supported in the space time ball of radius 1, and with integral equal to 1.
\vskip0.3cm
Formally, we can show that a weak solution is also a renormalized solution by multiplying the equation by ${\varphi}'(u)$. However, solutions of \eqref{eq_system_r} have a limited amount of regularity. This has to be performed carefully. Let us explain the difficulties. First let us focus on the term
$$
{\rm div} (\sqrt{\nu\rho}\S).
$$
One way to obtain the renormalized equation from the weak one, is to consider the family of test functions $\psi{\varphi}'(\overline{u}_\varepsilon)$. We need to pass to the limit in the expression
$$
\int_0^\infty\int_\O \psi\sqrt{\nu\rho}\S \nabla {\varphi}'(\overline{u}_{\varepsilon})\,dx\,dt.
$$
But we cannot pass into the limit for the term $\sqrt{\nu\rho} \nabla {\varphi}'(\overline{u}_{\varepsilon})=\sqrt{\nu\rho} {\varphi}''(\overline{u}_{\varepsilon})\nabla \overline{u}_{\varepsilon}$. Note that this term is different from $\overline{{\mathbb T}_\nu}_{\varepsilon}$. The problem is that $\nabla u$ is not bounded in any functional space.
\vskip0.3cm
An other difficulty is to obtain, in the sense of distribution, the equality
$$
{\varphi}'(u)\partial_t(\rho u)= \partial_t(\rho{\varphi}(u))+({\varphi}'(u)u-{\varphi}(u))\partial_t\rho.
$$
Indeed, we have absolutely no estimate available on $\partial_t u$. Following Di Perna and Lions, \cite{Lions1996}, this can be obtained using commutators estimates which requires more a priori estimates than can be formally intuited.
\vskip0.3cm
To solve these problems, we need to introduce a cut-off function in $\rho$, $\phi_m(\rho)$ where $\phi_m$ is defined for every $m>0$ as
\begin{equation}\label{eq_phirho}
\phi_m(y)=\left\{
\begin{array}{ll}
0,& \mathrm{for} \ \ 0\leq y\leq \dfrac{1}{2m},\\[0.3cm]
2 m y-1,& \mathrm{for} \ \ \dfrac{1}{2m}\leq y\leq \dfrac{1}{m},\\[0.3cm]
1, & \mathrm{for} \ \ \dfrac{1}{m}\leq y\leq m,\\[0.3cm]
2-y/m, & \mathrm{for} \ \ m\leq y\leq 2m,\\[0.3cm]
0, &\mathrm{for} \ \ y\geq 2m.
\end{array}
\right.
\end{equation}
We will now work on
\begin{equation}\label{def_vm}
v_m=\phi_m(\rho)u
\end{equation}
instead of $u$. Note that
$$
\nabla v_m =\frac{\phi_m(\rho)}{\sqrt{\nu\rho}}{\mathbb T}_\nu+4{\sqrt{\rho}}\phi'_m(\rho)\rho^{1/4} u \nabla \rho^{1/4}.
$$
For $m$ fixed, $\frac{\phi_m(\rho)}{\sqrt{\nu\rho}}$ and ${\sqrt{\rho}}\phi'_m(\rho)$ are bounded, and so $ \nabla v_m$ is bounded in $L^2$, thanks to the a priori estimates obtained from $\kappa>0$ and $r_1>0$ in \eqref{eq_a_priori}.
\vskip0.3cm
Obtaining the equation on $v_m$ is pretty standard, thanks to the extra regularity on the density provided by the quantum term, and the BD entropy. However, to highlight where are the difficulties, we will provide a complete proof.
\subsection{Preliminary lemmas}
In this subsection, we introduce two standard lemmas to clarify the issues. The second one use the commutator estimates of Di Perna and Lions \cite{Lions1996}.
\begin{lemm}\label{lemm_trivial}
Let $g\in L^p({\mathbb R}^+\times\O)$ and $h\in L^q({\mathbb R}^+\times\O)$ with $1/p+1/q=1$ and $H\in W^{1,\infty}({\mathbb R})$. We denote by $\partial$ a partial derivative with respect to one of the dimension (time or space). Then we have:
\begin{eqnarray*}
&&\int_0^\infty\int_\O \overline{g}_{\varepsilon} h \,dx\,dt=\int_0^\infty\int_\O g \overline{h}_{\varepsilon} \,dx\,dt,\\
&&\lim_{{\varepsilon}\to0}\int_0^\infty\int_\O \overline{g}_{\varepsilon} h \,dx\,dt=\int_0^\infty\int_\O g h \,dx\,dt,\\
&&\partial \overline{g}_{\varepsilon}=\overline{[\partial g]}_{\varepsilon},\\
&& \lim_{{\varepsilon}\to0}\|H(\overline{g}_{\varepsilon})-H(g)\|_{L^s_{\mathrm{loc}}({\mathbb R}^+\times\O)}=0,\qquad \mathrm{for \ any \ } 1\leq s<\infty,\\
&&\partial H(g)=H'(g)\partial g\in L^r({\mathbb R}^+\times\O) \qquad \mathrm{as \ long \ as \ } \ \partial g\in L^r({\mathbb R}^+\times\O).
\end{eqnarray*}
\end{lemm}
This lemma is very standard and then we omit the proof. We use also in the sequel the following lemma due to Lions (see \cite{Lions1996}).
\begin{lemm}\label{lemm_DPLions}
Let $\partial$ be a partial derivative in one direction (space or time).
Let $g, \partial g\in L^p({\mathbb R}^+\times\O),\,h\in L^{q}({\mathbb R}^+\times\O)$ with $1\leq p,q\leq \infty$, and $\frac{1}{p}+\frac{1}{q}\leq 1$. Then, we have
$$\|[\overline{\partial(gh)]}_{\varepsilon}-\partial(g\overline{h}_{\varepsilon})\|_{L^{r}({\mathbb R}^+\times\O)}\leq C\|\partial g\|_{L^{p}({\mathbb R}^+\times\O)}\|h\|_{L^{q}({\mathbb R}^+\times\O)}$$
for some constant $C\geq 0$ independent of $\varepsilon$, $g$ and $h$, and with $r$ given by $\frac{1}{r}=\frac{1}{p}+\frac{1}{q}.$ In addition,
$$\overline{[\partial(gh)]}_{\varepsilon}-\partial(g\overline{h}_{\varepsilon})\to0\;\;\text{ in }\,L^{r}({\mathbb R}^+\times\O)$$
as $\varepsilon \to 0$ if $r<\infty.$
\end{lemm}
\subsection{Equation on $v_m$}
This subsection is dedicated to show that for any $\psi\in C^\infty_c(\R^+\times\O)$, and any $m>0$, we have
\begin{equation}\label{eq_phi_m}
\int_0^\infty\int_\O\left\{ \partial_t\psi\phi_m(\rho)-\psi \left(u\cdot\nabla\phi_m(\rho)+\phi'_m(\rho)\frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{Tr}({\mathbb T}_\nu)\right)\right\} \,dx\,dt=0.
\end{equation}
\begin{equation}\label{eq_vm}
\begin{split}
&\int_0^\infty\int_\O\left\{ \partial_t\psi\rho v_m+\nabla\psi\cdot\left(\rho u\otimes v_m -\phi_m(\rho){\sqrt{\rho}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right)\right)\right.\\
&\qquad \left.+\psi\left(-\frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{tr}({\mathbb T}_\nu)\phi'_m(\rho)\rho u -{\sqrt{\kappa\rho}}\mathbb M\nabla\phi_m(\rho)+\phi_m(\rho) F\right)\right\}\,dx\,dt=0,
\end{split}
\end{equation}
where $F$ is defined in \eqref{eq_F}, $v_m$ in \eqref{def_vm}, ${\mathbb S}_\kappa$ is in \eqref{eq_quantic}, and $\S$ is the symmetric part of ${\mathbb T}_\nu$ defined in\eqref{eq_viscous}.
\vskip0.3cm
We begin with a list of a priori estimates. Note that they depends on all the parameters $r_0>0$, $r_1>0$, $\kappa>0$, and $m$.
\begin{lemm}\label{lemm_estimates}
There exists a constant $C>0$ depending only on the fixed solution $({\sqrt{\rho}},{\sqrt{\rho}} u)$, and $C_m$ depending also on $m$ such that
\begin{eqnarray*}
&&\|\rho\|_{L^5({\mathbb R}^+\times\O)}+ \|\rho u\|_{L^{5/2}({\mathbb R}^+\times\O)}+ \|\rho |u|^2+{\sqrt{\rho}}(|\S|+|{\mathbb S}_\kappa|+|\mathbb M|+|f|)\|_{L^{5/3}({\mathbb R}^+\times\O)}\\
&&\qquad\qquad\qquad+ \|\rho^{\gamma/2}\nabla\rho^{\gamma/2} \|_{L^{5/4}({\mathbb R}^+\times\O)}+\|r_0 u\|_{L^2({\mathbb R}^+\times\O)}+\|r_1\rho|u|^2u\|_{L^{5/4}({\mathbb R}^+\times\O)}\leq C\\
&& \|\nabla\phi_m(\rho)\|_{L^4({\mathbb R}^+\times\O)}+ \|\partial_t\phi_m(\rho)\|_{L^2({\mathbb R}^+\times\O)}\leq C_m.
\end{eqnarray*}
\end{lemm}
\begin{proof}
From \eqref{eq_a_priori}, {${\sqrt{\rho}} \in L^\infty({\mathbb R}^+,L^2(\O))$}, $\nabla{\sqrt{\rho}}\in L^\infty({\mathbb R}^+,L^2(\O))$ and $\nabla^2{\sqrt{\rho}}\in L^2({\mathbb R}^+\times\O)$. Hence, using
{Gagliardo-Nirenberg inequality}, ${\sqrt{\rho}}\in L^\infty({\mathbb R}^+;L^6(\O))$, and $\nabla{\sqrt{\rho}}\in L^2({\mathbb R}^+;L^6(\O))$. Since $\nabla\rho=2{\sqrt{\rho}}\nabla{\sqrt{\rho}}$, $\nabla\rho\in L^2({\mathbb R}^+,L^3(\O))$. And
$\rho^2 \in L^\infty({\mathbb R}^+;L^{3/2}(\O))$, so $\nabla(\rho^2)=2\rho\nabla\rho\in L^2({\mathbb R}^+;L^{3/2}(\O))$. This gives
$$
\rho^2\in L^2({\mathbb R}^+;L^{3}(\O))\cup L^\infty({\mathbb R}^+;L^{3/2}(\O)).
$$
By interpolation, $\rho^2$ lies in all the space $L^p(L^q)$ with
$$
\frac{\alpha}{2}=\frac{1}{p},\qquad \frac{\alpha}{3}+\frac{2(1-\alpha)}{3}=\frac{1}{q},
$$
for $0\leq \alpha\leq 1$. For $\alpha= 4/5$, we obtain $\rho^2\in L^{5/2}({\mathbb R}^+\times\O)$.
We have
$$
\rho u=\rho^{3/4} \rho^{1/4} u.
$$
From the $r_1$ term of $\mathcal{D}_\mathcal{E}^r$, $\rho^{1/4} u\in L^4({\mathbb R}^+\times\O)$, and we have $\rho^{3/4}\in L^{20/3}({\mathbb R}^+\times\O)$. Hence $\rho u\in L^{5/2}({\mathbb R}^+\times\O)$.
The term $|\S|+|{\mathbb S}_\kappa|+|\mathbb M|+|f|+{\sqrt{\rho}} |u|^2\in L^2({\mathbb R}^+\times\O)$ and ${\sqrt{\rho}}\in L^{10}({\mathbb R}^+\times\O)$, so $\rho |u|^2+{\sqrt{\rho}}(|\S|+|{\mathbb S}_\kappa|+|\mathbb M|+|f|)\in L^{5/3}({\mathbb R}^+\times\O)$.
From the a priori estimates, we have $\nabla\rho^{\gamma/2}\in L^2({\mathbb R}^+\times\O)$, and $\rho^{\gamma/2}\in L^\infty({\mathbb R}^+,L^2(\O))$. Using
{Galgliardo-Nirenberg inequality} we have $\rho^{\gamma/2}\in L^{10/3}({\mathbb R}^+\times\O)$. So $\rho^{\gamma/2}\nabla\rho^{\gamma/2} \in L^{5/4}({\mathbb R}^+\times\O)$.
\vskip0.3cm
{The estimate on $r_0u$ comes directly from the a priori estimates. Since $\sqrt{\rho}|u|^2$ is in $L^2({\mathbb R}^+,L^2(\O))$ and $\rho^{1/4}u$ is in $L^4({\mathbb R}^+ \times \O)$, we have $\rho^{1/4}u\sqrt{\rho}|u|^2 \in L^{4/3}({\mathbb R}^+\times \O)$. Then using $\rho^{1/4} \in L^{20}({\mathbb R}^+ \times \O)$ we obtain the result for $r_1\rho|u|^2u$.}
\vskip0.3cm
From Lemma \ref{lemm_trivial}, $\nabla\phi_m(\rho)=4\rho^{3/4}\phi'_m(\rho)\nabla\rho^{1/4}$. But $y\to 4y^{3/4}\phi'_m(y)$ is bounded, and {by \eqref{eq_a_priori}} $\nabla\rho^{1/4}\in L^4({\mathbb R}^+\times\O)$, so $\nabla\phi_m(\rho)\in L^4({\mathbb R}^+\times\O)$. From Lemma \ref{lemm_trivial}, and \eqref{eq_viscous},
\begin{eqnarray*}
&&\partial_t \phi_m(\rho)=\phi'_m(\rho)\partial_t\rho=-\phi'_m(\rho){\rm div}(\rho u)=-\phi'_m(\rho)(\sqrt{\rho/\nu}\mathrm{Tr}({\mathbb T}_\nu)+2{\sqrt{\rho}} u\cdot\nabla{\sqrt{\rho}})\\
&&\qquad\qquad\qquad\qquad =-{\sqrt{\rho}}\phi'_m(\rho) (\sqrt{1/\nu}\mathrm{Tr}({\mathbb T}_\nu)+4\rho^{1/4} u\cdot\nabla\rho^{1/4}).
\end{eqnarray*}
Hence $\partial_t\phi_m(\rho)\in L^2({\mathbb R}^+\times\O)$.
\end{proof}
We use the first equation in Definition \ref{def_weak} with test function $\overline{[\phi'_m(\overline{\rho}_{\varepsilon})\psi]}_{\varepsilon}$.From Lemma \ref{lemm_trivial}, and \eqref{eq_viscous} we get the following.
\begin{eqnarray*}
&&0=\int_0^\infty\int_\O\left\{ \partial_t\overline{\psi\phi'_m(\overline{\rho}_{\varepsilon})}_{\varepsilon} \rho +\rho u\cdot\nabla\overline{\psi\phi'_m(\overline{\rho}_{\varepsilon})}_{\varepsilon}\right\} \,dx\,dt\\
&&=-\int_0^\infty\int_\O\left\{ \psi\phi'_m(\overline{\rho}_{\varepsilon}) \partial_t\overline{\rho}_{\varepsilon} +{\rm div}(\overline{\rho u}_{\varepsilon})\psi\phi'_m(\overline{\rho}_{\varepsilon})\right\} \,dx\,dt\\
&&=\int_0^\infty\int_\O\left\{ \partial_t\psi\phi_m(\overline{\rho}_{\varepsilon}) -\psi\phi'_m(\overline{\rho}_{\varepsilon}) \left[\overline{\frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{Tr}({\mathbb T}_\nu)}_{\varepsilon}+2 \overline{{\sqrt{\rho}} u \cdot \nabla{\sqrt{\rho}}}_{\varepsilon}\right]\right\} \,dx\,dt.
\end{eqnarray*}
We have $\rho \in L^5({\mathbb R}^+\times\O)$ and $\partial_t [\phi'_m(\rho)\psi]\in L^{5/4}({\mathbb R}^+\times\O)$, since $\partial_t\rho\in L^2({\mathbb R}^+\times\O)$ and $\psi$ is $C^1$ compactly supported. From Lemma \ref{lemm_estimates}, $\rho u\in L^{5/2}({\mathbb R}^+\times\O)$, and $\nabla[\phi_m'(\rho)\psi]\in L^{5/3}({\mathbb R}^+\times\O)$, since $\psi$ is regular and compactly supported, and $\nabla\phi_m'(\rho)\in L^{4}({\mathbb R}^+\times\O)$. So we can pass to the limit ${\varepsilon}\to0$ using Lemma \ref{lemm_trivial}.
We obtain:
$$
0=\int_0^\infty\int_\O\left\{ \partial_t\psi\phi_m(\rho) -\psi\phi'_m(\rho) \left[\frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{Tr}({\mathbb T}_\nu)+2 {\sqrt{\rho}} u\cdot\nabla{\sqrt{\rho}}\right]\right\} \,dx\,dt.
$$
Since $\psi\nabla\phi_m(\rho)$ lies in $L^4({\mathbb R}^+\times\O)$ and is compactly supported, and $u$ is in $L^2({\mathbb R}^+\times\O)$, using the Lemma \ref{lemm_trivial} we find
$$
0=\int_0^\infty\int_\O\left\{ \partial_t\psi\phi_m(\rho) -\psi\left[\phi'_m(\rho) \frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{Tr}({\mathbb T}_\nu)+ u\cdot\nabla\phi_m(\rho)\right]\right\} \,dx\,dt,
$$
{which is \eqref{eq_phi_m}.}
\vskip0.3cm
For ${\varepsilon}$ small enough, we consider the function $\overline{\psi\phi_m(\rho)}_{\varepsilon}$ as test function in the Definition \ref{def_weak}. In the same way than above, from Lemma \ref{lemm_estimates} and Lemma \ref{lemm_trivial}, passing into the limit in ${\varepsilon}$, we get
\begin{eqnarray*}
&&\int_0^\infty\int_\O\left\{ \partial_t\psi\rho v_m+\nabla\psi\cdot\left(\rho u\otimes v_m -\phi_m(\rho){\sqrt{\rho}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right)\right)\right.\\
&&\qquad \left.+\psi \left(\partial_t \phi_m(\rho)+u\cdot\nabla\phi_m(\rho) \right)\rho u
+\psi\left( -{\sqrt{\kappa\rho}}\mathbb M\nabla\phi_m(\rho)+\phi_m(\rho) F\right)\right\}\,dx\,dt=0,
\end{eqnarray*}
Using \eqref{eq_phi_m} {this gives \eqref{eq_vm}}.
\subsection{Equation of renormalized solutions}
We use the function $\overline{\psi{\varphi}'(\overline{v_m}_{\varepsilon})}_{\varepsilon}$ as a test function in \eqref{eq_vm}. Using Lemma \ref{lemm_trivial}, we find:
\begin{eqnarray*}
&&\qquad\int_0^\infty\int_\O \left(\partial_t\left[\overline{\psi{\varphi}'(\overline{v_m}_{\varepsilon})}_{\varepsilon}\right]\rho v_m+\nabla\left[\overline{\psi{\varphi}'(\overline{v_m}_{\varepsilon})}_{\varepsilon}\right]:(\rho u\otimes v_m )\right)\,dx\,dt\\
&&=-\int_0^\infty\int_\O \psi{\varphi}'(\overline{v_m}_{\varepsilon})\left(\partial_t\overline{[\rho v_m]}_{\varepsilon}+{\rm div}\overline{(\rho u\otimes v_m )}_{\varepsilon}\right)\,dx\,dt
\end{eqnarray*}
Thanks to Lemma \ref{lemm_estimates}, we can use Lemma \ref{lemm_DPLions}, with $g=\rho$ and $h=v_m$, and then $g=\rho u$ and $h=v_m$. Note that $v_m \in L^4({\mathbb R}^+\times\O)$. So, the expression above as the same limit when ${\varepsilon}$ goes to zero than
\begin{eqnarray*}
&&-\int_0^\infty\int_\O \psi{\varphi}'(\overline{v_m}_{\varepsilon})\left(\partial_t\left(\rho\overline{v_m}_{\varepsilon}\right)+{\rm div}\left(\rho u\overline{ v_m }_{\varepsilon}\right)\right)\,dx\,dt.
\end{eqnarray*}
Thanks to the first equation in Definition \ref{def_weak} this is equal to
\begin{eqnarray*}
&&\qquad-\int_0^\infty\int_\O \psi{\varphi}'(\overline{v_m}_{\varepsilon})\left(\rho\partial_t\overline{v_m}_{\varepsilon}+\rho u\cdot\nabla\overline{ v_m }_{\varepsilon}\right)\,dx\,dt\\
&&=-\int_0^\infty\int_\O \psi\left(\rho\partial_t{\varphi}(\overline{v_m}_{\varepsilon})+\rho u\cdot\nabla{\varphi}(\overline{v_m}_{\varepsilon})\right)\,dx\,dt\\
&&=\int_0^\infty\int_\O {\varphi}(\overline{v_m}_{\varepsilon})\left(\rho\partial_t\psi+\rho u\cdot\nabla\psi)\right)\,dx\,dt,
\end{eqnarray*}
which converges, when ${\varepsilon}$ goes to 0, to
\begin{equation}\label{vianet1}
\int_0^\infty\int_\O {\varphi}(v_m)\left(\rho\partial_t\psi+\rho u\cdot\nabla\psi)\right)\,dx\,dt.
\end{equation}
{Note that
\begin{eqnarray*}
\nabla v_m&=&\nabla\left(\frac{\phi_m(\rho)}{\rho}\rho u\right)\\
&=&\nabla\left(\frac{\phi_m(\rho)}{\rho}\right)\rho u+\frac{\phi_m(\rho)}{\rho}\nabla(\rho u)\\
&=&\nabla\left(\frac{\phi_m(\rho)}{\rho}\right)\rho u+\frac{\phi_m(\rho)}{\sqrt{\nu}{\sqrt{\rho}}}{\mathbb T}_\nu+2\frac{\phi_m(\rho)}{\rho}{\sqrt{\rho}} u\cdot\nabla{\sqrt{\rho}}\\
&=&4{\sqrt{\rho}} \phi'_m(\rho) \rho^{1/4}u\cdot\nabla\rho^{1/4}.
\end{eqnarray*}
Note that ${\sqrt{\rho}}\phi_m'(\rho)$ is bounded.
So, thanks to the third line of \eqref{eq_a_priori}, we have $\nabla v_m\in L^2({\mathbb R}^+\times\O)$, and so $\nabla {\varphi}''(v_m)\in L^2({\mathbb R}^+\times\O)$.
}
So, thanks to Lemma \ref{lemm_trivial} we can pass into the limit in the other terms and find
\begin{eqnarray}
&&\qquad\int_0^\infty\int_\O\left\{ -\nabla\left(\psi{\varphi}'(v_m)\right))\phi_m(\rho){\sqrt{\rho}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right)\right. \nonumber \\
&& \left.+\psi{\varphi}'(v_m)\left(-\frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{Tr}({\mathbb T}_\nu)\phi'_m(\rho)\rho u -{\sqrt{\kappa\rho}}\mathbb M\nabla\phi_m(\rho)+\phi_m(\rho) F\right)\right\}\,dx\,dt. \label{vianet2}
\end{eqnarray}
Putting \eqref{vianet1} and \eqref{vianet2} together gives
\begin{eqnarray*}
&&\int_0^\infty\int_\O \left\{{\varphi}(v_m)\left(\rho\partial_t\psi+\rho u\cdot\nabla\psi)\right)\right.\\
&&\qquad -\nabla\psi{\varphi}'(v_m)\phi_m(\rho){\sqrt{\rho}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right)\\
&&\qquad -\psi {\varphi}''(v_m)\nabla v_m\phi_m(\rho){\sqrt{\rho}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right)\\
&&\qquad \left.+\psi{\varphi}'(v_m)\left(-\frac{{\sqrt{\rho}}}{\sqrt\nu}\mathrm{Tr}({\mathbb T}_\nu)\phi'_m(\rho)\rho u -{\sqrt{\kappa\rho}}\mathbb M\nabla\phi_m(\rho)+\phi_m(\rho) F\right)\right\}\,dx\,dt.
\end{eqnarray*}
We now pass into the limit $m$ goes to infinity.
From the a priori estimates \eqref{eq_a_priori}, $r_0 \ln\rho$ lies in $L^\infty({\mathbb R}^+;L^1(\O))$, so $\rho>0$ almost everywhere, and
\begin{eqnarray*}
&&\phi_m(\rho) \mathrm{ \ \ converges \ to \ 1,} \qquad \mathrm{for \ almost \ every \ }(t,x)\in {\mathbb R}^+\times\O, \\
&&v_m \mathrm{ \ \ converges \ to \ } u, \qquad \mathrm{for \ almost \ every \ }(t,x)\in {\mathbb R}^+\times\O, \\
&& | \rho \phi'_m(\rho) | \leq 2, \ \qquad \mathrm{and \ converges \ to \ }0 \qquad \mathrm{for \ almost \ every } \ (t,x)\in {\mathbb R}^+\times\O.
\end{eqnarray*}
Now, using that $\phi_m$ is compactly supported in ${\mathbb R}^+$, we get
\begin{eqnarray*}
{\sqrt{\rho}} \nabla v_m&=&\frac{\phi_m(\rho)}{{\sqrt{\rho}}}\rho\nabla u+4\rho^{1/4} u \rho\phi_m'(\rho)\nabla\rho^{1/4}\\
&=& \frac{\phi_m(\rho)}{{\sqrt{\rho}}}\left[\nabla(\rho u)-2{\sqrt{\rho}}{u \cdot}\nabla{\sqrt{\rho}}\right]+4\rho^{1/4} u \rho\phi_m'(\rho)\nabla\rho^{1/4}\\
&=&\phi_m(\rho) \frac{{\mathbb T}_\nu}{\sqrt\nu}+4\rho^{1/4} u \rho\phi_m'(\rho)\nabla\rho^{1/4},
\end{eqnarray*}
which, thanks to the dominated convergence theorem, converges to ${\mathbb T}_\nu / \sqrt{\nu}$ in $L^2({\mathbb R}^+\times\O)$. Hence, passing into the limit $m\to \infty$, we find
\begin{eqnarray*}
&&\int_0^\infty\int_\O \left\{{\varphi}(u)\left(\rho\partial_t\psi+\rho u\cdot\nabla\psi)\right) -\nabla\psi{\varphi}'(u){\sqrt{\rho}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right)\right. \\
&& \left.-\psi {\varphi}''(u)\frac{{\mathbb T}_\nu}{\sqrt{\nu}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right) +\psi{\varphi}'(u)F\right\}\,dx\,dt.
\end{eqnarray*}
This gives the second equation in the definition of renormalized solutions with
$$
R_{\varphi}= {\varphi}''(u)\frac{{\mathbb T}_\nu}{\sqrt{\nu}}\left({2}\sqrt{\nu}\S+{2}\sqrt{\kappa}{\mathbb S}_\kappa+{\sqrt{\kappa}}\mathbb M\right).
$$
\vskip0.3cm
We want now to show \eqref{eq_viscous_renormalise}. As above, multiplying \eqref{eq_viscous} by $\overline{[\phi_m(\rho){\varphi}_i'(u)]}_{\varepsilon}$, and passing into the limit ${\varepsilon}$ goes to 0, we find
\begin{eqnarray*}
&&\phi_m(\rho){\varphi}'_i(u)\sqrt{\nu\rho}{\mathbb T}_\nu={\nu}\nabla(\phi_m(\rho){\varphi}'_i(u)\rho u)-4\rho \phi'_m(\rho)\nabla\rho^{1/4} \rho^{1/4} u {\sqrt{\rho}} {\varphi}_i'(u)\\
&&-{\varphi}_i''(u)\frac{{\mathbb T}_\nu}{\sqrt{\nu}}{\sqrt{\rho}} u\phi_m(\rho)-2{\nu}\phi_m(\rho){\varphi}'_i(u){\sqrt{\rho}}{u\cdot}\nabla{\sqrt{\rho}}.
\end{eqnarray*}
Passing into the limit $m$ goes infinity, we recover \eqref{eq_viscous_renormalise}, with
$$
\overline{R}_{\varphi}= -{\varphi}_i''(u)\frac{{\mathbb T}_\nu}{\sqrt{\nu}}{\sqrt{\rho}} u.
$$
Hence, $({\sqrt{\rho}},{\sqrt{\rho}} u)$ is a renormalized solution.
\section{From renormalized solutions to weak solutions in the general case}
This section is dedicated to the proof of (2) in Theoren \ref{theo_df}. We consider $({\sqrt{\rho}},{\sqrt{\rho}} u)$, a renormalized solution as defined in Definition \ref{def_renormalise}, in the general case where $r_0\geq0$, $r_1\geq0$ and $\kappa\geq0$, but $\nu>0$. We want to show that it is also a weak solution as defined in Definition \ref{def_weak}. Let $\Phi:{\mathbb R}\to{\mathbb R}$ be a nonnegative smooth function compactly supported, equal to 1 on $[-1,1]$, and $\tilde{\Phi}(z)=\int_0^z\Phi(s)\,ds$. Then we define for $y\in {\mathbb R}^d$
$$
{\varphi_n}(y)=n\tilde{\Phi}(y_1/n)\Phi(y_2/n)\cdot\cdot\cdot\Phi(y_d/n).
$$
Note that ${\varphi_n}$ lies in $W^{2,\infty}({\mathbb R}^d)$ for any fixed $n$, ${\varphi_n}$ converges everywhere to $y\to y_1$, ${\varphi_n}'$ is uniformly bounded in $n$ and converges everywhere to $(1,0,\cdot\cdot\cdot,0)$, and
$\|{\varphi_n}''\|_{L^\infty({\mathbb R})}\leq C/n$ converges to 0, when $n$ converges to infinity. Hence $R_{{\varphi_n}}$ and $\overline{R}_{\varphi_n}$ both converge to 0 in the sense of measure when $n$ converges to infinity.
We use this function ${\varphi_n}$ in the second equation of the Definition \ref{def_renormalise}.
Using the Lebesgue's Theorem for the limit $n\to \infty$, we get the equation on $\rho u_1$ in the Definition \ref{def_weak}. permuting the directions, we get the full vector equation on $\rho u$ in Definition \ref{def_weak}.
We use again the Lebesgue's dominated convergence Theorem to pass into the limit in \eqref{eq_viscous_renormalise} with $i=1$ and the function ${\varphi_n}$ to obtain \eqref{eq_viscous}. Hence, the renormalized solution is also a weak solution.
\section{{Stability and existence of weak renormalized solutions}}
This section is dedicated to the proof of (4) {and (1)} in Theorem \ref{theo_df}. We consider sequences $r_{0,n}, r_{1,n}, \kappa_n, \nu_n, \rho_n, u_n$ as in the hypothesis of Theorem \ref{theo_df}. We begin to show the following lemma:
\begin{lemm}\label{lemm_stabilite}
{Up to} a subsequence, still denoted $n$, {the following properties hold.}
\begin{enumerate}
\item The sequence $\rho_n$ converges strongly to $\rho$ in $C^0({\mathbb R}^+;L^p_\mathrm{loc}(\O))$ for $1\leq p<\sup(3,\gamma)$.
\item The sequence $\rho_n u_n$ converges to $\rho u$ in $C^0({\mathbb R}^+; L^{3/2}(\O)-weak)$, and strongly in $L^p_{\mathrm{loc}}({\mathbb R}^+;L^q(\O))$ for $1\leq p<\infty$, and $1\leq q<3/2$.
\item The sequences ${{\mathbb T}_\nu}_{,n}, {\S}_{,n}, {{{\mathbb S}_\kappa}_{,n}} $ converge weakly in $L^2({\mathbb R}^+\times\O)$ to ${\mathbb T}_\nu, \S, {\mathbb S}_\kappa$.
\item For every function $H \in W^{2,\infty}({\mathbb R}^d)$, and $0<\alpha< 5\gamma/3$, we have that $\rho^\alpha_nH(u_n)$ converges
strongly in $L_{\mathrm{loc}}^p({\mathbb R}^+\times\O)$ to $\rho^\alpha H(u)$ for $1\leq p< 5\gamma/(3\alpha)$.
\item If $\lim_{n\to\infty}r_{1,n}=r_1>0$, then $\rho^{1/3}u_n$ converges to $\rho^{1/3} u$ in $L^p_{\mathrm{loc}}({\mathbb R}^+;L^q(\O))$ for $1\leq p<4$, and $1\leq q< 18/5$.
\item If $\lim_{n\to\infty}r_{0,n}=r_0>0$, then $r^{1/2}_{0,n}u_n$ converges to $r^{1/2}_{0} u$ in $L^p_{\mathrm{loc}}({\mathbb R}^+\times\O)$ for $1\leq p<2$.
\item Consider a smooth and increasing function $h:{\mathbb R}^+\to {\mathbb R}^+$ such that $h(y)=y^{3/4}$ for $y<1$ and $h(y)=y^{1/2}$ for $y>2$.
If $\lim_{n\to\infty}\kappa_{n}=\kappa>0$, then $\nabla h(\rho_n)$ converges to $\nabla h(\rho)$ in $L^2_{\mathrm{loc}}({\mathbb R}^+; L^p(\O))$, for $1\leq p<6$.
\end{enumerate}
\end{lemm}
\begin{proof}
From (1) to (4), we use only a priori estimates which are non dependent on $r_0$, $r_1$, and $\kappa$. Then $\rho_n u_n$ is uniformly bounded in $L^\infty({\mathbb R}^+;L^{3/2}(\O))$. From the continuity equation $\partial_t \rho_n$ is uniformly bounded in $L^\infty({\mathbb R}^+;W^{-1,3/2}(\O))$. Moreover $\nabla\rho_n$ is uniformly bounded in $L^\infty({\mathbb R}^+;L^{3/2}(\O))$, $\rho_n$ is uniformly bounded in $L^\infty({\mathbb R}^+;L^3(\O)\cap L^{\gamma}(\O))$. Hence, using Aubin Simon's Lemma $\rho_n$ is compact in $C^0({\mathbb R}^+;L^p_\mathrm{loc}(\O))$ for $1\leq p<\sup(3,\gamma)$.
\vskip0.3cm
From the second equation in the Definition \ref{def_weak}, the sequence $\partial_t(\rho_n u_n)$ is uniformly bounded in $L^2({\mathbb R}^+; H^{-N}(\O)) $ for a $N$ big enough. From \eqref{eq_viscous}, $\nabla(\rho_n u_n)$ is uniformly bounded in $L^2_\mathrm{loc}({\mathbb R}^+;L^{3/2}(\O))$. Together with $\rho_n u_n$ uniformly bounded in $L^\infty({\mathbb R}^+;L^{3/2}(\O))$ this gives the strong compactness of $\rho_n u_n$ in $L^p_{\mathrm{loc}}({\mathbb R}^+;L^q(\O))$ for $1\leq p<\infty$, and $1\leq q<3/2$.
\vskip0.3cm
The sequences ${{\mathbb T}_\nu}_{,n}, {\S}_{,n}, {{{\mathbb S}_\kappa}_{,n}} $ are uniformly bounded in $L^2({\mathbb R}^+\times\O)$, and so, up to a subsequence, converge weakly in $L^2({\mathbb R}^+\times\O)$ to functions ${\mathbb T}_\nu, \S, {\mathbb S}_\kappa$.
\vskip0.3cm
From (1) and (2), up to a subsequence still denoted $n$, $\rho_n$ and $\rho_n u_n$ converge almost everywhere respectively to $\rho$ and $\rho u$.
So, for almost every $(t,x)$ such that $\rho(t,x)>0$, $u_n(t,x)$ converges to $u(t,x)$, and so $\rho_n(t,x)^\alpha H(u_n(t,x))$ converges to $\rho(t,x)^\alpha H(u(t,x))$. But for almost every $(t,x)$ such that $\rho(t,x)=0$, $|\rho_n(t,x)^\alpha H(u_n(t,x))|\leq C \rho_n(t,x)^\alpha$ which converges to $0=\rho(t,x)^\alpha H(u(t,x))$ since $\alpha>0$. So, we have convergence almost everywhere. And $\rho^\alpha_nH(u_n)$ is uniformly bounded in $L^{5\gamma/(3\alpha)}({\mathbb R}^+\times\O)$. {Indeed $\rho^{\gamma/2}\in L^\infty({\mathbb R}^+,L^2(\O))\cap L^2({\mathbb R}^+,L^6(\O))$,and by interpolation,
$\rho^{\gamma/2}\in L^{10/3}(L^{10/3})$ .
} Hence, we have strong convergence in $L_{\mathrm{loc}}^p({\mathbb R}^+\times\O)$ to $\rho^\alpha H(u)$ for $1\leq p< 5\gamma/(3\alpha)$.
\vskip0.3cm
We assume that $\lim_{n\to\infty}r_{1,n}=r_1>0$.
{ We have
$\rho_n^{1/3}u_n=\rho_n^{1/4}u_n \rho_n^{1/12}$. Since ${\sqrt{\rho}}_n$ is uniformly bounded in $L^\infty({\mathbb R}^+,L^6(\O))$ (From Sobolev, since $\nabla {\sqrt{\rho}}_n$ is uniformly bounded in $L^\infty({\mathbb R}^+,L^2(\O))$), we have $\rho_n^{1/12}$ uniformly bounded in $L^\infty({\mathbb R}^+,L^{36}(\O))$. Moreover, $1/4+1/36=5/18$.
}
Then
the functions $\rho_n^{1/3} u_n $ are uniformly bounded in $L^4({\mathbb R}^+;L^{18/5}(\O))$.
We denote $\mathbf{1}_{\{\rho>0\}}$ the function which is equal to one on
$\{t,x | \rho(t,x)>0\}$ and zero on $\{t,x | \rho(t,x)=0\}$. The function $\mathbf{1}_{\{\rho>0\}}\rho_n^{1/3} u_n$ converges almost everywhere to $\mathbf{1}_{\{\rho>0\}}\rho^{1/3} u$ so the convergence holds in $L^p_{\mathrm{loc}}({\mathbb R}^+;L^q(\O))$ for $1\leq p<\infty$, and $1\leq q< {18/5}$. {Note that for almost every $(t,x)$ such that $\rho(t,x)=0$, we have
$\rho_n(t,x)^{1/12}=(\rho_n(t,x)-\rho(t,x))^{1/12}$.} So, for every $1\leq p<\infty$, and
$1\leq q <36$:
\begin{eqnarray*}
&& \|\mathbf{1}_{\{\rho=0\}}\rho_n^{1/3} u_n\|_{L^{p_1}_{\mathrm{loc}}({\mathbb R}^+;L^{q_1}(\O))}\leq \|\mathbf{1}_{\{\rho=0\}}\rho_n^{1/12}\|_{L^p_{\mathrm{loc}}({\mathbb R}^+;L^q(\O))}\|\rho_n^{1/4}u_n\|_{L^4({\mathbb R}^+\times\O)}\\
&&\qquad\qquad
\leq {C \|(\rho_n-\rho)^{1/12}\|_{L^{p}_{\mathrm{loc}}({\mathbb R}^+;L^{q}(\O))}=}C \|\rho_n-\rho\|^{1/12}_{L^{p/12}_{\mathrm{loc}}({\mathbb R}^+;L^{q/12}(\O))}
\end{eqnarray*}
converges to 0 when $n$ goes to {infinity}, where
$$
\frac{1}{p_1}=\frac{1}{4}+\frac{1}{p}, \qquad \frac{1}{q_1}=\frac{1}{4}+\frac{1}{q}.
$$
So $\rho_n^{1/3} u_n {=\mathbf{1}_{\{\rho=0\} }\rho_n^{1/3} u_n+ \mathbf{1}_{\{\rho>0\}}\rho_n^{1/3} u_n}$ converges to $\rho^{1/3} u$ in $L^p_{\mathrm{loc}}({\mathbb R}^+;L^q(\O))$ for $1\leq p<\infty$, and $1\leq q< {18/5}$.
\vskip0.3cm
Let us assume that $\lim_{n\to\infty}r_{0,n}=r_0>0$. Then $\ln\rho_n$ is uniformly bounded in $L^\infty({\mathbb R}^+;L^1(\O))$. The function $-\log$ is convex, so the limit $\rho$ verifies the same, and $\rho>0$ for almost every $(t,x)\in {\mathbb R}^+\times\O$. So $u_n$ converges almost everywhere to $u$, and $u_n$ is uniformly bounded in $L^2({\mathbb R}^+\times\O)$. Hence $u_n$ converges to $u$ in $L^p_{\mathrm{loc}}({\mathbb R}^+\times\O)$, for $1\leq p<2$.
\vskip0.3cm
Now we assume that $\lim_{n\to\infty}\kappa_{n}=\kappa>0$. We have
\begin{eqnarray*}
&&|\nabla (\nabla h(\rho_n))|=|2\sqrt{\rho_n}h'(\rho_n)\nabla^2\sqrt{\rho_n}+({8}\sqrt{\rho_n}h'(\rho_n)+16 h''(\rho_n)\rho_n\sqrt{\rho_n})(\nabla \rho_n^{1/4}\otimes\nabla\rho_n^{1/4})|\\
&&\qquad\qquad\qquad\leq C(|\nabla^2\sqrt{\rho_n}|+|\nabla \rho_n^{1/4}|^2).
\end{eqnarray*}
So $\nabla (\nabla h(\rho_n))$ is uniformly bounded in $L^2({\mathbb R}^+\times\O)$. Moreover , using the continuity equation and \eqref{eq_viscous}, we get
\begin{eqnarray*}
&&|\partial_t h(\rho_n)|=|h'(\rho_n)\sqrt{\rho_n}\mathrm{Tr}\frac{{{\mathbb T}_\nu}_{,n}}{\sqrt{\nu}}+4\sqrt{\rho_n} u_n\cdot\nabla\rho_n^{1/4} h'(\rho_n)\rho_n^{1/4}|\\
&&\qquad\qquad \leq C(|{{\mathbb T}_\nu}_n|+\rho |u|^2+|\nabla\rho^{1/4}|^2),
\end{eqnarray*}
which is uniformly bounded in $L^2_{\mathrm{loc}}({\mathbb R}^+,L^1(\O))$. Note that we cannot bound it in $L^2({\mathbb R}^+\times\O)$ as in the previous section, since we cannot use that $r_{1,n}$ is bounded by below. We have shown that $\partial_{t}\nabla h(\rho_n)$ is uniformly bounded in $L^2({\mathbb R}^+;W^{-2,6/5}(\O))$. Hence, using the Aubin Simon lemma, we find that $\nabla h(\rho_n)$ converges strongly to $\nabla h(\rho)$ in $L^2_{\mathrm{loc}}({\mathbb R}^+; L^p(\O))$, for $1\leq p<6$.
\end{proof}
\noindent{\underline{Proof of part (4) of Theorem \ref{theo_df}}}
We are now ready to show the part (4) in Theorem \ref{theo_df}.
Using (1) and (2) of Lemma \ref{lemm_stabilite}, we can pass into the limit in the continuity equation. Using (1) (2) (3) and (4) of Lemma \ref{lemm_stabilite} we can pass into the limit into the first line of the second equation of Definition \ref{def_renormalise}. The sequence $R_{n,{\varphi}}$ is uniformly bounded in measures, so it converges to a measure $R_{{\varphi}}$ with the same bound. The function $f_n$ converges weakly in $L^2({\mathbb R}^+ \times \O)$ to $f$ and, thanks to (4) of Lemma \ref{lemm_stabilite},
$\psi\sqrt{\rho_n} {\varphi}'(u_n)$ converges strongly in $L^2({\mathbb R}^+ \times \O)$ to $\psi\sqrt{\rho} {\varphi}'(u)$, so we can pass into the limit in this term. $\nabla \rho_n^{\gamma/2}$ converges weakly in $L^2({\mathbb R}^+ \times \O)$ to $\nabla\rho^{\gamma/2}$, and $\psi \rho_n^{\gamma/2}$ converges strongly to $\rho^{\gamma/2}$ thanks to (1) in Lemma \ref{lemm_stabilite}. So we can pass into the limit in the pressure term. If $r_{0,n}$ converges to 0, then $r_{0,n} u_n=r_{0,n}^{1/2}r_{0,n}^{1/2}u_n$ converges to 0 in $L^2({\mathbb R}^+ \times \O)$, since $r_{0,n}^{1/2}u_n$ is uniformly bounded in $L^2({\mathbb R}^+ \times \O)$. Otherwise, using (6) in Lemma \ref{lemm_stabilite}, it converges to $r_0 u$ in $L^1_{\mathrm{loc}}({\mathbb R}^+\times\O)$. We can treat the term $r_{1,n}$ in the same way using (5) in Lemma \ref{lemm_stabilite}. So the two equations of Definition \ref{def_renormalise} are verified at the limit. Thanks to (1) and (2) of Lemma \ref{lemm_stabilite}, we can pass into the limit for the initial values. It remains to pass into the limit in \eqref{eq_viscous_renormalise}, and \eqref{eq_quantic}. The measures $\overline{R}_{n,{\varphi}}$ are uniformly bounded in measures, so they converge to a measure with the same bound. The functions $\nabla\sqrt{\rho_n}$ converge weakly to $\nabla{\sqrt{\rho}}$ in $L^2_{\mathrm{loc}}({\mathbb R}^+ \times \O)$. So, using (1) (3) and (4) of Lemma \ref{lemm_stabilite}, we can pass into the limit in \eqref{eq_viscous_renormalise}. If $\kappa_n$ converges to 0, then $\sqrt{\kappa_n}{\mathbb S}_\kappa$ converges to 0 weakly in $L^2({\mathbb R}^+ \times \O)$. Otherwise, $\nabla^2\sqrt{\rho_n}$ converges weakly in $L^2({\mathbb R}^+ \times \O)$ to $\nabla^2{\sqrt{\rho}}$ and, thanks to (1) of Lemma \ref{lemm_stabilite}, $\sqrt{\rho_n}\nabla^2\sqrt{\rho_n}$ converges weakly to $\sqrt{\rho}\nabla^2\sqrt{\rho}$. Note that
$$
\sqrt{\rho_n} \nabla (\rho^{1/4}_n\otimes\rho^{1/4}_n)= g(\rho_n) \nabla (h(\rho_n)\otimes \rho^{1/4}_n),
$$
with $4 h'(\rho_n) \rho^{1/4}_n g(\rho_n)=1$. Especially, $|g(\rho_n)|\leq 1+\rho^{1/4}_n$. So, thanks to (1) of Lemma \ref{lemm_stabilite}, $g(\rho_n)$ converges strongly to $g(\rho)$ in $L^4({\mathbb R}^+,L^2(\O))$, $\nabla\rho^{1/4}_n$ converges weakly to $\nabla\rho^{1/4}$ in $L^4({\mathbb R}^+,L^4(\O))$, and $\nabla h(\rho_n)$ converges strongly to $\nabla h(\rho)$ in $L^2({\mathbb R}^+,L^4(\O))$ thanks to (7) in Lemma \ref{lemm_stabilite}. Hence we can pass into the limit in \eqref{eq_quantic}. This ends the proof of (4) in Theorem \ref{theo_df}.
\vskip0.3cm
\noindent{\underline{Proof of part (1) in Theorem \ref{theo_df}}}.
Consider sequences $r_{0,n}>0$, $r_{1,n}>0$ and $\kappa_n>0$, converging respectively to $r_0\geq0$, $r_1\geq0$ and $\kappa\geq0$. For $n$ fixed, thanks to \cite{VasseurYu2015}, there exists a weak solution in the sense of Definition \ref{def_weak} to the system. Thanks to (2) in Theorem \ref{theo_df}, these solutions are renormalized solutions in the sense of Definition \ref{def_renormalise}. Thanks to the stability result (4) in Theorem \ref{theo_df}, the limit is a renormalized solution for the system with coefficients $r_0$, $r_1$ and $\kappa$.
|
\section{Introduction}
The parameter $\epsilon_{K}^{\prime}/\epsilon_{K}$ is the ratio of
{the measures} of direct {and} indirect charge-parity ($CP$)
violation in the Kaon system. While {indirect} $CP$ violation is a
per-mille effect in the Standard Model (SM), {$\epsilon_{K}^{\prime}$ is
smaller by another three orders of magnitude than $\epsilon_{K}$,
with} $|\epsilon_{K}^{\prime}|\sim \mathcal{O}(10^{-6})$. A strong
suppression {by the Glashow-Iliopoulos-Maiani (GIM) mechanism} and an
accidental cancellation of leading contributions in the Standard Model
{makes} $\epsilon_{K}^{\prime}/\epsilon_{K}$ highly sensitive to new
physics. The first element of the SM prediction for
$\epsilon_{K}^{\prime}$ is the calculation of initial conditions for
Wilson coefficients and their renormalization group evolution from the
electroweak scale (of the order of $W$ and top mass) down to the
hadronic scale of order 1$\,$ GeV, at which hadronic matrix elements
are calculated. These steps purely involve perturbative methods and
have been carried out to leading order (LO) in the strong coupling
constant $\alpha_s$ in {Refs.\cite{Gilman:1978wm, Guberina:1979ix, Hagelin:1983rb, Buras:1987qa}.}
The next-to-leading order (NLO)
involves {the electromagnetic coupling $\alpha_{EM}\simeq 1/128$ \cite{Flynn:1989iu,Buchalla:1989we,Paschos:1991as, Lusignoli:1991bm},
the next higher order in $\alpha_s$ \cite{Buras:1991jm,Buras:1992tc, Ciuchini:1993vr}, and order $\alpha_{EM} \alpha_s $ \cite{Ciuchini:1993vr, Buras:1992zv ,Buras:1993dy}.}
In terms
of isospin amplitudes {$\epsilon_{K}^{\prime}$} is given by
({see e.g.\ Ref.}~\cite{Buras:2015yba})
%
\begin{equation}
\frac{\epsilon_{K}^{\prime}}{\epsilon_{K}} = \frac{\omega_{+}}{\sqrt{2}
\left|\epsilon_{K}\right| \textrm{Re} A_0} \left( \frac{1}{\omega_{+}}
\textrm{Im} A_2 - (1-\hat{\Omega}_{\textrm{eff}}) \textrm{Im} A_0
\right),
\label{eq:epsilonequation}
\end{equation}
where $A_I \equiv \langle (\pi \pi)_I | \mathcal{H}_{\textrm{eff}}^{|
\Delta S |=1} | K^0 \rangle$ are isospin amplitudes and $\omega_{+} =
(4.53 \pm 0.02) \times 10^{-2}$ (see
Refs.~\cite{Cirigliano:2003nn,Buras:2015yba} for the precise
definition), $|\epsilon_{K}|=(2.228 \pm 0.011)\cdot 10^{-3}$,
and $\textrm{Re} A_0 = ( 3.3201\pm 0.0018 )\times 10^{-7}\,\text{GeV}$
are taken from experiment. $\hat{\Omega}_{\textrm{eff}} = (14.8 \pm
8.0)\times 10^{-2}$ {parameterizes} isospin-violating contributions
\cite{Cirigliano:2003nn, Cirigliano:2003gt}.
The $| \Delta S |=1$ nonleptonic effective Hamiltonian for weak decays
in the Standard Model is given by \cite{Buras:1993dy
\beq
\mathcal{H}_{\textrm{eff}}^{ | \Delta S|=1} &= \frac{G_F}{\sqrt{2}}
\lambda_u \sum_{i=1}^{10} Q_i (\mu) \Bigl( \left( 1 - \tau \right) z_i
(\mu) + \tau v_i (\mu) \Bigr) + \textrm{H.c.} \\ &\equiv
\frac{G_F}{\sqrt{2}} \lambda_u \sum_{i=1}^{10} Q_i (\mu) \left( z_i
(\mu) + \tau y_i (\mu) \right) + \textrm{H.c.},
\label{eq:hamilton}
\eeq
where $\lambda_u = V_{us}^{\ast} V_{ud} $ and $\tau = - V_{ts}^{\ast}
V_{td} / \left( V_{us}^{\ast} V_{ud} \right)$. The operator basis $Q_i$
{comprises} ten operators which are defined in Ref.~\cite{Buras:1993dy};
the current-current operators $Q_1$ and $Q_2$
\beq
Q_1 = \left( \bar{s}_{\alpha} u_{\beta}\right)_{V-A} \left( \bar{u}_{
\beta} d_{\alpha} \right)_{V-A},~~~~~~ Q_2 = \left( \bar{s} u
\right)_{V-A} \left( \bar{u} d\right)_{V-A}, \label{eq:q1}
\eeq
the QCD-penguin
operators $Q_3$ to $Q_6$ \beq Q_3 = \left( \bar{s} d \right)_{V-A}
\sum_q \left( \bar{q} q \right)_{V-A},~~~
Q_4 = \left( \bar{s}_{\alpha} d_{\beta} \right)_{V-A} \sum_q \left( \bar{q}_{\beta} q_{\alpha} \right)_{V-A},\\
Q_5 = \left( \bar{s} d \right)_{V-A} \sum_q \left( \bar{q} q
\right)_{V+A},~~~ Q_6 = \left( \bar{s}_{\alpha} d_{\beta} \right)_{V-A}
\sum_q \left( \bar{q}_{\beta} q_{\alpha} \right)_{V+A}, \eeq and the
QED-penguin operators $Q_7$ to $Q_{10}$ \beq Q_7 = \frac{3}{2} \left(
\bar{s} d \right)_{V-A} \sum_q e_q \left( \bar{q} q \right)_{V+A},~~~
Q_8 = \frac{3}{2} \left( \bar{s}_{\alpha} d_{\beta} \right)_{V-A} \sum_q e_q \left( \bar{q}_{\beta} q_{\alpha} \right)_{V+A},\\
Q_9 = \frac{3}{2} \left( \bar{s} d \right)_{V-A} \sum_q e_q \left(
\bar{q} q \right)_{V-A},~~~ Q_{10} = \frac{3}{2} \left(
\bar{s}_{\alpha} d_{\beta} \right)_{V-A} \sum_q e_q \left(
\bar{q}_{\beta} q_{\alpha} \right)_{V-A},\label{eq:q10
\end{equation}
where $V \mp A$ represents $\gamma_{\mu} (1 \mp \gamma_5)$, $\alpha$ and
$\beta$ {denote} color indices, and $e_q$ is the electric charge of the
quark $q$. The corresponding Wilson coefficients $z_i$ and $v_i$ (or
$y_i$) serve as effective couplings to these effective operators.
By virtue of the framework of effective theories, the parameter $\mu$ splits short distance from long distance scales,
effectively separating the perturbative high energy regime from the non-perturbative realm of low energy QCD.
Taking up the perturbative part of the calculation, the Wilson
coefficients have been determined through matching calculations up to
next-to-leading order at the {scale $M_W$} \cite{Buras:1993dy}.
The calculation of the hadronic matrix elements, being non-perturbative quantities, is a major challenge and has recently been performed on the lattice with unprecedented accuracy \cite{Blum:2011ng,Blum:2012uk, Blum:2015ywa, Bai:2015nea}.
The combination of these calculations into a prediction for
$\epsilon_{K}^{\prime}/\epsilon_{K}$ requires a treatment within
renormalization group (RG) improved perturbation theory to
sum up large logarithms. However, it is known that the analytic
determination of the required evolution matrix at the next-to-leading
order suffers from singularities appearing in intermediate steps of the
calculation, which make a computational evaluation highly laborious and
complicated. The standard way to solve the NLO RG equations requires
the diagonalization of the LO anomalous dimension matrix
$\hat{\gamma}_s^{(0)}$ and the NLO correction involves fractions
whose denominators contain the differences of eigenvalues of
$\hat{\gamma}_s^{(0)}$. Some of these denominators vanish and are
usually regulated in the numerical evaluation \cite{Ciuchini:1992tj,
Ciuchini:1993vr}.
In Ref.~\cite{Huber:2005ig} an analytic solution for the
RG equations which is free of singularities is presented.
This solution involves the diagonalization of $\hat{\gamma}_s^{(0)}$
and gives explicit prescriptions to handle the different cases in which the
formulae of Refs.~\cite{Ciuchini:1992tj,Ciuchini:1993vr} develop
singularities.
In this paper, we present a new singularity-free solution which permits an easy
and convenient numerical implementation. Instead of singularities
our analytic formula has undetermined parameters. However, we will show
that these {spurious} parameters {cancel} and {leave} the evolution
matrix unambiguous. Unlike the solution of Ref.~\cite{Huber:2005ig}
our new formula requires neither the diagonalization of
$\hat{\gamma}_s^{(0)}$ nor a distinct treatment of the part of the RG
evolution which involves the spurious singularities. Using our
new RG evolution and the latest lattice results
\cite{Blum:2011ng,Blum:2012uk, Blum:2015ywa, Bai:2015nea}, we calculate
the $\epsilon_{K}'/\epsilon_{K}$ in the Standard Model at
next-to-leading order {to find a value which is} below the
experimentally measured quantity {by} 2.8\,$\sigma$.
The second objective of this paper is the derivation of a useful
formula for the calculation of new physics contributions to
$\epsilon_K^\prime / \epsilon_K$, in which we evaluate the evolution
matrices {for scales far above the electroweak scale. To this end
we identify a contribution of order $\alpha_{EM}^2/\alpha_s^2$ in the
evolution matrix which can become relevant for studies of TeV-scale new
physics, because $\alpha_s$ decreases with increasing scale.} We observe an
approximately logarithmic behavior of the evolution matrix as a function
of the energy scale above the electroweak scale.
This paper is organized as follows. In Sec.~\ref{section2}, we briefly
review the RG evolution of the $|\Delta S| = 1$ effective Hamiltonian at
the next-to-leading order. We give a detailed analysis of the
evolution matrix and its singularities and provide a new analytic
solution without singularities. Then we evaluate $\epsilon'_K /
\epsilon_K$ in the Standard Model at the next-to-leading order in
Sec.~\ref{sec:SM}. In Sec.~\ref{sec:NP}, we work out the evolution
matrices in the high-energy regime explicitly for calculations of new
physics contributions. The last section is devoted to conclusions and
discussion.
\section{Renormalization Group Evolution of the
$\boldsymbol{\Delta S=1}$ Hamiltonian}
\label{section2}
In this section, we review the singularities in the RG evolution of
the $|\Delta S| = 1$ effective Hamiltonian at the next-to-leading order.
Then we generalize the analytic ansatz of the RG evolution given in the
literature and present a solution, which is finite at all stages of
the calculation. Our solution contains free parameters, which we
show to cancel from the evolution matrix, and compare our
singularity-free solution with the standard results from the
literature.
\subsection{Singularities in the Evolution Matrix}
\label{sec:singexpl}
The evolution of the Wilson coefficients $v_i$ and $z_i$ from the $W$
boson mass and the charm mass respectively to the hadronic scale
$\mu$ are given by
\beq
\vec{v} (\mu) & = \hat{U}_3 (\mu, \mu_c)\hat{M}_c (\mu_c) \hat{U}_4 (\mu_c,m_b)\hat{M}_b (m_b) \hat{U}_5 (m_b,M_W) \vec{v}(M_W),
\label{eq:wilsonv}\\
\vec{z} (\mu) & = \hat{U}_3 (\mu, \mu_c) \vec{z}(\mu_c),
\label{eq:wilsonz}
\eeq where $\hat{U}_f (\mu_1,\mu_2)$ is the RG evolution matrix from
$\mu_2$ down to $\mu_1$ and $f $ is the number of the active
flavors between these two energy scales. The matrices
$\hat{M}_{c,b}$ represent matching matrices between effective theories
with different numbers of flavor and are given in
Ref.~\cite{Buras:1993dy}. Although the effect of the running of
$\alpha_{EM}$ is numerically negligible for
$\epsilon_{K}'/\epsilon_{K}$ in the Standard Model \cite{Buras:1993dy}, we consider this effect to cover new-physics scenarios with largely
separate scales.
The general form of the evolution matrix is given by
\cite{Buras:1979yt, Buchalla:1995vs},
%
\beq
\hat{U}_f (\mu_1,\mu_2) = T_{g_s} \exp \int^{g_s(\mu_1)}_{g_s(\mu_2)} d g_s' \frac{\hat{\gamma}^T \left(g_s' \right)}{\beta \left(g_s'\right)},
\label{eq:Uint}
\eeq
%
with the $g_s$-ordering operator $T_{g_s}$ and
the anomalous dimension matrix $\hat{\gamma}$
and the QCD $\beta$ function.
The expansions of the latter two quantities and $\alpha_{EM}$ up to NLO read:
\beq
\label{eq:anomalousNLO}
\hat{\gamma} \left(g_s (\mu) \right) &= \frac{\alpha_s (\mu)}{4 \pi }
\hat{\gamma}_s^{(0)} + \frac{\alpha_{EM}(\mu)}{4 \pi}
\hat{\gamma}_{e}^{(0)} + \frac{\alpha^2_s (\mu)}{\left( 4
\pi\right)^2} \hat{\gamma}_s^{(1)} + \frac{\alpha_{EM} (\mu)
\alpha_s(\mu)}{\left( 4 \pi\right)^2} \hat{\gamma}_{se}^{(1)},\\
\beta \left( g_s (\mu)\right) & = - g_s (\mu) \left(\frac{\alpha_s
(\mu)}{4 \pi } \beta_0 + \frac{\alpha^2_s
(\mu)}{\left( 4 \pi\right)^2} \beta_1 + \frac{\alpha_s (\mu) \alpha_{EM}(\mu)}{ ( 4 \pi)^2} \beta^{se}_{1}\right), \\
\alpha_{EM}(\mu) &= \alpha_{EM} (M) \left\{ 1 +
\frac{\alpha_{EM} (M)}{ \alpha_s (\mu)}
\frac{\beta^{e}_0}{\beta_0} \left( 1 -
\frac{\alpha_s(\mu)}{\alpha_s (M)} \right)\right\}^{-1}, \eeq
where $\beta_0 = 11 - 2 f / 3 $, $\beta_1 = 102 - 38f /3 $,
$\beta^{se}_{1} = -8/9( u + d/4) $, and $\beta^{e}_0 = - 4/3 (
4u /3 +$ $d/3 +\ell) $ are the leading and next-to-leading
coefficients of the QCD and QED beta functions,
and $u,\,d,\,\ell$ are the numbers of the active up-type-quark,
down-type-quark, and charged-lepton flavors ($f =
u+d$). $\hat{\gamma}^{(0)}_s$ is the LO QCD anomalous dimension
matrix, and the NLO corrections consist of the three remaining
matrices, $\hat{\gamma}^{(0)}_e$, $\hat{\gamma}^{(1)}_s$, and
$\hat{\gamma}^{(1)}_{se}$, which are the leading QED, next-to-leading
QCD, and combined QCD-QED anomalous dimension matrices, respectively.
The ansatz for the NLO evolution matrix (with
$\mu_1 < \mu_2$) is given by \cite{Ciuchini:1992tj, Ciuchini:1993vr}
\begin{equation}
\label{eq:UNLO}
\hat{U}_f (\mu_1,\mu_2) = \hat{K}(\mu_1) \hat{U}_0 (\mu_1,\mu_2)
\hat{K}^{\prime} (\mu_2),
\end{equation}
where
\begin{align}
\hat{K}(\mu_1) & = \left( \hat{1} + \frac{\alpha_{EM}}{4 \pi}
\hat{J}_{se}\right) \left( \hat{1} + \frac{\alpha_s (\mu_1)}{4 \pi}
\hat{J}_s \right) \left( \hat{1} + \frac{\alpha_{EM}}{\alpha_s (\mu_1)}
\hat{J }_{e}\right), \label{eq:K}\\
\label{eq:Kprime}
\hat{K}^{\prime} (\mu_2) & = \left( \hat{1} - \frac{\alpha_{EM}}{\alpha_s (\mu_2)} \hat{J }_{e} \right) \left( \hat{1} - \frac{\alpha_s (\mu_2)}{4 \pi} \hat{J}_{s} \right) \left( \hat{1} - \frac{\alpha_{EM}}{4 \pi} \hat{J}_{se}\right),
\end{align}
and the LO evolution matrix
\begin{equation}
\hat{U}_0 (\mu_1,\mu_2) = \hat{U}_0 \left(\alpha_s(\mu_1),
\alpha_s(\mu_2)\right) = \exp \left[ \frac{\hat{\gamma}^{(0)T}_s}{2
\beta_0} \ln \frac{\alpha_s(\mu_2)}{\alpha_s(\mu_1)}
\right],
\end{equation}
where the QED contributions to the beta functions ($
\beta^{se}_{1},~\beta^{e}_0 $) are discarded in this subsection
\ref{sec:singexpl}.
The matrices $\hat{K}(\mu_1 )$ and $\hat{K}' (\mu_2 )$ encode the
NLO corrections and depend on the number of active flavors through
the beta function and the anomalous dimension matrices. The matrices
$\hat{J}_{e}$, $\hat{J}_{s}$ and $\hat{J}_{se}$ govern the leading
electromagnetic, next-to-leading strong, and next-to-leading
combined strong-electromagnetic contributions to the RG evolution.
Differentiating Eqs.~\eqref{eq:UNLO} and \eqref{eq:Uint} with respect to $ g_s (\mu_1)$ yields the following differential equation for $\hat{K}(g_s (\mu_1))$ \cite{Buras:1979yt,Buras:1991jm},
\beq
\frac{\partial}{\partial g_s (\mu_1)} \hat{K}(g_s(\mu_1)) - \frac{1}{g_s (\mu_1) } \hat{K}(g_s(\mu_1)) \frac{\hat{\gamma}_{s}^{(0)T}}{\beta_0} = \frac{\hat{\gamma}^T (g_s (\mu_1))}{\beta (g_s (\mu_1))} \hat{K}(g_s(\mu_1)).
\label{eq:diffK}
\eeq The traditional ansatz in the literature is to take $\hat{J}_{e}$,
$\hat{J}_{s}$ and $\hat{J}_{se}$ as constant matrices for any fixed
number of flavors. The differential equation ~\eqref{eq:diffK} then
implies the following equations for the matrices
$\hat{J}_{e}$, $\hat{J}_{s}$ and $\hat{J}_{se}$ \cite{Ciuchini:1993vr},
\beq
\hat{J}_{s} - \left[ \hat{J}_{s} , \frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0} \right ] &= \frac{\beta_1}{\beta_0} \frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0} - \frac{\hat{\gamma}^{(1)T}_s}{2 \beta_0},
\label{eq:JsRome}\\
\hat{J }_{e} + \left[ \hat{J }_{e} , \frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0} \right ] &= \frac{\hat{\gamma}^{(0)T}_e}{2 \beta_0},
\label{eq:JeRome}\\
\left[ \hat{J}_{se} , \frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0} \right ] & =
\frac{\hat{\gamma}^{(1)T}_{se}}{2 \beta_0}
+ \left[ \frac{\hat{\gamma}^{(0)T}_e}{2 \beta_0}, \hat{J}_{s} \right]
- \frac{\beta_1}{\beta_0} \frac{\hat{\gamma}^{(0)T}_e}{2 \beta_0}.
\label{eq:JseRome}
\eeq
It is well known, however, that Eqs.~\eqref{eq:JsRome} and \eqref{eq:JeRome} develop singularities in the case of three flavors.
Furthermore,
Eq.~\eqref{eq:JseRome} is even singular for any number of flavors.
We now show how these singularities arise. For this purpose, it is
instructional to transform Eqs.~\eqref{eq:JsRome}--\eqref{eq:JseRome}
into the diagonal basis of $\hat{\gamma}^{(0)T}_s $. This is a common
procedure in the literature since it allows to isolate the singularities
and remove them ``by hand''. We stress that this is only for the purpose
of a better understanding of the origin of these singularities.
A numerical evaluation of
our solution does not require the diagonalisation of
$\hat{\gamma}^{(0)T}_s $.
Upon transforming Eqs.~\eqref{eq:JsRome}--\eqref{eq:JseRome} into the
basis where $\hat{\gamma}^{(0)T}_{s,D} =\hat{V}^{-1}
\hat{\gamma}^{(0)T}_{s} \hat{V}$ is diagonal, the solutions of
Eqs.~\eqref{eq:JsRome} and \eqref{eq:JeRome} take the form
\begin{equation}
\left( \hat{V}^{-1} \hat{J}_{s,e} \hat{V} \right)_{ij} = \frac{\cdots}{2
\beta_0 \mp \left( (\hat{\gamma}^{(0)T}_{s,D})_{jj} -
(\hat{\gamma}^{(0)T}_{s,D})_{ii} \right) }.
\label{eq:singularityA}
\end{equation}
We find singular solutions if the difference of two eigenvalues of
$\hat{\gamma}^{(0)T}_{s}$ is equal to $2 \beta_0$, which is the
case for three flavors: $\hat{\gamma}^{(0)T}_{s,D}$ has the
elements
$2$ and $-16$ and $2 \beta_0^{f=3} = 18$, so that one denominator
in Eq.~\eqref{eq:singularityA} vanishes with a generally non-zero numerator.
When we transform Eq.~\eqref{eq:JseRome} into the same basis
\begin{equation}
\left( \hat{V}^{-1} \hat{J}_{se} \hat{V} \right)_{ij} = \frac{\cdots}{
(\hat{\gamma}^{(0)T}_{s,D})_{jj} - (\hat{\gamma}^{(0)T}_{s,D})_{ii}},
\label{eq:singularityB}
\end{equation}
we find singular results for $i=j$ and also for degenerate eigenvalues.
Nonetheless, once all relevant terms have been joined together, all
these singularities cancel and the evolution matrix $\hat{U}_f
(\mu_1,\mu_2)$ becomes finite~\cite{Ciuchini:1993vr}. This procedure,
however, requires taking care of each singularity by hand by adopting
the aforementioned diagonal basis, then regularizing the singularities
and keeping track of them until the end of the calculation. Indeed,
Buras et al.\ have regulated some of the singularities by a logarithmic
term \cite{Buras:1993dy}. Subsequently, Adams and Lee have proposed a
systematical solution for all singularities \cite{Adams:2007tk}, which,
however, still requires the adoption of a certain diagonal basis. The
freedom of choosing the order of the eigenvalues on the diagonal of
$\hat{\gamma}^{(0)T}_{s,D}$ involves an ambiguity. This can pose a
problem in computational implementations, since it is absolutely
necessary to use the same diagonal basis as Adams and Lee do, which is
not the one which orders eigenvalues by their numerical value. The
solution in Ref.~\cite{Huber:2005ig} follows the same line, after
diagonalizing $\hat{\gamma}^{(0)T}_{s,D}$ several different cases must
be considered: whenever two eigenvalues differ by an integer multiple
of $2\beta_0$ a special implementation is required. In the next
subsection we propose a solution which does not rely on a specific basis
and permits a much faster, easier and, in particular, more stable
computational algorithm.
\subsection{Removing the Singularities}
In order to eliminate the singularities, we generalize the Roma
group's ansatz \cite{Ciuchini:1992tj, Ciuchini:1993vr} by
adding a logarithmic scale dependence to the $\hat{J}$ matrices
used in Eqs.~\eqref{eq:K},\,\eqref{eq:Kprime} in the following way
\begin{align}
\hat{J}_{s} ~\rightarrow ~& \hat{J}_{s}( \alpha_s(\mu)) = \hat{J}_{s,0}
+ \hat{J}_{s,1}\ln \alpha_s(\mu) , \nonumber\\
\hat{J}_{e} ~\rightarrow ~& \hat{J }_{e}( \alpha_s(\mu)) = \hat{J}_{e,0}
+ \hat{J}_{e,1} \ln \alpha_s(\mu) , \nonumber\\
\hat{J}_{se} ~\rightarrow ~& \hat{J}_{se}( \alpha_s(\mu)) =
\hat{J}_{se,0} + \hat{J}_{se,1} \ln \alpha_s(\mu) +
\hat{J}_{se,2} \ln^2 \alpha_s(\mu).
\label{eq:newJ}
\end{align}
In addition, we extend Eqs.~\eqref{eq:K},\,\eqref{eq:Kprime} as follows:
\begin{align}
\hat{K}(\mu_1, \mu_2) =& \left(\hat{1} + \frac{\alpha_{EM}}{4 \pi}
\hat{J}_{se} ( \alpha_s(\mu_1)) \right) \left(\hat{1} +
\frac{\alpha_s (\mu_1)}{4 \pi} \hat{J}_{s}( \alpha_s(\mu_1)) \right)
\nonumber \\ &\times \left(\hat{1} + \frac{\alpha_{EM}}{\alpha_s (\mu_1)}
\hat{J }_{e}( \alpha_s(\mu_1)) \right.
\nonumber \\ & ~~~~\left. + \left(\frac{\alpha_{EM}}{\alpha_s
(\mu_1)} \right)^2 \left( \hat{J}_{ee} ( \alpha_s(\mu_1)) - \frac{\beta^e_0}{\beta_0} \left( 1 - \frac{\alpha_s(\mu_1) }{\alpha_s (\mu_2)} \right) \hat{J}_{e}( \alpha_s(\mu_1)) \right)
\right), \label{eq:ext} \\ %
\hat{K}^{\prime} (\mu_2) =& \left( \hat{1} -
\frac{\alpha_{EM}}{\alpha_s (\mu_2)} \hat{J }_{e} ( \alpha_s(\mu_2))
- \left(\frac{\alpha_{EM}}{\alpha_s (\mu_2)} \right)^2 \left(
\hat{J}_{ee} ( \alpha_s(\mu_2)) - \left( \hat{J }_{e} (
\alpha_s(\mu_2))\right)^2 \right) \right) \nonumber \\ & \times \left(
\hat{1} - \frac{\alpha_s (\mu_2)}{4 \pi} \hat{J}_{s} (
\alpha_s(\mu_2)) \right) \left(\hat{1} - \frac{\alpha_{EM}}{4 \pi}
\hat{J}_{se} ( \alpha_s(\mu_2)) \right), \label{eq:ext2}
\end{align}
which somewhat resembles the NNLO {QCD} result of
Ref.~\cite{Gorbahn:2004my}. Here
we use the abbreviation $ \alpha_{EM} \equiv \alpha_{EM} (\mu_2)$ and
\beq
\hat{J}_{ee}( \alpha_s(\mu)) =
\hat{J}_{ee,0} + \hat{J}_{ee,1} \ln \alpha_s(\mu).
\label{eq:newJee}
\eeq
We systematically include $\mathcal{O}( \alpha_{EM}^2 / \alpha_s^2
)$ corrections in the RG evolution. This contribution has not been
considered in the literature. Although appearing as
$\mathcal{O}(\alpha_{EM}^2)$,
these
terms can become sizable at high energies because of the awkward $ 1/
\alpha_s^2 $ dependence, making them numerically comparable to
$\mathcal{O}(\alpha_s)$. We note
that this contribution does not receive contributions from higher
orders of the anomalous dimension matrix in Eq.~\eqref{eq:anomalousNLO},
but only appears at the next-to-leading order.
{With these generalizations we can now solve the} differential
equation in Eq.~\eqref{eq:diffK}. Inserting our ansatz into
Eq.~\eqref{eq:diffK} we obtain the following nine matrix equations for
the nine constant matrices $\hat{J}$:
\begin{align}
\label{eq:Js1full}
\hat{J}_{s,1} - \left[\hat{J}_{s,1},\frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & = 0, \\
\label{eq:Js0full}
\hat{J}_{s,0} - \left[\hat{J}_{s,0},\frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & = \frac{\beta_1}{\beta_0}\frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0} - \frac{\hat{\gamma}^{(1)T}_s}{2 \beta_0} - \hat{J}_{s,1}, \\
\label{eq:Je1}
\hat{J}_{e,1} + \left[\hat{J}_{e,1},\frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & = 0, \\
\label{eq:Je0}
\hat{J}_{e,0} + \left[\hat{J}_{e,0},\frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & = \frac{\hat{\gamma}^{(0)T}_e}{2 \beta_0} + \hat{J}_{e,1},\\
\label{eq:Jse2}
\left[\hat{J}_{se,2}, \frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & = 0,\\
\label{eq:Jse1}
\left[\hat{J}_{se,1}, \frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & =
\left[ \frac{\hat{\gamma}^{(0)T}_e}{ 2 \beta_0}, \hat{J}_{s,1} \right]
+ 2 \hat{J}_{se,2}, \\
\label{eq:Jse0}
\left[\hat{J}_{se,0},\frac{\hat{\gamma}^{(0)T}_s}{2 \beta_0}\right] & =
\frac{\hat{\gamma}^{(1)T}_{se}}{2 \beta_0}
+\left[ \frac{\hat{\gamma}^{(0)T}_e}{ 2 \beta_0}, \hat{J}_{s,0} \right]
- \frac{\beta_1}{\beta_0} \frac{\hat{\gamma}^{(0)T}_e}{2 \beta_0}
- \frac{\beta^{se}_{1}}{\beta_0} \frac{\hat{\gamma}_s^{(0) T}}{ 2 \beta_0} + \hat{J}_{se,1},
\\
%
\label{eq:Jee1}
\hat{J}_{ee,1} + \left[\hat{J}_{ee,1}, \frac{\hat{\gamma}^{(0)T}_s}{4 \beta_0}\right] & =
\frac{\hat{\gamma}^{(0)T}_e}{4 \beta_0} \hat{J}_{e,1} + \frac{1}{2} \frac{\beta^e_0}{\beta_0} \hat{J}_{e,1 },\\
\label{eq:Jee0}
\hat{J}_{ee,0} + \left[\hat{J}_{ee,0}, \frac{\hat{\gamma}^{(0)T}_s}{4
\beta_0}\right] & = \frac{\hat{\gamma}^{(0)T}_e}{4 \beta_0}
\hat{J}_{e,0} + \frac{1}{2} \frac{\beta^e_0}{\beta_0} \hat{J}_{e,0 } + \frac{1}{2} \hat{J}_{ee,1}.
\end{align}
These equations yield finite solutions for $\hat{J}$. As an effect of
the constant matrices $\hat{J}_{s(,e,se),1}$, the analytic singularities
of Eqs.~\eqref{eq:JsRome}--\eqref{eq:JseRome} do not occur, because for
the problematic matrix elements now both sides of the equations are
zero. We stress that one can solve Eqs.~\eqref{eq:Js1full} to
\eqref{eq:Jee0} without diagonalizing $\hat{\gamma}^{(0)T}_s$; these
equations are mere systems of linear equations for the $100$ elements of
$\hat{J}_{s,e,ee,0,1}$ and $\hat{J}_{se,0,1,2}$ each, which are
{quickly} solved by computer algebra programs
\cite{Lenz:1997aa}. However, there are {multiple solutions} in some
of the inhomogeneous equations, {because} the corresponding
homogeneous equations have a non-trivial null space. As a consequence,
these solutions for $\hat{J}$ {depend on} arbitrary parameters,
e.g. there are 16 undetermined components in the case of three active
flavors. These parameters, however, do not produce any ambiguity in
physical results. In the next subsection, we will show that they
completely drop out after combining terms of the same order and the
evolution matrix in Eq.~\eqref{eq:UNLO} does not depend on these
parameters. Therefore, one can set them to arbitrary values from the
beginning. In our calculation of $\epsilon_{K}'/\epsilon_{K}$ we kept
the parameters arbitrary as a crosscheck of the consistency of our
calculation.
The procedure to determine the evolution matrix from $\mu_2$ to $\mu_1$
requires algebraically solving the matrix equations
\eqref{eq:Js1full}--\eqref{eq:Jee0} for a given number of active
flavors and inserting the solutions into the full evolution matrix in
Eq.~(\ref{eq:UNLO}). We use $10\times10$ anomalous dimension matrices
$\hat{\gamma}^{(0)}_s$, $\hat{\gamma}^{(0)}_{e}$, $\hat{\gamma}^{(1)}_s$
and $\hat{\gamma}^{(1)}_{se}$
\cite{Buras:1992tc,Buras:1992zv,Ciuchini:1993vr,Buchalla:1995vs}. The
solutions for the matrices $\hat{J}$ in the case of three active
flavors (with two active leptons) in naive dimensional regularization (NDR) scheme with
$\overline{\textrm{MS}}$ subtraction, are given as follows:
\beq
&\hat{J}_{s,0} = \nonumber \\ & \scalebox{0.69}{$ \left(
\begin{array}{cccccccccc}
-55/324 & 223/108 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
223/108 & -55/324 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
-0.7392 & -0.3061 & -2.999 & -0.6652 & 1.457 & 0.2171 & 0 & 0 &
0.3061 & 0.7392 \\
0.3814 & -0.1853 & 2.838 & 1.037 & -0.05711 & -0.004122 & 0 & 0 &
0.1853 & -0.3814 \\
0.3990 & 0.3264 & 1.850 & 1.444 & -2.514 & 2.750 & 0 & 0 &
-0.3264 & -0.3990 \\
-1.181 & -1.776 & -7.095 & -6.691 & 0.6263 & 4.528 & 0 & 0 & 1.776
& 1.181 \\
0 & 0 & 0 & 0 & 0 & 0 & -679/648 & 67/24 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & t_s & 3749/648 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & -55/324 & 223/108 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 223/108 & -55/324 \\
\end{array}
\right) \label{Js0result},
$}\\[2mm]
&\hat{J}_{s,1} =
\scalebox{1.0}{$
\left(
\begin{array}{cccccccccc}
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & -10/27 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\end{array}
\right),
$} \label{Js1result}\\[2mm]
&\hat{J}_{e,0} = \nonumber \\
&
\scalebox{0.65}{$
\left(
\begin{array}{cccccccccc}
-4/27 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & -4/27 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
-0.001708 & 0.004962 & 0.002631 & 0.009301 & -0.03258 & -0.08924 & 0.0004431 &
0 & -0.07795 & 0.002792 \\
-0.004694 & -0.001225 & 0.007331 & 0.01080 & -0.02781 & -0.07666 & -0.006646 &
0 & -0.01071 & -0.08131 \\
0.0004270 & 0.003537 & -0.001407 & 0.001703 & -0.008641 & -0.02351 & 0.2102 & 2/5 &
0.001344 & 0.004454 \\
-0.001829 & -0.004273 & 0.002924 & 0.0004802 & 0.004780 & 0.01280 & -0.04904 & - 8/135 &
-0.004205 & -0.006649 \\
2/15 & -2/135 & 2/135 & -2/15 & 59/270 & 19/90 & t_e
& 3 t_e - 50/81 & 26/135 & 2/45 \\
-0.02605 & 0.005587 & -0.01083 & 0.02081 & -0.02530 & 0.06671 & - t_e/3 + 38/729
& - t_e + 8/27 & -0.03366 & -0.002023 \\
0.09942 & 0.02428 & -0.1174 & -0.04438 & -0.1994 & -0.5362 & 2/35 & -8/45 & 0.05967
& 0.05861 \\
0.02623 & 0.02072 & 0.04112 & -0.1125 & -0.1951 & -0.5158 & -2/35& - 4/15 & 0.01879
& -0.06080 \\
\end{array}
\right),
$}\label{Je0result}\\[2mm]
&\hat{J}_{e,1} =
\scalebox{1.0}{$
\left(
\begin{array}{cccccccccc}
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 4/243 & -4/81 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 4/729 & 4/243 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\end{array}
\right),
$}\label{Je1result}\\[2mm]
&\hat{J}_{se,0} = \nonumber \\
&\scalebox{0.46}{$
\left(
\begin{array}{cccccccccc}
3/8 & 9/8& 0 &0 &
0 & 0 &
0 &
0&0 &
0 \\
-9/8& -3/8 &0 & 0 &0 &0 &
0 &
0 & 0 &
0 \\
-26.08 & 20.94 & -25.20 & 22.07 & 4.847 & 8.717 & 16.02 & 0.00499 & -26.20 & 20.63 \\
21.87 & -25.07 & 31.46 & -15.23 & -5.751 & -8.314 & 7.459 & 0.05014 & 16.21 & -30.05 \\
2.409 & 2.535 & -1.122 & -0.9967 & 0.06192 & -0.1911 & 2 t_s/5 + 142.6 & 0.02577 & 4.175 & 4.300 \\
-1.581 & -1.594 & 0.7172 & 0.7036 & 0.1306 & 0.1116 & -8 t_s/135 -51.94 & -2.417 & -2.729 & -2.743 \\
-15.68 & -11.02 & -59.91 & -55.25 & -309.3 & 8.235 & 0.08482 & 0.2545 & 7.761 & 11.53 \\
- 2 t_s/15 +5.611 &2 t_s/135 + 2.955 & -2 t_s/135 + 19.78 & 2 t_s/15 + 17.12 & -59 t_s/270 + 102.8 &-19 t_s/90 -3.773 & -28 t_s/243+ 0.4857 & -0.08482 &-26
t_s/135 -1.473 &-2 t_s/45 -4.129 \\
27.12 & -19.23 & 45.81 & -0.03029 & -8.332 & -7.461 & -8 t_s/45 + 1.621 & -0.3044 & 18.81 & -27.48 \\
-21.04 & 26.43 & -13.67 & 34.30 & 2.682 & 10.09 & -4 t_s/15 + 3.035 & 0.8012 & -26.07 & 21.45 \\
\end{array}
\right)
$}\nonumber \\
&
\scalebox{1.0}{$+ \hat{V}
\left(
\begin{array}{cccccccccc}
t_{se1} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & t_{se2} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & t_{se3} & t_{se4} & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & t_{se5} & t_{se6} & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & t_{se7} & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & t_{se8} & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & t_{se9} & t_{se10} & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & t_{se11} & t_{se12} & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & t_{se13} & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & t_{se14} \\
\end{array}
\right)
\hat{V}^{-1}
$},\label{Jse0result}\\[2mm]
&\hat{J}_{se,1} = \nonumber \\
&
\scalebox{0.65}{$
\left(
\begin{array}{cccccccccc}
-1.485 & -0.2623 & 0 & 0 & 0 & 0 & 0 & 0 &
0 & 0 \\
-0.2623 & -1.485 & 0 & 0 & 0 & 0 &0 & 0 &
0 &0 \\
-0.3914 & 0.9178 & -0.5086 & 0.8458 & 0.1026 & 0.1994 & 0 & 0 & -1.075 & 0.8226 \\
0.9599 & -0.2650 & 1.225 & -0.04511 & -0.1655 & -0.1095 & 0 & 0 & 0.6962 & -1.117 \\
-0.002595 & -0.04387 & -0.09552 & -0.1368 & 0.1728 & -0.03447 & -0.1481 & 0 & 0.04387 & 0.002595 \\
0.05517 & 0.000282 & 0.1661 & 0.1112 & -0.2131 & -0.3630 & 0.02195 & 0 & -0.000282 & -0.05517 \\
0 & 0 & 0 & 0 & 0 & 0 & - 4 t_s/81 + 1.985 & 0 & 0 & 0 \\
0.04938 & -0.005487 & 0.005487 & -0.04938 & 0.08093 & 0.07819 & 8 t_s/243 -0.9268 & 4 t_s/81 -0.9234 & 0.07133 & 0.01646 \\
0.8624 & -0.3145 & 0 & 0 & 0 & 0 & 0.06584 & 0 & -0.1909 & -0.7342 \\
-0.3145 & 0.8624 & 0 & 0 & 0 & 0 & 0.09877 & 0 & -0.7342 & -0.1909 \\
\end{array}
\right),
$}\label{Jse1result}\\[2mm]
&\hat{J}_{se,2} =
\scalebox{1.}{$
\left(
\begin{array}{cccccccccc}
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 20/2187 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 40/6561 & - 20/2187 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
\end{array}
\right),
$}\label{Jse2result}
\eeq
\beq
&\hat{J}_{ee,0} = \nonumber \\
&
\scalebox{0.51}{$
\left(
\begin{array}{cccccccccc}
40/729 &0 & 0 & 0& 0& 0 & 0 & 0 & 0& 0\\
0& 40/729 &0&0 & 0&0 &0 & 0 & 0 & 0 \\
-0.002519 & -0.003958 & 0.000955 & -0.005971 & 0.03395 & 0.09227 &12 t_e/27083 + 0.001188 & 36 t_e/27083+ 0.01576 & 0.02318 & -0.002951 \\
0.000504 & -0.001464 & -0.006771 & -0.003253 & 0.03333 & 0.09097 & - 180 t_e/27083
+ 0.01104 & - 540 t_e/27083 + 0.03144 & 0.004142 & 0.02686 \\
0.005995 & -0.003625 & 0.002241 & -0.007379 & 0.01478 & 0.02625 & 8026 t_e/104463 -0.1123 &8026 t_e/34821 -0.2909 & 0.007872 & -0.001747 \\
-0.001130 & 0.002451 & -0.001997 & 0.001584 & -0.003039 & -0.002932 & - 9178 t_e/313389 + 0.03477 & - 9178 t_e/104463 + 0.08429 & -0.000697 & 0.002884 \\
-0.02801 & 0.01209 & -0.01239 & 0.02771 & -0.09800 & -0.1928 & - 94 t_e/243 + 0.06658 & - 94 t_e/81 +0.2660 & -0.03582 & 0.004286 \\
0.008577 & -0.003761 & 0.004577 & -0.007761 & 0.01725 & 0.01575 & 110 t_e/729 -0.03402 & 110 t_e/243 -0.1293 & 0.01058 & -0.001761 \\
-0.02099 & -0.01189 & 0.02183 & -0.01845 & 0.1185 & 0.2984 &22 t_e/189 - 0.005245 & 22 t_e/63 + 0.02511 & 0.01247 & -0.008604 \\
-0.009687 & -0.01325 & -0.02604 & 0.01978 & 0.1295 & 0.3422 & 2 t_e/63 + 0.03922 & 2 t_e/21 + 0.1510 & -0.001510 & 0.02511 \\
\end{array}
\right),
$}\label{Jee0result}
\\[2mm]
&\hat{J}_{ee,1} =
\scalebox{1.0}{$
\left(
\begin{array}{cccccccccc}
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 16/2193723 & - 16/731241 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 80/731241 & 80/243747 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 32104/25384509 & - 32104/8461503 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 36712/76153527 & 36712/25384509 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 376/59049 & 376/19683 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 440/177147 & - 440/59049 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 88/45927 & - 88/15309 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & - 8/15309 & - 8/5103 & 0 & 0 \\
\end{array}
\right),
$}\label{Jee1result}
\eeq
where $t_s$, $t_e$, and $t_{se1,2,\dots,14}$ are the arbitrary
parameters of the matrix equations. Our convention for the matrix
$\hat{V}$ is $(\hat{\gamma}_{s,D}^{(0) T})_{1,1} \leq
(\hat{\gamma}_{s,D}^{(0) T})_{2,2} \leq \dots \leq
(\hat{\gamma}_{s,D}^{(0) T})_{10,10}$. Although Eq.~\eqref{Jse0result}
{makes explicit reference to the} the diagonal basis, the term
{involving $\hat{V}$} completely drops {out from} the evolution matrix
(see next subsection), and thereby our solution for the latter
does not require any matrix diagonalisation. Our
Eqs.~(\ref{eq:ext})--(\ref{eq:Jee0}) hold in any operator basis.
Moreover, if an ordinary four-dimensional basis transformation is
applied to Eqs.~(\ref{eq:q1})--(\ref{eq:q10}), the corresponding RG
matrices $\hat J_{\ldots}$ can be simply found by transforming those
in Eqs.~(\ref{Js0result})--(\ref{Jee1result}) in the same way as
$\hat{\gamma}_s^{(0)T}$. If the basis transformation is
$D$-dimensional, meaning that it involves evanescent operators, the
$\hat J_{\ldots}$ matrices undergo an additional scheme transformation
\cite{Herrlich:1994kh,Gorbahn:2004my}. We collect the solutions for
more than three active flavors in Appendix~\ref{app:solutionJ}.
Substituting the generalized ansatz of Eqs.~(\ref{eq:ext}),\,(\ref{eq:ext2}) into Eq.~(\ref{eq:UNLO}), we find the full next-to-leading order evolution matrix,
\begin{align}
\hat{U}_f (\alpha_{1},\alpha_{2})=&~ \hat{U}_0 \left( \alpha_1,
\alpha_2 \right) + \frac{\alpha_{1}}{4 \pi} \hat{U}_{QCD} \left(
\alpha_1, \alpha_2 \right) + \frac{\alpha_{EM}}{\alpha_{1}}
\hat{U}_{QED} \left( \alpha_1, \alpha_2 \right) \nonumber \\ &+
\frac{\alpha_{EM}}{4 \pi} \hat{U}_{QCD\textrm{-}QED} \left( \alpha_1,
\alpha_2 \right) + \left( \frac{\alpha_{EM}}{\alpha_{1}}\right)^2
\hat{U}_{QED\textrm{-}QED} \left( \alpha_1, \alpha_2 \right) \nonumber \\ & +
\mathcal{O}\left(\frac{\alpha_{EM}^2}{\alpha_s}, \alpha_s^2, \alpha_s
\alpha_{EM}, \alpha_{EM}^2\right),
\label{eq:Ufull}
\end{align}
where we use the abbreviation $\alpha_{1,2} \equiv \alpha_s (\mu_{1,2}
)$ for $\mu_1 < \mu_2$ and $\alpha_{EM} \equiv \alpha_{EM} (\mu_2)$ with
\begin{align}
\label{eq:UQCD}
\hat{U}_{QCD} \left( \alpha_1, \alpha_2 \right) & = \hat{J}_{s}(\alpha_{1}) \hat{U}_0 \left( \alpha_1, \alpha_2 \right) - \frac{\alpha_{2}}{\alpha_{1}} \hat{U}_0 \left( \alpha_1, \alpha_2 \right) \hat{J}_{s}(\alpha_{2}), \\
\label{eq:UQED}
\hat{U}_{QED} \left( \alpha_1, \alpha_2 \right) & = \hat{J }_{e}(\alpha_{1}) \hat{U}_0 \left( \alpha_1, \alpha_2 \right) - \frac{\alpha_{1}}{\alpha_{2}} \hat{U}_0 \left( \alpha_1, \alpha_2 \right) \hat{J }_{e}(\alpha_{2}), \\
\label{eq:UQCDQED}
\hat{U}_{QCD\textrm{-}QED} \left( \alpha_1, \alpha_2 \right) & = \hat{J}_{se}(\alpha_{1}) \hat{U}_0 \left( \alpha_1, \alpha_2 \right) - \hat{U}_0 \left( \alpha_1, \alpha_2 \right) \hat{J}_{se}(\alpha_{2}) \nonumber\\ & ~~~ + \hat{J}_{s}(\alpha_{1}) \hat{U}_{QED} \left( \alpha_1, \alpha_2 \right) - \frac{\alpha_{2}}{\alpha_{1}} \hat{U}_{QED} \left( \alpha_1, \alpha_2 \right) \hat{J}_{s}(\alpha_{2}),\\
\label{eq:UQEDQED}
\hat{U}_{QED\textrm{-}QED} \left( \alpha_1, \alpha_2 \right) &=
\hat{J}_{ee} \left( \alpha_1 \right) \hat{U}_0 \left( \alpha_1, \alpha_2 \right) - \frac{\alpha_1}{\alpha_2}
\hat{U}_{QED} \left( \alpha_1, \alpha_2 \right) \hat{J}_{e} \left( \alpha_2 \right) \nonumber \\
& ~~~
- \left( \frac{\alpha_1}{\alpha_2} \right)^2 \hat{U}_0 \left( \alpha_1, \alpha_2 \right) \hat{J}_{ee} \left( \alpha_2 \right) - \frac{\beta^e_0}{\beta_0} \left( 1 - \frac{\alpha_1} {\alpha_2} \right) \hat{J }_{e}(\alpha_{1}) \hat{U}_0 \left( \alpha_1, \alpha_2 \right).
\end{align}
\subsection{Cancellation of {Spurious} Parameters}
We now present some details of the cancellation of the arbitrary parameters.
First, we take a look at the $\mathcal{O}(\alpha_s)$ part of the evolution matrix in Eq.~\eqref{eq:Ufull},
\begin{align}
\frac{\alpha_1}{4 \pi} \hat{U}_{QCD} \left( \alpha_1, \alpha_2 \right) =&~ \frac{\alpha_1}{4 \pi} \hat{J}_{s,0} \hat{U}_0 (\alpha_1,\alpha_2) - \frac{\alpha_2}{4 \pi} \hat{U}_0 (\alpha_1,\alpha_2) \hat{J}_{s,0} \nonumber\\
& + \frac{\alpha_1 \ln \alpha_1}{4 \pi} \hat{J}_{s,1} \hat{U}_0 (\alpha_1,\alpha_2) - \frac{\alpha_2 \ln \alpha_2 }{4 \pi} \hat{U}_0 (\alpha_1,\alpha_2) \hat{J}_{s,1}.
\label{eq:UQCDdetail}
\end{align}
In the three-flavor regime, the matrix $\hat{J}_{s,0}$ in
Eq.~\eqref{Js0result} contains an undetermined component $t_s$.
Since the first and second term of $\hat{U}_{QCD} $ in
Eq.~\eqref{eq:UQCDdetail} depend on different scales, one naively could
argue that the cancellation of any dependence has to take place for
each term independently of the other. However, we will show that this is
not the case.
We locate the undetermined parameter in $[ \hat{J}_{s,0} ]_{8,7} = t_s$.
The matrix product $\hat{J}_{s,0} \hat{U}_0 (\alpha_1,\alpha_2)$
naturally contains a dependence on $t_s$ in the 8th row. Actually, this
dependence does cancel for all elements except for $[ \hat{J}_{s,0} U_0
(\alpha_1,\alpha_2) ]_{8,7} \supset ( \alpha_2 / \alpha_1 )^{1/9}
t_s$. The matrix product $\hat{U}_0 (\alpha_1,\alpha_2) \hat{J}_{s,0}$
in the second term of $\hat{U}_{QCD} $ naturally obtains the parameter
$t_s$ in the 7th column, and again the product consistently cancels this
dependence for all entries except for $[ \hat{U}_0 (\alpha_1,\alpha_2)
\hat{J}_{s,0} ]_{8,7} \supset ( \alpha_2 / \alpha_1 )^{-8/9} t_s$.
The full cancellations is thus only achieved by taking both terms of the
first line of Eq.~\eqref{eq:UQCDdetail} into account and takes the form
\begin{align}
\left[\frac{\alpha_1}{4 \pi} \hat{U}_{QCD} \left( \alpha_1, \alpha_2 \right) \right]_{8,7} & \supset \left[\frac{\alpha_1}{4 \pi} \hat{J}_{s,0} \hat{U}_0 (\alpha_1,\alpha_2) - \frac{\alpha_2}{4 \pi} \hat{U}_0 (\alpha_1,\alpha_2) \hat{J}_{s,0}\right]_{8,7} \nonumber\\
& \supset \frac{1}{4 \pi} \left(\alpha_1 \left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{1}{9}} - \alpha_2 \left(\frac{\alpha_2}{\alpha_1}\right)^{- \frac{8}{9}} \right) t_s \nonumber \\
& = 0.
\label{eq:cancel}
\end{align}
The reason that causes the singularity to arise - eigenvalues of $\hat{\gamma}^{(0)T}_{s}$ differing by $2 \beta_0$ in Eq.~\eqref{eq:singularityA} - is also responsible for the cancellation of the undetermined parameter between the high and low scales.
The difference of two eigenvalues of $\hat{\gamma}^{(0)T}_{s}$ by $2 \beta_0$ causes a difference of 1 in the exponents of $( \alpha_2 / \alpha_1 )$ and indeed
the spectrum of $\hat{\gamma}^{(0)T}_{s} / 2 \beta_0$ contains both $1/9$ and $-8/9$ as eigenvalues.
Thus, this difference allows the prefactors $\alpha_1$ and $\alpha_2$ of the first two terms in Eq.~\eqref{eq:UQCDdetail} to exactly cancel these terms between the different scales and entirely independent on the actual size of the scales.
Next, we focus on the arbitrary parameter $t_e$ which appears in
the matrix $\hat{J}_{e,0}$ in Eq.~\eqref{Je0result} in the three flavor
regime and must cancel in the $\hat{U}_{QED}$ part of the evolution
matrix. Let us denote the $t_e$-dependent piece of
$\hat{J}_{e,0}$ with $\hat{t}_e$, where $[\hat{t}_e]_{7,7}=t_e$,
$[\hat{t}_e]_{7,8}=3 t_e$, $[\hat{t}_e]_{8,7}=- t_e/3$,
$[\hat{t}_e]_{8,8}=- t_e$, and the other components are zero. Using the
matrix $\hat{V}$ it can be written as $\hat{t}_e = \hat{V} \hat{t}_e^\prime
\hat{V}^{-1}$, where $[\hat{t}_e^\prime]_{10,1} = - t_e $ and the other
components are zero. Then, in the evolution matrix, the $t_e$
dependence takes the following form:
\beq
\frac{\alpha_{EM}}{\alpha_1} \hat{U}_{QED} \left( \alpha_1, \alpha_2
\right) & \supset \alpha_{EM} \left( \frac{1}{\alpha_1} \hat{J}_{e,0}
\hat{U}_0 \left( \alpha_1, \alpha_2 \right) - \frac{1}{\alpha_2}
\hat{U}_0 \left( \alpha_1, \alpha_2 \right) \hat{J}_{e,0} \right)\nonumber \\ &
\supset \alpha_{EM} \hat{V} \left( \frac{1}{\alpha_1} \hat{t}_e^\prime
\hat{U}_{0,D} \left( \alpha_1, \alpha_2 \right) - \frac{1}{\alpha_2}
\hat{U}_{0,D} \left( \alpha_1, \alpha_2 \right) \hat{t}_e^\prime \right)
\hat{V}^{-1},
\label{eq:tecancel}
\eeq
where $\hat{U}_{0,D} \left( \alpha_1, \alpha_2 \right) $ is defined as
\beq
\hat{U}_ 0 \left( \alpha_1, \alpha_2 \right) =& \hat{V} \textrm{diag}
\left(\left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{\left(\hat{\gamma}_{s,D}^{(0)T}\right)_{1,1}}{2
\beta_0}},
\left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{\left(\hat{\gamma}_{s,D}^{(0)T}\right)_{2,2}}{2
\beta_0}},\dots,
\left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{\left(\hat{\gamma}_{s,D}^{(0)T}\right)_{10,10}}{2
\beta_0}} \right) \hat{V}^{-1}\\ \equiv& \hat{V} \hat{U}_{0,D}
\left( \alpha_1, \alpha_2 \right) \hat{V}^{-1}.
\eeq
All components except for $(10,1)$ of the parenthesis in
Eq.~\eqref{eq:tecancel} are zero trivially. The cancellation of the
$(10,1)$ component then proceeds in the same way as in the QCD case:
\beq
\left[ \frac{1}{\alpha_1} \hat{t}_e^\prime \hat{U}_{0,D} \left( \alpha_1, \alpha_2 \right) - \frac{1}{\alpha_2} \hat{U}_{0,D} \left( \alpha_1, \alpha_2 \right) \hat{t}_e^\prime \right]_{10,1}
& = \left( \frac{1}{\alpha_1} \left( \frac{\alpha_2}{\alpha_1} \right)^{-\frac{8}{9}} - \frac{1}{\alpha_2} \left( \frac{\alpha_2}{\alpha_1}\right)^{\frac{1}{9}} \right) \cdot \left( - t_e \right)\nonumber \\
& = 0.
\eeq
Therefore, the $t_e$ dependence of $\hat{U}_{QED}$ vanishes.
The cancellation of the parameters
$t_{se1,2,\dots,14}$ in the second matrix
product of Eq.~\eqref{Jse0result} is more trivial.
Let us define the second matrix product as $\hat{V} \hat{t}_{se}
\hat{V}^{-1}$. In the evolution matrix, the matrix $\hat{t}_{se}$
appears only in the $ \hat{U}_{QCD\textrm{-}QED}$ part and the
cancellation can be understood in the following way:
\beq
\frac{\alpha_{EM}}{ 4 \pi } \hat{U}_{QCD\textrm{-}QED} \left( \alpha_1,
\alpha_2 \right) &\supset \frac{\alpha_{EM}}{4 \pi} \left(
\hat{J}_{se,0} \hat{U}_0 (\alpha_1,\alpha_2) - \hat{U}_0
(\alpha_1,\alpha_2) \hat{J}_{se,0} \right)\\ & \supset
\frac{\alpha_{EM}}{4 \pi} \hat{V}\left[ \hat{t}_{se}, ~\hat{U}_{0,D}
\left( \alpha_1, \alpha_2 \right) \right] \hat{V}^{-1}\nonumber \\ & = 0,
\eeq
where we use the fact that $(\hat{\gamma}_{s,D}^{(0) T})_{3,3} =
(\hat{\gamma}_{s,D}^{(0) T})_{4,4}$ and $(\hat{\gamma}_{s,D}^{(0)
T})_{7,7} = (\hat{\gamma}_{s,D}^{(0) T})_{8,8}$ are pairwise
degenerate eigenvalues for any number of active flavors.
On the contrary,
the cancellation of $t_s$ arising in $ \hat{U}_{QCD\textrm{-}QED}$ and $t_e$ in $ \hat{U}_{QED\textrm{-}QED}$ is highly non-trivial.
The $t_s$ dependence, for example, resides in $\hat{J}_{s,0}$, $\hat{J}_{se,0}$ and $\hat{J}_{se,1}$ which appear in the matrix $ \hat{U}_{QCD\textrm{-}QED}$. Logarithmic $\alpha_s$ terms are accompanied by $\hat{J}_{se,1}$ and by the matrix products $\hat{J}_{s} \hat{U}_{ QED} $ and $\hat{U}_{ QED} \hat{J}_{s} $.
Although we do not give an analytic explanation for these cancellations in this paper,
we have checked that taking the sum of all terms in Eqs.~\eqref{eq:UQCDQED} and \eqref{eq:UQEDQED} eliminates any $t_s$ and $t_e$ dependence of $ \hat{U}_{QCD\textrm{-}QED}$ and $ \hat{U}_{QED\textrm{-}QED}$.
Now we have shown that the evolution matrix in Eq.~\eqref{eq:Ufull}
is independent of the undetermined parameters, so that we can set
them to arbitrary values from the beginning. These parameters are
directly related to the singular components in
Eqs.~\eqref{eq:singularityA},\,\eqref{eq:singularityB} of the
standard solution in the literature. Therefore, our method
automatically regularizes all singularities and these parameters
correspond to the choices of the finite pieces of the regulated
expressions, which can therefore be viewed as scheme parameters.
We have also found that the cancellation of the parameters occurs
between the high and low scales. This insight is especially important
when considering new physics at a high scale. The Wilson coefficients
for a given model are typically calculated at leading order only.
In the evolution to the scale of 1$\,$\text{GeV}\ appropriate for Kaon
physics one then usually neglects the corrections to
$\hat K^\prime$ in Eq.~\eqref{eq:UNLO} justified by the smallness of
$\alpha_s(\mu_2)$ compared to $\alpha_s(\mu_1)$. In the typical
applications in flavor physics, which do not involve corrections of
order $\alpha_{EM}$, this procedure is scheme-independent. We here
show that such a treatment is inconsistent in view of the cancellation
of the singularity regulating scheme parameters.
This inconsistency does not appear in the QCD and QED parts
which are nonsingular at $f=4,5,6$.
However the combined QCD-QED part, in which singularities persist
for all numbers of flavors, will yield results depending on unphysical
arbitrary scheme parameters if parts of the evolution matrix are
discarded in the described way. Instead, the pieces of $\hat
K^\prime$ which depend on the scheme parameters $t_{se}$ must be
consistently retained.
\subsection{Validation of the Logarithmic Contribution}
Finally, let us comment on the logarithmic contributions $\hat{J}_{s,1}$ and $\hat{J}_{e,1}$.
At the $\mathcal{O} (\alpha_s)$ part, we have the following logarithmic contributions to the evolution matrix,
\beq
\hat{U}_f \left( \alpha_1, \alpha_2 \right) & \supset \frac{\alpha_1}{4 \pi} \hat{U}_{ QCD} \left( \alpha_1, \alpha_2 \right) \nonumber \\
& \supset \frac{1}{4 \pi} \left( \alpha_1 \ln \alpha_1 \hat{J}_{s,1} \hat{U}_0 \left( \alpha_1, \alpha_2 \right) - \alpha_2 \ln \alpha_2 \hat{U}_0 \left( \alpha_1, \alpha_2 \right) \hat{J}_{s,1} \right)
\label{eq:logalphas}\\
& = \frac{\alpha_1}{4 \pi} \left(\frac{\alpha_2}{\alpha_1}\right)^{\frac{1}{9} } \ln \frac{\alpha_1}{\alpha_2} \hat{J}_{s,1}.
\label{eq:logalphas2}
\eeq
In the $\hat{J}_{s,1}$ matrix, the only nonzero component is $[
\hat{J}_{s,1}]_{8,7} = - 10/27$. Using a calculation parallel to the
one in the previous subsection, we find that the only nonzero component
in the matrix product $\hat{J}_{s,1}\hat{U}_0 \left( \alpha_1, \alpha_2
\right) $ is $ [\hat{J}_{s,1}\hat{U}_0 \left( \alpha_1, \alpha_2
\right)]_{8,7} = ( \alpha_2 / \alpha_1 )^{1/9} \cdot (- 10/27)$, and
similarly $ [\hat{U}_0 \left( \alpha_1, \alpha_2 \right)
\hat{J}_{s,1}]_{8,7} = ( \alpha_2 / \alpha_1 )^{- 8/9} \cdot (- 10/27)$.
Then, the $(8,7)$ component in the parenthesis in
Eq.~\eqref{eq:logalphas} becomes $- (10/27) \alpha_1 (\alpha_2 /
\alpha_1)^{1/9 } \ln (\alpha_1 / \alpha_2)$, and we arrive at
Eq.~\eqref{eq:logalphas2}. We find that this result is consistent with
Eq.\,(40) of Ref.~\cite{Adams:2007tk}, where, in order to regulate the
singularity, a small regulator $\epsilon$ is introduced in the
eigenvalues of $\hat{\gamma}_{s}^{(0)T}$.
{With a similar calculation for} the $\mathcal{O}
(\alpha_{EM}/\alpha_s)$ part we obtain the following term
\beq
\hat{U}_f \left( \alpha_1, \alpha_2 \right) & \supset
\frac{\alpha_{EM}}{\alpha_1 } \hat{U}_{ QED} \left( \alpha_1, \alpha_2
\right)\nonumber \\ & \supset \alpha_{ EM} \left( \frac{1}{\alpha_1} \ln
\alpha_1 \hat{J}_{e,1} \hat{U}_0 \left( \alpha_1, \alpha_2 \right) -
\frac{1}{\alpha_2} \ln \alpha_2 \hat{U}_0 \left( \alpha_1, \alpha_2
\right) \hat{J}_{e,1} \right)\nonumber \\ & = \alpha_{EM}
\left(\frac{1}{\alpha_1} \ln \alpha_1 \left( \frac{\alpha_2}{\alpha_1}
\right)^{- \frac{8}{9}} - \frac{1}{\alpha_2} \ln \alpha_2 \left(
\frac{\alpha_2}{\alpha_1} \right)^{\frac{1}{9}} \right)
\hat{J}_{e,1} \nonumber \\ & = \frac{\alpha_{EM}}{\alpha_1} \left(
\frac{\alpha_2}{\alpha_1} \right)^{- \frac{8}{9}} \ln
\frac{\alpha_1}{\alpha_2} \hat{J}_{e,1}
\end{equation}
This logarithmic contribution is also consistent with Eq.~(2.28) of
Ref.~\cite{Buras:1993dy}.
\subsection{Higher orders in $\alpha_{EM}$ and comparison with
Ref.~\cite{Huber:2005ig}}
The RG evolution in the pioneering
papers~\cite{Flynn:1989iu,Buchalla:1989we,Buras:1993dy} discards all
terms which are quadratic or higher-order in $\alpha_{EM}$. Our solution in
Eq.~\eqref{eq:Ufull} is correct to order $\alpha_{EM}^2/\alpha_s^2$, but
neglects terms of order $\alpha_{EM}^2/\alpha_s$ and higher. The extra
term is numerically unimportant for the SM analysis, but matters in
studies of new-physics contributions generated at very high scales,
where $\alpha_s$ is small. We come back to this point in
Sec.~\ref{sec:co}. The RG evolution derived in
Ref.~\cite{Huber:2005ig} considers terms quadratic in
$\alpha_{EM}$, including terms of order
$\alpha_{EM}^2/\alpha_s$ which we neglect. In particular, the $\mu$
dependence of $\alpha_{EM}$ affects the RG evolution at order
$\alpha_{EM}^2/\alpha^2_s$ and is therefore also included in
Ref.~\cite{Huber:2005ig}.
While Ref.~\cite{Huber:2005ig} addresses
$B$ decays, the derived formulae equally apply to $\epsilon_K^\prime$
and were used in Ref.~\cite{Buras:2015yba}. We argue that the
inclusion of $\alpha_{EM}^2/\alpha_s$ {terms} in the RGE does not improve the
prediction of $\epsilon_K^\prime/\epsilon_K$, because other terms of
the same order are not included in the standard NLO solution: For
instance, at this order the two-loop pure QED anomalous dimension
matrix $\hat{\gamma}_{e}^{(1)}$ must be added to $\hat{\gamma}
\left(g_s (\mu) \right)$ in Eq.~\eqref{eq:anomalousNLO}.
{Another issue are the $\Delta I=1/2$ operators %
\beq %
Q_{11} = \left(
\bar{s}_{\alpha} d_{\alpha} \right)_{V-A} \, \left(
\bar{b}_{\beta} b_{\beta} \right)_{V-A}, \qquad \qquad
Q_{12} = \left(
\bar{s}_{\alpha} d_{\beta} \right)_{V-A} \, \left(
\bar{b}_{\beta} b_{\alpha} \right)_{V-A}
\eeq %
which are generated by electroweak box diagrams, so that their Wilson
coefficients are of order $\alpha_{EM}$. In agreement with
Ref.~\cite{Buchalla:1989we} we find a small impact of these operators,
{contributing ($-0.07 \times 10^{-4}$) to $\epsilon_K^\prime/\epsilon_K$}.
Furthermore, this contribution dominantly comes from $A_2$ which is entered
by $Q_{11,12}$ through RG mixing triggered by $\hat{\gamma}_{e}^{(0)}$
and is thus ${\cal O} (\alpha_{EM}^2/\alpha_s)$ and to be discarded.
While the contribution of $Q_{11,12}$ to $A_0$ is formally part of
the NLO solution for $\epsilon_K^\prime/\epsilon_K$, it is numerically
completely negligible (contributing $-0.01 \times 10^{-4}$).}
We close this section by comparing our solution of the RG equations
in Eqs.~(\ref{eq:newJ})--(\ref{eq:Jee0}) to the one in
Ref.~\cite{Huber:2005ig}. Actually, the latter also regulates all the singularities {by} logarithmic terms, and uses the diagonalisation
of $\hat{\gamma}_{s}^{(0)}$ as described before Eq.~\eqref{eq:singularityA}.
The matrices $\hat J_{\ldots}$ transform into
$\hat J_{\ldots,D}\equiv\hat{V}^{-1} \hat J_{\ldots}
\hat{V}$ when passing to the diagonal basis.
Therefore Eqs.~(\ref{eq:Js1full})--(\ref{eq:Jee0})
also hold with the replacements $\hat{\gamma}_{s}^{(0)}
\to \hat{\gamma}_{s,D}^{(0)}$ and $\hat J_{\ldots}
\to \hat J_{\ldots,D}$. In this form one can
most easily compare our result with Eq.~(47) of Ref.~\cite{Huber:2005ig}.
The $\hat{U}_0$, $\hat{U}_{QCD}$, $\hat{U}_{QED}$,
$\hat{U}_{QCD-QED}$, and $\hat{U}_{QED-QED}$ correspond to
$\mathcal{O}(\omega^0 \lambda^0)$, $\mathcal{O}( \omega)$,
$\mathcal{O}(\lambda)$, $\mathcal{O}(\omega \lambda)$, and
$\mathcal{O}( \lambda^2)$ terms in Ref.~\cite{Huber:2005ig},
respectively. We have checked that our formulae of
the RG evolution matrices are numerically equivalent to those in
Ref.~\cite{Huber:2005ig}. We find that our
solution is easier to implement and leads to a faster numerical
evaluation.
\section{$\boldsymbol{\epsilon_{K}^{\prime}/\epsilon_{K}}$ in the
Standard Model at Next-to-Leading Order}
\label{sec:SM}
In this section, we evaluate $\epsilon_{K}'/\epsilon_{K}$ in the
Standard Model at next-to-leading order, using the evolution matrix
derived in the previous section.
We calculate the Wilson coefficients $v_i$ and $z_i$ in
Eqs.~\eqref{eq:wilsonv} and \eqref{eq:wilsonz} with {the methodology
of} Ref.\,\cite{Buras:1993dy}. Throughout this paper, the
$\overline{\textrm{MS}}$--NDR regularization scheme is used. For the
next-to-leading order RG evolution of the Wilson coefficients, we use
the {singularity-free} evolution matrix in Eq.~\eqref{eq:Ufull} and
systematically discard higher-order contributions.
Table~\ref{tab:wilson} shows our result of the Wilson coefficients at
$\mu = 1.3~\text{GeV}$, where $y_i \equiv v_i - z_i$. We decompose $y_i$ into
the {LO} contribution $\mathcal{O}(1)$ and the four
$\mathcal{O}(\alpha_{EM}/\alpha_s,\,\alpha_s,\,\alpha_{EM},\alpha^2_{EM}/\alpha^2_s)$
{NLO} {terms}, where $\mathcal{O}(1)$ {refers to} tree-level
$W$-boson exchange {combined with the one-gluon anomalous dimension
matrix $\hat{\gamma}_{s}^{(0)}$ in the RG evolution.} Here we take $
\alpha_s (M_Z) = 0.1185$, $\alpha_{EM}(M_W) = 1/128$, $m_t = 163.3~\text{GeV}$,
$m_b = 4.18~\text{GeV}$, and $\mu_c = 1.4 \; \textrm{GeV}$, which is the
threshold {scale} between three and four flavor effective theories
in Eqs.~\eqref{eq:wilsonv} and \eqref{eq:wilsonz}. Note that we
include $\ln (m_c^2/\mu_c^2) $ contributions {in} the charm quark
threshold correction $z_i (\mu_c)$ in Eq.~\eqref{eq:wilsonz}, where we
use $m_c= 1.275~\text{GeV}$ \cite{Agashe:2014kda}. {To calculate} $\alpha_s
(\mu)$ we use \texttt{RunDec:v1.0} with two-loop accuracy
\cite{Chetyrkin:2000yt}.
\begin{table}[t]
\begin{center}
\caption{Wilson coefficients at $\mu = 1.3\,\textrm{GeV}$, where the
7--10th components are divided by $\alpha_{EM}(M_W)$. $y_i$ {is}
decomposed into the {LO} contribution and the {individual}
{NLO corrections}.}
\label{tab:wilson}
\small{
\begin{tabular}{l | c |cccccc}
\hline
\hline
$i $ & $ z_i \left(\mu \right)$ & $y_i \left(\mu \right)$ & $ \mathcal{O}(1)$&$ \mathcal{O}(\alpha_{ EM}/\alpha_s)$&$\mathcal{O}(\alpha_s)$&$\mathcal{O}(\alpha_{EM})$& $\mathcal{O}(\alpha^2_{ EM}/\alpha^2_s)$\\
\hline
$1$ & $-0.3903$ & $ 0$& $ 0$& $ 0$& $ 0$& $ 0$ &$0$\\
$2$ & $1.200$ & $ 0$& $ 0$& $ 0$& $ 0$& $ 0$ &$0$\\
$3$ & $0.0044$ & $ 0.0275$ & $0.0254$ & $0.0001$& $0.0007$&$0.0012$ & $0$\\
$4$ & $-0.0131 $ & $ -0.0566$& $-0.0485$ & $ {-0.0002}$ & $-0.0069$ &$-0.0009$&$0$\\
$5$ & $0.0039$ & $ 0.0068$ & $0.0124$ & $ 0.0001$ & $-0.0059$ & $0.0001$&$0$\\
$6$ & $-0.0128 $ & $ -0.0847$ & $-0.0736$& $-0.0003$ & $-0.0099$ &$-0.0008$&$0$\\
$7/\alpha_{EM}$ & $ {0.0040}$ & $ {-0.0321}$ & $ 0$ & $ {-0.1116}$& $ 0$ & $ {0.0760}$&$ {0.0035}$\\
$8/\alpha_{EM}$ & $ {0.0019} $ & $ {0.1148}$& $ 0$ & $ {-0.0227} $& $ 0$& $ {0.1366}$&$ {0.0009}$\\
$9/\alpha_{EM}$ & $ {0.0051}$ & $ { -1.3815}$& $ 0$& $ {-0.1267}$& $ 0$& $ {-1.2581}$&$ {0.0034}$\\
$10/\alpha_{EM}$ & $-0.0013 $ & $ {0.4883}$& $ 0$ & $ {0.0217}$& $ 0$&$ {0.4672}$&$ {-0.0006}$\\
\hline \hline
\end{tabular}
}
\end{center}
\end{table}
Next {we take} the hadronic matrix elements from {a recent lattice
QCD calculation} \cite{Blum:2011ng,Blum:2012uk, Blum:2015ywa,
Bai:2015nea}, {using the} real parts ($CP$-conserving parts) of the
isospin amplitudes $A_{I=0,2} = \langle (\pi \pi)_{I = 0,2} \left|
\mathcal{H}^{|\Delta S| =1}_{\textrm{eff}} \right| K^0 \rangle $
{as additional constraints \cite{Buras:1993dy}. These amplitudes}
have been measured very precisely \cite{Blum:2015ywa}
\begin{equation}
\textrm{Re}A_0 &= \left( 3.3201\pm 0.0018 \right)\times 10^{-7}~\text{GeV},
\label{eq:ReA0}\\
\textrm{Re}A_2 &= \left(1.4787 \pm 0.0031\right) \times 10^{-8}~\text{GeV}.
\eeq
Since the real parts are dominated by Standard-Model tree-level
{coefficients} $z_2$ (see Table~\ref{tab:wilson}), they can be used
to fix one of the hadronic matrix elements $\langle (\pi \pi)_{I} \left|
Q_i \left( \mu \right) \right| K^0 \rangle$ $\equiv \langle Q_i \left(
\mu \right) \rangle_I $. $\langle Q_2 \rangle_0 $ {dominates the
real part of $A_0$, but contributes} to the imaginary part only
through the operator Fierz relations\footnote{The {Fierz} relation
for $Q_4$ is modified by ${\cal O}(\alpha_s/ 4 \pi)$ corrections
\cite{Buras:1993dy}, {but} these contributions are numerically
small \cite{Buras:2015yba}. } \beq Q_4 = - Q_1 + Q_2 + Q_3,
~~~~~~Q_{10} = Q_2 + \frac{1}{2} \left( Q_1 - Q_3 \right).
\label{eq:Fierz}
\eeq $\langle Q_1 \rangle_0 $ is the second largest contribution and the
remaining matrix elements are almost negligible. The situation is {more
handy} in the case of $A_2$, where the real part is parameterized
entirely by $ \langle Q_2 \rangle_2$ due to the fact that $ \langle Q_1
\rangle_2 = \langle Q_2 \rangle_2$ in pure QCD \cite{Buras:1993dy,
Buchalla:1989we}. In our analysis we derive values of $\langle Q_2
\rangle_0 $ and $\langle Q_2 \rangle_2$ {at the scale $\mu$} from the
experimental measurements of $\textrm{Re}A_0$ and $\textrm{Re}A_2$,
respectively\footnote{ On the other hand, once one introduces the
ratio \beq q = \frac{z_{+}(\mu) \left( \langle Q_2 \rangle_0 +
\langle Q_1 \rangle_0 \right)}{z_{-}(\mu) \left( \langle Q_2
\rangle_0 - \langle Q_1 \rangle_0 \right)}~~~~~\textrm{with~}
z_{\pm} (\mu) = z_{2}(\mu) \pm z_{1}(\mu), \eeq one can calculate
Im$A_{I}/$Re$A_{I}$ without using the fit of $\langle Q_2
\rangle_{I}$ to the data. Ref.~\cite{Buras:2015yba} {uses} this
strategy with {the} parameter range $0 \leq q \leq 0.1$.
Basically, the difference with our method {(corresponding} to the
$q$-dependent terms in Ref.~\cite{Buras:2015yba}) {only affects}
{numerically} subleading contributions (the $i=3,\,4,\,9,\,10$
components of Im$A_{0}/$Re$A_{0}$). In either method
the hadronic uncertainties are reduced compared to the choice to
take all matrix elements from lattice. }.
The decay amplitude of $K \to (\pi \pi)_{I=0}$ has been computed using a
$2+1$ flavor lattice QCD simulation at the renormalization scale $\mu
=1.531$ GeV \cite{Bai:2015nea}. In order to combine these matrix
elements with the Wilson coefficients evaluated in {the}
three-flavor regime ---that is, {at} a scale below the charm quark
mass--- we need to evolve the hadronic matrix elements {down to a}
scale below $ \mu_c$. The isospin amplitude is given as \beq A_I & =
\frac{G_F}{ \sqrt{2}} \lambda_u \langle \vec{Q} ( \mu_1)^T \rangle_I
\vec{C} (\mu_1)\nonumber \\ & = \frac{G_F}{ \sqrt{2}} \lambda_u \langle \vec{Q}
( \mu_1)^T \rangle_I \hat{U}_3 \left( \mu_1,\mu_2 \right)\vec{C}
(\mu_2)\nonumber \\ & = \frac{G_F}{ \sqrt{2}} \lambda_u \langle \vec{Q} (
\mu_2)^T \rangle_I \vec{C} (\mu_2), \eeq where $\mu_1 < \mu_2$ and $C_i
(\mu) \equiv z_i (\mu)+ \tau y_i (\mu) $. In the final line, we use the
fact that the physical amplitude $A_I$ is independent of the
renormalization scale, {so that
\beq
\langle \vec{Q} ( \mu_1)^T \rangle_I = \langle \vec{Q} ( \mu_2)^T
\rangle_I \left( \hat{U}_3 \left( \mu_1,\mu_2 \right) \right)^{-1}.
\label{eq:RGHME}
\eeq
In practice, we first evaluate the hadronic matrix elements for the
$I=0$ states at $\mu = 1.3$ GeV from the lattice results
\cite{Bai:2015nea} using a three flavor evolution matrix,
cf. Eq.~\eqref{eq:RGHME}. {Here we use $\alpha_s^{(3)} (1.531\,\text{GeV}) = 0.353388
$ {as in the lattice calculation of} Ref.~\cite{Bai:2015nea}.}
Then we determine $\langle Q_{2}
(\mu)\rangle_0 $ (and $\langle Q_{4,10} (\mu)\rangle_0 $ through
Eq.~\eqref{eq:Fierz}) from the experimental value of Re$A_0$ using the
Wilson coefficients shown in Table \ref{tab:wilson}. We have taken the
CKM parameters from CKMfitter \cite{Charles:2015gya}. The results are
shown in Table \ref{tab:bfactors}(a).
\begin{table}
\begin{center}
\caption{ The hadronic matrix elements (a),\,(b) and $B$ parameters
(c) extracted from the lattice calculations {for} $I=0$
\cite{Bai:2015nea} and $I=2$ \cite{Blum:2015ywa}. The experimental
values of the real parts of the amplitudes have been used
\cite{Blum:2015ywa}. The large errors result from the quoted
lattice errors on the hadronic matrix elements. The experimental
errors are small in comparison. We take $\mu = 1.3 \; \textrm{GeV}$. }
\label{tab:bfactors}
\subtable[]{
\begin{tabular}{lc}
\hline
\hline
$i $ & $\langle Q_i \left( \mu \right) \rangle_{0}^{{\overline{\textrm{MS}}\textrm{--NDR}}}$$\left( \text{GeV} \right)^3$ \\
\hline
$1$ & $ {-0.144} \pm 0.046$\\
$2$ & $0.105 \pm 0.015$\\
$3$ & $ {-0.040} \pm {0.068}$\\
$4$ & $ {0.210} \pm {0.069} $\\
$5$ & $ {-0.179} \pm 0.068$\\
$6$ & $ {-0.338} \pm {0.121}$\\
$7$ & $ {0.154} \pm 0.065$\\
$8$ & $ {1.540} \pm {0.372}$\\
$9$ & $ {-0.197} \pm {0.070}$\\
$10$ & $ {0.053} \pm {0.038}$\\
\hline \hline
\end{tabular}
}
\hspace{1cm}
\subtable[]{
\begin{tabular}{lc}
\hline
\hline
$i $ & $\langle Q_i \left( \mu \right) \rangle_{2}^{{\overline{\textrm{MS}}\textrm{--NDR}}}$$\left( \text{GeV} \right)^3$ \\
\hline
$1$ & $0.01006 \pm 0.00002$\\
$2$ & $0.01006 \pm 0.00002$\\
$3$ & ---\\
$4$ & ---\\
$5$ & ---\\
$6$ & ---\\
$7$ & $ {0.127} \pm {0.012}$\\
$8$ & $ {0.852} \pm {0.052}$\\
$9$ & $0.01509 \pm 0.00003$\\
$10$ & $0.01509 \pm 0.00003$\\
\hline \hline
\end{tabular}
}
\\
\vspace{0.5cm}
\subtable[]{
\begin{tabular}{lr}
\hline
\hline
$B_1^{(1/2)} \left(\mu \right)$ & $ {35.5} \pm 11.2$\\
$B_2^{(1/2)} \left(\mu \right)$ & $ {5.17} \pm 0.71$\\
$B_3^{(1/2)} \left(\mu \right)$ & $ {-3.27} \pm {5.60}$\\
$B_5^{(1/2)} \left(\mu \right)$ & $ {0.88}\pm {0.33}$\\
$B_6^{(1/2)} \left(\mu \right)$ & $ {0.56} \pm {0.20}$\\
$B_7^{(1/2)} \left(\mu \right)$ & $ {0.24} \pm {0.10}$\\
$B_8^{(1/2)} \left(\mu \right)$ & $ {0.98} \pm {0.24}$\\
\hline
$B_1^{(3/2)} \left(\mu \right)$ & $0.437 \pm 0.001$\\
$B_7^{(3/2)} \left(\mu \right)$ & $ {0.37} \pm {0.03}$\\
$B_8^{(3/2)} \left(\mu \right)$ & $ {0.77} \pm {0.05}$\\
\hline \hline
\end{tabular}
}
\end{center}
\end{table}
The decay amplitude of $K \to \left( \pi \pi \right)_{I =2}$ has also
been computed using a $2+1$ flavor lattice QCD simulations, albeit at
the scale $\mu= 3.0$ GeV \cite{Blum:2011ng,Blum:2012uk, Blum:2015ywa}.
According to Ref.~\cite{Blum:2012uk}, one can extract the lattice
results in an operator basis renormalized by the
$\overline{\textrm{MS}}$--NDR regularization scheme. From
Ref.~\cite{Blum:2015ywa}, which is the latest lattice QCD calculation
for $ I = 2$, we obtain \beq
\mathcal{M}_{(27,1)}^{\overline{\textrm{MS}}\textrm{--NDR} } (3\,\text{GeV}) & = 3 \sqrt{3} \langle Q_1 (3\,\text{GeV}) \rangle_2 = 0.0502 \pm 0.0031~(\text{GeV})^3\\
\mathcal{M}_{(8,8)}^{\overline{\textrm{MS}}\textrm{--NDR} } (3\,\text{GeV}) &= 2 \sqrt{3} \langle Q_7 (3\,\text{GeV}) \rangle_2 = 0.993 \pm 0.038~(\text{GeV})^3,\\
\mathcal{M}_{(8,8)_{\textrm{mix}}}^{\overline{\textrm{MS}}\textrm{--NDR}
} (3\,\text{GeV}) &= 2 \sqrt{3} \langle Q_8 (3\,\text{GeV}) \rangle_2 = 4.547 \pm
0.275 ~(\text{GeV})^3, \eeq where the results of the
$(\Slash{\it{q}},\Slash{\it{q}})$ intermediate scheme are taken as
central value, while the results of the $(\gamma^{\mu}, \gamma^{\mu})$
scheme are taken as uncertainty. Using the three flavor evolution
matrix in Eq.~\eqref{eq:RGHME}, we obtain the hadronic matrix elements
at $\mu = 1.3$ GeV for the $I = 2$ states. Here, we use the lattice input $\alpha_s$ value: $\alpha_s^{(3)} (3\,\text{GeV}) = 0.24544
$ \cite{Blum:2012uk}.
Then, from the experimental
value of Re$A_2$ we determine $\langle Q_{2} (\mu)\rangle_2 $ (and
$\langle Q_{1,9,10} (\mu)\rangle_2 $ through Eq.~\eqref{eq:Fierz},
$\langle Q_{1} (\mu)\rangle_2 = \langle Q_{2} (\mu)\rangle_2 $ and $Q_9
= \frac{1}{2} \left( 3 Q_1 - Q_3\right)$ which is a Fierz relation).
The results are shown in Table~\ref{tab:bfactors}(b). Note that through
the evolution matrices $\hat{U}_{QED} $ and $\hat{U}_{QCD\textrm{-}QED}$
this procedure generates small nonzero values of $\langle
Q_{\textrm{3--6}} (\mu) \rangle_2$, which are regarded as
non-electroweak penguin contributions to Im$A_2$. Since the lattice
simulations have not calculated them at $3.0$ GeV, one should not use
them at the lower hadronic scale $\mu$. On the other hand, they have
been calculated {with} chiral perturbation theory
\cite{Cirigliano:2003nn, Cirigliano:2003gt} and are included in the
isospin-violating corrections $\hat{\Omega}_{\textrm{eff}}$ of
Eq.~\eqref{eq:epsilonequation}\footnote {The non-electroweak penguin
contributions are calculated at $\mu = 1.0 \pm 0.3$ GeV
\cite{Cirigliano:2003nn, Cirigliano:2003gt}. }. Therefore, we
{have} decided to omit these contributions at the hadronic scale $\mu$.
To compare with the literature, we also extract $B$ parameters
from the hadronic matrix elements in Table~\ref{tab:bfactors}(c). These
$B $ parameters are defined as {in Ref.~}\cite{Buras:2015yba}:
\beq
\langle Q_6 \left( \mu \right) \rangle_0 & = -4 \sqrt{\frac{3}{2}}
\left( \frac{m_K^2}{m_s (\mu) + m_d ( \mu) }\right)^2 \left( F_K - F_{\pi} \right) B^{(1/2)}_6 \left(\mu\right),\\
\langle Q_8 \left( \mu \right) \rangle_2 & = \sqrt{3} \left(
\frac{m_K^2}{m_s (\mu) + m_d ( \mu) }\right)^2 F_{ \pi} B^{(3/2)}_8
\left(\mu \right), \eeq
All other {{$B$ parameters} are defined in Ref.~\cite{Buras:1993dy}.
For running quark masses, we use {the} lattice results $m_s
(2\,\text{GeV})$ $= 93.8 (2.4)~\,{\rm MeV}$ and $m_d (2\,\text{GeV}) = 4.68 (16) $~MeV
with the three-flavor RG evolution \cite{Aoki:2013ldr}. Since the
uncertainty from the strange quark mass is already included in the
lattice results of $\langle {Q}_i \rangle_I$ as one of the
systematic errors, we {do} not include it in the estimation of
uncertainties of the $B$ parameters. The $B$ parameters are
consistent with Ref.~\cite{Buras:2015yba}, and we also confirmed
the
almost $\mu$-independent behavior of $B_6^{(1/2)}(\mu)$ and
$B_8^{(3/2)}(\mu)$ \cite{Buras:1993dy}. Note that in the following
analysis we {will} directly use the hadronic matrix elements $\langle
Q_i \rangle_I$ {rather than the} $B$ parameters.
Finally we combine the {short-distance} and long-distance contributions.
The master equation of $\epsilon_{K}'/\epsilon_{K}$ is given {in}
Eq.~\eqref{eq:epsilonequation}. Since the isospin-violating correction
by the electroweak penguins to Im$A_0$ are already subtracted from
$\hat{\Omega}_{ \textrm{eff}}$ as $\langle Q_{\textrm{7--10}}
\rangle_0$, {one should evaluate the last term in
Eq.~\eqref{eq:epsilonequation} as}
\begin{equation}
\left( 1- \hat{\Omega}_{\textrm{eff}}\right)
\textrm{Im}A_0 = \left( 1- \hat{\Omega}_{\textrm{eff}}\right) \left(
\textrm{Im}A_0\right)^{\textrm{others}} + \frac{1}{a} \left(
\textrm{Im}A_0\right)^{\textrm{EWP}}, ~~~~a = 1.017,
\label{eq:Omega}
\eeq {with the two terms representing the contributions from}
$\langle Q_{\textrm{3--6}}\rangle_0$ and $\langle
Q_{\textrm{7--10}}\rangle_0$, {respectively} \cite{Buras:2015yba}.
In addition, the experimental values of Re$A_0$ in Eq.~\eqref{eq:ReA0}
and $| \epsilon_{K} | = 2.228 \times 10^{-3}$ \cite{Agashe:2014kda} are
used. Our result for $ \epsilon_{K}'/\epsilon_{K}$ in the Standard
Model at the next-to-leading order is
\beq %
\left(\frac{\epsilon_{K}'}{\epsilon_{K}} \right)_{\textrm{SM--NLO}}=
\left( {1.06}\pm {4.66_{\rm{Lattice}}} \pm {1.91_{\rm{NNLO}}} \pm {0.59_{\rm{IV}}} \pm { 0.23_{m_t}} \right) \times 10^{-4}.
\label{eq:smnlo}
\eeq %
The first {error originates} from the lattice-QCD simulations
\cite{Blum:2015ywa, Bai:2015nea} {and is} dominated by the
uncertainty stemming from $ \langle Q_6 \rangle_0 $ {(which is} $\pm
{4.52} \times 10^{-4}$) (see Figure~\ref{fig:A0pies}(c)). The
uncertainties from $ \langle Q_3\rangle_0 $ through Eq.~\eqref{eq:Fierz}
and from $ \langle Q_8\rangle_2 $ are subleading ($\pm {0.77} \times
10^{-4}$ and $\pm {0.56} \times 10^{-4}$, respectively).
The second uncertainty comes from {perturbative} higher-order
corrections, {which we estimate in two ways.} Firstly, we estimate
uncertainties from higher-order corrections to the Wilson coefficients
{by calculating} the RG evolution of the Wilson coefficients with a
different method. Instead of using the analytic evolution matrices
formulated in Sec.~\ref{section2}, we solve the {corresponding} set
of differential equations numerically
\begin{equation}
\frac{d \vec{v} (\mu)}{d \ln \mu} = \hat{\gamma}^T(g_s (\mu)) \vec{v}
(\mu),~~~~\frac{d \vec{z} (\mu)}{d \ln \mu} = \hat{\gamma}^T(g_s (\mu))
\vec{z} (\mu)
\eeq
Since this RG evolution contains higher-order {(namely
$\mathcal{O}(\alpha_s^2, \alpha_s \alpha_{EM})$)} corrections, the
result is interpreted as {a conservative estimate of the}
uncertainty in the short-distance contributions. As a result, we find
that the Wilson coefficients are shifted by about 10 percent compared
with Table~\ref{tab:wilson}, and we obtain $ \epsilon_{K}' /\epsilon_{K}
= {-0.32}\times 10^{-4}. $ Hence, we estimate that the uncertainty from
higher-order corrections is $ \pm {1.38} \times 10^{-4}$.
\begin{figure}[t]
\begin{center}
\subfigure[$\mu_c$ dependence of $\epsilon'_K/\epsilon_K$ ]
{
\includegraphics[width=0.45\textwidth, bb = 0 0 360 234]{muc_v2.pdf}
}~~~
\subfigure[$\mu$ dependence of $\epsilon'_K/\epsilon_K$]
{
\includegraphics[width=0.45\textwidth, bb = 0 0 360 234]{mu_v2.pdf}
}
\\
\subfigure[$\mu$ dependence of Im$A_0$]
{
\includegraphics[width=0.45\textwidth, bb = 0 0 360 227]{A0_mu_v2.pdf}
}~~~
\subfigure[$\mu$ dependence of Im$A_2$]
{
\includegraphics[width=0.45\textwidth, bb = 0 0 360 227]{A2_mu_v2.pdf}
}
\caption{
{(a) The $\mu_c$ dependence of $\epsilon'_{K}/\epsilon_K$ in the range $1.3 < \mu_c < {3.0}$ GeV with $\mu = 1.3$ GeV.}
The $\mu$ dependence of $\epsilon'_{K}/\epsilon_K$ (b), Im$A_0$ (c) and Im$A_2$ (d) in the range $ {0.8} < \mu < 1.4 $ GeV with $\mu_c = 1.4$ GeV.
}
\label{fig:mucdep}
\end{center}
\end{figure}
{Secondly,} we have investigated the $\mu_c$ and $\mu$ dependences of $
\epsilon_{K}'/\epsilon_{K}$. In Fig.~\ref{fig:mucdep}(a), we show the
$\mu_c$ dependence of $ \epsilon_{K}'/\epsilon_{K}$ in the range $1.3 <
\mu_c < {3.0}$ GeV with fixed $\mu = 1.3$ GeV. In
Fig.~\ref{fig:mucdep}(b), we vary $\mu$ with $\mu_c$ fixed at $1.4$ GeV.
We find that the $\mu$ dependence is {small}, $\pm {0.77} \times
10^{-4}$, while the $\mu_c$ dependence is slightly larger, $\pm {1.09}
\times 10^{-4}$. The {scale $\mu$ enters the prediction in three ways:}
{First, the decomposition of the} isospin-violating corrections in
Eq.~\eqref{eq:Omega} {is imposed at this scale. Second, the omitted}
non-electroweak penguin contributions to Im$A_2$ {depend on $\mu$}, and
third, the experimental values of Re$A_0$ and Re$A_2$ to fix $\langle
Q_{2} (\mu)\rangle_2 $ and $\langle Q_{2} (\mu)\rangle_0 $ {are imposed
at} the hadronic scale $\mu$. In this process, we double-count the
uncertainty from the isospin-violating contributions, however, we find
that these uncertainties are very small compared with the uncertainties
stemming from lattice and thus we have not investigated them any
further. We show the $\mu$ dependences of Im$A_0$ {(and \emph{not} the
$\mu$ dependence} of $(1-\hat{\Omega}_{\textrm{eff}})$Im$A_0$) and
Im$A_2$ in Figs.~\ref{fig:mucdep}(c) and (d), respectively. We add
the three uncertainties in quadrature. Strictly speaking, this
double-counts some pieces of the unknown higher-order corrections.
The third uncertainty {in Eq.~\eqref{eq:smnlo}} stems from
isospin-violating corrections \cite{Cirigliano:2003nn,
Cirigliano:2003gt}, such as strong isospin violation $(m_u \neq m_d)
$, non-electroweak penguin transitions in the $I=2$ state and $\Delta I
= 5/2$ corrections \cite{Gardner:2000sb,Cirigliano:2000zw}. The
uncertainty is dominated by the non-electroweak penguin contributions to
Im$A_2$, however, the uncertainty in $\epsilon_{K}'/\epsilon_{K}$ is
small.
The {last} uncertainty {in Eq.~\eqref{eq:smnlo}} comes from the
running mass of the top quark~$m_t ( m_t)$ $= 163.3 \pm 2.7$ GeV
\cite{Alekhin:2012py}. Since the other uncertainties we have not
elaborated here are negligibly small according to
Ref.~\cite{Buras:2015yba}, we have omitted them in our error
{estimate}. Therefore, our final result is \beq
\left(\frac{\epsilon_{K}'}{\epsilon_{K}} \right)_{\textrm{SM--NLO}}=
\left( {1.06} \pm {5.07} \right) \times 10^{-4}, \eeq which is consistent
with Refs.~\cite{Buras:2015yba} and \cite{Bai:2015nea}.
On the other hand, it is well-known that the experimental {value} is much larger \cite{Gibbons:1993zq,Barr:1993rx,AlaviHarati:1999xp,Fanti:1999nm,Batley:2002gn, Abouzaid:2010ny}.
The current world average is \cite{Agashe:2014kda},
\beq
\textrm{Re}\left( \frac{\epsilon_{K}'}{\epsilon_{K}}\right)_{\textrm{exp}} = \left( 16.6 \pm 2.3 \right) \times 10^{-4}.
\eeq
We observe that our prediction of $\epsilon_{K}' /\epsilon_{K}$ in the
Standard Model is $ {2.8}\,\sigma$ below the experimental value. This
small Standard Model prediction and thus the large tension is supported
by the large-$N_c$ {``dual QCD''} approach \cite{Buras:1987wc,
Bardeen:1986uz,Bardeen:1986vz,Buras:2014maa,Buras:2015xba,Buras:2016fys},
which is an entirely different approach to low energy QCD than lattice
gauge theory. {There has been a dispute concerning the role of
final-state interactions (FSI) for the size of $ \langle Q_6 \rangle_0
$, with the chiral perturbation community favouring an enhancement of
$ \langle Q_6 \rangle_0 $ by FSI \cite{Pallante:1999qf} and an
opposing view of the large-$N_c$ community \cite{Buras:2016fys}.
Modern lattice calculations do include FSI \cite{Lellouch:2000pv} and
will speak the final word on FSI.} Since the main uncertainty {of the
SM prediction for $\epsilon_{K}' /\epsilon_{K}$} comes from
statistical and systematical errors in the lattice calculation of the
hadronic matrix elements for $A_0$, {the expected progress in this field
will sharpen the Standard Model prediction in the near future
\cite{Bai:2015nea}.}
We note that in absence of a lattice result for the hadronic matrix
element and the smallness of the corresponding Wilson coefficient, we
omit the contribution from the chromomagnetic penguin operators $Q_{8g}
= m_s g_s/(16 \pi^2) \overline{s} T^a \sigma_{\mu \nu} (1 - \gamma_5) d
G^{\mu \nu\,a}$ (and {the opposite-chirality analogue}
$\tilde{Q}_{8g}$). According to Ref.~\cite{Buras:2015yba},
{chromomagnetic penguins} contribute
$|0.2$--$0.7|\times10^{-4}$\footnote{The sign depends on the sign of the
hadronic matrix element. The preliminary lattice calculation of
$\langle \pi | Q_{8g} | K \rangle$ \cite{Lubicz:2014qfa} and
calculations in the chiral quark model \cite{Bertolini:1993rc,
Deshpande:1994vp, Bertolini:1994qk} imply that a contribution to
$\epsilon'_K / \epsilon_K$ is positive at the leading order. However,
next-to-leading order contributions to $\langle (\pi \pi)_{I=0} |
Q_{8g} | K^{0} \rangle$ are expected to mess up the leading order
estimate because of a parametric enhancement $\propto 1/N_c \cdot
m_K^2 / m_{\pi}^2$ \cite{Barbieri:1999ax, Buras:1999da}.} to
$\epsilon'_K/\epsilon_K$, which rather small compared with {the}
{QCD-penguin} and QED-penguin contributions (see
Figure~\ref{fig:A0pies}(c)). Even if we {add} this contribution as $
+ 0.7\times10^{-4}$ to the central value (to the higher-order
uncertainty) of $\epsilon'_K/\epsilon_K$, the discrepancy still persists
at $2.7\,(2.8)\,\sigma$.
\begin{figure*}[]
\subfigure[$ \textrm{Im}A_0 $ ]{
\includegraphics[width=0.32\textwidth, bb = 0 0 267 267]{pi_imA0_v2.pdf}
}
\hspace{-1.1cm}
\subfigure[$\textrm{Im}A_2 $]{
\includegraphics[width=0.32\textwidth, bb = 0 0 265 265]{pi_imA2_v2.pdf}
}
\hspace{-1.5cm}
\subfigure[$ \epsilon_{K}'/\epsilon_{K}$]{
\includegraphics[width=0.45\textwidth, bb = 0 0 372 262]{pi_epsilon_v2.pdf}
}
\caption{Composition of Im$A_0$, Im$A_2$ and $
\epsilon_{K}'/\epsilon_{K}$ with respect to the operator basis. We
take $\mu = 1.3$ GeV. In subfigure (c), the right (left) side of the
dashed line represents positive (negative) contributions. }
\label{fig:A0pies}
\end{figure*}
In Fig.~\ref{fig:A0pies} we show the composition of Im$A_0$, Im$A_2$
and $ \epsilon_{K}'/\epsilon_{K}$ with respect to the operator basis.
We observe that the positive dominant contribution to $
\epsilon_{K}'/\epsilon_{K}$ comes from $Q_6$ while $Q_9$ is
subdominant. The dominant negative contribution comes from $Q_8$ while
$Q_4$ is subdominant. Remarkably, their sum almost cancels at
next-to-leading order. This leads to an extremely small {central
value of the} Standard Model prediction for $
\epsilon_{K}'/\epsilon_{K}$.
Although the results of the Wilson coefficients
by themselves are slightly different when compared to the result of
Ref.~\cite{Buras:2015yba}, the products {with the hadronic matrix elements} are well consistent\footnote{
Indeed, the values of $ {y_6 \langle Q_6 \rangle_0 }$ and $ {y_8 \langle Q_8 \rangle_2}$ are in
good agreement with Ref.~\cite{Buras:2015yba}.}. The main difference
{between this reference and our analysis} is in the subleading
contributions. In Ref.~\cite{Buras:2015yba}, the hadronic matrix
elements $\langle Q_3 (\mu) \rangle_0$, $\langle Q_5 (\mu) \rangle_0$
and $\langle Q_7 (\mu) \rangle_0$ are set to be $0$ as central values,
while we have evaluated them from the lattice data. The numerical
difference in $\epsilon'_K/\epsilon_K$ is $\sim -1\times 10^{-4} $. We
also find that the contribution of
$\mathcal{O}(\alpha_{EM}^2/\alpha^2_s)$ terms, which has not been
considered in the literature so far, only contributes to $\epsilon'_K /
\epsilon_K $ as {little} as $ {-0.10} \times 10^{-4}$. {This term,
however, {can} be relevant in new-physics models with TeV-scale
isospin violation.}
\section{Beyond the Standard Model}
\label{sec:NP}
\subsection{Preliminaries}
Upon integrating out heavy degrees of freedom in models of new physics,
{new} contributions to Wilson coefficients of the Standard Model
operators $Q_i$ (and their opposite-chirality {analogues}
$\tilde{Q}_i$) arise.
As we have shown in the previous section, the Standard Model prediction
of $\epsilon'_K/\epsilon_K$ is significantly below the experimental
data. Although the discrepancy is only $ {2.8}\,\sigma$ at present, its
confirmation {with higher significance} by future lattice {results}
{may establish a} footprint of new physics.
Indeed, several new physics models can alleviate the
$\epsilon'_K/\epsilon_K$ tension, like generic flavor-violating $Z$ and
$Z^{\prime}$ models \cite{Buras:2014sba,Buras:2015yca,Buras:2015jaq},
331 models \cite{Buras:2014yna,Buras:2015kwd,Buras:2016dxz}, the
Littlest Higgs model with $T$-parity \cite{Blanke:2015wba},
flavor-violating additional pseudo-scalar models \cite{Goertz:2015nkp},
and the Minimal Supersymmetric Standard Model
\cite{Tanimoto:2016yfy,Kitahara:2016otd}.
Since $\epsilon_{K}^{\prime} / \epsilon_{K}$ is linear in the Wilson
coefficients, the {SM and new-physics contributions are simply
additive:}
\begin{align}
\frac{\epsilon_{K}^{\prime}}{\epsilon_{K}} & = \left(
\frac{\epsilon_{K}^{\prime}}{\epsilon_{K}} \right)_{\textrm{SM}} +
\left( \frac{\epsilon_{K}^{\prime}}{\epsilon_{K}} \right)_{\textrm{NP}} .
\end{align}
Using the following effective Hamiltonian for the new physics
contributions, %
\beq%
\mathcal{H}^{|\Delta S| = 1}_{\textrm{eff,~NP}} = \frac{G_F}{\sqrt{2}}
\sum_{i=1}^{10} \left( Q_i (\mu) s_i (\mu) + \tilde{Q}_i (\mu)
\tilde{s}_i (\mu) \right) + \textrm{H.c.},
\label{eq:effHamiNP}
\eeq %
where the opposite-chirality operators $\tilde{Q}_i $ are {found
from} $Q_i$ by interchanging $V-A \leftrightarrow V+A $, the new
physics contribution is given by
\begin{equation}
\left( \frac{\epsilon_{K}^{\prime}}{\epsilon_{K}} \right)_{\textrm{NP}}
&= \frac{G_F \omega_{+}}{2 \left| \epsilon_{K}^{\textrm{exp}}\right|
\textrm{Re} A_0^{\textrm{exp}}} \nonumber \\ & \times \left[
\frac{1}{\omega_{+}} \langle \vec{Q} ( \mu)^T \rangle_2 \textrm{Im}
\left[ \vec{s}(\mu) - \vec{\tilde{s}} (\mu) \right] - \langle \vec{Q}
( \mu)^T \rangle_0 (1-\hat{\Omega}_{\textrm{eff}}) \textrm{Im} \left[
\vec{s}(\mu) - \vec{\tilde{s}} (\mu) \right] \right]\nonumber \\ & =
\frac{G_F \omega_{+}}{2 \left| \epsilon_{K}^{\textrm{exp}}\right|
\textrm{Re} A_0^{\textrm{exp}}} \left[ \frac{1}{\omega_{+}} \langle
\vec{Q} ( \mu)^T \rangle_2 - \langle \vec{Q} ( \mu)^T \rangle_0
(1-\hat{\Omega}_{\textrm{eff}}) \right] \textrm{Im} \left[
\vec{s}(\mu) - \vec{\tilde{s}} (\mu) \right],\nonumber \\ & = \frac{G_F
\omega_{+}}{2 \left| \epsilon_{K}^{\textrm{exp}}\right| \textrm{Re}
A_0^{\textrm{exp}}} \langle \vec{Q}_{\epsilon'_K} ( \mu)^T \rangle
\textrm{Im} \left[ \vec{s}(\mu) - \vec{\tilde{s}} (\mu) \right] \nonumber \\ & =
\frac{G_F \omega_{+}}{2 \left| \epsilon_{K}^{\textrm{exp}}\right|
\textrm{Re} A_0^{\textrm{exp}}} \langle \vec{Q}_{\epsilon'_K} ( \mu)^T
\rangle \hat{U} \left(\mu, \mu_{\textrm{NP}} \right) \textrm{Im} \left[
\vec{s}(\mu_{\textrm{NP}}) - \vec{\tilde{s}} (\mu_{\textrm{NP}})
\right],
\label{eq:NPeps}
\eeq
where the isospin-violating correction in Eq.~\eqref{eq:Omega} is
\begin{eqnarray}
\left(1-\hat{\Omega}_{\textrm{eff}} \right)_{ij} =
\left\{ \begin{array}{ll}
0.852 & (i=j=\textrm{1--6}) \\
0.983 & (i=j=\textrm{7--10})\\
0 & (i \neq j),
\end{array} \right.
\end{eqnarray}
and we employed $\langle \tilde{Q}_i ( \mu) \rangle_I = - \langle Q_i ( \mu) \rangle_I$ and defined $ \langle \vec{Q}_{\epsilon'_K} \rangle$ as
\beq
\langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \equiv
\frac{1}{\omega_{+}} \langle \vec{Q} ( \mu)^T \rangle_2 - \langle \vec{Q} ( \mu)^T \rangle_0 (1-\hat{\Omega}_{\textrm{eff}}).
\eeq
The evolution matrix in Eq.~\eqref{eq:NPeps} is given by
\beq
\hat{U} \left(\mu, \mu_{\textrm{NP}} \right) \equiv \hat{U}_3 \left(\mu, \mu_{c} \right) \hat{M}_c \left(\mu_c\right) \hat{U}_4 \left(\mu_c,m_b \right) \hat{M}_b (m_b)
\hat{U}_5 \left(m_b, m_t \right) \hat{M}_t (m_t) \hat{U}_6 \left(m_t, \mu_{\textrm{NP}} \right),
\label{eq:U3456}
\eeq Since the matching matrices depend only on the difference of the
number of active up- and down-type quark flavors, we take $\hat{M}_t
(m) = \hat{M}_c (m)$. Note that the RG evolution {of the}
opposite-chirality operators is {the} same as for the Standard Model
operators and that these {two sets of operators do not mix with each
other}. We also note that the chromomagnetic operators are omitted in
our analysis.
In this section, we give a useful formula for the new physics
contributions to $\epsilon'_K / \epsilon_K$ considering the analytic
solutions of the next-to-leading order evolutions matrices and the
hadronic matrix elements we derived. We note that we omit the weak
boson exchanges in the RG evolutions from $\mu_{\textrm{NP}}$ to $M_W$,
where $\mu_{\textrm{NP}}$ represents the matching scale between the new
physics and the effective Hamiltonian in Eq.~\eqref{eq:effHamiNP}.
{Like the} photon exchanges {one should treat} weak boson exchanges as
next-to-leading contributions. {Note that} {large isospin violation in
new-physics models enters $\epsilon'_K / \epsilon_K$ through
the initial conditions of the Wilson coefficients and not through
the RG evolution.}
We also should comment on the running of $\alpha_{EM}$. Above $M_W$
scale, we use $e(\mu_{\textrm{NP}}) = g (\mu_{\textrm{NP}})
g'(\mu_{\textrm{NP}}) / \sqrt{g^2(\mu_{\textrm{NP}}) +
g^{'2}(\mu_{\textrm{NP}})}$, and $\beta^{e}_0 = \beta_0^{g'} / \cos
\theta_W^2(M_Z) $, where $\beta_0^{g'} = -53/9 ~(\mu < m_t)$ or $ -
41/6 ~(\mu > m_t) $. Strictly speaking, we have to consider the
running of $\theta_W$ for consistency. {However, we have checked that the
numerical effect for an $\mathcal{O}(10$ TeV) scale of new physics is
small. Therefore we use a fixed value: $\sin^2 \theta_W = 0.231$.}
\subsection{Counting of Orders\label{sec:co}}
In a full next-to-leading order estimation, we have to consider the
leading order term $\mathcal{O}(1)$ arising from the one-loop QCD RG
evolution as well as the terms defined as next-to-leading order, which
are: the one-loop QED correction $\mathcal{O}(\alpha_{EM}/ \alpha_s)$,
the QCD two-loop correction $\mathcal{O}(\alpha_s)$, and the two-loop
term including a photon and a gluon {at}
$\mathcal{O}(\alpha_{EM})$. The next-to-leading order RG evolution
matrix has an additional $\mathcal{O}(\alpha_{EM}^2/\alpha^2_s)$
correction, which appears only at this order. Hereafter, we will always
refer to these orders when labelling perturbative quantities of the
Wilson coefficients and the evolution matrices as
$\vec{s}_0,~\vec{s}_e,~\vec{s}_s,~\vec{s}_{se},$ {$\vec{s}_{ee}$} and
$\hat{U}_{0},~\hat{U}_{e},~\hat{U}_{s},~\hat{U}_{se},~\hat{U}_{ee}$,
respectively.
When we multiply two quantities which are given by a perturbation
series, we have to carefully keep track of and consistently discard
higher orders {of the perturbative series}. This is a subtle {and
cumbersome} feature which complicates mathematical expressions. In
this context, equations of the RG evolution should be more of a symbolic
character which are exact in the limit of expanding the corresponding
quantities to all orders. Since we necessarily truncate the
perturbation expansion of the Wilson coefficients as well as the
evolution matrices at some point, a product of them at next-to-leading
order is represented as follows:
\begin{align}
\label{eq:NPRGexpand}
&\langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \left( \vec{s} (\mu) - \vec{\tilde{s}} (\mu)\right) \nonumber \\
&~~= \langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \hat{U} (\mu,\mu_{\textrm{NP}})\left( \vec{s} (\mu_{\textrm{NP}})- \vec{\tilde{s}} (\mu_{\textrm{NP}})\right) \nonumber \\
& \stackrel{(\textrm{NLO})}{=} \langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \big( \hat{U}_{0} + \hat{U}_{e} + \hat{U}_{s} + \hat{U}_{se} + \hat{U}_{ee} \big) \big( \vec{s}_{0} + \vec{s}_{e} + \vec{s}_{s} + \vec{s}_{se} {+\vec{s}_{ee}} \big) \nonumber\\
&~~= \langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \left( \underbrace{\hat{U}_{0} \vec{s}_{0}}_{=: \vec{s}_{0} (\mu)} + \underbrace{\hat{U}_{0}\vec{s}_{e} + \hat{U}_{e} \vec{s}_{0}}_{=: \vec{s}_{e} (\mu)} + \underbrace{ \hat{U}_{0}\vec{s}_{s} + \hat{U}_{s} \vec{s}_{0}}_{=: \vec{s}_{s} (\mu)} \right. \nonumber \\
& ~~~~~~ \left.+ \underbrace{\hat{U}_{0}\vec{s}_{se} + \hat{U}_{e} \vec{s}_{s} + \hat{U}_{s}\vec{s}_{e} + \hat{U}_{se}\vec{s}_{0}}_{=: \vec{s}_{se} (\mu)}
+ \underbrace{\hat{U}_{e} \vec{s}_{e} + \hat{U}_{ee} \vec{s}_{0} {+ \hat{U}_{0} \vec{s}_{ee} } }_{=: \vec{s}_{ee} (\mu)} \right)
+ \mathcal{O} \left(\frac{\alpha_{EM}^2}{\alpha_s }, \alpha_s^2, \alpha_s \alpha_{EM}, \alpha_{EM}^2\right)\nonumber\\
&~~= \langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \left( \underbrace{\vec{s}_{0} (\mu) + \vec{s}_{e} (\mu) + \vec{s}_{s} (\mu) + \vec{s}_{se} (\mu) + \vec{s}_{ee} (\mu)}_{=:\vec{s}_{\textrm{NLO}} (\mu)} \right) + \mathcal{O} \left(\frac{\alpha_{EM}^2}{\alpha_s }, \alpha_s^2, \alpha_s \alpha_{EM}, \alpha_{EM}^2\right)\nonumber\\
&~~= \langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \vec{s}_{\textrm{NLO}} (\mu) + \mathcal{O}\left(\frac{\alpha_{EM}^2}{\alpha_s }, \alpha_s^2, \alpha_s \alpha_{EM}, \alpha_{EM}^2\right).
\end{align}
Here we {have} suppressed the opposite-chirality coefficients
$\vec{\tilde{s}}$ and the arguments of $\hat{U} (\mu,\mu_{\textrm{NP}})$
and $\vec{s} (\mu_{\textrm{NP}})$ for {better} readability. This procedure
defines $\vec{s}_{\textrm{NLO}} (\mu)$ as a next-to-leading order
quantity, where higher orders have been discarded consistently.
In view of undetermined Wilson coefficients, it is beneficial to arrange
{the} terms {above} according to the Wilson coefficients
{evaluated} at the new physics scale as
\beq
\langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \vec{s} (\mu) &
\stackrel{(\textrm{NLO})}{=} \langle \vec{Q}_{\epsilon'_K} ( \mu)^{T}
\rangle\left[ \left( \hat{U}_{0} + \hat{U}_{e} + \hat{U}_{s} +
\hat{U}_{se} + \hat{U}_{ee} \right) \vec{s}_{0} \right.\nonumber \\ &
~~~~~~~~~~~ \left.+ \left( \hat{U}_{0} + \hat{U}_{e} + \hat{U}_{s}
\right) \vec{s}_{e} + \left( \hat{U}_{0} + \hat{U}_{e} \right)
\vec{s}_{s} + \hat{U}_{0} \vec{s}_{se} {+ \hat{U}_{0} \vec{s}_{ee}
} \right],
\label{eq:QUs}
\eeq
where we {have} again suppressed $\vec{\tilde{s}}$ and the arguments of
$\hat{U} (\mu,\mu_{\textrm{NP}})$ and $\vec{s} (\mu_{\textrm{NP}})$.
{For} given numerical values for the hadronic matrix elements at a
low scale and {with our} evolution matrices {connecting}
$\mu_{\textrm{NP}}$ with the low scale $\mu$, we can determine the
\emph{weights}\ {which multiply the}
Wilson coefficients Im$[ \vec{s} (\mu_{\textrm{NP}})
-\vec{\tilde{s}} (\mu_{\textrm{NP}})]$ {in Eq.~\eqref{eq:QUs}}
for {any chosen} scale of new physics. \\
\subsection{Evolution Matrices at the TeV scale}
Above the electroweak scale we observe an approximately logarithmic
behavior of the evolution matrix $\hat{U}(\mu, \mu_{ \textrm{NP}})$ in
Eq.~\eqref{eq:U3456} with increasing energy scale. {This
observation} allows us to {derive} an approximation for the evolution
matrix in the high energy region, {which has an error of only a few
percent}.
We give approximate functions for all {components} of the evolution
matrix {linking} the new physics scale to the hadronic scale. Cast in
the form
\begin{equation}
\hat{U}_{0,e,s,se,ee}(\mu, \mu_{ \textrm{NP}}) = \hat{U}_{1,fit} + \hat{U}_{2,fit} \ln\frac{\mu_{\textrm{NP}}}{1\,\,{\rm TeV}},
\end{equation}
we combine them in terms of Eq.~\eqref{eq:QUs}.
Using the analytic evolution matrices evaluated in Sec.~\ref{section2} and the next-to-leading order matching matrices $\hat{M}_{c,b,t}$,
we obtain
\beq
\hat{U}_{0} \left( \mu, \mu_{\textrm{NP}} \right)+ \hat{U}_{e} \left( \mu, \mu_{\textrm{NP}} \right)+ \hat{U}_{s} \left( \mu, \mu_{\textrm{NP}} \right)+ \hat{U}_{se} \left( \mu, \mu_{\textrm{NP}} \right) + \hat{U}_{ee}\left( \mu, \mu_{\textrm{NP}} \right) \nonumber \\
\simeq \hat{U}_{0,1,fit} + \hat{U}_{ 0,2,fit} \ln \frac{\mu_{\textrm{NP}}}{1\,\,{\rm TeV}},
\label{eq:app0}
\eeq
for the $\mathcal{O}(1)$ Wilson coefficients at the $\mu_{\textrm{NP}}$ scale, and
\beq
\hat{U}_{0} \left( \mu, \mu_{\textrm{NP}} \right)+ \hat{U}_{e} \left( \mu, \mu_{\textrm{NP}} \right)+ \hat{U}_{s} \left( \mu, \mu_{\textrm{NP}} \right) &
\simeq \hat{U}_{e,1,fit} + \hat{U}_{e,2,fit} \ln \frac{\mu_{\textrm{NP}}}{1\,\,{\rm TeV}},\\
\hat{U}_{0} \left( \mu, \mu_{\textrm{NP}} \right)+ \hat{U}_{e} \left( \mu, \mu_{\textrm{NP}} \right) &
\simeq \hat{U}_{s,1,fit} + \hat{U}_{s,2,fit} \ln \frac{\mu_{\textrm{NP}}}{1\,\,{\rm TeV}},\\
\hat{U}_{0} \left( \mu, \mu_{\textrm{NP}} \right) &
\simeq \hat{U}_{se,1,fit} + \hat{U}_{se,2,fit} \ln \frac{\mu_{\textrm{NP}}}{1\,\,{\rm TeV}},
\label{eq:appse}
\end{equation}
for the
$\mathcal{O}(\alpha_{EM}/\alpha_s),~\mathcal{O}(\alpha_s),~\mathcal{O}(\alpha_{EM})$
{(or $\mathcal{O}(\alpha_{EM}^2/\alpha^2_s)$)} Wilson coefficients at
the $\mu_{\textrm{NP}}$ scale, respectively. Here $\mu = 1.3$ GeV and
$\mu_c = 1.4$ GeV are taken, and the fitting matrices $\hat{U}_{fit}$
are given in Appendix~\ref{app:approximation}.
We find that these approximate evolution matrices are highly accurate
in the range of $500\,\textrm{GeV}$--$10\,\textrm{TeV}$.
In order to estimate which Wilson coefficients are expected to gain large enhancements through the RG evolution, we calculate \emph{weights}\ for the Wilson coefficients at the $\mu_{ \textrm{NP}}$ scale.
We regard the coefficients of $\langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \sum_i$ $\hat{U}_{i} \left(\mu, \mu_{ \textrm{NP}} \right) $ $( \vec{s}(\mu_{\textrm{NP}}) - \vec{\tilde{s}} (\mu_{\textrm{NP}}) )$ in Eq.~\eqref{eq:QUs} as weights of the Wilson coefficients.
\input{tables} In Table~\ref{table:weight0}, we list the coefficient $
\langle \vec{Q}_{\epsilon^{\prime}_K}(\mu)^{T} \rangle ( \hat{U}_0 +
\hat{U}_{e} + \hat{U}_{s} + \hat{U}_{se} + \hat{U}_{ee} ) $ for the
$\mathcal{O}(1)$ Wilson coefficients at the scale $\mu_{\textrm{NP}} =
1,\,3,\,5$ and $10\,\,{\rm TeV}$ in units of $(\text{GeV})^{3}$, where the hadronic
matrix elements of Table~\ref{tab:bfactors} are taken. Similarly, the
weights of the
$\mathcal{O}(\alpha_{EM}/\alpha_s),~\mathcal{O}(\alpha_s)$, and
$\mathcal{O}(\alpha_{EM})$ {(or
$\mathcal{O}(\alpha_{EM}^2/\alpha^2_s)$)} Wilson coefficients are
given in Tables~\ref{table:weighte},\,\ref{table:weights}, and
\ref{table:weightse}, respectively. Note that these values are not
obtained by fitting but using the exact analytic evolution matrices. We
observe that these values are of course dominated by $\hat{U}_0$,
{with the sub-dominant contribution stemming from} $\hat{U}_{e}$ because of
the $1/\omega_{+}$ enhancement {and $\hat{U}_{s}$}. We also find, that the largest weights
come in the $7$ and $8$ components, and they are further enhanced
through the RG evolution in the high energy regime. Compared with the
coefficients at the weak scale,
\beq &\langle \vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \hat{U}_0 (\mu,
M_W) \nonumber \\ &~~= \left(
{0.37,\, -0.02,\, 0.12,\, -0.29,\, 0.34,\, 0.83,\, 15.33,\, 54.09,\, 0.53,\, 0.08}
\right),\\ &\langle
\vec{Q}_{\epsilon'_K} ( \mu)^{T} \rangle \left( \hat{U}_0 + \hat{U}_{e}
+ \hat{U}_{s} + \hat{U}_{se} + \hat{U}_{ee} \right) (\mu, 1\,\,{\rm TeV}) \nonumber \\
&~~={ \left(
{0.27,\, -0.06,\, 0.05,\,-0.19, \,0.08, \,0.31,\, 26.16,\, 88.61,\, 0.12,\,-0.08}
\right),} \label{eq:weight1}\\ &\langle \vec{Q}_{\epsilon'_K} (
\mu)^{T} \rangle \left( \hat{U}_0 + \hat{U}_{e} + \hat{U}_{s} +
\hat{U}_{se} + \hat{U}_{ee} \right) (\mu, 10\,\,{\rm TeV}) \nonumber \\ &~~={ \left(
{0.20,\, -0.11, \,-0.04, \,-0.15,\, -0.15,\, -0.08,\, 34.19,\, 113.60, -0.20,\, -0.22}
\right),}
\eeq
{the weights of the $7$ and $8$ components increase}
by $50$--$100$\,\% through the RG evolution {at the scale of
$1$--$10$\,TeV.} If one omits the NLO correction $
\hat{U}_{e}+\ldots\hat{U}_{ee} $ in Eq.~\eqref{eq:weight1}, one finds
$ {22.77} $ and $ {76.05}$ for the 7th and 8th element (see
Tab.~\ref{table:weightse}), which shows the impact of the NLO
corrections on these elements. Although the enhancement factor from the
RG evolution has been pointed out before in
Ref.~\cite{Buras:2015kwd,Buras:2015jaq} within a leading-order analysis,
it has not been considered in {most} of the literature. We emphasize
that this factor should be included when one {studies} TeV-scale
new-physics contributions to the QED-penguin operators in order to
alleviate the $\epsilon'_K/\epsilon_K$ discrepancy.
\section{Conclusions and Discussion}
\label{sec:summary}
Based on the first complete lattice calculation of the hadronic matrix
elements for the $K \to \pi \pi $ decay, we have evaluated the
Standard-Model prediction of $\epsilon^{\prime}_K/\epsilon_K$ at the
next-to-leading order. It is well known that the analytic RG evolution
matrices for the $\Delta S = 1$ nonleptonic effective Hamiltonian at the
next-to-leading order contains singularities in intermediate steps of
the calculation. These singularities make practical calculation
laborious even though {appropriate regulators} disappear {from} the
final (physical) result. {In this paper, we have generalized the
analytic ansatz {of the Roma group \cite{Ciuchini:1992tj,
Ciuchini:1993vr} to solve the RG equations and derive a
singularity-free solution by adding} logarithmic terms {to the
ansatz}. As a novel feature of our solution compared to
Refs.~\cite{Adams:2007tk,Huber:2005ig} we do neither require the
diagonalization of the LO anomalous dimension matrix nor
case-by-case implementations for different
eigenvalues of this matrix. Instead, the
different cases are encoded in the $\hat J$ matrices given in
Eqs.~\eqref{Js0result}--\eqref{Jee1result} and
Appendix~\ref{app:solutionJ}. The singular nature of the RG
equations leads to the presence of spurious parameters which cancel
between the high-scale and low-scale NLO terms in the RG evolution
matrix and thereby do not produce any ambiguity and play the role of
scheme parameters with respect to the regularization of the
singularities. {Thus} we have explicitly proven that all
singularities are automatically {treated in the proper way} without
the need for a manual regularization of the evolution matrix. This
feature also leads to a subtlety whenever the NLO evolution matrix is
combined with LO initial conditions for the Wilson coefficients, as
one usually does in studies of new-physics contributions to
$\epsilon_K^\prime$.}
Using the improved RG evolution matrices and applying the recent lattice
results, we have calculated $\epsilon_{K}'/\epsilon_{K}$ in the Standard
Model at the next-to-leading order. Our final results is
$\epsilon_{K}'/\epsilon_{K} = \left( {1.06} \pm {5.07} \right) \times
10^{-4} $, which is $ {2.8}\,\sigma$ below the measured value.
Our {result is} consistent with the {recent} literature and
highlights a tension between the Standard-Model prediction and
experiment. The uncertainty is dominated by the lattice result of
$\langle (\pi \pi)_0 | Q_6 | K^0 \rangle$. Therefore, {upcoming}
improvements of lattice calculations will reveal whether this tension
really {calls} for new physics or not.
We have also evaluated the evolution matrices in the high energy region
for calculations of new physics contributions to
$\epsilon^{\prime}_K/\epsilon_K$. {To this end} we have {further}
obtained an {easy-to-use} approximate formula for the RG evolution
matrices in the TeV region at the next-to-leading order and {have}
also calculated the weights for each of the Wilson coefficients at the
{scale of new physics}. We observe that the largest weights come in the
$7$ and $8$ components of the Wilson coefficients and that they are
further enhanced through the RG evolution {between electroweak and
TeV scales}. Here we confirm the feature noticed at LO in
Ref.~\cite{Buras:2015kwd,Buras:2015jaq} and find a further enhancement
by the NLO corrections to the evolution matrices.
Especially the Wilson coefficients of the QED-penguin
operators at the {scale of $1$--$10$\,TeV} increase by $50$--$100$\,\%
compared with the Wilson coefficients at the weak scale.
\section*{Acknowledgements}
The authors thank Andrzej Buras and Chris Sachrajda for illuminating
discussions and Christoph Bobeth for pointing our attention to
Ref.~\cite{Huber:2005ig}. {We are grateful to Andrzej Buras, Martin
Gorbahn, and Sebastian J\"ager for alerting us to a mistake in an
earlier version of this paper. We thank the referee {for guiding} us to a
simpler version of Eq.~(\ref{eq:JseRome}) and
Eqs.~(\ref{eq:Jse2})--(\ref{eq:Jse0})} {and for pointing out the issue of the running of $\alpha_{EM}$.}
The work of UN is supported by BMBF under
grant no.~05H15VKKB1. PT acknowledges support from the DFG-funded
doctoral school {\it KSETA}.
|
\section{Introduction}
\change{The loss or corruption of entire segments of audio data
is a highly important problem in music enhancement and restoration.
Such corruptions can range from short bursts in the range of few
milliseconds to extended distortions that persist over several hundred
or even thousands of milliseconds. Short distortions such as clicks
or clipping have seen extensive coverage in the literature~\cite{siedenburg2013audio,adler2011constrained,godsill2002digital}, while
the concealment of moderate length distortions, roughly in the range of
$10$ to at most $100$~ms, is treated in packet loss compensation~\cite{perkins1998survey,bahat2015self}
and previous work on audio inpainting~\cite{adler2012audio,siedenburg2013audio,lagrange2005long}. For such corruptions, it
is often reasonable to assume that the lost signal is almost stationary
for the duration of the corruption and/or can be inferred from the reliable
information surrounding the unreliable segment. For longer duration loss,
such an assumption is increasingly unrealistic and a restoration technique
cannot rely only on local information. Here, we propose a method to
compensate for such extended data loss by considering information from the
entirety of uncorrupted audio available.
Data loss or corruption in the range of seconds can have various causes,
e.g. partially damaged physical media, such as phonograph cylinders,
shellac or vinyl records or even magnetic tapes. In live music recordings,
imperfections due to unwanted noise sources originating from the audience,
the artists themselves or the environment are quite common. Even in audio
transmission, a short, but total, loss of the connection between transmitter
and receiver may lead to data loss beyond just a few hundred milliseconds.
In each of these scenarios, the data loss has highly unpleasant consequences
for a listener, and it is usually not feasible to reconstruct the lost content
from local information only.
Previous work on concealment of data loss in audio, though mostly considering shorter
corruption duration, has been performed under various names, depending on the target application
and the employed methodology:
Audio inpainting~\cite{adler2012audio}, audio interpolation~\cite{Etter1996:Interpolation_AR}, waveform
substitution~\cite{goodman1986waveform}, or imputation~\cite{smaragdis2009missing} to name but a few.
We will use the terminology of \emph{audio inpainting} in the remainder of this contribution. }
When missing parts have a length no longer than $50$ms, sparsity-based techniques can be
successful~\cite{adler2012audio,siedenburg2013audio,adler2011constrained}.
Otherwise, techniques relying on auto-regressive
modeling~\cite{Etter1996:Interpolation_AR}, sinusoidal
modeling~\cite{lagrange2005long,lukin2008:parametric.interp.gaps} or based on
self-content~\cite{bahat2015self} have been proposed. The latter provided promising results
for speech signals with distortions up to $0.25$ seconds, while the former rely on a
simple signal model that does not comply with complex music signals.
In this contribution, we propose a new algorithm, specifically
targeted at the concealment of long duration distortions in the range of several seconds
\change{given a single piece of music. The task of determining distortion locations
is highly application-dependent and may be anything from trivial to very difficult.
For the sake of focus, we assume the location of the distortion to be known}.
Our method arises from the assumption that, across many musical genres,
the repetition, or variation, of distinct and recurring patterns (themes, melodies, rhythms, etc)
is a central stylistic element and thus heavily featured. When listening to music,
we detect and memorize such \emph{internal redundancies}, thereby learning the mid-
and large-scale structures of a music piece \cite{mcadams1987}. The exploitation of
such redundancies in the computational analysis and processing of music seems only
natural and, indeed, has been proposed before, see e.g.\change{~\cite{foote1999visualizing,jehan2005creating,jehan2004event}
or~\cite{rafii2013repeating}. The latter also provides a more extensive discussion of
repetition as an essential element of many musical genres.}
Although music information retrieval (MIR) provides many sophisticated
methods for the analysis of micro- and macroscopic structures in music, properly handled, a simple \emph{time-frequency analysis} can provide
all the necessary information to uncover significant similarities in music
signals.
The contributions of this work are the design of appropriate time-frequency features and their
use for generating a map of similarities in music signals, as well as the use of the generated
similarity map to drive the automatic concealment of long duration data loss.
\subsection{Related Work}
\change{Self-similarity in music has previously been employed in several areas of music analysis and processing, e.g. beat estimation and segmentation, and is often based on similarity matrices, as proposed by Foote~\cite{foote1999visualizing}. The similarity matrix can be constructed from various features, see e.g. \cite{foote2001beat,bartsch2001catch,cooper2002automatic,foote2000automatic}. Self-similarity has also been successfully used for music/voice separation and speech enhancement~\cite{rafii2013repeating,rafii2013online}. Finally, the automatic analysis of musical structure based on
similarities is already found in~\cite{silva2016simple}, where it was used across songs for cover song detection. An alternate approach can be found in~\cite{jehan2005creating,jehan2004event}\footnote{These studies led to the
founding of "The Echo Nest", see \url{http://the.echonest.com/},
a company specialized into audio feature design. The idea of a similarity
graph already appears in the infinite jukebox: \url{http://labs.echonest.com/Uploader/index.html}.}.
There, the division of music into short, rhythm-dependent pieces is proposed,
each of which is supposed to correspond to a single \emph{beat}. Local features are obtained for
each piece by combining previously established rhythm, timbre and pitch features, but the implementation details of their method are not disclosed. In this contribution, we propose a simple time-frequency feature built from the short-time Fourier magnitude and phase that implicitly encodes rhythmic, timbral and pitch characteristics of the analyzed signal all at once. We build a sparse similarity graph based on this feature that highlights only the strongest connections in a music piece. This similarity graph can be seen as a post-processed variant of Foote's
similarity matrix and is used to perform data loss concealment by detecting suitable transitions between similar segments in a piece of music.}
The audio inpainting problem has mainly been addressed from a sparsity point of
view. \change{The hypothesis is that audio is often approximately sparse in a time-frequency
representation, i.e. it can be estimated using only a few
time-frequency atoms.} Using classical $\ell_0$ or $\ell_1$ optimization
techniques, algorithms have been designed to inpaint short audio
gaps~\cite{adler2012audio,siedenburg2013audio}. Such methods strive for approximate
recovery of the lost data by sparse approximation in a time-frequency representation such
as the short-time Fourier transform (STFT). Both their numerical and perceptual restoration quality
quickly degrade when the duration of lost data intervals exceeds $10$~ms. When applied to significantly longer gaps, these methods will simply fade out/in at the gap border and introduce silence in the inner gap region.
Audio inpainting is known as "waveform substitution"~\cite{goodman1986waveform} by the community addressing packet loss recovery techniques~\cite{perkins1998survey}. Most packet loss methods, however, are naturally designed for low delay processing and compromise computation speed over quality, see also~\cite{bahat2015self} for a short overview. In that contribution, Bahat et al. propose an algorithm searching for similar parts of the signal using time-evolving features, conceptually resembling our own contribution. The method in~\cite{bahat2015self}
is designed for packet loss concealment in speech transmission, however, and was tested only on gaps up to $0.25$ seconds. The reliance on Mel frequency cepstral coefficients (MFCC) is a good match for speech, but not optimally suited for music. In another approach, Martin et al.~\cite{martin2011exemplar} proposed an inpainting algorithm taking advantage of the redundancy in tonal feature sequences of a music piece. Their method is able to conceal defects with a length of several seconds, but performance of this algorithm depends on the amount of repetitive tonal sequences in a music piece \cite{martin2011exemplar} and it was only applied when a recurrence of the lost tonal sequence was present in the reliable signal. \change{It should be noted that parallel work on audio inpainting using self-similarity by Manilow and Pardo~\cite{manilow2017leveraging} has been presented while the present manuscript was under review.}
\subsection{Structure of the paper}
After the introduction, we introduce the idea of the similarity graph, Section \ref{sec:DNA}. The general method and construction of the graph is presented in Section \ref{sec:algo}. Technical details about the graph construction such as the exact choice of features and parameters are deferred to Section \ref{sec:TecDecGraph}. In Section \ref{sec:TecDecInp}, we detail how the similarity graph can be used for audio inpainting. Finally, the performance of the algorithm is discussed, based on both a basic verification experiment and though extensive listening tests, Section \ref{sec:evaluation}.
\section{A transition graph encoding music structures} \label{sec:DNA}
The problem we consider, i.e. how to restore a piece of music when an extended,
connected piece has been lost or corrupted, often requires us to abandon the
idea of exact recovery. In the case where only a short segment (up to about
$50$ms) has been lost~\cite{adler2012audio}, or the signal can be described by a
very simple structure~\cite{lagrange2005long}, it may be possible to infer the
missing information from the regions directly adjacent to the distortion with
sufficient quality. However, for complex music signals and corruptions of longer
duration, such inference remains out of reach. Instead, we employ an analysis of
the overarching medium- and large-scale structure of a music piece, determining
\emph{redundancies} in the signal to be exploited in the search for a replacement
for the distorted signal segment.
Conceptually, such analysis can be seen as a music segmentation
into chorus and verse, motifs and their variation, sections of
equal or different meters, etc \cite{macpherson2008form}. The main difference to our approach is that, instead of
working with high-level cognitive concepts such as meter and motifs, we instead consider
a basic time-frequency representation of the signal. In that representation, all the structures contained in a music recording are still preserved, although it is not always easily accessible to the human observer.
It is clear that repetition and less obvious redundancies do not occur to an equivalent
degree in every music piece. While they are an essential stylistic element to
pop and rock music, certain movements, e.g. in contemporary music, attempt the
active avoidance of the familiar.
But even if a pattern is not repeated in the
exactly same fashion, the conscious variation of previous structures, rhythmic,
harmonic or otherwise, is an integral part of most music. Note that the grade of
self-similarity inside a single recording may vary greatly.
Going back to the original problem of music restoration, it seems natural to
exploit this type of \emph{redundancy} in the musical piece to be restored.
The temporal evolution of
spectral content provides a surprisingly suitable first approximation of musical features.
Inspired by this observation, we construct an audio similarity graph. The
vertices of the graph represent small parts of musical content, while the edges
indicate the similarity between the segments in terms of local spectral
content. The crucial step towards good performance is the enforcement of
temporal coherence. This is achieved by selecting transitions that persist over
time, i.e. similarity is not instantaneous, but present for some period of time.
\section{Method} \label{sec:algo}
The ultimate goal of this contribution
is to provide a means for autonomous concealment of signal defects with a
duration of a few hundred to several thousand milliseconds.
\change{Here, we assume that the position of the defects is already known.}
\emph{The restoration
should sound natural and respect the overall structure of the signal under
scrutiny.} For short distortions, this implies, to some degree, the recovery
of the lost information in the defective region. For long gaps and dynamic signals,
we argue that it is of much greater importance than the transitions between the
reliable signal segments and the proposed restoration sound natural. The further
away from the transition points we are into the restored region, the less important
exact recovery becomes versus the restoration \emph{making sense} in the signal
context. Therefore, we suggest an analysis of the signal structure with the proposed
similarity graph, to determine the most natural fit for the distorted region from
unaffected portions of the signal. The resulting method is an abstracted and
autonomous version of manual restoration by searching the reliable signal for
a replacement for the defective region. Since the proposed method forgoes the
synthesis of new audio content and provided that enough reliable signal content is available, the proposed method can handle signal defects of arbitrary length without affecting audio quality.
We obtain from a short-time Fourier transform~\cite{allen1977unified,Grochening:2001a,ga46} simple similarity features carrying important temporal and spectral information. On the basis of these features, a similarity graph is constructed, representing the temporal evolution and structure of
the signal. If some signal segment is known to be defective, it is now sufficient to
determine another segment of similar length, such that the beginning and end of the
substitute resemble the signal before and after the defect. By placing the candidate
segment at the previously corrupted position, the defect can be concealed.
The proposed algorithm, illustrated in Figure~\ref{fig:main_idea}, searches for a replacement segment that optimally satisfies the three following criteria:
\begin{enumerate}
\item The transitions $T_1$ and $T_2$ (light green dashed lines) resulting
from the pasting operation should be perceptually transparent, i.e., the
listener should not be able to notice the transition, even if the
replacement segment does not correspond exactly to the missing data.
\item Some leeway is required for placing the transitions around the gap,
represented by $L_1$, $L_2$. However, the transition areas should not be
unnecessarily long.
\item The length of the piece should remain approximately the same, i.e.,
the replacement duration $D_2$ should be close to the gap plus its surroundings,
$D_1$.
\end{enumerate}
Some margin for compromise is, however, essential to the construction of a good
solution. Since the question of how strictly the reliable content is to be
preserved, i.e. how long $L_1$ and $L_2$ may be, is highly application-dependent,
a parameter in the optimization scheme enables the tuning of this property.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.95\linewidth]{figures/main_idea.png}
\end{center}
\caption{\textbf{Illustration of the proposed inpainting method.} The determined candidate
segment of duration $D_2$ is to be substituted for the gap. The optimal transition points
$T_1$ and $T_2$ are determined together with the candidate segment by jointly optimizing
(i) the similarity feature at $T_1$, $T_2$, (ii) the difference $|D_1-D_2|$ and the
length of the necessary transition areas $L_1$ and $L_2$.}
\label{fig:main_idea}
\end{figure}
In practice, at least for the inpainting problem, it is unnecessary to construct the full
similarity graph. Consequently, we construct a sparsified graph which considers unique and
strong matches only. Weak matches are discarded. Only the strongest from a cluster of (temporally close) matches are considered. Finally, only edges connected to at least one node in the vicinity of the gap are relevant, since $L_1$ and $L_2$ are supposed to be small, see Figure \ref{fig:main_idea}.
\subsection{Creation of the similarity graph}
The generation of the graph can be structured coarsely into $4$ distinct stages. In this section we disregard some technical details, instead concentrating on the general idea. The technical details of the individual steps of our method can be found in Sections~\ref{sec:TecDecGraph} and~\ref{sec:TecDecInp}.
\paragraph{1. Compute basic similarity features}
To determine temporal similarities in a signal, we have to settle on a \emph{feature}
that encodes the local signal behavior and a \emph{distance measure} that allows the
comparison of feature vectors. For simplicity, and because the results were comparable
to more sophisticated features, we settle here on a weighted combination of two features
obtained directly from a short-time Fourier (STFT) analysis of the signal. Let $\mathbf C$ be
the matrix of short-time Fourier coefficients, with $\mathbf C_{m,n}$ denoting the coefficient obtained at
the $n$-th time position in the $m$-th channel. $\mathbf C_{m,n}$, see Section \ref{sec:TecDecGraph}, can be decomposed uniquely into its magnitude $\mathbf M_{m,n}\geq 0$ and phase $\mathbf \Phi_{m,n}\in]-\pi,\pi]$ as
\[
\mathbf C_{m,n} = \mathbf M_{m,n}e^{i\mathbf \Phi_{m,n}}.
\]
Since the human auditory system perceives loudness approximately as a logarithmic function of sound pressure, the first part of our proposed feature is essentially a time slice of the dB-spectrogram, i.e. \[
\tilde{F}_n^1 := [ 20\log_{10}(\mathbf M_{0,n}),\ldots,20\log_{10}(\mathbf M_{M-1,n})].
\]
\change{Note that direct spectrogram features have already proven to be useful in other applications, e.g. repetition-based source separation, see~\cite{rafii2013repeating}.
Additionally,} the time-direction partial derivative of the phase provides an estimate of the local instantaneous frequency~\cite{augfla95:reassign,Holighaus:2016:RSG:2910117.2910282}. \change{Let $\mathbf \Phi^{td}$ denote the $M\times N$-matrix containing the values of the time direction partial derivative of the phase, i.e. $\mathbf \Phi^{td}_{m,n} = \partial \mathbf \Phi_{m,\cdot}[n]$. The second part of our proposed feature is essentially
\[
\tilde{F}_n^2 := \left[\mathbf \Phi^{td}_{0,n},\ldots,\mathbf \Phi^{td}_{M-1,n}\right],
\]
and $\tilde{F}_n = [\tilde{F}^1_n,\tilde{F}^2_n]^T$.} While $\tilde{F}^1$ puts a strong emphasis on signal components of high amplitude, $\tilde{F}^2$ attains large values for sinusoidal, or slowly frequency-varying, components independent of their magnitude, see also Figure \ref{fig:features}. This second part of the feature serves to emphasize low amplitude harmonic components, which may be highly important for perceived similarity. \change{The actual feature $\mathbf f_n$, defined in Section \ref{sec:TecDecGraph}, is conceptually equivalent to $\tilde{F}_n$, but implements some additional scaling.} Locality of the features is implied by obtaining the features from a STFT. The distance between two features at $l,k$ is simply the squared Euclidean distance of $\mathbf f_l$ and $\mathbf f_k$.
\begin{figure}[thb!]
\begin{center}
\includegraphics[width=0.95\linewidth]{figures/features.png}
\end{center}
\caption{Local audio features for an exemplary audio signal. The log-spectrogram $\tilde{F}_1$
(top) encodes the time-dependent intensity of frequency components. The smoothed partial phase derivative $\tilde{F}_2$ (bottom) has large values
in the area of stable, harmonic components, independent of the component magnitude.\label{fig:features}}
\end{figure}
\paragraph{2. Create a preliminary similarity graph}
The full (unprocessed) similarity graph determined from the given feature vectors would simply have all the time positions $n\in\{0,\ldots,N-1\}$ as vertices and edges connecting each vertex to every other vertex, with the associated weights derived from the distance between the associated features.
The creation of such a graph is not only very expensive, but we are further only interested in a small number of strongest connections for every vertex. Therefore, we only determine the $K$ nearest neighbors, in terms of feature distance. Since this operation is expensive, we use the FLANN library (Fast Library for Approximate Nearest Neighbors)~\cite{muja2014scalable} to efficiently provide an approximate solution. For the $K$ determined neighbors, the edge weights are recorded in the adjacency matrix as \change{
\begin{equation}\label{eq:Wnull}
\mathbf{W_0}(l,k) = \begin{cases}
e^{-{\textstyle \frac{\|\mathbf f_l - \mathbf f_k\|_2^2}{\sigma}}} & \text{if } k \text{ is among the } K \text{ n.n.s of } l\\
0 & \text{otherwise,}\\
\end{cases}
\end{equation}}
for some $\sigma>0$, following a traditional graph construction scheme, see also Figure \ref{fig:adjmatrices} (left).
\paragraph{3. Enhance time-persistent similarities}
The individual features obtained from the STFT usually characterize signal's properties on a local time interval and do not capture the long-term signal's spectral characteristics. In order to capture longer temporal structures of a signal, we refine the graph by emphasizing its edges whenever a sequence of features at consecutive time positions is similar to another. In practice,
this is achieved by convolving \change{the weight matrix $\mathbf{W_0}$} with a diagonal kernel $\mathbf D\in \RR^{L_K+1\times L_K+1}$, for some $L_K\in 2\mathbb{N}$, with
\[
\mathbf D_{l,l} = 1-\frac{|L_K-2l|}{L_K} \text{ and } \mathbf D_{l,k} = 0, \text{ if } l\neq k.
\]
The resulting adjacency matrix is given as
\begin{equation} \label{eq:weight_convolution}
\begin{split}
\mathbf W(l,k) & = (\mathbf{W_0}\ast \mathbf D)(l,k)\\
& = \sum_{l_0 =-L_K/2}^{L_K/2} \left(1-\left|\frac{2l_0}{L_K}\right|\right) \mathbf{W_0}(l+l_0,k+l_0),
\end{split}
\end{equation}
see also Figure \ref{fig:adjmatrices} (middle). Note that, in order to obtain an $N\times N$-matrix $\mathbf W$ and for the above equation to be valid $\mathbf W_0$ is implicitly extended to an $(N+L_K)\times(N+L_K)$-matrix with $L_K/2$ zeros on any side.
\begin{figure}[thb!]
\begin{center}
\includegraphics[width=0.32\linewidth]{figures/W_matrix.png}
\includegraphics[width=0.32\linewidth]{figures/Wconv_matrix.png}
\includegraphics[width=0.32\linewidth]{figures/Wsparse_matrix.png}
\end{center}
\caption{Weight matrix based on feature vectors calculated for an exemplary
audio signal \textit{without} a gap. Left panel: Preliminary weight matrix,
$\mathbf{W_0}$, of the initial graph. Center panel: Convolved weight matrix, $\mathbf{W}$.
Right panel: Excerpt of the weight matrix, $\mathbf{W_s}$, of the sparsified graph.}\label{fig:adjmatrices}
\end{figure}
\paragraph{4. Delete insignificant similarities/Merge clustered similarities}
After the convolution with the diagonal kernel, \change{the weight matrix $\mathbf{W}$} of our graph has been populated with a large number of nonzero entries, clustered around the entries of $\mathbf{W_0}$. The maxima of such clusters represent the strongest similarities between two regions of the signal. Moreover, only strong connections indicate significant similarities. Therefore, we delete all edges with weights below a certain threshold $t_w$ and select from every cluster of connections only the strongest, i.e the one with \emph{locally} the largest weight. \change{ This last step leads to the weight matrix $\mathbf{W_s}$ which is associated to the graph we use for our inpainting algorithm.} For an example of the final, sparsified adjacency matrix, see Figure~\ref{fig:adjmatrices} (right). Figure~\ref{fig:graphs} shows the difference between the original graph after Step 2 and part of the refined graph after Step 4.
\begin{figure}[thb!]
\begin{center}
\includegraphics[width=0.475\linewidth]{figures/full_graph.png}
\includegraphics[width=0.475\linewidth]{figures/max_graph.png}
\end{center}
\caption{Graphs based on feature vectors calculated for an exemplary audio
signal \textit{without} a gap. \change{Left panel: Initial non-sparse graph, $G_0$, corresponding
to the weight matrix $\mathbf{W_0}$, shown in Figure~\ref{fig:adjmatrices} (left). Right
panel: Sparse graph, $G_s$ (only local maximum weights above the threshold
considered), corresponding to the weight matrix $\mathbf{W_s}$, shown in
Figure~\ref{fig:adjmatrices} (right).}}\label{fig:graphs}
\end{figure}
\subsection{Application: Audio inpainting and the reduced similarity graph} \label{sec:audio_inpainting_reduced}
The usage of the similarity graph for solving an inpainting problem is
rather straightforward. According to the paradigm described in Figure~\ref{fig:main_idea},
we want to find two edges \change{$(l_0,k_0)$ and $(l_1,k_1)$}, such that
\begin{itemize}
\item $l_0$ is close to the beginning of the distorted region and $k_1$ is close to its end,
\item \change{$k_1-l_0$ is approximately equal to $l_1-k_0$} and
\item \change{$\mathbf{W_s}(l_0,k_0)$ and $\mathbf{W_s}(l_1,k_1)$} are large.
\end{itemize}
An appropriate choice of \change{$(l_0,k_0)$ and $(l_1,k_1)$} is determined by optimizing these $3$ criteria over all possible choices, for $l_0$ and $k_1$ in some limited range around the signal defect. The signal segment corresponding to the local features \change{$k_0,\ldots,l_1$} is then substituted for the original signal in the range corresponding to $l_0,\ldots,k_1$.
For the purpose of inpainting, we are only interested in edges that connect to at least one vertex either
shortly before, or shortly after, the signal defect. Hence, only a small horizontal (or vertical) slice of the sparse matrix $\mathbf{W_s}$ has to be computed, greatly reducing the complexity of the graph creation. \change{Figure~\ref{fig:subgraph} shows an example of such a reduced graph (not to be confused with the sparse graph) and the determined transitions \change{$T_1$ indexed by $(l_0,k_0)$ and $T_2$ indexed by $(l_1,k_1)$} for an exemplary signal and defect. In practice we use the reduced for graph all experiment of this paper.}
\begin{figure}[thb!]
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/subgraph.png}
\end{center}
\caption{Final reduced graph based on exemplary audio features calculated from an audio
signal \textit{with a gap}. \change{The
regions considered for the transitions are in gray with the gap in between them in white. All available transitions for the reconstruction are in light gray with the optimally selected $T_1$ and $T_2$ in blue. The nodes indexes $l_0,l_1$ and $k_0,k_1$ correspond the beginnings and the ends of the transitions $T_1$ and $T_2$} }\label{fig:subgraph}
\end{figure}
\section{The similarity graph in detail}\label{sec:TecDecGraph}
\subsection{Local audio features} \label{sec:audio_features}
\change{Building a similarity graph for full music pieces from STFT features is in practice challenging simply due
to the number and size of the obtained features. }To be efficient, the number of features has to remain small in contrast to the complexity of audio signals. Our solution leverages two techniques to obtain a good trade-off: 1) an adequate sub-sampling, and 2) a tight low-redundancy STFT.
While audio signals are often sampled at a very high rate, to compute
reliable audio features, a much lower rate is usually sufficient. We choose
a maximum sampling rate of $\xi_{\text{max}}$~Hz (default $12$~kHz, see Table~\ref{tab:parameters}
for all default parameters). If a given signal $s\in\RR^L$ is sampled at a higher rate $\xi_s$~Hz, $s$ is decimated with a decimation factor $d=\lceil \xi_s/\xi_{\text{max}} \rceil$, after the application of
an anti-aliasing filter. We denote the decimated signal by $s_d$.
The short-time Fourier transform (STFT) of $s_d$ with respect to a (real-valued) window function $g$,
hop size $a\in\mathbb{N}$ and $M$ channels is
defined as
\[
\mathbf C_{m,n}:= \sum_{l=0}^{L-1} s_d[l]g[l-na]e^{-2\pi iml/M},
\]
for $n\in\{0,\ldots,L/a-1\}$ and $m\in\{0,\ldots,M-1\}$. Recall the decomposition of $\mathbf C_{m,n}$ into magnitude and phase: $\mathbf C_{m,n} = \mathbf M_{m,n}e^{i\mathbf \Phi_{m,n}}$, $\mathbf M_{m,n}\geq 0$, $\Phi_{m,n}\in]-\pi,\pi]$. By default, we choose $g$ to equal a $1024$-point Itersine window~\cite{wesfreid1993adapted}, $a=128$ and $M=1024$. This particular construction leads to an $8$ redundant tight frame, hence preserving equally each signal frequency component.
The $2$ separate parts $\mathbf F^1_n$ and $\mathbf F^2_n$ of the feature vector $\mathbf f_n$ are obtained as follows.
\paragraph{dB-Spectrogram} Let $\mathbf S^{dB}_{m,n} := 20\log10(\mathbf M_{m,n})$, $n\in\{0,\ldots,L/a-1\}$ and $m\in\{0,\ldots,M-1\}$. For more convenient handling, $\mathbf S^{dB}$ is limited to a fixed range and peak-normalized, resulting in
\[
\mathbf F^{1}_{m,n} = t_s^{-1}\left(\mathbf S^{dB}_{m,n}-\max_{k,l}(\mathbf S^{dB}_{k,l})+t_s\right)_{+},
\]
where $(x)_+ = x$, if $x>0$, and $0$ otherwise. By default, $t_s=50$~dB. Figure~\ref{fig:features} (top) shows $\mathbf F^{1}$ for an exemplary audio signal.
\paragraph{Relative instantaneous frequency} In~\cite{augfla95:reassign},
the authors show that an instantaneous frequency estimate can be associated to
$C_{m,n}$ by
\begin{equation}\label{eq:instFreq}
\tilde{\xi}_{m,n} := \frac{\xi_sm}{M} -\mathbf{Im}(\mathbf C^{td}_{m,n}/\mathbf C_{m,n}),
\end{equation}
where $\mathbf C^{td}_{m,n}:= \sum_{l=0}^{L-1} s_d[l]g'[l-na]e^{-2\pi iml/M}$ and $g'$ is a discrete derivative of $g$.
The second term in the equation above is in fact an equivalent expression for the
partial derivative $\mathbf \Phi^{td}_{m,n}$ of $\mathbf \Phi$, with respect to $n$. $\tilde{\xi}_{m,n}$ might fluctuate quickly and its range depends on $m$. Both these properties are undesired for our purpose.
Therefore, we consider only its relative part, i.e. the second term in Eq. \ref{eq:instFreq}, and
perform a channel-wise smoothing of each $\tilde{\xi}_{m,\cdot}$, $m\in 0,\ldots,M-1$, by convolution with a localized kernel $v_{\text{ker}}$ (default: $8$-point Hann window).
Additionally, the expression for $\tilde{\xi}_{m,n}$ is unstable in regions of small magnitude $\mathbf M_{m,n}$ \cite{xxlbayjailsoend11}. With $t_p = \max_{m,n} |\mathbf{Im}(\mathbf C^{td}_{m,n}/\mathbf C_{m,n})|$, we define
\[
\mathbf F^{2}_{m,n} = \begin{cases}
-t_p^{-1}\left(\mathbf{Im}(\mathbf C^{td}_{m,\cdot}/\mathbf C_{m,\cdot})\ast v_{\text{ker}}\right)[n] & \text{ if } \mathbf F^{1}_{m,n}>0,\\
0 & \text{ else.}
\end{cases}
\]
The combined feature vector is obtained as
\begin{eqnarray*}
&\mathbf f_{n} = ( \mathbf F^{1}_{1,n},\ldots,\mathbf F^{1}_{M-1,n},\lambda\mathbf
F^{2}_{1,n},\ldots,\lambda\mathbf F^{2}_{M-1,n})^T,
\end{eqnarray*}
for $n=0,\ldots,L/a-1$. \change{We choose a default value of $\lambda = 3/2$, since this choice resulted in similar importance placed on both sub-features.}
\subsection{Creation of the similarity graph} \label{sec:similarity_graph}
When it comes to the graph creation, we desire an automatic parameter selection adapting to the audio features.
For the creation of the initial graph, we only need to determine the value of $\sigma$ in
the expression \eqref{eq:Wnull} for the preliminary weight matrix. Denoting as $K_n$ the
set of $K$ approximate nearest neighbors of the vertex $n$, our solution is to set
$\sigma$ to the average squared nearest neighbor distance
\[
\sigma = \frac{1}{NK}\sum_{n=0}^{N-1}\sum_{l\in K_n} \|\mathbf f_n - \mathbf f_l\|_2^2.
\]
Thus $\mathbf W_0(l,k)\approx 1$ if $\mathbf f_l$ and $\mathbf f_k$ are close, and decreasing towards $0$, the more
$\mathbf f_l$ and $\mathbf f_k$ differ. Our experiments showed that $K$ of $40$ is a good default value, which should be increased if the music is expected to be very redundant.
To obtain $\mathbf{ W}$ from $\mathbf{W_0}$ in~\eqref{eq:weight_convolution}, the length of the convolution kernel must be fixed. After
the convolution, the edges in the graph describe the similarity of
signal segments
of $\frac{aL_K}{\xi_s}\lceil \xi_s/\xi_{\text{max}}\rceil$~seconds duration. The choice of
$L_K$ determines the importance of \emph{long duration similarities} over such with short duration.
We used \change{$L_K = 40$} as a default value in order to consider roughly half-second segments for signals
sampled at $44.1$~kHz, see Figure \ref{fig:kernel}.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.475\linewidth]{figures/kernel.png}
\end{center}
\caption{Convolution kernel used to enhance the diagonal shape of the weight
matrix. \change{Here $L_k=40$.} \label{fig:kernel}}
\end{figure}
To transition from $\mathbf W$ to $\mathbf W_s$, we first perform a thresholding
by $t_w$. In $\mathbf{W_0}$, each entry can be 1 at maximum, see \eqref{eq:Wnull}. In $\mathbf{ W}$, solitary entries will be smaller than 1 and entries surrounded by other high-valued entries will be larger than 1. In order to suppress solitary entries, we used $t_w = 2$ as a default value.
The final step consisting of selecting the local maxima by choosing points that are equal to or larger than the four direct neighbors.
\change{In detail $\mathbf{W_s}$ is defined as
\begin{equation*}
\mathbf{W_s} (l,k)= \begin{cases}
\mathbf{W}(l,k) & \text{if } \mathbf{W} (l,k) \geq \\
& \indent \max\{ t_w, \mathbf{W} (l\pm 1,k\pm 1) \}\\
0 & \text{otherwise.}
\end{cases}
\end{equation*}
The notation $\pm$ should be interpreted as the collection of all possible choices, i.e. the maximum over all direct neighbors in $\mathbf{W}$
}
When applying the calculation of the transition graph to a signal where the
distorted area is known, the computational cost can be further reduced.
In particular, only a partial
transition graph needs to be computed because we are only interested in outgoing
connections within the short region before, and incoming connections
within a short region immediately after the distortion. Conceptually, we consider only
small $L_1$ and $L_2$, cp. Figure~\ref{fig:main_idea} and Section \ref{sec:transitions}. \change{Therefore, the $K$
nearest neighbors search and all the following operations, is not performed on all nodes, but only for a small
subset of features in the direct vicinity of the signal defect. This allows us to not only greatly reduce the computation cost, but also reduce the size of the optimization problem described in Section~\ref{sec:transitions}. An example of such resulting graph is given in Figure~\ref{fig:subgraph}.}
\section{The inpainting step in detail}\label{sec:TecDecInp}
\subsection{Selection of optimal transitions} \label{sec:transitions}
To select the optimal transition, we need to transform the three conditions of Section~\ref{sec:audio_inpainting_reduced} into a mathematical objective function.
Let $d_s,d_e$ denote the
index of the nodes corresponding to the start and end of the distorted region. \change{In the notation
of the previous section, only edge $(l_0,k_0),(l_1,k_1)$ with $l_0\in L_1:=\{[d_s-\epsilon_1,d_s]$
and $k_1\in L_2:=[d_e,d_e+\epsilon_2]\}$ are considered acceptable. In our experiments, we observed that setting $\epsilon_1=\epsilon_2$ to a length corresponding to approximately 5 seconds yielded good results. The region considered for possible transition can be seen as the red interval in Fig \ref{fig:subgraph}.}
Among all acceptable edges, we search for the solution that
minimizes the \emph{objective function}
\change{\begin{equation}
\begin{split}
f\left((l_0,k_0),(l_1,k_1)\right) & = \left|(k_1-l_0)-(l_1-k_0)\right| \\
& + \gamma_2 \left((d_s-l_0) + (k_1-d_e) \right) \\
& + \gamma_3 \left( \frac{1}{\mathbf{W_s}(l_0,k_0)} + \frac{1}{\mathbf{W_s}(l_1,k_1)}\right).
\end{split}
\end{equation}}
Compare the definition of $f$ with Figure \ref{fig:main_idea} to see that:
The first term controls the difference $D_2-D_1$, the second term the distances
$L_1,L_2$ from the defect and the third term controls the quality of the transitions.
By tuning $\gamma_2$ and $\gamma_3$, we can vary the importance of the individual terms.
In our experiments, $\gamma_2 = 1$ and $\gamma_3 = 100$ have provided good results.
Since the number of acceptable transitions is small, the computational benefit from
using a sophisticated optimization algorithm is negligible. Hence, we
solve the optimization problem by simply computing exhaustively the values of the
objective function for each set \change{$(l_0,k_0),(l_1,k_1)$ with $l_0\in L_1$
and $k_1\in L_2$.}
\subsection{Signal reconstruction} \label{sec:reconstruction}
When two audio signals are concatenated naively, discontinuities and phase jumps might
result in clicking artifacts. To reduce these effects, a smoothed
transition is clearly preferred. We propose the following:
Since the features are obtained from a STFT with time step $a$, with respect
to a possibly decimated signal, the \emph{time resolution} of the similarity
graph analysis equals \change{$\tilde{a}:= a\lceil \xi_s/\xi_{\text{max}}\rceil$} samples.
In other words, the preliminary solution obtained in the previous step suggests
the insertion of the signal samples \change{$s[\tilde{a}k_0,\ldots,\tilde{a}l_1-1]$} in
place of \change{$s[\tilde{a}l_0,\ldots,\tilde{a}k_1-1]$}.
To further improve the transition, we \change{allow to adjust the transition positions
$\tilde{a}l_0$ and $\tilde{a}k_1$ by up to half the similarity graph's time resolution, i.e. $\tilde{a}/2$
samples. The optimal adjustment is determined by maximizing a correlation, as
proposed in \cite{bahat2015self} and described below.} Denote by $L_w$ the length of the
analysis window
$g$ and \change{$\tilde{L}_w:= \lceil L_w/2\rceil \lceil \xi_s/\xi_{\text{max}}\rceil$}.
The final transitions are given
by \change{$(\tilde{l}_0,k_0),(\tilde{a}l_1,\tilde{k}_1)$},
where \change{
\[
\begin{split}
\tilde{l}_0 & = \argmax_{l\in{[}\tilde{a}l_0-\tilde{a}/2,\tilde{a}l_0+\tilde{a}/2[} \langle s_{l},s[\tilde{a}k_0-\tilde{L}_w,\ldots,\tilde{a}k_0+\tilde{L}_w-1]\rangle,\\
\tilde{k}_0 & = \argmax_{l\in{[}\tilde{a}k_1-\tilde{a}/2,\tilde{a}k_1+\tilde{a}/2[} \langle s_{l},s[\tilde{a}l_1-\tilde{L}_w,\ldots,\tilde{a}l_1+\tilde{L}_w-1]\rangle.
\end{split}
\]
Here, $s_l\in\RR^{2\tilde{L}_w}$ is the vector
\[
s_{l}[j] = \begin{cases}
0 & \text{ if } l-\tilde{L}_w+j \in[d_s,d_e],\\
s[l-\tilde{L}_w+j] & \text{ otherwise.}
\end{cases}
\]}
The obtained indices $\tilde{l}_0,\tilde{k}_1$ maximize the correlations between the original signal and the inpainting candidate.
\change{In order to obtain smooth transitions in the restored signal, we perform a time-frequency domain
cross-fading. Conceptually, this requires us to consider $3$ different arrays of short-time Fourier coefficients
with time step offsets $\tilde{l_0}-l_0\tilde{a}$, $0$ and $\tilde{k_1}-k_1\tilde{a}$, respectively:
\[
\begin{split}
\mathbf C^{(1)}_{m,n} & = V_{\tilde{g}}
\tilde{s}[(n-l_0)\tilde{a}+\tilde{l}_0,mL/M], \\
\mathbf C^{(2)}_{m,n} & = V_{\tilde{g}}
\tilde{s}[n\tilde{a},m\tilde{L}/\tilde{M}], \\
\mathbf C^{(3)}_{m,n} & = V_{\tilde{g}}
\tilde{s}[(n-k_1)\tilde{a}+\tilde{k}_1,mL/M].
\end{split}
\]
This ensures that in $C^{(1)}$, the $l_0$-th time frame is centered at the signal position $\tilde{l_0}$ and
in $C^{(3)}$, the $k_1$-th time frame is centered at position $\tilde{k_1}$. The analysis window
$\tilde{g}$ is chosen, such that on the undecimated signal $s$, it mimics $g$ acting on $s_d$. Hence, its length
and the number of channels $M$ are set to $2\tilde{L}_w$. Thus, by default, we choose $\tilde{g}$ to also be of Itersine shape.}
The restored signal can now be obtained by applying the inverse STFT
to the combined matrix,
\[
\begin{split}
\mathbf C^{\text{rec}} & = \Big(\mathbf C^{(1)}_{\cdot,1},\ldots,\mathbf C^{(1)}_{l_0-1},
\mathbf C^{(2)}_{\cdot,l_1},\ldots,\mathbf C^{(2)}_{\cdot,k_0-1}, \\
& \hspace{90pt} \mathbf C^{(3)}_{\cdot,k_1},\ldots,\mathbf C^{(3)}_{\cdot,L/\tilde{a}-1}
\Big).
\end{split}
\]
In practice, complexity is further reduced without altering the result, by computing $\mathbf C^{(j)}$, $j=1,2,3$, only for the time-positions relevant to the cross-fading, thus obtaining two small submatrices of $\mathbf C^{\text{rec}}$. \change{Note that any coefficient vector $\mathbf C^{(j)}_{\cdot,n}$, $j=1,2,3$, $n=0,\ldots,N-1$, only affects the reconstruction on an interval equal to the window length $2\tilde{L}_w$. Hence, both transitions have a duration of $2\tilde{L}_w$ and the first can be recovered from
\[
\Big(\mathbf C^{(1)}_{\cdot,l_0-r},\ldots,\mathbf C^{(1)}_{l_0-1},
\mathbf C^{(2)}_{\cdot,l_1},\ldots,\mathbf C^{(2)}_{\cdot,l_1+r-1}\Big),
\]
and similarly for the second transition. Here, $r:= 2\lceil \tilde{L}_w/\tilde{a}\rceil$ is a generous estimate of the ratio between the window length and the hop size.} The inverse STFT is then applied to these submatrices and the cross-fade regions, which are obtained as the central part of those inverse STFTs, are placed at the desired position in the signal. All other operations are performed in the time domain. To ensure equivalence with a complete STFT computation, the segments have to start/end $M$ samples before/after the cross-fading.
\section{Numerical evaluations}\label{sec:experiments} \label{sec:evaluation}
In this section we provide a numerical evaluation of the proposed algorithm.
First, we
verify the algorithm in a setting where the gap content is provided with the
remaining signal. A correct implementation should be able to perfectly replace
the gap by exactly the lost content. Second, we investigate algorithm's computational
performance in terms of average runtime.
For the evaluations, the algorithm was implemented in MATLAB. The implementation
is based on LTFAT \cite{ltfatnote030} for feature extraction, and on the GSPBox
\cite{perraudin2014gspbox} for graph creation. For non-commercial use, the
algorithm is available
online\footnote{\change{\url{https://epfl-lts2.github.io/rrp-html/audio_inpainting/}}}, alongside a
browser-based demonstration
\footnote{\change{\url{https://lts2.epfl.ch/web-audio-inpainting/}}}. Table
\ref{tab:parameters} provides a summary of the algorithm parameters used for the
evaluations.
\subsection{Verification}\label{ssec:experimentSanity}
Here, we address the question whether the algorithm perfectly recovers the gap
when an exact copy of the missing segment is present within the reliable signal.
For this purpose, we used a set of $16$ uncorrupted audio signals with various
content and at the sampling rate of $44100$ Hz. First, redundant signals were
created by repeating the signal, i.e. placing a copy of the signal at its end.
Then, each redundant signal was corrupted by creating a gap of $2$ seconds. For
each signal, the experiment was repeated five times with randomly chosen position of
a gap, yielding $80$ corrupted signals. Then the algorithm was applied on each
of the corrupted signal. \change{In all reconstructions, the $\ell^2$-norm difference
between the original and reconstructed signals was in the range of numerical
precision, implying that each corrupted signal was perfectly restored.}
Hence, we consider the implementation of the presented algorithm as verified.
\subsection{Computational complexity}\label{ssec:experimentComp}
The algorithm can be separated into different steps that all have different computational requirements.
Here, we investigated the individual costs of each step and their relative importance in the overall performance of the algorithm.
The evaluation was performed on a \change{modern notebook ($2.5$ GHz Intel i7, 2 cores, $16$ GB RAM)} for the same set of corrupted signals \change{as in
Sec.~\ref{ssec:experimentMain}}. Table \ref{tab:timing} shows mean and standard
deviation of the computation time per minute of audio signal. On average, each
minute of audio signal required $2.47$-s computation time for the reconstruction.
The feature computation, graph creation and the selection of the optimal
transition\change{, performed on the reduced sparse graph (see Figure~\ref{fig:subgraph}), }
scale linearly with the length of the provided reliable data, in
terms of both storage and time complexity. As a result, our result consists of the timing per minute of analyzed music.
In all our experiments, the reliable
data was given by a full song, without the corrupted segment.
\change{Note that linear complexity can only be achieved by considering the
reduced graph. For the full sparse graph $\mathbf W_0$, complexity of the graph creation
is $\mathcal O(N \log N)$ and the transition selection would even scale roughly quadratically, i.e.
be $\mathcal O(N^2)$. Even if the selection is restricted to the range considered in the
reduced graph, linear complexity would be out of reach. Therefore, the computation
time per minute is not a reliable indicator anymore. We just remark that on the dataset used,
the graph construction was on average $~8$ times slower, while the average duration for
transition selection increased by a factor of $~40$, when performed on the full sparse graph.
Although we did not systematically evaluate memory usage of the method, it should be noted that
restricting to the reduced graph is considerably more efficient in that regard, as well.}
If multiple
corruptions are to be removed using the same set of reliable data, the algorithm
benefits from the fact that features only need to be computed once. Since the
feature computation is the bottleneck of the method (this can be seen in Table
\ref{tab:timing}), this may lead to significant boosts of computational performance in the case
of multiple gaps.
\begin{table}[ht!]
\begin{center}
\change{\begin{tabular}{|l|c|c|}
\hline
{\bf Processing step} & {\bf Reduced graph} (Mean) & {\bf Reduced graph} (STD)\\% & {\bf Full graph} \\
\hline
\hline
Feature extraction & $1.84$ & $0.18$\\% & $1.70$ / $0.11$ \\
\hline
Graph construction & $0.50$ & $0.06$\\% & $3.89$ / $2.30$ \\
\hline
Transition selection & $0.02$ & $0.004$\\% & $0.84$ / $2.29$ \\
\hline
Signal reconstruction & $0.04$ & $0.007$\\% & $0.05$ / $0.02$ \\
\hline
\hline
{\bf Total} & $2.47$ & $0.18$ \\% & $6.56$ / $3.60$ \\
\hline
\end{tabular}}
\end{center}
\caption{\label{tab:timing}\change{Average execution time of the proposed methods per minute of provided audio (Based on a database of $16$ songs) for the reduced graphs, see Fig.~\ref{fig:subgraph}.
}
}
\end{table}
\section{Perceptual evaluation}\label{ssec:experimentMain}
In order to estimate the potential of the proposed algorithm for music, we conducted a psychoacoustic test, in which we evaluated the impact of the artifacts occurring from inpainting various songs from a music database. In particular, we were interested in addressing the following questions:
\begin{enumerate}
\item How often are subjects able to detect an alteration (detectability)? The answer gives us access to how often our algorithm is able to fool the listener.
\item How precisely can subjects pinpoint the alteration? The answer gives us an indication of the inpainting quality and of the confidence of the test subject.
\item How disturbing are the detected artifacts (severity)? The answer provides some good {insights into} the reconstruction quality even when the listener is not fooled.
\item Is the familiarity of the song correlated with the detectability or the severity? The answer gives some {intuition about} the quality of the reconstruction and ensures that we are not only fooling the non-familiar test subjects.
\end{enumerate}
In order to ensure that our experiment provides meaningful results truly describing the potential of the proposed algorithm, our subjects were familiar with the tested music genres and we have collected ratings for familiarity and liking the songs.
\subsection{Testing methodology}
\paragraph{Material}
The sound material consisted of songs from the following genres: \change{pop, rock,
jazz, classical. These genres were selected to cover the most common listening habits and, with respect to music structure also include many other, similar genres like blues, country, folk, oldies, hip-hop, etc}. Six songs per genre were selected from hundreds of songs with the
aim to well-represent the genre.
\paragraph{Subjects}
In order to test subjects familiar with our material, in a self-assessment questionnaire, \change{a candidate had to provide the average weekly listening duration (in hours) to the genres pop, rock, jazz, classical, and others}. For the evaluation, only candidates listening at least 4 hours per week to music from \change{all four main genres in total} were considered. In total, 15 subjects were selected for the test. They were paid on an hourly basis.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=1.0\linewidth]{figures/Fig_Screenshot.png}
\end{center}
\caption{\change{The interface used in the experiment. See text for more details.
\label{fig:Screenshot}}}
\end{figure}
\paragraph{Task}
In each trial, subject listened to a sound stimulus and was asked to pay attention to a potential
artifact \change{(see Fig.~\ref{fig:Screenshot})}. A slider scrolled horizontally while the sample was played indicating the current position within a stimulus. The subject was asked to tag the artifact's position by aligning a second slider with the begin of the perceived artifact. Then, while listening again to the same stimulus, the subject was asked to confirm (and re-align if required) the slider position and answer three questions:
\begin{enumerate}
\item \emph{Severity (S)}: How poor was it ("Wie schlimm ist es")? The possible answers were: (0) no issue ("Kein Fehler"), (1) not disturbing ("Nicht störend"), (2) mildly disturbing ("Leicht störend"), and (3) not acceptable ("Nicht akzeptabel").
\item \emph{Familiarity (F)}: How familiar are you with this song ("Wie gut kennen Sie dieses Stück"): (0) never heard before ("Noch nie gehört"), (1) I have heard it before ("Schon mal gehört"), (2) I often listen to ("Höre ich öfters"), (3) I know it well ("Kenne ich gut"), and (4) I can play/sing it ("Kann ich spielen/singen").
\item \emph{Liking (L)}: How do you like this song ("Wie gefällt Ihnen dieses Stück"): (0) not at all ("Gar nicht"), (1) I can not tell ("Kann nicht sagen"), (2) nice ("Nett"), (3) very nice ("Sehr nett"), and (4) amazing ("Bin begeistert").
\end{enumerate}
The questions were answered by tapping on the corresponding category. Then, the subject continued with the next trial by tapping the "next" button.
Before the experiment, the subject was informed about the purpose and procedure of the experiment and an exemplary reconstruction was presented. Any questions with respect to the procedure were clarified.
\paragraph{Conditions}
Three conditions were tested. For the \textit{inpainting condition}, the song was corrupted at a random place with the gap of 1~s duration and then reconstructed with the default parameters from Tab.~\ref{tab:parameters}. The
reconstructed song was cropped 2 to 4 seconds (randomly varying) before and
after the gap resulting in samples of 5 to 9-s duration. \change{The gap was not allowed to be within the first and last 30 s of the song, but the inpainting was allowed to use the full song for processing.} For the \textit{reference
condition}, the song was cropped at a random place with a duration varying from 5
to 10 seconds. The reference condition did not contain any artifact and was used to estimate the sensitivity of a subject. For the \textit{click condition}, a click was superimposed to the song at
a random position and the result was cropped 2.5 to 4.5 s before and after the
click's position resulting in samples of 5 to 9-s duration. The artifact in this condition was used as a reference artifact and was clearly audible.\change{\footnote{\change{For other music genres like electronic music, the click might not be always audible and an other type of reference artifact would have been required.}}}
In total, three inpainted, one reference, and one click conditions were created per song.
\change{The combination of genres}, songs-per-genre, and conditions-per-song resulted in a
block of 120 stimuli. All stimuli were normalized in the level (the click
condition was normalized before superimposing the click). Within the block, the
order of the stimuli and conditions was random.
Each subject was tested with two blocks, resulting in 240 trials per subject in
total. Subjects were allowed to take a break at any time, with one planned break
per block. For each subject, the test lasted approximately 2.5 hours.
\subsection{Results}
\paragraph{Detection rate of the artifacts}
The detection results are shown in the left panel of Fig.~\ref{fig:EvalPercentage}. The average detection rates for the click, inpainting, and reference conditions were $95.6\pm5.0\%$, $40.1\pm19.2\%$, and $28.6\pm17.9\%$, respectively. The high detection rate and small variance in the click condition demonstrate a good attention of our subjects, for whom even a single click was clearly audible. The clearly non-zero rate in the reference condition shows that our subjects were highly motivated in finding artifacts. The detection rate in the inpainted condition was between those from the reference and click conditions. Note that the reference condition did not contain any artifacts, thus, the artifact's detection rate in that condition is here referred to as the false-alarm rate.
The large variance of the false-alarm rate shows that it is listener-specific. Thus, for further analysis, the detection rates from the inpainted condition were related to the listener-specific false-alarm rate, i.e., the sensitivity index $d'$ was used \cite{macmillan2004detection}. The false-alarm rate can be considered as a reference for guessing, thus, $d'=1$ indicates that the artifacts was detected at the level of chance rate. The right panel of Fig.~\ref{fig:EvalPercentage} shows the statistics of $d'$ for the inpainting and the click conditions. For the click condition, the average across all subjects was $4.36\pm1.91$, again demonstrating a good detectability of the clicks. For the inpainting condition, the average $d'$ was $1.49\pm0.42$, i.e., slightly above guessing ($d'= 1$). A t-test performed on listener's $d'$s showed a significant ($p = 0.0005$) difference from guessing, \change{indicating that the our listeners, as a group, were able to often detect the artifacts better than guessing. A
listener-specific analysis, however, showed that only seven out of our 15 subjects were able to detect the inpainting better than chance, as revealed by a 2-by-2 contingency table analysis with the false-alarm and inpainting-detection rates evaluated at a significance level of 0.05. }
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.75\linewidth]{figures/Fig7_DetectabilityArtifacts.png}
\end{center}
\caption{Detectability of artifacts is much lower than those of clicks but slightly higher than guessing. Left: Statistics of the rate of perceived artifacts across all subjects. Right: Statistics of the sensitivity index $d'$, i.e., the inpainting-detection rate relative to the false-alarm rate, across all subjects. $d'$ of 1 corresponds to the chance rate. Condition: Reference (R), inpainted (I), and click (C). Statistics: Median (circle), 25\% and 75\% quartiles (thick lines), coverage of 99.3\% (thin lines, assuming normal distribution), outliers (crosses, horizontally jittered for a better visibility). \label{fig:EvalPercentage}}
\end{figure}
\paragraph{Influence of familiarity on the detectability}
A natural question that arises for this method is, in how far familiarity with a song will influence the detectability of the artifacts. While a comprehensive answer to this question is beyond the scope of this paper and would require a whole new study, here we aim at a brief impression for our subject pool.
Fig.~\ref{fig:EvalDetectFam} shows the detection rate (left panel) and the $d'$ (right panel) as functions of the familiarity ratings.
While there seems to be a correlation of detectability and familiarity, surprisingly the link is not very strong. Arguably, there seems to be nearly no difference in the detection rates between songs rated with familiarity rating between 2 ("I often listen to") and of 4 ("I can sing/play it"), while there seems to be some difference to the other ratings of less familiarity. Interestingly, even for the very familiar songs the detection rate is much lower than for clicks and the $d'$ is only twice as large as that for the the chance rate.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.75\linewidth]{figures/Fig8_Familiarity.png}
\end{center}
\caption{Detectability is not much related with the familiarity. Left: Statistics of the rate of perceived artifacts across all subjects as a function of the familiarity rating. Right: Statistics of the sensitivity index $d'$ as a function the familiarity rating. All other conventions as in Fig.~\ref{fig:EvalPercentage}.
\label{fig:EvalDetectFam}}
\end{figure}
\paragraph{Detection of the artifact position}
\change{
Subjects who successfully detected an artifact should be able to provide an information about its position within the stimulus, i.e., the perceived position of the artifact should correlate with its actual position. The left panel of Fig.~\ref{fig:EvalCorr} shows the perceived positions plotted versus the actual positions of the artifacts, for an exemplary average listener. The reported perceived artifact's position might refer to gap's begin or end, with the choice even varying from stimulus to stimulus. Thus, we correlated the reported position with the begin, the end, and the nearer of the two positions (referred to as "best choice"). The "best choice" positions are highlighted by triangles.
Across all subjects, correlation coefficients' statistics is shown in the center panel of Fig.~\ref{fig:EvalCorr}. The moderate correlations indicate that as soon as our subjects detected an artifact, they had some estimate of its position within the stimulus. In contrast, for the clicks, the high correlation indicates that our subjects were able to exactly determine and report the position of the click artifact.
In order to determine the precision in the reporting the artifact's position, we also calculated the difference between the perceived and actual artifact's position. The standard deviation of these differences calculated for a subject is referred to as the precision error. Their statistics across subjects is shown in the right panel of Fig.~\ref{fig:EvalCorr}. For the click condition, the average precision error across all subjects was $157\pm130$\,ms. It describes the procedural precision of subjects within our task. For the inpainting condition, the average precision error considering the artifact's begin, end, and "best choice" as the actual position was $1232\pm180$\,ms, $1247\pm199$\,ms, and $1069\pm115$\,ms, respectively. The "best choice" shows the lowest precision errors, being more than six times larger than the procedural precision error. This indicates that even if detected, our subjects had large difficulties to determine the artifact's position and these difficulties did not originate from the
task.
}
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.98\linewidth]{figures/Fig9_PositionCorrelation.png}
\end{center}
\caption{The position of perceived artifacts is weakly correlated with their actual position. Left: Perceived versus actual artifact's begin and positions (blue squares and green circles, respectively) for an exemplary subject. Triangles show the "best choice", i.e., perceived positions being nearer to either begin or end actual positions. Center: Statistics of the correlation coefficients for all subjects. Right: Statistics of the precision error for all subjects. B, E: perceived position versus begin and end of the artifact, respectively, in the inpainting condition. X: perceived position versus "best choice" in the inpainting condition. C: perceived position of the click in the click condition. $CC_B$, $CC_E$, $CC_X$: cross-correlation coefficient for the condition B, E, and X, respectively, of the exemplary listener. All other conventions as in Fig.~\ref{fig:EvalPercentage}). \label{fig:EvalCorr}}
\end{figure}
\paragraph{Disturbance rate of detected artifacts}
Finally, we have analyzed the ratings we have collected. The left panel of Fig.~\ref{fig:Ratings} shows the statistics of the severity ratings reported in the inpainted and click conditions. For the click condition, most of the ratings were between 1 ("not disturbing") and 3 ("not acceptable") with an average across all subjects of $2.00\pm0.55$. This indicates that on average, our subjects rated the clicks as disturbing. In contrast, for the inpainted condition, most of the ratings were between 0 ("no issue") and 1 ("not disturbing") with an average of $0.60\pm0.33$. This indicates that on average, our subjects rated the inpainting results halfway between "no issue" and "not disturbing".
This analysis considered all inpainted stimuli so far, ignoring the fact that for some of them our subjects detected the artifact and for some not. A statistic of severity ratings considering detected artifacts only (i.e., $S>0$) is shown in the center part of the left panel in Fig.~\ref{fig:Ratings}. The average across all subjects was $1.46\pm0.35$. This is higher than the average considering all severity ratings, but still significantly ($p=0.0002$) lower than the severity of the clicks as revealed by a paired t-test calculated between the ratings for clicks and inpainted but detected artifacts. This indicates that even when the inpainting artifacts were perceived, their severity was rated significantly lower than that of the clicks.
\begin{figure}[htb!]
\begin{center}
\includegraphics[width=0.95\linewidth]{figures/Fig10_Ratings.png}
\end{center}
\caption{Statistics of ratings across all subjects. Left: severity ratings (S). Center: Familiarity (F) and liking (L) ratings. Condition: Inpainted (I), click (C), ratings considering perceived artifacts only (S>0). Right: Statistics of Pearson's correlation coefficients between S and F (SF), S and L (SL), as well as F and L (FL). All other conventions as in Fig.~\ref{fig:EvalPercentage}. \label{fig:Ratings}}
\end{figure}
\paragraph{Influence of the familiarity on the severity}
The stimulus' familiarity and liking might also have influenced our experimental outcome. The average ratings for the familiarity and liking are shown in the center panel of Fig.~\ref{fig:Ratings}. Most of the familiarity ratings were between category 1 ("I have heard it several times") and 2 ("I often listen to"), with an across-subject average of $1.23\pm0.41$. Considering the perceived artifacts only (i.e., $S>0$), the average increased to $1.44\pm0.45$. This increase was significant ($p=0.022$, paired t-test on all and the perceived only ratings), indicating that our subjects were slightly more familiar with stimuli containing detectable artifacts. The liking ratings were mostly between 1 ("I cannot tell") and ("very nice"), with an average of $2.10\pm0.43$. Considering perceived artifacts only, the average increased to $2.13\pm0.44$. This increase was not significant ($p=0.68$, paired t-test between all and the perceived only ratings). As it seems, the artifact's detectability was not related to the
song liking.
The link between the severity of an artifact and the familiarization and/or liking ratings was further investigated by calculating the Pearson's correlation coefficients between the severity and other ratings. The right panel of Fig.~\ref{fig:Ratings} shows the group statistics of the correlation coefficients, which, on average, were $0.14\pm0.16$ and $0.09\pm0.17$ for the correlation of severity with familiarity and liking, respectively. Such low correlations indicate that neither the familiarity nor liking was clearly linked with the perceived artifact's severity. Out of curiosity, also the correlation between the familiarity and liking was calculated, resulting in an across-subject average of $0.54\pm0.26$. This correlation indicates a good link between the familiarity and liking of our stimuli, but also raises evidence that familiarity and liking are not fully equivalent.
\section{Conclusions}
We have introduced a method for the restoration of audio signals in the presence
of corruption/loss of data over an extended, connected period of time. Since,
for complex audio signals, the length of the lost segment usually prohibits
the inference of the correct data purely from the adjacent reliable data, our
solution is based on the larger scale structure of the underlying audio
signal. The reliable data is analyzed, detecting spectro-temporal
similarities, resulting in a graph representation of the signal's temporal
evolution that indicates strong similarities. Inpainting of the lost data is
then achieved by determining two suitable transitions between the border
regions around the corrupted signal segment and a region that is considered
to be similar. In other words, the algorithm jumps from shortly before the
\emph{gap} to a similar section of the audio signal and, after some time,
back to a position shortly after the gap, effectively exchanging the
corrupted piece with a suitable substitute. Consequently, the algorithm is
capable of efficiently exploiting naturally occurring redundancies in the
reliable data.
In order to test the efficiency of our algorithm, we have conducted a psychoacoustic evaluation. The results show that our listeners were able to detect $40\%$ of the artifacts implying that our method completely fooled our listeners more than $60\%$ of the time. Our listeners showed a false-alarm rate of $28\%$, indicating that sensitivity of correctly detecting a gap was with $d'=1.49$ rather low (as compared with $d'=4.36$ for well-detectable clicks and with $d'=1$ for the chance rate). \change{In fact, listener-specific analysis showed that only seven out of 15 tested listeners were able to detect the inpainting on a statistical significant level. } Our study showed two additional quality signs of our method. First, the detected artifacts were rated on average between ``not disturbing'' and ``mildly disturbing''. Second, even though detected, our subjects \change{only vaguely determined the artifact's position, with the six-fold detection precision error than that in the reference condition. While our
test was limited to four music genres, they covered many music structures usually found in other genres. However, inpainting performed on a very different genre like the contemporary electronic music might have led to different results, both numeric and perceptual.}
Besides having built and tested a novel audio inpainting algorithm, it is
worth noting that the graph constructed with our method gives an intuitive
analysis of the signal at hand, exposing self-similarities and global
structure and can be used for a number of different purposes. For example, a
song can be re-composed by following the edges of the graph while respecting
the global music structure. \change{Multiple matches in a highly repetitive song can be used as a tool for further song modifications, offering a creative tool for algorithmic composing, e.g., in the field of contemporary electronic music. }
\change{Similarity graphs can be used in many
applications, thus, it is important to further improve this kind of signal representation.}
Hence, future work includes closing the gap between the internal similarity
measures and human hearing by incorporating perceptually motivated similarity
measures derived, possibly, from a perceptually-motivated
representation~\cite{DBLP:journals/corr/NecciariHBP16} or a computational
model of the auditory system~\cite{Irino:2006b}. Such a modification will
greatly improve the reliability of the algorithm and its results. It seems
worth noting, however, that even after considering an auditory model,
reliable retrieval of strongly context-sensitive data such as speech and
singing voice will require additional contextual information and might be
better achieved by a generative approach~\cite{ saino2006hmm}, applied after
separating voice and music in the signal~\cite{li2007separation}.
\section*{Acknowledgment}
\change{We thank the reviewers and the editor for their review this publication and their helpful suggestions.} We thank Pierre Vandergheynst for his support during this project. His ideas and
suggestions have helped significantly to this contribution.
This work has been supported by the Swiss Data Science Center and by the Austrian Science Fund (FWF) projects FLAME
(\textit{Frames and Linear Operators for Acoustical Modeling and Parameter
Estimation}; Y 551-N13) and MERLIN (\textit{Modern methods for the restoration of
lost information in digital signals}; I 3067-N30)
\begin{table*}[ht!]
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
\bf{ Quantity} & \bf{Variable used }& \bf{Default value} & \bf{Unit }\\
\hline
\hline
\multicolumn{4}{|c|}{\emph{Audio features}} \\
\hline
Maximum sampling frequency & $\xi_{s,\text{max}}$ & $12'000$ &$Hz$ \\
\hline
Size of the patch & $a$ & $128$ & samples \\
\hline
Number of frequencies & $M$ & $1024$ & - \\
\hline
Length of the window & $L_w$ & $M$ & samples \\
\hline
Type of window & - & 'Itersine' & - \\
\hline
Dynamic range & $p$ & $50$ & dB \\
\hline
Trade-off between the amplitude and phase & $\lambda$ & $3/2$ & - \\
\hline
\multicolumn{4}{|c|}{\emph{Graph}} \\
\hline
Initial number of neighbors & $K$ & $40$ & - \\
\hline
Kernel length & $L_k$ & $40$ & - \\
\hline
Hard threshold for the weight matrix & $t_w$ & $2$ & - \\
\hline
\multicolumn{4}{|c|}{\emph{Optimization}} \\
\hline
Regularization parameter 1 & $\gamma_1$ & $1$ & - \\
\hline
Regularization parameter 2 & $\gamma_2$ & $100$ & - \\ \hline
\end{tabular}
\end{center}
\caption{\label{tab:parameters}Default parameters of the algorithm}
\end{table*}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
Current experiments generally rely on neutral beams, directed toroidally, in order to induce the plasma to rotate. This toroidal rotation has been experimentally proven to stabilize resistive wall modes (a class of MHD instabilities that can cause disruptions) thereby enabling sustained discharges with a plasma $\beta$ that violates the Troyon limit \cite{StraitExpRWMstabilizationD3D1995, SabbaghExpRWMstabilizationNSTX2002, deVriesRotMHDStabilization1996, ReimerdesRWMmachineComp2006}. Since most designs of reactor-scale devices violate the Troyon limit \cite{PolevoiITERscenario2002} and are intolerant to disruptions \cite{SugiharaITERdisruptionVDE2007}, driving fast toroidal rotation in large devices is critical. Throughout this paper we will use the ITER design \cite{AymarITERSummary2001} as an example of a large device in order to provide realistic numbers and a frame of reference. Here we note that numerical analysis indicates that the slowest rotation able to stabilize resistive wall modes in ITER has an on-axis Alfv\'{e}n Mach number around $0.5\% - 5\%$ \cite{LiuITERrwmStabilization2004}. The precise value depends on the exact numerical model used, but is significantly lower for broader rotation profiles.
Driving rotation in large plasmas is difficult because they have more inertia and require more energetic neutral beams to penetrate to the magnetic axis. Because of the velocity scalings of momentum versus energy, more energetic neutral beams inject less momentum per unit power \cite{ParraMomentumTransitions2011}. This explains why the neutral beams in ITER are not expected to drive substantial toroidal rotation \cite{LiuITERrwmStabilization2004, DoyleITERconfinement2007}. Therefore, unless the momentum pinch effect can be used to dramatically amplify the driven rotation \cite{WeisenMomPinch2012} or bring in rotation from the edge \cite{OmotaniEdgeNeutralRot2016, TalaMomPinch2011}, we must turn to ``intrinsic'' rotation (i.e. spontaneous rotation that is observed in the absence of external momentum injection \cite{RiceExpIntrinsicRotMeas2007}). This rotation is generated by the plasma through turbulent transport of momentum. Because it is generated by the plasma itself, intrinsic rotation would be expected to scale well to large devices. However, the gyrokinetic equation, which is thought to govern turbulence in the core of tokamaks, possesses a particular symmetry \cite{ParraUpDownSym2011, SugamaUpDownSym2011, CamenenPRLSim2009} that implies this intrinsic momentum flux must be small in $\rho_{\ast} \equiv \rho_{i} / a \ll 1$, the ratio of the ion gyroradius to the tokamak minor radius. Fortunately, there is one mechanism that breaks this symmetry and is capable of spontaneously generating rotation in the core of a stationary plasma: up-down asymmetry in the magnetic geometry.
If the flux surfaces in a tokamak are up-down asymmetric (i.e. do not have mirror symmetry about the midplane), then the momentum flux is no longer constrained to be small in $\rho_{\ast} \ll 1$. In principle, up-down asymmetric flux surfaces are no more difficult to create than up-down symmetric surfaces, but all existing devices have been designed with nearly up-down symmetric flux surface shapes in mind. Hence, the ability of a device to create a particular up-down asymmetric surface depends strongly on the specifics of the shaping coils and the vacuum vessel. The TCV tokamak \cite{HofmannTCVOverview1994}, which was designed to accommodate strong shaping, has been used to experimentally investigate flux surfaces with a single up-down asymmetric shaping mode \cite{CamenenPRLExp2010}. As expected, a large change in the rotation profile was observed when the up-down asymmetry of the plasma shape was varied. Subsequent gyrokinetic simulations \cite{BallMomUpDownAsym2014}, which give results consistent with the TCV experiments, indicate that up-down asymmetry is a feasible method to generate the current experimentally-measured rotation levels in reactor-sized devices.
Configurations with only a single up-down asymmetric shaping mode drive rotation through the direct interaction of toroidicity (which defines up versus down) and the shaping mode. Recent analytic work \cite{BallMirrorSymArg2016, BallDoctoralThesis2016} demonstrates that adding a second shaping effect introduces two new physical mechanisms that have the potential to enhance the rotation. First, the tilting symmetry presented in reference \cite{BallMirrorSymArg2016} shows that flux surfaces with only a single shaping mode $m$ must have momentum flux that is exponentially small in $m \gg 1$. Including two shaping effects allows them to beat together to produce an up-down asymmetric envelope on the connection length-scale that can interact with toroidicity to drive rotation. This breaks the tilting symmetry and permits the momentum flux to have a stronger scaling (i.e. polynomially small in $m \gg 1$). These scalings indicate that using low order shaping effects and combining different shaping effects to make asymmetric envelopes can effectively drive fast intrinsic rotation. Physically, high order shaping effects do not effectively drive rotation because the turbulent eddies, which are extended along the magnetic field line, average over small-scale variation in the magnetic equilibrium. Second, looking in the screw pinch limit (i.e. large aspect ratio limit) of a tokamak we learn that flux surfaces with mirror symmetry about any line in the poloidal plane do not drive any intrinsic rotation \cite{BallDoctoralThesis2016}. Including a second shaping effect can break mirror symmetry, allowing rotation to be driven through the direct interaction between the two shaping effects (completely independently of toroidicity). These two mechanisms dominate in certain regimes (i.e. the $m \gg 1$ and large aspect ratio limits) and bring in fundamentally new physics, but their importance in more realistic geometries is still unclear.
Together all of these results indicate that low order shaping effects are optimal for maximizing intrinsic rotation and it is important to explore non-mirror symmetric configurations with an up-down asymmetric envelope. In this context, there are two options. The first is to introduce up-down asymmetric elongation using external poloidal field coils and then rely on the Shafranov shift (i.e. the shift in the magnetic axis due to toroidicity) to break the mirror and tilting symmetries. This appears optimal because it makes use of the lowest possible shaping modes (i.e. $m = 1$ and $m = 2$). However, this strategy has the drawbacks that the effect of the Shafranov shift is formally small in aspect ratio and the direction and magnitude of the shift is a consequence of the plasma $\beta$ profile and the global MHD equilibrium. Hence it is not independently controlled by external coils. The second option is to use external coils to introduce both elongation and triangularity (i.e. $m = 2$ and $m = 3$ shaping) into the flux surface shape in order to directly break mirror symmetry and create an envelope that breaks the tilting symmetry. Both modes are lowest order in aspect ratio and can be directly controlled by external shaping magnets, but this relies on higher order shaping modes than the first option. Practically speaking these two strategies are intertwined as the divertor geometry nearly always introduces some triangularity into the flux surfaces and the Shafranov shift exists regardless of the shape of flux surfaces. Nevertheless, for simplicity it is useful to distinguish them and examine each option independently. In this work we will explore the former: the influence of the Shafranov shift and the effect of the $\beta$ profile on the turbulent momentum flux in the core of tokamaks.
In section \ref{sec:MHDequil} we use the Grad-Shafranov equation to estimate the magnitude and direction of the Shafranov shift in a tokamak with a tilted elliptical boundary. To do so we start in section \ref{subsec:analyticCalc} by expanding the Grad-Shafranov equation in the large aspect ratio limit to write the lowest and next order analytic solutions for a linear toroidal current profile as a Fourier series in poloidal angle. In section \ref{subsec:numCoeffCalc}, we calculate the Fourier coefficients needed to match the tilted elliptical boundary condition. In section \ref{subsec:ECOMcomparison}, we find the dependence of the Shafranov shift on the boundary tilt angle and show that the shift is insensitive to the shape of both the current and pressure profiles (when the geometry, total plasma current, and average pressure gradient are kept fixed). These analytic results are verified using equilibrium calculations performed with the numerical Grad-Shafranov solver ECOM \cite{LeeECOM2015}. Section \ref{sec:gyrokineticSims} contains the results from nonlinear gyrokinetic simulations of the equilibria calculated in section \ref{sec:MHDequil}. Section \ref{subsec:inputParameters} starts by using the results of the MHD analysis to generate local equilibria for the gyrokinetic simulations. Section \ref{subsec:results} details the results of several numerical scans aimed at illuminating the effect of the Shafranov shift and the $\beta$ profile on momentum transport. In section \ref{subsec:betaPrime} we discuss the sensitivity of the momentum transport to changes in the magnetic equilibrium caused by altering the local gradient of $\beta$. Furthermore, in section \ref{subsec:betaProfile} we consider the impact of changing the shape of the radial profile of $\beta$. Section \ref{sec:conclusions} contains a summary of the results and some concluding remarks.
\section{MHD equilibrium calculation of the Shafranov shift}
\label{sec:MHDequil}
In this section we will calculate a general analytic solution to the Grad-Shafranov equation for a linear (in poloidal flux) toroidal current profile to lowest and next order in an expansion in large aspect ratio. The zeroth and first order solutions are needed because the Shafranov shift does not appear to lowest order. The analytic solution will contain Fourier coefficients, which in general must be calculated numerically to achieve a tilted elliptical boundary flux surface. Making use of our numerically calculated Fourier coefficients, we will argue that varying the shape of the current profile and the shape of the pressure profile (while keeping the geometry, total plasma current, and average pressure gradient fixed) does not significantly affect the Shafranov shift. These theoretical results are verified against the equilibrium code ECOM.
Due to the insensitivity of the Shafranov shift to the exact current and pressure profiles, we are free to use the constant current case for input into the gyrokinetic simulations of section \ref{sec:gyrokineticSims}. This is helpful as the Fourier coefficients in the constant current equilibria can be calculated analytically.
\subsection{Analytic solution for a linear current profile}
\label{subsec:analyticCalc}
The geometry of a tokamak equilibrium is governed by the Grad-Shafranov equation \cite{GradGradShafranovEq1958},
\begin{align}
R^{2} \ensuremath{\vec{\nabla}} \cdot \left( \frac{\ensuremath{\vec{\nabla}} \psi}{R^{2}} \right) = - \mu_{0} R^{2} \frac{d p}{d \psi} - I \frac{d I}{d \psi} , \label{eq:gradShaf}
\end{align}
where $R$ is the tokamak major radial coordinate, $\psi$ is the poloidal magnetic flux divided by $2 \pi$, $\mu_{0}$ is the permeability of free space, $p$ is the plasma pressure, $I \equiv R B_{\zeta}$ is the toroidal magnetic field flux function, $\vec{B}$ is the magnetic field, and $\zeta$ is the toroidal angle. Note that the effect of the $\beta \equiv 2 \mu_{0} p / B^{2}$ profile only enters through the gradient of the pressure. In order to investigate the behavior of the Shafranov shift in a tilted elliptical geometry we will expand in the large aspect ratio limit, i.e. $\epsilon \equiv a / R_{0} \ll 1$ where $a$ is the tokamak minor radius and $R_{0}$ is the major radial location of the center of the boundary flux surface. We will take the typical orderings for a low $\beta$, ohmically heated tokamak \cite{FreidbergIdealMHD1987pg126}:
\begin{align}
\frac{B_{p}}{B_{0}} \sim \epsilon, \hspace{10pt}
\frac{2 \mu_{0} p}{B_{0}^{2}} \sim \epsilon^{2} , \label{eq:gradShafOrderings}
\end{align}
where $B_{0}$ is the on-axis toroidal magnetic field and $B_{p} = | \ensuremath{\vec{\nabla}} \psi | / R$ is the poloidal magnetic field. Also, we must expand $\psi = \psi_{0} + \psi_{1} + \ldots$, $I = I_{0} + I_{1} + I_{2}$, and $p = p_{2}$, where the subscripts indicate the order of the quantity in $\epsilon$ relative to the lowest order contributions of $\psi_{0} \sim a R_{0} B_{p} \sim \epsilon^{2} R_{0}^{2} B_{0}$ and $I_{0} \sim R_{0} B_{0}$. To $O \left( \epsilon^{-1} B_{0} \right)$ we find that the Grad-Shafranov equation is
\begin{align}
- I_{0} \frac{d I_{1}}{d \psi_{0}} - I_{1} \frac{d I_{0}}{d \psi_{0}} = 0 . \label{eq:gradShafNegOrder}
\end{align}
Since $I_{0} = R_{0} B_{0}$ is a constant, this requires that $I_{1}$ also be a constant. We are free to absorb $I_{1}$ into $I_{0}$ and set $I_{1} = 0$. Hence, using $r \sim a$ we find to $O \left( B_{0} \right)$ that
\begin{align}
\frac{1}{r} \frac{\partial}{\partial r} \left( r \frac{\partial \psi_{0}}{\partial r} \right) + \frac{1}{r^{2}} \frac{\partial^{2} \psi_{0}}{\partial \theta^{2}} = - \mu_{0} R_{0}^{2} \frac{dp_{2}}{d \psi_{0}} - I_{0} \frac{dI_{2}}{d \psi_{0}} \label{eq:gradShafLowestOrder}
\end{align}
and to $O \left( \epsilon B_{0} \right)$ that
\begin{align}
\frac{1}{r} \frac{\partial}{\partial r} \left( r \frac{\partial \psi_{1}}{\partial r} \right) &+ \frac{1}{r^2} \frac{\partial^{2} \psi_{1}}{\partial \theta^{2}} = \psi_{1} \frac{d}{d \psi_{0}} \left( -\mu_{0} R_{0}^{2} \frac{d p_{2}}{d \psi_{0}} - I_{0} \frac{d I_{2}}{d \psi_{0}} \right)\label{eq:gradShafNextOrder} \\
& - 2 \mu_{0} r R_{0} \frac{d p_{2}}{d \psi_{0}} \Cos{\theta} + \frac{\Cos{\theta}}{R_{0}} \frac{\partial \psi_{0}}{\partial r} - \frac{\Sin{\theta}}{r R_{0}} \frac{\partial \psi_{0}}{\partial \theta} , \nonumber
\end{align}
where $r \equiv \sqrt{\left( R - R_{0} \right)^{2} + Z^{2}}$ is the distance from the center of the boundary flux surface, $\theta \equiv \ArcTan{Z / \left( R - R_{0} \right)}$ is the usual cylindrical poloidal angle, and the axial location of the center of the boundary flux surface is assumed to be at $Z = 0$.
Like references \cite{BallMomUpDownAsym2014, BallMastersThesis2013}, we will develop our intuition by investigating how the Shafranov shift changes with three simple, but realistic toroidal current profiles: constant, linear peaked, and linear hollow (in poloidal flux). Using Ampere's law and $\vec{B} = I \ensuremath{\vec{\nabla}} \zeta + \ensuremath{\vec{\nabla}} \zeta \times \ensuremath{\vec{\nabla}} \psi$ one can show that the toroidal current is related to the right-hand side of the Grad-Shafranov equation through
\begin{align}
- \mu_{0} R^{2} \frac{d p}{d \psi} - I \frac{d I}{d \psi} = \mu_{0} j_{\zeta} R , \label{eq:toroidalCur}
\end{align}
where $j_{\zeta}$ is the toroidal current density in the plasma. We will parameterize all three profiles (i.e. constant, peaked, and hollow) by
\begin{align}
- \mu_{0} R_{0}^{2} \frac{dp_{2}}{d \psi_{0}} - I_{0} \frac{dI_{2}}{d \psi_{0}} = \mu_{0} j_{\zeta 0} R_{0} = j_{N} \left( 1 - f_{N} \psi_{0} \right) , \label{eq:currentProfiles}
\end{align}
where $j_{\zeta 0}$ is the lowest order current density in the aspect ratio expansion, $j_{N}$ is a positive constant, $f_{N} \in \left[ - \psi_{0 b}^{-1}, \psi_{0 b}^{-1} \right]$ determines the slope of the current profile, and $\psi_{0 b}$ is the lowest order value of the poloidal flux on the boundary flux surface. The constant current case is achieved by setting $f_{N} = 0$, while the hollow current case arises from allowing $f_{N}$ to be negative.
Additionally, from equation \refEq{eq:gradShafNextOrder} we see that it will be necessary to distinguish the contributions to the current from the pressure and magnetic field terms in equation \refEq{eq:toroidalCur}. Like the toroidal current, we will assume the pressure gradient has the form of
\begin{align}
- \mu_{0} R_{0}^{2} \frac{dp_{2}}{d \psi_{0}} =& ~ j_{N p} \left( 1 - f_{N p} \psi_{0} \right) , \label{eq:pressureProfShape}
\end{align}
where $j_{N p}$ and $f_{N p} \in \left[ - \psi_{0 b}^{-1}, \psi_{0 b}^{-1} \right]$ are constants. By equation \refEq{eq:currentProfiles}, this implies that the toroidal magnetic field flux function term must be
\begin{align}
- I_{0} \frac{dI_{2}}{d \psi_{0}} =& ~ j_{N I} \left( 1 - f_{N I} \psi_{0} \right) ,
\end{align}
where
\begin{align}
j_{N I} \equiv& ~ j_{N} - j_{N p} \\
f_{N I} \equiv& ~ \frac{1}{j_{N I}} \left( j_{N} f_{N} - j_{N p} f_{N p} \right)
\end{align}
are constants.
\subsubsection{Solutions to the $O \left( B_{0} \right)$ Grad-Shafranov equation.}
\label{subsubsec:GradShafSolLowestOrder}
Like references \cite{GreeneEquilibrium1971, LaoGradShafExpansion1981, RodriguesGradShafAspectRatio2004, RodriguesGradShafNoniterative2009}, we will solve the $O \left( B_{0} \right)$ Grad-Shafranov equation by Fourier analyzing the magnetic flux in poloidal angle as
\begin{align}
\psi_{0} \left( r, \theta \right) =& ~ \psi_{0, 0}^{C} \left( r \right) + \sum_{m = 1}^{\infty} \left[ \psi_{0, m}^{C} \left( r \right) \Cos{m \theta} + \psi_{0, m}^{S} \left( r \right) \Sin{m \theta} \right] , \label{eq:gradShafLowestOrderSols}
\end{align}
where $m$ is an integer representing the poloidal flux surface shaping mode number. Using equation \refEq{eq:gradShafLowestOrderSols} we can rewrite equation \refEq{eq:gradShafLowestOrder} as
\begin{align}
\frac{1}{r} \frac{d}{dr} \left( r \frac{d \psi_{0, m}^{T}}{dr} \right) &+ \left( f_{N} j_{N} - \frac{m^{2}}{r^{2}} \right) \psi_{0, m}^{T} \left( r \right) = j_{N} \delta_{m, 0} , \label{eq:gradShafFourierLowest}
\end{align}
where $m \geq 0$, $\delta_{i, j}$ is the Kronecker delta, and $T = C, S$ is a superscript that indicates the sine or cosine mode. The solutions to this equation with zero poloidal flux at the magnetic axis are
\begin{align}
\psi_{0, 0}^{C} \left( r \right) =& - \frac{1}{f_{N}} \left( J_{0} \left( \sqrt{f_{N} j_{N}} r \right) - 1 \right) \label{eq:gradShafSol0modeFourierLowest} \\
\psi_{0, m}^{C} \left( r \right) =& ~ C_{0, m} \frac{m! ~ 2^{m}}{\left( f_{N} j_{N} \right)^{m/2}} J_{m} \left( \sqrt{ f_{N} j_{N}} r \right) \\
\psi_{0, m}^{S} \left( r \right) =& ~ S_{0, m} \frac{m! ~ 2^{m}}{\left( f_{N} j_{N} \right)^{m/2}} J_{m} \left( \sqrt{ f_{N} j_{N}} r \right) , \label{eq:gradShafSolMmodeFourierLowest}
\end{align}
where $m > 0$, $J_{m}$ is the $m^{\text{th}}$ order Bessel function of the first kind. The Fourier coefficients $C_{0, m}$ and $S_{0, m}$ are determined by the boundary conditions at the plasma edge, which is physically controlled by the locations and currents of external plasma shaping coils. Using trigonometric identities, equation \refEq{eq:gradShafLowestOrderSols} and equations \refEq{eq:gradShafSol0modeFourierLowest} through \refEq{eq:gradShafSolMmodeFourierLowest} can be rewritten as
\begin{align}
\psi_{0} \left( r, \theta \right) =& - \frac{1}{f_{N}} \left( J_{0} \left( \sqrt{f_{N} j_{N}} r \right) - 1 \right) \label{eq:gradShafLowestOrderSolsTilt} \\
&+ \sum _{m=1}^{\infty} N_{0, m} \frac{m! ~ 2^{m}}{\left( f_{N} j_{N} \right)^{m/2}} J_{m} \left( \sqrt{f_{N} j_{N}} r \right) \Cos{m \left( \theta + \theta_{t 0, m} \right)} , \nonumber
\end{align}
where $N_{0, m} \equiv \sqrt{ C_{0, m}^{2} + S_{0, m}^{2}}$ is the magnitude of the Fourier mode and $\theta_{t 0, m} \equiv - \ArcTan{S_{0, m} / C_{0, m}} / m$ is the Fourier mode tilt angle.
Note that for the constant current case (i.e. $f_{N} = 0$), equation \refEq{eq:gradShafLowestOrderSolsTilt} reduces to
\begin{align}
\psi_{0} \left( r, \theta \right) =& ~ \frac{j_{N}}{4} r^{2} + \sum _{m=1}^{\infty} N_{0, m} r^{m} \Cos{m \left( \theta + \theta_{t 0, m} \right)} . \label{eq:gradShafLowestOrderSolsTiltConst}
\end{align}
To understand the hollow current case, it is useful to note the identity
\begin{align}
J_{m} \left( i x \right) = i^{m} I_{m} \left( x \right) , \label{eq:modBesselFnFirstKind}
\end{align}
where $I_{m}$ is the $m^{\text{th}}$ order modified Bessel function of the first kind. From this we can demonstrate that equation \refEq{eq:gradShafLowestOrderSolsTilt} is equivalent to
\begin{align}
\psi_{0} \left( r, \theta \right) =& \frac{1}{- f_{N}} \left( I_{0} \left( \sqrt{- f_{N} j_{N}} r \right) - 1 \right) \label{eq:gradShafLowestOrderSolsTiltHollow} \\
&+ \sum _{m=1}^{\infty} N_{0, m} \frac{m! ~ 2^{m}}{\left( - f_{N} j_{N} \right)^{m/2}} I_{m} \left( \sqrt{- f_{N} j_{N}} r \right) \Cos{m \left( \theta + \theta_{t 0, m} \right)} , \nonumber
\end{align}
which can be more easily applied to hollow toroidal current profiles (i.e. $f_{N} < 0$).
\subsubsection{Solutions to the $O \left( \epsilon B_{0} \right)$ Grad-Shafranov equation.}
\label{subsubsec:GradShafSolNextOrder}
In order to solve the $O \left( \epsilon B_{0} \right)$ equation we must first Fourier analyze the magnetic flux in poloidal angle. The lowest order Fourier-analyzed flux is given by equation \refEq{eq:gradShafLowestOrderSols} and equations \refEq{eq:gradShafSol0modeFourierLowest} through \refEq{eq:gradShafSolMmodeFourierLowest}. To next order, we can write
\begin{align}
\psi_{1} \left( r, \theta \right) &= \psi_{1, 0}^{C} \left( r \right) + \sum_{m = 1}^{\infty} \left[ \psi_{1, m}^{C} \left( r \right) \Cos{m \theta} + \psi_{1, m}^{S} \left( r \right) \Sin{m \theta} \right] , \label{eq:gradShafNextOrderSols}
\end{align}
but we still must solve for $\psi_{1, m}^{C} \left( r \right)$ and $\psi_{1, m}^{S} \left( r \right)$ by substituting equations \refEq{eq:gradShafLowestOrderSols} and \refEq{eq:gradShafNextOrderSols} into equation \refEq{eq:gradShafNextOrder}. Since $\psi_{1, m}^{C} \left( r \right)$ and $\psi_{1, m}^{S} \left( r \right)$ do not depend on $\theta$, we can take each Fourier component of equation \refEq{eq:gradShafNextOrder} as a separate equation. This gives
\begin{align}
\frac{1}{r} \frac{d}{dr} \left( r \frac{d \psi_{1, m}^{T}}{dr} \right) &+ \left( f_{N} j_{N} - \frac{m^{2}}{r^{2}} \right) \psi_{1, m}^{T} \left( r \right) = \Lambda_{m}^{T} \left( r \right) \label{eq:gradShafFourierNext}
\end{align}
for each Fourier mode $m$, where the inhomogeneous terms are given by $\Lambda_{m}^{T} \left( r \right)$. For $m = 0$ and $T = C$
\begin{align}
\Lambda_{0}^{C} \left( r \right) &\equiv \frac{1}{R_{0}} \left[ \frac{1}{2} \frac{d \psi_{0, 1}^{C}}{dr} + \left( \frac{1}{2 r} - r f_{N p} j_{N p} \right) \psi_{0, 1}^{C} \left( r \right) \right] , \label{eq:inhomoTermsC0}
\end{align}
for $m = 1$ and $T =C$
\begin{align}
\Lambda_{1}^{C} \left( r \right) &\equiv \frac{1}{R_{0}} \left[ \frac{1}{2} \frac{d \psi_{0, 2}^{C}}{dr} + \left( \frac{1}{r} - r f_{N p} j_{N p} \right) \psi_{0, 2}^{C} \left( r \right) \right. \label{eq:inhomoTermsC1} \\
&+ \left. \frac{d \psi_{0, 0}^{C}}{dr} + 2 r j_{N p} \left( 1 - f_{N p} \psi_{0, 0}^{C} \left( r \right) \right) \right] , \nonumber
\end{align}
for $m = 1$ and $T = S$
\begin{align}
\Lambda_{1}^{S} \left( r \right) &\equiv \frac{1}{R_{0}} \left[ \frac{1}{2} \frac{d \psi_{0, 2}^{S}}{dr} + \left( \frac{1}{r} - r f_{N p} j_{N p} \right) \psi_{0, 2}^{S} \left( r \right) \right] , \label{eq:inhomoTermsS1}
\end{align}
and for all other $m$ and $T = C, S$
\begin{align}
\Lambda_{m}^{T} \left( r \right) &\equiv \frac{1}{R_{0}} \left[ \frac{1}{2} \frac{d \psi_{0, m + 1}^{T}}{dr} + \left( \frac{m+1}{2 r} - r f_{N p} j_{N p} \right) \psi_{0, m + 1}^{T} \left( r \right) \right. \label{eq:inhomoTermsTm} \\
&+ \left. \frac{1}{2} \frac{d \psi_{0, m - 1}^{T}}{dr} - \left( \frac{m-1}{2 r} + r f_{N p} j_{N p} \right) \psi_{0, m - 1}^{T} \left( r \right) \right] . \nonumber
\end{align}
Equation \refEq{eq:gradShafFourierNext} can be solved using the method of variation of parameters, yielding
\begin{align}
\psi_{1, m}^{T} \left( r \right) &= - \frac{\pi}{2} J_{m} \left( \sqrt{f_{N} j_{N}} r \right) \int_{0}^{r} d r' ~ r' Y_{m} \left( \sqrt{f_{N} j_{N}} r' \right) \Lambda_{m}^{T} \left( r' \right) \nonumber \\
&+ \frac{\pi}{2} Y_{m} \left( \sqrt{f_{N} j_{N}} r \right) \int_{0}^{r} d r' ~ r' J_{m} \left( \sqrt{f_{N} j_{N}} r' \right) \Lambda_{m}^{T} \left( r' \right) \label{eq:psiFourierNextOrderCoeffs} \\
&+ T_{1, m} \frac{m! ~ 2^{m}}{\left( f_{N} j_{N} \right)^{m/2}} J_{m} \left( \sqrt{f_{N} j_{N}} r \right) , \nonumber
\end{align}
where we have imposed regularity at the origin, $Y_{m}$ is the $m^{\text{th}}$ order Bessel function of the second kind, and $T_{1, m} = C_{1, m}, S_{1, m}$ are Fourier coefficients determined by the boundary conditions at the plasma edge. Combining equations \refEq{eq:gradShafNextOrderSols}, \refEq{eq:inhomoTermsC0} through \refEq{eq:inhomoTermsTm}, and \refEq{eq:psiFourierNextOrderCoeffs} gives the complete solution to the $O \left( \epsilon B_{0} \right)$ Grad-Shafranov equation for an arbitrary boundary condition.
To understand the hollow current case (i.e. $f_{N} < 0$), we will use equation \refEq{eq:modBesselFnFirstKind} and the identity
\begin{align}
Y_{m} \left( i x \right) = i^{m+1} I_{m} \left( x \right) - \frac{2}{\pi} i^{-m} K_{m} \left( x \right) ,
\end{align}
where $K_{m}$ is the $m^{\text{th}}$ order modified Bessel function of the second kind. This enables equation \refEq{eq:psiFourierNextOrderCoeffs} to be reformulated as
\begin{align}
\psi_{1, m}^{T} \left( r \right) =& ~ I_{m} \left( \sqrt{- f_{N} j_{N}} r \right) \int_{0}^{r} d r' ~ r' K_{m} \left( \sqrt{ - f_{N} j_{N}} r' \right) \Lambda_{m}^{T} \left( r' \right) \nonumber \\
&- K_{m} \left( \sqrt{- f_{N} j_{N}} r \right) \int_{0}^{r} d r' ~ r' I_{m} \left( \sqrt{- f_{N} j_{N}} r' \right) \Lambda_{m}^{T} \left( r' \right) \label{eq:psiFourierNextOrderCoeffsHollow} \\
&+ T_{1, m} \frac{m! ~ 2^{m}}{\left( - f_{N} j_{N} \right)^{m/2}} I_{m} \left( \sqrt{- f_{N} j_{N}} r \right) . \nonumber
\end{align}
For a constant current profile (i.e. $f_{N} = 0$), we can take the limit of equations \refEq{eq:gradShafNextOrderSols}, \refEq{eq:inhomoTermsC0} through \refEq{eq:inhomoTermsTm}, and \refEq{eq:psiFourierNextOrderCoeffs} as $f_{N} j_{N} \to 0$ to find
\begin{align}
\psi_{1} \left( r, \theta \right) =& \frac{1}{4 R_{0}} \left[ \left( \frac{j_{N} + 4 j_{N p}}{4} r^{3} - \frac{j_{N} f_{N p} j_{N p}}{12} r^{5} \right) \Cos{\theta} \right. \nonumber \\
&+ \sum_{m = 2}^{\infty} \left( r^{m+1} - \frac{f_{N p} j_{N p}}{2 \left( m + 1 \right)} r^{m+3} \right) N_{0, m} \Cos{ \left( m - 1 \right) \theta + m \theta_{t 0, m}} \nonumber \\
&- \left. \sum_{m = 2}^{\infty} \frac{f_{N p} j_{N p}}{m + 2} r^{m+3} N_{0, m} \Cos{\left( m + 1 \right) \theta + m \theta_{t 0, m}} \right] \label{eq:psiNextOrderSolConst} \\
&+ \sum_{m = 0}^{\infty} r^{m} N_{1, m} \Cos{m \left( \theta + \theta_{t 1, m} \right)} , \nonumber
\end{align}
where $N_{1, m} \equiv \sqrt{ C_{1, m}^{2} + S_{1, m}^{2}}$ is the magnitude of the next order Fourier mode, $\theta_{t 1, m} \equiv - \ArcTan{S_{1, m}/C_{1, m}} / m$ is the next order Fourier mode tilt angle, and we have used equation \refEq{eq:gradShafLowestOrderSolsTiltConst} along with
\begin{align}
\lim_{f_{N} j_{N} \to 0} \frac{m! ~ 2^{m}}{\left( f_{N} j_{N} \right)^{m/2}} J_{m} \left( \sqrt{f_{N} j_{N}} r \right) =& ~ r^{m} \label{eq:BesselFnExpansionOn} \\
\lim_{f_{N} j_{N} \to 0} Y_{m} \left( \sqrt{f_{N} j_{N}} r \right) =& - \frac{1}{m \pi} \frac{m! ~ 2^{m}}{\left( f_{N} j_{N} \right)^{m/2}} r^{-m} \label{eq:modifiedBesselFnExpansionOn}
\end{align}
for $m \neq 0$. The first line of equation \refEq{eq:psiNextOrderSolConst} contains the direct effect of toroidicity on the equilibrium, i.e. the Shafranov shift. The second and third lines show that a zeroth order shaping mode $m$ splits into two modes, $m-1$ and $m+1$, at first order. The last line contains the homogeneous solution, which enables an arbitrary boundary condition to be satisfied.
\subsection{Solution for a tilted elliptical boundary condition}
\label{subsec:numCoeffCalc}
In order to model realistic tilted elliptical tokamaks in our gyrokinetic simulations we must know how the Shafranov shift depends on the tilt angle of the elliptical boundary flux surface (parameterized by $\theta_{\kappa b}$ as shown in figure \ref{fig:thetaKappaDef}). We will argue that the Shafranov shift is insensitive to the shape of the current and pressure profiles (using linear profiles parameterized by equations \refEq{eq:currentProfiles} and \refEq{eq:pressureProfShape} respectively) when the geometry, plasma current, and average $d p / d \psi$ is kept fixed. Doing so makes the gyrokinetic simulations presented in section \ref{sec:gyrokineticSims} more widely applicable, as they use equilibria derived assuming constant current and pressure gradient profiles.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{figs/thetaKappaDefMod}
\caption{An illustration of the boundary flux surface (black, solid) with the untilted boundary surface (black, dashed) and the axis of axisymmetry (black, dash-dotted) shown for reference. Here $a$ is the tokamak minor radius (i.e. the minimum radial position on the flux surface of interest), $b$ is the maximum radial position, $\kappa_{b} \equiv b / a$ is the boundary elongation, and $\theta_{\kappa b}$ is the boundary tilt angle.}
\label{fig:thetaKappaDef}
\end{figure}
Together equations \refEq{eq:gradShafLowestOrderSolsTilt}, \refEq{eq:gradShafNextOrderSols}, \refEq{eq:inhomoTermsC0} through \refEq{eq:inhomoTermsTm}, and \refEq{eq:psiFourierNextOrderCoeffs} give the general solution of the Grad-Shafranov equation to $O \left( \epsilon B_{0} \right)$, which is sufficient to capture the behavior of the Shafranov shift. However, we still must determine the Fourier coefficients $N_{0, m}$, $\theta_{t 0, m}$, $C_{1, m}$, and $S_{1, m}$ in order to create a tilted elliptical boundary flux surface. To do so we require the poloidal flux to be constant on the boundary, parameterized in polar form by
\begin{align}
r_{b} \left( \theta \right) = \frac{\sqrt{2} \kappa_{b} a}{\sqrt{\kappa_{b}^{2} + 1 + \left( \kappa_{b}^{2} - 1 \right) \Cos{2 \left( \theta + \theta_{\kappa b} \right)}}} , \label{eq:boundarySurf}
\end{align}
where figure \ref{fig:thetaKappaDef} shows the definitions of the various geometric parameters. Note that the tilt angle of the boundary $\theta_{\kappa b}$ is defined to increase in the clockwise direction, in contrast to the poloidal angle $\theta$.
To calculate $N_{0, m}$ and $\theta_{t 0, m}$ we substitute equation \refEq{eq:boundarySurf} into equation \refEq{eq:gradShafLowestOrderSolsTilt} to give
\begin{align}
\psi_{0} \left( r_{b} \left( \theta \right), \theta \right) = \psi_{0 b} . \label{eq:gradShafBoundaryCondLowestOrder}
\end{align}
Since $\psi_{0 b}$ is a constant we know that $\psi_{0} \left( r_{b} \left( \theta \right), \theta \right)$ does not depend on $\theta$. In theory, ensuring that this is true for all values of $\theta$ determines all of the lowest order Fourier coefficients. However, the exact solution for these coefficients is not analytic, so we will resort to a numerical solution. Before we do so we will note that, because the lowest order Grad-Shafranov equation has cylindrical symmetry, the only angle intrinsic to the problem is $\theta_{\kappa b}$, which is introduced by the boundary condition. This implies that
\begin{align}
\theta_{t 0, m} = \theta_{\kappa b} \label{eq:tiltAngleSol}
\end{align}
for all $m$, which suggests that it will be useful to define a new poloidal angle
\begin{align}
\theta_{s} \equiv \theta + \theta_{\kappa b} . \label{eq:shiftThetaDef}
\end{align}
Furthermore, since an ellipse has mirror symmetry about exactly two axes, we know that $N_{0, m} = 0$ for odd $m$.
To determine $N_{0, m}$ for even $m$ we will take the Fourier series of $\psi_{0} \left( r_{b} \left( \theta_{s} \right), \theta_{s} \right) - \psi_{0 b}$. Truncating the series at a large mode number $m_{\text{max}}$ gives a long series of cosine terms. Requiring that the coefficient of each term must individually vanish gives a numerical approximation for all $N_{0, m}$ with $m \leq m_{\text{max}}$. In the limit that $m_{\text{max}} \to \infty$ this approximation approaches the exact solution, though in practice $m_{\text{max}} \approx 10$ was found to achieve sufficient precision for our purposes. This was determined by ensuring that the magnetic axis did not move significantly when $m_{\text{max}}$ was changed by $40\%$.
To next order we must determine $C_{1, m}$ and $S_{1, m}$ such that
\begin{align}
\psi_{1} \left( r_{b} \left( \theta \right), \theta \right) = \psi_{1 b} \label{eq:gradShafBoundaryCondNextOrder}
\end{align}
is true, where $\psi_{1 b}$ is the next order value of the poloidal flux on the boundary flux surface. This is done in a similar manner to the lowest order equations, except the Grad-Shafranov equation no longer has cylindrical symmetry and we must evaluate the integrals in equation \refEq{eq:psiFourierNextOrderCoeffs}. The lack of symmetry means that we do not automatically know the tilt angle of the modes. However, since $\psi_{0}$ only has even Fourier mode numbers, it can be shown that equation \refEq{eq:gradShafNextOrder} only has odd Fourier modes. Hence, $C_{1, m} = S_{1, m} = 0$ for even $m$.
To calculate $C_{1, m}$ and $S_{1, m}$ for odd $m$ we construct $\psi_{1} \left( r, \theta \right)$ from equations \refEq{eq:gradShafNextOrderSols} and \refEq{eq:inhomoTermsC0} through \refEq{eq:psiFourierNextOrderCoeffs}. Taylor expanding this in $f_{N} j_{N} a^{2} \ll 1$ to $O \left( \left( f_{N} j_{N} a^{2} \right)^{f_{\text{max}}} \right)$ allows us to analytically calculate the integrals appearing in equation \refEq{eq:psiFourierNextOrderCoeffs} because the Bessel functions become summations of polynomials. We can now substitute equation \refEq{eq:boundarySurf} and find the Fourier series of $\psi_{1} \left( r_{b} \left( \theta \right), \theta \right) - \psi_{1 b}$ to mode number $m_{\text{max}}$. Again, we require that all of the Fourier coefficients must individually vanish, which produces a numerical approximation for each $C_{1, m}$ and $S_{1, m}$ with $m \leq m_{\text{max}}$. A value of $f_{\text{max}} \approx 10$ was found to give a sufficiently accurate solution. This was determined by ensuring that the magnetic axis did not move significantly when $f_{\text{max}}$ was changed by $40\%$.
For a hollow current profile, we repeat the entire above process except for using equation \refEq{eq:gradShafLowestOrderSolsTiltHollow} instead of equation \refEq{eq:gradShafLowestOrderSolsTilt} and equation \refEq{eq:psiFourierNextOrderCoeffsHollow} instead of equation \refEq{eq:psiFourierNextOrderCoeffs}. While the above process also works for the case of a constant toroidal current profile, this case actually has an analytic solution, which we derive in \ref{app:exactMagAxisLoc}.
In order to understand the effect of changing the current and pressure profiles in a single experimental device, we will choose to keep the major radial location of the center of the boundary flux surface ($R_{0}$), the minor radius ($a$), the edge elongation ($\kappa_{b}$), the total plasma current ($I_{p}$), and an estimate of the average pressure gradient ($p_{\text{axis}}/\psi_{0 b}$, i.e. the on-axis pressure divided by the edge poloidal flux) fixed. In order to keep these parameters fixed as we change the current and pressure profiles we must calculate how they enter into both $j_{N}$ and $j_{N p}$. Calculating $j_{N p}$ is straightforward, as we can directly integrate equation \refEq{eq:pressureProfShape} over poloidal flux to find
\begin{align}
j_{N p} = \mu_{0} R_{0}^{2} \frac{p_{\text{axis}}}{\psi_{0 b}} \left( 1 - \frac{f_{N p} \psi_{0 b}}{2} \right)^{-1} . \label{eq:jNpEstimate}
\end{align}
To calculate $j_{N}$ we start with the definition of the plasma current,
\begin{align}
I_{p} \equiv \int d S j_{\zeta} = \int_{0}^{2 \pi} d \theta_{s} \int_{0}^{r_{b} \left( \theta_{s} \right)} d r j_{\zeta} r , \label{eq:totalPlasmaCurrent}
\end{align}
where $S$ is the poloidal cross-sectional surface. Since we are only searching for a simple estimate, we will use equation \refEq{eq:currentProfiles} to rewrite equation \refEq{eq:totalPlasmaCurrent} as
\begin{align}
I_{p} = \int_{0}^{2 \pi} d \theta_{s} \int_{0}^{r_{b} \left( \theta_{s} \right)} d r \frac{j_{N}}{\mu_{0} R_{0}} \left( 1 - f_{N} \psi_{0} \right) r ,
\end{align}
which is accurate to lowest order in aspect ratio. Substituting the boundary shape (i.e. equation \refEq{eq:boundarySurf}) and the constant current solution for $\psi_{0} \left( r, \theta_{s} \right)$ (i.e. equations \refEq{eq:gradShafLowestOrderSolsTiltConst}, \refEq{eq:tiltAngleSol}, \refEq{eq:boundaryFluxConst}, and \refEq{eq:fourierShapingConst}) allows us to directly take the integral to find
\begin{align}
j_{N} = \mu_{0} \frac{I_{p}}{\pi a^{2} \kappa_{b}} R_{0} \left( 1 - \frac{f_{N} \psi_{0 b}}{2} \right)^{-1} + \Order{f_{N}^{2} j_{N}^{2} a^{4}} . \label{eq:jNestimate}
\end{align}
The $\Order{f_{N}^{2} j_{N}^{2} a^{4}}$ error arises from the fact that we used the constant current solution for $\psi_{0} \left( r, \theta_{s} \right)$, which is only accurate to lowest order in $f_{N} j_{N} a^{2} \ll 1$. This means that as we change $f_{N p}$ and $f_{N}$ we must change $j_{N p}$ and $j_{N}$ according to equations \refEq{eq:jNpEstimate} and \refEq{eq:jNestimate} respectively.
In figure \ref{fig:gradShafCalcSol} we plot the calculated flux surfaces resulting from three different current profiles, setting $f_{N p} = f_{N}$. We use inputs of $R_{0} = 3$, $a = 1$ (where we have normalized all lengths to the minor radius), $\kappa_{b} = 2$, and
\begin{align}
\frac{j_{N p}}{j_{N}} \approx \frac{\pi a^{2} \kappa_{b} R_{0}}{I_{p}} \frac{p_{\text{axis}}}{\psi_{0 b}} \approx 0.7 \label{eq:jNconstantsRatio}
\end{align}
using projections for ITER \cite{AymarITERSummary2001}. Additionally, we choose to plot the case of $\theta_{\kappa b} = \pi / 8$ because nonlinear gyrokinetic simulations have shown this value to be optimal for generating rotation \cite{BallMomUpDownAsym2014}. Note that the $\psi_{0 b}$ appearing in equation \refEq{eq:jNconstantsRatio} is part of $p_{\text{axis}} / \psi_{0 b}$, so it is fixed for all three profiles and can be calculated for a constant current profile from equation \refEq{eq:boundaryFluxConst}. In figure \ref{fig:gradShafCalcSol} we see that the current profile has an effect on the penetration of elongation from the boundary to the magnetic axis. This indicates that hollower current profiles better support elongation throughout the plasma, which is consistent with previous theoretical work \cite{BallMomUpDownAsym2014, BallMastersThesis2013, RodriguesMHDupDownAsym2014, BizarroUpDownAsymGradShafEq2014, BallShapingPenetration2015} as well as EFIT equilibrium reconstruction on simulated experimental data (see figure 5(b) of reference \cite{LaoShapeAndCurrent1985}). However, given these parameters, the Shafranov shift is not visibly altered, even with the extreme changes in the current profile.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{figs/calcFluxSurfSol.pdf}
\caption{Calculated flux surfaces for $f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0$ (black, solid), $f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0.4$ (red, dotted), and $f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = - 0.4$ (blue, dashed).}
\label{fig:gradShafCalcSol}
\end{figure}
In order to verify our calculation, we compared our results with the ECOM code \cite{LeeECOM2015}, a fixed boundary equilibrium solver capable of modeling up-down asymmetric configurations. In figure \ref{fig:gradShafSolComp} we see a direct graphical comparison between ECOM and the results of our calculation that were shown in figure \ref{fig:gradShafCalcSol}. The two sets of results agree well, especially for the constant and hollow current profile cases. The most significant source of error is expected to be finite aspect ratio effects in our analytic calculation, which arise from the assumption that $\epsilon = 1 / 3 \ll 1$. Hence, since we carried out the analytic calculation to lowest and next order in the aspect ratio expansion, we expect to have an $\epsilon^{2} \sim 10\%$ error. We also note that we do not expect the $\Order{\epsilon^{2} B_{0}}$ solution (i.e. the largest order that we omitted) to modify the Shafranov shift in a configuration with an elliptical boundary. This is because reference \cite{HakkarainenEquilibrium1990} demonstrates that toroidicity only introduces $m = 0$ and $m = 2$ modes at order $\Order{\epsilon^{2} B_{0}}$. Furthermore, equation \refEq{eq:psiNextOrderSolConst} demonstrates that a lowest order shaping effect $m$ introduces only $m - 1$ and $m + 1$ modes to order $\Order{\epsilon B_{0}}$. This suggests that only $m - 2$, $m$, and $m + 2$ modes will appear to $\Order{\epsilon^{2} B_{0}}$. Therefore, we expect that the $m = 1$ mode will not appear at $\Order{\epsilon^{2} B_{0}}$, so the Shafranov shift will not be changed.
\begin{figure}
\begin{center}
(a) \hspace{0.36\textwidth}
\includegraphics[height=0.45\textwidth]{figs/ECOMcompFluxSurfConst.pdf}
(b) \hspace{0.37\textwidth} (c) \hspace{0.36\textwidth}
\includegraphics[height=0.45\textwidth]{figs/ECOMcompFluxSurfPeaked.pdf}
\includegraphics[height=0.45\textwidth]{figs/ECOMcompFluxSurfHollow.pdf}
\end{center}
\caption{Flux surfaces calculated by both ECOM (dotted) and analytically (solid) for (a) $f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0$ (black), (b) $f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0.4$ (red), and (c) $f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = - 0.4$ (blue).}
\label{fig:gradShafSolComp}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figs/geoMagAxisLocMod.pdf}
\caption{Example flux surfaces showing the geometric meaning of the parameters $r_{\text{axis}}$ and $\theta_{\text{axis}}$, the minor radial and poloidal locations of the magnetic axis respectively.}
\label{fig:geoMagAxisLoc}
\end{figure}
\subsection{Location of the magnetic axis}
\label{subsec:ECOMcomparison}
We can obtain the Shafranov shift from our calculation by numerically solving the equation
\begin{align}
\left. \ensuremath{\vec{\nabla}} \left( \psi_{0} \left( r, \theta \right) + \psi_{1} \left( r, \theta \right) \right) \right|_{r = r_{\text{axis}}, \theta = \theta_{\text{axis}}} = 0 \label{eq:magAxisCondition}
\end{align}
using equations \refEq{eq:gradShafLowestOrderSolsTilt}, \refEq{eq:gradShafNextOrderSols}, \refEq{eq:inhomoTermsC0} through \refEq{eq:inhomoTermsTm}, \refEq{eq:psiFourierNextOrderCoeffs}, and \refEq{eq:tiltAngleSol} as well as our numerical solutions for $N_{0, m}$, $C_{1, m}$, and $S_{1, m}$. Here $r_{\text{axis}}$ and $\theta_{\text{axis}}$ are the minor radial and poloidal location of the magnetic axis respectively, as indicated in figure \ref{fig:geoMagAxisLoc}. For the special case of a tilted elliptical boundary with a constant toroidal current profile (i.e. $f_{N} = 0$) we can exactly solve equation \refEq{eq:magAxisCondition} as shown in \ref{app:exactMagAxisLoc}. Equations \refEq{eq:magAxisPoloidalLocExact} and \refEq{eq:magAxisRadialLocExactCondition} give the exact location of the magnetic axis when considering the poloidal flux to lowest order and next order in $\epsilon \ll 1$.
\begin{figure}
\hspace{0.04\textwidth} (a) \hspace{0.4\textwidth} (b) \hspace{0.25\textwidth}
\begin{center}
\includegraphics[width=0.45\textwidth]{figs/rShafShiftLoc.pdf}
\includegraphics[width=0.45\textwidth]{figs/thetaShafShiftLoc.pdf}
\end{center}
\caption{The (a) minor radial and (b) poloidal location of the magnetic axis for constant ($f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0$) (black, solid, circles), linear peaked ($f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0.4$) (red, dotted, squares), and linear hollow ($f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = - 0.4$) (blue, dashed, triangles) current/pressure gradient profiles, calculated analytically (lines) and by ECOM (points).}
\label{fig:ShafShiftLoc}
\end{figure}
In figure \ref{fig:ShafShiftLoc} we show the location of the magnetic axis as we vary the shape of the current/pressure profile (by changing $f_{N}$ and keeping $f_{N p} = f_{N}$), while holding the geometry, $I_{p}$, and $p_{\text{axis}} / \psi_{0 b}$ fixed. For the most part, we see reasonable quantitative agreement between our theoretical results and ECOM. However, the two calculations disagree on the trend of $r_{\text{axis}}$ with $f_{N} \psi_{0 b}$ at large tilt angles. We do not think this is significant as it appears to be a breakdown in our inverse aspect ratio expansion. The two calculations become consistent if the aspect ratio is directly increased or if smaller tilt angles are used (where the effective aspect ratio is larger).
An important property of figure \ref{fig:ShafShiftLoc}, which is supported by both the analytic and ECOM calculations, is the insensitivity of the Shafranov shift to extreme changes in the shape of the current profile. Both the magnitude and the direction of the Shafranov shift change very little between the different current profiles. This is especially true in the domain of $\theta_{\kappa b} \in \left[ 0, \pi / 4 \right]$, which is the range of tilt angles that seem most promising for implementing in an experiment \cite{CamenenPRLExp2010, BallMomUpDownAsym2014}. This result allows us to simplify our treatment of the Shafranov shift. The gyrokinetic simulations we will present in section \ref{sec:gyrokineticSims} are formally inconsistent because they do not assume constant current and pressure gradient profiles, but they use the Shafranov shift of equilibria with constant current and pressure gradient profiles. However, this inconsistency is not important because the Shafranov shift only depends weakly on the shape of the current and pressure gradient profiles. As we will see in figure \ref{fig:momHeatFluxRatioWithShift}, the turbulent momentum flux driven by the Shafranov shift is approximately linear in the size of the shift, so small errors in the shift will only lead to small errors in the momentum flux.
Also from figure \ref{fig:ShafShiftLoc}, we learn that the tilt angle has a large effect, not just on the direction of the Shafranov shift, but also its magnitude. This is intuitive because we know that, for an ellipse with $\kappa = 2$, the midplane chord length is twice as long in the $\theta_{\kappa b} = \pi / 2$ geometry as it is in the $\theta_{\kappa b} = 0$ geometry. Lastly, we see that the direction of the Shafranov shift varies considerably, but it is purely outwards for the $0$ and $\pi / 2$ tilt angles as expected. Importantly, it does not align with the tilt angle of the ellipse, so it breaks the mirror symmetry of the configuration.
\begin{figure}
\hspace{0.04\textwidth} (a) \hspace{0.4\textwidth} (b) \hspace{0.25\textwidth}
\begin{center}
\includegraphics[width=0.45\textwidth]{figs/rShafShiftLocRadialVariation.pdf}
\includegraphics[width=0.45\textwidth]{figs/thetaShafShiftLocRadialVariation.pdf}
\end{center}
\caption{The (a) minor radial and (b) poloidal location of the magnetic axis for constant ($f_{N p} \psi_{0 b} = 0$) (black, solid, circles), linear peaked ($f_{N p} \psi_{0 b} = 0.4$) (red, dotted, squares), and linear hollow ($f_{N p} \psi_{0 b} = - 0.4$) (blue, dashed, triangles) pressure gradient profiles, calculated analytically (lines) and by ECOM (points) for a constant current profile.}
\label{fig:ShafShiftLocPressureProf}
\end{figure}
In figure \ref{fig:ShafShiftLocPressureProf} we show the location of the magnetic axis as we vary the shape of the pressure profile (by changing $f_{N p}$) with a constant current profile (i.e. $f_{N} = 0$), while holding the geometry, $I_{p}$, and $p_{\text{axis}} / \psi_{0 b}$ fixed. We see good quantitative agreement between the calculation given in \ref{app:exactMagAxisLoc} and ECOM. Figure \ref{fig:ShafShiftLocPressureProf} indicates that varying the shape of the pressure profile has little effect on the Shafranov shift. We note that, in general, varying the pressure profile has a large effect on the magnitude of the Shafranov shift, but not when $I_{p}$ and $p_{\text{axis}} / \psi_{0 b}$ are held constant. This is important as it justifies using our MHD results for the Shafranov shift with a constant $d p / d \psi$ profile as input for gyrokinetic simulations that are based on ITER, which has a constant $d p / d r_{\psi}$ profile \cite{AymarITERSummary2001}. Even though this is formally inconsistent, our analysis suggests the Shafranov shift in a configuration with constant $d p / d \psi$ will be a reasonable estimate of the Shafranov shift in a configuration with constant $d p / d r_{\psi}$ (as long as the geometry, $I_{p}$, and $p_{\text{axis}} / \psi_{0 b}$ are the same). As we will show in figure \ref{fig:momHeatFluxRatioWithShift}, the momentum flux is approximately linear in the size of the Shafranov shift. Hence, the small error introduced by using the Shafranov shift calculated with a constant pressure gradient profile (in $\psi$) will not lead to large differences in the momentum flux.
\section{Nonlinear gyrokinetic simulations}
\label{sec:gyrokineticSims}
In this section we will use the results from section \ref{sec:MHDequil} in order to perform nonlinear gyrokinetic simulations that include the effect of a realistic Shafranov shift on plasma turbulence. Since the size of the Shafranov shift is closely connected to the plasma pressure, we also included the effect of $\beta'$ on the magnetic equilibrium. We will use a modified version of GS2 \cite{KotschenreutherGS21995} to self-consistently calculate the time-averaged radial flux of toroidal angular momentum $\left\langle \Pi_{\zeta i} \right\rangle_{t}$ and the time-averaged radial flux of energy $\left\langle Q_{i} \right\rangle_{t}$ for ions. These calculations use a local equilibrium specified by an up-down asymmetric generalization of the Miller geometry model \cite{MillerGeometry1998}.
\subsection{Input parameters}
\label{subsec:inputParameters}
In this work, we will use a flux surface of interest with Cyclone base case parameters (unless otherwise specified) \cite{DimitsCycloneBaseCase2000}: a minor radius of $\rho_{0} = 0.54$, a major radius of $R_{c 0} / a = 3$ (i.e. the major radial location of the center of the flux surface of interest), a safety factor of $q = 1.4$, a magnetic shear of $\hat{s} \equiv \left( \rho_{0} / q \right) d q / d \rho = 0.8$, a temperature gradient of $d \Ln{T_{s}} / d \rho = - 2.3$, and a density gradient of $d \Ln{n_{s}} / d \rho = - 0.733$ (where the subscript $s$ indicates either the ion or electron species). Here $\rho \equiv r_{\psi} / a$ is the normalized minor radial flux surface label, $r_{\psi}$ is a real-space flux surface label that indicates the minimum distance of each flux surface from its center, $\rho_{0} \equiv r_{\psi 0} / a$ is the value of $\rho$ on the flux surface of interest, and $r_{\psi 0}$ is the value of $r_{\psi}$ on the flux surface of interest. We note that taking $\hat{s} \neq 0$ can be formally inconsistent with a constant toroidal current profile (as it is in the large aspect ratio limit for circular flux surfaces). However, from figure \ref{fig:ShafShiftLoc} we know that we can vary the current profile without affecting the Shafranov shift much, as long as we keep $I_{p}$ and $p_{\text{axis}} / \psi_{0 b}$ fixed. Because of this freedom, we can use the Shafranov shift calculated assuming constant current and pressure gradient profiles for the Cyclone base case. Many of our simulations will model elliptical flux surfaces, all of which have an elongation of $\kappa = 2$. Furthermore, all turbulent fluxes calculated by GS2 will be normalized to gyroBohm values of
\begin{align}
\Pi_{gB} &\equiv \rho_{\ast}^{2} n_{i} a m_{i} v_{th, i}^{2} \\
Q_{gB} &\equiv \rho_{\ast}^{2} n_{i} T_{i} v_{th, i} ,
\end{align}
where $\rho_{\ast} \equiv \rho_{i} / a$ is the ratio of the ion gyroradius to the tokamak minor radius, $n_{i}$ is the ion density, $m_{i}$ is the ion mass, $T_{i}$ is the local ion temperature, and $v_{th, i} \equiv \sqrt{2 T_{i} / m_{i}}$ is the local ion thermal speed. All simulations used at least 48 grid points in the poloidal angle, 127 grid points in the wavenumber of the radial direction, 22 grid points in the wavenumber of the direction within the flux surface (but still perpendicular to the magnetic field), 12 grid points in the energy, and 10 grid points in the untrapped pitch angle. The large number of poloidal grid points was needed to properly resolve the strong flux surface shaping.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figs/geoFluxSurfCenterMod.pdf}
\caption{Three example flux surfaces (black, solid) at different values of $r_{\psi}$ with their geometric center (red, crosses). This illustrates the meaning of the parameters $R_{c} \left( r_{\psi} \right)$ (red, dashed, vertical) and $Z_{c} \left( r_{\psi} \right)$ (red, dashed, horizontal), the major radial and axial locations of the center of each flux surface respectively.}
\label{fig:geoFluxSurfCenter}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{figs/linearShift.pdf}
\caption{The shift in the center of flux surfaces (relative to the center of the boundary flux surface $R_{0}$) as a function of normalized poloidal flux for geometries with $\theta_{\kappa b} = 0$. The points are calculated by ECOM for a constant current profile ($f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0$) (black, circles), a linear peaked current profile ($f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = 0.4$) (red, pluses), a linear hollow current profile ($f_{N} \psi_{0 b} = f_{N p} \psi_{0 b} = - 0.4$) (blue, pluses), a linear peaked pressure profile ($f_{N} \psi_{0 b} = 0$ and $f_{N p} \psi_{0 b} = 0.4$) (red, crosses), and a linear hollow pressure profile ($f_{N} \psi_{0 b} = 0$ and $f_{N p} \psi_{0 b} = - 0.4$) (blue, crosses). Also shown is our analytic solution (solid line) and a linear best fit (dashed line).}
\label{fig:shiftWithRadius}
\end{figure}
The Miller geometry specification in GS2 captures the Shafranov shift through local values of $d R_{c} / d r_{\psi}$ and $d Z_{c} / d r_{\psi}$. Here $R_{c} \left( r_{\psi} \right)$ and $Z_{c} \left( r_{\psi} \right)$ indicate the location of the center of each flux surface as shown in figure \ref{fig:geoFluxSurfCenter}. In order to model a realistic geometry, we will calculate local values of $d R_{c} / d r_{\psi}$ and $d Z_{c} / d r_{\psi}$ for arbitrary tilt angle from our global MHD results. Specifically, we will use the dependence of the global Shafranov shift on tilt angle calculated for constant current and $d p / d \psi$ profiles (i.e. the solid black line shown in figure \ref{fig:ShafShiftLoc}).
First we will assume that $d R_{c} / d \psi$ and $d Z_{c} / d \psi$ are constant from the boundary flux surface to the magnetic axis. In figure \ref{fig:shiftWithRadius}, we plot our analytic solution (using the coefficients calculated in \ref{app:exactMagAxisLoc}) and ECOM results to show that this assumption holds, regardless of the shape of the pressure and current profiles. Additionally, using equations \refEq{eq:gradShafLowestOrderSolsTiltConst} and \refEq{eq:fourierShapingConst} we see that
\begin{align}
\psi \propto r_{\psi}^{2}
\end{align}
for a constant current profile and an exactly elliptical boundary. Therefore, using that $\psi = \psi_{b}$ at $r_{\psi} = a$, one can calculate the constant of proportionality and show
\begin{align}
\frac{d \psi}{d r_{\psi}} = 2 \frac{\psi_{b}}{a} \rho .
\end{align}
Hence, the local Shafranov shift can be written as
\begin{align}
\left. \frac{d R_{c}}{d r_{\psi}} \right|_{r_{\psi 0}} =& \left. \frac{d \psi}{d r_{\psi}} \right|_{r_{\psi 0}} \frac{d R_{c}}{d \psi} = \left( 2 \frac{\psi_{b}}{a} \rho_{0} \right) \frac{R_{0} - R_{c} \left( 0 \right)}{\psi_{b} - 0} = - 2 \rho_{0} \frac{r_{\text{axis}}}{a} \Cos{\theta_{\text{axis}}} \label{eq:radialShafShiftLocal} \\
\left. \frac{d Z_{c}}{d r_{\psi}} \right|_{r_{\psi 0}} =& \left. \frac{d \psi}{d r_{\psi}} \right|_{r_{\psi 0}} \frac{d Z_{c}}{d \psi} = \left( 2 \frac{\psi_{b}}{a} \rho_{0} \right) \frac{0 - Z_{c} \left( 0 \right)}{\psi_{b} - 0} = - 2 \rho_{0} \frac{r_{\text{axis}}}{a} \Sin{\theta_{\text{axis}}} , \label{eq:axialShafShiftLocal}
\end{align}
where the coordinate system is defined such that the boundary flux surface is centered at $\left( R = R_{0}, Z = 0 \right)$. Therefore, we are able to calculate $\left. d R_{c} / d r_{\psi} \right|_{r_{\psi 0}}$ and $\left. d Z_{c} / d r_{\psi} \right|_{r_{\psi 0}}$ for an ITER-like pressure profile using equations \refEq{eq:radialShafShiftLocal} and \refEq{eq:axialShafShiftLocal} as well as the constant current results shown in figure \ref{fig:ShafShiftLoc}.
GS2 also requires a local value of
\begin{align}
\beta' \equiv \frac{2 \mu_{0} a}{B_{0}^{2}} \frac{d p}{d r_{\psi}} \label{eq:betaPrimeDef}
\end{align}
because it constructs the poloidal magnetic field to be consistent with the Grad-Shafranov equation. We will find that the momentum transport is quite sensitive to $\beta'$, so it is an important parameter. In keeping with rough projections for ITER \cite{AymarITERSummary2001}, we use a pressure profile that is linear in $r_{\psi}$. This allows us to estimate that
\begin{align}
\beta' \approx - \frac{2 \mu_{0} p_{\text{axis}}}{B_{0}^{2}} \approx - 0.06 , \label{eq:betaPrimeEstimate}
\end{align}
using an ITER-like value for $p_{\text{axis}}$. Since we are running electrostatic simulations the value of $\beta$ itself has no effect.
We note that assuming a constant $\beta'$ (i.e. $d p / d r_{\psi}$) profile is formally inconsistent with the constant $d p / d \psi$ profile used in the MHD calculation of the Shafranov shift. Hence, using the results shown in figure \ref{fig:ShafShiftLoc} together with equation \refEq{eq:betaPrimeEstimate} is not formally valid. However, figure \ref{fig:ShafShiftLocPressureProf} shows that the magnitude and direction of the Shafranov shift is insensitive to large changes in the shape of the pressure profile at constant $R_{0}$, $a$, $\kappa_{b}$, $I_{p}$, and $p_{\text{axis}} / \psi_{0 b}$. This suggests that, since we have kept the proper parameters fixed, the mismatch between the pressure profile of the simulation and the pressure profile used to calculate the Shafranov shift will not have much effect.
\subsection{Parameter scan results}
\label{subsec:results}
A total of four scans in $\theta_{\kappa}$, the tilt angle of the flux surface of interest, were performed at
\begin{enumerate}
\item[(1)] $\beta' = 0$ with no Shafranov shift, \\
\item[(2)] $\beta' = 0$ with a modest Shafranov shift (approximately half the ITER-like Shafranov shift), \\
\item[(3)] $\beta' = 0$ with an ITER-like Shafranov shift, and \\
\item[(4)] an ITER-like $\beta' = -0.06$ with an ITER-like Shafranov shift.
\end{enumerate}
These scans were chosen to directly determine the independent influences of the Shafranov shift and $\beta'$, while minimizing the total number of simulations. The magnitude and direction of the local ITER-like Shafranov shift was kept consistent with equations \refEq{eq:radialShafShiftLocal} and \refEq{eq:axialShafShiftLocal}. Additionally, a single simulation was performed with $\beta' = -0.06$ and no Shafranov shift in order to isolate the effect of $\beta'$.
Four scans in $\rho_{0}$, the minor radial coordinate of the flux surface of interest, were performed at
\begin{enumerate}
\item[(1)] $\beta' = 0$ with no Shafranov shift, \\
\item[(2)] $\beta' = 0$ with an ITER-like Shafranov shift, \\
\item[(3)] an ITER-like $\beta' = -0.06$ with no Shafranov shift, and \\
\item[(4)] an ITER-like $\beta' = -0.06$ with an ITER-like Shafranov shift.
\end{enumerate}
All simulations had elliptical flux surfaces with $\theta_{\kappa} = \pi / 8$. These scans were done in order to investigate the balance between the Shafranov shift, which we expect to enhance the momentum transport, and $\beta'$, which our GS2 simulations will reveal to reduce the momentum transport. For these scans we kept $\beta'$ constant to be consistent with ITER (according to equation \refEq{eq:betaPrimeEstimate}) and again calculated the local Shafranov shift at each minor radius according to equations \refEq{eq:radialShafShiftLocal} and \refEq{eq:axialShafShiftLocal}.
Lastly a small scan was performed with circular flux surfaces in which $\theta_{\text{axis}}$, the direction of the Shafranov shift, was varied. This is unphysical, but it was done to explicitly isolate the effect of a pure flux surface Shafranov shift.
\subsubsection{Elliptical boundary tilt scans.}
\label{subsubsec:tiltedEllipticalScans}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/nonlinearAvgIonRatio.pdf}
\caption{The ion momentum transport for flux surfaces with no shift (black, circles), a modest shift (blue, triangles), and an ITER-like shift (red, squares) for $\beta' = 0$ (filled) and an ITER-like $\beta'$ (empty).}
\label{fig:momHeatFluxRatio}
\end{figure}
Figure \ref{fig:momHeatFluxRatio} shows the ratio of the time-averaged ion momentum flux to the time-averaged ion energy flux, calculated by GS2 for the tilted elliptical scans. As we will show in section \ref{subsec:betaProfile}, this quantity indicates the strength of momentum transport and is roughly proportional to the level of rotation (see equation \refEq{eq:rotationGradEst}). Figure \ref{fig:momHeatFluxRatio} also provides an estimate of the statistical error in the data. This error arises from performing a finite time-average over noisy turbulent quantities. It was estimated by repeating several simulations and computing the average difference between the corresponding results.
Figure \ref{fig:momHeatFluxRatio} demonstrates that the presence of an ITER-like Shafranov shift increases the momentum transport, here by approximately $30 \%$. As discussed in section \ref{sec:introduction}, this is expected because the Shafranov shift provides an additional source of up-down asymmetry and breaks both the mirror and tilting symmetry of the flux surfaces. However, we see that a non-zero $\beta'$ significantly reduces the momentum transport. We will investigate this result in section \ref{subsec:betaPrime} by studying at the magnitude of the up-down symmetry-breaking in the gyrokinetic equation. These two effects counteract one another and for ITER-like values at $\theta_{\kappa} = \pi / 8$ and $\rho_{0} = 0.54$ the shift is overshadowed by $\beta'$, leading to a net reduction in the momentum transport of about $30 \%$. In performing this scan, we added two simulations at $\pi / 16$ in order to better resolve the steep gradient that appears at small tilt angles. Additionally, we removed two simulations at $\pi / 2$ to save computational time because we had already confirmed that up-down symmetric shapes drive no rotation, even with a Shafranov shift.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/nonlinearAvgIonHeatFlux.pdf}
\caption{The ion energy flux for flux surfaces with no shift (black, circles), a modest shift (blue, triangles), and an ITER-like shift (red, squares) for $\beta' = 0$ (filled) and an ITER-like $\beta'$ (empty). In this and subsequent figures, whenever a single set of error bars is shown, it gives a representative estimate of the error for each data point.}
\label{fig:heatFlux}
\end{figure}
Figure \ref{fig:heatFlux} shows the ion energy flux. We see that it is fairly insensitive to the effects of both the Shafranov shift and $\beta'$ in the domain of $\theta_{\kappa} \in \left[ 0, \pi / 8 \right]$. At more extreme tilt angles we see that $\beta'$ dramatically increases the energy flux, as does the shift (albeit to a lesser extent).
\subsubsection{Minor radial scans.}
\label{subsubsec:minorRadialScans}
These scans keep $\beta'$, $d \Ln{T_{s}} / d \rho$, $d \Ln{n_{s}} / d \rho$, $q$, and $\hat{s}$ constant with minor radius. We chose to keep $\beta'$ constant to be consistent with ITER (according to equation \refEq{eq:betaPrimeEstimate}). The others were kept fixed in order to make comparisons with previous results more straightforward. However, constant values for $d \Ln{T_{s}} / d \rho$ and $d \Ln{n_{s}} / d \rho$ is not an unreasonable approximation to many experiments, especially in the core of tokamaks \cite{BarnesTrinity2010}. The local shift is calculated at each minor radius to be consistent with equations \refEq{eq:radialShafShiftLocal} and \refEq{eq:axialShafShiftLocal}, which result from the global MHD calculation.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/nonlinearAvgIonRatioWithMinorRadius.pdf}
\caption{The radial dependence of the momentum transport for flux surfaces with no shift (black, circles) and a strong shift (red, squares) varied according to equations \refEq{eq:radialShafShiftLocal} and \refEq{eq:axialShafShiftLocal}, for $\beta' = 0$ (filled) and an ITER-like $\beta'$ (empty).}
\label{fig:momHeatFluxRatioWithMinorRadius}
\end{figure}
The minor radial dependence of the momentum flux is shown in figure \ref{fig:momHeatFluxRatioWithMinorRadius}. Note that at $\rho_{0} = 1$ the momentum transport in the shifted configurations with and without $\beta'$ are indistinguishable. Comparing the two scans with $\beta' = 0$, we see that the difference in the momentum transport from the two scans increases with minor radius. The only difference between the scans is the presence of the local Shafranov shift, which also increases with minor radius. Hence, this reinforces a result of figure \ref{fig:momHeatFluxRatio}: the Shafranov shift increases the momentum transport. Similarly, comparing the two scans with no shift reinforces the fact that $\beta'$ reduces the momentum transport (which we also observed in figure \ref{fig:momHeatFluxRatio}). Additionally, comparing the no shift, $\beta' = 0$ case to the ITER-like shift, ITER-like $\beta'$ case demonstrates the counteracting effects of the shift and $\beta'$ on the momentum transport. Because the shift is weak at small values of $\rho_{0}$, the net effect of the shift and $\beta'$ is to lower the momentum transport. However, at large values of $\rho_{0}$ the shift is stronger, but $\beta'$ remains the same. Here the net effect of the shift and $\beta'$ is to enhance the momentum transport.
Lastly, a dominant trend appearing in figure \ref{fig:momHeatFluxRatioWithMinorRadius} is the roughly linear decrease of the momentum transport with minor radius. It is most clearly seen in the data series with no shift and $\beta' = 0$, because the only difference between the four simulations is the value of the minor radius. This trend (i.e. an increase in the momentum transport with increasing aspect ratio) is not currently understood as nearly all simulations of intrinsic rotation from up-down asymmetry were performed using $r_{\psi 0} / R_{c 0} \approx 1 / 6$. However, it was also observed in several simulations performed at $r_{\psi 0} / R_{c 0} \approx 1 / 12$ and $r_{\psi 0} / R_{c 0} \approx 1 / 3$ in reference \cite{BallMomUpDownAsym2014}.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/nonlinearAvgIonRatioWithShift}
\caption{The change in the momentum transport caused by introducing a local Shafranov shift with a magnitude of $-d r_{c} / d r_{\psi}$ for $\beta' = 0$ (filled) and an ITER-like $\beta'$ (empty).}
\label{fig:momHeatFluxRatioWithShift}
\end{figure}
In figure \ref{fig:momHeatFluxRatioWithShift} we show $\Delta_{\text{shift}} \left[ \left( v_{th, i} / R_{c 0} \right) \left\langle \Pi_{\zeta i} \right\rangle_{t} / \left\langle Q_{i} \right\rangle_{t} \right]$, the change in the momentum transport due to the Shafranov shift, where $r_{c} \equiv \sqrt{ \left( R_{c} - R_{0} \right)^{2} + Z_{c}^{2}}$ and $\Delta_{\text{shift}} \left[ x \right]$ is defined to be the value of $x$ when the Shafranov shift is included minus the value of $x$ when the Shafranov shift is omitted. This figure uses the same data as figure \ref{fig:momHeatFluxRatioWithMinorRadius}, but more clearly demonstrates that the momentum transport is not sensitive to small changes in the Shafranov shift. Rather it increases smoothly and fairly linearly with the strength of the Shafranov shift, irrespective of the value of $\beta'$.
\subsubsection{Circular flux surface scan.}
\label{subsubsec:circularScan}
\begin{figure}
(a) \hspace{0.27\textwidth} (b) \hspace{0.28\textwidth} (c)
\begin{center}
\includegraphics[height=0.24\textwidth]{figs/circScanHoriz.pdf}
\includegraphics[height=0.24\textwidth]{figs/circScanDiag.pdf}
\includegraphics[height=0.24\textwidth]{figs/circScanVert.pdf}
\end{center}
\caption{The magnetic geometry for circular flux surfaces with an ITER-like (a) horizontal shift, (b) diagonal shift, or (c) vertical shift.}
\label{fig:circScanGeo}
\end{figure}
To completely isolate the effect of the Shafranov shift on momentum transport we also ran simulations with shifted circular flux surfaces as shown in figure \ref{fig:circScanGeo}. To create up-down asymmetry and drive momentum transport we varied the direction of the tilt by changing the parameter $\theta_{\text{axis}}$ with the magnitude of the shift fixed at $\sim 30 \%$ larger than an untilted ITER-like machine. Scanning $\theta_{\text{axis}}$ is unphysical because circular flux surfaces can only ever have a shift in the outboard radial direction, which corresponds to $\theta_{\text{axis}} = 0$. Though unphysical, this scan will help clarify the influence of the Shafranov shift.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/nonlinearAvgIonHeatFluxCirc.pdf}
\caption{The energy flux for circular flux surfaces with no shift (black, dotted line) and an ITER-like shift (red, square points) as a function of the direction of the Shafranov shift. All simulations have $\beta' = 0$.}
\label{fig:heatFluxCirc}
\end{figure}
Figure \ref{fig:heatFluxCirc} shows that the presence and direction of the Shafranov shift has little effect on the ion energy flux from circular flux surfaces. This behavior is similar to the tilted elliptical results (see figure \ref{fig:heatFlux}) in the range of $\theta_{\kappa} \in \left[ 0, \pi / 8 \right]$, but different from the tilted elliptical results in the range of $\theta_{\kappa} \in \left[ \pi / 8, \pi / 2 \right]$. This is consistent because the magnitude of the shift in the circular equilibria is similar to that of the elliptical equilibria in the range of $\theta_{\kappa} \in \left[ 0, \pi / 8 \right]$, but considerably less than the magnitude of the shift present in the elliptical equilibria with larger tilt angles. Therefore, both figures indicate that the shift present in the circular and minimally-tilted elliptical flux surfaces is not strong enough to modify the energy flux significantly.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/nonlinearAvgIonRatioCircThesis.pdf}
\caption{The momentum flux for circular flux surfaces with no shift (black, dotted line) and an ITER-like shift (red, square points) as a function of the direction of the Shafranov shift. All simulations have $\beta' = 0$. Note that we have kept the range of the vertical axis the same as in figures \ref{fig:momHeatFluxRatio} and \ref{fig:momHeatFluxRatioWithMinorRadius} for ease of comparison.}
\label{fig:momHeatFluxRatioCirc}
\end{figure}
Figure \ref{fig:momHeatFluxRatioCirc} shows the effect of a strong Shafranov shift on momentum transport. We see that a pure shift in circular flux surfaces (even when it is diagonal or vertical) drives minimal rotation compared to that generated by elliptical flux surfaces (as shown in figure \ref{fig:momHeatFluxRatio}). This is somewhat surprising since the shift is an $m = 1$ shaping effect and we expect the momentum flux to scale as $\Exp{-m}$ in mirror symmetric configurations \cite{BallMirrorSymArg2016}. However, there are two important caveats. Firstly, the exponential scaling is only true in the limit of $m \gg 1$, which is clearly not satisfied for $m = 1$. Secondly, the Shafranov shift has a relatively minor effect on the magnetic equilibrium compared with elongating the flux surfaces to $\kappa = 2$ (even when the shift is $30 \%$ stronger than that expected in ITER). This can be quantified by looking at the geometric coefficients that appear in the gyrokinetic equations (see \ref{app:geoCoeff} and reference \cite{BallMirrorSymArg2016} for more details on these coefficients). The geometric coefficients are the only way the magnetic geometry enters the local gyrokinetic model, so we know they must control the momentum transport. Plotting the geometric coefficient $\left| \ensuremath{\vec{\nabla}} \psi \right|^{2}$ as an example produces figure \ref{fig:gds22}, which shows that elongating an unshifted circular configuration to $\kappa = 2$ causes a $300 \%$ change, while introducing the Shafranov shift only causes a $50 \%$ change. To fairly compare the ability of the Shafranov shift and elongation to drive rotation we should control for the effect on the magnetic equilibrium. From figure \ref{fig:gds22} we see that an elliptical configuration with $\kappa = 1.2$ has a similar effect on $\left| \ensuremath{\vec{\nabla}} \psi \right|^{2}$ as the pure Shafranov shift. Performing a nonlinear gyrokinetic simulation of a tilted elliptical configuration with $\kappa = 1.2$ and $\theta_{\kappa} = \pi / 8$ demonstrates that, like a pure Shafranov shift, it generates little momentum transport. This suggests that the Shafranov shift and elongation drive similar levels of rotation when they alter the geometric coefficients to a similar degree. Elongation is capable of driving much more rotation than a pure Shafranov shift, because it can have a much larger effect on the geometric coefficients. The effect of the Shafranov shift on the geometric coefficients is constrained through a practical limit on the maximum value of $\beta$. This proves to be more restrictive than the vertical stability limit, which constrains the externally-applied elongation.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{figs/gds22.pdf}
\caption{The geometric coefficient $\left| \ensuremath{\vec{\nabla}} \psi \right|^{2}$ for unshifted circular flux surfaces (black, solid), circular flux surfaces with a strong vertical shift (blue, dashed), and unshifted flux surfaces with a vertical elongation of $\kappa = 2$ (red, dotted) or $\kappa = 1.2$ (red, dash-dotted) normalized to the unshifted circular value.}
\label{fig:gds22}
\end{figure}
\subsection{Effect of the value of $\beta'$}
\label{subsec:betaPrime}
\begin{figure}
\begin{center}
(a) \hspace{0.4\textwidth} (b) \hspace{0.35\textwidth}
\includegraphics[height=0.3\textwidth]{figs/drifts_0_3.pdf}
\includegraphics[height=0.3\textwidth]{figs/drifts_0_54.pdf}
(c) \hspace{0.4\textwidth} (d) \hspace{0.35\textwidth}
\includegraphics[height=0.35\textwidth]{figs/drifts_0_8.pdf}
\includegraphics[height=0.35\textwidth]{figs/drifts_1_0.pdf}
\end{center}
\caption{The geometric coefficient $\vec{v}_{d s} \cdot \ensuremath{\vec{\nabla}} \alpha$ in units of $v_{th, s}^{2} / \left( a^{2} \Omega_{s} \right)$ without Shafranov shift at (a) $\rho_{0} = 0.3$, (b) $\rho_{0} = 0.54$, (c) $\rho_{0} = 0.8$, and (d) $\rho_{0} = 1$ for no $\beta'$, $w_{||}^{2} = v_{th, s}^{2}$, $w_{\perp}^{2} = 0$ (black, solid); an ITER-like $\beta'$, $w_{||}^{2} = v_{th, s}^{2}$, $w_{\perp}^{2} = 0$ (red, dotted); and an ITER-like $\beta'$, $w_{||}^{2} = 0$, $w_{\perp}^{2} = 2 v_{th, s}^{2}$ (blue, dashed).}
\label{fig:driftCoeff}
\end{figure}
In section \ref{subsec:results} we included the effect of the Shafranov shift in nonlinear, local gyrokinetic simulations and found that it enhanced momentum transport as expected. Since the magnitude of the shift depends on the plasma pressure, we also included a non-zero $\beta'$. While the Shafranov shift alters the spacing between flux surfaces, $\beta'$ enters through the right-hand side of the Grad-Shafranov equation and alters the local magnetic shear (i.e. the radial derivative of the magnetic field line pitch angle). We found that $\beta'$ strongly reduced the momentum flux, often entirely canceling the enhancement due to the Shafranov shift. Consequently, it is important to understand how $\beta'$ alters the geometric coefficients of gyrokinetics.
In \ref{app:geoCoeff} we discuss how $\beta'$ enters into the analytic expressions for the geometric coefficients. We show that $\beta'$ vanishes in the large aspect ratio limit (for the orderings of equation \refEq{eq:gradShafOrderings}), like the Shafranov shift. This means that for large aspect ratio tokamaks $\beta'$ can be ignored and the results of reference \cite{BallMomUpDownAsym2014} (which ignores $\beta'$) apply. However, the Shafranov shift also vanishes in this limit, so it cannot be used to enhance the momentum transport.
Figure \ref{fig:driftCoeff} uses the geometries from figure \ref{fig:momHeatFluxRatioWithMinorRadius} to show the quantitative effect of $\beta'$ on the geometric coefficient $\vec{v}_{d s} \cdot \ensuremath{\vec{\nabla}} \alpha$ (defined by equation \refEq{eq:driftVelAlpha}) with different values of $w_{||}$ and $w_{\perp}$. Here $\vec{w}$ is the velocity coordinate in the frame rotating with the background plasma flow, the $||$ subscript indicates parallel to the magnetic field, the $\perp$ subscript indicates perpendicular to the magnetic field, $\vec{v}_{d s}$ is the guiding center particle magnetic drifts,
\begin{align}
\alpha \equiv& ~ \zeta - I \left( \psi \right) \left. \int_{\theta_{\alpha} \left( \psi \right)}^{\theta} \right|_{\psi} d \theta' \left( R^{2} \vec{B} \cdot \ensuremath{\vec{\nabla}} \theta' \right)^{-1} \label{eq:alphaDef}
\end{align}
is the coordinate within the flux surface and perpendicular to the magnetic field, and $\theta_{\alpha} \left( \psi \right)$ is a free function. Previous work seems to indicate that $\vec{v}_{d s} \cdot \ensuremath{\vec{\nabla}} \alpha$ may be the most important geometric coefficient for understanding intrinsic rotation transport due to up-down asymmetry \cite{BallMomFluxScaling2016}. We see that including a non-zero $\beta'$ tends to reduce the up-down asymmetry of $\vec{v}_{d s} \cdot \ensuremath{\vec{\nabla}} \alpha$, which is consistent with the observed reduction in momentum transport.
\subsection{Effect of the $\beta$ profile}
\label{subsec:betaProfile}
In order to estimate a realistic value for $\beta'$, we used the on-axis value of $\beta$ predicted for ITER and assumed $\beta$ was linear with minor radius $r_{\psi}$. This gave a reasonable order of magnitude estimate. However, since the momentum transport is strongly and adversely affected by $\beta'$ it is worthwhile to discuss the implications of different radial profiles of $\beta'$. For example, we expect that in H-mode operation $\beta'$ would be larger at the plasma edge and smaller in the core compared to L-mode. Unfortunately, since intrinsic rotation is ultimately driven by the gradients in density and temperature, $\beta'$ is necessary, even though including the effect of $\beta'$ in the geometric coefficients reduces the momentum flux. To see the relationship between $\beta'$ and the rotation gradient we will follow the analysis of reference \cite{BallMomUpDownAsym2014}.
First, we neglect the momentum pinch (which can only ever enhance the level of rotation) and assume that diffusion is the only mechanism balancing the intrinsic source to get
\begin{align}
\left\langle \Pi_{\zeta i} \right\rangle_{t} \approx D_{\Pi i} n_{i} m_{i} R_{c}^{2} \frac{d \Omega_{\zeta i}}{d r_{\psi}} , \label{eq:momFluxEstimate}
\end{align}
where $\left\langle \Pi_{\zeta i} \right\rangle_{t}$ is the time-averaged intrinsic ion momentum flux source term arising from up-down asymmetry (i.e. the momentum flux calculated by GS2 for $\Omega_{\zeta i} = d \Omega_{\zeta i} / d r_{\psi} = 0$), $D_{\Pi i}$ is the momentum diffusivity (i.e. the kinematic viscosity), $R_{c}$ is the major radial location of the center of a given flux surface, $\Omega_{\zeta i} \equiv u_{\zeta i} / R$ is the ion rotation frequency, and $u_{\zeta i}$ is the ion bulk toroidal velocity. We take the energy flux to be the diffusion of a temperature gradient \cite{FreidbergFusionEnergy2007pg452} according to
\begin{align}
\left\langle Q_{i} \right\rangle_{t} \approx - D_{Q i} n_{i} \frac{d T_{i}}{d r_{\psi}} , \label{eq:heatFluxEstimate}
\end{align}
where $\left\langle Q_{i} \right\rangle_{t}$ is the time-averaged energy flux calculated by GS2. Combining these two equations through the turbulent ion Prandtl number $Pr_{i} \equiv D_{\Pi i} / D_{Q i} \approx 0.7$ \cite{BallMomUpDownAsym2014} gives
\begin{align}
\frac{1}{v_{th, i}} \frac{d u_{\zeta i}}{d r_{\psi}} \approx \frac{-1}{2 Pr_{i}} \left( \frac{v_{th, i}}{R_{c}} \frac{\left\langle \Pi_{\zeta i} \right\rangle_{t}}{\left\langle Q_{i} \right\rangle_{t}} \right) \frac{d}{d r_{\psi}} \Ln{T_{i}} . \label{eq:rotationGradEstSimple}
\end{align}
where we used that $T_{i} = m_{i} v_{th, i}^{2} / 2$. Doing this is useful because the Prandtl number is expected to be unaffected by changes in tokamak parameters. We will introduce the Alfv\'{e}n Mach number, $M_{A} \equiv \left| u_{\zeta i} \right| \sqrt{\mu_{0} n_{i} m_{i}} / B_{0}$, because it is the relevant quantity for stabilizing MHD modes, such as resistive wall modes. Neglecting the density gradient (because $d \Ln{T_{i}} / d r_{\psi}$ is three times larger than $d \Ln{n_{i}} / d r_{\psi}$) as well as assuming $n_{e} = n_{i}$ and $T_{e} = T_{i}$ allows equation \refEq{eq:rotationGradEstSimple} to be rewritten as
\begin{align}
M_{A} \left( \rho \right) \approx \left| \int_{1}^{\rho} d \rho' \frac{1}{2 \sqrt{2} Pr_{i} \left( \rho' \right)} \left( \frac{v_{th, i} \left( \rho' \right)}{R_{c} \left( \rho' \right)} \frac{\left\langle \Pi_{\zeta i} \left( \rho' \right) \right\rangle_{t}}{\left\langle Q_{i} \left( \rho' \right) \right\rangle_{t}} \right) \frac{\beta' \left( \rho' \right)}{\sqrt{\beta \left( \rho' \right)}} \right| . \label{eq:rotationGradEst}
\end{align}
We wrote this expression in terms of $\left( v_{th, i} / R_{c} \right) \left\langle \Pi_{\zeta i} \right\rangle_{t} / \left\langle Q_{i} \right\rangle_{t}$ because it is the normalized parameter that indicates how strongly a given geometry drives rotation.
\begin{figure}
\begin{center}
(a) \hspace{0.4\textwidth} (b) \hspace{0.35\textwidth}
\includegraphics[width=0.43\textwidth]{figs/betaProfile.pdf}
\includegraphics[width=0.47\textwidth]{figs/betaPrimeProfile.pdf}
(c) \hspace{0.4\textwidth} (d) \hspace{0.35\textwidth}
\includegraphics[width=0.45\textwidth]{figs/MachNumberProfileWithShift.pdf}
\includegraphics[width=0.45\textwidth]{figs/MachNumberProfileWithoutShift.pdf}
\end{center}
\caption{Example (a) $\beta$ profiles with their corresponding (b) $\beta'$ and (c,d) Alfv\'{e}n Mach number profiles, estimated using the data from figure \ref{fig:momHeatFluxRatioWithMinorRadius} both (c) with and (d) without the effects of the pressure profile on the magnetic equilibrium, for constant $\beta'$ (black, solid), linear peaked $\beta'$ (red, dotted), and linear hollow $\beta'$ (blue, dashed) profiles.}
\label{fig:MachNumberProfile}
\end{figure}
Equation \refEq{eq:rotationGradEst} shows several competing dependencies on $\beta$ and $\beta'$, both explicitly and through $\left( v_{th, i} / R_{c} \right) \left\langle \Pi_{\zeta i} \right\rangle_{t} / \left\langle Q_{i} \right\rangle_{t}$. Hence, it is difficult to analytically determine the $\beta$ profile that maximizes rotation. However, we can perform a bilinear interpolation of the data in figure \ref{fig:momHeatFluxRatioWithMinorRadius} to approximate the functional form of $G \left( \rho, \beta' \right) \equiv \left( v_{th, i} / R_{c} \right) \left\langle \Pi_{\zeta i} \right\rangle_{t} / \left\langle Q_{i} \right\rangle_{t}$. We note that the dependence on $\rho$ also includes a change in the strength of the Shafranov shift according to equations \refEq{eq:radialShafShiftLocal} and \refEq{eq:axialShafShiftLocal}. To estimate the function $G \left( \rho, \beta' \right)$ between data points at $\left( \rho_{1}, \beta'_{1} \right)$, $\left( \rho_{1}, \beta'_{2} \right)$, $\left( \rho_{2}, \beta'_{1} \right)$, and $\left( \rho_{2}, \beta'_{2} \right)$ we use
\begin{align}
G \left( \rho, \beta' \right) &\approx \frac{\beta'_{2} - \beta'}{\beta'_{2} - \beta'_{1}} \left( \frac{\rho_{2} - \rho}{\rho_{2} - \rho_{1}} G \left( \rho_{1}, \beta'_{1} \right) + \frac{\rho - \rho_{1}}{\rho_{2} - \rho_{1}} G \left( \rho_{2}, \beta'_{1} \right) \right) \\
&+ \frac{\beta' - \beta'_{1}}{\beta'_{2} - \beta'_{1}} \left( \frac{\rho_{2} - \rho}{\rho_{2} - \rho_{1}} G \left( \rho_{1}, \beta'_{2} \right) + \frac{\rho - \rho_{1}}{\rho_{2} - \rho_{1}} G \left( \rho_{2}, \beta'_{2} \right) \right) . \nonumber
\end{align}
In the region $\rho < 0.3$ we do not have data, so we assume that $\left( v_{th, i} / R_{c} \right) \left\langle \Pi_{\zeta i} \right\rangle_{t} / \left\langle Q_{i} \right\rangle_{t}$ is constant in $\rho$. This assumption is conservative compared to a linear extrapolation using the data at $\rho = 0.3$ and $\rho = 0.54$. Furthermore, it leads to zero slope on-axis, which is consistent with constant asymptotic behavior in the large aspect ratio limit. To calculate the rotation profile for geometries with a strong Shafranov shift (shown in figure \ref{fig:MachNumberProfile}(c)) we used only the red square points in figure \ref{fig:momHeatFluxRatioWithMinorRadius}. Alternatively, if we assume that $\beta$ is very low (i.e. $\beta' = 0$) we can ignore the effect of the pressure profile on the magnetic geometry (i.e. neglect the Shafranov shift and assume pressure gradient term in the Grad-Shafranov equation is much smaller than the toroidal field flux function term). In this case the rotation profile (shown in figure \ref{fig:MachNumberProfile}(d)) can be calculated by a 1-D interpolation of the filled black circles because the turbulent transport becomes independent of $\beta'$.
Figure \ref{fig:MachNumberProfile} shows that both the Shafranov shift and the shape of the $\beta$ profile have a significant effect on the rotation profile. A broader $\beta$ profile consistently produces a broader rotation profile, but with a lower on-axis Mach number. This means that the $\beta$ profile that maximizes the {\it on-axis} Mach number is not necessarily optimal because broad rotation profiles are expected to be significantly more effective at stabilizing resistive wall modes \cite{LiuITERrwmStabilization2004}. Additionally, figures \ref{fig:MachNumberProfile}(c) and (d) indicate that stronger plasma pressure effects (i.e. Shafranov shift and $\beta'$) will cause up-down asymmetry to drive broader intrinsic rotation profiles. The reason for this can be seen in figure \ref{fig:momHeatFluxRatioWithMinorRadius}. Adding both the Shafranov shift and $\beta'$ (to go from the filled black circles to the empty red squares) reduces the core momentum transport, while enhancing the edge momentum transport. Lastly, we see that the largest rotation gradient occurs at the edge of the peaked pressure profile because the integral over the momentum flux in equation \refEq{eq:rotationGradEst} is weighted towards regions with small $\beta$ and large $\beta'$. This indicates that, even though the up-down asymmetry of a single-null divertor is usually limited to the edge, it may still drive significant rotation (especially in H-mode operation).
\section{Conclusions}
\label{sec:conclusions}
This paper focuses on two competing effects influencing the momentum transport: the Shafranov shift and $\beta'$. Together the two effects reduce momentum transport in the core, enhance it near the edge, and roughly cancel when averaged over the entire device. Using the nonlinear gyrokinetic simulations shown in figure \ref{fig:momHeatFluxRatioWithMinorRadius}, we estimate the rotation profile when these two effects are included (i.e. figure \ref{fig:MachNumberProfile}(c)) and when they are omitted (i.e. figure \ref{fig:MachNumberProfile}(d)). Comparing these profiles demonstrates that the on-axis value of the rotation is roughly unchanged, but the rotation profile is broadened (which is expected to be advantageous for stabilizing resistive wall modes). The magnitude of the on-axis rotation was found to be $\sim 1 \%$ (without including any enhancement due to the momentum pinch effect), which is in the range of what is needed to stabilize resistive wall modes in a large device like ITER (i.e. $0.5\% - 5\%$) \cite{LiuITERrwmStabilization2004}.
As anticipated a strong Shafranov shift was found to enhance the momentum transport in up-down asymmetric configurations because the shift itself becomes up-down asymmetric. The magnitude and direction of the shift was found to be insensitive to the shape of both the toroidal current (for a pressure profile that is a uniform fraction of the current profile) and pressure (for a uniform current profile) profiles at fixed geometry, plasma current, and average $d p / d \psi$.
On the other hand, it was found that the effect of $\beta'$ on the magnetic equilibrium significantly reduces the momentum transport, often entirely canceling the effect of the Shafranov shift. Consequently, the shape and magnitude of the rotation profile is sensitive to the radial profile of $\beta$. By studying the geometric coefficients, we found that, like the Shafranov shift, $\beta'$ appears to $\Order{\epsilon}$. However, unlike the Shafranov shift it tends to reduce the up-down asymmetry of the geometric coefficients.
\ack
J.B. and F.I.P. were funded in part by the RCUK Energy Programme (grant number EP/I501045). Computing time for this work was provided by the Helios supercomputer at IFERC-CSC under the projects SPIN, TRIN, MULTEIM, and GKMSC. The authors also acknowledge the use of ARCHER through the Plasma HEC Consortium EPSRC grant number EP/L000237/1 under the projects e281-gs2 and e281-rotation. J.P.L. was supported by the U.S. Department of Energy, Office of Science, Fusion Energy Sciences under Award No. DE-FG02-91ER-54109. A.J.C. was supported by the U.S. Department of Energy, Office of Science, Fusion Energy Sciences under Award Nos. DE-FG02-86ER53223 and DE-SC0012398.
|
\section{introduction}
Gibbs \cite{Gibbs}, using entropy method, proved that the temperature distribution of an object in an external gravitational field is uniform. The Boltzmann equation produces the same result \cite{BoltzmannEquation} for a dilute gas. But our intuition tells us that a molecule moving up in a gravitational field loses energy and therefore should lead to a lower temperature. With the similar reasoning, Josef Loschmidt claimed that {\it the equilibrium temperature of a gas column subject to gravity should be lower at the top of the column and higher at its base.} This is known as Loschmidt's gravito-thermal effect \cite{Trupp,book}, about which some experiments \cite{Graeff, Liao} have been carried out. However, as Maxwell noted, this effect violates the second law of thermodynamics. Then a paradox was raised by Sheehan {\it et al.} \cite{Sheehan} saying that gravitational field leads to effect violating the second law of thermodynamics. Wheeler \cite{Wheeler} examined the paradox and found that there is no violation.
Our intuition works better on kinetic theory instead of entropy theory. Therefore Coombes and Laue \cite{cccc} derived a relation about molecular distribution function with respect to height and velocity, with which, they were able to explain why the temperature remains uniform intuitively. However, their relation was obtained without taking into account the molecular collisions.
To take into account molecular collisions, we present a new derivation. We do this in two steps. In step one, we use the principle of detailed balance \cite{Tolman} to study only those processes involving no collisions. In step two, we use the principle of microscopic reversibility \cite{Lewis} and a thought experiment to study collisions.
\section{detailed balance principle with no collisions}
Figure~\ref{gas} shows a dilute gas in equilibrium in a gravitational field with a constant acceleration $g$. Let us study a molecule at position $\vec{r}_1(x_1,y_1,z_1)$ with velocity $\vec{v}_1(v_{1x},v_{1y},v_{1z})$. In case of no collisions, the molecule reaches position $\vec{r}_2(x_2,y_2,z_2)$ with corresponding velocity $\vec{v}_2(v_{2x},v_{2y},v_{2z})$ where
\begin{equation}\label{first}
\begin{array}{l}
v_{2x}=v_{1x},\\
v_{2y}=v_{1y},\\
\frac{1}{2}mv_{2z}^2+mgz_2=\frac{1}{2}mv_{1z}^2+mgz_1,
\end{array}
\end{equation}
and $m$ is the mass of the molecule.
\begin{figure*}
\subfloat[
A molecule at position $\vec{r}_1$ with velocity $\vec{v}_1$. When no collisions happen on the path, it reaches $\vec{r}_2$ with velocity $\vec{v}_2$\label{sfig:testa}]{%
\includegraphics[width=8.2cm]{gasrr2.pdf}
}
\hfill
\subfloat[
A molecule at position $\vec{r}_2$ with velocity $-\vec{v}_2$. When no collisions happen on the path, it reaches $\vec{r}_1$ with velocity $-\vec{v}_1$\label{sfig:testa}]{%
\includegraphics[width=8.2cm]{gasrrv2.pdf}
}
\caption{A dilute gas in a gravitational field.
Two cross sections $S_1$ and $S_2$ are also shown.
}
\label{gas}
\end{figure*}
Let us introduce a distribution function $f(\vec{r},\vec{v})$ that is defined in such a way that $f(\vec{r},\vec{v})d^3\vec{r}d^3\vec{v}$ is the number of molecules within phase space volume element $d^3\vec{r}d^3\vec{v}$ in the neighbourhood of ($\vec{r},\vec{v}$). The system being in equilibrium means that the distribution function $f(\vec{r},\vec{v})$ must obey some relation which we will find in two steps.
In step one, we study only those processes involving no collisions for a molecule moving from $\vec{r}_1$ to $\vec{r}_2$. For that, we write the distribution function into two parts,
\begin{eqnarray}\label{f12}
f(\vec{r}_1,\vec{v}_1)= &&f(\vec{r}_1,\vec{v}_1, \mathop{\vec{r}_1\longrightarrow \vec{r}_2}_{\text{collisions}}^{\text{no}})\nonumber\\
+&& f(\vec{r}_1,\vec{v}_1, \mathop{\vec{r}_1\longrightarrow \cdots }_{\text{collisions}}^{\text{with}}).
\end{eqnarray}
When no collisions happen on the path, a molecule from $\vec{r}_1$ with $\vec{v}_1$ reaches $\vec{r}_2$ with $\vec{v}_2$. When collisions happen before it reaches $\vec{r}_2$, the molecule will deviate from the path and go somewhere else. In the same way, we write
\begin{eqnarray}\label{f21}
f(\vec{r}_2,-\vec{v}_2)=&& f(\vec{r}_2,-\vec{v}_2, \mathop{\vec{r}_2\longrightarrow \vec{r}_1}_{\text{collisions}}^{\text{no}})\nonumber\\
+&& f(\vec{r}_2,-\vec{v}_2, \mathop{\vec{r}_2\longrightarrow \cdots }_{\text{collisions}}^{\text{with}}).
\end{eqnarray}
In order to use the principle of detailed balance, imagine two cross sections $S_1$ and $S_2$ at heights $z_1$ and $z_2$, respectively (Fig.~\ref{gas}).
Then we can introduce two quantities
\begin{eqnarray}\label{1}
&&\Delta F(\vec{v}_1, \mathop{S_1\longrightarrow S_2}_{\text{collisions}}^{\text{no}})\nonumber\\
=&& \iint\limits_{S_1} f(\vec{r}_1,\vec{v}_1, \mathop{\vec{r}_1\longrightarrow S_2}_{\text{collisions}}^{\text{no}}) \Delta v_{1x}\Delta v_{1y}\Delta v_{1z} dx_1dy_1
\end{eqnarray}
and
\begin{eqnarray}\label{2}
&& \Delta F(-\vec{v}_2, \mathop{S_2\longrightarrow S_1}_{\text{collisions}}^{\text{no}} )\nonumber \\
=&& \iint\limits_{S_2} f(\vec{r}_2,-\vec{v}_2, \mathop{\vec{r}_2\longrightarrow S_1}_{\text{collisions}}^{\text{no}} )\Delta v_{2x}\Delta v_{2y}\Delta v_{2z} dx_2dy_2,
\end{eqnarray}
where
\begin{equation}\label{3}
\Delta v_{2x}=\Delta v_{1x},\ \ \Delta v_{2y}=\Delta v_{1y},\ \ \Delta v_{2z}=\frac{v_{1z}}{v_{2z}}\Delta v_{1z},
\end{equation}
which is from Eq.~(\ref{first}).
One can introduce a flux density \cite{Jeans, Chapman, Wheeler} $j(\vec{r},\vec{v})=v_z f(\vec{r},\vec{v})$ to construct a continuity equation. But here, given that the system is in equilibrium, we can use the principle of detailed balance \cite{Tolman} to write
\begin{equation}\label{4}
{v}_{1z}\Delta F(\vec{v}_1, \mathop{S_1\longrightarrow S_2}_{\text{collisions}}^{\text{no}})= v_{2z}\Delta F(-\vec{v}_2, \mathop{S_2\longrightarrow S_1}_{\text{collisions}}^{\text{no}}),
\end{equation}
which means that in unit time, the average number of those molecules crossing $S_1$ with velocity $\vec{v}_1$ reaching $S_2$ with velocity $\vec{v}_2$ equals the average number of those molecules crossing $S_2$ with velocity $-\vec{v}_2$ reaching $S_1$ with velocity $-\vec{v}_1$.
Combining all the equations together, we get
\begin{eqnarray}\label{int}
&&\iint\limits_{S_1}f(\vec{r}_1,\vec{v}_1, \mathop{\vec{r}_1\longrightarrow S_2}_{\text{collisions}}^{\text{no}}) dx_1dy_1 \nonumber\\
=&&\iint\limits_{S_2}f(\vec{r}_2,-\vec{v}_2, \mathop{\vec{r}_2\longrightarrow S_1}_{\text{collisions}}^{\text{no}} )dx_2dy_2 .
\end{eqnarray}
In fact, the function $f(\vec{r},\vec{v})$ is uniform on $S_1$ and $S_2$ separately, so Eq. (\ref{int}) means
\begin{equation} \label{5}
f(\vec{r}_1,\vec{v}_1, \mathop{\vec{r}_1\longrightarrow \vec{r}_2}_{\text{collisions}}^{\text{no}}) =f(\vec{r}_2,-\vec{v}_2,\mathop{\vec{r}_2\longrightarrow \vec{r}_1}_{\text{collisions}}^{\text{no}}).
\end{equation}
So far we have only considered those processes involving no collisions. We will show in the next section that a similar relation holds for all processes.
\section{Microscopic reversibility and collisions}
Now we study those processes involving collisions. We need to use a thought experiment. First, we look at the dilute gas in Fig.~\ref{gas} and observe how the molecules move and collide as time goes on. Then, at a random time, we put a dark matter \cite{darkmatter} molecule at $\vec{r}_1$ with $\vec{v}_1$. Since the dark matter molecule does not collide with any other molecules, it would always reach $\vec{r}_2$ with velocity $\vec{v}_2$. We do the thought experiment $N$ times. Suppose $N_1$ times the dark matter molecule passes through at least one ordinary molecule on the path.
In the same way, we do thought experiment of a dark matter molecule going from $\vec{r}_2$ with $-\vec{v}_2$ to $\vec{r}_1$. Again, we do the experiment $N$ times, and $N_2$ times the dark matter molecule passes through at least one ordinary molecule on the path.
In the limit of $N\to \infty$, we have
\begin{equation} \label{NN}
\lim_{N\to \infty}\frac{N_1}{N}= \lim_{N\to \infty}\frac{N_2}{N}.
\end{equation}
This is guaranteed by the principle of microscopic reversibility \cite{Lewis}: {\it Corresponding to every individual process there is a reverse process, and in a state of equilibrium the average rate of every process equals the average rate of its reverse process}.
In the limit of $N\to \infty$, for every microscopic process counted in $N_1$, there is a time reversal microscopic process counted in $N_2$, and vice versa.
Now, we replace the dark matter molecule by an ordinary molecule and replicate each thought experiment. Whenever the dark matter molecule passes through the first molecule on its path, the ordinary molecule would collide with it and deviate from the path. Therefore in $N$ experiments about the ordinary molecule going from $\vec{r}_1$ with $\vec{v}_1$ to $\vec{r}_2$, $N_1$ of them would have collisions. So we know that the probability for a molecule to collide with other molecules on its path going from $\vec{r}_1$ with $\vec{v}_1$ to $\vec{r}_2$ is
\begin{equation}
P_1=\lim_{N\to \infty}\frac{N_1}{N}.
\end{equation}
In the same way, the probability for a molecule to collide with other molecules on its path going from $\vec{r}_2$ with $-\vec{v}_2$ to $\vec{r}_1$ is
\begin{equation}
P_2=\lim_{N\to \infty}\frac{N_2}{N}.
\end{equation}
Combining with Eq.~(\ref{NN}), we know that
\begin{equation}\label{psymmetry}
P_1=P_2.
\end{equation}
From Eq.~(\ref{f12}) and Eq.~(\ref{f21}), we can write
\begin{equation} \label{e1}
f(\vec{r}_1,\vec{v}_1, \mathop{\vec{r}_1\longrightarrow \vec{r}_2}_{\text{collisions}}^{\text{no}}) = f(\vec{r}_1,\vec{v}_1)(1-P_1),
\end{equation}
and
\begin{equation}\label{e2}
f(\vec{r}_2,-\vec{v}_2, \mathop{\vec{r}_2\longrightarrow \vec{r}_1}_{\text{collisions}}^{\text{no}}) = f(\vec{r}_2,-\vec{v}_2)(1-P_2).
\end{equation}
Substituting them into Eq.~(\ref{5}) and considering Eq.~(\ref{psymmetry}), we get
\begin{equation}\label{new_relation}
f(\vec{r}_1,\vec{v}_1)=f(\vec{r}_2,-\vec{v}_2),
\end{equation}
where $\vec{r}_1$, $\vec{r}_2$, $\vec{v}_1$ and $\vec{v}_2$ are related to each other as shown in Fig.~\ref{gas} and Eq.~(\ref{first}). This relation can be interpreted as a specific version of the principle of detailed balance for a dilute gas in a gravitational field with collisions taken into account.
\section{temperature and density}
The relation (\ref{new_relation}) determines the temperature distribution and the density distribution.
In one-dimensional situation, it becomes
\begin{equation} \label{zz}
f(z_1,v_{1z})=f(z_2,-v_{2z}).
\end{equation}
If the Maxwell distribution holds locally, the relation further becomes $f(z_1,v_{1z})=f(z_2,v_{2z})$, which is the result that Coombes and Laue \cite{cccc} have obtained in a different way without taking into account the molecular collisions.
With their relation, Coombes and Laue were able to explain intuitively why the temperature is uniform and why the barometric formula \cite{Berberan-Santos} holds.
We can do the same. If the Maxwell distribution holds, we can write
\begin{equation}\label{ff}
\begin{array}{l}
f(z_1,v_{1z})=n(z_1)\sqrt{\frac{m}{2\pi kT_1}}\exp\left(-\frac{mv_{1z}^2}{2kT_1}\right),\\
f(z_2,-v_{2z})=n(z_2)\sqrt{\frac{m}{2\pi kT_2}}\exp\left(-\frac{mv_{2z}^2}{2kT_2}\right),\\
\frac{1}{2}mv_{2z}^2+mgz_2=\frac{1}{2}mv_{1z}^2+mgz_1,
\end{array}
\end{equation}
where $n$ is the molecular number density and $T_1$ and $T_2$ are the temperatures at heights $z_1$ and $z_2$, respectively. Substituting them into Eq.~(\ref{zz}), we get
\begin{eqnarray}
&&\frac{n(z_1)}{n(z_2)}\exp\left[-\frac{mg(z_2-z_1)}{kT_2}\right] \nonumber\\
&&\times\sqrt{\frac{T_2}{T_1}}\exp\left[-\frac{mv_{1z}^2}{2k}\left(\frac{1}{T_1}-\frac{1}{T_2}\right)\right]=1.
\end{eqnarray}
This relation must hold for all $v_{1z}$. Here $z$ and $v_z$ are independent variables. So we have $T_1=T_2$. This means that the temperature $T$ is uniform and the barometric formula holds,
\begin{equation}
\frac{n(z_2)}{n(z_1)}=\exp\left[-\frac{mg(z_2-z_1)}{kT}\right].
\end{equation}
According to Maxwell distribution, the temperature is determined by the shape of the velocity distribution function. A molecule going up against the gravitational field indeed loses energy. But this does not change the shape of velocity distribution function. So the temperature does not change.
About the Maxwell distribution, it holds locally for an equilibrium state whether or not a gravitational field presents \cite{cccc}. But note that Rom\'{a}n, White and Velasco \cite{bbbb,aaaa} argued that the Maxwell distribution should hold only in the thermodynamic limit.
\begin{acknowledgments}
The authors are grateful to Shu-Hang Li and Jie Ren for discussions.
\end{acknowledgments}
|
\section{Introduction: The Ericksen-Leslie model with general Leslie stress}
In their pioneering articles, Ericksen \cite{Eri62} and Leslie \cite{Les68} developed a continuum theory for the flow of nematic liquid crystals. Their theory models nematic
liquid crystal flow from a hydrodynamical point of view and describes the evolution of the underlying system under the influence of the velocity $u$ of the fluid and the
orientation configuration $d$ of rod-like liquid crystals. We already observe here that the modulus $|d|_2$ of the director field $d$ must equal 1 pointwise,
as $d$ represents a direction field. For more information see e.g. \cite{Cha92,DGP95,Ste04,ASS04} or \cite{LW14b}, \cite{HP15b}.
The original derivation \cite{Eri62,Les68} is based on the conservation laws for mass, linear and angular momentums
as well as on constitutive relations given by Leslie in \cite{Les68}.
Following arguments from thermodynamics and employing the entropy principle, we proposed in \cite{HP15a}, \cite{HP15b} thermodynamically consistent models of Ericksen-Leslie type,
even in the case of compressible fluids. Let us emphasize that these models {\em contain the classical Ericksen-Leslie model in its general form as a special case}.
A related class of models also dealing with the non-isothermal situation was presented by Feireisl, Rocca and Schimperna \cite{FRS11} as well as by Feireisl, Fr\'emond, Rocca and Schimperna
in \cite{FFRS12}. Their models include stretching as well as rotational terms and are consistent with the fundamental laws of thermodynamics. The equation for the director $d$ is, however,
given in the penalized form, which does not seem to be physical. They show that the presence of the term $|\nabla d|_2^2$ in the internal energy as well as the stretching term
$d \cdot \nabla u$ give rise, in order to respect the laws of thermodynamics, to two new non dissipative contributions in the stress tensor $S$ and in the flux $q$. It is interesting
to note that these two new contributions coincide with the extra terms derived by Sun and Liu \cite{SL09} by different methods. It seems that these extra terms are non physical and arise there
for purely mathematical reasons.
Given a bounded domain $\Omega \subset \R^n$, $n \geq 2$, with smooth enough boundary, the general Ericksen-Leslie model in the non-isothermal situation derived as in
\cite{HP15a} and \cite{HP15b} reads as
\begin{align}\label{eq:elgeneral}
\left\{
\begin{array}{rll}
\partial_t \rho +{\rm div}(\rho u)&=0\quad &\mbox{in } \Omega,\\
\rho\cD_tu +\nabla \pi &= {\rm div}\, S \quad &\mbox{in } \Omega,\\
\rho\cD_t\epsilon +{\rm div}\, q &= S:\nabla u -\pi{\rm div}\, u+{\rm div}(\rho\partial_{\nabla d}\psi\cD_td)\quad &\mbox{in } \Omega,\\
\gamma \cD_td-\mu_V Vd&= P_d\big( {\rm div}(\rho\frac{\partial\psi}{\partial\nabla d})-\rho\nabla_d\psi\big)+ \mu_D P_d Dd \quad &\mbox{in } \Omega,\\
u=0,\quad q\cdot \nu&=0 \quad &\mbox{on } \partial\Omega,\\
\rho(0)=\rho_0,\quad u(0)=u_0,\quad \theta(0)&=\theta_0,\quad d(0)=d_0\quad &\mbox{in } \Omega.
\end{array}\right.
\end{align}
Here the unknown variables $\rho, u$ denote the density and velocity of the fluid, respectively, $\epsilon$ the
internal energy and $d$ the so called director, which - we recall - must have modulus 1. Moreover, $D = \frac{1}{2}(\nabla u]^T+ \nabla u)$ denotes the rate of deformation tensor, the vorticity tensor $V$ defined by
$V=\frac{1}{2}(\nabla u -[\nabla u]^T)$ is skew-symmetric, and $q$ denotes the heat flux. By $\cD_t=\partial_t +u\cdot\nabla$ we denote the Lagrangian derivative and $P_d$ is defined as $P_d=I-d\otimes d$.
Note that the condition $|d|_2=1$ is preserved by smooth solutions, as $P_d d=0$ as well as $(Vd|d)=0$, hence $\cD_t|d|^2_2=0$.
In addition, $\psi$ denotes the free energy, which, following Oseen \cite{Ose33} and Frank \cite{Fra58}, see also Virga \cite{Vir94}, is given by the Oseen-Frank functional $\psi^F(d,\nabla d)$
as
\begin{equation}\label{eq:psif}
\psi^F(d,\nabla d) := k_1(\dive d)^2 + k_2 (d \cdot \curl d)^2 + k_3|d \times \curl d|^2 + (k_2+k_4)(tr(\nabla d)^2- (\dive d)^2),
\end{equation}
where $k_1,\ldots,k_4$ are the so called {\em Frank coefficients}, which may depend on $\rho$ and $\theta$.
These equations have to be supplemented by the thermodynamical laws
\begin{align}\label{eq:thermo}
\epsilon = \psi +\theta \eta, \quad \eta = -\partial_\theta \psi, \quad \kappa = \partial_\theta \epsilon,\quad \pi =\rho^2 \partial_\rho\psi,
\end{align}
and by the constitutive laws
\begin{align}\label{eq:const}
\left\{
\begin{array}{lll}
S &= S_N + S_E + S_L^{stretch} + S_L^{diss}, \\
S_N&= 2\mu_s D + \mu_b {\rm div}\, u \, I, \\
S_E&= -\rho \frac{\partial \psi}{\partial\nabla d}[\nabla d]^{\sf T},\\
S_L^{stretch}& = \frac{\mu_D+\mu_V}{2\gamma}{\sf n} \otimes d + \frac{\mu_D-\mu_V}{2\gamma} d \otimes {\sf n},\quad
{\sf n}= \mu_V Vd +\mu_D P_dDd-\gamma\cD_td,\\
S_L^{diss}& =\frac{\mu_P}{\gamma} ({\sf n}\otimes d + d\otimes {\sf n}) + \frac{\gamma\mu_L+\mu_P^2}{2\gamma}(P_dDd\otimes d + d\otimes P_dDd) +\mu_0 (Dd|d)d\otimes d,\\
\end{array}\right
.\end{align}
and
\begin{equation}\label{eq:q1}
q = -\tilde\alpha_0 \nabla\theta -\tilde\alpha_1(d|\nabla\theta)d.
\end{equation}
Here all coefficients $\mu_j,\tilde\alpha_j$ and $\gamma$ are functions of $\rho,\theta,d,\nabla d$.
For thermodynamical consistency we require the conditions
\begin{align}\label{eq:pos1}
\mu_s\geq0,\quad 2\mu_s+n\mu_b\geq0,\quad \tilde\alpha_0\geq0,\quad \tilde\alpha_0+\tilde\alpha_1\geq0,\quad \mu_0,\mu_L\geq0,\quad \gamma>0.
\end{align}
We also note that the natural boundary condition at $\partial\Omega$ for $d$ becomes
\begin{equation}
\nu_i\nabla_{\partial_i d}\psi = 0\quad \mbox{ on }\partial\Omega.
\end{equation}
Observe that this condition is {\em fully nonlinear}, in general. Physically, it means that the boundary does not interact with the director field.
Otherwise one would have to model such interactions and it seems to be unclear whether this could be done in a physically consistent way by simply imposing Dirichlet boundary conditions.
For this reason we employ the {\em Neumann condition} for $d$ throughout this paper. Actually, we can prove local well-posedness also in the case of Dirichlet boundary conditions, but the
set of equilibria becomes more complicated in this case. Our results on stability and long time behaviour are only valid for constant equilibria.
In the case of {\em isotropic elasticy} with constant density and temperature one has $k_1=k_2=k_3=1$ and $k_4=0$ and so the Oseen-Frank energy reduces to the Dirichlet energy, i.e.
$$
\psi(d,\nabla d):= \psi^F(d,\nabla d) = \frac{1}{2}|\nabla d|^2,
$$
and thus
$
\dive ( \frac{\partial \psi^F}{\partial(\partial \nabla d)}) = \Delta d$.
Then the Ericksen stress tensor simplifies to
\begin{equation}\label{eq:ericksenstress}
S_E= -\lambda \nabla d[\nabla d]^{\sf T},
\end{equation}
where $\lambda = \rho\partial_\tau\psi$,
and the natural boundary condition at $\partial\Omega$ for $d$ becomes the Neumann condition $\partial_\nu d = 0$ on $\partial\Omega$.
It is the aim of this article to investigate the above Ericksen-Leslie system analytically, in the case of isotropic elasticity and for incompressible as well as compressible fluids.
To this end, Part I of this article will concentrate on the case of incompressible fluids, whereas Part II will investigate the compressible case.
Here {\em incompressibility} means that the density $\rho$ is constant and {\em isotropy} means that the free energy $\psi$ is a function of $\varrho$, $\theta$ and $\tau=|\nabla d|_2^2/2$, only.
For the convenience of the reader, we rewrite the model in the incompressible and isotropic case, which then reads as follows.
\begin{align}\label{eq:el}
\left\{
\begin{array}{rll}
\rho\cD_tu +\nabla \pi &= {\rm div}\, S \quad &\mbox{in } \Omega,\\
{\rm div}\, u&=0\quad &\mbox{in } \Omega,\\
\rho\cD_t\epsilon +{\rm div}\, q &= S:\nabla u+ {\rm div}(\lambda\nabla d\cD_t d)\quad &\mbox{in } \Omega,\\
\gamma\cD_td -\mu_V Vd- {\rm div}[ \lambda\nabla]d&=\lambda|\nabla d|^2d + \mu_D P_d Dd\quad &\mbox{in } \Omega,\\
u=0,\quad q\cdot \nu=0,\quad \partial_\nu d&=0 \quad &\mbox{on } \partial\Omega,\\
\rho(0)=\rho_0,\quad u(0)=u_0,\quad \theta(0)&=\theta_0,\quad d(0)=d_0\quad &\mbox{in } \Omega.
\end{array}\right.
\end{align}
These equations have to be supplemented by the thermodynamical laws for the internal energy $\epsilon$, entropy $\eta$, heat capacity $\kappa$
given in \eqref{eq:thermo}, by Ericksen's tension $\lambda =\rho\partial_\tau\lambda$, and by the constitutive laws \eqref{eq:const} as above, with Ericksen stress of the form \eqref{eq:ericksenstress}, and with $q$ satifying
\begin{equation}\label{eq:q2}
q = -\alpha \nabla\theta.
\end{equation}
Note that all coefficients $\mu_j,\alpha$ and $\gamma$ are functions of $\theta$ and $\tau$, by the principle of equi-presence.
For further purposes, it is convenient to write the equation for the internal energy as an equation for the temperature $\theta$. It reads as
$$
\rho\kappa\cD_t \theta +{\rm div}\, q = S:\nabla u + {\rm div}(\lambda \nabla)d\cdot \cD_t d + (\theta \partial_\theta\lambda) \nabla d \nabla\cD_t d.
$$
Observe the appearance of unusual third order terms due to the presence of $\cD_t d$ in the Leslie stress $S_L$ as well as in the last term of the energy balance. This alludes a peculiarity of the system, which has to be overcome in the analysis.
Let us emphasize that in the case $\mu_V=\gamma$, our parameters $\mu_s,\mu_0,\mu_V,\mu_D,\mu_P,\mu_L$ are in one-to-one correspondence to the celebrated
Leslie parameters $\alpha_1,\ldots,\alpha_6$ given in the Leslie stress $\sigma_L$ defined by
\begin{equation}\label{eq:leslie}
\sigma_L := \alpha_1 (d^TDd)d\otimes d +\alpha_2N \otimes d + \alpha_3 d\otimes N + \alpha_4 D + \alpha_5(Dd)\otimes d + \alpha_6 d \otimes (Dd),
\end{equation}
where $D$ denotes the deformation tensor as above and
$$
N:=N(u,d):=\partial_td + (u \cdot \nabla)d - Vd,
$$
with $V$ as above. This shows that our model \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const} {\em contains the classical isothermal and isotropic
Ericksen-Leslie model} given by
\begin{align}\label{eq:el-isotropic}
\left\{
\begin{array}{rlll}
u_t+ (u \cdot \nabla) u -\Delta u +\nabla \pi&\!=\!& -\dive(\nabla d[\nabla d]^T) + \dive \sigma_L & \text{in } (0,T) \times \Omega, \\
\dive u &\!=\!& 0 & \text{in } (0,T) \times \Omega, \\
d_t + u \nabla d - Vd + \frac{\lambda_2}{\lambda_1}Dd &\!=\!& -\frac{1}{\lambda_1}(\Delta d + |\nabla d|^2d) + \frac{\lambda_2}{\lambda_1}(d^TDd)d & \text{in } (0,T) \times \Omega,\\
(u,\partial_\nu d) &\!=\!& (0,0)& \text{on } (0,T) \times \partial \Omega, \\
(u,d)_{\vert t=0} &\!=\!& (u_0,d_0) & \text{in } \Omega. \end{array}\right.
\end{align}
{\em as a special case}; here we have $\lambda_1= -\gamma/\lambda$ and $\lambda_2 = \mu_D/\lambda$. Note that in the simplified model no third order terms appear, which considerably simplifies this problem.
It was shown by Parodi \cite{Par70} in 1970 that, {\em assuming Onsager's reciprocal relation}, one is lead to the relation
\begin{equation}\label{eq:parodi}
\alpha_2 + \alpha_3 = \alpha_6 - \alpha_5,
\end{equation}
where the coefficients $\alpha_j$ denote the Leslie coefficients introduced above.
The analysis of the Ericksen-Leslie system began by the pioneering work of Lin \cite{Lin89}, \cite{Lin91} and Lin and Liu \cite{LL95}, \cite{LL00}, who introduced and studied
the nowadays called {\em isothermal simplified model}. They studied the situation where the nonlinearity in the equation for $d$ is replaced by a Ginzburg-Landau energy functional
and proved the existence of global weak solutions under suitable assumptions on the intial data. Wang proved in \cite{Wan11} global well-posedness for the simplified system for
initial data being small in $BMO^{-1}\times BMO$ in the case of a whole space $\Omega=\R^n$ by combining techniques of Koch and Tataru with methods from
harmonic maps.
Concerning the situation of {\em bounded domains}, a rather complete understanding of the well-posedness as well as the dynamics of the simplified system subject to
Neumann conditions for $d$ was obtained in \cite{HNPS14}. First results on the existence of global weak solutions to the simplified system subject to Dirichlet boundary conditions
in two dimensions go back to F.Lin, J. Lin and Wang \cite{LLW10}. Recently, considering the simplified system in three dimensions subject to Dirichlet conditions $d=d_b$ on $\partial\Omega$,
Huang, Lin, Liu and Wang \cite{HLLW15} constructed examples of small initial data for which one has finite time blow up of $(u,d)$.
It seems that Coutard and Shkoller \cite{CS01} were the first who considered so called
{\em stretching terms} analytically in the equation for $d$. More precisely, they replaced the equation for $d$ in the {\em simplified model}
by a Ginzburg-Landau type approximation {\em including stretching} of the form
\begin{equation}\label{eq:introstretch}
\gamma(\partial_td + u \cdot \nabla d + d \cdot \nabla u)= \Delta d - \frac{1}{\varepsilon^2}(|d|_2^2 -1)d \quad \text{in } (0,T) \times \Omega.
\end{equation}
They proved local well-posedness for \eqref{eq:introstretch} as well as a global existence result for small data.
Note, however, that in this case the presence of the stretching term $d \cdot \nabla u$ causes loss of total energy balance and, moreover, the
condition $|d|_2= 1$ in $(0,T) \times \Omega$, is not preserved anymore.
For recent results on the general Ericksen-Leslie model with vanishing Leslie and general Ericksen stress we refer to the articles
\cite{HLX13} and \cite{MLG14}. For results including stretching terms for $d$, we refer to the articles
\cite{LLW10}, \cite{HLW13}, \cite{WZZ13}, \cite{WXL13}, and \cite{HP15a,HP15b}, which contain well-posedness criteria for the
general system under various assumptions on the Leslie coefficients.
The main new idea in investigating the general Ericksen-Leslie equations analytically in the strong sense is, similiarly to the situation of the simplified and isothermal
system treated in \cite{HNPS14}, to regard them as a quasilinear parabolic evolution equation.
Restricting ourselves in Part I to the case of {\em incompressible fluids} and to the case of isotropic elasticity,
we present in the following a rather complete dynamic theory for the underlying equations. It seems to be the first well-posedness result for the {\em Ericksen-Leslie equations
dealing with general Leslie stress $S_L$ without assuming additional conditions on the Leslie coefficients}.
The results given in the three main theorems below answer all questions concerning well-posedness, stability and longtime behaviour for the {\em general} Ericksen-Leslie system subject
to Neumann boundary conditions for $d$ in a very satisfactory way.
It is proved by means of techniques involving maximal $L_p$-regularity and quasilinear parabolic evolution equations. For these methods, we refer to the booklet by
Denk, Hieber, and Pr\"{u}ss \cite{DHP03}, the articles \cite{Pru03}, \cite{PSZ09}, \cite{KPW10}, \cite{LPW14} and to the monograph by Pr\"uss and Simonett \cite{PS16}. For the convenience of the reader we have summarized the relevant results from these papers in Section3.
Let us also emphasize that for obtaining our well-posedeness results in the strong sense {\em no structural conditions} on the Leslie coefficients are imposed and that in
particular {\em Parodi's relation} \eqref{eq:parodi} on the Leslie coeffcients is {\em not} being assumed.
Moreover, the equilibria of the system have been identified in our recent paper [14] - which are zero velocities and constant density, temperature and director - and there it also has been proved that
these are thermodynamically stable. The negative total entropy has been shown to be a strict Lyapunov functional; in particular, the model is thermodynamically consistent.
We further mention at this point the series of articles \cite{FRS11},\cite{FFRS12},\cite{FSRZ12},\cite{FSRZ13}, in which it is shown that their particular
systems admit a global weak solution for a natural class of initial data.
For more information on modeling and analysis of the Ericksen-Leslie system, we refer e.g. to \cite{Cha92}, \cite{DGP95} and to the survey articles \cite{HP15b} and \cite{LW14b}.
\section{Thermodynamical Stability and Consistency}
In this short section we recall from \cite{HP15a} and \cite{HP15b} that the above model \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const}, \eqref{eq:q2} has the following thermodynamical
properties. To this end, we introduce the following \\
{\bf Assumption (P)}: \\
\begin{align}\label{eq:pos2}
\mu_s>0,\quad \alpha>0,\quad \mu_0,\mu_L\geq0,\quad \kappa,\gamma>0,\quad \lambda, \lambda +2\tau \partial_\tau\lambda>0.
\end{align}
\begin{theorem}(\cite{HP15a}, Theorem 1). \label{prop:equi}
Assume that condition (P) holds. Then the incompressible and isotropic model \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const}, \eqref{eq:q2} has the following properties:
\begin{itemize}
\item[i)] Along smooth solutions the total energy
$$
{\sf E}:= \int_\Omega [\frac{\rho}{2}|u|_2^2 +\rho\epsilon]dx
$$
is preserved.
\item[ii)] Along smooth solutions the total entropy
$${\sf N}:= \int_\Omega \rho\eta dx$$
is non-decreasing.
\item[iii)] The negative total entropy $-{\sf N}$ is a strict Lyapunov functional.
\item[iv)] The condition $|d|_2=1$ is preserved along smooth solutions.
\item[v)] The equilibria of the system are given by the set of constants
\begin{equation}\label{eq:equilibria}
\cE=\{(0,\theta_*,d_*):\, \theta_*\in(0,\infty), \; d_*\in\RR^n,\; |d_*|_2=1\}.
\end{equation}
Here $\theta_*$ is uniquely determined by the identity
$$ \epsilon(\theta_*,0) = {\sf E}_0/\rho|\Omega|.$$
\item[vi)] The equilibria are the critical points of the total entropy with prescribed energy.
\item[vii)] The second variation of ${\sf N}$ with prescribed energy at an equilibrium is negative semi-definite.
\end{itemize}
\end{theorem}
\section{Background: Quasilinear Parabolic Evolution Equations}
\noindent
In this section we briefly recall some results on abstract quasilinear parabolic problems
\begin{equation}\label{Int1}
\dot{v}+A(v)v=F(v),\quad t>0,\quad v(0)=v_0,
\end{equation}
which are employed in the proofs of our main theorems. These results are due to Pr\"uss \cite{Pru03}, Pr\"uss and Simonett \cite{PS04}, K\"ohne, Pr\"uss and Wilke \cite{KPW10}, and
Pr\"uss, Simonett and Zacher \cite{PSZ09}; a convenient reference for this theory is the monograph by Pr\"uss and Simonett \cite{PS16}, Chapter 5.
Assume that $(A,F):V_\mu\to \cB(X_1,X_0)\times X_0$ and $v_0\in V_\mu$. Here the spaces $X_1,X_0$ are Banach spaces such that $X_1\hookrightarrow X_0$ with dense embedding and
$V_\mu$ is an open subset of the real interpolation space
$$X_{\gamma,\mu}:=(X_0,X_1)_{\mu-1/p,p},\quad \mu\in(1/p,1].$$
We are mainly interested in solutions $v$ of \eqref{Int1} having \emph{maximal} $L_{p}$-\emph{regularity}, i.e.
$$v\in H_{p}^1(J;X_0)\cap L_{p}(J;X_1)=:\EE_1(J), \mbox{ where } J = (0,T).$$
The trace space of this class of functions is given by $X_\gamma:=X_{\gamma,1}$.
However, to see and exploit the effect of parabolic regularization in the $L_p$-framework it is also useful to consider solutions in the class of weighted spaces
$$v\in H_{p,\mu}^1(J;X_0)\cap L_{p,\mu}(J;X_1)=:\EE_{1,\mu}(J),\quad \mbox{which means } \; t^{1-\mu}v\in \EE_1(J). $$
The trace space for this class of weighted spaces is given by $X_{\gamma,\mu}$.
In our approach it is crucial to know that the operators $A(v)$ have the property of {\em maximal $L_{p}$-regularity}. Recall that an operator
$A_0$ in $X_0$ with domain $X_1$ has maximal $L_p$-regularity, if the linear problem
$$ \dot{v} +A_0v =f, \quad t\in J,\; v(0)=0,$$
admits a unique solution $v\in \EE_1(J)$, for any given $f\in L_p(J;X_0)=:\EE_0(J)$. It has been proved in \cite{PS04} that in this case maximal regularity also holds in the weighted case.
\begin{proposition}\label{LWPthm}
Let $p\in (1,\infty)$, $v_0\in V_\mu$ be given and suppose that $(A,F)$ satisfies
\begin{equation}\label{LWP2}
(A,F)\in C^{1}(V_\mu;\cB(X_1,X_0)\times X_0),
\end{equation}
for some $\mu\in (1/p,1]$. Assume in addition that $A(v_0)$ has maximal $L_p$-regularity.
Then there exist $a=a(v_0)>0$ and $r=r(v_0)>0$ with $\bar{B}_{X_{\gamma,\mu}}(v_0,r)\subset V_\mu$ such that problem \eqref{Int1} has a unique solution
$$v=v(\cdot,v_1)\in \EE_{1,\mu}(0,a)\cap C([0,T];V_\mu),$$
on $[0,a]$, for any initial value $v_1\in \bar{B}_{X_{\gamma,\mu}}(v_0,r)$. In addition,
$$ t\partial_t v \in \EE_{1,\mu}(0,a),$$
in particular, for each $\delta\in(0,a)$ we have
$$v\in H^2_p((\delta,a);X_0)\cap H^1_p((\delta,a);X_1)\hookrightarrow C^1([\delta,a];X_\gamma)\cap C^{1-1/p}([\delta,a];X_1),$$
i.e.\ the solution regularizes instantly.
\end{proposition}
\noindent
The next result provides information about the continuation of local solutions.
\begin{corollary}\label{LWPcor1}
Let the assumptions of Theorem \ref{LWPthm} be satisfied and assume that $A(v)$ has maximal $L_p$-regularity for all $v\in V_\mu$. Then the solution $v$ of \eqref{Int1} has a maximal
interval of existence $J(v_0)=[0,t_+(v_0))$, which is characterized by the following alternatives:
\smallskip\\
{ (i)} Global existence: $t_+(v_0)=\infty$; \smallskip\\
{(ii)} $\liminf_{t\to t_+(v_0)} {\rm dist}_{X_{\gamma,\mu}}(v(t),\partial V_\mu)=0$; \smallskip\\
{(iii)} $\lim_{t\to t_+(v_0)} v(t)$ does not exist in $X_{\gamma,\mu}$.
\end{corollary}
\noindent
Next we assume that there is an open
set $V\subset X_\gamma$ such that
\begin{equation}
\label{AF}
(A,F)\in C^1(V,\cB(X_1,X_0)\times X_0).
\end{equation}
Let $ \cE\subset V\cap X_1$ denote the set of equilibrium solutions of (\ref{Int1}), which means that
$$
v\in\cE \quad \mbox{ if and only if }\quad v\in V\cap X_1 \mbox{ and } A(v)v=F(v).
$$
Given an element $v_*\in\cE$, we assume that $v_*$ is
contained in an $m$-dimensional {\em manifold of equilibria}. This means that there
is an open subset $U\subset\R^m$, $0\in U$, and a $C^1$-function $\Psi:U\rightarrow X_1$, such that
\begin{equation}
\label{manifold}
\begin{aligned}
& \bullet\
\text{$\Psi(U)\subset \cE$ and $\Psi(0)=v_*$,} \\
& \bullet\
\text{the rank of $\Psi^\prime(0)$ equals $m$, and} \\
& \bullet\
\text{$A(\Psi(\zeta))\Psi(\zeta)=F(\Psi(\zeta)),\quad \zeta\in U.$}
\end{aligned}
\end{equation}
We suppose that
the operator $A(v_*)$ has the property of maximal $L_p$-regularity, and define the full linearization of \eqref{Int1} at $v_*$ by
\begin{equation}
\label{A0}
A_0w=A(v_*)w+(A^\prime(u_*)w)v_*-F^\prime(v_*)w\quad\hbox{for }
w\in X_1.
\end{equation}
\noindent
After these preparations we can state the following result on convergence of solutions starting near $v_*$ which is called the {\em generalized principle of linearized stability}.
\begin{proposition}\label{th:1}
Let $1<p<\infty$. Suppose $v_*\in V\cap X_1$ is an
equilibrium of (\ref{Int1}), and suppose that the functions
$(A,F)$ satisfy \eqref{AF}. Suppose further that $A(v_*)$ has the
property of maximal $L_p$-regularity and let $A_0$ be defined in
\eqref{A0}.
Suppose that $v_*$ is {\rm normally stable},
which means
\begin{description}
\item[{\rm i)}] near $v_*$ the set of equilibria $\cE$ is a $C^1$-manifold in $X_1$ of dimension $m\in\N$,
\vspace{-1mm}
\item[{\rm ii)}] \, the tangent space for $\cE$ at $v_*$ is isomorphic to ${\sf N}(A_0)$,
\vspace{-1mm}
\item[{\rm iii)}] \, $0$ is a {\em semi-simple eigenvalue} of
$A_0$, i.e.\ $ {\sf N}(A_0)\oplus {\sf R}(A_0)=X_0$,
\vspace{-1mm}
\item[{\rm iv)}] \, $\sigma(A_0)\setminus\{0\}\subset \C_+=\{z\in\C:\, {\rm Re}\, z>0\}$.
\end{description}
Then $v_*$ is stable in $X_\gamma$, and there exists $\delta>0$ such
that the unique solution $v$ of (\ref{Int1}) with initial
value $v_0\in X_\gamma$ satisfying $|v_0-v_*|_{\gamma}\leq\delta$
exists on $\R_+$ and converges at an exponential rate in $X_\gamma$
to some $v_\infty\in\cE$ as $t\rightarrow\infty$.
\end{proposition}
\noindent
The next result contains information on bounded solutions in the presence of compact embeddings and of a strict Lyapunov functional.
\begin{proposition}\label{LTBthm}
Let $p\in (1,\infty)$, $\mu\in (1/p,1)$, $\bar{\mu}\in (\mu,1]$, with $V_\mu\subset X_{\gamma,\mu}$ open.
Assume that $(A,F)\in C^1(V_\mu;\cB(X_1,X_0)\times X_0)$, and that the embedding
$X_{\gamma,\bar{\mu}}\hookrightarrow X_{\gamma,\mu}$ is compact. Suppose furthermore that $v$ is a maximal solution which is bounded in $X_{\gamma,\bar{\mu}}$ and satisfies
\begin{equation}\label{GDC}
{\rm dist}_{X_{\gamma,\mu}}(v(t),\partial V_{\mu})\ge\eta>0,\; \mbox{ for all }t\geq 0.
\end{equation}
Suppose that $\Phi\in C(V_{\mu}\cap X_\gamma;\R)$ is a strict Lyapunov functional for \eqref{Int1}, which means that $\Phi$ is strictly decreasing along non-constant solutions.
Then $t_+(v_0)=\infty$, i.e.\ $v$ is a global solution of \eqref{Int1}. Its $\omega$-limit set $\omega_+(v_0)\subset \cE$ in $X_\gamma$ is nonempty, compact and connected.
If, in addition, there exists $v_*\in\omega_+(v_0)$ which is normally stable, then $\lim_{t\to\infty}v(t)=v_*$ in $X_\gamma$.
\end{proposition}
\section{Main Results}
In order to formulate the main well-posedness result for the Ericksen-Leslie system in the incompressible and isotropic case and to have access to the tools presented in Section 3,
we introduce a functional analytic setting as follows. Denote the principal variable by $v=(u,\theta,d)$ and
let us rewrite system \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const}, \eqref{eq:q2} as a quasi-linear evolution equation of the form
\begin{equation} \label{eq:QLP}
\dot{v} +A(v) v = F(v),\quad t>0,\; v(0)=v_0,
\end{equation}
replacing $\cD_td$ appearing in the the equations for $u$ and $\theta$ by the equation for $d$. We also apply the {\em Helmholtz projection} $\PP$ to the equation for $u$.
Then $v$ belongs to the base space $X_0$ defined by
$$
X_0 :=L_{q,\sigma}(\Omega)\times L_q(\Omega;\R)\times H^1_q(\Omega;\R^n),
$$
where $1<p,q<\infty$ and $\sigma$ indicates solenoidal vector fields.
The regularity space will be
$$
X_1 :=\{ u\in H^2_q(\Omega;\RR^n)\cap L_{q,\sigma}(\Omega):\, u=0 \mbox{ {on} } \partial\Omega\}\times Y_1,
$$
with
$$
Y_1 := \{ (\theta,d)\in H^2_q(\Omega)\times H^3_q(\Omega;\RR^n):\, \partial_\nu \theta=\partial_\nu d =0 \mbox{ on } \partial\Omega\}.
$$
We consider solutions $v$ within the class
$$
v\in H^1_{p,\mu}(J;X_0)\cap L_{p,\mu}(J;X_1)=\EE_{1,\mu}(J),
$$
where $J=(0,a)$ with $0<a\leq\infty$ is an interval and $\mu\in (1/p,1]$.
The time trace space of this class is given by
\begin{equation}\label{eq:trace}
X_{\gamma,\mu} := \{ u\in B_{qp}^{2(\mu-1/p)}(\Omega)^n\cap L_{q,\sigma}(\Omega):\, u=0 \mbox{ on } \partial\Omega\}\times Y_{\gamma,\mu},
\end{equation}
where
$$
Y_{\gamma,\mu} = \{ (\theta,d)\in B^{2(\mu-1/p)}_{qp}(\Omega)\times B^{1+2(\mu-1/p)}_{qp}(\Omega;\R^n):\, \partial_\nu\theta=\partial_\nu d=0 \mbox{ on } \partial\Omega\},$$
whenever the boundary traces exist.
Note that
$$
X_{\gamma,\mu}\hookrightarrow B_{qp}^{2(\mu-1/p)}(\Omega)^{n+1}\times B_{qp}^{1+2(\mu-1/p)}(\Omega)^n\hookrightarrow C(\overline{\Omega})^{n+1}\times C^1(\overline{\Omega})^n,
$$
provided
\begin{equation}\label{eq:pq}
\frac{1}{p} +\frac{n}{2q}<\mu\leq 1.
\end{equation}
For brevity we set $X_\gamma:=X_{\gamma,1}$. Finally, the state manifold of the problem is defined by
$$
\cSM = \{ v\in X_{\gamma}:\, \theta(x)>0,\, |d(x)|_2 =1 \mbox{ in } \Omega\}.
$$
We assume the following regularity assumptions on the parameter functions:
\vspace{.2cm}\noindent
{\bf Regularity assumption (R):} \\
The parameter functions are assumed to satisfy
\begin{equation}\label{eq:reg}
\mu_j, \alpha, \gamma\in C^2((0,\infty)\times[0,\infty)) \mbox{ for } j=S,V,D,P,L,0, \mbox{ and } \psi\in C^4((0,\infty)\times [0,\infty)).
\end{equation}
\vspace{.2cm}
The fundamental well-posedness results regarding the general isotropic incompressible Ericksen-Leslie system reads as follows.
\medskip
\begin{theorem}{\rm (Local Well-Posedness).} \label{thm:1} \\
Let $\Omega \subset \R^n$ be a bounded domain with boundary of class $C^{3-}$ and assume conditions (R) and (P).
Assume that $J=(0,a)$, $1<p,q,<\infty$ and $\mu \in (1/p,1]$ are subject to \eqref{eq:pq} and $v_0 \in X_{\gamma,\mu}$. Then for some $a=a(v_0)>0$, there is a unique solution
$$
v\in H^1_{p,\mu}(J,X_0)\cap L_{p,\mu}(J;X_1),
$$
of \eqref{eq:QLP}, i.e. of \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const}, \eqref{eq:q2} on $J$. Moreover,
$$
v\in C([0,a];X_{\gamma,\mu})\cap C((0,a];X_\gamma),
$$
i.e.\ the solution regularizes instantly in time. It depends continuously on $v_0$ and exists on a maximal time interval $J(v_0) = [0,t^+(v_0))$.
Moreover,
\begin{align*}
t\partial_t v \in H^1_{p,\mu}(J;X_0)\cap L_{p,\mu}(J;X_1),\quad a<t^+(v_0).
\end{align*}
Furthermore,
$$
|d(\cdot,\cdot)|_2 \equiv 1, \quad {\sf E}(t)\equiv {\sf E}_0, \quad t \in J,
$$
and $-N$ is a strict Lyapunov functional.
In addition, problem \eqref{eq:QLP} generates a local semi-flow in its natural state manifold $\cSM$.
\end{theorem}
Below, we denote by $\bar{\cE}$ the set
$$\bar{\cE} =\{ (0,\theta_*,d_*): \theta_*>0,\, d_*\in\RR^n\}$$
of constant equilibria of the system when ignoring the constraints $|d|_2=1$ and ${\sf E}={\sf E}_0$. The next result concerns stability of equilibria.
\begin{theorem}{\rm (Stability of Equilibria).} \label{thm:2}\\
Assume conditions (R) and (P). Then any equilibrium $v_*\in \overline{\cE}$ of \eqref{eq:QLP} is stable in $X_\gamma$. Moreover, for each $v_*\in \overline{\cE}$
there is $\delta > 0$ such that if $|v_0 -v_*|_{X_{\gamma,\mu}} \leq \delta$, then the solution $v$ of \eqref{eq:QLP} with initial value $v_0$ exists
globally in time and converges at an
exponential rate in $X_\gamma$ to some $v_\infty\in \overline{\cE}$.
\end{theorem}
\noindent
The third result concerns global existence and convergence of solutions to equilibria in the topology of the state manifold $\cSM$.
\begin{theorem}{\rm (Long-Time Behaviour).}\label{thm:3}\\
Assume conditions (R) and (P) and let $v$ be the solution of equation \eqref{eq:QLP} with $v_0 \in \cSM$. Then the following assertions hold. \\
{\bf a)} Suppose that for some $\bar{\mu}\in ( 1/p+n/2q+1/2,1]$ we have
\begin{equation}\label{eq:longtime}
\sup_{t \in (0,t^+(v_0))} [|v(t)|_{X_{\gamma,\mu}} + |1/\theta(t)|_{L_\infty}]<\infty.
\end{equation}
Then $t^+(v_0)=\infty$ and $v$ is a global solution.\\
{\bf b)} If $v$ is a global solution, bounded in $X_{\gamma,\bar{\mu}}$ and with $1/\theta$ bounded, then $v$ converges exponentially in $\cSM$ to an
equilibrium $v_\infty\in \mathcal E$ of \eqref{eq:QLP} as $t\to \infty$.\\
{\bf c)} If $v$ is global solution of \eqref{eq:QLP} which converges to an equilibrium in $\cSM$, then \eqref{eq:longtime} valid.
\end{theorem}
\begin{remark}
Let us emphasize that the above theorem holds true {\em without any structural assumptions} besides condition {(P)} on the Leslie coefficients. In particular, the above
well-posedness results hold true {\em without} assuming Parodi's relation \eqref{eq:parodi}.
\end{remark}
\begin{remarks}
{a)} Wu, Xu and Liu considered in \cite{WXL13} the {\em isothermal penalized} Ericksen-Leslie model and gave a formal physical derivation of the Ericksen-Leslie model based on an
energy variational approach assuming Parodi's relation. Then they prove that, under certain assumptions on the data and the Leslie coefficients, the isothermal penalized Ericksen-Leslie
system admits a unique, global solution provided the viscosity is large enough and study as well it longtime behaviour. Moreover, {\em assuming Parodi's relation}, but not largeness of the
viscosity, they show global well-posedness and Lypunov stability for the {\em penalized} Ericksen-Leslie system near local energy minimizers. \\
{b)} Wang, P. Zhang and Z. Zhang \cite{WZZ13} proved local well-posedness of the {\em isothermal} general Ericksen-Leslie system as well as global well-posedness for small initial data
under various conditions on the Leslie coefficients, which ensure that the energy of the system is dissipated.\\
{c)} It is interesting to compare our above results with a recent result due to Huang, Lin, Liu and Wang \cite{HLLW15}, where they considered the simplified system subject to
Dirichlet boundary conditions $d=d_b$ on $\partial\Omega$ and where they constructed examples of small initial data for which one has finite time blow up of $(u,d)$ for the solution
of the simplified system.
\end{remarks}
\section{Maximal $L^p$-Regularity of the Linearization}
The main task to apply the results in Section 3 is to establish maximal $L_p$-regularity of the linearized problem. To prove this,
we linearize equation \eqref{eq:el} at an initial value $v_0=(u_0,\theta_0,d_0)$ and drop all terms of lower order. This yields the principal linearization
\begin{align}\label{eq:prlin}
\left\{
\begin{array}{rll}
\cL_\pi(\partial_t,\nabla)v_\pi &= f & \mbox{ in } J\times \Omega,\\
u=\partial_\nu\theta=\partial_\nu d &= 0 & \mbox{ on } J \times \partial\Omega,\\
u=\theta = d &= 0 & \mbox{ on } \{0\}\times\Omega.
\end{array}\right.
\end{align}
Here $J=(0,a)$, $v_\pi=(u,\pi,\theta,d)$ is the unknown, and $f=(f_u,f_\pi,f_\theta,f_d)$ are the given data. Denote the spatial co-variable by $\xi$ and by
$z$ that in time. Assume that $z \in \Sigma_\phi:=\{z \in \C\backslash\{0\}: 0 \leq |arg z| < \pi\}$. Then
the differential operator $L=\cL_\pi(\partial_t,\nabla)$ is defined via its symbol $\cL_\pi(z,i\xi)$, which is
\begin{equation}\label{princ-symb-pi}
\cL_\pi(z,i\xi) :=
\left[ \begin{array}{cccc}
M_u(z,\xi) & i\xi & 0& izR_1(\xi)^{\sf T}\\
i\xi^{\sf T}&0&0&0\\
0&0&m_\theta(z,\xi)& -iz \theta_0b a(\xi)\\
-iR_0(\xi)&0&-iba(\xi)& M_d(z,\xi)
\end{array}\right], \quad \xi \in \R^n, z \in \Sigma_\phi,
\end{equation}
with $b =\partial_\theta \lambda$, and $\lambda_1=\partial_\tau\lambda$. We also introduce the parabolic part of this symbol by dropping pressure gradient and divergence, i.e.
\begin{equation}\label{princ-symb}
\cL(z,i\xi) = \left[ \begin{array}{ccc}
M_u(z,\xi) & 0& izR_1(\xi)^{\sf T}\\
0&m_\theta(z,\xi)& iz \theta_0b a(\xi)\\
-iR_0(\xi)&ib a(\xi)& M_d(z,\xi)
\end{array}\right], \quad \xi \in \R^n, z \in \Sigma_\phi.
\end{equation}
The entries of these matrices are given by
\begin{align*}
m_\theta(z,\xi) &:= \rho\kappa z +\alpha|\xi|^2,\\
a(\xi) &:=\xi\cdot\nabla d_0, \\
M_d(z,\xi) &:= \gamma z + \lambda|\xi|^2 +\lambda_1 a(\xi)\otimes a(\xi)= m_d(z,\xi) +\lambda_1 a(\xi)\otimes a(\xi),\\
R_0(\xi)&:= \frac{\mu_D +\mu_V}{2} P_0\xi\otimes d_0 + \frac{\mu_D -\mu_V}{2}(\xi| d_0) P_0,\\
R_1(\xi) &:= (\frac{\mu_D +\mu_V}{2} +\mu_P)P_0\xi\otimes d_0 + (\frac{\mu_D -\mu_V}{2}+\mu_p)(\xi| d_0) P_0,\\
M_u(z,\xi) &:= \rho z + \mu_s|\xi|^2 + \mu_0(\xi|d_0)^2d_0\otimes d_0 +a_1(\xi|d_0)P_0\xi\otimes d_0 \\
& \quad + a_2 (\xi|d_0)^2 P_0 + a_3 |P_0\xi|^2 d_0\otimes d_0 + a_4 (\xi|d_0) d_0\otimes P_0\xi.
\end{align*}
Here $P_0=P_{d_0}=I-d_0\otimes d_0$, and $a_j$ are certain coefficients. Note that the above coefficients depend on $x$ through the dependence of the parameter functions on the
initial value $v_0(x)$. The maximal regularity result for \eqref{eq:prlin} employed below reads as follows.
\begin{theorem}\label{thm:mr}
Let $J=(0,a)$, $1<p,q<\infty$, and assume condition (R) and (P). Then equation \eqref{eq:prlin} admits a unique solution $v_\pi=(u,\pi,\theta,d)$ satisfying
\begin{align*}
(u,\theta)&\in {_0H}^1_p(J;L_q(\Omega))^{n+1} \cap L_p(J; H^2_q(\Omega))^{n+1},\\
\pi&\in L_p(J; \dot{H}^1_q(\Omega)),\\
d&\in {_0H}^1_p(J;H^1_q(\Omega))^{n} \cap L_p(J; H^3_q(\Omega))^{n},
\end{align*}
if and only if
\begin{align*}
\begin{array}{rl}
(f_u,f_\theta) &\in L_p(J;L_q(\Omega))^{n+1},\\
f_d &\in L_p(J;H^1_q(\Omega))^n, \\
f_\pi &\in {_0H}^1_p(J;H^{-1}_q(\Omega))\cap L_p(J;H^1_q(\Omega)).\\
\end{array}
\end{align*}
Further, the solution map $f\mapsto v_\pi$ is continuous between the corresponding spaces.
\end{theorem}
Let us remark that if we replace $\partial_t $ by $\partial_t+\omega$, where $\omega>0$ is a sufficiently large constant, then the assertion of Theorem \ref{thm:mr} holds true also for
$J=(0,\infty)$.
\vspace{.2cm}\noindent
{\em Proof}. We subdivide the proof into 5 steps.\\
{\em Step 1: The Principal Symbol with Constant Coefficients in $\Omega=\RR^n$.}\\
To extract the structure of $\cL$, we introduce the symbols
$$
R(\xi):= (\xi|d_0) P_0 + P_0\xi\otimes d_0,\quad R_\mu(\xi):= \mu_-(\xi|d_0) P_0 + \mu_+P_0\xi\otimes d_0,\quad \mu_\pm := \mu_D \pm \mu_V+\mu_P.
$$
Then $M_u$ simplifies to
$$
M_u= m_u + \mu_0(\xi|d_0)^2d_0\otimes d_0 + \frac{\mu_L}{4} R^{\sf T}R + \frac{1}{4\gamma} R_\mu^{\sf T}R_\mu + \frac{\mu_P\mu_V}{2\gamma}(\xi|d_0)(R-R^{\sf T}),
$$
and we also have
$$
R_1 = R_\mu -R_0 \mbox{ and } m_u(z,\xi) =\rho z +\mu_s|\xi|^2.
$$
Next, we set $v=(u,w)$, $v_\pi=(v,\pi,w)$ and $w=(\theta,d)$. Then, setting
$$
J={\rm diag}(I,1/\theta_0,zI),
$$
and the second line of $\cL$ by $1/\theta_0$ as well as the last line with $\bar{z}$, we obtain the estimate
\begin{align*}
{\rm Re} (\cL v| Jv)& = {\rm Re}\, m_u |u|_2^2 + {\rm Re}\, m_\theta|\theta|^2 + {\rm Re}\, z(\lambda_0|\xi|^2|d|_2^2 + \lambda_1|(a(\xi)|d)|^2)\\
& \quad + \frac{\mu_L}{4} |Ru|_2^2 + \frac{1}{4\gamma} |R_\mu u|_2^2 + {\rm Re}[ i z (d|R_\mu u) + \gamma|z|^2|d|_2^2\\
&\geq c [ {\rm Re}\, z( |u|_2^2 +|\theta|^2 + |\xi|^2|d|_2^2) + |\xi|^2(|u|_2^2+|\theta|^2) + (2\gamma|z||d|_2-|R_\mu u|_2)^2],
\end{align*}
provided
$$
\rho,\mu_s,\kappa,\gamma, \alpha, \lambda,\lambda +2\tau\partial_\tau\lambda>0 \mbox{ and } \mu_0,\mu_L\geq0.
$$
One could even relax the assumptions on $\mu_0$ and $\mu_L$ to $2\mu_s+\mu_0>0$ and $2\mu_s +\mu_L>0$, but we will not do this here.
This means that the symbol $\bar{J}\cL$ is accretive for ${\rm Re}\,z>0$, i.e. it is strongly elliptic.
Let us emphasize that we do not need any structural conditions on the coefficients $\mu_D,\mu_V,\mu_P, \partial_\theta\lambda$.
\vspace{.2cm}\noindent
{\em Step 2: Schur Reductions}. \\
In this step we perform a Schur reduction to reduce the above symbol to a symbol only for $u$. To this end, we consider the subsystem for $w$, i.e.\ the equation
\begin{align*}
\left[\begin{array}{cc}
m_\theta(z,\xi) & -iz \theta_0b a(\xi)^{\sf T}\\
-ib a(\xi)& m_d(z,\xi) + \lambda_1 a(\xi)\otimes a(\xi)
\end{array}\right]
\left[\begin{array}{c}
\theta\\ d
\end{array}\right]= \left[ \begin{array}{c}
f_\theta\\ f_d + i R_0(\xi)u
\end{array}\right].
\end{align*}
To solve this system, we follow the strategy developed in \cite{HP15a} and introduce the new variable $\delta = (a(\xi)|d)$. Then, multiplying the second equation with
$a(\xi)$ we obtain the system
\begin{align*}
\left[\begin{array}{cc}
m_\theta(z,\xi) & -iz \theta_0b \\
-ib |a(\xi)|^2& m_d(z,\xi) + \lambda_1| a(\xi)|^2
\end{array}\right]
\left[\begin{array}{c}
\theta\\ \delta
\end{array}\right]= \left[ \begin{array}{c}
f_\theta\\ (f_d|a(\xi)) + i (R_0(\xi)u|a(\xi))
\end{array}\right].
\end{align*}
This system is easily solved to the result
\begin{align*}
\left[ \begin{array}{c} \theta\\
\delta\end{array}\right] =
\frac{1}{det(z,\xi)} \left[\begin{array}{cc}
m_d(z,\xi) + \lambda_1| a(\xi)|^2& iz \theta_0b \\
ib |a(\xi)|^2& m_\theta(z,\xi)
\end{array}\right]
\left[ \begin{array}{c}
f_\theta\\ (f_d + i R_0(\xi)u|a(\xi))
\end{array}\right],
\end{align*}
where
$$
det(z,\xi) = m_\theta(z,\xi)(m_d(z,\xi)+\lambda_1 |a(\xi)|^2) + z\theta_0b^2|a(\xi)|^2.
$$
Note that this symbol behaves like $(z+|\xi|^2)^2$ as soon as $\rho,\kappa,\lambda, \lambda+2\tau\partial_\tau\lambda>0$.
Knowing $\delta=(a(\xi)|d)$ and $\theta$, we are now able to determine $d$. As a result we obtain
$$
d = m_d^{-1} [ f_d +iR_0u + i b a(\xi)\theta - \lambda_1 a(\xi)\delta].
$$
Following the arguments given in \cite{HP15a}, we see that $\theta$ and $d$ belong to the right regularity classes,
whenever $f_\theta,f_d$ and $u$ are so.
In order to extract the Schur complement for $u$, we set $f_\theta=f_d=0$ and compute $d$. This yields
\begin{equation}\label{eq:schurc}
d= i[\frac{1}{m_d}(I-a_0\otimes a_0)+ \frac{m_\theta}{det}a_0\otimes a_0]R_0 u=i[\frac{1}{m_d}P_{a_0}+ \frac{m_\theta}{det}Q_{a_0}]R_0 u.
\end{equation}
with $a_0(\xi)=a(\xi)/|a(\xi)|$ if $a(\xi)\neq0$ and $a_0(\xi)=0$ otherwise. This is the representation of $d$ needed for the Schur complement of $u$.
\vspace{.2cm}\noindent
{\em Step 3: The Generalized Stokes Symbol}.\\
We insert \eqref{eq:schurc} into the equation for $u$ to obtain the generalized Stokes symbol for $(u,\pi)$ and obtain
\begin{align}\label{eq:genstokes}
M(z,i\xi) &= M_u(z,\xi) - zR_1^{\sf T}(\xi)[\frac{1}{m_d(z,\xi)}P_{a_0}(\xi)+ \frac{m_\theta(z,\xi)}{det(z,\xi)}Q_a(\xi)]R_0(\xi)
\end{align}
As the Schur reduction preserves accretivity, even with the same accretivity constant, we obtain
$$
{\rm Re}(M(z,i\xi)u|u)\geq {\rm Re}\, m_u(z,\xi)|u|^2 =\big(\rho {\rm Re}\, z +\mu_s |\xi|^2\big)|u|^2.
$$
This shows that $M$ is strongly elliptic. For this reason we may now apply the method developed by Bothe and Pr\"uss \cite{BP07} or Pr{\"u}ss and Simonett \cite{PS16}, Section 7.1,
to prove maximal $L_p$-regularity of the resulting generalized Stokes problem. In these references we need to replace $\lambda$ by $z$ and the symbol $z +\cA(\xi)$ by $M(z,i\xi)$.
We will not do this here in detail and refer the reader to Section 7.1 of \cite{PS16} for this analysis.
\vspace{.2cm}\noindent
{\em Step 4: The Lopatinskii-Shapiro Condition}.\\
In order to guarantee the solvability of the above problem in a half-space, we need to replace the co-variable $\xi$ by the one-dimensional differential operator
$\xi-i\nu\partial_y$, where $(\xi|\nu)=0$. The Lopatinskii-Shapiro condition then means that the problem
\begin{align}\label{eq:lopshi}
\cL(z,i\xi+\nu\partial_y) v&=0,\quad y>0,\\
u(0)=\partial_y \theta=\partial_y d&=0,\nn
\end{align}
admits only the zero solution in $L_2(\RR_+)^{2n+1}$, for all $(z,\xi)\neq(0,0)$.
In order to prove this condition, suppose that $v(y)$ is a solution of the ODE system \eqref{eq:lopshi}, which belongs to $L_2(\R_+)^{2n+1}$. Taking the inner product with
$v$ in $L_2(\R_+)$ , taking real parts, integrating by parts with respect to \ $y$ and employing the boundary conditions, we obtain the estimate
\begin{align*}
c{\rm Re}(\cL(z,i\xi+\nu\partial_y) v|v)_{L_2}&\geq {\rm Re}\, z [ |u|_{L_2}^2 +|\theta|_{L_2}^2 +|\xi|^2|d|_{L_2}^2+|\partial_y d|^2_{L_2} ]+|z|^2|d|_{L_2}^2\\
& \quad + |\xi|^2(|u|_{L_2}^2 +|\theta|_{L_2}^2) + |\partial_yu|_{L_2}^2 + |\partial_y\theta|_{L_2}^2.
\end{align*}
This shows that the Lopatinskii-Shapiro is valid. Hence, to prove maximal $L_p$-regularity in the half space case, we may proceed in the following way. First we perform the
same Schur reductions as in Step 2 and as in \cite{HP15a}. This yields the unique existence of $\theta$ and $d$ in the right regularity class. We then employ the half-space theory for
the generalized Stokes symbol $M$ by the methods in Bothe and Pr\"uss \cite{BP07} or Pr\"uss and Simonett \cite{PS16}, Section 7.2, to obtain maximal $L_p$-regularity for the
half-space case.
\vspace{.2cm}\noindent
{\em Step 5: General Domains and Variable Coefficients}.\\
The results of Step 3 and Step 4 extend by a perturbation argument to a bent half-space, and to the case of variable coefficients with small deviation from constant ones.
We then may apply a localization procedure to cover the case of general domains with smooth boundaries and variable coefficients. For details we refer at this point e.g.\ to
Sections 6.3 and 7.3. of the monograph \cite{PS16} by Pr\"uss and Simonett. This completes the proof of Theorem \ref{thm:mr}.
\rightline{$\Box$}
\section{Proofs of the Main Results}
In this section we present the proofs of the above three main results. They are based on the theory of quasilinear parabolic evolution equations, see Section 3.
\vspace{.1cm}\noindent
{\em Proof of Theorem \ref{thm:1}}. \\
As already discussed above, we rewrite the system \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const}, \eqref{eq:q2} as a quasi-linear evolution equation of the form
\begin{equation} \label{NLCF}
\dot{v} +A(v) v = F(v),\quad t>0,\; v(0)=v_0,
\end{equation}
replacing $\cD_td$ appearing in the the equations for $u$ and $\theta$ by the equation for $d$.
Here $v=(u,\theta,d)$. We further apply the Helmholtz projection $\PP$ to the equation for $u$ and recall the base space
$$
X_0= L_{q,\sigma}(\Omega)\times Y_0,
$$
with $Y_0=L_q(\Omega)\times H^1_q(\Omega;\R^n)$ as well as the regularity space $X_1$ as above, i.e.
$$
X_1 =\{ u\in H^2_q(\Omega;\RR^n)\cap L_{q,\sigma}(\Omega):\, u=0 \mbox{ {on} } \partial\Omega\}\times Y_1,
$$
with
$$
Y_1 = \{ (\theta,d)\in H^2_q(\Omega)\times H^3_q(\Omega;\RR^n):\, \partial_\nu \theta=\partial_\nu d =0 \mbox{ on } \partial\Omega\}.
$$
In order to prove local well-posedness of the system \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const}, \eqref{eq:q2} we may now resort to the abstract theory presented in Section 3.
Note first that by Theorem \ref{thm:mr} the quasi-linear part $A(v)$ has maximal $L_p$-regularity. A result by Pr{\"u}ss and Simonett \cite{PS04}, Theorem 2.4, implies that $A(v)$
also admits maximal regularity in $L_{p,\mu}(J;X_0)$, hence also in the situation of time weights.
Recalling the solution space
$\EE_\mu(J)= H^1_{p,\mu}(J;X_0)\cap L_{p,\mu}(J;X_1)$, we see that the time-trace space $X_{\gamma,\mu}$ of $\EE_\mu(J)$ is given as in \eqref{eq:trace}, and the embedding
\begin{equation}\label{eq:embedding}
X_{\gamma,\mu}\hookrightarrow B_{qp}^{2(\mu-1/p)}(\Omega)^{n+1}\times B_{qp}^{1+2(\mu-1/p)}(\Omega)^n\hookrightarrow C^1(\overline{\Omega})^{n+1}\times C^2(\overline{\Omega})^n,
\end{equation}
holds, provided
\begin{equation*}
\frac{1}{p} +\frac{n}{2q} +1/2<\mu\leq 1.
\end{equation*}
Here $B^s_{pq}(\Omega)$ denote as usual the Besov spaces; see e.g.\ Triebel \cite{Tri92}. Therefore, the mappings $A$ and $F$ satisfy the assumptions
of the local existence theorem Theorem 3.1, as well as of Corollary 3.2, hence we obtain local well-posedness for \eqref{NLCF} and strong solutions on a maximal time interval.
Even more, if only $$\frac{1}{p} +\frac{n}{2q}<\mu\leq 1$$ holds, also the results in
LeCrone, Pr\"{u}ss and Wilke \cite{LPW14}, Theorem 2.1, apply, and we obtain local strong solutions if the initial values only satisfy
$$ (u_0,\theta_0,d_0)\in B^{2(\mu-1/p)}_{qp}(\Omega)^{n+1}\times B_{qp}^{1+2(\mu-1/p)}(\Omega)^n$$
plus compatibility conditions, which means that it is enough to assume that $u_0,\theta_0, d_0,\nabla d_0$ are H\"older continuous, choosing $\mu$ close to $1/p$ which is possible if $q$ is large enough.
Recalling that the state manifold of (\ref{NLCF}) is given by
$$
\cSM=\{ (u,\theta, d)\in X_{\gamma}: \, \theta>0,\; |d|_2=1\},
$$
where $X_\gamma:=X_{\gamma,1}$, we see that by these results that $\cSM$ is locally positive invariant for the semi-flow, the total energy ${\sf E}$ is preserved and the negative total entropy $-{\sf N}$ is a strict Lyapunov functional for the semi-flow on $\cSM$. This completes the proof of Theorem \ref{thm:1}.
\rightline{$\Box$}
\vspace{.1cm}\noindent
{\em Proof of Theorems \ref{thm:2} and \ref{thm:3}}. \\
The linearization of the system \eqref{eq:el}, \eqref{eq:thermo}, \eqref{eq:const} at an equilibrium $v_*=(0,\theta_*,d_*)$ is given by the operator
$A_*=A(v_*)$ defined in the base space $X_0$ with domain $D(A_*)=X_1$. This operator has maximal $L_p$-regularity. Moreover,
$A_*$ is the negative generator of a compact analytic $C_0$-semigroup having compact resolvent, due to the compact embedding of $X_1=D(A_*)$ into $X_0$.
Hence, its spectrum consists only of countably many eigenvalues of finite multiplicity.
\begin{lemma}
Let $z \ne 0$ be an eigenvalue of $A_*$. Then ${\rm Re}\,z < 0$.
\end{lemma}
{\em Proof}.
Suppose that $z\in \C\backslash\{0\}$ is an eigenvalue of $A_*$ with ${\rm Re}\, z\geq 0$. Then
\begin{align*}
\cL_\pi(z,\nabla)v_\pi &= 0\quad \mbox{ in } \Omega,\\
u=\partial_\nu \theta=\partial_\nu d &=0\quad \mbox{ on } \partial\Omega.
\end{align*}
where $(v_\pi)=(u,\theta,d)$ as above. Multiplying the equation for $d$ with $\bar{z}$ and taking the inner product of this equation with $v_\pi$ in $L_2(\Omega)$ yields by integration by parts
the estimate
\begin{align*}
0={\rm Re}(\cL_\pi(z,\nabla)v_\pi|v_\pi)_{L_2} &\geq c\Big[ {\rm Re}\,z( |u|_{L_2}^2 + |\theta|_{L_2}^2 + |\nabla d|_{L_2}^2)\\
&\quad + |z|^2|d|_{L_2} + |\nabla u|_{L_2}^2 +|\nabla \theta|_{L_2}^2\Big].
\end{align*}
This implies $u=\theta=d=0$. Hence $A_*$ does not have eigenvalues in the $L_2$-setting with nonnegative real parts, except for $z=0$.
Due to elliptic regularity, eigenvalues are independent of $p$, and so the assertion follows for $A_*$ defined in $X_0$.
\rightline{$\Box$}
The above lemma states that all eigenvalues of $A_*$ expect for $0$ are stable. In addition, the eigenvalue $0$ is semi-simple. Its
eigenspace is given by
$$
{\sf N}(A_*)=\{(0,\vartheta, {\sf d}):\, \vartheta\in\RR, {\sf d}\in\RR^n\},
$$
and hence coincides with the set of constant equilibria $\bar{\cE}$ determined in Theorem \ref{prop:equi} when ignoring the constraint $|d|_2=1$ and
conservation of energy. Therefore each such equilibrium is normally stable.
Hence, the assertion of Theorem \ref{thm:2} follows by means of the generalized principle of linearized stability, Theorem 3.3.
Finally, we note that the assertion of Theorem \ref{thm:3} follows from Theorem 3.4, as we have compact embeddings and $-{\sf N}$ serves as a strict Lyapunov functional.
\rightline{$\Box$}
|
\section{Introduction}
\label{sec:introduction}
\input{sections/intro.tex}
\section{Background on ART and {\large\textbf{\textit{DEX2OAT}\xspace}}~}
\label{sec:background}
\input{sections/background.tex}
\section{The Case for Compiler-Assisted Security on Android}
\label{sec:dilemma}
\input{sections/dilemma.tex}
\section{{\large\textbf{\textit{ARTist}\xspace}}~Design}
\label{sec:architecture}
\input{sections/architecture.tex}
\section{Use Cases}
\label{sec:usecases}
\input{sections/usecases.tex}
\section{Discussion}
\label{sec:discussion}
\input{sections/discussion.tex}
\section{Conclusion}
\label{conclusion}
\input{sections/conclusion.tex}
{
\bibliographystyle{abbrv}
\subsection{{\large\textbf{\textit{ARTist}\xspace}}}
\label{sec:discussion:artist}
We first evaluate and discuss general limitations of the \textit{ART Instrumentation and Security Toolkit}\xspace.
\subsubsection{Robustness}
\label{sec:discussion:artist:eval}
We briefly evaluate \textit{ARTist}\xspace in terms of its robustness by applying an instrumentation routine to a subset of the top apps from different Google play store categories and observing their execution. For the experiment, \textit{ARTist}\xspace injects into each method of a target application calls to a carefully crafted tracing method that is merged into the app's code. The tracing method uses stack inspection to determine its caller and prints the corresponding method name to the log. All tests are conducted on a real device (rooted Nexus 5 running Android 6 factory image). Out of 85 non-multidex apps (see Section \ref{sec:discussion:artist:impllimit}), 83 apps were successfully instrumented and remained stable when tested (97.64\% success rate), indicating the robustness of \textit{ARTist}\xspace's instrumentation capabilities.
Because runtime overhead for instrumented apps largely depends on the concrete code that is injected, we refer to the concrete benchmarks of the \textit{Modules} in Sections \ref{sec:discussion:permmodule:eval} and \ref{sec:discussion:taintmodule:eval} for a performance evaluation.
\subsubsection{Conceptual Limitations}
\label{sec:discussion:artist:conceptlimit}
\paragraph{Native code support.} \textit{Optimizing}\xspace operates by design on \textsf{dex}\xspace input only. Bundled native libraries (i.e., C/C++) that are connected via JNI are never transformed into \textit{Optimizing}\xspace's IR and therefore neither instrumented nor inspected by our prototype. Native code components are a limitation of the attacker model of not only our concept but indeed an open challenge for most of the solutions by Android security research, e.g. code analysis as well as IRM solutions in particular
\paragraph{Potential fallback to \textsf{dex}\xspace.} The \textsf{oat}\xspace files produced by \textit{dex2oat}\xspace still contain the original \textsf{dex}\xspace byte code of the app to allow fallback to interpretation mode. Naturally, fallback to interpretation would render our instrumentation of the compiled \textsf{dex}\xspace byte code futile. This fallback is currently limited to app debugging, however, no guarantees exist that such a fallback cannot be triggered maliciously. Similarly, dynamically loaded \textsf{dex}\xspace code~\cite{kruegel:ndss14:dynloading,Grace2012} (e.g., via the \hformat{DexClassLoader}) is by default compiled to native bytecode, but no guarantee can be given that dynamically loaded code cannot fall back to interpretation.
\subsubsection{Implementation Limitations}
\label{sec:discussion:artist:impllimit}
\paragraph{Permanence of instrumentation.} Instrumentation of an app's \textsf{oat}\xspace file might be reverted through an application update or a firmware update where apps are re-compiled. Thus, there exists a window of opportunity for an attacker to start an uninstrumented app after a system or app update. Apps, however, cannot be started programmatically after install/update until the user has started the app manually and both scenarios can be detected by \textit{deployment app}\xspace via system notifications (i.e., broadcasts). Assuming that the system notifies the \textit{deployment app}\xspace fast enough in order to re-instrument the updated app before the user manually starts the app, the window of opportunity in which an uninstrumented app is started can be closed.
\paragraph{Deployment strategy.} In order to create a pure application layer solution, our prototype currently relies on the na{\"i}ve approach of requesting elevated privileges to replace the installed app \textsf{oat} file with the instrumented version. We can eliminate this requirement by integrating \textit{ARTist}\xspace with an application layer only sandboxing solution that provides file system virtualization, such as \textit{Boxify}~\cite{backes2015boxify} or \textit{NJAS}~\cite{bianchi2015njas}, or by resetting the execution environment and replacing loaded libraries using reference hijacking \cite{RefHijacking}. Both approaches enable the manipulation of file paths from the original to the instrumented \textsf{oat} file at application startup time.
\paragraph{Multidex support.} Currently our \textit{ARTist}\xspace prototype does not support multidex\footnote{\url{http://developer.android.com/tools/building/multidex.html}} apps. Enabling support for such apps would not only improve coverage of our prototype for more complex apps, but also simplify the merging of additional code by shifting the merge process from the pre-processing step in \textit{deployment app}\xspace into the \textit{sec-compiler}\xspace by simply providing the additional \textsf{dex}\xspace files as input to the compilation process.
\subsection{Dynamic Permission Module}
\label{sec:discussion:permmodule}
We briefly evaluate the performance impact of our dynamic permission module and discuss limitations.
\subsubsection{Evaluation}
\label{sec:discussion:permmodule:eval}
The additional security checks inlined by our \textit{Module} are only inserted before permission-protected SDK method calls, so we cannot rely on benchmark apps, because they rarely trigger the added functionality. Therefore, we evaluate the performance impact of our permission checking code using custom microbenchmarks. Table \ref{table:PermMicrobenchmarks} depicts the results of our measurements for calls that are protected by 3 distinct permissions. The overhead encountered in the microbangemarks ranges between 1.18\% and 30.65\%, thus showing the feasibility of our prototype.
\input{tables/PermMicrobenchmarks}
\subsubsection{Limitations}
\label{sec:discussion:permmodule:limit}
\paragraph{Restriction to synchronous calls.} In order to demonstrate the straightforward implementation of an \textit{ARTist}\xspace \textit{Module}, we opted for a simple instrumentation strategy that only covers synchronous permission-protected method calls. In result, the current prototype does not support callbacks or asynchronicity and its implementation should therefore be considered a proof-of-concept only.
\paragraph{Best effort permission map.} In order to direct \textit{ARTist}\xspace to the instrumentation targets, i.e., the application's permission-protected method calls, we utilize a map of methods calls to the permissions enforced by those calls. While the PScout project \cite{au2012pscout} provides exact API method to permission mappings up to Android version 5.1.1, \textit{ARTist}\xspace operates on Marshmallow that requires mappings for Android 6. Consequently, the permission map utilized by our \textit{Module} is a hand picked subset of methods from the v5.1.1 PScout map that did not change for Marshmallow.
\subsection{Taint Tracking Module}
\label{sec:discussion:taintmodule}
We evaluate our taint tracking \textit{Module} in terms of feasibility, performance, and the limitations of its current prototypical implementation.
\subsubsection{Evaluation}
\label{sec:discussion:taintmodule:eval}
\paragraph{Runtime overhead.}
We leverage an Android microbenchmark application to evaluate the performance of our prototype. Since our taint-instrumentation only affects the performance of Java code, we specifically chose the \textit{Passmark} benchmark, which does not contain native libraries and implements all benchmarks in Java. Table~\ref{table:PerfBench} compares the results of the baseline benchmark with a non-instrumented \textit{Passmark} app to those of an instrumented and taint-aware version. The results show an overhead ranging between 7.74\% and 30.73\%, which is within an acceptable range for a taint tracking approach that is not fully tuned for performance. This result is also roughly comparable to microbenchmark results of TaintDroid's~\cite{EnGiBy_10:Taindroid} interpreter-based approach. However, as stated in \cite{managedTT}, microbenchmarks are not very representative in user-driven scenarios such as Android apps, so we take this result with a grain of salt.
Overall performance can be enhanced by introducing custom optimizations specifically tailored towards improving taint tracking code. One approach would be to eliminate \textit{taint-sets} and \textit{taint-gets} that are based on stack operations and cancel each other out, e.g., alternating \textit{pushs} and \textit{pops} of the same tag as seen for methods that return the return value of another method call. Moreover, the analysis phase allows to abstain from instrumenting apps that do not contain any global taint sinks in order to not impact performance at all in this case.
\input{tables/PerfBench}
\paragraph{Functional Evaluation.} We conducted this case study to research whether intra-application taint tracking can be achieved with an compiler-based instrumentation framework such as \textit{ARTist}\xspace, so our functional evaluation focuses on detecting different kinds of data leaks in apps. However, to the best of our knowledge, there is no standardized test suite specifically tailored towards evaluating dynamic taint tracking systems for Android apps, and testing real applications is not feasible because they lack the required ground truth. In order to overcome this unsatisfactory situation, we decided to exploit an open source suite called \textit{DroidBench}~\cite{Arzt:2014:FPC:2594291.2594299,droidbenchwebsite} that was initially created to benchmark static taint tracking systems. Even though this does not immediately apply to a dynamic system such as ours, we can still leverage the fact that it provides us with an assortment of applications with different but well-defined leakage behavior. Table~\ref{table:DroidBenchShort} summarizes our \textit{Module's} results for those tests and categories within scope. Tests for implicit flows, inter-component communication, and reflection are omitted because they currently exceed the scope of our proof-of-concept taint tracking. As we are \textit{abusing} the benchmark suite, we need to be careful which conclusions we draw from the test results. The first insight we gain, however, is that our case study succeeded in showing that intra-app taint tracking can be implemented as a pure application layer solution using compiler-driven instrumentation. The second insight we derive is that, as indicated by lower results such as those for the \textit{Android Specifics} category, our proof-of-concept does not yet catch up with previous works such as TaintDroid. Nonetheless, our work not only shows the feasibility of the approach but also lays the foundation for creating a full-fledged taint tracking system for Android versions above Marshmallow that utilizes compiler-based instrumentation and does not require modification of the operating system.
\input{tables/DroidBenchShort}
\newpage
\subsubsection{Limitations}
\label{sec:discussion:taintmodule:limit}
\paragraph{No tracking of implicit flows.} Like TaintDroid~\cite{EnGiBy_10:Taindroid}, our system currently does not track implicit flows (i.e., data leakage using control flow dependencies) and malevolent apps could exfiltrate data in a way that is unnoticeable by our prototype. As the TaintDroid authors discuss, mitigating leakage through control flows would require static analysis and access to the app's source code---both of which TaintDroid could not provide. \textit{ARTist}\xspace however is already provided with the full app code and it would be highly interesting future work to investigate to which extent the structural program information of the IR and analytical features of the compiler backend (e.g., \textit{Optimizing}\xspace) can help to remedy the limitations of customary taint tracking solutions on Android.
\paragraph{Taint tracking boundaries.} The compiler is restricted to the app's code base, which introduces imprecision when leaking information through SDK methods, where a \textit{taint-set} at the caller side (developer code) but not the \textit{taint-get} at the callee side (SDK) can be inlined. In particular, and in contrast to object types, storing primitives or strings in collections or sharing them across threads are corner cases where the taints will not be propagated appropriately. This shortcoming can be solved by using pre-computed control-flow models for framework methods~\cite{cao2015edgeminer} to generate corresponding \textit{taint-set} and \textit{taint-get} pairs that model the transition of data through the framework. A preferable technical solution in the future, which removes the potential over-approximations of SDK internal states in control-flow models~\cite{cao2015edgeminer} and which could be of interest beyond taint tracking, is the instrumentation of the \textit{core image}. The core image is a pre-compiled \textsf{oat}\xspace file of the framework classes that is pre-loaded into every application process via Zygote. Since the core image is created with \textit{dex2oat}\xspace during the device startup once after each system update, it can be instrumented using a \textit{sec-compiler}\xspace as in \textit{ARTist}\xspace. However, in either case and as in the original work~\cite{EnGiBy_10:Taindroid}, data that already left the phone (e.g., through a network socket) cannot be tracked.
\newpage
\paragraph{Inter-application communication.} Our prototype is currently limited to \textit{intra}-application tracking and lacks support for \textit{inter}-application tracking, for instance, through the file system or Binder IPC. This opens the possibility of confused deputy~\cite{PoWaMoHaCh_11:PermRedeleg,DaDmSa_10:PrivEscalation} or collusion attacks~\cite{ScZhZhInKaWa_11:Soundcomber,BuDaDm_12:TowardsT} to exfiltrate data. Assuming that all installed apps are instrumented, a fix to this problem would be the instrumentation of the I/O method calls in order to write out taints together with the data (e.g., into a file or Binder Parcel) and restore the taints at receiver side. When abandoning the requirement for a pure application-layer solution, our system could also be complemented with the original TaintDroid file system and IPC infrastructure, which is unaffected by the loss of DVM, in order to track taints across applications.
\subsection{Android Runtime}
\label{sec:back:general}
Android is essentially a Linux-based operating system with an extensive middleware software-stack on top of the kernel. The middleware provides native libraries, a feature-rich application framework that implements the Android SDK, and a managed runtime on top of which system as well as third-party applications and a small number of framework services are executed. The runtime executes bytecode generated from Java-based applications and Android's SDK components. The runtime provides the code executed within its environment the necessary hooks to interact with the rest of the system, such as the operating system, the application framework services, or the native Android user space (i.e., components running outside the managed runtime). Every process executing an application runtime environment is usually forked from a warmed-up process, called Zygote, which has all necessary libraries and a skeleton runtime for the app code preloaded.
\paragraph{Runtime prior to Android 5.} On Android devices prior to version 5, the runtime consisted of the DEX bytecode interpreter (or Dalvik virtual machine), which was specifically designed for devices with constrained resources (e.g., register-based execution model instead of stack-based). It executes Dalvik executable bytecode (short \textsf{dex}\xspace), which is created from the Java bytecode of applications at application build time. Thus, every application package ships the \textsf{dex}\xspace bytecode compiled from the application Java sources. Additionally, since Android version 2.2, Dalvik uses just-in-time compilation of hotspot code segments in order to improve the runtime performance of applications.
\paragraph{Runtime since Android 5.}
With Android 5, Google moved over from an interpreter-based app execution to an on-device, ahead-of-time compilation of apps' \textsf{dex}\xspace bytecode to native code that is executed in a newly introduced managed runtime called \textit{ART}. This shift in the runtime model was intended to address the app performance needs of Android's user and developer base. The new compiler suite was designed from scratch to allow for compile time optimizations that improve application performance, start up time, battery lifespan, and also to solve some well-known limitations of the previous interpreter-based runtime, such as the 65k method limit\footnote{\url{http://developer.android.com/tools/building/multidex.html}}. In particular, Google made the \textit{Optimizing} compiler backend, which was introduced as opt-in feature in Android 5, the default backend in Android 6. In the following Sections~\ref{sec:back:dex2oat} and \ref{sec:back:optimizing}, we will elaborate in more technical details on this new compiler suite and in particular on the \textit{Optimizing}\xspace backend.
\paragraph{Prior documentation of ART.} Even though the Android source code is publicly available as a part of the Android Open Source Project~(AOSP), little attention has yet been given to ART from a security researcher's perspective. Paul Sabanal had an early look~\cite{hidingart} at the Android Runtime right after its silent introduction as a developer option on Android 4.4 KitKat. Beside providing information on the ART executable file formats, the paper discusses the idea of hiding rootkits in framework or app code, assuming root access has already been granted. However, especially in its early phase, the Android Runtime has undergone frequent changes, which by now make this documentation unfortunately outdated.\footnote{E.g., compare the documented \textsf{oat}\xspace version 45 and the current version 63 at the time of this writing.}
Another work~\cite{fuzzingart} focuses on fuzzing the new runtime with automatically generated input files in order to detect bugs and vulnerabilities. While it provides some high-level overview on the compiler structure and its backends, it unfortunately omits any deeper information on the \textit{Optimizing}\xspace backend.
Thus, this background section servers also the purpose of filling a gap in the technical documentation of those new Android features.
\begin{figure}
\includegraphics[width=\linewidth]{figures/background}
\vspace*{+0.1cm}
\caption{A high-level overview of the \textit{dex2oat}\xspace compiler using \textit{Optimizing}\xspace backend including the transformation to the IR, optimizations, and native code generation. }
\label{figure:dex2oat}
\end{figure}
\subsection{\textbf{\textit{DEX2OAT}\xspace}~ Compiler Suite}
\label{sec:back:dex2oat}
Android's on-device compiler \textit{dex2oat}\xspace is responsible for the validation of applications and their compilation to native code. It was designed from scratch to be highly flexible and of modular structure, providing numerous configuration possibilities, compiler backends, and native code generators for supported Android platforms. The general workflow of the compiler suite is depicted in Figure~\ref{figure:dex2oat} and its steps will be explained in the remainder of this section. Providing a full technically documentation of the entire compiler suite and all its intricacies would unfortunately exceed the space limitations of this paper. Therefore, in this section we only focus on the parts relevant to this paper.
\subsubsection{Input File Format}
As an input format, \textit{dex2oat}\xspace expects the very same \textsf{dex}\xspace files that DVM used to interpret.
This strategical decision ensured that neither developers nor app store operators needed to adapt their code to ART. Developers still upload their apps as Android Application Package (\textsf{APK}) files that bundle the app's code with its resources. When a new app is installed on the device, \textit{dex2oat}\xspace compiles the app \textsf{dex}\xspace bytecode and the ART runtime executes it, which is completely transparent for the end user. Using this strategy, ART is still compatible with the old Android app base without enforcing a fallback to interpretation, which would loose all benefits that the new compiler provides.
\subsubsection{Compilation}
Before the actual compilation is performed, each input \textsf{dex}\xspace file is checked for validity. Those checks are more extensive and stricter than those implemented in the DVM in order to allow for state-of-the-art code optimizations. The compilation itself is done on a per-method base and can be
parallelized. \textit{dex2oat}\xspace delegates the actual compilation completely to the backend and only writes the results of the compilation to an \textsf{oat} file along with the original \textsf{dex}\xspace code. There are three compilation phases shared between all backends:
\begin{description}[font=\it]
\item[Transformation:] A graph-based intermediate representation (IR) is created from the \textsf{dex}\xspace code. Depending on the actual backend, multiple IRs are possible.
\item[Optimization:] Given a populated IR graph, the code is optimized. Each backend provides its own set of optimization measures, ranging from very basic techniques to state-of-the-art algorithms.
\item[Native code generation:] The IR nodes are transformed to native code using a code generator for the specific CPU architecture of the current platform. The level of sophistication of the register allocation algorithm and implementation of the code generator depend on the backend.
\end{description}
\paragraph{Backends.} On an Android stock device running version 5 (Lollipop) or higher, \textit{dex2oat}\xspace can choose between two different backends, \textit{Quick}\xspace and \textit{Optimizing}\xspace. Although \textit{Quick}\xspace was \textit{dex2oat}\xspace's default backend until Android 6, we focus in the remainder of this section and paper on the newer \textit{Optimizing}\xspace backend. This choice is not only motivated by the fact that \textit{Optimizing}\xspace is the default backend since Android 6 but also by the fact that \textit{Quick}\xspace is essentially derived from Dalvik and lacks a sophisticated IR that can support state-of-the-art compiler optimizations---including sophisticated security-oriented algorithms. However, \textit{Optimizing}\xspace is designed completely from scratch and little is yet known about its internal structure and design.
In Figure~\ref{figure:dex2oat} the compilation steps of \textit{Optimizing}\xspace are depicted. More insights on the inner workings of the new default compiler backend will be provided in Section~\ref{sec:back:optimizing}.
\subsubsection{Oat File Format}
\textsf{Oat} files are Android’s new file format for apps that are ready to be loaded and executed by the ART runtime. Even though the format was newly created for the Android platform, technically speaking \textsf{oat}\xspace files are specialized ELF shared objects that are loaded into processes, i.e., loading a compiled app into an application process is comparable to loading an (ELF) shared library into the process space of a dynamically linked executable. Besides the native code generated with \textit{dex2oat}\xspace, \textsf{oat}\xspace files contain the complete original \textsf{dex}\xspace code, which is required to hold up consistency between the code that the developer wrote in Java, the \textsf{dex}\xspace code that used to be interpreted, and the compiled code, or to allow fall back to interpretation mode when debugging apps.
\subsection{Optimizing Intermediate Representation}
\label{sec:back:optimizing}
\begin{figure}[t]
\centering
\includegraphics[width=.75\linewidth]{figures/codesnippet}
\vspace*{+0.2cm}
\caption{An example code snippet containing a leak of the device's phone number to the logging facility.}
\label{figure:codesnippet}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=.8\linewidth]{figures/ir}
\vspace{+0.2cm}
\caption{Generated IR in SSA form for the \texttt{getID()} method in Figure~\ref{figure:codesnippet}.}
\label{figure:ir}
\end{figure}
We introduce insights into \textit{dex2oat}\xspace's \textit{Optimizing}\xspace backend, which we derived mainly from the AOSP source code of the ART project's master branch between June 2015 and February 2016. \textit{Optimizing}\xspace's intermediate representation is essentially a control flow graph on the method level, which the Android developers denote as \textsf{HGraph}\xspace. It is further enriched with structural data about the program and populated with instruction nodes, denoted as \textsf{HInstructions}\xspace. Figure~\ref{figure:codesnippet} presents an example Java code and Figure~\ref{figure:ir} presents the resulting\footnote{Presented code is simplified and limited to relevant instructions for the sake of readability.} \textsf{HGraph}\xspace of the \texttt{getID} function in the \textit{Optimizing}\xspace IR. We will come back to this example in our case study in Section~\ref{sec:usecases:casestudy}.
\subsubsection{\textsf{HGraph}\xspace}
The \textsf{HGraph}\xspace serves as the single intermediate representation of the app code. When the graph is created, \textsf{dex}\xspace instructions of the app's bytecode are scanned one by one and the corresponding \textsf{HInstructions}\xspace are created and interlinked with the current basic block and the graph. In order to allow for complex optimizations, the graph is transformed into a \textit{single static assignment form}~(SSA). Pairs of value definitions and usage, so-called \textit{def-use-pairs}, are created during a liveness analysis and explicitly interlinked afterwards. At this point, \textit{phi nodes} are introduced where static analysis cannot reliably decide which value will be assigned at a given position.
In this form, the graph is amenable to a multitude of possible optimizations. The available optimizations includes algorithms such as \textit{BoundsCheckElimination} to remove redundant bounds checks, \textit{GVNOptimization} to remove duplicate code, dead code elimination, or \textit{loop invariant code motion} to optimize hotspot code in loops. In the following Section~\ref{sec:architecture}, we will show how this form is also amenable to security-oriented instrumentation, thus supporting compiler-based security solutions on Android, such as dynamic taint tracking (see Section~\ref{sec:usecases:casestudy}).
\subsubsection{\textsf{HInstructions}\xspace}
The \textsf{HGraph}\xspace nodes roughly correspond to \textsf{dex}\xspace instructions. The supplementary material provides an overview of the \textsf{dex}\xspace instructions and their corresponding \textsf{HInstructions}\xspace counterparts. Beside this transformation, nodes in the \textsf{HGraph}\xspace have additional attributes that have no equivalent in \textsf{dex}\xspace bytecode (e.g., an SSA index). The \textsf{HInstructions}\xspace distinguish between arguments and inputs. While the former corresponds to the arguments given to an operator or method, the latter encodes additional dependencies that may not be immediately observable given only the underlying \textsf{dex}\xspace code, as in the case of static method invocations that in addition to their arguments have an \hformat{HLoadClass} or \hformat{HClinit} as their input. All \textsf{HInstructions}\xspace share a basic set of information: Type, inputs, uses, id, and further data is attached to each node in order to ease the creation of and working with the \textsf{HGraph}\xspace. Each node is uniquely identified within the graph by its id that is assigned and incremented continuously during node creation. The type can be \textsf{void} for methods that have no return value, \textsf{not} for strings and object types of any kind, and additionally any of the Java primitive types. In order to get the actual object type, a fallback to the original \textsf{dex}\xspace file is required. This loose coupling between \textsf{HInstructions}\xspace and \textsf{dex}\xspace instructions as well as the presence of a method local \textsf{dex}\xspace program counter in each node show that the IR nodes are not completely independent of the original \textsf{dex}\xspace file.
\paragraph{Semantic consistency.} In addition to the instructions that represent the original application logic, the \textsf{HGraph}\xspace also contains meta-instructions to preserve the semantic consistency between the original Java code of the developer, the \textsf{dex}\xspace bytecode shipped with APKs, and the native bytecode actually executed in ART. First, additional instructions are inlined into the graph to support meaningful debugging (e.g., to map from segmentation faults in ART to actual stack traces) and to conduct various forms of runtime checks (e.g., checking type casting, bounds checking, division-by-zero checks, or null pointer exceptions). Second, instructions to represent so-called suspension points are added, which effectively subdivide the application code into multiple chunks. Each suspension point between two chunks acts as a synchronization point between native code and original \textsf{dex}\xspace bytecode in the program execution and also serves as an entry point for garbage collectors or debuggers.
\subsection{IRM for Dynamic Permission Enforcement}
\label{sec:usecases:IRM}
In the literature, Inline Reference Monitoring (IRM) is mostly implemented by modifying the bytecode before the installation~\cite{JeMiVa_11:DrAndroid,DaSaKh_12:IARMDroid}~or by hooking into an application's method at the caller or callee side at runtime~\cite{backes13TACAS}. By utilizing a security-instrumented compiler, IRM can be implemented without the need to resign and repackage apps as it is required by established approaches. Moreover, \textit{dex2oat}\xspace-based IRM can operate at instruction granularity instead of at the method level. Those capabilities are showcased by our IRM injection \textit{module} that allows for dynamic permission enforcement, as shown by~\cite{backes13TACAS,JeMiVa_11:DrAndroid,XuSaAn_12:Aurasium} on Android versions before Marshmallow.
The module is split into two distinct parts, the code injection routine that will inline permission enforcement code and the accompanying library that acts as a policy decision point. While the first directs the instrumentation process at installation time, the latter enforces the user's policy at runtime.
\paragraph{Code injection.} We first utilize \textit{ARTist}\xspace to locate the call sites of permission-protected SDK methods that are defined in a policy configuration file. Afterwards, \textit{ARTist}\xspace injects additional calls to our companioning library right before the call sites to check whether the critical method invocations should be allowed. This ensures that the control flow is diverted to our policy decision point before the execution of permission-protected methods.
The limitations imposed by the choice of this rather basic strategy are discussed in Section \ref{sec:discussion:permmodule:limit}.
\paragraph{Policy decision point.} The library that our \textit{Module} injects into target apps provides methods to check their current state of permissions. Based on the given user permission policy, the library either allows or rejects the execution of a protected SDK method.
\subsection{Case study: Taint Tracking}
\label{sec:usecases:casestudy}
Established approaches for dynamic taint tracking on Android~\cite{EnGiBy_10:Taindroid} rely on instrumenting the by now scrapped DVM for intra-application taint tracking or directly rewrite bytecode \cite{schutte2014appcaulk}. In this case study, we explore the applicability of a compiler-based instrumentation framework like \textit{ARTist}\xspace to re-instantiate intra-app taint tracking for applications on Android version 6 and higher. That is, through a prototypical implementation, we want to investigate whether inlining taint tracking logic into the application code base with \textit{ARTist}\xspace at compilation time can be a surrogate for solutions prior to Android version 5. Please note, that this case study does not aim at a full replacement of existing solutions like TaintDroid~\cite{EnGiBy_10:Taindroid}, but at demonstrating a new potential foundation for future taint-tracking on Android.
\subsubsection{Module Design}
\label{sec:usecase:arch}
In general, we want to track information as it flows through the code using tracking logic inlined by a new \hformat{HOptimization} in the \textit{Optimizing}\xspace backend. However, simply assigning each single value that should be tracked a taint tag and updating the tag for each single instruction operating on it will incur a major performance penalty. To minimalize the runtime impact, we split our approach into two phases: \textit{analysis} and \textit{instrumentation}. In the \textit{analysis} phase, we identify flows of tainted information between \textit{sources} and \textit{sinks}. By restricting ourselves only to those relevant flows of the values we are interested in, we avoid generating irrelevant but costly taint tracking code for parts of the method that never actually influence the data that is obvserved and gain noticable performance improvements over more na{\"i}ve taint tracking. In the \textit{instrumentation} phase, code will be inlined that creates, propagates, and checks the taint values along the identified data flows. Our combined analysis and instrumentation achieves flow-, path-, object-, and context-sensitive taint tracking.
While \cite{schutte2014appcaulk} and \cite{livshits2013towards} also utilize static analysis to optimize and guide the instrumentation process, both assume a holistic view on the application in form of a control or data flow graph. In contrast, \textit{dex2oat}\xspace backends operate on a per-method level, leaving the primary challenge for our taint tracking \textit{Module} to achieve similar tracking properties while inspecting one method at a time.
A na{\"i}ve solution to this problem would be to retrofit the compiler suite to provide an application-wide view and instrumentation. However, our prototype demonstrates how we can still achieve taint tracking for the whole application while restricting ourselves to a per-method view and instrumentation. To this end, we introduce in the following a new design for storing and propagating taint tags, in particular we have to refine the definitions of \textit{sink} and \textit{source}.
\subsubsection{Analysis Phase}
\label{sec:usecase:analphase}
In order to optimize the instrumentation with taint tracking code, we exploit the processing features (e.g., \textsf{HGraph}\xspace's Visitor~\cite{Gamma:1995:DPE:186897} pattern support) of the \textit{dex2oat}\xspace compiler to detect the data flow sources and sinks and afterwards use its static analysis features to identify the relevant data flows and the operations along those flows that have to be instrumented.
\subparagraph{Refining source and sink definition.} The literature on taint tracking for Android defines sources and sinks as the API methods that input privacy-sensitive information into the application process (e.g., framework functions that return sensitive data, such as the location or telephony API) or, respectively, leak privacy-sensitive information from the application process (e.g., file handles, Internet sockets, or logging facilities). Since \textit{dex2oat}\xspace is operating on a per-method level, we cannot assume that our analysis is able to always connect a sink and a source (e.g., when they are located in different methods). To address this problem, we have to connect the data flows of tainted variables across the different methods while maintaining the per-method-based analysis. To this end, we introduce in addition to the above mentioned sinks and sources from the literature---in the following denoted as \textit{global sinks/sources}---new \textit{method-local sinks/sources}, more precisely \textsf{HInstructions}\xspace, which form the entry and exit points for inter-procedural data flows. Thus, global sinks and sources are points of interest for taint tag creation and check, respectively, while local sinks and sources are for inter-procedural tag propagation. For local sinks and sources we differentiate between three categories each: \textit{local sources} include arguments provided to the current method~(\textsf{LSO1}\xspace), return values from method invocations~(\textsf{LSO2}\xspace), and values read from fields~(\textsf{LSO3}\xspace). Conversely, \textit{local sinks} are method invocations that leak values through its parameters from the current method~(\textsf{LSI1}\xspace), return statements of the current method~(\textsf{LSI2}\xspace), and field setting instructions~(\textsf{LSI3}\xspace). At the beginning of the analysis phase we collect all sinks within all methods and in a subsequent step detect all relevant sources for those sinks (see next paragraph).
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{figures/tracking}
\vspace{+0.1cm}
\caption{Tracking tainted variable \textit{id} from example of Figure~\ref{figure:codesnippet}. All discovered sinks and sources are marked. Solid lines indicate intra-procedural data flows of tainted variables, dashed lines inter-procedural data flows between local sink-source pairs. Right hand side depicts the inlined taint tracking code to propagate taint tags.}
\label{fig:tracking}
\end{figure*}
\subparagraph{Creating intra-procedural data flows.} After collecting all local and global sinks, we create for each sink a backward slice within the currently analyzed method by inspecting the current instruction and recursively tracing back its input parameters depending on the concrete instruction type to discover all sources that influence this sink. For instance, Figure~\ref{fig:tracking} continues the example code from Figure~\ref{figure:codesnippet}. In the Java code on the left hand side, all sinks have been identified (i.e., the parameter \texttt{id} passed to function \texttt{prefixID} in line~9 is a local sink of type \textsf{LSI1}\xspace, the return statements in lines~14 and 21 form local sinks of type \textsf{LSI2}\xspace, and in line~20 the call to \texttt{Log} forms a global sink). Using backwards slicing (solid lines \circledgreen{1}, \circledblue{3}, and \circledblue{5}) the local sources in lines~17 (\textsf{LSO1}\xspace) and 9 (\textsf{LSO2}\xspace) as well as the global source in line 6 (\texttt{getDeviceID} call to retrieve device's phone number) have been identified. Each resulting backward slice is defined by its starting point (i.e., the sink) and all found endpoints (i.e., the sources). Constants cannot be tainted and are therefore explicitly omitted as sources. Together those backward slices form the input for the instrumentation phase.
Note that high precision of this analysis is desirable but \textit{not} required for secure taint tracking. Higher precision of this backwards slicing
helps in removing irrelevant taint tracking code for the observed values and hence improves performance of the instrumented application's process, but actual taint propagation occurs at \textit{runtime} and the set of our slices contains a superset\footnote{The use of reflection is an exception to this rule since this poses serious problems for static analyses.} of the relevant flows for complete tracking of tainted values.
\subsubsection{Instrumentation Phase}
\label{sec:usecase:analphase}
In the instrumentation phase, we inline code that at runtime creates taint tags for global sources and that checks taints at global sinks. Additionally, we inline code that inter-procedurally propagates taints at runtime from a local sink to a local source, ensuring the data flow of a tainted value across multiple methods correctly propagates the taints.
\subparagraph{TaintLib.} In order to improve the flexibility of our solution by not restricting our system to a specific implementation for storing, updating, and checking taints, we make use of \textit{ARTist}\xspace's modular design and deploy the instrumentation code in form of a new \textit{companion library} called \textit{TaintLib}\xspace that is merged by \textit{ARTist}\xspace into the app code at compile time. \textit{TaintLib}\xspace in turn relies on a policy file that defines the global and local sources/sinks as well as the sources' taints tags. \textit{TaintLib}\xspace provides source-type-specific \textit{taint-set} methods, calls to which are inlined at all sources, and sink-type-specific \textit{taint-get} methods, calls to which are inlined at all sinks. By injecting \textit{TaintLib}\xspace method calls instead of concrete taint tracking logic, we decouple the instrumentation from the taint management code. For global sources, \textit{taint-get} retrieves and sets the taint tag according to the policy and \textit{taint-set} at global sinks checks\footnote{While a na{\"i}ve check halts the program when tainted data is about to leak, invoking a sanitizer as suggested by \cite{livshits2013towards} can easily be implemented in \textit{ARTist}\xspace.} the taint tag. In contrast, for local sinks \textit{taint-set} propagates the tag together with the tainted value to the next local source, where it is retrieved with \textit{taint-get}. By instrumenting all methods alike, an implicit contract between all methods is established and fulfilled, i.e., every time a \textit{taint-get} tries to obtain the taint value of a method parameter on the callee side, we assume that the corresponding \textit{taint-set} has been executed in the calling method to provide the taint data. In case the slice contains multiple \textit{sources}, the output of their corresponding \textit{taint-gets} is combined by injecting a call to a combination method that will return the merged taint tag.
To continue our running example, the right hand side of Figure~\ref{fig:tracking} presents the IR of the code snippet with \textit{taint-set} and \textit{taint-get} calls inlined. For instance, the \texttt{setArgTaint} call for \textsf{LSI1}\xspace in basic block~2 of \texttt{getID} (\textsf{HInstructions}\xspace~52) precedes the local sink in \textsf{HInstructions}\xspace~31 that invokes the \texttt{prefixID} function. The \texttt{setArgTaint} instruction transfers the taint of \texttt{id} inter-procedurally to the \texttt{getArgTaint} instruction in \textsf{HInstruction}\xspace~77 of basic block~0 of \texttt{prefixID} (dashed line \circledblack{2}), from where it is intra-procedurally propagated using the backwards slicing information (solid line \circledblue{3}). Similarly, the taint is propagated back from \texttt{prefixID} to \texttt{getID} through the return statement and variable assignment (dashed line \circledblack{4}).
\subparagraph{Inter-procedural taint tag propagation channel.} In the case of parameters (\textsf{LSO1}\xspace and \textsf{LSI1}\xspace) and method returns (\textsf{LSO2}\xspace and \textsf{LSI2}\xspace), there are at runtime always pairs of \textit{taint-sets} and \textit{taint-gets}, given by the fact that for each callee method, there is a caller method that also has been instrumented. Combining this with the observation that a caller-callee method pair is always executed by the same thread, the taint propagation can be realized using \textsf{thread local storage} for a \textsf{taint stack}. At the caller side, the taint information is \textsf{pushed} onto a per-thread stack and at the callee side it is \textsf{popped} again, vaguely resembling the x86 calling convention for passing arguments to methods. Keeping in mind that almost every injected \textit{TaintLib} method call accesses the taint information, replacing more straightforward approaches for taint storage (like a single \textsf{HashMap}) with cheaper stack operations also benefits the overall performance of our taint tracking solution.
In the case of field operations (\textsf{LSO3}\xspace and \textsf{LSI3}\xspace), we can neither assume them to appear in pairs nor to be executed on the same thread and therefore employ a thread-safe mapping in the form of a \textsf{ConcurrentHashMap}. This, however, raises the challenge of providing easily computable, stable and unique keys. If we consider our taint tags not to store the taint value of a certain value, but of a certain location, we can compute stable identifiers for fields and use them as keys. For \textsf{static} class fields, identifying the specific class and field is sufficient and can be precomputed during compilation. The current implementation injects the computed key as a constant into the \textsf{HGraph}\xspace and provides it as an argument to a field \textit{taint-set} or \textit{taint-get}. For object fields, we do not only need to identify classes but concrete objects, which requires runtime information. In this case, we only inject the field identifier as a constant and provide it together with the field's concrete object to a newly added \textit{TaintLib} function. The returned key is robust to object aliasing, so we do not loose track of objects in e.g. collections. Afterwards, we can use this key in a \textit{taint-set} or \textit{taint-get} for the object field.
It is important to note that our approach to taint tracking depends not only on the entity for which we store taints (i.e., variable locations instead of values), but also on the type of data to which we assign taint values. In our model, we track taints only for primitive types and the taint tag of objects is transitively given by their field's tags. In case of non-primitive fields, the rule applies recursively because eventually all objects can be decomposed to primitives. This design decision is motivated by the fact that tracking all \textit{taint-set} and \textit{taint-get} operations on fields and on all method invocations is more fine-grained than storing taint information at the object level.
\subsection{Further Use Cases}
\paragraph{Dynamic analysis.} Compiler-based solutions are inherently well-suited for white box approaches that require an understanding of the application's internals. One example is the taint tracking \textit{Module} described above that re-instantiates TaintDroid-inspired intra-app taint tracking. Other examples are existing works on commodity systems~\cite{compilerTT,araujo2015compiler} that already utilize compilers for information flow control, which can now be realized on Android as well.
\paragraph{Container solutions.} Modifications of the Android runtime environment have been used in the past (for instance \textit{Divide}\footnote{http://www.divide.com}, now part of Google Android for Work) to establish container solutions that, e.g., encrypt file system I/O of apps or restrict inter-application communication. Using a compiler-based approach such as \textit{ARTist}\xspace, similar container solutions can be established by replacing the corresponding method invocations (e.g., calls to Java's I/O classes) with calls to injected security-enhanced versions of the same.
\paragraph{Code replacement and compile-time patching.} Google has recently started separating security-critical libraries, such as the notorious WebKit, from application packages into stand-alone apps that are called by apps on-demand. This allows Google to maintain those libraries on an ecosystem-wide scale and roll out security patches more effectively. Since \textit{ARTist}\xspace is not only able to inject but also to replace or remove code from an app's code base, \textit{ARTist}\xspace can also be used to apply \textit{compile-time patches} by replacing vulnerable libraries within apps with fixed versions. In an extreme case, this mechanism could allow for removing entire libraries by mocking all their method invocations (e.g., removing ads), or moving code partitions behind a strong security boundary, such as a dedicated process, and reconnect the code through inter-process communications (e.g., as done in the AdSplit~\cite{Shekhar:2012:ASS:2362793.2362821} or AdDroid~\cite{Pearce:2012:APS:2414456.2414498} solutions).
\paragraph{Beyond security: profiling and debugging.} Besides its application in the security domain, using \textit{ARTist}\xspace to inject tracing, debugging or profiling code allows to gain additional insights into third party applications. A basic example is the method call-tracing we employ in our robustness evaluation in Section \ref{sec:discussion:artist:eval}.
\subsection{Robustness}
\subsection{Performance}
\subsubsection{Install Time}
\subsubsection{Runtime}
\subsection{Security-Instrumented Compiler}
\paragraph{Choice of instrumentation point.} The general concept of security-instrumented compilers is not restricted in its modifications of the compilation code. Given \textit{dex2oat}\xspace's modular design, there are immediately multiple possibilities apparent where app modifying code could be placed. For instance, \textit{dex2oat}\xspace's design would easily allow porting bytecode and binary rewriting approaches (\textsf{Instr}$_{\text{DEX}}$\xspace \& \textsf{Instr}$_{\text{BIN}}$\xspace) into the compiler infrastructure (cf.~Figure~\ref{figure:dex2oat3}). Of the different choices, \textit{ARTist}\xspace's \textit{sec-compiler}\xspace is concretely designed to operate on the intermediate representation of \textit{dex2oat}\xspace's \textit{Optimizing}\xspace backend (\textsf{Instr}$_{\text{OPT}}$\xspace), where the existing optimization infrastructure and static code information in the \textit{Optimizing}\xspace IR allow for efficient and precise code modification. More precisely, our app instrumentation code is realized as an \hformat{HOptimization}, which, as a result, is no different than other optimizations in the sense that they are provided with required information, such as the current method's \hformat{HGraph}, and are modularly integrated into the optimization workflow. As \hformat{HOptimization}, our security instrumentation logic has full control over the ordering and execution of optimizations, which opens up the opportunity to optimize the already instrumented code by creating or applying compatible optimizations that improve the performance of the security code.
Generally, using the \hformat{HOptimization} interface one can extend the compiler with custom functionality that is decoupled from \textit{dex2oat}\xspace's code base. We will refer to those independent extensions as \textit{Modules} for the remainder of this paper.
\paragraph{Spotting instrumentation targets.} \hformat{HGraph} supports the visitor pattern~\cite{Gamma:1995:DPE:186897} that enables us to iterate over, inspect, and modify each single \hformat{HInstruction} of the app's code. In contrast to method hooking techniques, we can therefore operate at the instruction level. In \textit{ARTist}\xspace, \hformat{HGraphVisitors} are primarily used to identify instrumentation targets and apply the desired modification. However, they can also be utilized to bootstrap static analysis. We will see concrete implementations using a visitor to collect instrumentation sites for our dynamic permission enforcement system in Section~\ref{sec:usecases:IRM} and starting points for backward slicing in our taint tracking case study in Section~\ref{sec:usecases:casestudy}.
\paragraph{Modification capabilities.} Given an instrumentation target in the form of an \textsf{HInstruction}\xspace, there are several possibilities for modification like changing types, inputs, or even removing/replacing the instruction altogether. It is also possible to instantiate new instructions of any type and inline them before or after the current target. \textit{ARTist}\xspace provides a new API dedicated to automate such modifications if provided with the information which methods, where and what to instrument. Since we are only generating nodes in the form of \textsf{HInstruction}\xspace objects and insert them into the \textsf{HGraph}\xspace, we do not have to modify the generation of native code from the IR. This means that the code generator is agnostic towards our changes and receives no unexpected instructions or structures. The result of this integrated solution is that we still take advantage of the robustness of \textit{Optimizing}\xspace's code generators, which are well-tested, constantly improved, and in productive use on every stock Android phone running version 6+.
\paragraph{Configuration.} Using \textit{Modules}, the instrumentation and modification process is already flexible. To further increase flexibility, \textit{Modules} can, in turn, depend on policy configuration files that govern the instrumentation process. While the design of such policy files highly depends on the concrete \textit{Module}, there are recurring and common patterns, for instance, the amount and type of instrumentation targets that should be detected, as demonstrated in Section~\ref{sec:usecases:IRM}. In general, this allows adaptation of existing instrumentation solutions to new targets or provisioning them with new security policies.
\subsection{Compiler Deployment App}
\label{sec:architecture:app}
Responsibility of the \textit{deployment app}\xspace is to deploy the \textit{sec-compiler}\xspace at application layer in addition to the system's \textit{dex2oat}\xspace binary. Using \textit{deployment app}\xspace, one can create security-instrumented versions of installed applications by re-compiling the apps' bytecode with \textit{sec-compiler}\xspace and replacing the \textsf{oat}\xspace files stored\footnote{Located at \path{/data/app/<package-name>-1/oat/arm/base.odex}} on filesystem. To make the Android runtime agnostic to this instrumentation, two particular challenges had to be overcome. First, Android has mechanisms in place to verify that \textsf{oat}\xspace files correspond to their respective apps and that the paths of the \textsf{oat}\xspace files are correct. Our implementation solves this challenge by rewriting paths and checksums to match those that the system \textit{dex2oat} would have generated. Second, the \textsf{oat}\xspace files are by default stored at and loaded from a protected location to which 3$^{rd}$ party apps have no access. A na{\"i}ve solution to this problem would be to require extended privileges for our \textit{deployment app}\xspace (e.g., a dedicated SELinux type or root on security-relaxed after-market ROMs). We discuss alternatives to the na{\"i}ve approach in Section~\ref{sec:discussion}, which abstain from extended privileges by using app virtualization or reference hijacking. After solving those challenges, the Android default runtime will load the instrumented \textsf{oat}\xspace file while remaining agnostic to the fact that we replaced it.
\paragraph{Executing the compiler.} Instead of shipping \textit{deployment app}\xspace with a statically linked \textit{dex2oat}\xspace binary that includes our \textit{ARTist}\xspace extensions, we opted for utilizing a copy of Android's default \textit{dex2oat}\xspace binary and leveraging its modularity to ship our extensions to the compiler suite as separate libraries. We use the \textsf{LD\_LIBRARY\_PATH} environment variable to ensure that our \textit{dex2oat}\xspace loads and dynamically links our \textit{ARTist}\xspace libraries, such as \textsf{libart-compiler.so}, from the assets directory of the \textit{deployment app}\xspace.
\paragraph{Inlining custom code.} While the instrumentation with \textit{ARTist}\xspace already provides powerful tools to modify the application, most security solutions require an additional custom code library within the app (e.g., additional taint tracking logic in Section~\ref{sec:usecases:casestudy}). To facilitate adding custom code to an instrumented app, \textit{deployment app}\xspace has a preprocessing step that is executed before the app's bytecode is compiled. This step utilizes the \textsf{DexMerger} utility to combine the app's original bytecode with the additional code library. During compilation, connections between original and new code are built in form of invocations of the added code's methods.
|
\section{Introduction}\label{sec:int}
The analysis of complex structures has become a highly attractive field in methodological as well as applied research within different disciplines. Here, one very prominent topic are relationships that are present in network structures. Most generally, the objective of interest here lies on the investigation of structures between different entities of the network. For an intense description of different aspects of network analysis we refer the interested reader to \cite{Carrington2005}, \cite{Kolaczyk:2009} and \cite{Goldenberg:EtAl:2010}.
Although most of this research has been conducted within the field of social network analysis or statistical physics, a growing number of papers considers network structures in the context of spatial processes including the investigation of spatial point patterns. The statistical analysis of such spatial point pattern data that occurs on linear networks has been pioneered by \cite{Okabe:Yamada:2001}. This paper has later been extended by \cite{Ang:2010}, \cite{Ang:Baddeley:Nair2012} as well as \cite{Baddeley:Jammalamadaka:Nair:2014} who derived a geometrically-adjusted extension of Ripleys' $K$-function \citep{ripley:76} to planar point processes on linear networks. As the underlying assumption, these models treat a point pattern $x$ on a linear network $\mathfrak{L}$ as a realization of a point process that occurs randomly on a linear network in a bounded region on the linear network space $\mathfrak{L}\subset\mathbb{R}$. Further contributions with respect to this linear network formalism are presented by \cite{Okabe:Satoh:2009}, \cite{Okabe:Sugihara:2012} and \cite{Borruso:2005,Borruso:2008}.
An alternative approach for the analysis of spatial point patterns that is not restricted to the linear network formalism or to simple point processes has recently been introduced by \cite{Eckardt:Mateu2016}. This alternative formalism allows to take undirected, directed or partially directed graph structures as well as temporal dynamics into consideration. In contrast to \cite{Ang:2010} or \cite{Ang:Baddeley:Nair2012}, this alternative approach highlights the possibility to achieve several different graph-based intensity formulations and related statistics for network structures. Different from the linear network formalism, \citet{Eckardt:Mateu2016} assume that a point pattern appears randomly between pairs of geo-referenced nodes whose location within a planar region is treated as fixed. In this respect, the intensity of a process appears as an edgewise counting measure adjusted for the geodesic distance between the endpoints of the edge.
This leads to the non-circular definition of three different classes of intensity measures, and finally to different versions of Ripleys' $K$-function based on subsets of adjacent network segments. The first class contains various intensity measures and related statistics that are calculated with respect to distinct edges. In contrast to this edgewise calculation, all measures can be formulated with respect to sequences of distinct nodes and distinct edges leading to pathwise statements. This pathwise perspective forms the second class of intensity functions. Lastly, the third class of such measures consists of average nodewise intensities which are obtained as an average intensity over subsets of adjacent nodes. These nodewise intensities are computed as means over the different sets of edge intensity functions.
Although the results of \citet{Eckardt:Mateu2016} allow for deeper insights in the behavior of possibly marked point patterns and the relational information contained in the network structure, the possible impact of additional covariates is not yet captured. Principally, this limitation might affect all three classes of intensity measures: the edgewise, the pathwise and the nodewise mean intensity function. Here, possible influences by different types of covariates seem to be a plausible assumption. To approach this limitation, we treat a generic network intensity function $\lambda(\phi) $ as the conditional expectation of a point pattern $\phi$ over a set of fixed structural elements of a network within a structured network regression model. These network elements include distinct edges for edgewise intensities, paths for pathwise intensities, and nodes in case of nodewise mean intensity functions.
This paper is in general the first which relates the intensity function to a regression model within the context of network structures where we assume that the joint distribution of a point pattern $\phi$ associated with a specific set of network elements given a set of explanatory variables and parameters belongs to an exponential family. This formulation is different from exponential random graph models (ERGMs) in which the adjacency of a network is treated as the outcome of a regression. In this sense, ERGMs aim to explore the effects on the network structure. For a detailed discussion of ERGMs we refer the interested reader to \cite{Koskinen:Lusher:Robins:2011}. Differently from ERGMs, our focus here lies on the effects on structurally calculated generic intensity functions where the network structure is treated as fixed.
This paper is organized as follows: the main results of \citet{Eckardt:Mateu2016} and different possible network intensity functions are described in Section \ref{sec:prior}. Hereafter, Section \ref{sec:reg} introduces a generalized structured regression model for network mean intensity functions. Applications of the regression model to neighbor and community disturbances data are discussed in Section \ref{sec:apl}. Finally, the concluding Section \ref{sec:conclusion} comments on the major results and impacts on future research.
\section{Network intensity functions for spatial point patterns }\label{sec:prior}
This section focuses on the introduction of different edgewise and averaged nodewise intensity functions as proposed in \cite{Eckardt:Mateu2016} and only briefly covers the underlying graph theoretical concepts. Pathwise intensity functions which are generalizations of edgewise intensities to sequences of distinct nodes and edges will not explicitly be covered here. For a comprehensive treatment of graph theory and further network related intensity functions as well as a linkage of graphical modeling to planar point processes we refer to \cite{Eckardt:2016} and \citet{Eckardt:Mateu2016} and the references therein.
Formally, a network is expressed in terms of a graph $G=(V,E)$ in which every node $v_i(s_{v_i})$ is indexed with a pair of fixed coordinates $s_{v_i}=(x_{v_i},y_{v_i})$ encoding locations of interest such as crossings in a traffic network. These fixed positions are pairwise joint by edges which are treated as edge intervals $s_{e_i}=(s_{v_i},s_{v_j})$ of arbitrary length that are spanned between two nodes. In this context, \citet{Eckardt:Mateu2016} define the set of all $k$ edge intervals in the graph as $\mathcal{S}_{E(G)}=\lbrace s_{e_1},\ldots,s_{e_k}\rbrace$. The realization of a point process $X$ with respect to a network is then understood as a random event that occurs on a location within a closed interval belonging to $\mathcal{S}_{E(G)}$. This random location is formally expressed as $\tilde{s}=(\tilde{x}, \tilde{y})$.
The number of points that fall into the edges could then be expressed by a counting measure $N(\cdot)$. If the graph only consists of undirected edges, this counting measure is defined as
\[
N(s_{e_i})=\sum\mathbbm{1}_{\lbrace x_{v_i}\leq \tilde{x}\leq x_{v_j},y_{v_i}\leq \tilde{y}\leq y_{v_j}\rbrace}X(\tilde{s}), x_{v_i}<x_{v_j},y_{v_i}<y_{v_j}.
\]
To define nodewise mean intensity functions, we first introduce a pairwise intensity function
\[
\lambda(s_{e_i})=\lim_{|ds_{e_i}|\rightarrow 0}\left\{\frac{\mathbbm{E}\left[N(ds_{e_i})\right]}{|ds_{e_i}|}\right\}, s_{e_i}\in\mathcal{S}_{E(G)}
\]
where $ds_{e_i}$ is an infinitesimal interval. From this, an intensity related to neighboring vertices is obtained as
\[
\lambda(v_i)=\frac{1}{|\dg(v_i)|}\sum_{v_j\in\nach(v_i)}\lambda(s_{e_i}).
\]
where $ e_i=(v_i,v_j),\nach(v_i)$ is the set of all $k$-nearest neighboring nodes connected to $v_i$ and $\dg(v_i)$ is the size of $\nach(v_i)$ .
Similar intensity functions for alternative graph structures can be achieved as modification of the previous definitions. For a graph which only consists of directed edges two possible counting measures exist. These are,
\[
N(s_{e_i}^{in})=\sum\mathbbm{1}_{\lbrace\pa(x_{v_i}\leq \tilde{x}\leq x_{v_j},y_{v_i}\leq \tilde{y}\leq y_{v_j})\rbrace}X(\tilde{s}), x_{v_i}<x_{v_j}, y_{v_i}<y_{v_j}
\]
and
\[
N(s_{e_i}^{out})=\sum\mathbbm{1}_{\lbrace\child(x_{v_i}\leq \tilde{x}\leq x_{v_j},y_{v_i}\leq \tilde{y}\leq y_{v_j})\rbrace}X(\tilde{s}), x_{v_i}<x_{v_j}, y_{v_i}<y_{v_j}
\]
where $\pa(x_{v_i})$ are all the edges pointing to $v_i$ and $\child(x_{v_i})$ are all the edges departing from $v_i$. Again, these counting measures can be used to define two types of directed nodewise mean intensities functions. For the set of edges which point to $v_i$ this leads to
\[
\lambda^{in}(v_i)=\frac{1}{|\dg^+(v_i)|}\sum_{v_j\in\pa(v_i)}\lambda(s_{e_i}^{in})
\]
where
\[
\lambda(s_{e_i}^{in})=\lim_{|ds_{e_i}^{in}|\rightarrow 0}\left\{\frac{\mathbbm{E}\left[N(ds_{e_i}^{in})\right]}{|ds_{e_i}^{in}|}\right\}, s_{e_i}^{in}\in\mathcal{S}_{E(G)}
\]
and $\dg^+(v_i)$ is the size of $\pa(x_{v_i})$. The mean related to the opposite relation, the set of arrows which point from $v_i$, can be expressed as
\[
\lambda(s_{e_i}^{out})=\lim_{|ds_{e_i}^{out}|\rightarrow 0}\left\{\frac{\mathbbm{E}\left[N(ds_{e_i}^{out})\right]}{|ds_{e_i}^{out}|}\right\}, s_{e_i}^{out}\in\mathcal{S}_{E(G)}.
\]
where
\[
\lambda^{out}(v_i)=\frac{1}{|\dg^-(v_i)|}\sum_{v_j\in\child(v_i)}\lambda(s_{e_i}^{out}).
\]
and $\dg^-(v_i)$ is the size of $\child(x_{v_i})$.
Besides intensity measures related to networks which only consist of directed or undirected edges, a third graph configuration might be of interest where directed and undirected edges appear simultaneously. In this case, all previously described counting measures and related statistics generally remain applicable with respect to subsets of distinct edges. Besides, additional measures could be derived as a combination of undirected and directed versions. In this respect, \citet{Eckardt:Mateu2016} presented an extended intensity measure related to the union of undirected and directed edges as
\[
\lambda^{cg}(v_i)=\frac{1}{|\dg^{cg}(v_i)|} \lambda^{out}(v_i)\cup \lambda^{in}(v_i)\cup \lambda(v_i).
\]
Alternatively, several re-definitions of $\lambda^{cg}(v_i)$ can be considered which only take certain unions of distinct subsets of edges such as $\pa(\cdot)\cup\child(\cdot)$ or $\nach(\cdot)\cup\child(\cdot)$ into account.
\section{Structured network regression model}\label{sec:reg}
Besides the edgewise, pathwise or mean nodewise intensity functions of point processes that occur on differently shaped network graphs, we now consider the situation where we want to estimate or predict the edgewise, pathwise or mean nodewise intensity functions based on additional information. This additional information is treated as a generic list of $l$ exploratory variables including various network statistics as well as additional covariates of different type and different dimension.
To begin, we assume that $\phi_i$ is the realization of a planar point process $\Phi_i$ related to the $i$-th structural element, $i=1,\ldots,n$, of a network. These structural elements could be edges, paths or nodes of a undirected, directed or partly directed network. Additionally, we denote a generic network intensity function related to the distinct types of $\lambda_i$ as introduced in \citet{Eckardt:Mateu2016} as $\lambda_i(\phi)$. That is, $\lambda_i(\phi)$ could either be related to edges, paths or nodes in undirected, directed or partially directed networks as described in Section \ref{sec:prior}.
The general idea which will be elaborated in this section is to write $\lambda_i(\phi)$ as the outcome of additive combinations of structural covariates. A principle regression model in this spirit is the generalized linear model as introduced by \cite{McCullagh1989}. Here, the distribution of an observation $\phi_i$, given a set of covariates $\mathbf{z}_i$ and unknown parameters $\gamma_i$, is assumed to belong to an exponential family
\begin{equation}
f_{\Phi_i}(\phi_i|\vartheta_i,\psi)=\exp\left(\frac{\phi_i\vartheta_i-b(\vartheta_i)}{a_i(\psi)}+c(\phi_i,~\psi)\right),~i=1,\dots,n
\label{eq:expofam}
\end{equation}
where $\psi$ is a scale parameter and $a(\cdot), b(\cdot)$ and $c(\cdot)$ are unknown functions. The conditional expectation $\mathbb{E}\left[\phi_i|\mathbf{z}_i, \gamma_i\right]=\mu_i$ is modeled by $g(\mu_i)=\eta_i$ or $\mu_i=h(\eta_i)$ where $g$ (resp. $h$) is a known link function (resp. response function) and $\eta_i=\mathbf{z}_i^{\mathsf T}\gamma_i$ is a linear predictor.
For our purpose, this linear predictor $\eta_i$ seems to be too restrictive and unable to capture all, possibly nonlinear, effects including for example graph statistics as well as spatial or temporal correlation among observations, and heterogeneity. A more flexible regression framework which unifies several extensions of the generalized linear model such as the generalized additive model of \cite{Hastie1990} or the geoadditive model of \cite{Kammann2003} is the structured additive regression (STAR) model. In this model, the linear predictor $\eta_i$ is replaced by a structured additive predictor $\eta^{\star}_i=\sum^p_{j=1} f_j(\nu_{ij})+\mathbf{z}_i^{\mathsf T}\gamma_i$ where $f_j$ are not necessary smooth functions of generic covariates $\nu_j$ of different type and dimension. To ensure the identifiability of $f_j(\cdot)$, the functions are constraint to have a zero mean. For a detailed discussion of the structured additive formalism in general we refer the interested reader to \cite{Fahrmeir2004} and \cite{Brezger2006} and the references therein. Structured additive regression models for count data have been presented in \cite{Fahrmeir2003, Fahrmeir2006} and, focussing on geoadditive survival models, in \cite{Hennerfeld2006}.
To model a generic network intensity function $\lambda_i(\phi)$ related to the $i$-th network element, we integrate a set of fixed graph statistics $\mathbf{w}_i$ into the structured additive predictor. In more detail, setting $\mu_i=\lambda_i(\phi)$ we assume that the generic network intensity function is linked to a structured network predictor $\eta^{\circ}_i=\beta_0+\eta^{\star}_i+\mathbf{w}_i^{\mathsf T}\xi_i$ where $\mathbf{w}_i$ is a set of graph statistics, $\xi_i$ is a set of unknown parameters and $\eta^{\star}_i$ is a generic structured additive predictor related to the $i$-th network element.
Precisely, the generic structured network predictor for the $i$-th structural network element is defined as
\begin{equation}
\eta^{\circ}_i=\beta_0+\sum^p_{j=1} f_j(\nu_{ij})+f(\alpha_s)+\mathbf{z}_i^{\mathsf T}\gamma_i+\mathbf{w}_i^{\mathsf T}\xi_i
\label{Eq:SNR}
\end{equation}
where $\beta_0$ represents a possible offset parameter, $f_j(\nu_{ij})$ and $\mathbf{z}_i^{\mathsf T}\gamma_i$ model the nonlinear and fixed effects aggregated at the $i$-th structural network level and $f(\alpha_s)$ encode coarser information recorded at a spatial lattice data level. Thus, apart from the spatial lattice information, all information collected in the structured network predictor is recorded at an identical level of structural network elements as $\lambda_i(\phi)$. The class of model with a predictor in form of \eqref{Eq:SNR} is called {\em{structured network regression model}}. Using a log-link yields to
\[
\lambda_i(\phi)=\exp(\eta^{\circ}_i).
\]
Inference for the structured network regression is carried out using an empirical Bayesian framework. Here, the unknown functions $f_j(\nu_{ij})$ and the spatial lattice information $f(\alpha_i)$ as well as the fixed parameters $\gamma$ and graph statistics $\xi$ in \eqref{Eq:SNR} are treated as random and are supplemented by priors.
For the fixed linear effects $\gamma$ and the graph statistics $\xi$ we consider flat priors such that $p(\gamma)$ and $p(\xi)$ are proportional to a constant $c$. For the unknown functions, the prior choice depends on the type of covariate $\nu_i$ and on smoothness assumptions. To define a general form of the prior for the unknown functions and the spatial lattice information, we reformulate \eqref{Eq:SNR} as
\begin{equation}
\boldsymbol{\eta^{\circ}}=\mathbf{\iota}\boldsymbol{\beta}_0+\sum^p_{j=1}\mathbf{X}_j\boldsymbol{\beta}_j+\mathbf{Z}\gamma+\mathbf{W}\xi
\label{eq:matrix:form}
\end{equation}
with $\mathbf{\iota}$ as a vector of ones and $\boldsymbol{\beta}_0$ as vector of possible offset parameters. $\mathbf{X}_j\boldsymbol{\beta}_j$ restates the vector of function evaluations $f_j=(f_j(\nu_{1j}),\ldots,f_j(\nu_{nj}))^{\mathsf T}$ in form of a matrix product. Here, $\mathbf{X}_j$ is a design matrix and $\boldsymbol{\beta}_j$ is a vector of unknown coefficients. The general form of the prior for the unknown functions and the spatial lattice information can then be expressed as
\begin{equation}\label{eq:genpriori}
p(\boldsymbol{\beta}_j|\sigma^2_j)\propto\exp\left(-\frac{1}{2\sigma_j^2}\boldsymbol{\beta}^{\mathsf T}_j\mathbf{K}_j\boldsymbol{\beta}_j\right).
\end{equation}
Here, $\sigma^2_j$ is a variance parameter which controls the trade-off between flexibility and smoothness and $\mathbf{K}_j$ is a penalty matrix.
As smoothness prior for the metrical covariates we consider P-splines which has been introduced by \cite{Eilers1996} in a frequentist setting. Here, the underlying assumption is that the unknown function $f_j$ can be approximated by means of a spline of degree $l$ defined on a set of equally space knots $\nu_j^{min} = \zeta_0 < \zeta_1 < \ldots < \zeta_{\tau-1} < \zeta_\tau = \nu_j^{max}$ within the domain of the covariate $\nu_j$. Such a spline can then be represented as a linear combination of $m=l+r$ $B$-spline basis functions $B_m$
\[
f_j(\nu_j)=\sum^m_{j=1}\boldsymbol{\beta}_{jm}B_m(x_j)
\]
where $\boldsymbol{\beta}_j$ is a vector of unknown regression coefficients. For a detailed discussion of alternative prior choice we refer the interested reader to \cite{Brezger2006}, \cite{Fahrmeir2004} and \cite{kneib2006}.
For the empirical Bayesian estimation we reparametrize the SNR model in terms of a generalized linear mixed models (GLMM).
As described in \cite{green1987} for splines and in \cite{Fahrmeir2004} and \cite{kneib2006} for the STAR model, we decompose $\boldsymbol{\beta}_j$ into a penalized ($p$) and a unpenalized part ($q$) such that $\boldsymbol{\beta}_j=\mathbf{X}^{(p)}_j\boldsymbol{\beta}_j^{(p)}+\mathbf{X}^{(q)}_j\boldsymbol{\beta}_j^{(q)}$. From this, we can rewrite $\boldsymbol{\beta}^{\mathsf T}_j\mathbf{K}_j\boldsymbol{\beta}_j$ of \eqref{eq:genpriori} as $\boldsymbol{\beta}_j^{(p){\mathsf T}}\boldsymbol{\beta}_j^{(p)}$. Then, setting $\mathbf{X}^{+}_j=\mathbf{X}_j\mathbf{X}^{(p)}_j$ and $\mathbf{X}^{-}_j=\mathbf{X}_j\mathbf{X}^{(q)}_j$ we can rewrite our structured network predictor as
\[
\eta^{\circ}_j=\mathbf{X}^{+}_j\boldsymbol{\beta}^{(p)}_j+\mathbf{X}^{-}_j\boldsymbol{\beta}^{(q)}_j.
\]
Here, as prior we assume $p(\beta_j^{(q)})\propto c$ and
$p(\beta_j^{(p)})\sim N(0,\sigma_j^2I))$.
Using this result, the estimation of the regression and variance parameters can be performed using GLMM techniques, namely iteratively weighted least squares (IWLS) and restricted maximum likelihood (REML).
\section{Application: urban disturbances-related data}\label{sec:apl}
For illustration of the SNR model, we considered the dependence of the nodewise mean intensity of neighbor and community disturbances on a set of various covariates and graph statistics.
The data was obtained from the local officials of the city of Castell\'{o}n (Spain) and contains the georeferenced coordinates of phone calls received by the police station as well as a set of $32$ additional covariate information. The listed calls were received at the local police call centre or transferred by 112 emergency service to the local police call centre. Geo-codification was performed indirectly by local officials based on precise address information provided by the caller. The calls comprise up to nine different types of crimes or anti-social behavior. From this data, we pre-selected a sample of $N=9790$ events classified as neighbor and community disturbances.
The city of Castell\'{o}n is divided into $108$ census sub-areas with an overall surface of $108659 km^2$. According to the information given by the city hall, the total amount of inhabitants is $181616$ of people at the end of 2010. Here, the analysis is based on a subset of phone calls received from the city center that has an overall surface of $8616 Km^2$ divided in $89$ census sub-areas and 130294 inhabitants.
For the analysis, we selected in total $1611$ segmenting locations of the traffic network treated as the vertex set of our network graph. So, the vertex set contains $1611$ single nodes, two isolated nodes were excluded. The corresponding traffic network is shown in Figure \ref{fig:1} where we depicted some events as black dots.
\begin{figure}[H]
\begin{center}
\includegraphics[width=3in]{castellon.eps}
\caption{Castell\'{o}n traffic network where solid pink lines indicate streets and events are plotted as black dots. \label{fig:1}}
\end{center}
\end{figure}
In this graph, $34$ nodes have a degree of one, $30$ nodes a degree of two and $368$ vertices a degree of three. Additionally, we observed four adjacent nodes that have been reported for $181$ locations while a vertex degree of five only appears once. The mean degree in this network is $3.14$. The graph consists of $21$ components. In addition, the length of the longest path is $64$.
To each vertex we attribute the precise georeferenced coordinates of the segmenting location. For any edge in the edge set we calculated the interval length as the squared geodesic distance between pairs of these coordinated vertices. Form this procedure we obtained nodewise mean intensity values for $614$ out of $1611$ vertices.
Similarly, we calculated nodewise mean values and proportion for all covariates. Besides, the degree and the betweenness centrality measures of the graph were also used as covariates. The betweenness centrality measure expresses the number of shortest paths passing through a certain node. In our graph we observed a mean betweenness of $18210$. The maximum betweenness was $204000$.
The betweenness centrality measures was also used to detect communities structures in the graph where we recursively extracted the edge with the highest betweenness value \citep[cf.][]{Newman2004}. This results in $45$ disjoint community groups.
In a first step, we perform a hierarchical cluster analysis to detect similarity structures within our data. A four cluster solution using the Ward algorithm is depicted in Figure \ref{fig:2} where the color red highlights nodes which were treated as missing values in our analysis.
\begin{figure}[H]
\begin{center}
\includegraphics[width=3in]{clusternet.eps}
\end{center}
\caption{Four cluster solution of covariate effects for the complete Castell\'{o}n network. Colour red indicates nodes for which the mean nodewise intensity has been zero. \label{fig:2}}
\end{figure}
Here, we observed that most of the calls which have been classified as neighbor and community disturbances were located in the city centre of Castell\'{o}n.
In a second step, we performed variable selection procedures using a generalized cross-validated Lasso and also classification and regression trees. Based on this analysis, we selected a covariate set of $5$ metric covariates and also the degree statistic. As continuous covariates we included the soil value indicator as well as the closest distances to pharmacies, parks, education and also health centres. The summary statistics of the nodewise means of all continuous covariates measures are reported in Table \ref{tab:1}.
\begin{table}
\caption{Summary statistics for the covariates of the SNR model. \label{tab:1}}
\begin{center}
\begin{tabular}{rrrrrrr}
& Min & 1st Q & Median & Mean & 3rd Q & Max \\ \hline
distance to the closest pharmacy & $7.62$ & $71.36$ & $107.15$ & $145.17$ & $169.48$ & $859.17$ \\
distance to the closest health center & $36.46$ & $214.41$ & $316.47$ & $350.68$ & $439.77$ & $1161.20$ \\
distance to the closest park & $0.00$ & $73.52$ & $127.37$ & $146.77$ & $199.86$ & $521.76$ \\
distance to the closest education center & $0.50$ & $106.80$ & $159.26$ & $185.61$ & $219.77$ & $719.72$ \\
soil value indicator &$171.0$ &$450.0$ & $537.0$ & $551.7$ & $632.0$ & $946.0$ \\
\end{tabular}
\end{center}
\end{table}
\subsection{Results of the SNR model}
We now discuss the results of our final SNR model. For selection, we implemented different SNR models and investigated the Akaike information criterion (AIC), Bayesian information criterion (BIC) and generalized cross-validation statistic (GCV) for model selection between competing models. As covariate set we considered the degree, the distance to the closest pharmacy, the distance to the closest park, the distance to the closest health center, the distance to the closest education center and the soil value indicator.
To evaluate the performance of the SNR model, we considered four alternative models ($mod1$ - $mod4$). For $mod1$ and $mod2$ we excluded the degree measures from the covariate set. As alternative, $mod3$ and $mod4$ were computed for all $6$ covariates. Additionally, we only considered P-splines for the distance to the closest park and the soil value indicator in $mod2$ and $mod4$. The AIC, BIC and GCV values for the null model and all four competitive models are reported in Table \ref{tab:one}.
\begin{table}
\caption{Information criteria and generalized cross-validation for different
models \label{tab:one}}
\begin{center}
\begin{tabular}{rrrr}
Model & AIC & BIC & GCV \\ \hline
$null$ & $ 1042.85$ & $1047.27$ & $1.66705$\\
$mod1$ & $876.094$ & $902.614$ & $1.40099$ \\
$mod2$ & $807.698$ &$896.986$ &$1.30024$\\
$mod3$ & $683.808$ &$728.008$ & $1.08252$\\
$mod4$ & $672.84$ &$737.003$ &$1.06472$\\
\end{tabular}
\end{center}
\end{table}
Here, we observed that the inclusion of the degree measure improves the model performance. When considering the AIC and the GCV statistics, we found that the inclusion of nonlinear terms in the SNR model leads to an improved fit.
Based on these results, we selected $mod4$ as our final model. Thus, our final model includes the vertex degree and the distance to the closest pharmacy, the distance to the closest health center and the distance to closest education center were chosen as fixed effects. The results for the fixed effects are reported in Table \ref{tab:2} where we chose a vertex degree of 1 as the reference category.
\begin{table}
\caption{Parametric coefficients of the SNR model. \label{tab:2}}
\begin{center}
\begin{tabular}{rrrrr}
& Estimate & Std. Error & t value & Pr($>|t|$) \\
\hline
(Intercept) & $1.88$ & $ 0.16$ &$ 12.08$ & $<2e-16$ \\
node degree = $2$ &$-0.82$ & $0.15$ &$-5.46 $ & $<2e-16$ \\
node degree = $3$ &$-1.24$ & $0.09$ &$-13.9$ & $<2e-16$ \\
node degree = $4$ &$-1.45$ & $0.11$ &$-13.39$ & $<2e-16$ \\
node degree = $5$ &$-3.22$ & $2.36$ &$ -1.36$ & $0.1740$ \\
distance to the closest pharmacy &$-0.0019$ & $0.0004$ &$-4.2171 $ & $<2e-16$ \\
distance to the closest health center &$-0.0006$ & $0.0002$ &$-2.8411 $ & $0.0046$ \\
distance to the closest education center &$ 0.0008$ & $0.0004$ &$ 2.0211 $ & $0.0437$ \\
\end{tabular}
\end{center}
\end{table}
Except for degree = $5$, we observe significant effects for all covariates.
For the impact of the degree measure, we observe that the nodewise mean intensity of neighbor and community disturbances decreases if the node degree increases. This might indicate that the subjective perception of neighbor and community disturbances strictly depends on the number of adjacent street segments in the Castell\'{o}n traffic net. This could mean that the subjective threshold of e.g. noise pollution is less strict for inhabitants of highly structured traffic areas, as these areas are commonly expected to be noisier.
A similar effect is shown for the nodewise mean distance to the closest pharmacy and the nodewise mean distance to the closest health center. Interestingly, for the nodewise mean distance to the closest education center, a positive effect is reported.
In addition, we modeled the effect of the distance to the closest park and the soil value indicator on the nodewise mean intensity of neighbor and community disturbances nonlinearly using P-splines. The estimated nonlinear effects for the covariates distance to the closest park and soil value indicator together with 80\% and 95\% credible intervals are visualized in Figure \ref{fig:3}.
\begin{figure}
\begin{center}
\includegraphics[width=3in]{Fig3.eps}
\end{center}
\caption{Nonparametric effects of the distance to the closest park and of the soil value indicator with pointwise 80\% and 95\% credible intervals. \label{fig:3}}
\end{figure}
Here, for the left panel we observed that the nonlinear effects of the distance to the closest park on our outcome variable followed an U-shape relationship. We found that the nodewise mean intensity of neighbor and community disturbances strongly decreases for distances up to 300 meters. Thereafter, a steadily increase is depicted.
A different effect is shown in the right panel of Figure \ref{fig:3}. Here, the smooth curve expressed an inverse U-shape impact on the nodewise mean intensity of neighbour and community disturbances.
\section{Conclusions}\label{sec:conclusion}
In this paper, we proposed a structured network regression model which provides a flexible toolbox for analyzing the impact of a set of different covariates and various graph statistics on generic intensity functions in the context of spatial network structures. This SNR model combines generalized structured additive regression models, graph theoretical statistics and the formalism of point patterns which occur on spatial network structures where generic intensity functions are treated as regressand. By doing so, the SNR model explicitly controls for the network structures which will not be addressed by classical regression techniques for spatial point patterns. Neither the network structure nor the structural relations contained in the graph will be captured by classical point pattern methodology and regression techniques.
The unified framework of the SNR model offers new insights in spatial point patterns that occur on complex domains. Various graph statistics and well-known concepts from network analysis can be chosen and integrated in a general regression framework. On the one hand, this leads to a strong increase of possible impact factors and various regression models. On the other hand, the unified framework provides several fit characteristics such as AIC, BIC or GCV for the evaluation of competitive models.
In general, the structured network predictor can be enlarged to allow for heterogeneity including random effect terms. One possibility would be to relate this random effects to the network structure such as edgewise random effects. Extensions to quantile, expectile or mode SNR models are straightforward. In addition, the estimation of the SNR model is performed using well-established fitting algorithms. Finally, the whole estimation could also been carried out applying a fully Bayesian approach and MCMC-techniques instead of the proposed REML and IWLS procedures.
\bibliographystyle{Chicago}
|
\section{Introduction} \label{Intro
\section{Introduction} \label{Intro}
\subsection{}The seminal conjecture of
Stark predicts that canonical elements constructed (unconditionally) from the leading terms at zero of the Artin $L$-series of complex linear characters should belong to
the rational vector spaces that are spanned by the $r$-th exterior powers of suitable groups of algebraic units, where $r$ denotes the order of vanishing at zero of the
relevant $L$-series.
It is believed that if Stark's conjecture is true, then these `Stark elements' should constitute a higher rank Euler system for the multiplicative group $\mathbb{G}_m$ over number fields and so there is considerable interest in studying their detailed arithmetic properties, especially the integrality.
The basic integral properties of Stark elements were first studied by Stark himself in \cite{stark4}, and then by Tate in \cite{tatebook}, for the case $r=1$ and then subsequently by Rubin in \cite{rubinstark} where the so-called `Rubin-Stark Conjecture' was formulated in the setting of general order of vanishing.
More recently, we formulated a strong refinement of the Rubin-Stark Conjecture in \cite[Conj. 7.3]{bks1}, obtained a natural interpretation of this conjecture in terms of a general theory of `arithmetic zeta elements' that was motivated by an earlier approach of Kato to the formulation of generalized Iwasawa main conjectures and derived a series of consequences of our conjecture concerning the detailed algebraic properties of number fields.
In this article we shall use the leading terms at arbitrary integer points of the $L$-series of linear characters to unconditionally define for non-negative integers $r$ and even integers $w$ canonical `(generalized) Stark elements of rank $r$ and weight $w$'. In this context the `rank' relates to the exterior power of the arithmetic module in which the element is constructed and the `weight' to the integer point $j_w := -w/2$ at which one takes the leading term of the $L$-series (so that $w$ is the weight of the associated motive $h^0(\Spec L)(j_w)$). In particular, in weight $0$, our construction recovers the classical theory of Rubin-Stark elements, and hence in weight $0$ and rank $1$ recovers the original constructions of Stark.
Our approach will show that generalized Stark elements of any fixed rank and weight should encode (in a very explicit way) detailed arithmetic information concerning the Galois structure of important \'etale cohomology groups.
In addition, our approach leads naturally to the simultaneous study of generalized Stark elements of differing weights and thereby introduces the perspective of $p$-adic families to the investigation. We remark that this philosophy of $p$-adic families has not hitherto been used in the setting of the Rubin-Stark Conjecture and is we feel worthy of further consideration.
To be a little more precise about the results proved here we fix an odd prime $p$ and a finite abelian extension $L/K$ of number fields. We assume that $L/K$ is
unramified outside a finite set of places $S$ of $K$ containing all places that are either archimedean or $p$-adic and we set $G := {\rm Gal}(L/K)$.
Then, as a first step, in Conjecture \ref{higherfitt} we predict that
generalized Stark elements of weight $w$ (and of appropriate rank) over $L$ explicitly determine the initial Fitting ideal of the \'etale cohomology group $H^2(\mathcal{O}_{L,S},\ZZ_p(1-j_w))$, regarded as a $\ZZ_p[G]$-module in the natural way.
This conjecture constitutes a natural extension of \cite[Conj. 7.3]{bks1} from the case of weight $0$ to the case of arbitrary weight. In addition, an interpretation of generalized Stark elements in terms of the theory of arithmetic zeta elements allows us to prove Conjecture \ref{higherfitt} for all absolutely abelian fields and for the minus part of CM-extensions of totally real fields (see Theorems \ref{evidence} and \ref{thm1}).
Next we write $p^{n}$ for the number of $p$-power roots of unity in $L$ and note that the Galois modules $H^2(\mathcal{O}_{L,S},\ZZ/p^{n}(1-j))(j-k)$ and $H^2(\mathcal{O}_{L,S},\ZZ/p^{n}(1-k))$ are isomorphic for any choice of integers $j$ and $k$. In particular, since Conjecture \ref{higherfitt} implies that Stark elements of weight $w$ over $L$ determine the initial Fitting ideal of the $(\ZZ/p^n\ZZ)[G]$-module $H^2(\mathcal{O}_{L,S},\ZZ/p^{n}(1-j_w))$, it suggests Stark elements of differing weights (and fixed rank) should be related by congruences modulo $p^{n}$.
In Conjecture \ref{congruence conjecture} we use algebraic techniques developed in \cite{bks1} to formulate, modulo the assumed validity of a weak version of Conjecture \ref{higherfitt}, a precise and explicit family of congruence relations between Stark elements of differing weights over arbitrary number fields $L$.
We show that this very general family of conjectural congruences recovers upon appropriate specialization a wide variety of results in the literature ranging from the classical congruences of Kummer concerning Bernoulli numbers to the results of Beilinson and H\"uber-Wildeshaus concerning the cyclotomic elements of Deligne-Soul\'e and a more recent conjecture of Solomon concerning certain `explicit reciprocity laws' for Rubin-Stark elements. These various connections allow us, in particular, to derive strong evidence, both theoretical and numerical, in support of Conjecture \ref{congruence conjecture} (see Theorem \ref{cong evidence} and Remark \ref{cong evid rem}).
The main contents of this article is as follows. In \S\ref{section2} we give the definition of generalized Stark elements and then in \S\ref{statement of conjs} we formulate our central conjectures concerning these elements (in Conjectures \ref{higherfitt} and \ref{congruence conjecture}). In \S\ref{zeta evidence} we give a natural interpretation of generalized Stark elements in terms of the theory of zeta elements and then use this interpretation to prove Conjecture \ref{higherfitt} in some important cases. Finally, in \S\ref{cong evi proof}, we relate special cases of Conjecture \ref{congruence conjecture} to well-known results in the literature and thereby deduce some supporting evidence for it.
\subsection{}\label{notation} For the reader's convenience we end the introduction by collecting together details concerning notation and conventions that are used in the sequel.
\subsubsection{Algebra} Let $E$ be a field of characteristic $0$. For any abelian group $A$, we denote $E\otimes_\ZZ A$ by $EA$. If $A$ is a $\QQ$-vector space, we sometimes denote $E\otimes_\QQ A$ also by $EA$. Similarly, if $E$ is an extension of $\QQ_p$ ($p$ is a prime number) and $A$ is a $\ZZ_p$-module, we denote $E\otimes_{\ZZ_p}A$ and $E\otimes_{\QQ_p} A$ also by $EA$.
For any integer $m$, we denote $A/mA$ simply by $A/m$.
For a commutative ring $R$ and an $R$-module $M$ we set
$M^\ast := \Hom_R(M,R)$.
If $M$ is a free $R$-module with basis $\{b_1,\ldots,b_r\}$, then for each $i$ with $1\leq i \leq r$ we write $b_i^\ast$ for the homomorphism $M\to R$ that sends $b_j$ to $1$ if $i=j$ and to $0$ if $i\neq j$.
For any field $E$, the absolute Galois groups is denoted by $G_E$. Let $c\in G_\RR$ denote the complex conjugation. For a $\ZZ[G_\RR]$-module $M$, let $M^\pm$ be the submodule $\{ a \in M \mid c\cdot a =\pm a\}$ of $M$. We also use the idempotent $e^\pm :=\frac{1\pm c}{2}$ of $\ZZ[\frac12][G_\RR]$ and the decomposition $M=M^+\oplus M^-$ with $M^\pm=e^\pm M$ for any $\ZZ[\frac12][G_\RR]$-module $M$.
\subsubsection{Arithmetic}
Fix an algebraic closure $\overline \QQ $ of $\QQ$. For any non-negative integer $m$, we denote by $\mu_{m}$ the subgroup of all $m$-th roots of unity in $\overline \QQ^\times$. As usual, we denote $\mu_{p^n}$ ($p$ is a prime number) by $\ZZ/p^n(1)$, and $\varprojlim_n \mu_{p^n}$ by $\ZZ_p(1)$. For any integer $j$, $\ZZ_p(j)$ and $\QQ_p(j)$ are defined in the usual way.
For a number field $K$, i.e. a finite extension of $\QQ$ in $\overline \QQ$, we write $S_\infty(K)$, $S_\CC(K)$ and $S_p(K)$ for the set of archimedean,
complex and $p$-adic places of $K$ respectively. We write $S_\infty$ for $S_\infty(K)$ if there is no danger of confusion. The ring of integers of $K$ is denoted by $\co_K$. For a finite set $S$ of places of $K$, the ring of $S$-integers of $K$ is denoted by $\co_{K,S}$. If $L$ is a finite extension of $K$, then the set of places of $K$ which ramify in $L$ is denoted by $S_{\rm ram}(L/K)$ and the set of places of $L$
lying above any given set of places $S$ of $K$ is denoted by $S_L$. The ring of $S_L$-integers of $L$ is denoted by $\co_{L,S}$ instead of $\co_{L,S_L}$.
Let $L/K$ be a finite abelian extension with Galois group $G$. Let $S$ and $T$ be finite disjoint sets of places of $K$ such that $S_\infty(K) \cup S_{\rm ram}(L/K) \subset S$.
Then, for a character $\chi \in \widehat G:=\Hom_\ZZ(G,\CC^\times)$, the $S$-truncated $T$-modified $L$-function is defined by
$$L_{K,S,T}(\chi,s):=\prod_{v\in T}(1-\chi({\rm Fr}_v) {\N}v^{1-s})\prod_{v\notin S}(1- \chi({\rm Fr}_v){\N}v^{-s})^{-1} \ ({\rm Re}(s)>1)$$
where ${\rm Fr}_v \in G$ is the Frobenius automorphism at a place of $L$ above $v$, and ${\N}v$ is the cardinality of the residue field $\kappa(v)$ of $v$. The function
$L_{K,S,T}(\chi,s)$ continues meromorphically to the whole complex plane and its leading term at an integer $j$ is denoted by
$L_{K,S,T}^\ast(\chi,j)$. The $S$-truncated $T$-modified $L$-function for $L/K$ is defined by setting
$$\theta_{L/K,S,T}(s):=\sum_{\chi \in \widehat G}L_{K,S,T}(\chi^{-1},s)e_\chi \,\,\,\text{ with } \,\,\, e_\chi:=\frac{1}{\# G}\sum_{\sigma \in G}\chi(\sigma) \sigma^{-1}$$
and has leading term at $s=j$ equal to $\theta_{L/K,S,T}^\ast(j):=\sum_{\chi \in \widehat G}L_{K,S,T}^\ast(\chi^{-1},j)e_\chi \in \CC[G]^\times$. When $T=\emptyset$, we omit it from notations (so we denote $L_{K,S,\emptyset}(\chi,s)$ by $L_{K,S}(\chi,s)$, for example). Note that
$$\theta_{L/K,S,T}(s)= \delta_{L/K,T}(s)\cdot \theta_{L/K,S}(s) $$
with $\delta_{L/K,T}(s) := \prod_{v \in T}(1-{\N}v^{1-s}{\rm Fr}_v^{-1})$.
\section{Generalized Stark elements} \label{section2}
\subsection{The general set up}\label{setup}
Let $L/K$ be a finite abelian extension of number fields. Set $G:=\Gal(L/K)$. Fix an {\it odd} prime number $p$. For each place $w$ of $L$,
we fix an algebraic closure $\overline L_w$ of $L$ and an embedding $\overline \QQ \hookrightarrow \overline L_w$. From this, we regard $G_{L_w}$ as a subgroup of $G_L$,
and the localization map of Galois cohomology $H^i(L, \cdot)\to H^i(L_w,\cdot)$ is defined by the restriction map. Also, for each
place $w$ in $S_\infty(L)$, we identify $\overline L_w$ with $\CC$.
For each integer $j$ we set
\[ S_{\infty}^j(L) := \begin{cases} S_\infty(L) &\text{if $j$ is even,}\\
S_\CC(L) &\text{if $j$ is odd,}\end{cases} \]
and note that
\begin{equation*}\label{first decomp} Y_L(j):=\bigoplus_{w\in S_\infty(L)} H^0(L_w,\ZZ_p(j))= \bigoplus_{w \in S^j_\infty(L)}\ZZ_p(j).\end{equation*}
In particular, setting $\xi:=(e^{2\pi \sqrt{-1}/p^n})_n \in \ZZ_p(1)$ one obtains a $\ZZ_p$-basis $\{w(j)\}_{w \in S^j_\infty(L)}$ of $Y_L(j)$,
which is defined by $w(j)=(w(j)_{w'})_{w'}$ where
$$w(j)_{w'}:=\begin{cases}
\xi^{\otimes j} &\text{if $w'=w$,}\\
0 &\text{if $w'\neq w.$}
\end{cases}
$$
Next we note that the complex conjugation $c$ in $G_\RR$ acts on the Betti cohomology
$$H_L(j):=H_B^0(\Spec L(\CC),\QQ(j)) =\bigoplus_{\iota: L\hookrightarrow \CC}(2\pi\sqrt{-1})^j \QQ$$
by $c\cdot (a_\iota)_\iota:=(c\cdot a_\iota)_{c\circ \iota}$ for each $a_\iota$ in
$(2\pi\sqrt{-1})^j\QQ$ and we set
$$H_L(j)^+:=e^+ H_L(j) \,\,\,\text{ with } \,\,\, e^+:=\frac{1+c}{2}.$$
Note that the natural decomposition $\CC= \RR \oplus \RR(-1)$ induces an isomorphism
\begin{equation} \label{period map}
\RR \otimes_{\QQ}L
\simeq \RR H_L(j)^+ \oplus \RR H_L(j-1)^+ .
\end{equation}
For each embedding $\iota': L \hookrightarrow \CC$ we define $\iota'_{j}=(\iota'_{j,\iota})_\iota$ in $H_L(j)$ by setting
$$\iota'_{j,\iota}:=\begin{cases}
(2\pi\sqrt{-1})^j &\text{if $\iota=\iota'$,}\\
0 &\text{if $\iota\neq \iota'.$}
\end{cases}
$$
Then, if for each place $w$ in $S_\infty(L)$ we write $\iota_w: L \to \CC$ for the embedding induced by the fixed embedding
$\overline \QQ \hookrightarrow \overline L_w=\CC$, we obtain an isomorphism of $\QQ_p[G]$-modules
\begin{eqnarray}
Y_L(j) \otimes_{\ZZ_p}\QQ_p \stackrel{\sim}{\to} H_L(j)^+ \otimes_\QQ \QQ_p \label{beta}
\end{eqnarray}
that sends each element $w(j)$ to $e^+\iota_{w,j}.$
For each place $v$ in $S_\infty(K)$ we now fix a place $w_v$ in $S_\infty(L)$ that lies above $v$ and set $S_\infty(L)/G := \{w_v \mid v \in S_\infty(K) \}$. For any idempotent $\varepsilon$ in $\ZZ_p[G]$ we set
$$W_j^\varepsilon:=\{ w \in S_\infty^j(L)\cap (S_\infty(L)/G) \mid \varepsilon \cdot w(-j) \neq 0 \}$$
and then define $r_j^\varepsilon:=\# W_j^\varepsilon$.
\begin{lemma} \label{lemma idempotent}
If $\varepsilon$ is a primitive idempotent of $\ZZ_p[G]$, then $\varepsilon Y_L(-j)$ is a free $\ZZ_p[G]\varepsilon$-module of rank $r_j^\varepsilon$ with basis $\{ \varepsilon \cdot w(-j) \mid w \in W_j^\varepsilon\}$.
\end{lemma}
\begin{proof}
For each $w \in S_\infty^j(L)$, the $\ZZ_p[G]$-submodule of $Y_L(-j)$ generated by $w(-j)$ is projective, since $p$ is odd. Thus, if $\varepsilon$ is a primitive idempotent (so that the ring $\ZZ_p[G]\varepsilon$ is local), then $ \ZZ_p[G]\varepsilon \cdot w(-j)$ is either zero or a free $\ZZ_p[G]\varepsilon$-module of rank one and so the decomposition
\[ \varepsilon Y_L(-j) = \bigoplus_{w\in S_\infty^j(L)\cap S_\infty(L)/G} \ZZ_p[G] \varepsilon\cdot w(-j)\]
implies that $\varepsilon Y_L(-j)$ is free with basis $\{ \varepsilon \cdot w(-j) \mid w \in W_j^\varepsilon\}$.
\end{proof}
The algebra $\ZZ_p[G]$ is semilocal and so every idempotent of $\ZZ_p[G]$ is
a sum of primitive idempotents. By Lemma \ref{lemma idempotent},
we may consider, without any loss of generality,
an idempotent satisfying the following condition.
\begin{hypothesis}\label{hypo}
$\varepsilon$ is an idempotent of $\ZZ_p[G]$ such that
the $ \ZZ_p[G]\varepsilon$-module $\varepsilon Y_L(-j)$ is free of rank $r_j^\varepsilon$ and has as basis the set $\{\varepsilon \cdot w(-j) \mid w \in W_j^\varepsilon\}$.\end{hypothesis}
\begin{example} \label{example} Suppose $K$ is totally real and $L$ is CM and write $c$ for the complex conjugation in $G$.
For each integer $j$ we obtain idempotents of $\ZZ_p[G]$ by setting $e_j^\pm := (1\pm (-1)^{j}c)/2$ and we abbreviate
$W_j^{e_j^\pm}$ to $W_j^\pm$. Then Hypothesis \ref{hypo} is satisfied in each of the following cases.
\begin{itemize}
\item[(i)] If $\varepsilon = e_j^+$, then $\varepsilon\cdot w(-j) = w(-j)$ for each $w$ in $S_\infty^j(L)\cap S_\infty(L)/G = S_\infty(L)/G$ and so we have $W_j^+=S_\infty(L)/G$ and $r_j^\varepsilon =
\# S_\infty(L)/G = \# S_\infty(K) = [K:\QQ]$.
\item[(ii)] If $\varepsilon = e_j^-$, then $\varepsilon\cdot w(-j) = 0$ for each $w$ in $S_\infty^j(L)$ so $W_j^-$ is empty and $r_j^\varepsilon =0$.
\end{itemize}
\end{example}
\subsection{The period-regulator isomorphisms}\label{period-regulator} In this section we assume the idempotent $\varepsilon$ satisfies Hypothesis \ref{hypo} with respect to $j$.
In the sequel we fix a finite set $S$ of places of $K$ which contains $S_\infty(K) \cup S_p(K) \cup S_{\rm ram}(L/K)$. We also fix (and do not explicitly mention) an isomorphism of fields $\CC \simeq \CC_p$.
We write $\widehat G^\varepsilon$ for the subset of $\widehat G$ comprising characters $\chi$ for which
$\varepsilon\cdot e_\chi \not= 0$.
We define a subset of $\widehat G^\varepsilon$ by setting
\[ \widehat G_j^\varepsilon := \{ \chi\in \widehat G^\varepsilon \mid {\rm dim}_{\CC_p}(e_\chi \CC_p H^1(\co_{L,S},\QQ_p(1-j))) = r_j^{\varepsilon} \}\]
and then obtain an idempotent of $ \QQ_p[G]\varepsilon$ by setting
\[ \varepsilon_j := \sum_{\chi \in \widehat G_j^{\varepsilon}}e_\chi .\]
\begin{remark}\label{exp idem} Lemma \ref{prop complex}(ii) below implies that for each
$\chi \in \widehat G^\varepsilon$ one has
\begin{eqnarray*}
\chi\in \widehat G^\varepsilon_j \Longleftrightarrow
\begin{cases}
e_\chi \CC_p H^2(\co_{L,S},\QQ_p(1-j))
\,\,\, \text{vanishes}, &\text{if $j \neq 1$},\\
e_\chi ( \CC_p \oplus \CC_pH^2(\co_{L,S},\QQ_p)) \,\,\, \text{vanishes}, &\text{if $j=1$}.
\end{cases}
\end{eqnarray*}
By using this description one can deduce the following facts.
\begin{itemize}
\item[(i)] If $j<0$, then $\widehat G^\varepsilon_j = \widehat G^\varepsilon$ (by Soul\'e \cite[Th. 10.3.27]{NSW}) and so $\varepsilon_j = \varepsilon$.
\item[(ii)] If $j=0$, then
\begin{eqnarray*}
\widehat G_0^\varepsilon &=& \{\chi \in \widehat G^\varepsilon \mid e_\chi \CC_p X_{L,S\setminus S_\infty} =0\}\\
&=& \{\chi\in \widehat G^\varepsilon \mid \ord_{s=0}L_{K,S}(\chi,s) = r^\varepsilon_0 \}.
\end{eqnarray*}
Here (and in the sequel), for any finite set $\Sigma$ of places of $K$ we write $X_{L,\Sigma}$ for the kernel of the homomorphism $\bigoplus_{w \in \Sigma_L} \ZZ_p w \to \ZZ_p$ sending each $w$ to $1$. The first displayed equality then follows by noting that, by class field theory, there is a canonical isomorphism $H^2(\co_{L,S},\QQ_p(1)) \simeq \QQ_p X_{L,S\setminus S_\infty}$, and the second equality follows directly from \cite[Chap. I, Prop. 3.4]{tatebook}.
\item[(iii)] If $j=1$, then Leopoldt's Conjecture for $L$ is equivalent to the vanishing of $H^2(\co_{L,S},\QQ_p)$ and hence implies that
$$
\widehat G^\varepsilon_1 = \{ \chi \in \widehat G^\varepsilon \mid \chi \neq {\bf 1} \},
$$
where we write ${\bf 1}$ for the trivial character of $G$, and so $\varepsilon_1 =\varepsilon (1-e_{\bf 1})$.
\item[(iv)] If $j>1$, then Schneider's Conjecture \cite{ps} for $L$ is equivalent to the vanishing of $H^2(\co_{L,S},\QQ_p(1-j))$ and hence implies $\widehat G^\varepsilon_j = \widehat G^\varepsilon$ and so $\varepsilon_j = \varepsilon$.
\end{itemize}
\end{remark}
In the remainder of this section we define for each integer $j$ a canonical isomorphism of $\CC_p[G]$-modules
\[ \lambda_{j} : \varepsilon_j\CC_p {\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j} H^1(\co_{L,S},\ZZ_p(1-j))
\stackrel{\sim}{\to} \varepsilon_j
\CC_p {\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j} Y_{L}(-j).\]
The explicit definition that we give is motivated by the Tamagawa number conjecture of Bloch and Kato (see, in particular, the proof of Corollary \ref{etnc proof} below).
\subsubsection{The case $j<0$} In this case the known validity of the Quillen-Lichtenbaum Conjecture (which follows from the recent proof by
Rost and Voevodsky of the Bloch-Kato Conjecture) gives a canonical Chern character isomorphism
\[ {\rm ch}_j: K_{1-2j}(\co_{L})\otimes_\ZZ\ZZ_p \stackrel{\sim}{\to} H^1(\co_{L,S},\ZZ_p(1-j))\]
where we write $K_*(-)$ for Quillen's higher algebraic $K$-theory functor.
One also has $\varepsilon_j = \varepsilon$ (by Remark \ref{exp idem}(i)) and
we define $\lambda_{j}$ to be the $r^\varepsilon_j$-th exterior power of the composite isomorphism of $\CC_p[G]$-modules
\begin{equation*}\label{regulator1} \varepsilon\CC_p H^1(\co_{L,S},\ZZ_p(1-j)) \stackrel{\sim}{\to} \varepsilon\CC_p K_{1-2j}(\co_L) \stackrel{\sim}{\to} \varepsilon\CC_p H_L(-j)^+ \stackrel{\sim}{\to}
\varepsilon\CC_p
Y_{L}(-j),\end{equation*}
where the first map is induced by the inverse of the isomorphism ${\rm ch}_j^{-1}$, the second by $(-1)$-times the Borel regulator map
\[ b_{j}: \RR K_{1-2j}(\co_L) \stackrel{\sim}{\to} \RR H_L(-j)^+\]
and the third by the isomorphism in (\ref{beta}).
\subsubsection{The case $j = 0$} We note that $ H^1(\co_{L,S},\ZZ_p(1))$ is identified with $\ZZ_p \co_{L,S}^\times$ via Kummer theory and we define
$\lambda_{0}$ to be the $r^\varepsilon_0$-th exterior power of the composite isomorphism of $\CC_p[G]$-modules
\begin{equation}
\varepsilon_0 \CC_p H^1(\co_{L,S},\ZZ_p(1)) = \varepsilon_0 \CC_p \co_{L,S}^\times \stackrel{\sim}{\to} \varepsilon_0 \CC_pX_{L,S} \stackrel{\sim}{\to} \varepsilon_0\CC_p Y_{L}(0)
\label{log}\end{equation}
where the first map is the restriction of the Dirichlet regulator (sending each $a$ in $\co_{L,S,T}^\times$ to $- \sum_{w \in S_L} \log |a|_w w$)
and the second isomorphism follows from the vanishing of $\varepsilon_0 \CC_p X_{L,S\setminus S_\infty}$ (see Remark \ref{exp idem}(ii)).
\subsubsection{The case $j = 1$} \label{subsection j=1}
We write $\Gamma_{L,S}$ for the Galois group of the maximal abelian pro-$p$ extension of $L$ unramified outside $S$. Then the module $H^1(\co_{L,S},\QQ_p)$ identifies with $\Hom_{\rm cont}(\Gamma_{L,S},\QQ_p)$ and so, by combining the
global class field theory with Remark \ref{exp idem}(iii), one obtains a canonical short exact sequence of $\CC_p[G]$-modules
\begin{equation}\label{j=1 iso}
\xymatrix{ \varepsilon_1 \CC_p H^1(\co_{L,S},\QQ_p) \ar@{^{(}->}[r] & \varepsilon_1{\bigoplus}_{w\in S_p(L)} (\CC_p \mathcal{O}_{L_w}^\times)^\ast \ar[d]^{\simeq}_{{\rm exp}^\ast_p} \ar@{->>}[r] & \varepsilon_1(\CC_p\mathcal{O}_L^\times)^\ast \\
& \varepsilon_1(\CC_p\otimes_{\QQ}L)^\ast & \varepsilon_1(\CC_p H_L(0)^+)^\ast .\ar[u]_{\simeq}
}
\end{equation}
Here the first vertical isomorphism is induced by the linear dual of the $p$-adic exponential map homomorphisms $L_w\to \QQ_p\mathcal{O}_{L_w}^\times$ for $w$ in $S_p(L)$ and the second by the linear dual of the isomorphism $\CC_p\mathcal{O}_L^\times \simeq \CC_p X_{L,S_\infty}$ induced by the Dirichlet regulator map and the fact that
$\varepsilon_1 \QQ_p X_{L,S_\infty}$ is equal to $\varepsilon_1\QQ_p Y_{L}(0)$ and hence isomorphic to $\varepsilon_1 \QQ_p H_L(0)^+$ by (\ref{beta}).
Abbreviating ${\rm det}_{\CC_p[G]}(-)$ to ${\rm D}(-)$, we then define $\lambda_{1}$ to be the composite isomorphism of $\CC_p[G]$-modules
\begin{multline*} \varepsilon_1 \CC_p{\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_1} H^1(\co_{L,S},\ZZ_p) = \varepsilon_1{\rm D}(\CC_p H^1(\co_{L,S},\ZZ_p))\\ \simeq \varepsilon_1\left({\rm D}((\CC_p\otimes_{\QQ}L)^\ast)\otimes_{\CC_p[G]}{\rm D}^{-1}((\CC_p H_L(0)^+)^\ast)\right)\\
\simeq \varepsilon_1 {\rm D}((\CC_p H_L(1)^+)^\ast) \simeq \varepsilon_1{\rm D}(\CC_pY_L(1)^\ast) \simeq \varepsilon_1 \CC_p{\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_1}Y_L(-1).\end{multline*}
Here the first isomorphism is the canonical isomorphism induced by (\ref{j=1 iso}), the second is induced by the linear dual of (\ref{period map}) (with $j =1$), the third by (\ref{beta}) and the last by the canonical identification $Y_L(1)^\ast \simeq Y_L(-1)$.
\subsubsection{The case $j>1$}
In this case the vanishing of $\varepsilon_j H^2(\co_{L,S},\QQ_p(1-j))$ (see Remark \ref{exp idem}) combines with the
local and global duality theorems to give a canonical short exact sequence of $\CC_p[G]$-modules
{\small
\begin{equation*}
\xymatrix{ \varepsilon_j \CC_p H^1(\co_{L,S},\ZZ_p(1-j)) \ar@{^{(}->}[r] & \ar[d]^{\simeq}_{{\rm syn}_{p}}\varepsilon_j \bigoplus_{w\in S_p(L)} \CC_pH^1(L_w,\ZZ_p(j))^\ast \ar@{->>}[r]
& \varepsilon_j \CC_p H^1(\co_{L,S},\ZZ_p(j))^\ast \\
& \varepsilon_j(\CC_p\otimes_{\QQ}L)^\ast & \varepsilon_j\CC_p H_L(j-1)^{+,\ast} \ar[u]_{\simeq} }
\end{equation*}
}
in which the second vertical homomorphism is induced by the dual of $-b_{1-j} \circ {\rm ch}_{1-j}^{-1}$ and the
first is induced by the linear duals for each $w$ in $S_p(L)$ of the canonical composite homomorphisms $L_w \to H^1_{\rm syn}(\mathcal{O}_{L_w},j) \to H^1(L_w,\QQ_p(j))$ involving syntomic cohomology that are discussed by Besser in \cite[(5.3) and Cor. 9.10]{besser}.
We then define $\lambda_{j}$ to be the isomorphism of $\CC_p[G]$-modules
obtained from the above diagram in just the same way that $\lambda_{1}$ is obtained from
(\ref{j=1 iso})
\begin{remark}\label{besser rem} In \cite[Prop. 9.11]{besser} Besser proves that for $w$ in $S_p(L)$ the composite homomorphism $L_w \to H^1_{\rm syn}(\mathcal{O}_{L_w},j) \to H^1(L_w,\QQ_p(j))$ used above coincides with the exponential map of Bloch and Kato for $\QQ_p(j)$ over $L_w$. In this way the definition of $\lambda_{j}$ for $j > 1$ is naturally analogous to the definition of
$\lambda_{1}$. \end{remark}
\begin{remark} \label{remark schneider}
A closer analysis of the discussions used to define $\lambda_{j}$ for $j>0$ shows that, in this case, if $\varepsilon Y_L(1-j)$ vanishes, then $\varepsilon_j=\varepsilon$.
\end{remark}
\subsection{The definition of generalized Stark elements}
\begin{definition}\label{gen stark def} Fix an integer $j$, an idempotent $\varepsilon$ of $\ZZ_p[G]$ that satisfies Hypothesis \ref{hypo} (with respect to $j$). Fix also finite sets $S$ and $T$ of places of $K$ satisfying
\begin{itemize}
\item $S_\infty(K)\cup S_p(K)\cup S_{\rm ram}(L/K) \subset S$;
\item $S\cap T=\emptyset$;
\item $T=\emptyset$ if $j=1$.
\end{itemize}
Then the `Stark element of rank $r^\varepsilon_j$ and weight $-2j$' for $(L/K,S,T,\varepsilon)$ is the unique element $\eta_{L/K,S,T}^{\varepsilon}(j)$ of
$\varepsilon_j\CC_p {\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j} H^1(\co_{L,S},\ZZ_p(1-j))$ that satisfies
\[ \lambda_{j}(\eta_{L/K,S,T}^{\varepsilon}(j)) = \varepsilon_j\theta_{L/K,S,T}^{\ast}(j)\cdot
{\bigwedge}_{w \in W_j^{\varepsilon}}w(-j) \ \ \text{ in }\ \ \varepsilon_j\CC_p{\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j}Y_L(-j). \
\end{definition}
\begin{remark}
It is natural to regard $\eta_{L/K,S,T}^\varepsilon(j)$ to be of weight $-2j$ since it is associated to the motive $h^0(\Spec L)(j)$.
\end{remark}
\begin{example}\label{example2} Definition \ref{gen stark def} generalizes the classical notion of Rubin-Stark element introduced by Rubin in \cite{rubinstark}. In fact, we have
$$\varepsilon_0\theta_{L/K,S,T}^{\ast}(0) =
\varepsilon \cdot \lim_{s \to 0 }s^{-r_0^\varepsilon}\theta_{L/K,S,T}(s)$$
by Remark \ref{exp idem}(ii) and \cite[Chap. I, Prop. 3.4]{tatebook} and so $\eta_{L/K,S,T}^{\varepsilon}(0)$ coincides with (the `$\varepsilon$-component' of) the Rubin-Stark element for the data
$(L/K,S,T,W_0^\varepsilon)$.
\end{example}
The following proposition is a natural analogue of \cite[Prop. 6.1]{rubinstark}.
\begin{proposition} \label{property stark}
Suppose that $(L'/K,S',T',\varepsilon')$ is another collection of data as in Definition \ref{gen stark def} (with respect to $j$) for which all of the following properties are satisfied: $L \subset L'$, $S\subset S'$, $T \subset T'$, with $G':=\Gal(L'/K)$ the natural surjection $\ZZ_p[G'] \to \ZZ_p[G]$ sends $\varepsilon'$ to $\varepsilon$ and $W_j^{\varepsilon}$ is the set of places of $L$ obtained by restricting places in $W_j^{\varepsilon'}$.
Then $r_j^{\varepsilon'}=r_j^\varepsilon=:r$ and the map
$$\varepsilon_j' \CC_p {\bigwedge}_{\ZZ_p[G']}^{r} H^1(\co_{L',S'},\ZZ_p(1-j)) \to \varepsilon_j\CC_p {\bigwedge}_{\ZZ_p[G]}^{r} H^1(\co_{L,S'},\ZZ_p(1-j))$$
induced by the corestriction map ${\rm Cor}_{L'/L}: H^1(\co_{L',S'}, \ZZ_p(1-j)) \to H^1(\co_{L,S'},\ZZ_p(1-j))$ sends $\eta_{L'/K,S',T'}^{\varepsilon'}(j)$ to
$$ \delta_{L/K,T'\setminus T}(j) \cdot \left( \prod_{v \in S' \setminus S}(1-{\N}v^{-j}{\rm Fr}_v^{-1}) \right) \cdot \eta_{L/K,S,T}^\varepsilon(j).$$
\end{proposition}
\begin{proof} This follows easily from the fact that the natural surjection $\CC[G'] \to \CC[G]$ sends $\varepsilon_j' \theta_{L'/K,S',T'}^\ast(j)$ to $$\varepsilon_j \theta_{L/K,S',T'}^\ast(j)=\varepsilon_j \delta_{L/K, T'\setminus T}(j) \cdot \left( \prod_{v \in S' \setminus S}(1-{\N}v^{-j}{\rm Fr}_v^{-1}) \right) \cdot \theta_{L/K,S,T}^\ast(j).$$
\end{proof}
\section{Statement of the conjectures}\label{statement of conjs}
\subsection{A Rubin-Stark Conjecture in arbitrary weight}
\subsubsection{Exterior power biduals and pairings} Fix a commutative ring $R$ and a finitely generated $R$-module $M$. In the following, we abbreviate $\bigwedge_R^r (M^\ast)$ to $\bigwedge_R^r M^\ast$.
For non-negative integers $r$ and $s$ with $r\le s$ there is a canonical pairing
$${\bigwedge}_R^s M \times {\bigwedge}_R^r M^\ast \to {\bigwedge}_R^{s-r} M$$
defined by
$$(a_1\wedge\cdots\wedge a_s,\varphi_1\wedge\cdots\wedge \varphi_r)\mapsto
\sum_{\sigma \in {\mathfrak{S}_{s,r}} }{\rm{sgn}}(\sigma)\det(\varphi_i(a_{\sigma(j)}))_{1\leq i,j\leq r} a_{\sigma(r+1)}\wedge\cdots\wedge a_{\sigma(s)} ,$$
with $\mathfrak{S}_{s,r}:= \{ \sigma \in \mathfrak{S}_s \mid \sigma(1) < \cdots < \sigma(r) \text{ and } \sigma(r+1) <\cdots<\sigma(s) \}.$ We denote the image of $(a,\Phi)$ under the above pairing by $\Phi(a)$.
We also use the following construction (compare \cite[\S 1.2]{rubinstark}).
\begin{definition} For each non-negative integer $r$ the $r$-th exterior bidual of $M$ is the module
$${\bigcap}_R^r M:=\left({\bigwedge}_R^r M^\ast \right)^\ast.$$
\end{definition}
\begin{remark} If $R=\ZZ_p[G]$ with a finite abelian group $G$, then the map $a \mapsto (\Phi \mapsto \Phi(a))$ induces an identification
\[ \left\{ a \in \QQ_p{\bigwedge}_{\ZZ_p[G]}^r M \ \middle| \ \Phi(a)\in \ZZ_p[G] \text{ for every $\Phi \in {\bigwedge}_{\ZZ_p[G]}^r M^\ast$} \right\} \simeq {\bigcap}_{\ZZ_p[G]}^r M\]
and so we may regard ${\bigcap}_{\ZZ_p[G]}^r M$ as a subset of $\QQ_p{\bigwedge}_{\ZZ_p[G]}^r M$. \end{remark}
\begin{lemma} \label{lemma phi}
Suppose that $R=\ZZ_p[G]$ with a finite abelian group $G$ and that $M$ is $\ZZ_p$-free. Let $H$ be a subgroup of $G$ and denote the natural surjection $\QQ_p[G] \to \QQ_p[G/H]$ by $\pi_H$. Then, for any $a \in \QQ_p \bigwedge_{\ZZ_p[G]}^r M$, we have
$$\pi_H\left( \left\{ \Phi(a) \ \middle| \ \Phi \in {\bigwedge}_{\ZZ_p[G]}^r M^\ast \right\}\right)=\left\{ \Psi ({\N}_H^r (a)) \ \middle| \ \Psi \in {\bigwedge}_{\ZZ_p[G/H]}^r (M^H)^\ast \right\},$$
where
$${\N}_H^r : \QQ_p {\bigwedge}_{\ZZ_p[G]}^r M \to \QQ_p {\bigwedge}_{\ZZ_p[G/H]}^r M^H$$
is the map induced by the norm map
$$M \to M^H; \ m \mapsto \sum_{\sigma \in H} \sigma \cdot m.$$
\end{lemma}
\begin{proof}
This follows from \cite[Rem. 2.9 and Lem. 2.10]{sano}.
\end{proof}
\subsection{$T$-modified cohomology} \label{ST} Let $j$ be an integer, and $S$ and $T$ sets of places of $K$ as in Definition \ref{gen stark def}.
Let now $R$ denote any of the rings $\ZZ_p$, $\QQ_p$ and $\ZZ/p^n$ for some
natural number $n$. Then, as $T$ is disjoint from $S$,
for each $w$ in $T_L$
there is a natural momorphism of \'{e}tale cohomology complexes
$R\Gamma(\co_{L,S},R(1-j)) \to R\Gamma(\kappa(w),R(1-j))$.
We define
$R\Gamma_{T}(\co_{L,S},R(1-j))$ to be a complex that lies in an exact triangle in the derived category $D(R[G])$ of complexes of $R[G]$-modules of the form
\begin{equation}\label{T-mod def} R\Gamma_{T}(\co_{L,S},R(1-j)) \to R\Gamma(\co_{L,S},R(1-j)) \to
\bigoplus_{w\in T_L}R\Gamma(\kappa(w),R(1-j)) \to , \end{equation}
where the second arrow is the diagonal map induced by the morphisms described above. In each degree $i$ we then set
\[ H^i_T(\co_{L,S},R(1-j)) := H^i(R\Gamma_{T}(\co_{L,S},R(1-j)))\]
and we note that $H_T^i(\co_{L,S},\QQ_p(1-j))=H^i(\co_{L,S},\QQ_p(1-j))$ (this follows from the fact that $R\Gamma(\kappa(w),\QQ_p(1-j))$ is acyclic if $j\neq 1$ and the assumption that $T=\emptyset$ if $j=1$). In particular, we can regard $\bigcap_{\ZZ_p[G]}^r H^1_T(\co_{L,S},\ZZ_p(1-j))$ as a lattice of $\QQ_p \bigwedge_{\ZZ_p[G]}^r H^1(\co_{L,S},\ZZ_p(1-j))$.
\begin{example} \label{ex ST}
Kummer theory identifies
$H_T^1(\co_{L,S},\ZZ_p(1))$ with the $p$-completion of the $(S,T)$-unit group $\co_{L,S,T}^\times:=\ker(\co_{L,S}^\times \to \bigoplus_{w\in T_L}\kappa(w)^\times)$
of $L$.
\end{example}
\subsubsection{Statement of the conjecture}
In the sequel for each non-negative integer $i$ we write ${\rm Fitt}_{\ZZ_p[G]}^i(M)$ for the $i$-th Fitting ideal of a $\ZZ_p[G]$-module $M$. We also write $I_G$ for the augmentation ideal of $\ZZ_p[G]$.
\begin{conjecture}\label{higherfitt} Fix an integer $j$, an idempotent $\varepsilon$ of $\ZZ_p[G]$ satisfying Hypothesis \ref{hypo} (with respect to $j$) and sets of places $S$ and $T$ as in Definition \ref{gen stark def}. Assume $\varepsilon H_T^1(\co_{L,S},\ZZ_p(1-j))$ is $\ZZ_p$-free and in addition that $\varepsilon \in I_G$ if $j = 1$.
Then, with $\eta = \eta_{L/K,S,T}^\varepsilon(j)$, one has
\begin{equation}\label{bks eq}
\left\{\!\Phi(\eta) \middle| \Phi\! \in\!
{\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j}H^1_T(\co_{L,S},\ZZ_p(1-j))^\ast\!\right\}\!=\!\varepsilon\cdot{\rm Fitt}_{\ZZ_p[G]}^{0}(H^2_{T}(\co_{L,S},\ZZ_p(1-j)))
\end{equation}
and hence also
\begin{equation}\label{rs inclusion} \eta \in {\bigcap}_{\ZZ_p[G]}^{r^\varepsilon_j} H^1_T(\co_{L,S},\ZZ_p(1-j)).\end{equation}
\end{conjecture}
\begin{remark}\label{injectivity}
One sees that $H^1(\co_{L,S},\ZZ_p)$ is always $\ZZ_p$-free. When $j \neq 1$, we see that
$\varepsilon H_T^1(\co_{L,S},\ZZ_p(1-j))$ is $\ZZ_p$-free if and only if the
composite map
\[ \varepsilon H^0(L,\QQ_p/\ZZ_p(1-j)) \to
\varepsilon H^1(\co_{L,S},\ZZ_p(1-j)) \to \varepsilon \bigoplus_{w\in T_L}H^1(\kappa(w),\ZZ_p(1-j))\]
is injective, where the first map is the natural boundary homomorphism.
\end{remark}
\begin{remark}
With the assumption that $\varepsilon \in I_G$ if $j=1$, we can use the complex $R\Gamma_T(\co_{L,S},\ZZ_p(1-j))[1] \oplus Y_L(-j)[-1]$ to construct an exact sequence of $\ZZ_p[G] \varepsilon$-modules of the form
$$0 \to \varepsilon H^1_T(\co_{L,S},\ZZ_p(1-j)) \to F \to F \to \varepsilon (H_T^2(\co_{L,S},\ZZ_p(1-j))\oplus Y_L(-j) )\to 0$$
where $F$ is both finitely generated and free (see \S \ref{zeta evidence}). This sequence is a natural analogue of classical `Tate sequences' (as discussed, for example, in \cite[\S 2.3]{bks1}) and plays a key role in our analysis. (For $j=1$ and the trivial character, a Tate sequence of similar kind is studied in \cite{GreitherK2} by Greither and the second author.)
\end{remark}
\begin{example} \label{example rubin stark}
If we identify $H_T^1(\co_{L,S},\ZZ_p(1))$ with the $p$-completion of the $(S,T)$-unit group $\co_{L,S,T}^\times$ of $L$ (see Example \ref{ex ST}) then in the case of $j =0$ the equality (\ref{bks eq}) recovers the `$\varepsilon$-component' of
the $p$-completion of \cite[Conj. 7.3]{bks1} and thus constitutes a refinement of a range of well-known conjectures
in the literature. For the same reason, in the setting of Example \ref{example2}, the $j=0$ case of the containment (\ref{rs inclusion}) recovers the
`$\varepsilon$-component' of the $p$-completion of the Rubin-Stark Conjecture \cite[Conj. B$'$]{rubinstark} for the data $(L/K,S,T,W_0^\varepsilon)$.
\end{example}
\begin{example} \label{minus example}Assume $K$ totally real, $L$ CM, $j\le 0$ and take $\varepsilon$ to be the idempotent $e_j^-$ in Example
\ref{example}(ii).
\noindent{}(i) In this case the inclusion (\ref{rs inclusion}) is unconditionally valid. To see this, note $r_j^\varepsilon = 0$ so
\[\eta_{L/K,S,T}^\varepsilon(j)=e^-_j\theta_{L/K,S,T}^\ast(j)=\theta_{L/K,S,T}(j)=\delta_{L/K,T}(j)\theta_{L/K,S}(j).\]
In addition, there is a natural exact sequence
\[ 0 \to \bigoplus_{v \in T} \ZZ_p[G] \xrightarrow{(1-{\N}v^{1-j}{\rm Fr}^{-1}_v)_v } \bigoplus_{v \in T}\ZZ_p[G] \to \bigoplus_{w\in T_L}H^1(\kappa(w),\ZZ_p(1-j)) \to 0\]
which implies
\begin{equation}\label{T equality} \delta_{L/K,T}(j)\cdot \ZZ_p[G]\varepsilon = \varepsilon {\rm Fitt}_{\ZZ_p[G]}^0\left(\bigoplus_{w\in T_L}H^1(\kappa(w),\ZZ_p(1-j)) \right)\end{equation}
and the assumed injectivity of the displayed map in Remark \ref{injectivity} implies this ideal is contained in $\varepsilon
{\rm Ann}_{\ZZ_p[G]}(H^0(L,\QQ_p/\ZZ_p(1-j)))$.
The claimed inclusion (\ref{rs inclusion}) thus follows from the fact that Deligne and Ribet \cite{DR} have shown that
$a \cdot \theta_{L/K,S}(j)$ belongs to $\ZZ_p[G]$ for any $a$ in ${\rm Ann}_{\ZZ_p[G]}(H^0(L,\QQ_p/\ZZ_p(1-j)))$.
\noindent{}(ii) If $j < 0$, then the conjectural equality (\ref{bks eq}) implies
\begin{multline*} \ZZ_p[G]\cdot\theta_{L/K,S,T}(j) =\varepsilon{\rm Fitt}_{\ZZ_p[G]}^0(H_T^2(\co_{L,S},\ZZ_p(1-j))) \\
\subset \varepsilon{\rm Fitt}_{\ZZ_p[G]}^0(H^2(\co_{L,S},\ZZ_p(1-j))) = \varepsilon\ZZ_p{\rm Fitt}_{\ZZ[G]}^0(K_{-2j}(\co_{L,S})),\end{multline*}
where the inclusion is true as $H^2(\co_{L,S},\ZZ_p(1-j))$ is a quotient of $H^2_T(\co_{L,S},\ZZ_p(1-j))$ and the
final equality because the validity of the
Quillen-Lichtenbaum Conjecture gives a canonical isomorphism $H^2(\co_{L,S},\ZZ_p(1-j)) \simeq \ZZ_p K_{-2j}(\co_{L,S})$. Noting that $K_{-2j}(\co_{L}) \subset K_{-2j}(\co_{L,S})$ (see \cite[Prop. 5.7]{CKPS}), this displayed inclusion shows (\ref{bks eq}) refines the classical Coates-Sinnott Conjecture, which predicts $\ZZ_p[G]\cdot \theta_{L/K,S,T}(j) \subset \ZZ_p {\rm Ann}_{\ZZ[G]}(K_{-2j}(\co_L)).$
\end{example}
In \S\ref{zeta evidence} we interpret generalized Stark elements in terms of the theory of arithmetic zeta elements and use this connection to obtain the following evidence in support of Conjecture \ref{higherfitt}.
\begin{theorem}\label{evidence}
Conjecture 3.5 is valid in both of the following cases.
\begin{itemize}
\item[(i)] $L$ is an abelian extension of $\QQ$.
\item[(ii)] $K$ is totally real, $L$ is CM, $j\le 0$, $\varepsilon$ is
the idempotent $e_j^-$ in Example \ref{example}(ii), and
the Iwasawa $\mu$-invariant vanishes for the cyclotomic
$\ZZ_{p}$-extension $L_{\infty}/L$.
\end{itemize}
\end{theorem}
We end this section by stating some functorial properties of Conjecture \ref{higherfitt}.
\begin{proposition} \label{functorial1}
Let $(L'/K,S',T', \varepsilon')$ be as in Proposition \ref{property stark}.
\begin{itemize}
\item[(i)] Suppose $S'=S$ and $T'=T$. Then (\ref{bks eq}) (resp. (\ref{rs inclusion})) for $(L'/K,S,T,\varepsilon' ,j )$ implies (\ref{bks eq}) (resp. (\ref{rs inclusion})) for $(L/K,S,T,\varepsilon,j )$.
\item[(ii)] Suppose that $L'=L$, $T'=T$ and $\varepsilon'=\varepsilon$. Then (\ref{rs inclusion}) for $(L/K,S,T,\varepsilon,j)$ implies that for $(L/K,S',T,\varepsilon,j)$.
\item[(iii)] Suppose that $L'=L$, $S'=S$ and $\varepsilon'=\varepsilon$. Then (\ref{rs inclusion}) for $(L/K,S,T,\varepsilon,j)$ implies that for $(L/K,S,T',\varepsilon,j)$.
\end{itemize}
\end{proposition}
\begin{proof}
We know that $\varepsilon' R\Gamma_T(\co_{L',S},\ZZ_p(1-j))$ is a perfect complex of $\ZZ_p[G']$-modules and acyclic outside degrees one and two (see Lemma \ref{prop complex} below), and that
$$R\Gamma_T(\co_{L',S},\ZZ_p(1-j))\otimes_{\ZZ_p[G']}^{\mathbb{L}} \ZZ_p[G] \simeq R\Gamma_T(\co_{L,S},\ZZ_p(1-j))$$
(see \cite[Prop. 1.6.5]{FK}, for example). From this
we see that
$$\varepsilon' H^1_T(\co_{L',S},\ZZ_p(1-j))^{\Gal(L'/L)} \simeq \varepsilon H^1_T(\co_{L,S},\ZZ_p(1-j))$$
and that
$$\varepsilon' H^2_T(\co_{L',S},\ZZ_p(1-j)) \otimes_{\ZZ_p[G']} \ZZ_p[G] \simeq \varepsilon H^2_T(\co_{L,S},\ZZ_p(1-j)).$$
Noting this, claim (i) follows from Proposition \ref{property stark} and Lemma \ref{lemma phi}.
Next, we show claim (ii). We have an exact sequence
$$0 \to H_T^1(\co_{L,S},\ZZ_p(1-j)) \to H^1_T(\co_{L,S'},\ZZ_p(1-j)) \to \bigoplus_{w \in (S'\setminus S)_L} H^1_{/f}(L_w, \ZZ_p(1-j)).$$
Since the last term is $\ZZ_p$-free, we see that the restriction map
$$H^1_T(\co_{L,S'},\ZZ_p(1-j))^\ast \to H^1_T(\co_{L,S},\ZZ_p(1-j))^\ast$$
is surjective. From this, we see that
$${\bigcap}_{\ZZ_p[G]}^r H^1_T(\co_{L,S},\ZZ_p(1-j)) \subset {\bigcap}_{\ZZ_p[G]}^r H^1_T(\co_{L,S'},\ZZ_p(1-j)).$$
Now the assertion in claim (ii) is clear by Proposition \ref{property stark}.
One can prove claim (iii) in the same way as \cite[Prop. 5.3.1]{popescuBC} by using the exact triangle (\ref{T-mod def}) and the equality (\ref{T equality}).
\end{proof}
\subsection{Congruences between Stark elements of differing weights}\ \
\
The `refined class number formula for $\mathbb{G}_m$', as independently conjectured by Mazur-Rubin \cite{MRGm} and the third author \cite{sano},
constitutes a family of congruence relations between Stark elements of weight zero and differing ranks. In this section we formulate precise
families of conjectural congruences between Stark elements of fixed rank and different weights.
At the outset we fix an integer $j$, an idempotent $\varepsilon$ of $\ZZ_p[G]$ satisfying Hypothesis \ref{hypo} (with respect to $j$) and sets of places $S$ and $T$ as in \S\ref{ST}.
We set $r := r^\varepsilon_j$ and $W:=W_j^\varepsilon$ . We also fix a labeling $W=\{w_1,\ldots,w_r\}$ and use this to define the wedge product ${\bigwedge}_{w \in W}$.
We fix a positive integer $n$ such that $\mu_{p^n}\subset L^\times$ and use the cyclotomic character
\[ \chi_{\rm cyc}: G \to \Aut(\mu_{p^n})\simeq (\ZZ/p^n)^\times.\]
For each integer $a$ we also write ${\rm tw}_a$ for the ring automorphism of $\ZZ/p^n[G]$ that sends
each element $\sigma$ of $G$ to $\chi_{\rm cyc}(\sigma)^a \sigma$. We then fix an integer $k$ and define $\delta$ to be the unique idempotent of $\ZZ_p[G]$ which projects to ${\rm tw}_{k-j}(\overline \varepsilon)$ in $\ZZ/p^n[G]$, where $\overline \varepsilon$ denotes the image in $\ZZ/p^n[G]$ of $\varepsilon$.
Then, by an explicit computation, one checks that $W^{\delta}_{k} = W$ (and hence $r_k^{\delta}=r$) and that $\delta Y_L(-k)$ is a free $\ZZ_p[G]\delta$-module of rank $r$ with basis $\{\delta \cdot w(-k) \mid w \in W\}$.
We next define a ${\rm tw}_{k-j}$-semilinear homomorphism
\begin{eqnarray*}
{\rm tw}_{j,k}: \varepsilon {\bigcap}_{\ZZ_p[G]}^r H^1_T(\co_{L,S},\ZZ_p(1-j)) \to \delta{\bigcap}_{\ZZ/p^n[G]}^r
H^1_T(\co_{L,S},\ZZ/p^n(1-k)) \label{twist lattice}
\end{eqnarray*}
that will play a key role in our conjectural congruences.
For simplicity, for each integer $a$ we set
$$H(\ZZ_p(a)):= H^1_T(\co_{L,S},\ZZ_p(a))\,\,\,\text{ and } \,\,\, H(\ZZ/p^n(a)):= H^1_T(\co_{L,S},\ZZ/p^n(a)).$$
Note that the natural map $H(\ZZ_p(1-j))\to H(\ZZ/p^n(1-j))$ induces a homomorphism
\begin{eqnarray}
{\bigcap}_{\ZZ_p[G]}^r H(\ZZ_p(1-j)) \to {\bigcap}_{\ZZ/p^n[G]}^r H(\ZZ/p^n(1-j)). \label{mod p^n}
\end{eqnarray}
Recall that each $w \in S_\infty(L)$ determines the embedding $\iota_{w}: L \hookrightarrow \CC$ (see \S \ref{setup}) and for
each integer $i$ with $1\leq i \leq r$, we set
\begin{eqnarray}
\xi_i:=\iota_{w_i}^{-1}(e^{2\pi\sqrt{-1}/p^n}) \in H^0(L,\ZZ/p^n(1)). \label{def xi}
\end{eqnarray}
We write
\begin{eqnarray*}
c_i: H(\ZZ/p^n(1-k))^\ast \to H(\ZZ/p^n(1-j))^\ast \label{def ci}
\end{eqnarray*}
for the map induced by cup product with $\xi_i^{\otimes (j-k)}$ and
$${\bigwedge}_{\ZZ/p^n[G]}^r H(\ZZ/p^n(1-k))^\ast \to {\bigwedge}_{\ZZ/p^n[G]}^r H(\ZZ/p^n(1-j))^\ast$$
for the map sending each element $a_1\wedge\cdots \wedge a_r$ to $c_1(a_1)\wedge \cdots \wedge c_r(a_r).$
Taking the $\ZZ/p^n$-dual of the last map we obtain a homomorphism
$${\bigcap}_{\ZZ/p^n[G]}^r H(\ZZ/p^n(1-j)) \to {\bigcap}_{\ZZ/p^n[G]}^r H(\ZZ/p^n(1-k))$$
and we define ${\rm tw}_{j,k}$ to be the composite of this homomorphism with (\ref{mod p^n}).
\begin{conjecture}\label{congruence conjecture} Fix an integer $j$, an idempotent $\varepsilon$ of $\ZZ_p[G]$ satisfying Hypothesis \ref{hypo} (with respect to $j$) and sets of places $S$ and $T$ as in \S\ref{ST}. Assume that the integer $k$, and associated idempotent $\delta$ defined above, are such that
\begin{itemize}
\item $\varepsilon H^1_T(\co_{L,S},\ZZ_p(1-j))$ and $\delta H^1_T(\co_{L,S},\ZZ_p(1-k))$ are both $\ZZ_p$-free;
\item $T=\emptyset$ if either $j=1$ or $k=1$;
\item $\varepsilon\in I_G$ if $j=1$;
\item $\delta\in I_G$ if $k=1$.
\end{itemize}
Then, if the containment (\ref{rs inclusion}) is valid for both pairs $(\varepsilon,j)$ and
$(\delta, k)$, one has
$${\rm tw}_{j,k}(\eta_{L/K,S,T}^{\varepsilon}(j)) = \eta_{L/K,S,T}^{\delta}(k)$$
in the finite module $\delta{\bigcap}_{\ZZ/p^n[G]}^r H^1_T(\co_{L,S},\ZZ/p^n(1-k))$.
\end{conjecture}
We discuss evidence for this conjecture in \S\ref{cong evi proof} and, in particular, prove the following result. This result (and its proof)
shows that the conjecture incorporates a wide selection of results ranging from classical explicit reciprocity law due to Artin-Hasse and Iwasawa
to the classical congruences of Kummer. In this result we use the notation of Example \ref{example}.
\begin{theorem}\label{cong evidence} Assume $K$ is totally real and $L$ is CM.
\begin{itemize}
\item[(i)] If $K = \QQ$, then for all integers $j$ and $k$ Conjecture \ref{congruence conjecture} is valid with $\varepsilon = e^+_j$.
\item[(ii)] For all non-positive integers $j$ and $k$
Conjecture \ref{congruence conjecture} is valid with $\varepsilon = e_j^-$.
\item[(iii)] Conjecture \ref{congruence conjecture} for the data $T = \emptyset, j = 0$, $\varepsilon = e^+$ and
$k = 1$ is a refinement of the `Congruence Conjecture' \cite[{${\rm CC}(L/K,S,p,n-1)$}]{solomon} of Solomon.
\end{itemize}
\end{theorem}
\begin{remark}\label{cong evid rem} The proof of claim (i) relies both (if $j < 0$) on results of Beilinson and H\"uber-Wildeshaus on the cyclotomic elements of
Deligne-Soul\'e and (if $j > 0$) on Kato's generalized explicit reciprocity law, whilst claim (ii) relies on results of Deligne and Ribet.
Claim (iii) is of interest both because Solomon's Conjecture (recalled in \S\ref{solomon section} below) is formulated as an explicit reciprocity
law for Rubin-Stark elements (extending that of Artin-Hasse and Iwasawa \cite{iwasawa}) and also because it allows us to interpret
the extensive numerical evidence in support of Solomon's conjecture in \cite{rs} as evidence for Conjecture \ref{congruence conjecture}.\end{remark}
\begin{remark} In a subsequent article we intend to explore connections between Conjecture \ref{congruence conjecture} and the very general
(conjectural) formalism discussed by Fukaya and Kato in \cite{FK}.\end{remark}
\begin{proposition} \label{functorial2}
Let $(L'/K,S',T',\varepsilon')$ be as in Proposition \ref{property stark}.
\begin{itemize}
\item[(i)] Suppose $S'=S$ and $T'=T$. Then Conjecture \ref{congruence conjecture} for $(L'/K,S,T,\varepsilon',j,k)$ implies that for $(L/K,S,T,\varepsilon,j,k)$.
\item[(ii)] Suppose that $L'=L$ and $\varepsilon'=\varepsilon$. Then Conjecture \ref{congruence conjecture} for $(L/K,S,T,\varepsilon,j,k)$ implies that for $(L/K,S',T',\varepsilon,j,k)$.
\end{itemize}
\end{proposition}
\begin{proof}
Note that, since $\mu_{p^n} \subset L$, ${\rm tw}_{j,k}$ is a $\ZZ_p[\Gal(L'/L)]$-homomorphism. Note also that ${\N}_{L'/L}^r (\eta_{L'/K,S,T}^{\varepsilon'}(j))=\eta_{L/K,S,T}^\varepsilon(j)$ and ${\N}_{L'/L}^r (\eta_{L'/K,S,T}^{\delta'}(j))=\eta_{L/K,S,T}^\delta(j)$ by Proposition \ref{property stark}, where $\N_{L'/L}^r=\N_{\Gal(L'/L)}^r$. Assuming Conjecture \ref{congruence conjecture} for $(L'/K,S,T,\varepsilon',j,k)$, we have
\begin{multline*}
{\rm tw}_{j,k}(\eta_{L/K,S,T}^\varepsilon(j))={\rm tw}_{j,k}({\N}_{L'/L}^r (\eta_{L'/K,S,T}^{\varepsilon'}(j))) \\
= {\N}_{L'/L}^r ({\rm tw}_{j,k}(\eta_{L'/K,S,T}^{\varepsilon'}(j))) \equiv {\N}_{L'/L}^r(\eta_{L'/K,S,T}^{\delta'}(j) ) = \eta_{L/K,S,T}^\delta(j) \text{ (mod $p^n$)}.
\end{multline*}
Hence we have proved claim (i).
Since we have
$${\rm tw}_{k-j}(\delta_{L/K,T'\setminus T}(j)) \equiv \delta_{L/K,T'\setminus T}(k) \text{ (mod $p^n$)}$$
and
$${\rm tw}_{k-j}\left(\prod_{v \in S'\setminus S}(1-{\N} v^{-j} {\rm Fr}_v^{-1})\right) \equiv \prod_{v \in S'\setminus S}(1-{\N} v^{-k} {\rm Fr}_v^{-1}) \text{ (mod $p^n$)},$$
claim (ii) follows from the fact that ${\rm tw}_{j,k}$ is ${\rm tw}_{k-j}$-semilinear, and Propositions \ref{property stark} and \ref{functorial1}.
\end{proof}
\begin{proposition} Suppose that $v \notin S\cup T$ splits completely in $L$, and assume (\ref{rs inclusion}) for both $(L/K,S, T,\varepsilon,j)$ and $(L/K,S, T, \delta,k)$. Then Conjecture \ref{congruence conjecture} is valid for $(L/K,S\cup\{v\},T,\varepsilon,j,k)$.
\end{proposition}
\begin{proof} If $v$ is any such place, then $p^n $ divides $\#\kappa(v)^\times={\N}v-1$ (since $\mu_{p^n}\subset L$) and hence also both $(1-{\N}v^{-j})$ and $(1-{\N}v^{-k})$.
The stated assumptions and Proposition \ref{property stark} therefore imply that
\[ \eta_{L/K,S\cup\{v\},T}^\varepsilon(j)=(1-{\N}v^{-j})\eta_{L/K,S,T}^\varepsilon(j)\in p^n\cdot {\bigcap}_{\ZZ_p[G]}^r H^1_T(\co_{L,S},\ZZ_p(1-j))\]
and that
\[ \eta_{L/K,S\cup \{v\},T}^{\delta}(k)=(1-{\N}v^{-k})\eta_{L/K,S,T}^{\delta}(k) \in p^n\cdot
{\bigcap}_{\ZZ_p[G]}^r H^1_T(\co_{L,S},\ZZ_p(1-k))\]
and so both sides of the displayed equality in Conjecture \ref{congruence conjecture} for $(L/K,S\cup\{v\}, T,\varepsilon,j,k)$ vanish.
\end{proof}
\section{Zeta elements and the proof of Theorem \ref{evidence}}\label{zeta evidence}
In this subsection, we interpret generalized Stark elements in terms of the theory of arithmetic zeta elements and use this connection to prove Theorem \ref{evidence}.
\subsection{Perfect complexes}
Let $\varepsilon \in \ZZ_p[G]$ be any idempotent (we do not need to assume Hypothesis \ref{hypo} in this subsection). With $Z$ denoting either $\ZZ_p$ or $\ZZ/p^n$ for some natural number $n$ we define an object of $D(Z[G]\varepsilon)$ by setting
\[ C_{L,S,T}^{\varepsilon}(j)_Z := Z[G]\varepsilon \otimes^{\mathbb{L}}_{\ZZ_p[G]}( R\Gamma_{T}(\co_{L,S},\ZZ_p(1-j))[1] \oplus Y_L(-j)[-1] ) .\]
The properties of these complexes that we use are recorded in the following result
We write $D^{\rm perf}(Z[G]\varepsilon)$ for the full triangulated subcategory of
$D(Z[G]\varepsilon)$ comprising complexes that are `perfect' (that is, isomorphic to a bounded complex of finitely generated projective $Z[G]\varepsilon$-modules).
\begin{lemma} \label{prop complex} The following claims are valid for all integers $j$.
\begin{itemize}
\item[(i)] $C_{L,S,T}^{\varepsilon}(j)_Z$ belongs to $D^{\rm perf}(Z[G]\varepsilon)$ and is acyclic outside degrees $-1, 0$ and $1$.
\item[(ii)] Assume that $\varepsilon H^1_T(\co_{L,S},\ZZ_p(1-j))$ is $\ZZ_p$-free if $Z=\ZZ/p^n$ and that $\varepsilon\in I_G$ if $j=1$. Then we have
$$H^{i}(C_{L,S,T}^\varepsilon(j)_Z)=\begin{cases}
0 &\text{if $i=-1$},\\
\varepsilon H^1_T(\co_{L,S},Z(1-j)) &\text{if $i=0$},\\
\varepsilon H^2_T(\co_{L,S},Z(1-j)) \oplus \varepsilon (Y_L(-j)\otimes_{\ZZ_p}Z) &\text{if $i=1$}.
\end{cases}$$
Furthermore, we have a (non-canonical) isomorphism of $\QQ_p[G]$-modules
$$\QQ_p H^0(C_{L,S,T}^{\varepsilon}(j)_{\ZZ_p}) \simeq \QQ_p H^1(C_{L,S,T}^{\varepsilon}(j)_{\ZZ_p}).$$
\end{itemize}
\end{lemma}
\begin{proof} Since $p$ is odd, it is well-known that $R\Gamma(\co_{L,S},\ZZ_p(1-j))$ belongs to $D^{\rm perf}(\ZZ_p[G])$ and is acyclic outside
degrees zero, one and two (see, for example, \cite[Prop. 1.6.5]{FK}). Claim (i) follows from this and the
fact that the complex $\bigoplus_{w\in T_L}R\Gamma(\kappa(w),Z(1-j))$ in the triangle (\ref{T-mod def}) belongs to
$D^{\rm perf}(Z[G])$ and is acyclic outside degrees zero and one.
To prove the first assertion of claim (ii) it suffices to show $\varepsilon H^0_T(\co_{L,S}, Z(1-j))$ vanishes under the stated assumptions.
If $j\neq 1$, then $H^0(\co_{L,S}, \ZZ_p(1-j))$, and hence also, $H_T^0(\co_{L,S}, \ZZ_p(1-j))$ vanishes. If $\varepsilon H_T^1(\co_{L,S}, \ZZ_p(1-j))$ is $\ZZ_p$-free, then the exact triangle
$$R\Gamma_T(\co_{L,S},\ZZ_p(1-j)) \stackrel{p^n}{\to} R\Gamma_T(\co_{L,S},\ZZ_p(1-j)) \to R\Gamma_T(\co_{L,S},\ZZ/p^n(1-j)) \to$$
implies $\varepsilon H_T^0(\co_{L,S}, \ZZ/p^n(1-j))$ vanishes.
Next, we consider the case when $j=1$. Recall that we set $T=\emptyset$ in this case (see \S \ref{ST}). Since $\varepsilon \in I_G$ by assumption, we have
$$\varepsilon H^0_T(\co_{L,S},Z)=\varepsilon H^0(\co_{L,S},Z)=\varepsilon \cdot Z=0.$$
To prove the remaining assertion of claim (ii) we write $R\Gamma_c(\co_{L,S},\ZZ_p(j))$ for the compactly supported cohomology complex of $\ZZ_p(j)$ and note that Artin-Verdier duality (as expressed, for example, in \cite[(6)]{BFetnc2}) combines with the triangle (\ref{T-mod def}) to give a
canonical exact triangle in $D^{\rm perf}(\ZZ_p[G])$ of the form
\begin{equation*}\bigoplus_{w\in T_L}R\Gamma(\kappa(w),\ZZ_p(1-j)) \to C_{L,S,T}^{1}(j)_{\ZZ_p} \to R\Hom_{\ZZ_p}(R\Gamma_c(\co_{L,S},\ZZ_p(j)),
\ZZ_p)[-2] \to . \end{equation*}
Then, since $C_{L,S,T}^{\varepsilon}(j)_{\ZZ_p}$ is acyclic outside degrees zero and one, the final assertion of claim (ii) follows from this triangle and the fact that the $\QQ_p[G]$-equivariant Euler characteristics of both $\varepsilon \bigoplus_{w\in T_L}R\Gamma(\kappa(w),\QQ_p(1-j))$ and $R\Gamma_c(\co_{L,S},\QQ_p(j))$ vanish. \end{proof}
\subsection{Zeta elements} We quickly review the definition of zeta elements in the context of Conjecture \ref{higherfitt}. To do this we fix
notation $L/K,G,p,S,T, j, \varepsilon$ and $\varepsilon_j$ as in \S \ref{section2}. We often abbreviate $C_{L,S,T}^\varepsilon(j)_{\ZZ_p}$ to $C_{L,S,T}^{\varepsilon}(j)$. When $\varepsilon=1$, we omit it from notations (so we denote $C_{L,S,T}^1(j)$ by $C_{L,S,T}(j)$, for example).
The definition of $\varepsilon_j$ combines with Lemma \ref{prop complex}(ii) to imply $\QQ_p[G] \varepsilon_j \otimes_{\ZZ_p[G]}^\mathbb{L} C_{L,S,T}(j) $
is acyclic outside degrees zero and one and that there are canonical isomorphisms
\[ \varepsilon_j\QQ_p H^i(C_{L,S,T}(j))) \simeq \begin{cases}
\varepsilon_j H_T^1(\co_{L,S},\QQ_p(1-j)) &\text{if $i = 0$,}\\
\varepsilon_j\QQ_p Y_{L}(-j) &\text{if $i = 1$.}\end{cases}\]
Since these $\QQ_p[G]\varepsilon_j$-modules are both free of rank $r^\varepsilon_j$ there is a canonical `passage to cohomology' isomorphism
of $\QQ_p[G]\varepsilon_j$-modules
\begin{multline}\label{passage}
\pi_j: \varepsilon_j \QQ_p {\det}_{\ZZ_p[G]}(C_{L,S,T}(j)) \\
\xrightarrow{\sim} \varepsilon_j \QQ_p \left( {\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j} H_T^1(\co_{L,S},\ZZ_p(1-j)) \otimes_{\ZZ_p[G]}
{\bigwedge}_{\ZZ_p[G]}^{r^\varepsilon_j} Y_{L}(j) \right) .\end{multline}
Here we identify $Y_{L}(j)$ with $Y_L(-j)^\ast$.
\begin{definition} \label{definition zeta element}
The zeta element associated to the data $(L/K,S,T,\varepsilon,j)$ is the unique element $z_{L/K,S,T}^{\varepsilon}(j)$ of
$\varepsilon_j\CC_p {\det}_{\ZZ_p[G]}(C_{L,S,T}(j))$ that satisfies
$$\pi_j(z_{L/K,S,T}^\varepsilon(j))= \eta_{L/K,S,T}^\varepsilon(j) \otimes {\bigwedge}_{w \in W_j^{\varepsilon}} w(j),$$
or equivalently,
\[ ({\rm ev}_L \circ (\lambda_{j} \otimes {\rm id}) \circ \pi_j)(z_{L/K,S,T}^\varepsilon(j)) =
\varepsilon_j \theta_{L/K,S,T}^\ast(j), \]
where ${\rm ev}_{L}$ denotes the standard `evaluation' isomorphism
\[ {\bigwedge}_{\CC_p[G]}^{r^\varepsilon_j}\CC_pY_{L}(-j) \otimes_{\CC_p[G]}
{\bigwedge}_{\CC_p[G]}^{r^\varepsilon_j} \CC_p Y_{L}(-j)^\ast
\simeq \CC_p[G]. \]
\end{definition}
\subsection{The proof of Theorem \ref{evidence}} In this section we prove the following results.
\begin{theorem} \label{thm1} If there exists a $\ZZ_p[G] \varepsilon$-basis $z$ of $\varepsilon {\det}_{\ZZ_p[G]}(C_{L,S,T}(j))$ with
$\varepsilon_j z = z_{L/K,S,T}^\varepsilon(j)$, then Conjecture \ref{higherfitt} is valid.
\end{theorem}
\begin{corollary}\label{etnc proof} Theorem \ref{evidence} is valid.
\end{corollary}
\begin{proof} The first point to note is that the maps that are used in the explicit definition of the isomorphism $\lambda_{j}$ given in \S\ref{period-regulator} coincide with the maps that occur in the statement of the equivariant Tamagawa number conjecture for the pair $(h^0(\Spec L)(j),\ZZ_p[G]\varepsilon)$ (see \cite[Conj. 4(iv)]{BFetnc}). This fact is clear if $j \le 1$ and follows in the case $j > 1$ from the result of Besser recalled in Remark \ref{besser rem}.
Given this, and our definition of the element $z_{L/K,S,T}^\varepsilon(j)$, the latter conjecture implies the existence of
a $\ZZ_p[G]\varepsilon$-basis of $\varepsilon\cdot{\det}_{\ZZ_p[G]}(C_{L,S,T}(j))$ with the property stated in Theorem \ref{thm1}. We note that this conjecture is usually formulated without using the set $T$, but as noted in \cite[Prop. 3.4]{bks1} one can formulate a natural $T$-modified version of this conjecture, whose validity is independent of the choice of $T$.
The result of Theorem \ref{evidence}(i) now follows directly from Theorem \ref{thm1} and the fact that if $L$ is abelian over $\QQ$, then the equivariant Tamagawa number conjecture for
$(h^0(\Spec L)(j),\ZZ_p[G])$ is known to be true (by work of the first author and Greither \cite{BG}, and of Flach \cite{flach2part}).
Theorem \ref{evidence}(ii) can be proved by the same method as in \cite[Cor. 3.18]{bks2}
by using the Iwasawa main conjecture proved by Wiles.
\end{proof}
The proof of Theorem \ref{thm1} occupies the rest of this section (and is motivated by the argument used to prove \cite[Th. 7.5]{bks1}). We assume that $\varepsilon H^1_T(\co_{L,S},\ZZ_p(1-j))$ is $\ZZ_p$-free and that $\varepsilon \in I_G$ if $j=1$.
We set $\mathcal{A} := \ZZ_p[G]\varepsilon$, $A := \QQ_p[G]\varepsilon$, $W:=W^\varepsilon_j$ and $r := r^\varepsilon_j$. We also label (and thereby order) the elements of $W$ as $\{w_i\}_{1\le i\le r}$.
Then Lemma \ref{prop complex}(ii) implies that $C_{L,S,T}^\varepsilon(j)$ is acyclic outside degrees zero and one and
we can therefore choose a representative of $C_{L,S,T}^\varepsilon(j)$ of the form $F \stackrel{\psi}{\to} F$
with $F$ a free $\mathcal{A}$-module with basis $\{b_1,\ldots,b_d\}$ for some sufficiently large integer $d$
so that
the natural surjection
\[ F \to \coker (\psi) = H^1(C_{L,S,T}^{\varepsilon}(j))= \varepsilon H^2_T(\co_{L,S},\ZZ_p(1-j)) \oplus \varepsilon Y_L(-j) \]
sends $b_i$ with $1\leq i \leq r$ to $\varepsilon \cdot w_i(-j)$ and $\{b_{r+1},\ldots ,b_d\}$ to a set of generators of $\varepsilon H^2_T(\co_{L,S},\ZZ_p(1-j))$.
See \cite[\S 5.4]{bks1} for the detail of this construction. Note that the representative chosen in loc. cit. is of the form $P \to F$ with $P$ projective and $F$ free, but in the present case we can identify $P$ with $F$ by Swan's theorem (see \cite[(32.1)]{CR}). Also, note that the assumption that $\varepsilon H^1_T(\co_{L,S},\ZZ_p(1-j))$ is $\ZZ_p$-free is needed here.
We may therefore identify ${\det}_\mathcal{A}(C_{L,S,T}^{\varepsilon}(j))$ with ${\bigwedge}_\mathcal{A}^d F \otimes_\mathcal{A}
{\bigwedge}_\mathcal{A}^d F^\ast$. With respect to this identification, any
$\mathcal{A}$-basis of ${\det}_\mathcal{A}(C_{L,S,T}^{\varepsilon}(j))$ has the form
\[ z_x := x \cdot b_1\wedge\cdots\wedge b_d \otimes b_1^\ast \wedge \cdots \wedge b_d^\ast \]
with $x \in \mathcal{A}^\times$, where we write $b_i^\ast$ for the $\mathcal{A}$-linear dual of $b_i$.
Next we write
\[ \pi_j': \QQ_p {\rm det}_\mathcal{A}( C_{L,S,T}^{\varepsilon}(j)) \to \varepsilon_j \QQ_p {\rm det}_\mathcal{A}( C_{L,S,T}^{\varepsilon}(j)) \stackrel{\sim}{\to}
\varepsilon_j{\bigwedge}_{\QQ_p[G]}^{r}H_T^1(\co_{L,S},\QQ_p(1-j))\]
for the composite homomorphism of $A$-modules in which the first map is `multiplication by $\varepsilon_j$' and the second is the composite of the isomorphism $\pi_j$ in (\ref{passage}) and
the isomorphism of $A$-modules $\varepsilon_j \QQ_p{\bigwedge}_{\ZZ_p[G]}^{r} Y_{L}(j) \stackrel{\sim}{\to} A\varepsilon_j$ that sends the element
$\varepsilon_j \cdot w_1(j) \wedge \cdots \wedge w_r(j)$ to $\varepsilon_j$.
Then, with this notation, the argument of \cite[Lem. 4.3]{bks1} implies that
\begin{eqnarray}\label{first step}\pi_j'( z_x) &=&(-1)^{r(d-r)}x \left( {\bigwedge}_{r< i \leq d}\psi_i \right)(b_1\wedge\cdots \wedge b_d) \\
&=& (-1)^{r(d-r)} x \sum_{\sigma \in \mathfrak{S}_{d,r}}
{\rm sgn}(\sigma)\det(\psi_i(b_{\sigma(k)}))_{r < i,k\leq d} b_{\sigma(1)}\wedge \cdots \wedge b_{\sigma(r)}\nonumber
\end{eqnarray}
with $\psi_i:=b_i^\ast \circ \psi \in F^\ast$ for each index $i$.
In particular, note that the element $( {\bigwedge}_{r< i \leq d}\psi_i )(b_1\wedge\cdots \wedge b_d) $ of ${\bigwedge}_\mathcal{A}^r F$ lies in $\varepsilon_j{\bigwedge}_{\QQ_p[G]}^{r}H_T^1(\co_{L,S},\QQ_p(1-j))$, which is regarded as a submodule of $\QQ_p \bigwedge_\mathcal{A}^r F$ via the inclusion
\begin{eqnarray} \label{psi inclusion}
\varepsilon H_T^1(\co_{L,S},\ZZ_p(1-j))=H^0(C_{L,S,T}^\varepsilon(j)) =\ker \psi \hookrightarrow F.
\end{eqnarray}
Next we note that the matrix of the endomorphism $\psi$ with respect to the basis $\{b_1,\ldots,b_d\}$ of $F$ is $(\psi_i(b_k))_{1\leq i,k\leq d}$ and
that $\psi_i=0$ for each $i$ with $1\leq i \leq r$ since the elements $\{ \varepsilon \cdot w_i(-j)\}_{1\le i\le r}$ are an $\mathcal{A}$-basis of $\varepsilon Y_{L}(-j)$. The
matrices $\{ \det(\psi_i(b_{\sigma(k)}))_{r<i,k\leq d}\}_{\sigma \in \mathfrak{S}_{d,r}}$ are therefore a set of generators of the $\mathcal{A}$-module
\[ {\rm Fitt}_\mathcal{A}^r(H^1(C_{L,S,T}^{\varepsilon}(j))) = {\rm Fitt}_\mathcal{A}^0(\varepsilon H_T^2(\mathcal{O}_{L,S},\ZZ_p(1-j))) = \varepsilon {\rm Fitt}_{\ZZ_p[G]}^0(H^2_T(\mathcal{O}_{L,S},\ZZ_p(1-j))),\]
where the first equality is valid since the $\mathcal{A}$-module $H^1(C_{L,S,T}^{\varepsilon}(j))$ is the direct sum of $\varepsilon H_T^2(\mathcal{O}_{L,S},\ZZ_p(1-j))$ and a free module $\varepsilon Y_L(-j)$ of rank $r$.
Note that the restriction map $F^\ast \to \varepsilon H_T^1(\co_{L,S},\ZZ_p(1-j))^\ast$ is surjective since the cokernel of (\ref{psi inclusion}) is $\ZZ_p$-free. This fact combines with the equality (\ref{first step})
to imply that
\begin{multline}\label{first version}
\left\{\Phi(\pi_j'(z_x)) \ \middle| \ \Phi\! \in\! \varepsilon {\bigwedge}_{\ZZ_p[G]}^r H^1_T(\mathcal{O}_{L,S},\ZZ_p(1-j))^\ast \right\} \\=
\varepsilon {\rm Fitt}_{\ZZ_p[G]}^0(H^2_T(\mathcal{O}_{L,S},\ZZ_p(1-j))).
\end{multline}
Now suppose that $\varepsilon_j \cdot z_x= z_{L/K,S,T}^\varepsilon(j)$. Then the definition of $z_{L/K,S,T}^\varepsilon(j)$ implies
$\pi_j'(z_x) =\eta_{L/K,S,T}^\varepsilon(j)$ and, given this,
the result of Theorem \ref{thm1} follows directly from the equality (\ref{first version}).
\section{Some evidence for Conjecture \ref{congruence conjecture}}\label{cong evi proof}
In this section, we give a proof of Theorem \ref{cong evidence}. Throughout this section we assume $K$ totally real and $L$ CM.
\subsection{Deligne-Soul\'e elements, explicit reciprocity and Theorem \ref{cong evidence}(i)} In this subsection we assume $K=\QQ$ and $\varepsilon=e_j^+$. In this case we have $r:=r_j^{\varepsilon}=\# S_\infty(L)/G=\# S_\infty(\QQ)=1$ (see Example \ref{example}(i)).
By Proposition \ref{functorial2}, we may assume $L =\QQ(\mu_f)$ with $f \in \ZZ_{>0}$ such that $f \not\equiv 2 \ (\text{mod } 4)$. Also, we may assume $p^n \mid f$ and that $S$ is equal to the minimal set $\{\infty\}\cup \{\ell \mid f\}$, with $\infty$ the archimedean place of $\QQ$. Finally, we may assume $T=\emptyset$ since $p$ is odd. We note that $\varepsilon \in I_G$ is satisfied when $j=1$, and that $\varepsilon_j=\varepsilon$ holds when $j\neq 0$ (see Remarks \ref{exp idem}(i) and \ref{remark schneider}). In the following, we often omit $T$. (For example, we denote $\eta_{L/\QQ,S,\emptyset}^\varepsilon(j)$ by $\eta_{L/\QQ,S}^\varepsilon(j)$.) For simplicity, we denote $\eta_{L/\QQ,S}^{e_j^+}(j)$ by $\eta_{L/\QQ,S}^+(j)$.
Recall that $w \in S_\infty(L)/G$ determines the embedding $\iota_w : \overline \QQ \hookrightarrow \CC $ (see \S \ref{setup}). We set $\zeta_m:=\iota_w^{-1}(e^{2\pi \sqrt{-1}/m})$ for any integer $m$.
Recall the definition of `cyclotomic elements' of Deligne-Soul\'e.
First, for a positive integer $m$, define
$$c_{1-j}(\zeta_f)_m:={\rm Cor}_{\QQ(\mu_{p^m f})/L}((1-\zeta_{p^mf} )\otimes \zeta_{p^m}^{\otimes (-j)})\in H^1(\co_{L,S},\ZZ/p^m(1-j)). $$
Here we regard $(1-\zeta_{p^m f} )\otimes \zeta_{p^m}^{\otimes (-j)}$ as an element of $H^1(\ZZ[\mu_{p^m f},1/p],\ZZ/p^m (1-j))$ via the Kummer map
\begin{eqnarray*}
\ZZ \left[\mu_{p^m f}, \frac1p \right]^\times \otimes \ZZ/p^m (-j)
&\to& H^1\left(\ZZ\left[\mu_{p^m f},\frac1p \right],\ZZ/p^m (1) \right)\otimes_\ZZ \ZZ/p^m (-j)\\
&\simeq& H^1\left(\ZZ\left[\mu_{p^m f},\frac1p\right],\ZZ/p^m (1-j)\right).
\end{eqnarray*}
The cyclotomic element is defined by the inverse limit
$$c_{1-j}(\zeta_f):=\varprojlim_m c_{1-j}(\zeta_f)_m \in \varprojlim_m H^1(\co_{L,S},\ZZ/p^m(1-j)) \simeq H^1(\co_{L,S},\ZZ_p(1-j)). $$
Noting that
$$\eta_{L/\QQ,S}^{+}(0)=2^{-1} \cdot (1-\zeta_f)(1-\zeta_f^{-1}) \in \ZZ_p \co_{L^+,S}^\times=e^+ H^1(\co_{L,S},\ZZ_p(1))$$
(see Example \ref{example2} and \cite[p.79]{tatebook}),
we see by definition that ${\rm tw}_{0,j}(\eta_{L/\QQ,S}^{+}(0))=e_j^+ c_{1-j}(\zeta_f)_n$. From this, we have
$${\rm tw}_{j,k}(e_j^+ c_{1-j}(\zeta_f))=e_k^+ c_{1-k}(\zeta_f)_n$$
for arbitrary integers $j$ and $k$.
Hence, it is sufficient to show that
\begin{eqnarray}
e_j^+c_{1-j}(\zeta_f)= \eta_{L/\QQ,S}^{+}(j). \nonumber
\end{eqnarray}
We may assume $j \neq 0$. Suppose first that $j <0$. In this case, by the definition of $\eta_{L/\QQ,S}^{+}(j)$, it is sufficient to show that the image of $e_j^+ c_{1-j}(\zeta_f)$ under the isomorphism
$$\lambda_{j}: e_j^+\CC_p H^1(\co_{L,S},\ZZ_p(1-j)) \simeq e_j^+\CC_p K_{1-2j}(\co_L) \simeq e_j^+\CC_p Y_L(-j),$$
is $e_j^+\theta_{L/\QQ,S}^\ast(j) \cdot w(-j)$. This is a direct consequence of the results of Beilinson and H\"uber-Wildeshaus \cite[Cor. 9.7]{HW} (see also \cite[Th. 5.2.1 and 5.2.2]{HK}).
Next, suppose $j >0$. Again, it is sufficient to show that the image of $e_j^+ c_{1-j}(\zeta_f)$ under the isomorphism $\lambda_{j}$ is $e_j^+ \theta_{L/\QQ,S}^\ast(j) \cdot w(-j)$. Recalling Remark \ref{besser rem}, we note that in this case $\lambda_{j}$ coincides with the map
$$e_j^+\CC_p H^1 (\co_{L,S},\ZZ_p(1-j)) \stackrel{{\rm exp}_p^\ast}{\simeq} e_j^+( \CC_p \otimes_\QQ L)^\ast \stackrel{\alpha_j^\ast}{\simeq} e_j^+ \CC_p H_L(j)^{+,\ast}\simeq e_j^+\CC_p Y_L(-j),$$
where ${\rm exp}_p^\ast$ is the dual exponential map, $\alpha_j^\ast$ is induced by (\ref{period map})
and the last isomorphism is induced by (\ref{beta}) with the identification $Y_L(j)^\ast=Y_L(-j)$.
By using the explicit reciprocity law due to Kato \cite[Th. 5.12]{katoiwasawa} (see also \cite[Chap. II, Th. 2.1.7]{katolecture} and \cite[Th. 3.2.6]{HK}), we have
$${\rm exp}_p^\ast (c_{1-j}(\zeta_f))=\left(x \mapsto -\frac{1}{f^j} {\rm Tr}_{L/\QQ}(x d_j(\zeta_f))\right) \in L^\ast,$$
where $d_j$ is the polylogarithmic function
$$d_j(t):=\frac{(-1)^j}{(j-1)!} {\rm Li}_{1-j}(t)=\frac{(-1)^j}{(j-1)!} \left(\frac1t \frac{d}{dt}\right)^{j-1}\left(\frac{t}{1-t}\right). $$
From this and the classical formula
$$e_j^+\theta_{L/\QQ,S}^\ast(j)=\frac14 \left(\frac{2\pi \sqrt{-1}}{f}\right)^j \sum_{1\leq a \leq f, \ (a,f)=1}(d_j(e^{2\pi \sqrt{-1}a/f})+(-1)^j d_j (e^{-2\pi \sqrt{-1}a/f})) \sigma_a^{-1},$$
where $\sigma_a \in \Gal(L/\QQ)$ is the automorphism sending $\zeta_f$ to $\zeta_f^a$, we see by computation that
$\alpha_j^\ast\circ {\rm exp}_p^\ast(e_j^+ c_{1-j}(\zeta_f))=e_j^+ \theta_{L/\QQ,S}^\ast(j) \cdot w(-j)$ and this completes the proof.
\subsection{Generalized Kummer congruences and Theorem \ref{cong evidence}(ii)} In the setting of Theorem \ref{cong evidence}(ii) one has $\eta_{L/K,S,T}^{e_j^-}(j)=\theta_{L/K,S,T}(j)$ for any non-positive integer $j$ (see Example \ref{minus example}) and $r_j^{e_j^-}=0$ and the map ${\rm tw}_{j,k}$ coincides with the composite homomorphism
$$e_j^-\ZZ_p[G] \to e_j^- \ZZ/p^n[G] \stackrel{{\rm tw}_{k-j}}{\to} e_k^- \ZZ/p^n[G]$$
which sends each $\sigma \in G$ to $\chi_{\rm cyc}(\sigma)^{k-j}\sigma$. Hence, it is sufficient to show that
$${\rm tw}_j(\theta_{L/K,S,T}(0)) \equiv \theta_{L/K,S,T}(j) \text{ (mod $p^n$)}$$
for any non-positive integer $j$.
But, since ${\rm tw}_j(\delta_{L/K,T}(0)) \equiv \delta_{L/K,T}(j) \text{ (mod $p^n$)},$ the above congruence follows directly from the well-known result due to Deligne-Ribet \cite{DR} that for any element $a$ of ${\rm Ann}_{\ZZ_p[G]}(H^0(L,\QQ_p/\ZZ_p(1)))$ one has ${\rm tw}_j(a \cdot \theta_{L/K,S}(0)) \equiv {\rm tw}_j(a) \theta_{L/K,S}(j) \text{ (mod $p^n$)}.$
\begin{remark} The above argument shows that Conjecture \ref{congruence conjecture} constitutes a wide-ranging extension of the classical Kummer congruences. To see this take $K=\QQ$, $L=\QQ(\mu_{p^n})$, $S=\{\infty,p\}$ and $T=\emptyset$. Then write $\Delta$ for the subgroup of $G$ of order $p-1$ and set
$e_{\Delta}:=\frac{1}{p-1}\sum_{\sigma \in \Delta} \sigma \in \ZZ_p[G].$
Let $j,k$ be odd negative integers such that $j \equiv k$ (mod $p^{n-1}(p-1)$) and $1-j\not\equiv 0$ (mod $p-1$). The first condition implies that ${\rm tw}_{j-k}$ is the identity map, and the second that $e_{\Delta}H^0(L,\QQ_p/\ZZ_p(1-j))$ vanishes and hence that $e_\Delta H^1(\co_{L,S},\ZZ_p(1-j))$ is $\ZZ_p$-free. The same holds for $e_\Delta H^1(\co_{L,S},\ZZ_p(1-k))$ since $j \equiv k \text{ (mod $p-1$)}$.
By Theorem \ref{cong evidence}(ii), we deduce $e_{\Delta}\theta^\ast_{L/\QQ,S}(j) \equiv e_\Delta \theta^\ast_{L/\QQ,S}(k) \text{ (mod $p^n$)}$ and hence $(1-p^{-j})\zeta(j) \equiv (1-p^{-k})\zeta(k) \text{ (mod $p^n$)},$ where $\zeta(s)$ denotes the Riemann zeta function. This is exactly the formulation of Kummer's congruence.
\end{remark}
\subsection{Solomon's Congruence Conjecture and Theorem \ref{cong evidence}(iii)}\label{solomon section} In this subsection we first review the explicit statement of Solomon's Conjecture and then prove Theorem \ref{cong evidence}(iii).
\subsubsection{}To review Solomon's conjecture we set $L_p:=L\otimes_\ZZ \ZZ_p \simeq \prod_{w \in S_p(L)}L_w$ and $U_{L_p}^1:=(\co_{L}\otimes_\ZZ \ZZ_p)^\times \otimes_\ZZ \ZZ_p \simeq \prod_{w \in S_p(L)}U_{L_w}^1$, where $U_{L_w}^1$ denotes the group of principal units of $L_w$.
As in \S \ref{subsection j=1}, we denote by $\Gamma_{L,S}$ the Galois group of the maximal abelian pro-$p$ extension of $L$ unramified outside $S$.
We denote $\Gal(L^+/K)$ by $G^+$. We write $S_\infty(L)/G=\{w_1,\ldots,w_r\}$. Note that $r=[K:\QQ]$.
Note that $L/K$ corresponds to $K/k$ in \cite{solomon}. In \cite[\S 2.2]{solomon}, a representative $\{\tau_1,\ldots,\tau_r\}$ of the coset space $G_\QQ/G_K$ is fixed. To fit this choice into our setting, we set
$$\tau_i:=\iota_{w_1}^{-1}\circ \iota_{w_i}, $$
where each $\iota_{w_i}$ is regarded as an embedding $\overline \QQ \hookrightarrow \overline L_{w_i}=\CC$ (see \S \ref{setup}).
In \cite{solomon}, the algebraic closure of $\QQ$ is considered to be in $\CC$. This means that an embedding $\overline \QQ \hookrightarrow \CC$ is fixed. We take $\iota_{w_1}$ for this fixed choice.
Also, an embedding $j: \overline \QQ \hookrightarrow \overline \QQ_p$ is chosen in \cite[\S 2.4]{solomon}. Since we fixed an isomorphism $\CC \simeq \CC_p$, we take $j$ to be the composition $\overline \QQ \stackrel{\iota_{w_1}}{\hookrightarrow} \CC \simeq \CC_p.$
In \cite[Def. 2.14]{solomon}, Solomon defined a map
$$\mathfrak{s}_{L/K,S} \in \Hom_{\ZZ_p[G]^-}\left({\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-},\QQ_p[G]^-\right)$$
by using the zeta value $e^-\theta_{L/K,S}^\ast(1)$ (we will give the definition in the proof of Proposition \ref{solomon stark} below). Solomon's Integrality Conjecture \cite[{${\rm IC}(L/K,S,p)$}]{solomon} is equivalent to the containment
\begin{eqnarray}
\mathfrak{s}_{L/K,S} \in \Hom_{\ZZ_p[G]^-}\left({\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-}, \ZZ_p[G]^-\right). \nonumber
\end{eqnarray}
Next, we explain the formulation of Solomon's Congruence Conjecture. Assume that $\mu_{p^n} \subset L$. In \cite[\S 2.3]{solomon}, Solomon constructed a pairing
$$H_{L/K,S,n-1}: {\bigcap}_{\ZZ_p[G^+]}^r \ZZ_p\co_{L^+,S}^\times \times {\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-} \to \ZZ/p^n[G]^-$$
by using the Hilbert symbol $L_w^\times \times L_w^\times \to \mu_{p^n}$ for $p$-adic places $w\in S_p(L)$ (see the proof of Proposition \ref{solomon stark} below).
Assuming the validity of (\ref{rs inclusion}) for the data $(L/K,S,\emptyset,e^+,0)$ (or equivalently, the $p$-part of the Rubin-Stark Conjecture for the data $(L^+/K,S,\emptyset, S_\infty(K))$, see Example \ref{example rubin stark}), Solomon's Congruence Conjecture \cite[{${\rm CC}(L/K,S,p,n-1)$}]{solomon} asserts that for every $u \in {\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-}$ one has
$$\mathfrak{s}_{L/K,S}(u) \equiv (-1)^r \chi_{\rm cyc}(\tau_1\cdots \tau_r)H_{L/K,S,n-1}(\eta_{L/K,S}^{e^+}(0),u) \text{ (mod $p^n$)}.$$
(The sign $(-1)^r$ appears here since we use $(-1)$-times the usual logarithm
(see (\ref{log}))).
\subsubsection{}We now prove Theorem \ref{cong evidence}(iii). To do this we note that $H^1(\co_{L,S},\ZZ_p)^- = \Hom_{\rm cont}(\Gamma_{L,S}^-,\ZZ_p)$. The dual of the map ${\rm rec}_p: U_{L_p}^1 \to \Gamma_{L,S}$ that sends $u$ to $\prod_{w \in S_p(L)} {\rm rec}_w(u),$
where ${\rm rec}_w$ denotes the local reciprocity map at $w$, therefore induces a homomorphism
\begin{multline*}
{\rm rec}^\ast_p : {\bigcap}_{\ZZ_p[G]^-}^r H^1(\co_{L,S},\ZZ_p)^- \to {\bigcap}_{\ZZ_p[G]^-}^r \Hom_{\ZZ_p}(U_{L_p}^{1,-},\ZZ_p)\\
\simeq \Hom_{\ZZ_p[G]^-} \left({\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-},\ZZ_p[G]^-\right)^\#,
\end{multline*}
where for a $\ZZ_p[G]$-module $M$ we denote by $M^\#$ the module $M$ on which $G$ acts via the involution $\sigma \mapsto \sigma^{-1}$, and the last isomorphism follows from
$$\Hom_{\ZZ_p}(U_{L_p}^{1,-},\ZZ_p) \simeq \Hom_{\ZZ_p[G]^-}(U_{L_p}^{1,-},\ZZ_p[G]^-)^\#; \ f \mapsto \sum_{\sigma\in G}f(\sigma(\cdot))\sigma^{-1}$$
and the definition of $r$-th exterior bidual. The proof of Theorem \ref{cong evidence}(iii) is thus reduced to the following result.
\begin{proposition} \label{solomon stark}One has ${\rm rec}_p^\ast (\eta_{L/K,S}^{e^-}(1))=(-1)^r\mathfrak{s}_{L/K,S}$ and
\begin{multline*}
{\rm rec}_p^\ast ( {\rm tw}_{0,1}(a))= \chi_{\rm cyc}(\tau_1\cdots \tau_r) H_{L/K,S,n-1}(a, \cdot)\\ \text{ in }\Hom_{\ZZ_p[G]^-}\left({\bigwedge}_{\ZZ_p[G]^-}^r U _{L_p}^{1,-},\ZZ/p^n[G]^- \right)^\#
\end{multline*}
for every $a\in {\bigcap}_{\ZZ_p[G]^+}^r H^1(\co_{L,S},\ZZ_p(1))^+={\bigcap}_{\ZZ_p[G^+]}^r \ZZ_p \co_{L^+,S}^\times$.
\end{proposition}
\begin{proof}
We review the definition of $\mathfrak{s}_{L/K,S}$. For an integer $i$ with $1\leq i \leq r$, we define
$$\log_p^{(i)}: U_{L_p}^{1,-} \to \overline \QQ_p$$
by $\log_p^{(i)}(u):=\log_p(j(\tau_i u))=\log_p(\iota_{w_i}(u)),$ where $\log_p$ denotes the $p$-adic logarithm defined on $\{a\in \overline \QQ_p \mid |a-1|_p<1\}$. We define
$${\rm Log}_p^{(i)}:=\sum_{\sigma \in G}\log_p^{(i)} (\sigma (\cdot ))\sigma^{-1} \in \Hom_{\ZZ_p[G]^-}(U_{L_p}^{1,-},\overline \QQ_p[G]^-).$$
Put
$$a_{L/K,S}^-:=\iota_{w_1}^{-1}\left( \left(\frac{\sqrt{-1}}{\pi}\right)^r e^-\theta_{L/K,S}^\ast(1) \right)\in \overline \QQ[G]^-.$$
We define $\mathfrak{s}_{L/K,S}$ by
$$
\mathfrak{s}_{L/K,S}:=j(a_{L/K,S}^-) \cdot {\rm Log}_p^{(1)} \wedge \cdots \wedge {\rm Log}_p^{(r)}
\in \Hom_{\ZZ_p[G]^-}\left({\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-},\overline \QQ_p[G]^-\right)^\#.
$$
Solomon proved that the image of $\mathfrak{s}_{L/K,S}$ lies in $\QQ_p[G]^-$, and that $\mathfrak{s}_{L/K,S}$ is independent of the choice of $j$ (see \cite[Prop. 2.16]{solomon}).
Now we prove the first assertion of the proposition.
By the definition of $\eta_{L/K,S}^{e^-}(1)$, it is sufficient to prove that the composition map
\begin{eqnarray}
{\bigwedge}_{\CC_p[G]^-}^r \CC_p Y_L(1) \stackrel{\sim}{\to} {\bigwedge}_{\CC_p[G]^-}^r (\CC_p \otimes_\QQ L)^- \stackrel{\exp_p}{\stackrel{\sim}{\to}} \CC_p {\bigwedge}_{\ZZ_p[G]^-}^r U_{L_p}^{1,-} \stackrel{(-1)^{r}\mathfrak{s}_{L/K,S}}{\to} \CC_p[G]^- \nonumber
\end{eqnarray}
coincides with $ e^- \theta_{L/K,S}^\ast(1) \cdot w_1(1)^\ast \wedge \cdots \wedge w_r(1)^\ast$ in $\Hom_{\CC_p[G]^-}(\bigwedge_{\CC_p[G]^-}^r \CC_p Y_L(1), \CC_p[G]^-)^\#$, where the first map is induced by
$$\CC_p Y_L(1)\stackrel{(\ref{beta})}{\simeq} \CC_p H_L(1)^+ \stackrel{(\ref{period map})}{\simeq} (\CC_p\otimes_\QQ L)^- ,$$
and the second by the $p$-adic exponential map. We denote by $\beta$ and $\alpha$ the maps induced by (\ref{beta}) and (\ref{period map}) respectively.
Noting that the equality $(-1)^{r}j(a_{L/K,S}^-)= (\pi \sqrt{-1})^{-r}e^-\theta_{L/K,S}^\ast(1)$ holds in $\CC_p[G]^-$ (via the isomorphism $\CC \simeq \CC_p$), we see that
$$(-1)^{r}\mathfrak{s}_{L/K,S}= (\pi \sqrt{-1})^{-r}e^-\theta_{L/K,S}^\ast(1) \cdot {\rm Log}_p^{(1)}\wedge \cdots \wedge {\rm Log}_p^{(r)}.$$
Hence, it is sufficient to prove that $ \iota_{w_i} \circ \alpha \circ \beta (w_i(1))= \pi \sqrt{-1} $, but this is straightforward to check by definition.
Thus we have proved the first assertion of the proposition.
To prove the second assertion we review the definition of $H_{L/K,S,n-1}$. In \cite[\S 2.3]{solomon}, Solomon defined a map $f_u \in \Hom_{\ZZ}(\co_{L^+,S}^\times,\ZZ/p^n)$ for any $u\in U_{L_p}^1$ as follows. For each $w\in S_p(L)$, we denote the Hilbert symbol
$$L_w^\times \times L_w^\times \to \mu_{p^n}; \ (x,y)\mapsto \frac{{\rm rec}_w(y) x^{1/p^n}}{x^{1/p^n}}$$
by $(x,y)_{w,n}$. The map $f_u$ is then defined by setting $f_u(a):=\sum_{w\in S_p(L)}\xi_1^\ast((a,u_w)_{w,n}),$
where $u_w$ is the $w$-component of $u \in U_{L_p}^1$ and $\xi_1^\ast$ is the isomorphism $\mu_{p^n} \stackrel{\sim}{\to} \ZZ/p^n$ sending $\xi_1$ to $1.$ (For the definition of $\xi_i$, see (\ref{def xi}).)
We define the ring isomorphism
$$\chi_{\rm cyc}^\# : \ZZ/p^n[G^+] \stackrel{\sim}{\to} \ZZ/p^n[G]^-; \ \sigma\mapsto 2^{-1} \sum_{\widetilde \sigma}\chi_{\rm cyc}(\widetilde \sigma)\widetilde \sigma^{-1},$$
where $\widetilde \sigma\in G$ runs over the lifts of $\sigma\in G^+$.
The pairing $H_{L/K,S,n-1}$ is defined by
$$H_{L/K,S,n-1}(a,u_1\wedge\cdots\wedge u_r):=2^r \chi_{\rm cyc}^\#((\widetilde F_{u_1}\wedge\cdots\wedge \widetilde F_{u_r} )(a)),$$
where $\widetilde F_{u_i}:=\sum_{\sigma \in G^+} \widetilde f_{u_i}(\sigma (\cdot))\sigma^{-1}$, and $\widetilde f_{u_i} \in \Hom_{\ZZ_p}(\co_{L^+,S}^\times,\ZZ_p)$ is a lift of $f_{u_i}$.
We compare Solomon's pairing $H_{L/K,S,n-1}$ with the twisting map
\begin{multline*}
{\rm tw}_{0,1}: {\bigcap}_{\ZZ_p[G^+]}^r \ZZ_p\co_{L^+,S}^\times ={\bigcap}_{\ZZ_p[G]^+}^r H^1(\co_{L,S},\ZZ_p(1))^+ \\
\to {\bigcap}_{\ZZ/p^n[G]^-}^r H^1(\co_{L,S},\ZZ/p^n)^-=\Hom_{\ZZ_p[G]^-} \left({\bigwedge}_{\ZZ_p[G]^-}^r \Gamma_{L,S}^-,\ZZ/p^n[G]^- \right)^\#.
\end{multline*}
For each $i$ with $1\leq i \leq r$, we define $h_i: \Gamma_{L,S} \to \Hom_\ZZ(\co_{L,S}^\times,\ZZ/p^n)$ by
$$h_i(\gamma)(a):=\xi_i^\ast\left( \frac{\gamma a^{1/p^n}}{a^{1/p^n}} \right),$$
where $\xi_i^\ast: \mu_{p^n}\stackrel{\sim}{\to} \ZZ/p^n$ is defined by $\xi_i \mapsto 1$. Put $H_i:=\sum_{\sigma \in G^+}h_i(\sigma(\cdot))\sigma^{-1}$. One checks that
$${\rm tw}_{0,1}(a)(\gamma_1\wedge \cdots \wedge \gamma_r)=2^r \chi_{\rm cyc}^\# ((\widetilde H_1 \wedge \cdots \wedge\widetilde H_r)(a)),$$
where $\widetilde H_i\in \Hom_{\ZZ[G^+]}(\co_{L^+,S}^\times,\ZZ[G^+])$ is a lift of $H_i$. Note also that for all $a\in \co_{L^+,S}^\times$ and $u \in U_{L_p}^{1,-}$ one has
\begin{multline*}
h_{i}({\rm rec}_p(u))(a)=\xi_i^\ast\left(\frac{{\rm rec}_p(u)a^{1/p^n}}{a^{1/p^n}}\right) = \xi_i^\ast\left(\prod_{w \in S_p(L)}\frac{{\rm rec}_w(u_w) a^{1/p^n}}{a^{1/p^n}}\right) \\
=\sum_{w\in S_p(L)}\xi_i^\ast((a,u_w)_{w,n}) = \sum_{w\in S_p(L)}\xi_1^\ast(\tau_i (a,u_w)_{w,n}) =\chi_{\rm cyc}(\tau_i)f_u(a). \end{multline*}
Hence we have
$${\rm rec}_p^\ast ( {\rm tw}_{0,1}(a))= \chi_{\rm cyc}(\tau_1\cdots \tau_r) H_{L/K,S,n-1}(a, \cdot) $$
in $\Hom_{\ZZ_p[G]^-}({\bigwedge}_{\ZZ_p[G]^-}^r U _{L_p}^{1,-},\ZZ/p^n[G]^-)^\#$, as required.
\end{proof}
|
\section{Introduction}
\label{sec:Introduction}
While flow problems are inherently unsteady, computer flow simulations have traditionally focused mainly on steady-state flow problems because they reduce the computational effort dramatically. Nevertheless, in many practical applications it is important to quantify the impact of unsteady flow phenomena on the forces and moments exerted on a body. These phenomena impact performance characteristics such as the lift and drag of a body, or the dynamic response of a control system. Historically, in aircraft design, these unsteady effects have required additional analyses to mitigate undesirable aeroelastic effects such as wing flutter and undesirable stall characteristics, among other issues \cite{ja:yurkovich2003flutter,ja:leerausch1996flutter, ja:niewoehner2005stall}. Sometimes the unsteady effects are beneficial, e.g., when using leading edge extensions (LEX) to improve high angle of attack performance \cite{cp:hirato2015lex}. Studies of low Reynolds number unsteady flows have become much more relevant today with the development of micro air vehicles (MAV) \cite{ja:nakata2011mav, ja:wang2011mav, ja:yang2013mav, cp:percin2012mav, ja:sun2002fruitfly, ja:huang1995lowRe,ja:choi2015lowRe}, with typical sizes as small as 15 cm. At these low Reynolds numbers viscous forces dominate the flow characteristics, leading to unsteady viscous effects such as laminar separation and von K\'{a}rm\'{a}n vortices, as commonly demonstrated in the flow over a cylinder \cite{bk:vonkarman1963aerodynamics}.
The use of CFD allows for preliminary analyses of these designs to determine whether any undesirable unsteady effects will be present, before committing to the expensive development and testing of a physical system. However, a major limitation of unsteady flow analysis using CFD is the prohibitive amount of computational time required to simulate a time-accurate solution with the time integration schemes commonly used to solve the Navier-Stokes equations. This coupled set of nonlinear partial differential equations has to be solved iteratively to determine the solution for each time step. A fine mesh resolution typically required to capture the length and time scales of the flow. For making the computations feasible it is necessary to use a time discretization scheme that maximizes convergence speed without prohibitively restricting the time steps due to stability constraints. It is also important that the time integration scheme provides an accurate solution at every time step.
Explicit time integration methods have been used for time-accurate solutions of unsteady flow problems due to their low computational cost per step and moderate memory requirements. For example, in \cite{kennedy_low-storage_2000} a number of embedded high-order explicit Runge-Kutta methods with minimal memory storage have been developed for the compressible Navier-Stokes equations based on van der Houwen's technique \cite{van1972explicit} for stage memory storage reduction. However, stability constraints restrict the maximum time steps that explicit methods can employ.
Implicit time-stepping methods have better stability properties than explicit methods, and therefore they can use very large time steps. However, the computational costs per step are also larger. The overall computational efficiency is given by the tradeoff between the computational cost per step and the total number of steps required to carry out the simulation.
The main cost of implicit methods is associated with solving a large system of nonlinear equations at each step \cite{Bijl_2002_NSimplicit}. Newton type methods for the solution of nonlinear systems are commonly used in the CFD literature in conjunction with preconditioned Krylov-based solvers for the inherent linear systems \cite{jothiprasad_higher_order_2003}. The popular Jacobian-free Newton Krylov (JFNK) methods employ finite difference approximations of the Jacobian-vector products required by Krylov solvers \cite{Birken_2012_timeDG}. Studies of JFNK methods applied to solve Navier-Stokes equations \cite{Qin_2000_preconditionNS} have shown that error tolerances of Krylov space solvers need to be carefully optimized for performance and accuracy.
There is considerable interest in developing numerical schemes that provide a suitable level of implicitness for time integration of stiff flow problems, such as to allow relatively large time steps while keeping the cost per time step comparable to that of explicit methods.
In this paper we study the efficiency of several different matrix-free, both explicit and implicit, time integration methods applied to computational fluid dynamics problems of moderate to large dimensions. In addition to standard techniques, we also examine a new class of lightly-implicit time integration schemes, called Rosenbrock-Krylov methods which are particularly well suited to employ approximate Jacobian-vector products.
The remaining part of this paper is structured as follows. Section \ref{sec:Time_Integration_CFD} reviews numerical methods accessible for the time integration of large systems of ordinary differential equations arising from flow problems. The numerical methods investigated here and their implementation are presented in Section \ref{sec:software}. Section \ref{sec:Num_Exp} applies these methods to a number of test problems and studies their effectiveness in terms of their numerical accuracy, stability, and computational efficiency in case of high dimensional problems. Conclusions and future work directions are discussed in Section \ref{sec:Conclusions}.
\section{Numerical time integration for CFD applications}
\label{sec:Time_Integration_CFD}
Consider the autonomous initial value problem:
\begin{equation}
\label{eqn:ode}
\frac{d\vec{y}}{dt} = \vec{f}(\vec{y}), \quad y(t_0) = y_0, \quad t_0 \le t \le t_F, \quad y(t) \in \mathbb{R}^N,
\quad f : \mathbb{R}^N \to \mathbb{R}^N.
\end{equation}
In this paper equation \eqref{eqn:ode} represents the system of ODEs resulting from the spatial semi-discretization of the Navier-Stokes equations for flow problems in the method-of-lines framework. The system is considered autonomous without loss of generality: any system can be written in autonomous form by appending the time variable to the solution vector. With only time derivatives remaining in equation \eqref{eqn:ode}, it is the choice of time-stepping method that determines the stability, accuracy, and efficiency of the numerical solution as the solution is propagated in time. This paper is concerned with the study of high-order implicit time marching schemes and their performance in large CFD applications.
We next review several important classes of numerical time integration algorithms.
\subsection{Runge-Kutta methods}
\label{sec:RKmethods}
The historically well-known time integration schemes attributed to Runge and Kutta are well-studied \cite{Butcher_1996_history,Enright_1994_RKsurvey} and extensively utilized in flow applications \cite{jorgenson1989explicit,Kennedy_2000_lowStorageNS}. Let $y_n \approx y(t_n)$ be a numerical approximation of the solution of the system \eqref{eqn:ode}. An $s$-stage Runge-Kutta method(advances the numerical solution to the next time step $t_{n+1}=t_{n}+h$ as follows:
\begin{subequations}
\label{eqn:RK}
\begin{align}
\vec{k}_i& = f\left( \vec{y}_n+ h \sum_{j=1}^{s} a_{i,j}\, \vec{k}_j \right ),\quad i=1,\dots,s; \label{eqn:RK_stage} \\
\vec{y}_{n+1}&= \vec{y}_n + h \sum_{j=1}^{s} b_j\, \vec{k}_j. \label{eqn:RK_y1}
\end{align}
\end{subequations}
The method coefficients
\[
\vec{a}={[a_{i,j}]}_{1 \leq i,j \leq s} \quad \vec{b}={[b_i]}_{1 \leq i \leq s} \quad \vec{c}={[c_i]}_{1 \leq i \leq s},
\]
are determined such that the method \eqref{eqn:RK} has the desired accuracy and stability properties \cite[II.1]{Hairer_book_I}.
Explicit Runge-Kutta (ERK) methods are characterized by coefficients $a_{i,j}=0$ for any $j \le i$. This means that each stage value $k_i$ \eqref{eqn:RK_stage} depends only on previously stage vectors $k_1,\dots,k_{i-1}$. This leads to the convenient result that explicit Runge-Kutta methods need only one ODE right-hand-side function evaluation per stage, and no linear or nonlinear systems of equations are solved in the process. The stability requirements due to CFL conditions limit the step size $h$, and therefore impact the efficiency of the method.
Singly Diagonally Implicit Runge-Kutta methods (SDIRK) \cite[IV.6]{Hairer_book_II} are characterized by coefficients $a_{i,j}=0$ for any $j < i$, and $a_{i,i}= \gamma > 0$ for all stages $i=1,\dots,s$. Solving for the stage vector $\vec{k}_i$ requires the solution of a nonlinear system of equations at each stage
\begin{align}
\vec{F}_i(\vec{k}_i)&= \vec{k}_i - f\left( \xi_i+ h \gamma \vec{k}_i \right )=0 \quad \text{for} \quad i=1, \dots, s , \label{eqn:Newton}
\end{align}
which makes the computational cost per step significantly larger than for ERK. However, this also leads to improved stability properties and the ability to use much larger time steps.
The nonlinear equation \eqref{eqn:Newton} is solved using Newton-type iterations:
\begin{equation}
\label{eqn:SDIRK_Newton}
\Delta \vec{k}_i^{\{\ell\}} = - {\left(\frac{\partial {F_i}}{\partial k_i} \right)}^{-1} \vec{F_i}\left(k_i^{\{\ell\}}\right),
\quad
\vec{k}_i^{\{\ell+1\}} = \vec{k}_i^{\{\ell\}} + \Delta \vec{k}_i^{\{\ell\}} , \quad \ell = 0,1,\dots
\end{equation}
where
\begin{equation}
\label{eqn:Newton_Jac_Matrix}
\frac{\partial F_i}{\partial k_i} =\mathbf{I}_N- h\,\gamma\, \mathbf{J}_n,
\end{equation}
and $\mathbf{J}_n$ is the Jacobian of the ODE right-hand-side function:
\begin{equation}
\label{eqn:ODE_Jacobian}
\mathbf{J}_n = \left. \frac{\partial f(y)}{\partial y} \right|_{y=y_n}.
\end{equation}
The fact that $a_{i,i}=\gamma$ for all stages allows re-using the LU decomposition of \eqref{eqn:Newton_Jac_Matrix} in the solution of linear systems appearing in equation \eqref{eqn:SDIRK_Newton} for all stage vectors $i=1,\dots,s$.
\subsection{Rosenbrock methods}
\label{sec:ROS}
Linearly implicit methods avoid the nonlinear systems \eqref{eqn:SDIRK_Newton} and solve only linear systems at each stage. One step of a Rosenbrock (ROS) method \cite[IV.7]{Hairer_book_II} reads:
\begin{subequations}
\label{eq:ROS}
\begin{align}
\vec{Y}_i &= \vec{y}_n+ h \sum_{j=1}^{i-1} a_{i,j} \,\vec{k}_j, \quad
\vec{k_i} = f\Bigl(\vec{Y}_i \Bigr) + h\, \mathbf{J}_n\, \sum_{j=1}^{i} \gamma_{i,j}\, \vec{k}_j, \quad i=1, \dots, s; \\
\vec{y_{n+1}}&= \vec{y_n} + h \sum_{i=1}^{s} b_j \vec{k}_j. \label{eqn:ROS_y1}
\end{align}
\end{subequations}
Therefore, the stage vectors $k_i$ are found in succession by solving linear systems of the form:
\begin{equation}
\label{eqn:ROS_stage_eq}
\left( \mathbf{I}_N - h\, \gamma_{i,i}\, \mathbf{J}_n \right)\, \vec{k}_i = f\left( \vec{Y}_i \right ) + h\, \mathbf{J}_n\, \sum_{j=1}^{i-1} \gamma_{i,j} \vec{k}_j, \qquad i=1, \dots, s.
\end{equation}
As with SDIRK methods, choosing $\gamma_{i,i} = \gamma > 0$ for $i=1,\dots,s$ helps reduce the computational costs by allowing to reuse the same LU factorization \eqref{eqn:Newton_Jac_Matrix} for all stages.
\subsubsection{Rosenbrock-Wanner methods}
\label{sec:ROW}
Rosenbrock methods simplify the computational effort necessary to solve the stage vector equations by limiting the implicitness to linear terms containing the exact Jacobian right-hand-side products \cite{Lubich_1995_linearlyImplicit}. As a consequence, the accuracy of the method depends on the availability of the exact Jacobian. In many practical cases an exact Jacobian is difficult to compute, however some approximation of the Jacobian may be available at reasonable computational cost. The Rosenbrock-Wanner (ROW) methods are Rosenbrock schemes that retain the order of accuracy for any matrix $\mathbf{A}_n$ used in place of the exact Jacobian $\mathbf{J}_n$ in \eqref{eq:ROS}. The preservation of accuracy is possible by imposing additional order conditions on the method coefficients \cite{Rang_2005_Wmethods}. Better approximations of the Jacobian $\mathbf{A}_n\approx \mathbf{J}_n$ will ensure better numerical stability. We note that while the formal definition for a ROW method is the same as in equation \eqref{eq:ROS}, the method coefficients are different due to the additional order conditions.
\subsubsection{Rosenbrock-Krylov methods}
\label{sec:ROK}
The stage vectors in Rosenbrock-type methods are computed by solving the linear system of dimension $N$ in equation \eqref{eqn:ROS_stage_eq}. For large problems the solutions of these linear systems is best obtained via a Krylov-space iterative linear algebra solver such as GMRES \cite{saad2003iterative}.
Instead of using a Krylov-based iterative solver such as GMRES, Rosenbrock-Krylov (ROK) methods developed in \cite{Sandu_2014_ROK} reformulate the method \eqref{eq:ROS} using implicitness only in the Krylov subspace of dimension $M$ constructed using modified Arnoldi iteration \cite{saad2003iterative}
\[
\mathcal{K}_M\big(\mathbf{J}_n, f(y_n)\big) = \textnormal{range}\{\mathbf{V_n} \}, \quad \mathbf{V_n} \in \mathbb{R}^{N \times M}, \quad \mathbf{V_n}^T\, \mathbf{V_n} = \mathbf{I}_M, \quad \mathbf{V_n}^T\, \mathbf{J}_n\, \mathbf{V_n} = \mathbf{H_n}.
\]
Here $\bf H$ and $\bf V$ are the upper Hessenberg and the orthogonal basis of the Krylov space, respectively, and are results of Arnoldi process.
A single time step of a ROK method is constructed as follows \cite{Sandu_2014_ROK}:
\begin{subequations}
\label{eq:ROK}
\begin{align}
F_i &= f \left( y_n +\displaystyle \sum_{j=1}^{i-1} \alpha_{i,j}k_j\right), \label{eq:ROK_stage_breakdown}\\
\phi_i &= \mathbf{V_n}^T \, F_i, \\
\lambda_i &= \left( \mathbf{I}_{M} - h\,\gamma\, \mathbf{H_n}\right)^{-1}\, \left( h\,\phi_i + h\, \mathbf{H_n} \,\displaystyle\sum_{j=1}^{i-1} \gamma_{i,j}\, \lambda_j\right), \label{eqn:ROK_Lambda_system}\\
k_i &= \mathbf{V_n}\, \lambda_i + h\, (F_i - \mathbf{V_n}\, \phi_i), \label{eqn:ROK_full_space_stage_vectors}\\
y_{n+1} &= y_n + \displaystyle\sum_{i=1}^s b_i\, k_i.
\end{align}
\end{subequations}
For $M \ll N$ the linear system \eqref{eqn:ROK_Lambda_system} is easily solvable using direct methods, and the stage vectors in full space can be recovered projecting the reduced space stage values back to full space\cite{Sandu_2014_ROK}. The minimum dimension of the Krylov space is determined by the desired order of the numerical scheme. Readers interested in the derivation of the order conditions for this method may refer to \cite{Sandu_2014_ROK}. In the extreme case $M=0$ the method \eqref{eq:ROK} reduces to an explicit Runge-Kutta method. In practice, however, we need the Krylov space to be large enough to capture some of the dominant eigenvalues of the full Jacobian, corresponding to fast-changing modes of \eqref{eqn:ode}, such as to alleviate the restrictions on step size of explicit methods imposed by the stiffness of the problem. An important question is whether the additional computational cost required by Arnoldi iteration can be compensated by the increases in step size due to the implicit nature of the method. The numerical results in Section \ref{sec:software} provide answers to this question.
\subsection{Matrix-free implementations}
\label{sec:matfree}
As discussed in Sections \ref{sec:RKmethods} and \ref{sec:ROS}, implicit time-stepping methods use the Jacobian matrix \eqref{eqn:ODE_Jacobian} to solve linear or nonlinear systems of equations at each step. The dimension of the state vector in \eqref{eqn:ode} may become significantly large when a highly refined spatial discretization is required, whether to capture fine details of flow or when CFD is used in large data-driven applications such as climate research. Computation and storage of Jacobian matrices, even in sparse form, is not favorable in such scenarios.
In some cases the complexity of the spatial discretization scheme impedes construction of analytic Jacobian matrices. In CFD codes such as SENSEI-Lite \cite{derlaga2013sensei}, the use of complex upwind flux schemes including Roe's and Van Leer's flux scheme \cite[V]{laney1998computational} in addition to MUSCL \cite{ja:vanleer1979limiters}reconstruction with flux limiters make the generation of an analytic Jacobian challenging. Furthermore, to allow for modularity as a research code, SENSEI-Lite allows these different flux schemes and limiters to be used interchangeably, dependent upon the test problem configuration. This interchangeability would necessitate the formulation of an analytic Jacobian for every possible configuration, eliminating the modularity of the code structure. The framework of matrix-free methods allows us to exploit the benefits of advanced time-stepping methods without forming the Jacobian matrix directly.
Krylov space iterative methods for solving the stage equations \eqref{eqn:SDIRK_Newton} or \eqref{eqn:ROS_stage_eq} rely on computing Jacobian-vector products. Instead of computing the Jacobian matrix and then multiplying, it is possible to approximate directly Jacobian-vector products using the finite-difference approximation of a directional derivative:
\begin{equation}
\label{eqn:matvecFD}
\mathbf{J}_n \cdot v= \left.\frac{\partial f(y)}{\partial y}\right|_{y_n}\cdot v \approx \frac{f(y_n+\varepsilon v)-f(y_n)}{\varepsilon}.
\end{equation}
The optimum $\varepsilon$ is chosen considering the trade-off between truncation and round-off errors \cite{KnollKeyes}. Higher-order approximations are not favorable here as they require more right-hand-side function evaluations, which, considering the large dimensions of the problem, are costly to compute.
A more in-depth analysis of different strategies to compute Jacobian-vector products, and their effects on convergence and efficiency of implicit time integration methods, can be found in \cite{Tranquilli2016}. Of special interest is using exact Jacobian-vector products instead of finite difference approximations. Numerical experiments applied to discretizations of PDEs in \cite{Tranquilli2016} indicate that exact Jacobian-vector products provide more robustness in observed convergence orders, and increased runtime efficiency over methods using approximate products \eqref{eqn:matvecFD}.
\section{The software infrastructure for numerical investigations}
\label{sec:software}
\subsection{Time-stepping schemes and their implementation}
The time integration software used in the numerical experiments is \textsc{matlode} \cite{matlodeWeb}, a Matlab library for integration of ODE systems including implicit and explicit Runge-Kutta methods, Rosenbrock methods and Krylov based methods. The package also supports forward, adjoint, and tangent linear models, enabling sensitivity analysis applications. Aside from the methods available in \textsc{matlode} package, Matlab's explicit time-stepping scheme based on Dormand and Prince \cite{Dormand80} is also included in tests for comparison. Table \ref{tab:numerical_methods_overview} summarizes the methods used in numerical experiments and their properties.
\begin{table}[htb]
\centering
\caption{Overview of time stepping methods used in numerical experiments. }
\label{tab:numerical_methods_overview}
\begin{tabular}{llccl}
\hline
\rowcolor[HTML]{EFEFEF}
Method & Family & Stages & Order & Stability \\ \hline
ERK & Explicit Runge-Kutta & 5 & 4 & Conditionally stable \\
DOPRI5 & Explicit Runge-Kutta & 7 & 5 & Conditionally stable \\
DOPRI853 & Explicit Runge-Kutta & 12 & 8 & Conditionally stable \\
SDIRK & Implicit Runge-Kutta & 5 & 4 & L-stable \\
ROS4 & Rosenbrock & 4 & 4 & L-stable \\
ROW & Rosenbrock-W & 4 & 3 & L-stable \\
ROK & Rosenbrock-Krylov & 5 & 4 & Conditionally stable \\
ODE45 & Explicit Runge-Kutta & 5 & 4 & Conditionally stable \\ \hline
\end{tabular}
\end{table}
\subsection{The fluid flow simulation code}
SENSE-Lite, the CFD code employed in this paper, can solve both the Euler and Navier-Stokes equations for compressible flow \cite{derlaga2013sensei}:
\begin{subequations}
\label{eqn:Nav_Stokes}
\begin{align}
\frac{\partial \rho}{\partial t} +
\frac{\partial \rho v_1}{\partial x_1} + \frac{\partial \rho v_2}{\partial x_2} &= 0 , \\
\frac{\partial \rho v_1}{\partial t} +
\frac{\partial \rho v_1^2}{\partial x_1} + \frac{\partial \rho v_1 v_2}{\partial x_2} + \frac{\partial p}{\partial x_1} &= \frac{\partial \tau_{11}}{\partial x_1} + \frac{\partial \tau_{12}}{\partial x_2} ,\\
\frac{\partial \rho v_2}{\partial t} +
\frac{\partial \rho v_1 v_2}{\partial x_1} + \frac{\partial \rho v_2^2}{\partial x_2} + \frac{\partial p}{\partial x_2} &= \frac{\partial \tau_{12}}{\partial x_1} + \frac{\partial \tau_{22}}{\partial x_2} ,\\
\frac{\partial E_t}{\partial t} +
\frac{\partial \rho v_1 E_t}{\partial x_1} + \frac{\partial \rho v_2 E_t}{\partial x_2} + \frac{\partial v_1 p}{\partial x_1} + \frac{\partial v_2 p}{\partial x_2} &= \notag \\
k\,\left(\frac{\partial q_1}{\partial x_1} + \frac{\partial q_2}{\partial x_2}\right) + \mu\, \Big(\frac{\partial}{\partial x_1} (v_1 \tau_{11} + v_2 \tau_{12}) &+ \frac{\partial}{\partial x_2} (v_1 \tau_{12} + v_2 \tau_{22}) \Big) ,
\end{align}
where
\begin{align}
q_i &= -k\, \frac{\partial T}{\partial x_i} , \\ \tau_{i,j} &= \mu \,\left(\frac{\partial v_j}{\partial x_i} + \frac{\partial v_i}{\partial x_j}\right) - \frac{2}{3}\, \mu \, \left(\vec{\triangledown} \cdot \vec{v}\right)\, \delta_{i,j}.
\end{align}
\end{subequations}
SENSEI-Lite uses a curvilinear, structured-grid, finite volume method. Second-order spatial accuracy is achieved through a standard total variation diminishing scheme consisting of MUSCL reconstruction and selectable flux limiters \cite{ja:vanleer1979limiters}. The code is written in C++ and MEX interfaces are used to call it from within time integrators implemented in Matlab. The primary function of the MEX code is to return the spatial residual for a given solution state; this is returned as a vector that is independent of any temporal information and can be used for building arbitrary time integration methods such as multi-stage ERK. This residual vector can optionally be returned as multiple vectors split according to the underlying equations, e.g., viscous and inviscid contributions from the Navier-Stokes equations as shown in \eqref{eqn:Nav_Stokes}; these vectors sum to the full residual and can also be used to integrate the equations independently. The Matlab client code is responsible for storing multiple solution state and update vectors as necessary, and can apply time-dependent source terms if desired.
\section{Numerical results}
\label{sec:Num_Exp}
\subsection{Experimental setting}
\label{sec:Setting}
This section details the numerical experiments setup using \textsc{matlode} and SENSEI-Lite packages to study the performance of matrix-free time-stepping methods on unsteady flow problems. A reference solution is computed and stored by integrating the model using an explicit method with tight tolerances ($\sim 10^ {-9}$) on errors. In each numerical experiment, the solution at the final integration time is compared against this reference solution, and the error is measured using the $\mathcal{L}_2$ norm:
\begin{align*}
Error = \left\Vert{y_N - y_{\rm ref}}\right\Vert_2 \qquad \text{where} \qquad y=\left[ \rho , \rho v , \rho u, E_t \right]^T
\end{align*}
Where $y_N$ and $y_{ref}$ are numerical and reference state vectors at the final integration time respectively.
All experiments use matrix-free time-stepping methods. For each test problem we evaluate the eigenvalues of the Jacobian matrix in equation \eqref{eqn:ODE_Jacobian} to get information about the dynamic modes of the state variable evolution. A wide spread of the eigenvalues indicates the existence of both slow and fast dynamics, in other words, of ``stiff'' dynamics. The largest eigenvalue gives an estimate of the largest stable step-size in explicit methods. Eigenvalues are computed using Matlab's implementation of the implicitly restarted Arnoldi method that estimates the largest 1000 eigenvalues in magnitude for each test problem; this computation is also performed in matrix-free form.
\subsection{Vortex-shedding cylinder test problem}
\label{sec:Cylinder_Intro}
The vortex-shedding cylinder test problem consists of a two-dimensional circular cylinder in a low subsonic flow (Mach 0.1) using a gas model for air at 5,000 ft altitude standard atmospheric conditions (278K, 84.31kPa). The viscosity for air is calculated based on local flow conditions using Sutherland's law. The modeled is a free-stream flow, using far-field boundary conditions set at over 100 chord lengths away from the surface of the cylinder to minimize interactions with the boundary. The default cylinder diameter is $8 \times 10^{-5}\, m$, which yields a Reynolds number of approximately 200 and results in a steady and predictable two-dimensional shedding of alternating vortices behind the cylinder. At this low Reynolds number there are no sub-grid-scale turbulence effects, so all physically accurate spatial and temporal scales in the solution can be directly modeled; therefore, the CFD code can solve the laminar Navier-Stokes equations with no underlying turbulence model. For all test problems, the flux scheme used is Roe's approximate Riemann solver and no flux limiter is employed in the MUSCL scheme as flux limiters were observed to be unnecessary and to generally reduce the accuracy of the FVM reconstruction in the continuous and smooth flow field around the cylinder. The parameters of this problem are modified to provide different tests, as summarized in Table \ref{tab:cylinder_param}. Care is taken so that the qualitative solution behavior (and appropriate Reynolds number) is maintained for all tests.
{\renewcommand{\arraystretch}{1.2}
\begin{table}[htp]
\centering
\caption{Parameters of the vortex-shedding cylinder test problems.}
\label{tab:cylinder_param}
\begin{tabular}{|l|l|l|l|}
\hline
\multicolumn{1}{|c|}{Parameter} & \multicolumn{1}{c|}{Description} & Test problem 1 & Test problem 2 \\
& & \multicolumn{1}{c|}{value} & \multicolumn{1}{c|}{value} \\\hline
$\rho (kg/m^3)$ & Reference fluid density & 1.0565 & 1.0565 \\
$S$ & Sutherland's coefficient & 1.45E-6 & 2.9E-6 \\
$T (K)$ & Temperature & 278 & 278 \\
$v (m/s)$ & Reference velocity & 340 & 340 \\
$L (m)$ & Diameter & 8E-5 & 8E-5 \\
$Re$ & Reynold's number & 165.90 & 82.95 \\ \hline
\end{tabular}
\end{table}
The first experiment is performed on the vortex-shedding cylinder test problem 1 with parameters given in Table \ref{tab:cylinder_param}. %
Figure \ref{fig:cyl_flow} illustrates snapshots of the density component of the flow for cylinder test problem 1 at different times, showing the cyclic development of vortices behind the cylindrical object.
\begin{figure}
\centering
\includegraphics[width =\textwidth]{./images/cylinder_flow.pdf}
\vspace*{6px}
\includegraphics[width =\textwidth]{./images/Long.pdf}
\caption{Snapshots of density component of the flow for vortex-shedding cylinder test problem 1 at different times.}
\label{fig:cyl_flow}
\end{figure}
This experiment uses fixed step sizes to study the temporal orders of convergence for each method, and the results are plotted in Figure \ref{fig:cyl_order}. Numerical orders of convergence calculated for each method are reported in Table \ref{tab:Cyl_conv}. One notable observation is the significantly lower numerical order of convergence for SDIRK method as a result of poor convergence of the Newton iteration, especially in the absence of preconditioners for the solution of linear systems. Readers interested in numerical experiments on a smaller problem that verify the theoretical order of convergence may consult the Appendix.
\begin{figure}
\centering
\includegraphics[width=0.5 \linewidth]{./images/CylConv.pdf}
\caption{Convergence diagrams for the matrix-free methods of Table \ref{tab:numerical_methods_overview} using fixed step sizes over integration window $T=[0,10^{-7}]$ second for cylinder test problem 1.}
\label{fig:cyl_order}
\end{figure}
\begin{table}[h]
\centering
\caption{Numerical orders of convergence for various methods applied to the cylinder test problem 1.}
\label{tab:Cyl_conv}
\begin{tabular}{lcc}
\hline
Method & Numerical order & Theoretical order\\
\hline
ERK4 & 4.02 & 4\\
ERK5 (DOPRI5) & 5.29 & 5\\
ERK5 (DOPRI853) & 5.97 & 8 \\
SDIRK & 2.94 & 4 \\
ROS4 & 3.11 & 4 \\
ROW & 2.96 & 3 \\
ROK & 3.85 & 4 \\
ODE45 & 5.39 & 5\\
\hline
\end{tabular}
\end{table}
The cylinder test problem 2 with parameters given in Table \ref{tab:cylinder_param}, uses a different value for Sutherland's Law coefficient that translates into increased kinetic viscosity compared to vortex-shedding cylinder test problem 1. Figure \ref{fig:cyl_eigs} shows the numerical approximation of the first 1000 eigenvalues for the two test problems. As indicated in Figure \ref{fig:cyl_eigs} the more viscous test problem shows larger negative eigenvalues, therefore, we expect stricter stability bounds on the step sizes for this test problem. For both problems the cluster of eigenvalues with the largest negative real parts consists of only a limited number of modes. This fact becomes relevant when we consider ROK methods that use a reduced order Jacobian for implicit integration. Throughout the numerical experiments the ROK method employs a reduced space of dimension four, unless otherwise specified.
\begin{figure}[htbp]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/cyl_def_eigs.pdf}
\caption{Cylinder test problem 1}
\label{fig:cyl_eigs_test-1}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/cyl_2x_eigs.pdf}
\caption{Cylinder test problem 2}
\label{fig:cyl_eigs_test-2}
\end{subfigure}
\caption{Eigenvalue distribution of the Jacobians of the right-hand-side functions for vortex-shedding cylinder test problems.}
\label{fig:cyl_eigs}
\end{figure}
The stability of a method is related to its choice of timesteps in the adaptive time stepping framework. Figure \ref{fig:cyl_time_steps} compares the step sizes of fourth order explicit Runge-Kutta method (ERK) to Rosenbrock-Krylov method of the same order for two different error tolerances. We can verify that the step sizes for the explicit method quickly reaches the upper bound set by the stability constraints regardless of the accuracy tolerance chosen for the method. On the other hand, by implicitly treating some of the stiff modes, the ROK method is able to achieve stable numerical integration for larger step sizes as is clear from Figures \ref{fig:cyl_def_ROK_time_steps} and \ref{fig:cyl_2x_suth_ROK_time_steps}. Furthermore, we notice that the step sizes scale well relative to the required accuracy.
\begin{figure}[htbp]
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/ERK_def_steps.pdf}
\caption{ERK on cylinder test problem 1}
\label{fig:cyl_def_ERK_time_steps}
\end{subfigure}
~~
\begin{subfigure}{0.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/ROK_def_steps.pdf}
\caption{ROK on cylinder test problem 1}
\label{fig:cyl_def_ROK_time_steps}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/ERK_2x_steps.pdf}
\caption{ERK on cylinder test problem 2}
\label{fig:cyl_2x_suth_ERK_time_steps}
\end{subfigure}
~~
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/ROK_2x_steps.pdf}
\caption{ROK on cylinder test problem 2}
\label{fig:cyl_2x_suth_ROK_time_steps}
\end{subfigure}
\caption{Adaptive time steps taken by the explicit Runge-Kutta method and by the ROK method for vortex shedding cylinder test problems. }
\label{fig:cyl_time_steps}
\end{figure}
Figure \ref{fig:cyl_M4_steps_timing} shows the Work-precision diagrams for the vortex-shedding cylinder test problems. Integration is performed over time window $T=[0, 2 \times 10^{-6}]$ seconds. These results lead to the following conclusions:
\begin{itemize}
\item As the adaptive time stepping method uses tighter tolerances, the integrator takes smaller steps leading to an increased number of total steps. This is the case for all integration methods presented here. However, the total number of steps for explicit methods does not change considerably for a wide range of tolerances, an effect observable in Figure \ref{fig:cyl_M4_steps_timing} where the lines for explicit methods are nearly vertical. This is a result of the fact that for stiff problems the adaptive time steps are bounded by stability requirements rather than by accuracy constraints.
\item Implicit methods are able to take fewer steps when the required solution tolerances are low, due to their improved stability properties. This is the case for ROS, ROW and SDIRK methods in Figure \ref{fig:cyl_def_M4_steps} and \ref{fig:cyl_2x_suth_M4_steps}. However, inspecting the runtime diagrams on Figures \ref{fig:cyl_def_M4_timing} and \ref{fig:cyl_2x_suth_M4_timing} reveals that the the increased cost of these methods makes them considerably less efficient.
\item The effect of stiffness of the problem can also be seen in the timing reported in Figures \ref{fig:cyl_def_M4_timing} and \ref{fig:cyl_2x_suth_M4_timing}. We notice that while the explicit methods take about the same amount of time to complete the integration for all choices of solution tolerance, the runtime scales better with tolerances for implicit methods.
\item ROK is the most efficient method for the cylinder test problems for error tolerances below $10^{-6}$. This is an indication that ROK is able to capture sufficiently many stiff components in its Krylov subspace, and that by treating them implicitly the method is able to take larger time steps.
\end{itemize}
\begin{figure}[htbp]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/cyl_def_M4_steps.pdf}
\caption{Convergence diagram, cylinder test problem 1}
\label{fig:cyl_def_M4_steps}
\end{subfigure}
~~
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/cyl_def_M4_timing.pdf}
\caption{Work-precision diagram, cylinder test problem 1}
\label{fig:cyl_def_M4_timing}
\end{subfigure}
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/cyl_2x_M4_steps.pdf}
\caption{Convergence diagram, cylinder test problem 2}
\label{fig:cyl_2x_suth_M4_steps}
\end{subfigure}
~~
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/cyl_2x_M4_timing.pdf}
\caption{Work-precision diagram, cylinder test problem 2}
\label{fig:cyl_2x_suth_M4_timing}
\end{subfigure}
\caption{Relative performance of different integration methods applied to the cylinder test problems.}
\label{fig:cyl_M4_steps_timing}
\end{figure}
\subsection{Flow over \textsc{NACA0012} wing test problem }
\label{sec:naca_Intro}
An analysis of unsteady flow over a \textsc{NACA0012} airfoil is performed in addition to the cylinder test problem. Similar free-stream flow conditions are applied to this test problem, using atmospheric conditions at 5,000 ft. The Mach number is increased to 0.25; however, this flow is still within the subsonic flow regime. The chord length of this symmetric airfoil geometry is set to 0.001 meters, producing flow with a Reynolds number close to 5,000. Experimental analysis suggests that for the \textsc{NACA0012} airfoil, a free-stream flow at this Reynolds number produces predominantly laminar flow over the airfoil and in its wake \cite{ja:huang1995lowRe}, allowing for turbulent effects to be neglected in this analysis. At an angle of attack of 15 degrees, vortices shed into the wake due to laminar separation of the flow over the upper surface of the airfoil generating an unsteady flow solution. As with the cylinder case, Roe's flux scheme is used with no flux limiter.
Figure \ref{fig:naca_flow} shows snapshots of the density field at different time moments. The distribution of Jacobian eigenvalues is shown in Figure \ref{fig:naca_eigs}, and indicates that a large number of fast eigenvalues are clustered together. The performance diagrams in Figure \ref{fig:naca3_M4_steps_timing} indicate that the fastest methods here are Rosenbrock and ROW methods, until ROK and SDIRK become most efficient for errors below $10^{-4}$. This test problem converges to a steady state solution, and this particular dynamics favors the implicit Rosenbrock methods, since the fast modes need only to be only damped out and not to be accurately solved.
\begin{figure}
\centering
\includegraphics[width = \textwidth]{./images/naca_flow.pdf}
\caption{Snapshots of density component of the flow for \textsc{NACA0012} at different times.}
\label{fig:naca_flow}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.7 \linewidth]{./images/eigs_naca_zoomed.pdf}
\caption{Eigenvalue distribution of the Jacobian for the \textsc{NACA0012} test problem.}
\label{fig:naca_eigs}
\end{figure}
\begin{figure}[htbp]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/naca3_M4_steps.pdf}
\caption{Convergence diagram}
\label{fig:naca3_M4_steps}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/naca3_M4_timing.pdf}
\caption{Work-precision diagram}
\label{fig:naca3_M4_timing}
\end{subfigure}
\caption{Relative performance of different integration methods applied to the \textsc{NACA0012} test problem.}
\label{fig:naca3_M4_steps_timing}
\end{figure}
\subsection{Vortex shedding cylinder with iso-thermal conditions test problem}
The final test problem uses the vortex shedding cylinder problem 2, with iso-thermal conditions at cylinder boundaries for a wall temperature of $33^{\circ} K$. Among each class of explicit and implicit integrators we have selected the fastest representative methods. We have also chosen three variations of ROK method with 4, 8, and 12 Krylov basis vectors, respectively. The results in Figure \ref{fig:cool_ring_steps_timing} demonstrate that Krylov methods retain their computational superiority for a wide range of error tolerances. Furthermore, we observe that adding more basis vectors to the Krylov subspace requires extra cost (Figure \ref{fig:cooled_ring_timing}) and in turn increases step sizes slightly (Figure \ref{fig:cooled_ring_steps}), but ultimately the most efficient method is the one with minimum basis size, i.e. 4 vectors. Finally, it is worthwhile to point out that once the adaptive error controller of the step size reaches the GMRES tolerance, the error of JFNK methods does not decrease any further. This is seen in Figure \ref{fig:cool_ring_steps_timing} forthe Rosenbrock-W method where the error curve flattens at an error level of about 5E-4.
\begin{figure}[htbp]
\centering
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width=\linewidth]{./images/cool_M4_steps.pdf}
\caption{Convergence diagram}
\label{fig:cooled_ring_steps}
\end{subfigure}%
\begin{subfigure}{.45\textwidth}
\centering
\includegraphics[width= \linewidth]{./images/cool_M4_timing.pdf}
\caption{Work-precision diagram}
\label{fig:cooled_ring_timing}
\end{subfigure}
\caption{Relative performance of different integration methods applied to cooled cylinder test problem with $T_\textrm{wall} = 33^{\circ} K$.}
\label{fig:cool_ring_steps_timing}
\end{figure}
\section{Conclusions}
\label{sec:Conclusions}
This paper studies the performance of several types of high-order matrix-free time stepping methods when applied to solve unsteady flow problems. The methods under consideration are explicit Runge-Kutta, diagonally implicit Runge-Kutta and Rosenbrock schemes paired with iterative linear algebra solvers, and Rosenbrock-Krylov schemes. All implementations of implicit methods are matrix-free where the necessary Jacobian-vector products are approximated by finite differences; this is a typical setting for solving large-scale CFD applications.
Favorable properties of explicit methods include their easier implementation overhead and the low computational cost per step. As expected, our numerical experiments show that the overall performance of explicit methods deteriorates quickly in the presence of mild stiffness, and for such problems they are not competitive with implicit methods.
Traditional implicit methods such as SDIRK and Rosenbrock can take large steps on stiff problems, as expected. However, their overall performance depends on how well the underlying linear systems are solved at each step. We observed that inaccurate linear solutions lead to loss of convergence, and that in absence of well-tuned linear algebra preconditioners the computational costs of matrix-free implicit methods are quite large.
Rosenbrock-Krylov methods are the most effective when applied to flow problems with a limited number of stiff modes. These methods are a suitable choice when exact full Jacobians are not available, as is the case in large CFD problems. The order conditions theory of Krylov-based methods accounts for the errors associated with linear algebra; indeed, our numerical experiments confirm that they show full order where other implicit methods suffer from reduced temporal convergence caused by inexact Jacobian approximations or by poorly converged linear system solutions. On the other hand, as the test problems become more stiff, Krylov-based methods lose their performance superiority due to the increased computational cost of creating a large Krylov basis required for numerical stability.
The numerical experiments in this study were performed on a 2D Navier-Stokes problem. Extension of the investigation to 3D unsteady flow problems is a future direction to pursue. Furthermore, finding automatic strategies for the cost-effective construction of Krylov basis is an important question for the advancement of Krylov-based family of methods.
\section{Acknowledgments}
This work was funded by the Air Force Office of Scientific Research (AFOSR) computational mathematics program via grant No. FA9550-12-1-0442 and by Computational Science Laboratory (CSL) in the Department of Computer Science at Virginia Tech.
\section{Appendix: Order reduction with matrix-free methods }
\label{sec:apex}
In this section we confirm the orders of convergence for the time-stepping methods used in this paper by applying them to Lorenz--96 test problem \cite{lorenz1996predictability}. Proposed by Edward Lorenz, this model is often used to model chaotic behavior of atmospheric systems:
\begin{align*}
\frac{\partial u_i} {\partial t} = \left( u_{i+1} - u_{i-2} \right) \, u_{i-1} - u_i + F \qquad \text{for} \quad i=1, \cdots, 40.
\end{align*}
The Jacobian of this system is a banded matrix that can be implemented in sparse format and used in the implicit methods of Section \ref{sec:Time_Integration_CFD}. Setting the external force factor $F=8$ and using a range of fixed time steps to propagate the model forward, we observe nearly full order for all of the methods as reported in Table \ref{tab:Lorenz--96_conv}. One source of local error causing order reduction can be traced back to the truncation errors made by replacing the Jacobian-vector product in equation \eqref{eqn:matvecFD} with a first-order finite difference approximation. Poor convergence of the Newton's iteration as well as the Krylov-based solver for the linear system in the implicit methods are other sources of error contributing to inexact stage vectors and ultimately to order reduction. It is notable, however, that in all the convergence tests the Rosenbrock-Krylov method retains its full order of convergence. Unlike fully and linearly implicit methods, the system solved in equation \eqref{eqn:ROK_Lambda_system} is formed using the Arnoldi iteration with an inexact Jacobian from the beginning and is solved using a direct method and therefore the issues arising from the convergence of the iterative solvers is avoided altogether. Interested readers may consult \cite{Sandu_2014_ROK} for a detailed analysis of this phenomenon.
\begin{table}[h]
\centering
\caption{Orders of convergence for methods applied to the Lorenz--96 problem.}
\label{tab:Lorenz--96_conv}
\begin{tabular}{lcc}
\hline
Method & Numerical order & Theoretical order\\
\hline
ERK4 & 4.00 & 4 \\
ERK5 (DOPRI5) & 5.58 & 5 \\
ERK5 (DOPRI853) & 6.14 & 8 \\
SDIRK & 3.89 & 4 \\
ROS4 & 3.91 & 4 \\
ROW & 2.94 & 3 \\
ROK & 3.85 & 4\\
ODE45 & 5.47 & 5 \\
\hline
\end{tabular}
\end{table}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5 \linewidth]{./images/LorenzConv.pdf}
\caption{Convergence diagram for the Lorenz--96 test problem.}
\label{fig:Conv_Lorenz}
\end{figure}
\section*{References}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.